Analysis in Nutrition Research: Principles of Statistical Methodology and Interpretation of the Results 0128145560, 9780128145562

Table of contents :
Content: Design of observational nutrition studies / George Pounis --
Study design in experimental settings / Monica Dinu, George Pounis and Francesco Sofi --
Collection and management of dietary data / Pauline M. Emmett, Louise R. Jones, Kate Northstone, George Pounis, Caroline M. Taylor --
Dietary Pattern Analysis / Claudia Agnoli, George Pounis and Vittorio Krogh --
Statistical analysis of retrospective health and nutrition data / George Pounis --
Statistical analysis of prospective health and nutrition data / George Pounis --
Meta-analysis of nutrition studies / Emmanouil Bouras, Konstantinos K. Tsilidis, George Pounis, Anna-Bettina Haidich --
Principles of research publication / Gregory S. Patience, George Pounis, Paul A. Patience, Daria C. Boffito --
Mediterranean diet: a health-protective dietary pattern for modern times / Dimitra Mastorakou, Mikael Rabaeus, Patricia Salen, George Pounis, Michel de Lorgeril --
Polyphenol-rich diets in cardiovascular disease prevention / Junichi Sakaki, Melissa Melough, Sang Gil Lee, George Pounis, Ock K. Chun --
Hydration and health / Adam D. Seal, Hyun-Gyu Suh, Lisa T. Jansen, LynnDee G. Summers, Stavros A. Kavouras --
Diet, healthy aging, and cognitive function / Krasimira Aleksandrova, George Pounis and Romina di Giuseppe --
Diet and bone health / Kate Maslin and Elaine Dennison --
Diet and lung health / Foteini Malli, Themis Koutsioukis, George Pounis, Konstantinos I. Gourgoulianis.

Citation preview

Analysis in Nutrition Research

This page intentionally left blank

Analysis in Nutrition Research Principles of Statistical Methodology and Interpretation of the Results

Edited by

George Pounis Nutrition Consultant in the Greek Food Industry, Athens, Greece

Academic Press is an imprint of Elsevier 125 London Wall, London EC2Y 5AS, United Kingdom 525 B Street, Suite 1650, San Diego, CA 92101, United States 50 Hampshire Street, 5th Floor, Cambridge, MA 02139, United States The Boulevard, Langford Lane, Kidlington, Oxford OX5 1GB, United Kingdom Copyright © 2019 Elsevier Inc. All rights reserved. No part of this publication may be reproduced or transmitted in any form or by any means, electronic or mechanical, including photocopying, recording, or any information storage and retrieval system, without permission in writing from the publisher. Details on how to seek permission, further information about the Publisher’s permissions policies and our arrangements with organizations such as the Copyright Clearance Center and the Copyright Licensing Agency, can be found at our website: www.elsevier.com/permissions. This book and the individual contributions contained in it are protected under copyright by the Publisher (other than as may be noted herein). Notices Knowledge and best practice in this field are constantly changing. As new research and experience broaden our understanding, changes in research methods, professional practices, or medical treatment may become necessary. Practitioners and researchers must always rely on their own experience and knowledge in evaluating and using any information, methods, compounds, or experiments described herein. In using such information or methods they should be mindful of their own safety and the safety of others, including parties for whom they have a professional responsibility. To the fullest extent of the law, neither the Publisher nor the authors, contributors, or editors, assume any liability for any injury and/or damage to persons or property as a matter of products liability, negligence or otherwise, or from any use or operation of any methods, products, instructions, or ideas contained in the material herein. Library of Congress Cataloging-in-Publication Data A catalog record for this book is available from the Library of Congress British Library Cataloguing-in-Publication Data A catalogue record for this book is available from the British Library ISBN: 978-0-12-814556-2 For information on all Academic Press publications visit our website at https://www.elsevier.com/books-and-journals

Publisher: Charlotte Cockle Acquisition Editor: Megan Ball Editorial Project Manager: Michelle Kublis Production Project Manager: Paul Prasad Chandramohan Cover Designer: Christian Bilbow Typeset by TNQ Technologies

Contents List of Contributors ............................................................................................................................. xv Preface ................................................................................................................................................ xix Acknowledgments............................................................................................................................... xxi

PART 1 ANALYSIS IN NUTRITION RESEARCH CHAPTER 1 Design of Observational Nutrition Studies........................................... 3 George Pounis 1.1 Introduction to Observational Nutrition Studies ........................................................3 1.2 Ecological Nutrition Studies.......................................................................................5 1.2.1 Description........................................................................................................ 5 1.2.2 Challenges ........................................................................................................ 5 1.2.3 Example of Ecological Study on Diet and Cancer.......................................... 6 1.3 Cross-Sectional Nutrition Studies ..............................................................................6 1.3.1 Description........................................................................................................ 6 1.3.2 Sampling........................................................................................................... 8 1.3.3 Challenges ...................................................................................................... 10 1.3.4 Example of the National Health and Nutrition Examination Survey ........... 11 1.4 CaseeControl Nutrition Studies...............................................................................12 1.4.1 Description...................................................................................................... 12 1.4.2 Sampling......................................................................................................... 12 1.4.3 Challenges ...................................................................................................... 14 1.4.4 An Example in the Study of Hodgkin Lymphoma........................................ 15 1.5 Cohort Nutrition Studies...........................................................................................15 1.5.1 Description...................................................................................................... 15 1.5.2 Sampling......................................................................................................... 16 1.5.3 Challenges ...................................................................................................... 17 1.5.4 Example of the Nurses’ Health Study ........................................................... 18 References........................................................................................................................ 19

CHAPTER 2 Study Design in Experimental Settings .............................................. 23 Monica Dinu, George Pounis and Francesco Sofi 2.1 Introduction...............................................................................................................24 2.2 Experimental Designs ...............................................................................................24 2.2.1 Single-Arm Studies ........................................................................................ 25 2.2.2 Parallel Studies ............................................................................................... 25 2.2.3 Crossover Studies ........................................................................................... 27 2.3 Planning a Dietary Intervention ...............................................................................29 2.3.1 Definition of the Research Question, Hypothesis, and Main Objectives...... 30

v

vi

CONTENTS

2.3.2 Primary and Secondary Outcomes................................................................. 30 2.3.3 Study Population ............................................................................................ 30 2.3.4 Selection of the Study Design........................................................................ 35 2.3.5 Measurements................................................................................................. 35 2.3.6 Study Protocol ................................................................................................ 35 2.4 Conducting the Dietary Intervention........................................................................37 2.4.1 Recruitment and Screening of Participants.................................................... 37 2.4.2 Compliance of Participants ............................................................................ 37 2.4.3 Data Management........................................................................................... 38 2.4.4 Statistical Analysis ......................................................................................... 39 2.5 Conclusions ...............................................................................................................39 References.........................................................................................................................39 Further Reading ................................................................................................................41

CHAPTER 3

Collection and Management of Dietary Data ..................................... 43 Pauline M. Emmett, Louise R. Jones, Kate Northstone, George Pounis and Caroline M. Taylor

3.1 Introduction...............................................................................................................44 3.2 Dietary Information ..................................................................................................45 3.2.1 Methods for Collecting Dietary Information................................................. 45 3.2.2 Choice of Method........................................................................................... 54 3.2.3 Types of Collected Dietary Information ........................................................ 54 3.3 Data Management and Dietary Analysis..................................................................57 3.3.1 Portion Sizes................................................................................................... 57 3.3.2 Estimations of Nutrient Content of Foods..................................................... 58 3.3.3 Configuration and Labeling of Datasets ........................................................ 60 3.3.4 Handling of Food Group Data ....................................................................... 61 3.4 Data Manipulation ....................................................................................................64 3.4.1 Validation, Reproducibility, Calibration, and Biomarkers ............................ 64 3.4.2 Measurement Error, Misreporting, and Outliers............................................ 65 3.4.3 Energy Adjustment......................................................................................... 67 References ........................................................................................................................ 67

CHAPTER 4

Dietary Pattern Analysis ..................................................................... 75 Claudia Agnoli, George Pounis and Vittorio Krogh

4.1 Introduction...............................................................................................................76 4.2 Types of Dietary Pattern Analysis............................................................................77 4.3 A Priori Dietary Pattern Analysis.............................................................................77 4.3.1 Description...................................................................................................... 77 4.3.2 Key Aspects for the Development of a Dietary Score .................................. 79 4.3.3 Dietary Assessment Methodologies for a Priori Dietary Pattern Analysis... 81

CONTENTS

vii

4.3.4 Evaluation of the Quality of an a Priori Dietary Pattern .............................. 81 4.3.5 Examples of a Priori Dietary Patterns ........................................................... 82 4.4 A Posteriori Dietary Pattern Analysis ......................................................................86 4.4.1 Description...................................................................................................... 86 4.4.2 Factor Analysis............................................................................................... 86 4.4.3 Examples of Factor Analysis ......................................................................... 89 4.4.4 Cluster Analysis ............................................................................................. 89 4.4.5 Examples of Cluster Analysis ........................................................................ 91 4.4.6 Comparison of Factor and Cluster Analysis.................................................. 91 4.4.7 Reproducibility and Validity of a Posteriori Dietary Pattern Analysis......... 92 4.5 Hybrid Dietary Pattern Analysis ..............................................................................94 4.5.1 Description...................................................................................................... 94 4.5.2 Examples of Hybrid Dietary Pattern Analysis .............................................. 94 4.6 Challenges in Dietary Pattern Analysis....................................................................95 References.........................................................................................................................96

CHAPTER 5 Statistical Analysis of Retrospective Health and Nutrition Data.... 103 George Pounis 5.1 Introduction...........................................................................................................104 5.2 Hypothesis Testing................................................................................................105 5.3 Descriptive Statistics ............................................................................................106 5.3.1 Categorical Data .........................................................................................107 5.3.2 Continuous Data .........................................................................................107 5.4 Assessment of Normality .....................................................................................111 5.5 Confidence Interval...............................................................................................113 5.6 Pearson Chi-Square Test.......................................................................................115 5.7 Statistical Tests for Comparison of Means ..........................................................117 5.7.1 t Test ...........................................................................................................117 5.7.2 One-Way Analysis of Variance ..................................................................118 5.8 Pearson Correlation Coefficient ...........................................................................120 5.9 Nonparametric Tests .............................................................................................121 5.10 Linear Regression Analysis..................................................................................122 5.10.1 Simple Linear Regression ........................................................................122 5.10.2 Multiple Linear Regression Analysis.......................................................130 5.11 Logistic Regression Analysis ...............................................................................135 5.11.1 Odds Ratio ................................................................................................135 5.11.2 Simple Binary Logistic Regression Analysis ..........................................136 5.11.3 Multiple Binary Logistic Regression Analysis........................................137 5.11.4 Examples ..................................................................................................138 References.......................................................................................................................140

viii

CONTENTS

CHAPTER 6

Statistical Analysis of Prospective Health and Nutrition Data ....... 145 George Pounis

6.1 Introduction.............................................................................................................145 6.2 Descriptive Statistics...............................................................................................146 6.3 Measures to Calculate the Occurrence of a Health Outcome................................146 6.3.1 Incidence....................................................................................................... 146 6.3.2 Relative Risk ................................................................................................ 148 6.4 Survival Analysis ....................................................................................................149 6.4.1 Basic Concepts ............................................................................................. 149 6.4.2 KaplaneMeier Analysis ............................................................................... 150 6.4.3 Log-Rank Test .............................................................................................. 152 6.4.4 Cox Regression Analysis ............................................................................. 155 References ...................................................................................................................... 159

CHAPTER 7

Meta-Analysis of Nutrition Studies .................................................. 163 Emmanouil Bouras, Konstantinos K. Tsilidis, George Pounis and Anna-Bettina Haidich

7.1 Introduction.............................................................................................................164 7.2 Methodology of Meta-Analysis in Nutrition Research..........................................165 7.2.1 Defining the Search Strategy........................................................................ 166 7.2.2 Study Selection Procedure ........................................................................... 169 7.2.3 Quality Assessment ...................................................................................... 169 7.2.4 Data Extraction............................................................................................. 170 7.3 Statistical Methodologies Applied in Meta-Analysis of Nutrition Studies ...........170 7.3.1 Statistical Measures of Effect Included in Meta-Analysis .......................... 171 7.3.2 Choice of Meta-Analytical Method ............................................................. 173 7.3.3 Statistical Heterogeneity .............................................................................. 175 7.3.4 Small-Study Effects...................................................................................... 177 7.3.5 Software for Meta-Analysis ......................................................................... 178 7.4 Presentation and Interpretation of Results .............................................................178 7.4.1 Study Selection............................................................................................. 179 7.4.2 Study Characteristics.................................................................................... 179 7.4.3 Forest Plot..................................................................................................... 179 7.4.4 Assessing Heterogeneity .............................................................................. 182 7.4.5 Risk for Bias................................................................................................. 184 7.4.6 Funnel Plot ................................................................................................... 184 7.5 Limitations and Biases............................................................................................187 7.5.1 Challenges .................................................................................................... 189 References ...................................................................................................................... 191

CONTENTS

ix

CHAPTER 8 Principles of Research Publication ................................................. 197 Gregory S. Patience, George Pounis, Paul A. Patience and Daria C. Boffito 8.1 Introduction.............................................................................................................198 8.2 Citation Impact and Metrics ...................................................................................198 8.3 Article Elements .....................................................................................................200 8.3.1 Title............................................................................................................... 200 8.3.2 Abstract......................................................................................................... 203 8.3.3 Introduction .................................................................................................. 204 8.3.4 Methods and Results .................................................................................... 205 8.3.5 Discussion and Conclusions......................................................................... 207 8.4 Web Tools for Writing............................................................................................207 8.4.1 Word Choices ............................................................................................... 208 8.4.2 Thesauri ........................................................................................................ 208 8.4.3 Grammar ....................................................................................................... 210 8.4.4 Translation .................................................................................................... 211 8.5 Reporting Data and Analysis..................................................................................211 8.5.1 Dietary Evaluation and Analysis.................................................................. 212 8.5.2 Statistical Analysis ....................................................................................... 212 8.5.3 Graphs........................................................................................................... 213 8.5.4 Tables............................................................................................................ 217 8.6 Publishing Process ..................................................................................................218 8.6.1 Selecting the Journal .................................................................................... 218 8.6.2 A Winning Cover Letter ............................................................................... 218 8.7 Authorship Criteria and Acknowledgments ...........................................................220 8.8 Strengthening the Reporting of Observational Studies in Epidemiology and Consolidated Standards of Reporting Trials Statements........................................223 8.9 Conclusions .............................................................................................................224 References.......................................................................................................................225

PART 2 CHALLENGES IN NUTRITION SCIENCE CHAPTER 9 Mediterranean Diet: A Health-Protective Dietary Pattern for Modern Times .............................................................................. 233 Dimitra Mastorakou, Mikael Rabaeus, Patricia Salen, George Pounis and Michel de Lorgeril 9.1 Introduction.............................................................................................................234 9.2 Scientific Definition of Mediterranean Diet...........................................................234 9.2.1 Historical Overview...................................................................................... 234 9.2.2 The Traditional Mediterranean Diet ............................................................ 235

x

CONTENTS

9.3 Evidence on Health Benefits of the Mediterranean Diet .......................................235 9.3.1 Cardiovascular Diseases ............................................................................... 235 9.3.2 Cancer........................................................................................................... 236 9.3.3 Cognitive Function ....................................................................................... 237 9.3.4 Nonalcoholic Fatty Liver Disease................................................................ 237 9.3.5 Rheumatoid Arthritis.................................................................................... 238 9.3.6 Pulmonary Function ..................................................................................... 238 9.4 Food Components of Mediterranean Diet..............................................................239 9.4.1 Fish and Marine Omega-3 Polyunsaturated Fats......................................... 239 9.4.2 Plant Omega-3 Fats ...................................................................................... 240 9.4.3 Olive Oil ....................................................................................................... 241 9.4.4 Fruits and Vegetables ................................................................................... 241 9.4.5 Nuts and Seeds ............................................................................................. 242 9.4.6 Dietary Fiber................................................................................................. 243 9.4.7 Wine.............................................................................................................. 244 9.5 Mediterranean Diet Adherence in Modern Times .................................................244 9.5.1 Level of Adherence in Modern Populations................................................ 244 9.5.2 Challenges of Adhering to the Mediterranean Diet .................................... 246 9.6 Shifting to the Mediterranean Diet in the Modern Context ..................................246 9.6.1 Updated Mediterranean Diet Recommendations ......................................... 246 9.6.2 Focus on Sustainability ................................................................................ 248 9.7 Conclusions .............................................................................................................250 References.......................................................................................................................250

CHAPTER 10

Polyphenol-Rich Diets in Cardiovascular Disease Prevention ..... 259 Junichi Sakaki, Melissa Melough, Sang Gil Lee, George Pounis and Ock K. Chun

10.1 Introduction...........................................................................................................260 10.2 Polyphenols: Classes, Structures, and Chemical Properties ................................261 10.3 Polyphenols in Human Diet .................................................................................261 10.3.1 Major Dietary Sources .............................................................................261 10.3.2 Analytical Techniques to Determine Phenolic Contents .........................264 10.3.3 Estimation of Polyphenol Intakes in Human Populations.......................265 10.4 Dietary Polyphenols and Cardiovascular Disease ...............................................267 10.4.1 Cardiovascular Disease Incidence and Mortality: Evidence From Epidemiological Studies...........................................................................267 10.4.2 Cardiovascular Disease Biomarkers: Evidence From Epidemiological Studies...........................................................................272 10.4.3 Cardiovascular Disease Biomarkers: Evidence From Interventional Studies ......................................................................................................274

CONTENTS

xi

10.4.4 Biological Functions of Polyphenols in Cardiovascular Disease Prevention .................................................................................................280 10.5 Challenges in Polyphenol Research .....................................................................287 10.6 Conclusion ............................................................................................................288 References.......................................................................................................................289 Chemspider Websites......................................................................................................298

CHAPTER 11

Hydration and Health...................................................................... 299 Adam D. Seal, Hyun-Gyu Suh, Lisa T. Jansen, LynnDee G. Summers and Stavros A. Kavouras

11.1 Introduction...........................................................................................................300 11.1.1 Terminology and Measurement of Hydration Status...............................300 11.2 Hydration and Kidney Health ..............................................................................301 11.2.1 Introduction ..............................................................................................301 11.2.2 Hydration and Kidney Function...............................................................302 11.2.3 Arginine Vasopressin and Chronic Kidney Disease ................................302 11.2.4 Mesoamerican Nephropathy.....................................................................303 11.2.5 Other Kidney Diseases .............................................................................303 11.2.6 Water Intake Intervention in Chronic Kidney Disease............................304 11.3 Hydration and Glucose Regulation ......................................................................304 11.3.1 Introduction ..............................................................................................304 11.3.2 Implications of Arginine Vasopressin in Glucose Regulation.................304 11.3.3 Other Considerations in Hydration and Glucose Regulation ..................306 11.4 Implications of Fluid Balance and Obesity .........................................................306 11.4.1 Introduction ..............................................................................................306 11.4.2 Sugar-Sweetened Beverages ....................................................................307 11.4.3 Hydration Status and Obesity ..................................................................307 11.5 Hydration and Cardiovascular Health ..................................................................309 11.5.1 Introduction ..............................................................................................309 11.5.2 Cardiovascular Effects..............................................................................310 11.6 Hydration and Oxidative Stress............................................................................310 11.6.1 Introduction ..............................................................................................310 11.6.2 Implications of Poor Fluid Balance .........................................................311 11.7 Hydration and Risk of Bladder Cancer................................................................311 11.7.1 Introduction ..............................................................................................311 11.7.2 Possible Causes ........................................................................................312 11.8 Conclusion ............................................................................................................312 References.......................................................................................................................313

xii

CONTENTS

CHAPTER 12

Diet, Healthy Aging, and Cognitive Function................................. 321 Krasimira Aleksandrova, George Pounis and Romina di Giuseppe

12.1 Introduction...........................................................................................................321 12.2 Definition and Epidemiology of Cognitive Decline ............................................322 12.3 Bioactive Components of a Healthy Diet and Cognitive Decline.......................323 12.3.1 Antioxidant Vitamins ...............................................................................323 12.3.2 Polyphenols ..............................................................................................323 12.3.3 Vitamins of the B Complex .....................................................................324 12.3.4 Unsaturated Fatty Acids ...........................................................................325 12.3.5 Minerals ....................................................................................................326 12.4 Dietary Patterns and Cognitive Decline...............................................................326 12.4.1 The Mediterranean Diet ...........................................................................327 12.4.2 The DASH Diet Plan................................................................................327 12.4.3 The MIND Diet ........................................................................................327 12.5 Modern Approaches in the Study of Diet and Cognitive Health ........................328 12.6 Challenges.............................................................................................................329 12.7 Conclusion ............................................................................................................330 References.......................................................................................................................330

CHAPTER 13

Diet and Bone Health...................................................................... 337 Kate Maslin and Elaine Dennison

13.1 Introduction...........................................................................................................338 13.2 Brief Overview of Bone Physiology ....................................................................338 13.2.1 Overview of Bone Structure and Function ..............................................338 13.2.2 Bone Modeling and Remodeling .............................................................338 13.2.3 Assessment of Bone Outcomes in Research............................................339 13.3 Calcium and Vitamin D........................................................................................339 13.3.1 Calcium and Bone, an Overview .............................................................339 13.3.2 Vitamin D and Bone, an Overview..........................................................340 13.3.3 Phosphorus................................................................................................341 13.3.4 Magnesium ...............................................................................................342 13.3.5 Potassium..................................................................................................343 13.3.6 Vitamin K .................................................................................................343 13.3.7 Other Nutrients .........................................................................................343 13.4 Life Course Perspective on Nutrition and Bone ..................................................344 13.4.1 Osteoporosis: Burden and Epidemiology ................................................344 13.4.2 Etiology of Osteoporosis and Role of Diet .............................................344 13.4.3 Maternal Nutrition and Bone Outcomes..................................................345 13.4.4 Childhood and Adolescence.....................................................................346 13.4.5 Bone Health in the Older Adult: Menopause and Beyond .....................348 13.5 Challenges in Nutrition and Bone Research ........................................................351 References.......................................................................................................................352

CONTENTS

CHAPTER 14

xiii

Diet and Lung Health ...................................................................... 355 Foteini Malli, Themis Koutsioukis, George Pounis and Konstantinos I. Gourgoulianis

14.1 Introduction ........................................................................................................356 14.2 Diet and Pulmonary Function ............................................................................357 14.3 Diet and Asthma.................................................................................................357 14.3.1 Dietary Patterns ......................................................................................358 14.3.2 Vitamins..................................................................................................358 14.3.3 Minerals ..................................................................................................359 14.3.4 Fatty Acids..............................................................................................359 14.3.5 Probiotics ................................................................................................359 14.3.6 Phytochemicals.......................................................................................360 14.4 Diet and Chronic Obstructive Pulmonary Disease ............................................360 14.4.1 Dietary Patterns ......................................................................................360 14.4.2 Vitamins..................................................................................................360 14.4.3 Minerals ..................................................................................................361 14.4.4 Fatty Acids..............................................................................................361 14.4.5 Probiotics and Dietary Fibers.................................................................361 14.4.6 Phytochemicals.......................................................................................362 14.5 Diet and Lower Respiratory Tract Infections ....................................................362 14.5.1 Vitamins..................................................................................................363 14.5.2 Minerals ..................................................................................................364 14.5.3 Fatty Acids..............................................................................................364 14.5.4 ProbioticsePrebiotics .............................................................................364 14.5.5 Phytochemicals.......................................................................................365 14.6 Diet and Tuberculosis.........................................................................................365 14.6.1 Dietary Patterns ......................................................................................365 14.6.2 Vitamins..................................................................................................365 14.6.3 Minerals ..................................................................................................366 14.6.4 Fatty Acids..............................................................................................366 14.6.5 Probiotics ................................................................................................366 14.6.6 Phytochemicals.......................................................................................366 14.7 Diet and Lung Cancer ........................................................................................367 14.7.1 Dietary Patterns ......................................................................................367 14.7.2 Vitamins..................................................................................................367 14.7.3 Minerals ..................................................................................................368 14.7.4 Fatty Acids..............................................................................................368 14.7.5 Phytochemicals.......................................................................................368 14.8 Diet and Cystic Fibrosis.....................................................................................369 14.8.1 Vitamins..................................................................................................369

xiv

CONTENTS

14.8.2 Minerals ..................................................................................................369 14.8.3 Fatty Acids..............................................................................................370 14.8.4 Probiotics ................................................................................................370 14.8.5 Phytochemicals.......................................................................................370 14.9 Diet and Interstitial Lung Diseases....................................................................370 14.10 Maternal Diet in Early Life and Lung Health ...................................................371 14.10.1 Dietary Patterns ....................................................................................371 14.10.2 Vitamins................................................................................................371 14.10.3 Fatty Acids............................................................................................372 14.10.4 Probiotics ..............................................................................................372 14.11 Challenges in Diet and Lung Health Research..................................................372 References.......................................................................................................................373 Further Reading ..............................................................................................................381 Index ................................................................................................................................................. 383

List of Contributors Claudia Agnoli Epidemiology and Prevention Unit, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy Krasimira Aleksandrova Nutrition, Immunity and Metabolism Senior Scientist Group, Department of Nutrition and gerontology, German Institute of Human Nutrition Potsdam-Rehbruecke (DIfE), Nuthetal, Germany Daria C. Boffito Department of Chemical Engineering, Polytechnique Montre´al, Montre´al, QC, Canada Emmanouil Bouras Department of Medicine, Aristotle University of Thessaloniki, Thessaloniki, Greece Ock K. Chun Department of Nutritional Sciences, University of Connecticut, Storrs, CT, United States Michel de Lorgeril Laboratoire Cœur et Nutrition, TIMC-IMAG, School of Medicine, University of Grenoble-Alpes, Grenoble, France Elaine Dennison MRC Lifecourse Epidemiology Unit University of Southampton, United Kingdom Romina di Giuseppe Institute of Epidemiology, University Kiel, Kiel, Germany Monica Dinu Department of Experimental and Clinical Medicine, University of Florence, Florence, Italy Pauline M. Emmett Centre for Academic Child Health, Population Health Sciences, Bristol Medical School, University of Bristol, Bristol, United Kingdom Konstantinos I. Gourgoulianis Respiratory Medicine Department, University of Thessaly, School of Medicine, Larissa Greece Anna-Bettina Haidich Department of Medicine, Aristotle University of Thessaloniki, Thessaloniki, Greece Lisa T. Jansen University of Arkansas, Fayetteville, AR, United States Louise R. Jones Population Health Sciences, Bristol Medical School, University of Bristol, Bristol, United Kingdom Stavros A. Kavouras Hydration Science Lab, Arizona State University, Phoenix, AZ, United States

xv

xvi

LIST OF CONTRIBUTORS

Themis Koutsioukis Respiratory Medicine Department, University of Thessaly, School of Medicine, Larissa Greece Vittorio Krogh Epidemiology and Prevention Unit, Fondazione IRCCS Istituto Nazionale dei Tumori, Milan, Italy Sang Gil Lee Department of Food Science and Nutrition, Pukyong National University, Busan, Korea Foteini Malli Respiratory Medicine Department, University of Thessaly, School of Medicine, Larissa Greece Technological Institute of Thessaly, Nursing Department, Larissa, Greece Kate Maslin MRC Lifecourse Epidemiology Unit University of Southampton, United Kingdom Dimitra Mastorakou Leicester, United Kingdom Melissa Melough Department of Nutritional Sciences, University of Connecticut, Storrs, CT, United States Kate Northstone Population Health Sciences, Bristol Medical School, University of Bristol, Bristol, United Kingdom Gregory S. Patience Department of Chemical Engineering, Polytechnique Montre´al, Montre´al, QC, Canada Paul A. Patience Department of Electrical Engineering, Polytechnique Montre´al, Montre´al, QC, Canada George Pounis Alimos, Athens, Greece Mikael Rabaeus Geneva, Switzerland Junichi Sakaki Department of Nutritional Sciences, University of Connecticut, Storrs, CT, United States Patricia Salen Laboratoire Cœur et Nutrition, TIMC-IMAG, School of Medicine, University of Grenoble-Alpes, Grenoble, France Adam D. Seal University of Arkansas, Fayetteville, AR, United States Hydration Science Lab, Arizona State University, Phoenix, AZ, United States Francesco Sofi Department of Experimental and Clinical Medicine, University of Florence, Florence, Italy Unit of Clinical Nutrition, Careggi University Hospital, Florence, Italy Don Carlo Gnocchi Foundation, Onlus IRCCS, Florence, Italy

LIST OF CONTRIBUTORS

Hyun-Gyu Suh University of Arkansas, Fayetteville, AR, United States Hydration Science Lab, Arizona State University, Phoenix, AZ, United States LynnDee G. Summers University of Arkansas, Fayetteville, AR, United States Caroline M. Taylor Centre for Academic Child Health, Population Health Sciences, Bristol Medical School, University of Bristol, Bristol, United Kingdom Konstantinos K. Tsilidis University of Ioannina School of Medicine, Ioannina, Greece

xvii

This page intentionally left blank

Preface Analysis in Nutrition Research: Principles of Statistical Methodology and Interpretation of the Results describes, in a comprehensive manner, the methodologies of the quantitative analysis of data originating specifically from nutrition studies. The book summarizes various study designs in nutrition research, research hypotheses, the proper management of dietary data, and analytical methodologies, with a specific focus on how to interpret the results of any given study. In addition, it provides a comprehensive overview of the methodologies used in study design and the management and analysis of collected data, paying particular attention to all of the available, modern methodologies and techniques. Readers will find an overview of the recent challenges and debates in the field of nutrition research that will define major hypotheses for research in the next 10 years. Nutrition scientists, researchers, and undergraduate and postgraduate students will benefit from this thorough publication on the topic.

xix

This page intentionally left blank

Acknowledgments It is a pleasure to thank all of the people who made this title possible. I am grateful to the contributors who kindly participated in preparing the various chapters of this book and shared with us their knowledge, experience, and expertise. I would not have achieved this ambitious goal without the support of Elsevier and I wish to thank Megan Ball, who trusted me with the editorial role. I want to extend my sincere gratitude to my colleagues from Elsevier, who supported me in the various steps of this publication. My final thoughts go to my family and Eleftheria. I would not have gone so far without their love and encouragement. Georgios

xxi

This page intentionally left blank

PART

ANALYSIS IN NUTRITION RESEARCH

1

This page intentionally left blank

CHAPTER

1

DESIGN OF OBSERVATIONAL NUTRITION STUDIES

George Pounis Alimos, Athens, Greece

CHAPTER OUTLINE 1.1 Introduction to Observational Nutrition Studies...................................................................................3 1.2 Ecological Nutrition Studies..............................................................................................................5 1.2.1 Description ..................................................................................................................5 1.2.2 Challenges...................................................................................................................5 1.2.3 Example of Ecological Study on Diet and Cancer.............................................................6 1.3 Cross-Sectional Nutrition Studies ......................................................................................................6 1.3.1 Description ..................................................................................................................6 1.3.2 Sampling.....................................................................................................................8 1.3.3 Challenges.................................................................................................................10 1.3.4 Example of the National Health and Nutrition Examination Survey .................................11 1.4 CaseeControl Nutrition Studies .......................................................................................................12 1.4.1 Description ................................................................................................................12 1.4.2 Sampling...................................................................................................................12 1.4.3 Challenges.................................................................................................................14 1.4.4 An Example in the Study of Hodgkin Lymphoma ...........................................................15 1.5 Cohort Nutrition Studies..................................................................................................................15 1.5.1 Description ................................................................................................................15 1.5.2 Sampling...................................................................................................................16 1.5.3 Challenges.................................................................................................................17 1.5.4 Example of the Nurses’ Health Study ...........................................................................18 References ............................................................................................................................................19

1.1 INTRODUCTION TO OBSERVATIONAL NUTRITION STUDIES The science of human nutrition that is frequently described as “nutrition science” or “nutrition” is the science of food, the nutrients and other substances therein, their action, interaction and balance in relation to health and disease, and the processes by which the human organism ingests, absorbs, transports, utilizes and excretes food substances [1,2]. Analysis in Nutrition Research. https://doi.org/10.1016/B978-0-12-814556-2.00001-4 Copyright © 2019 Elsevier Inc. All rights reserved.

3

4

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

Systematic observation along with accurate and systematic measurement has been the basis for the development of science in various fields from mathematics and physics to sociology and medicine. Although nutrition science is relatively young compared with others, it has been firmly tied to observation and the measurement since the mid-1950s, when the first evidence-based conclusions regarding the association of nutrition and health were determined [1e13]. As part of nutrition science, nutritional epidemiology incorporates valuable information related to the methodologies and techniques of observation and measurement applied to generate evidence-based conclusions regarding the interaction between diet and health, among others. A considerable proportion of dietary risk factors for various diseases have been identified and studied through epidemiological health and nutrition studies. An observational nutrition study may be defined as the detailed investigation and analysis of information provided by a retrospective or prospective systematic observation and measurement of a sample’s dietary factors (exposures) and health characteristics (outcomes), in which the researcher does not willingly influence the collected information. Observational nutrition studies have an epidemiological character and aim to generalize conclusions derived from the investigation and analysis of sample data from the reference population [6,14]. They are often named population-based studies. According to the nature of the collected data, they are divided into retrospective and prospective studies [6,14,15]. Types of retrospective surveys include ecological, cross-sectional, and caseecontrol studies whereas prospective surveys include follow-up and longitudinal studies, which are generally labeled cohort studies. All of these study designs provide valuable descriptive information about adherence to dietary patterns, the consumption of foods and nutrients, the presence of certain dietary behaviors, and other dietary characteristics in a population base. In addition, they aim to test hypotheses related to the association of dietary exposures with health outcomes. Although by themselves these studies are not enough to prove a cause-and-effect relationship between a dietary factor and a health outcome, associations observed mainly in cohort surveys assign the potential existence of causality. In these cases, further investigation through intervention trials and meta-analysis is essential when it is ethically and technologically feasible. Evidence-based nutrition science or practice is well-promoted as the preferable and most accurate methodology to make decisions in all related disciplines; it aims to maintain or improve the health of individuals, groups, and populations [10]. Observational nutrition studies are a first and important step in the evidence-based concluding chain of nutrition science. All of the various types of observational nutrition studies have certain advantages and limitations that derive from their nature, and which should be considered throughout the stages of the survey (i.e., design, implementation, analysis, presentation, and publication of results and conclusions). In addition, the methodology of the collection, the management and statistical analysis of data, and the presentation and interpretation of the results vary among the different designs. These issues are addressed in other chapters of the book (i.e., Chapters 3e8). The following sections offer an overview of the various types of observational nutrition studies illustrating their basic concepts and design methodologies, providing survey examples from the literature.

1.2 ECOLOGICAL NUTRITION STUDIES

5

1.2 ECOLOGICAL NUTRITION STUDIES 1.2.1 DESCRIPTION An ecological nutrition study is the first level of systematic and comparative observation and measurement of the dietary characteristics of large populations (mainly geographically oriented), usually in parallel with the systematic and comparative observation and measurement of health-related indices in the same populations [6]. The experience of variations in an index of health status between populations with variations in the average value of a dietary factor introduces the concept of the potential association between the dietary exposure and the health outcome. This remains to be confirmed in other observational and experimental settings. Implementation of this type of survey requires at least that a population-based measure of a dietary factor and an index of health status be available for two or more populations. For instance, a country with 20% per capita intake of calories from fats has a lower incidence of colon cancer than does a country with 45% per capita intake of calories from fats. This introduces the concept that the consumption of fat in these populations may have a role in explaining the variation presented in their incidences of colon cancer. Although the major aim of ecological studies is to characterize populations over a dietary factor and an index of health status rather than evaluating the association between a dietary exposure and a health outcome [14]. More often, this survey type is used to explore geographical differences in the diet and health of large populations. It is also used to compare changes in diet and the health status of populations over time. Sometimes an ecological study is the only design that can be planned if the dietary data of the study population are unavailable at an individual level. An advantage of this survey type is that the average population-based diets and the per capita consumption of food tend to be stable over time [14]. Moreover, the indices of diseases are mainly derived from large samples and are under limited biases. For these reasons, it can be hypothesized that variations in health indices and dietary factors in the studied populations may have longer time of preoccurrence. However, a serious limitation is that the same variations can also be attributed to determinants of disease other than dietary, such as genetic, environmental, clinical, and lifestyle. This poses a major disadvantage of ecological studies along with the limited opportunity to reproduce results, especially for international surveys. Despite the inability of this study design to validate a cause-and-effect relationship between a dietary exposure and a health outcome, it has been proved to be effective in generating scientific hypotheses for further observational and experimental studies.

1.2.2 CHALLENGES Among the major concerns for the designers of an ecological nutrition study is the identification of appropriate population-based measures of the dietary characteristics and health status of the populations under investigation. Estimates of the population dietary intake are often retrieved from preexisting data generated by systematic measures and evaluations performed for other reasons (i.e., economic studies, census). Sources of dietary information include national and international data about the per capita

6

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

consumption of foods, data from household surveys, and individual survey data from representative sample populations. An important consideration for the survey’s investigators is a study of the methodology used for data collection and management. The use harmonized national and international population-based dietary data ensures their comparability. This purpose is assisted by using common data collection and management methodologies. Another challenge to conducting an ecological study is evaluating the level of accuracy of population-based dietary data. Biases related to the methodology of data acquisition should be limited. For instance, frequently, the data for the per capita consumption of food in a country are retrieved by economic studies for food sales, imports, and exports. Because of their aims and methodology, these surveys cannot control for the amount of food that is wasted. It may be true that a significant amount of food that has been sold to consumers was not actually consumed. The percentage of wasted food among countries with different socioeconomical standings may vary significantly. This may lead to biased conclusions regarding variations observed in food consumption and the health status of populations under investigation.

1.2.3 EXAMPLE OF ECOLOGICAL STUDY ON DIET AND CANCER In 1975, Armstrong and Doll published findings of their ecological study on the association of dietary factors with the incidence of various types of cancer and cancer mortality rates [16]. This was one of the first observational studies on the topic, and despite the methodological limitations derived from its design, it has been a reference work for further studies on the topic (i.e., more than 1700 citations in Scopus Metrics). The researchers used data on the incidence rate of 27 cancer types in 23 countries for individuals aged 35e64 years, derived from the Union for International Cancer Control in 1966 and 1970 [17,18]. Data on the cancer mortality rates for 14 cancer types in 32 countries were taken from Segi et al. and the World Health Organization (1967e69, 1970) [19e21]. This information was studied in association with data on the international per capita consumption of various foods derived from various reference sources. The results of the study were presented grouped by the type of cancer, and graphs showing related trends were used. Various correlations were noted between dietary variables and cancer incidence or mortality. The researchers pointed out in their conclusions that the most strong associations were of meat and total fat with colon, rectum, and breast cancer. Interestingly, they also stressed the limitations of their study and noted that other population characteristics might act as confounding factors in some of the observed associations.

1.3 CROSS-SECTIONAL NUTRITION STUDIES 1.3.1 DESCRIPTION Cross-sectional nutrition studies are often named descriptive surveys; they are a basic type of observational study. They mainly collect and analyze dietary and health data from a sample population at the certain time point in the participants’ recruitment [22,23]. For this reason, information provided by participants is a snapshot of the sample’s dietary or other characteristics at the time of recruitment.

1.3 CROSS-SECTIONAL NUTRITION STUDIES

7

Cross-sectional studies in nutrition science provide descriptive results for the degree of exposure to dietary factors of a sample population. They evaluate current dietary habits of participants and offer valuable information about the consumption of foods and other dietary components. Energy, water, and nutrient intake, adherence to dietary patterns, and other dietary factors can be estimated [6,24e27]. Moreover, by collecting biological samples and biobanking [28,29], markers of food and nutrient bioavailability and metabolism can be measured in the sample population base. Frequently, participants undergo anthropometric screening and the related information is collected and analyzed [30e32]. Because these studies run in parallel with other health-related studies, they provide important information for the prevalence of a disease or health condition [33]. It is crucially important that the diagnostic criteria used in the survey are the most up-to-date and in accordance with the recent literature. There are several expressions of prevalence; a simple one is: Number of diagnosed cases of a disease or a health condition at a certain time point in the sample population Prevalence ¼  100 sample size For example, when we read that the prevalence of obesity was 25% in November 2015 in a sample population, we may understand that of the 100 participants who were screened at that specific time point, 25 were obese. During a cross-sectional nutrition survey, the degree of exposure to various potential risk factors such as lifestyle, socioeconomic, environmental, genetic, and clinical is usually measured in parallel at the same time point of the study sample’s recruitment [8,34e36]. The choice of the factor that will be assessed depends on the aims of the survey. It is highly essential that all screening tools used to assess the risk factors be valid and accurate. The time frame of a cross-sectional study is limited to a time point, so it is impossible to identify cause-and-effect relationships between dietary exposures and health outcomes [33,37,38]. The definition of causality in an observed association requires at least that the dietary exposure occurred before the diagnosis of a disease or health condition [6,14,15]. This condition is not confirmed by a crosssectional design. Statistical methodologies are available to evaluate the association of a dietary exposure and a health outcome and they are also widely used by researchers, such as linear regression modeling and generalized linear modeling (see Chapter 5). However, the interpretation of significant associations in cross-sectional settings should be limited by the study’s inability to express causality. When these associations are supported by literature data (i.e., in vitro or other studies) that offer biological plausibility, they may guide further hypothesis testing and studies in other survey designs (i.e., cohort studies, meta-analyses, and randomized control trials). Weakness points of this type of observational study also relate to dietary assessment methodologies and tools, which sometimes fail to evaluate the dietary habits of participants accurately. In particular, bias related to the recall of dietary information from sample members is evident whereas potential alterations of long-term dietary habits resulting from the previous presence of a disease or a health condition cannot be thoroughly investigated. Thus, the observed statistical associations between dietary exposures and health outcomes may be biased. These issues are addressed in detail in Chapter 3, where the collection and management of dietary data are discussed.

8

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

Despite the limitations presented by the cross-sectional design, the data extracted by this type of nutrition study provide useful information for public health professionals. In fact, cross-sectional surveys offer comprehensive screening of the nutritional and health status of sample populations at certain time points. For this reason, many health institutions worldwide regularly plan, fund, conduct, analyze, and publish this type of study (related examples are presented in Section 1.3.4).

1.3.2 SAMPLING Given the importance of cross-sectional nutrition surveys in public health perspectives, the need exists to generalize cross-sectional results derived by analyzing sample data to the reference population [39]. The proper choice of the reference population and the methodology for recruiting a representative sample are crucial. The process of selecting the reference population requires a precise definition of the main hypothesis testing, dietary exposures, health outcomes, and other studied factors (Fig. 1.1). The calculation of the sample size, which will ensure the statistical power of the study’s results, is the next important step [40]. There is available free or paid software that allows these calculations with a certain level of accuracy. When the reference population and the sample size have been well-defined, a decision should be made about the sampling methodology or sample design. This is the method that will be used for the representative selection of sample members from the reference population. Sample members can be individuals, couples, households, or even schools. The finest approach is to use random sampling [15]. In this method, theoretically the probability of a member of the population being included in the study’s sample is the same for all members of the population. In this scenario, possible selection biases are limited [33]. Several sample designs are available: simple random sampling, systematic sampling, multistage sampling, and stratified sampling [6,15,23]. Application of the sampling methodology depends on available resources and the decisions of the investigators regarding how the sample will best represent the reference population. Simple random sampling is a sample design that requires an available comprehensive sample frame (i.e., a list of all potential sample members of the reference population) [41,42]. For instance, a

FIGURE 1.1 Sampling of cross-sectional nutrition studies.

1.3 CROSS-SECTIONAL NUTRITION STUDIES

9

complete sample frame for the adult population of a region can be the electoral register. Once the sample frame has been identified, a number is given to all of its members. The random selection of sample members who will be invited to participate in the study is based on a random sequence of numbers, which can be generated by software. Systematic sampling is another sample design based on the conditions that potential participants are arranged in a type of sequence and sample members are selected according to a random starting point and a fixed periodic interval [41e43]. Examples of the potential participants’ sequence are a series of index cards, houses along the side of a street, and patients who arrive at a hospital or clinic. A periodic interval is commonly named sampling interval and is calculated by the following formula: Sampling interval ¼

Reference population size Desired sample size

For example, if we want wanted to select a random group of 200 sample members out of 2000 members of the reference population, the sampling interval is 10. Staring from a random point of the population sequence (e.g., the 18th member), every 10th member of the sequence should be invited to participate in the study (i.e., the 18th, 28th, 38th.) until 200 sample members will be reached. To ensure that the sample is random, the starting point should be random and the sampling interval should not correspond with any repeated pattern in the sequence of the reference population. For instance, if we have a list of couples and the women are always placed first, if we start from the third member of the list and the sampling interval is 8, the sample members will be all women. This example can be easily understood, although there are might be repeated patterns that are difficult to identify and may be associated with specific characteristics of the sample. Multistage sampling is one of the most popular sample designs for cross-sectional nutrition studies because it is more realistic and ensures the representativeness of a sample deriving from a large geographical area [44]. Frequently, cross-sectional nutrition studies aim to evaluate the current dietary habits of large populations such as a country. It would be unrealistic to select a sample of 1000 participants from a country using a simple random sample design. The first problem would be the existence of a comprehensive sample frame. Even if this were available, the sample members indicated by random numbers might live far from each other, which would make it impractical in terms of time and resources to recruit them. To apply multistage sampling in this case, we may divide the country into regions, and from that list we may select a random sample of regions (first-stage sampling). In each of the randomly selected regions a list of towns should be drawn and a random sample of them can be selected (second-stage sampling). In each randomly selected town, a sample frame may be available to make a random selection of sample members (third-stage sampling). Even if the sample frame is incomplete, further sampling stages can be used with geographical or other related criteria until a random, representative, and feasible sample in terms of recruitment practices can be selected. Stratified sampling is a sample design that is used in parallel with simple or the multistage sampling [12]. A stratified random sample occurs after dividing the population into specific subgroups (i.e., strata) often according to demographic characteristics such as gender, age, and socioeconomic status. In every subgroup, a random sample is selected by applying simple or multistage sampling. This type of sample design is appropriate when there may be indications that dietary exposures or health outcomes vary between strata.

10

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

1.3.3 CHALLENGES A major concern of researchers regarding the final sample for all types of sample design is to achieve an optimal level of randomness and representativeness of the reference population. In this effort, a serious challenge is to identify an up-to-date and complete sample frame, because many crosssectional nutrition studies refer to large populations. The use of random techniques in selecting sample members from a relatively complete sample frame increases the probability that the final study sample will be representative. The census of a population may provide valuable information to recruiters about demographics and other characteristics of the reference population. During various stages of the study’s sample recruitment, it is a common practice for some of the collected data to be assessed compared with census data to determine significant differences. The absence of significant variations between them indicates representativeness. Data from previous studies in the reference population could be also used for the same reason. Another challenge for the designers of cross-sectional nutrition studies is the response rate. To calculate the response rate, we may use the following formula: Response rate ¼

Number of the interviewed sample members  100 Number of the invited sample members

For example, from a sample frame of 2000 sample members, 250 randomly selected sample members were invited to participate in the study and 150 were finally interviewed. Thus, the response rate is 60%. There are several reasons why a sample member might not participate in the study. The designers of a cross-sectional survey should plan techniques that deals with this issue (e.g., prompting by telephone and mail, study advertisements, letters outlining the importance of replying). All possible efforts should be made to ensure a high response rate to minimize possible selection bias. A common dilemma regards what response rate is acceptable. There are empirical rules proposing that a response rate over 80% might set a level of confidence for the recruiters. However, consider a situation in which the response rate was 90% but 10% of nonresponders had very specific and distinct characteristics from responders that made them decide not to respond. In this example, selection bias might be present despite the high response rate. To evaluate bias introduced by nonresponse, it is important to obtain some information from sample members who refused to participate or could not be reached. The design and application of a short questionnaire acquiring important information from a random sample of nonresponders might be useful. These data could be later compared with those from responders and help the researchers to identify significant differences. Commonly, demographics (e.g., gender, age, socioeconomic status) or other information might be available for the members of the sample frame. These data could be comparatively evaluated for responders and nonresponders. Equally important to selection bias that is related to the response rate is biased responses deriving from the use of inappropriate or inaccurate assessment methodologies or tools. Investigators face myriad challenges when decisions need to be made regarding the use of valid methodologies and tools to measure dietary exposures, health outcomes, and other studied factors. The same is for the

1.3 CROSS-SECTIONAL NUTRITION STUDIES

11

application of appropriate data management and statistical analysis methodologies. These challenges are evident in all study designs; they are thoroughly presented and discussed in Chapters 3e6. Generally, compared with other observational study designs (i.e., caseecontrol and cohort studies) or randomized control trials, cross-sectional nutrition studies are less expensive and time-consuming [6]. Moreover, in parallel with the evaluation of dietary data, valuable descriptive information on health parameters and other characteristics of the population can be assessed. The prevalence of diseases can be estimated especially for longer-lasting diseases (i.e., chronic diseases) because diseases lasting for a short period may be underestimated (i.e., the Neyman bias) [7,45]. Considering that a cross-sectional study refers to the current status of a population, it gains great value for public health planning means. Although associations between dietary exposures and health outcomes that are identified through a cross-sectional design do not have a cause-and-effect character, they guide further hypothesis testing and studies in the field.

1.3.4 EXAMPLE OF THE NATIONAL HEALTH AND NUTRITION EXAMINATION SURVEY The National Health and Nutrition Examination Survey (NHANES) is a unique rolling research program that collects cross-sectional information for the nutritional status and health of children and adults in the United States [46]. It was initiated in the early 1960s, and from the 1999 onward, it became a continuous program focusing on a variety of health and nutrition aspects. As a project of National Center for Health Statistics (NCHS), NHANES collects data from a representative sample of 5000 persons each year. The NCHS is part of the Centers for Disease Control and Prevention and has the responsibility of producing vital and health statistics for the nation. Participants are located in counties across the country, 15 of which are visited each year. To produce reliable statistics, NHANES oversamples persons aged 60 years and older who are African American and Hispanic. For all participants, the NHANES data collection includes an in-person household interview that is performed by highly trained field investigators. In a second stage, all sample members are asked to complete a comprehensive health examination, which consists of clinical tests, laboratory studies, and additional interviews. These examinations are performed by experienced scientific personnel in specially designed and equipped mobile centers, which travel to locations throughout the country. The third stage includes postexamination interviews and questionnaires that are administered by telephone or mail. Three major types of dietary data are collected in NHANES: dietary behavior, a 24-h dietary recall, and a food frequency questionnaire (FFQ). During the household interview, information is obtained on dietary behavior including topics such as dietary modifications owing to health conditions and dietary supplement use. During the in-person examination, a 24-h dietary recall is administered. These dietary interviews are conducted by dietary interviewers. Survey participants aged 12 years and older complete the dietary interview on their own. Proxy respondents report for children who are aged 5 years and younger and for other persons who cannot self-report. Proxy-assisted interviews are conducted with children aged 6e11 years. A second 24-h dietary recall is administered to all participants by telephone 3e10 days after the in-person health examination. The collected data provide valuable information about the nutritional and health status of the participants. They are used to estimate the prevalence of various major diseases and the distribution of

12

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

dietary and other risk factors in the population. In addition, the association of the nutritional status with health promotion and disease prevention is evaluated. The NHANES results and outputs set the basis for further studies and assist in developing meaningful public health policies.

1.4 CASEeCONTROL NUTRITION STUDIES 1.4.1 DESCRIPTION The caseecontrol nutrition study is another type of observational study with retrospective character. In this survey type, participants with a newly diagnosed disease (cases) are compared with matched healthy individuals (controls) for the past exposure to dietary factors [6,14]. Observation of significant differences in the dietary exposures among groups of cases and controls suggests that the exposure may be associated with decreased or increased prevalence of the disease [47]. This study type allows an estimation of the odds ratio (OR) (see Chapter 5), which represents the odds that a health outcome will occur given a particular dietary exposure, compared with the odds of the outcome occurring in the absence of that exposure [22,23]. Study of the ORs introduces the concept of the association between dietary exposure and health outcome. However, data related to the dietary habits of participants for the period previous to the diagnosis of the disease are retrieved mainly by memory recall (using various dietary assessment methodologies) and are under recall bias [15,48,49]. In this context, interpretation of the results should be always performed with caution. A basic characteristic of this survey type is the matching of cases and controls for some confounding factors (e.g., gender, age, socioeconomic status). Conceptually, a confounder is a variable that influences both the dietary exposure and the health outcome, causing a false association between them [33]. Matching is a kind of control for confounding effects in an observed association between a dietary exposure and a health outcome [14]. In addition, statistical methodologies that are used to analyze caseecontrol data allow further controlling for potential confounding (see Chapter 5) [11,50]. Caseecontrol nutrition studies are generally efficient for studying the dietary implications of rare diseases (i.e., diseases with a low incidence in the reference population that need a long induction period) [15]. They also demand less time and resources to be conducted compared with cohort studies (see Section 1.5) because they require smaller samples and no follow-up is necessary.

1.4.2 SAMPLING A major aspect of sampling methodologies applied to caseecontrol studies is the proper choice of cases and controls. A variety of approaches are available to recruit sample members for both groups that aims to ensure a certain level of representativeness of the reference populations and avoid sampling bias. The first dilemmas for caseecontrol study designers are related to the methodology that will be used to select the cases. The precise and accurate definition of the disease or the health condition under investigation is the first, crucial step [9]. The diagnostic criteria that will be applied should ensure the highest level of accuracy possible to identify the cases, and they should be the most up-to-date. In the first stages of design, the recruiters should also decide on the inclusion of prevalent or incidence cases. Prevalent cases are existing cases that are present in the population during the time

1.4 CASEeCONTROL NUTRITION STUDIES

13

frame of the study’s sample recruitment .. This definition includes both new and old cases. The incidence cases are the newly diagnosed individuals within the period of the sample recruitment .. The major advantage of selecting incidence instead of prevalent cases in a caseecontrol nutrition study is that dietary information recalled from newly diagnosed individuals tends to be more representative of past exposure to the investigated dietary factors. The dietary habits of prevalent cases may have been amended in the time between the diagnosis of the disease and the participant’s recruitment, as an effort to adhere to a healthier lifestyle. The selection of prevalent cases is sometimes the adopted methodology because it needs less time and resources. In this scenario, the check for differences in the investigated dietary exposures between groups of prevalent cases with various times of diagnosis may help in understanding the level of biases induced in the study. The choice of incidence cases should be highlighted as the preferred methodology in caseecontrol nutrition studies because the exposure is always a dietary factor that is sensitive to alterations after the diagnosis of a disease. Moreover, by recruiting only newly diagnosed cases and selecting controls with a methodology to be representative of the population in which the cases arise makes identification of an association between a dietary factor and a disease mimic those of a cohort study. Cases that will be recruited into the study can be selected from the population of patients attending a hospital or a network of hospitals. If these health care units are the major points of treatment for patients of a disease under investigation, the sample members will be more representative of the diseased population. Another methodology is the recruitment of cases from the population over a fixed period of time. In this instance, there should be evident difficulties in tracing the subjects and the refusal rate may be greater. These factors may lower the representativeness of the diseased population. There are certain situations in which this method is applied; they are related to the existence of a comprehensive registration system for patients of the disease under investigation .. This allows the identification of a complete sample frame of the diseased population. Equally crucial as the choice of the cases is the appropriate selection of controls that fit the same eligibility criteria as cases apart from those that related to the diagnosis of the disease under investigation. The study designers should take into account any special characteristic or circumstance of the study. A first approach is to select the controls randomly from the sample frame of the population from which the incident cases were identified [51,52]. It is a time- and resource-consuming methodology, requiring a complete sample frame that is challenging to find, and the selection of cases may be under selection bias of nonresponse. A second methodology is to select controls from the close environment of the diseased person, such as relatives and friends [51,52]. In this instance, the level of the exposure of the controls to dietary factors may have similarities with that of the cases, leading to overmatching and making it impossible to identify significant differences. The same challenge is evident when the controls are drawn from the sample frame of the neighborhoods of the cases. People living in the same areas interact each other in a way that may lead to overmatching of cases and controls, especially for the investigated exposures (i.e., dietary habits). Another common approach is to select controls from the same hospitals from which the cases have been drawn [51,52]. This methodology requires the controls to have been hospitalized for a reason different from the disease under investigation. Many health conditions have common soils in terms of

14

CHAPTER 1 DESIGN OF OBSERVATIONAL NUTRITION STUDIES

dietary risk factors that are attributed to their development and progression. The selection of the sample members in a caseecontrol nutrition study, however, requires that the level of dietary exposure of the cases and controls not adhere to a common pattern. In this way, the potential differences in dietary factors between cases and controls corresponding to a variation in the prevalence of the disease may be assessed. Some selective factors may be present that bring patients to a specific hospital (i.e., social status, residence). When controls are chosen from the same hospitals where cases have been drawn, both groups become more homogeneous with respect to these selective factors, limiting their effects. The selection of controls from a hospital base also seems to enhance the completeness and accurateness of the collected information. Controls experiencing a recent hospitalization and illness tend to be more cooperative with the study’s procedures because they better appreciate and respect the significance of the aims of a health survey.

1.4.3 CHALLENGES During the design of a cross-sectional nutrition study, the investigators may force various challenges related to recruiting the cases and controls and the method applied to match the two groups. The first challenge is the uncertainty presented in classifying patients into the diseased group. It is sometimes evident that the diagnostic criteria do not provide with a definite decision regarding the presence of a disease. In these instances, further classification of cases into groups of level of certainty for the presence of the disease (i.e., high, moderate, low) may assist in data analysis. Ideally, when the definition of the cases is performed with a high level of certainty, this results in a homogeneous group of patients with a common pathology. The observation of significant associations in these studies may provide a better etiological background. There are some occasions, especially for newly diagnosed cases, in which patients are too ill to respond or to recall information about diet or other risk factors. This generates the need for some eligibility criteria that should be decided with respect to general hypothesis testing, always considering avoiding selection bias. They should be common for cases and controls apart from those related to the diagnosis of the disease under investigation. The number of the invited eligible participants who refused to participate in the survey should also be recorded along with the reason for nonparticipation and some other information (i.e., age, gender, sociodemographic characteristics) if possible. Another challenge for the study designers regards the methodology to be used to match the cases and controls. A first, common approach is that for every case, one or more identical controls, in terms of some characteristics (matching factors), is recruited [33]. The matching factors are confounders of the association between the dietary exposure (see Section 1.4.1) and the disease and should be carefully chosen. Their selection should be based on strong evidence supporting their confounding role. Frequently used factors are gender, age, residence, and nationality. Researchers should avoid matching with many factors because this may cause significant similarities in terms of dietary exposures between cases and controls. Overmatching makes it difficult to study dietary exposures as potential causes for the disease or the health condition under investigation [33]. Researchers can perform further adjustments for potential confounders during statistical analysis of the data. The second matching approach requires the group of healthy controls to be recruited with an equal proportion of some matching factors with the group of cases [15]. For instance, if in the group of cases

1.5 COHORT NUTRITION STUDIES

15

35% are females, 55% are at age a. Now suppose that in Example 5.1, after quantifying the hypothesis testing, choosing the appropriate statistical test, and calculating the test statistic, we get p  value ¼ 0:01 < 0:05 ¼ a. The condition P  a is confirmed, so it may be concluded that there is strong statistical evidence to reject H0 and support the concept that the consumption of citrus fruits prevents colon cancer.

5.3 DESCRIPTIVE STATISTICS The descriptive phase of data analysis is the first to be performed after ensuring the quality of the study dataset. It describes the characteristics of the sample, which are relevant to the scopes of the study and provides an overview of them. The descriptive statistics appear in the beginning of the results section of any scientific report of a nutrition study [9]. Text, tables, and figures are used for the comprehensive presentation of the descriptive results. Often the data are illustrated stratified by the health outcome under investigation (i.e., the presence or absence of hypertension) or a confounding factor or effect modifier (i.e., sex, age group, educational status).

5.3 DESCRIPTIVE STATISTICS

107

During this phase, the researchers are also able to assess the distribution of the data [10e12]. This is a useful procedure because it assists in checking for any violation in the assumptions underlying various statistical tests that will be presented later in this chapter. The following sections introduce the most popular descriptive statistics for categorical and continuous data, with their definitions and appropriate examples.

5.3.1 CATEGORICAL DATA Categorical variables are mainly analyzed and presented as relative frequencies [13]. Categorical data can be the degree of a dietary exposure (i.e., red meat consumption per day), a health outcome (i.e., the presence or absence of hypertension), or a confounding factor (i.e., gender). Example 5.2. In a study of red meat consumption, 200 participants were grouped into three categories: 1. 0e30 g/day, n ¼ 50 participants 2. 30e60 g/day, n ¼ 100 participants 3. 60 g/day, n ¼ 50 participants The descriptive analysis of this variable can be presented as: 25% of participants were reported to consume 0e30 g/day. 50% of participants were reported to consume 30e60 g/day. 25% of participants were reported to consume more than 60 g/day. Sometimes both the relative frequency and the number of participants in each group (i.e., the absolute frequency) are illustrated.

5.3.2 CONTINUOUS DATA There are various descriptive statistics for the analysis and presentation of continuous data. The most frequently used ones will be introduced in the following sections. A continuous variable can be the level of a dietary exposure (i.e., the intake of red meat in grams per day), a health outcome (i.e., blood glucose levels), or a confounding factor (i.e., age).

5.3.2.1 Arithmetic Mean The arithmetic mean (or mean) is the most popular and easily understood measure of a central tendency; it provides descriptive information for normally distributed continuous data [14,15] (see Section 5.4). Consider a random continuous variable c and its values c1,c2,c3,.,cn for an n sample population. The arithmetic mean c is defined as: n P

c1 þ c2 þ c3 þ . þ cn i¼1 ¼ c¼ n n

ci

108

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

Example 5.3. The intake of red meat (g/day) was assessed in a cross-sectional study that enrolled 1P 00  ci ¼ 7500 g day. The mean intake of red meat in 100 participants and the sum of its values was i¼1

this sample population is estimated as: 1P 00

c¼

i¼1

n

ci ¼

7500 ¼ 75 g=day 100

One way to interpret this arithmetic mean is that the consumption of red meat in a randomly selected participant from the sample population is expected to be about 75 g/day. Many hypothesis tests and statistical tests are based on the arithmetic mean. As a measure of central tendency, it is affected by the outliers (i.e., observations that are distant from other observations in the sample) [14].

5.3.2.2 Range, Deviation, Mean Absolute Deviation, Variance, and Standard Deviation A measure of central tendency alone cannot give a full description of a continuous variable unless it is followed by a measure of statistical dispersion (i.e., a measure of variability). Consider a random continuous variable c and its values c1,c2,c3,.,cn for an n sample population. The most commonly used measures of variability are presented in Table 5.2. Example 5.4. Descriptive statistics will be calculated for the consumption of citrus fruits (g/day) in a sample of 10 participants. Sample observations are given in the table. Participant

Consumption of Citrus Fruits (g/day)

ci Lc

1 2 3 4 5 6 7 8 9 10

130 120 220 50 30 150 120 90 30 60

30 20 120 50 70 50 20 10 70 40



ci Lc

2

900 400 14,400 2500 4900 2500 400 100 4900 1600

The arithmetic mean is: 10 X i¼1

ci ¼ 130 þ 120 þ 220 þ 50 þ 30 þ 150 þ 120 þ 90 þ 30 þ 60 ¼ 1000

g day

5.3 DESCRIPTIVE STATISTICS

109

Table 5.2 Common Measures of Variability Range

R¼cmaximumcminimum

Deviation

D i ¼ ci  c

Mean absolute deviation

MAD ¼

  n P   ci c

Standard deviation

s2 ¼

s ¼

·

i¼1

n

 n P Variance

· Simplest expression of statistical dispersion. does not give information about how values · Range spread between maximum and minimum is signed distance of sample observation · Deviation from mean · Sum of all sample deviations is always 0 value of deviation (i.e., absolute distance of a · Absolute sample value from mean) is a useful measure for

2

calculatiing mean absolute deviation (MAD) MAD expresses average distance of sample observation from mean

is average squared distance of sample · Variance observation from mean of variance are square units of variable · Units is useful for calculating standard deviation · Variance that goes back to units of the variable as measured in

ci c

i¼1

n1

the sample

pffiffiffiffi s2

deviation has the same units of the variable as · Standard measured in the sample deviation is most-often used measure of · Standard variability for normal distributed continuous data (see Section 5.4) because an empirical rule defines that: 68% of sample observations lie within 1 standard deviation of the mean 95% sample observations lie within 2 standard deviations of the mean 99.7% sample observations lie within 3 standard deviations of the mean

· · ·

10 P

ci

c ¼ i¼1 n The range is

¼

1000 g ¼ 100 10 day

R ¼ cmaximum  cminimum ¼ 220  30 ¼ 190

The mean absolute deviation is

g day

 10  X   ci  c ¼ 30 þ 20 þ 120 þ j50j þ j70j i¼1

þ 50 þ 20 þ j10j þ j70j þ j40j g ¼ 480 day

110

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

 n  P   ci  c MAD ¼ i¼1 The variance is

n  X

ci  c

n

2

¼

480 g ¼ 48 10 day

¼ 900 þ 400 þ 14; 400 þ 2500 þ 4900

i¼1

þ 2500 þ 400 þ 100 þ 4900 þ 1600  2 g ¼ 32; 600 day 2 n  P ci  c s2 ¼ i¼1

n1

The standard deviation is

 2 32; 600 g ¼ 3622:22 10  1 day pffiffiffiffi pffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi g s ¼ s2 ¼ 3622:22 z 60:18 day

¼

The standard deviation is a measure of variability that mainly follows the arithmetic mean in the presentation of descriptive results for normally distributed continuous data [14,15]. In Example 5.4, and assuming that the consumption of citrus fruits in the sample population follows a normal distribution (see Section 5.4), the mean consumption can be reported as: 100  60.18 g/day (mean  standard deviation).

5.3.2.3 Median and Percentiles The median is a measure of central tendency that is efficient for presenting skewed continuous variables (i.e., variables that are not normally distributed) (see Section 5.4) [7,16]. To identify the median, sample observations should be ordered from the smallest to the largest. If n number of observations is odd, the median is the middle observed value and if n is even, the median is the average of the two middle values. In Example 5.4, we have n ¼ 10 (i.e., an even number of observations), so the median is the average of the fifth and sixth ordered observations (middle observations) from the smallest to the largest. In this case, the following order occurs: 30; 30; 50; 60; 90; 120; 120; 130; 150; 220 median ¼

90 þ 120 g ¼ 105 2 day

The main interpretation of the median is that 50% of sample observations lie below the median and the other 50% lie above it. The median is less sensitive to extreme values compared with the arithmetic mean. However, the mean has more sample information and is the basis for many hypothesis and statistical tests. The percentile is a statistical measure indicating the value below which a given percentage of sample observations lies. The 30th percentile, for instance, is the value below which 30% of observations may fall. The 50th percentile is the median (introduced earlier) and the 25th and 75th percentiles are also known as the first and third quartiles, respectively.

5.4 ASSESSMENT OF NORMALITY

111

When continuous data are presented descriptively as the median, it is preferable to report them along with the first and the third quartiles. This allows the reader to obtain a more comprehensive view of the data distribution.

5.3.2.4 Other Descriptive Statistical Measures There are a few special occasions in the statistical analysis of dietary data when other statistical measures of central tendency or dispersion are used, such as the geometric mean, harmonic mean, weighted arithmetic mean, mode, geometric median, standard error of the mean, coefficient variation, and interquartile range [17]. For the purposes of this chapter, we will not emphasize them.

5.4 ASSESSMENT OF NORMALITY Normal distribution (i.e., Gaussian distribution) is the most common continuous probability distribution; it is extremely useful in medical statistics [14,18e20]. Consider a random continuous variable N < c < N; the probability function of normal distribution is defined as: ðcmÞ2 1 f ðcÞ ¼ pffiffiffiffiffiffiffiffiffiffiffi e 2s2 2ps2 where N < m < N and s > 0 are respectively the mean and standard deviation of the distribution. Fig. 5.1 illustrates the normal curve N(m, s2). In a normal distribution, 68% of the area under curve lies within (ms, mþs), 95% lies within (m2s, mþ2s), and 99.7% lies within (m3s, mþ3s). Often, we need to calculate specific areas under the curve, which represent certain probabilities. A wide range of statistical techniques assume that the sample data follow a normal distribution. Many health- and nutrition-related variables seem to be normally distributed. However, it is essential

FIGURE 5.1 Normal curve N (m, s2).

112

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

for the nutrition researcher to run some normality tests after ensuring the quality of the data to assess the distribution of continuous variables. ShapiroeWilk’s test is one of the most common normality tests [21e23]. Suppose there is a random continuous variable c and its ordered values are c1,c2,c3,.,cn for an n sample population. The ShapiroeWilk’s test assesses the hypotheses: H0: The c is normally distributed. HA: The c is not normally distributed. ShapiroeWilk’s test can be calculated by:  n 2 P ai ci i¼1 W¼ n  2 P ci  c i¼1

where ai are constants generated from the sample data. Most statistical software calculates the value of W and the related pvalue. When pvalue > a (i.e., the statistical significance level), there is statistical evidence not to reject H0 and it supports the concept that the random continuous variable c is normally distributed. KolmogoroveSmirnov test is another goodness-of-fit test frequently applied to assess the normality of continuous data [24]. Hypothesis testing is similar to that for the ShapiroeWilk test; when pvalue > a, there is statistical evidence that the random continuous variable is normally distributed. This test is also generated by most available statistical software. Studies have found that even in its best form, the KolmogoroveSmirnov test is less efficient for assessing normality than the ShapiroeWilk test [25]. In addition, some visual methods are particularly useful for assessing the normality of data and confirming previous tests. A first, common approach is to generate a histogram of the continuous variable and observe the distribution of the data, which should be bell-shaped and resemble the normal distribution [14,18,20]. Another graphical test is the quantile-quantile (Q-Q) plot, i.e., the plot of the standardized sample data against the standard normal distribution [26]. For normally distributed data, ideally the generated Q-Q plot points should fall into a reasonably straight diagonal line from bottom left to bottom right. A third visual method is the PeP plot i.e., the plot of the cumulative probability of a variable against the cumulative probability of the normal distribution [27]. Ideally, PeP plot points of normally distributed data should also fall into a reasonably straight diagonal line from bottom left to bottom right. When the results of statistical tests assessing the normality of sample data agree with the outcome of visual methods for the same variables, the researcher can be more confident about his decision. Usually, dietary data such as the consumption of foods or nutrients cannot approximate a normal distribution [28e31]. This may also occur for other collected continuous data (i.e., blood biomarkers, clinical or lifestyle characteristics) or in situations in which the sample size is relatively small. Thus, it is essential to run a comprehensive normality check before analyzing data. More precisely, it is recommended to apply a combination of normality statistical tests and visual methods. When the distribution of a continuous variable is skewed (i.e., not normal), the researchers should pay special attention to the way they treat these data during other statistical procedures. In addition, as mentioned in Section 5.3.2.3, the median should be preferred for a descriptive presentation of the data instead of the mean.

5.5 CONFIDENCE INTERVAL

113

5.5 CONFIDENCE INTERVAL In Section 5.3, the concept was introduced of various descriptive statistical measures that can be calculated in a sample population. Some are also known as point estimators. This term describes a single value calculated from sample data, which is the best estimate of an unknown population parameter [7]. For instance, the arithmetic mean c is the point estimator of the unknown mean m of the reference population. In reality, it is impossible to measure the true value of m in a large reference population. For this reason, after appropriate sampling, the point estimator c is calculated from the collected sample data. The basic idea in the statistical inference is to generate a CI for the point estimator of a population parameter [32]. This interval represents an estimated range of values for the point estimator, which is likely to include the unknown population parameter with a certain level of confidence. Generally, a CI has the following form: CI ¼ Point estimator  Margin of error To facilitate the presentation of the CI, let us assume that for a normally distributed variable with sample mean c, its standard deviation s in the reference population is already known. In most real-life statistical analyses, the population s is unknown; instead, the sample standard deviation s is used. For a (1a)100% level of confidence in which a is the statistical significance level, the CI can be calculate by: s c  za=2  pffiffiffi n s ffiffi p In this situation, c is the point estimator, za=2  n is the margin of error, and n is the sample size. za/2 is often called the critical point; it very much depends on a and its value can be found in tables of the standard normal distribution [33]. Fig. 5.2A shows where the values za/2 and za/2 are placed in a standard normal distribution. Because the statistical significance level is usually set at a ¼ 0.05, the confidence level is calculated as (10.05)100% ¼ 95%. A 95% CI is most frequently reported in the literature, although sometimes a 99% CI is provided. For a 95% confidence level, the critical values za/2 and za/2 from standard normal distribution tables are z0.025 ¼ 1.96 and z0.025 ¼ 1.96 [33], as shown in Fig. 5.2B. Suppose for the same normally distributed variable with a mean c, standard deviation s is unknown. Then, the (1a)100% CI is calculated using the sample’s standard deviation s: s c  ta=2  pffiffiffi n Here, for the point estimator c, the margin of error is ta=2  psffiffi. The ta/2 critical value comes from n

the t distribution tables with n1 degrees of freedom. The t distribution is a useful and frequently applied theoretical continuous data distribution in statistical inference; it has a bell curve form similar to that of a normal distribution. For a 95% CI, the value of ta/2 ¼ t0.05/2 ¼ t0.025 can be found from the t distribution tables with n1 degrees of freedom, i.e., the value depends on the n sample size [33]. The same concept of the CIs is applied to a variety of point estimators, some of which will be introduced later in this chapter and in Chapter 6, such as the linear regression bb coefficient, the OR, and the risk ratio. For the desired confidence level, most statistical software calculates the related CIs for various point estimators.

114

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

FIGURE 5.2 Values za/2 and za/2 of (A) standard normal distribution and (B) confidence intervals.

Example 5.5. The mean consumption of citrus fruits in a sample of 16 participants was 100 g/day. Suppose the population’s standard deviation s is known and equal to 40 g/day. The researchers are interested in calculating a 95% CI for the mean. Because the standard deviation of the population is known, the 95% CI can be calculated: s c  za=2  pffiffiffi n For za/2 ¼ 1,96 given from the standard normal distribution tables, the 95% CI is: 40 40 100  1:96  pffiffiffiffiffi ¼ 100  1:96  pffiffiffiffiffi ¼ 100  19:6 16 16 Thus, the 95% CI is (10019.6, 100þ19.6) or (80.4 g/day, 119.6 g/day). This means that with a 95% confidence level, the mean m of the reference population ranges within (80.4 g/day, 119.6 g/day).

5.6 PEARSON CHI-SQUARE TEST

115

5.6 PEARSON CHI-SQUARE TEST The Pearson chi-square (c2) test is used to evaluate relations among categorical data [34e38]. More precisely, it can provide us with statistical evidence to support the independence or not of two categorical variables. Suppose there are two random categorical variables, k and l, with p and q categories, respectively, measured in an n sample population. Table 5.3 presents the number of observations Xij recorded in each cell of a q  p table of the variables. The Pearson chi-square test evaluates the hypotheses: H0: The k and l variables are independent. HA: The k and l variables are not independent (i.e., they are related) The chi-square test is calculated by: S¼

q X p X ðXij  Eij Þ2 Eij i¼1 j¼1

where Eij is the expected number of observations in each cell of the q  p table with the hypothesis that k and l are independent (Table 5.3). Eij is calculated by: Table 5.3 Pearson’s Chi-Square Test Frequencies Tables Observations Recorded in p and q Categories of k and l Variables. q 3 p Table Categories of Variable l

Categories of Variable k

1 2 3 . q Sum

1

2

3

.

p

Sum

X11 X21 X31 . Xq1 V1

X12 X22 X32 . Xq2 V2

X13 X23 X33 . Xq3 V3

. . . . . .

X1p X2p X3p . Xqp Vp

S1 S2 S3 . Sq n

Estimated observations in p and q categories of k and l variables, under the hypothesis of the independence of k and l. q 3 p Table Categories of Variable l

1 2 3 . q

Categories of variable k 1

2

3

.

E11 ¼ S1nV1 E21 ¼ S2nV1 E31 ¼ S3nV1 .

E12 ¼ S1nV2 E22 ¼ S2nV2 E32 ¼ S3nV2 .

E13 ¼ S1nV3 E23 ¼ S2nV3 E33 ¼ S3nV3 .

. . . . .

Eq1 ¼

Sq V1 n

Eq2 ¼

Sq V2 n

Eq3 ¼

Sq V3 n

p S V

E1p ¼ 1n p S V E2p ¼ 2n p S V E3p ¼ 3n p . Eqp ¼

Sq Vp n

116

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

S i Vj i ¼ 1; 2; 3; .; q and j ¼ 1; 2; 3; .; p n Most statistical software calculates the value of Eij, S, and the related pvalue. When pvalue < a (i.e., the statistical significance level), there is statistical evidence to reject H0 and to support the concept that the two categorical variables are dependent. In other words, there is statistical evidence that the two categorical variables are related. To apply the Pearson chi-square test, we need two random categorical variables. To trust the results of this test, an assumption is made that more than 80% of cells in the q  p table should have five or more observations (Eij). For a 1  2, 2  1, or 2  2 table, it is recommended that the expected number of observations (Eij) be at least 10. When this assumption is violated in a 2  2 table, many statistical software packages calculate the Fisher exact probability test, which can be used instead [17]. Eij ¼

Example 5.6. The presence of diabetes (binary outcome: yes/no) and the consumption of leafy vegetables (binary exposure: consumers/nonconsumers) were assessed in a cross-sectional study that enrolled 1000 participants. The following 2  2 table illustrates the distribution of observed data within the categories of the two variables. The researchers are interested in evaluating the relation between the two categorical variables.

Diabetic participants Nondiabetic participants Sum

Consumers of Leafy Vegetables

Nonconsumers of Leafy Vegetables

Sum

50 600 650

150 200 350

200 800 1000

The following 2  2 table for the expected number of observations Eij ¼

Si V j n

can be formulated:

Consumers of Leafy Vegetables

Nonconsumers of Leafy Vegetables

Diabetic participants

E11 200,650 1000 ¼ 130

E12 200,350 1000 ¼ 70

Nondiabetic participants

E21 800,650 1000 ¼ 520

E22 800,350 1000 ¼ 280

Then the Pearson chi-square test is: S¼

2 X 2 X ðXij  Eij Þ2 ð50  130Þ2 ð150  70Þ2 ð600  520Þ2 ð200  280Þ2 þ þ þ ¼ 175:82 ¼ Eij 130 70 520 280 i¼1 j¼1

Suppose that the statistical software provided for this value of the chi-square test is a related pvalue ¼ 0.001 < 0.05 ¼ a. The results indicated that there is statistical evidence to reject H0 and to support the concept that the two categorical variables are dependent. It can be also stated that there is statistical evidence to support that the presence of diabetes is related to the consumption or

5.7 STATISTICAL TESTS FOR COMPARISON OF MEANS

117

nonconsumption of leafy vegetables. In fact, the presence of diabetes among consumers of leafy 50 ,100 ¼ 7:7%, much greater than this among non-consumers 150,100 ¼ 42:9%. vegetables was 560 350 Pearson chi-square test is useful for analyzing data from caseecontrol nutrition studies because for both cases and controls (i.e., binary outcome), the level of past exposition to a dietary factor (categorical variable with two or more categories) is evaluated.

5.7 STATISTICAL TESTS FOR COMPARISON OF MEANS 5.7.1 T TEST A number of t tests are available. The independent-samples t test, also known as the Student t test [39e41], is most frequently applied in retrospective nutrition data analysis and will be discussed here. Suppose there is a random continuous variable c that is normally distributed, and c1 and c2 are the mean values of two independent sample groups, n1 and n2. The independent-samples t test assesses the hypotheses: H0 : c1 ¼ c2 HA : c1 sc2 Simply speaking, it assesses whether the means of these two independent sample groups are statistically equal. The independent-samples t test can be calculated by: t¼

c1  c2 rffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1 1 sp þ n1 n2

where sp is estimates the pooled standard deviation of the two samples groups: sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi ðn1  1Þs21 þ ðn2  1Þs22 sp ¼ n1 þ n2  2 Most statistical software calculates the value of t and the related pvalue. When pvalue7 times/week” (pvalue < 0.05). An ANOVA test suitable this time for dependent samples, the repeated-measures ANOVA, will be discussed in Chapter 6. It is mainly applied to compare means for repeated measures (i.e., prospective data analysis) of a random, normally distributed continuous variable measured three or more different times in the same sample.

5.8 PEARSON CORRELATION COEFFICIENT The concept of correlation analysis may be appropriate when the relation between continuous variables is studied. The statistical measure usually applied to control for the correlation between two random, normally distributed continuous variables is the Pearson correlation coefficient [16,44,45]. Suppose there are n observations of two random, normally distributed continuous variables, k and l. The Pearson correlation coefficient is defined as: P  n  n  P ki  k li  l i¼1 i¼1 s ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi rkl ¼ sffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi 2 P 2 n  n  P ki  k li  l i¼1

i¼1

The rkl has values from 1 to 1. When rkl > 0, k and l are positively correlated, i.e., as one variable increases, so, too, does the other. When rkl < 0, k and l are negatively correlated, i.e., as one variable increases, the other decreases. When rkl z 1 or rkl z 1, there is perfect positive or negative linear correlation between the respective variables. On the other hand, when rkl z 0, there is no correlation. The test of the following hypotheses can provide evidence regarding the statistical significance of the Pearson correlation coefficient: H0 : rkl ¼ 0 HA : rkl s0 The statistic that is used to test this hypothesis is: pffiffiffiffiffiffiffiffiffiffiffi rkl n  2 S ¼ qffiffiffiffiffiffiffiffiffiffiffiffiffiffi 1  r 2kl Most statistical software calculates the Pearson correlation coefficient rkl, the value of the S statistic, and the related pvalue. When pvalue < a (i.e., the statistical significance level), there is statistical evidence to reject H0 and to support the concept that rkl s 0. In other words, there is statistical evidence to support that there is a significant correlation between k and l. The sign of the rkl defines the direction of the correlation, i.e., positive or negative.

5.9 NONPARAMETRIC TESTS

121

To apply the Pearson correlation coefficient test, we need two random continuous variables that are normally distributed. We always need to remember that this statistic cannot be applied in skewed (not normally distributed) data. In that case, it is better to use nonparametric statistical tests (see Section 5.9). Example 5.9. The fasting blood glucose levels (mg/dL) and consumption of leafy vegetables (g/day) were measured in a cross-sectional study of 1000 participants. The researchers are interested in evaluating the correlation between the two continuous variables. Note: The normality tests for the two variables showed evidence to support that both are normally distributed in our sample. Now suppose that the analysis of the sample data provided a Pearson correlation coefficient of r ¼ 0.65 and a related pvalue ¼ 0.004. The results indicated that there is statistical evidence of a negative (r ¼ 0.65 < 0) significant correlation between the fasting blood glucose levels and leafy vegetable consumption (pvalue ¼ 0.004 < 0.05 ¼ a). It can be also stated that an increase in leafy vegetable consumption was associated with a decrease in fasting blood glucose levels.

5.9 NONPARAMETRIC TESTS In Sections 5.7 and 5.8, the application of statistical tests assumed that the continuous data were normally distributed (see Section 5.6). Often in retrospective data analysis, this assumption is not fulfilled by dietary or other collected data. On these occasions, corresponding nonparametric tests should be applied [46e48]. The same should be followed when the sample size is very small (i.e., empirically, n < 30). Table 5.4 presents parametric tests and their nonparametric equivalents. The calculation of most nonparametric tests does not use the exact values of continuous sample data but rather their corresponding ranks (i.e., ranked data). Hypothesis testing and the interpretation of the results follow the same pattern as the corresponding parametric tests. Most statistical software generates these tests and the related pvalues.

Table 5.4 Parametric and Nonparametric Equivalent Tests Parametric Test Student t test

One-way analysis of variance

Pearson correlation coefficient

Type of Data random binary variable that describes · One the two independent sample groups · One random continuous variable · One random categorical variable with

· ·

more than two categories that describes independent-sample groups One random continuous variable Two random continuous variables

Nonparametric Equivalent Test ManneWhitney U test

KruskaleWallis test

Spearman rank correlation coefficient

122

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

For instance, Example 5.8, the fasting blood glucose levels they measured in three independentsample groups of leafy vegetables consumption were also normally distributed. If this assumption was not fulfilled, the researchers should use the KruskaleWallis test instead of the one-way ANOVA F test. For a related pvalue < a, the researchers could then state that there is statistical evidence to reject H0 and support the concept that the fasting glucose levels are significantly different among the groups of leafy vegetable consumption. Researchers should not avoid using nonparametric tests when the circumstances require doing so.

5.10 LINEAR REGRESSION ANALYSIS Earlier in this chapter, the concept of the correlation or relation between two random variables was introduced. Linear regression analysis is a statistical method that can be employed to explore and model the relations comprehensively between a random continuous variable, which is called a dependent variable, and one or more random variables (continuous or categorical), which are called independent variables or predictors [49e52]. It is a major statistical methodology that can be applied to a retrospective study’s dataset and evaluate potential associations between a dietary exposure and a health outcome [4,5]. It requires the health outcome under investigation to be measured on a continuous scale (i.e., fasting blood glucose levels, systolic blood pressure, body mass index [BMI]). In this analysis format, the health outcome has the role of the dependent variable, and dietary exposure the role of the predictor. The major objective of linear regression analysis when it is applied in health and nutrition data is to evaluate the degree of the association between a dietary factor and a health outcome. The predictive ability of the dietary exposure on the health outcome can also be estimated. These cannot be achieved by correlation coefficients (Pearson and Spearman; see Sections 5.8 and 5.9), which may provide us with information only about the presence and the direction (i.e., positive or negative) of an association between two random continuous variables. Moreover, compared with correlation coefficients, linear regression modeling is a more comprehensive statistical technique with many options that allows the assessment of possible confounding effects or interactions in an observed association by other sample characteristics. Interpretation of the results generated from this analysis should be performed with caution considering the nature of the collected data. As mentioned earlier in Chapter 1, naturally, a retrospective study cannot provide evidence about the presence of a cause-and-effect association owing to the limitations derived from its design. Nevertheless, the results from using linear regression analysis in retrospective surveys are highly influential and can direct further data analysis and hypothesis generation in the field. In the following discussion, the simple and the multiple forms of linear regression analysis are presented.

5.10.1 SIMPLE LINEAR REGRESSION 5.10.1.1 Definition Consider the observed Yi and Xi values in an n sample population for the random continuous variables Y and X of the reference population. A scatter bidimensional plot of these values can be, for instance, similar to that presented in Fig. 5.3A.

123

Simple linear regression analysis (fit and interpretation). (A) Scatter bidimensional plot of Yi and Xi (B) Fitted linear regression function for scatterplot points ðXi ; Yi of (A) (C) Residuals, as the signed vertical distances between the points and the regression line (D) Graphical interpretations of bb0 and bb1 .

5.10 LINEAR REGRESSION ANALYSIS

FIGURE 5.3

124

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

A simple linear regression analysis is the simplest version of the linear regression family, which estimates a linear function (i.e., linear model) of Y and X, given the sample’s pairs of observations (Xi,Yi) [53]. A major objective of the estimation of this function is to explore the degree and statistical significance of the association between Y and X. A secondary aim is to use this function to forecast an unobserved value of Y for a given value of X. The theoretical form of the linear function is: Y ¼ b0 þ b1 X þ ε where b0and b1 are values estimated from applying linear regression analysis to the sample data; and εis the random error or noise representing the i sample values of Y, which vary in the line Y ¼ b0 þ b1X. The model estimated from simple linear regression analysis is: Ybi ¼ bb0 þ bb1 Xi where Ybi are the predicted values of Yi, i ¼ 1,2,3,.,n calculated by linear regression analysis; and bb0 and bb1 are the estimated values of b0 and b1 calculated by linear regression analysis. The bb1 is also called the coefficient of X and the bb0 is known as the constant of the model. The random error term ε is not required in this formula because the predicted Yei values all fall into the predicted line. Fig. 5.3B illustrates the fitted linear regression function for the scatterplot points of Fig. 5.3A.

5.10.1.2 Least-Squares Approach

To fit the linear regression model in the sample data and estimate the values for be0 and be1 , the most commonly applied method is the least-squares approach [7,16,54]. Consider ei, the residuals of Yi observed values, and their definition to be: ei ¼ Yi  Yei Fig. 5.3C helps us understand that the residuals are the signed vertical distances between the points (Xi,Yi) and the regression line. The best-fitted line for the least-squares approach is one that minimizes 2 n n  P P e2i ¼ the sum of the squared vertical distances, i.e., the Yi  Yei . i¼1

i¼1

With this rule, it can be proved that be0 and be1 are calculated as: n P

be0 ¼ i¼1

Xi2 n

n P

i¼1 n P

i¼1

n be1 ¼

Yi 

n P

Xi2 

Xi Y i 

i¼1 n P

n

i¼1

Xi2 

n P

Xi

n P

Xi Y i

i¼1 i¼1  n 2 P Xi i¼1 n P

Xi

n P

Yi

i¼1 i¼1  n 2 P Xi i¼1

5.10 LINEAR REGRESSION ANALYSIS

125

Simple linear regression analysis with the least-squares approach has specific assumptions for the error term ε. It is assumed that ε: 1. is a random variable with a mean of 0, 2. is normally distributed, and 3. has a constant variance s2 at every value of X. The error terms are also assumed to be independent. It can also be stated that because ε should be normally distributed, the same should be true for the Y variable.

5.10.1.3 Interpretation of be0 and be1

be0 expresses the expected value of Y for X ¼ 0; depending on the definition of X and the range of the observed data, it may or may not have a practical meaning. be1 is the slope of the regression line; it also expresses an increase or decrease in the mean of Y for a one-unit increase in X when X is a random continuous variable. For a two-unit increase in X, the increase or decrease in the mean of Y is 2  be1 , and so on. The sign of be1 defines whether there is an increase or decrease (i.e., an increase for “” and a decrease for “ “). The interpretations of be0 and be1 are presented graphically in Fig. 5.3D. When X is a categorical variable with l categories, an l  1 number of be1 terms are estimated from linear regression analysis. One of the l categories is set as the reference category (usually the first or the last) by the researcher. In this analysis format, each of the estimated be1 terms shows the increase or decrease in the mean of Y between each category of X and the reference category. These interpretations of be1 are explained better in Example 5.10 (see Section 5.10.1.6). be1 is also a measure of the strength and significance of the association between X and Y. If be1 z0, Y is equal to the constant be0 and there is no evident linear association between X and Y. The following hypotheses are always tested in a simple linear regression analysis: H0 : be1 ¼ 0 HA : be1 s0 A t test statistic is calculated along with a related p  value [7,16,54]. When pvalue < a (i.e., the statistical significance level), there is statistical evidence to reject H0 and support the concept that be1 s0. In other words, there is statistical evidence that there is a significant association between Y and X. The sign of be1 informs us about whether there is a positive or negative association. The concept of the CI introduced earlier (see Section 5.5) can be applied to be0 and be1 as point estimators derived from the sample data. Most statistical software calculates the CI for various confidence levels whereas a 95% CI is most frequently reported. The CI is also used as a proxy of the statistical significance of the b coefficients. When the critical value 0 does not fall within the CI, there is evidence that the b coefficient is statistically significant. The opposite occurs when the CI overlaps with the null value.

5.10.1.4 Assumptions and Coefficient of Determination Before interpreting the results, it is essential to check that all assumptions for applying a simple linear regression analysis are fulfilled by the sample data. The first step is to draw a scatter bidimensional plot of X and Y, which can give us valuable information about the existence of linearity in the association of the two variables (Fig. 5.4A and B). By observing the distribution of the (Xi,Yi) points, we may

126

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

FIGURE 5.4 Control of assumptions of simple linear regression analysis. (A) Scatter bidimensional plot of the X and Y, nonlinear pattern (B) Scatter bidimensional plot of the X and Y, linear pattern (C) P-P plot of standardized residuals, non-normaly distributed residuals (D) P-P plot of standardized residuals, normally distributed residuals (E) Plot of the Xi values against the standardized residuals, graph points non-equally distributed below and above the straight line starting from 0 (F) Plot of the Xi values against the standardized residuals, graph points equally distributed below and above the straight line starting from 0.

5.10 LINEAR REGRESSION ANALYSIS

127

understand whether a nonlinear (Fig. 5.4A) or linear pattern is evident (Fig. 5.4B). If a nonlinear pattern is observed (i.e., logarithmic, exponential), the linear form of regression is inappropriate and results from the analysis cannot be interpreted properly. The same scatter can inform us about the presence of outliers that may have not been identified during the data management procedures. The next step is to assess assumptions related to the distribution of the residuals, ei, which should be normally distributed with a mean of 0 and a constant variance s2 (see Section 5.10.1.2). These checks can be facilitated by generating appropriate graphs. To perform graphical tests, we first need to ask the statistical software to generate the corresponding residuals from an estimation of the linear function of the simple regression analysis. In addition to the basic definition of residuals (see Section 5.10.1.2), other more effective forms are available, such as standardized, studentized, and jackknife residuals. Among them, standardized residuals are the most popular form used in nutrition research; most statistical software packages can calculate them [17,55e57]. A P-P plot of standardized residuals (Fig. 5.4C and D) informs us about whether they are normally distributed. If all points on the graph fall into a reasonably straight diagonal line from bottom left to bottom right (Fig. 5.4D), the standardized residuals are normally distributed. An unforeseen situation is illustrated in Fig. 5.4C, in which most points fall below the diagonal line. The normality of residuals may also be evaluated by applying the ShapiroeWilk test, as described in the section on Assessment of Normality. Furthermore, a plot of Xi values against standardized residuals (Fig. 5.4E and F) can provide information about the variance and the mean of the residuals. It is expected that all points on the graph will be equally distributed below and above the straight line starting from 0 (Fig. 5.4F). This expectation originates from the assumption that the residuals are expected to sum to 0, and so to have a mean of 0. In addition, the residuals should have a constant variance, which means that constant variability of the residuals should be observed for all values of X. This assumption in covered in Fig. 5.4F, in which all the points of graphs lie within the two light straight lines that have been drawn. On the contrary, in Fig. 5.4E, the variance of the residuals increases with X, which violates the constant variation assumption. This situation is common and we should be careful before interpreting the results. In addition to verifying linear regression analysis assumptions, it is common to calculate another statistical measure called the coefficient of determination, or simply Rsquared (R2). This expresses the proportion of variance of the dependent variable that can be predicted from the independent variable [7,16,54]. This statistic ranges from 0 to 1 and measures the goodness-of-fit for the predicted regression line. Theoretically, when R2 has a value of 1, all sample points (Xi,Yi) should fall exactly into a predicted line revealing a perfect fit. Generally, greater values of R2 show a better fit of the regression line to the sample data. Often, R2 is expressed as a percentage; for instance, R2 ¼ 0.14 shows that 14% of variance of the dependent variable can be explained by the independent variable.

5.10.1.5 Data Transformations Unfortunately, assumptions about applying linear regression analysis may not be fulfilled by the sample data. The symmetric normal bell-shaped distribution may not adequately describe the dietary data, or the health outcomes may happen to have various skewed forms. To overcome these problematic situations, researchers often use techniques to transform a skewed variable to one that approximates the normal symmetric bell shape. Transformation of the original

128

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

data requires to a mathematical operation to be performed on each observation and transformed values to be used for the purposes of the analysis. The most popular approach is the log-transformation of a skewed continuous variable, in which the log-values of the observations are calculated and used as a log-transformed variable in the linear regression analysis. Usually, the variable that happens to be transformed is the dependent variable; however, in more complex models and in multiple linear regression analysis (see Section 5.10.2), it is useful to transform both the dependent variable and the exposures. The log-transformed variables often seem to be effective in meeting the assumptions of linear regression analysis. However, interpretation of the results should be performed with special attention considering that a log-transformed variable does not keep its original measurement units. When these data are employed, interpretation of the estimated parameters of the model refers to the logtransformed variables and not the original forms. Another approach to meeting the assumptions of the linear regression is the square root transformation of the skewed positive data. The square rootetransformed variables can be used in modeling procedures, although again interpretation of the results needs special attention. When skewedness is present in the dietary exposure, continuous data can be transformed into categorical data expressing, for instance, groups of food consumption or nutrient intake. The cutoff points for categorizing the observations into groups may follow some literature data, or the variable’s percentiles can be used (see Section 5.3.2.3). This procedure often appears to be helpful and always ensures a solution. An alternative to dietary exposures with skewed distributions is first to analyze them together by applying dietary pattern analysis methods; the extracted dietary patterns may be introduced later into the linear regression analysis (see Chapter 4). Moreover, when skewedness is present in the health outcome (for instance, in a blood biomarker as the c-reactive protein), appropriate cutoff points defining the presence or severity of a disease or health condition may be used to group sample members into two or more categories. Later, this new variable can be used in other statistical methods such as the logistic regression analysis (see Section 5.11).

5.10.1.6 Application to Dietary Data and Examples Simple linear regression analysis is the most frequently adopted statistical method in cross-sectional observational settings, when a health outcome given in the form of a continuous variable and a dietary factor are tested for their association. In this analysis format, Y represents the health outcome and X represents the dietary exposure. For instance, the health outcome can be blood glucose levels (mg/dL), systolic blood pressure (mmHg), or BMI (kg/m2) and the dietary exposure can be daily food group consumption (grams), daily macronutrient or micronutrient intake (various units applied), categories of the frequency of the recorded consumption of a food group or the dietary pattern extracted by an “a priori” or “a posteriori” dietary pattern analysis (see Chapter 4). Example 5.10. In a cross-sectional study of 1000 participants, fasting blood glucose levels (mg/dL) and consumption of leafy vegetables (g/day) were measured. (a) The researchers are interested in modeling their association. The fasting blood glucose level is a health parameter measured on a continuous scale; thus, we may assume that simple linear regression analysis may be used.

5.10 LINEAR REGRESSION ANALYSIS

129

The first step is to draw a bidimensional scatter plot of actual sample measurements of glucose levels and leafy vegetable consumption. From this graph, we can observe whether a linear pattern is present that correlates between the health outcome and the dietary factor. For simplicity, let us assume that this is true. In this example, simple linear regression analysis may be used to model the investigated association. The basic results of the analysis are given in the following table.

Leafy vegetables consumption (g/day) Constant (be0 )

bb

P Value

0.20 110.1

.003 .002

95% Confidence Interval 0.35 98.3

0.10 121.2

R2 0.025

Note: The P values are derived from the corresponding t tests.

The estimated model is: Yei ¼ 85:1  0:2Xi where Xi is leafy vegetable consumption (g/day) evaluated in the sample population; i ¼ 1,2,3,.,1000 (i.e., n ¼ 1000); and Yei are the corresponding predicted values of the fasting blood glucose levels (mg/dL). Before interpreting the results, it is essential to examine whether all assumptions for applying simple linear regression analysis are fulfilled. For this reason, we need to ask the statistical software to predict the residuals of the analysis. For simplicity, let assume that the graphical test (P-P plot) and ShapiroeWilk test indicated that the residuals are normally distributed. Also, the plot of the Xi values against the standardized residuals showed that all points on the graph were equally distributed below and above the straight line starting from 0 and that the residuals have a constant variance. Thus, we are allowed to proceed in interpreting the results. The analysis of the data showed that leafy vegetable consumption was significantly negatively associated with the fasting blood glucose levels (b coefficient ¼ 0.02 < 0; P ¼ .003 < 0.05). More precisely, a 1-g/day increase in leafy vegetable consumption was associated with a 0.2-mg/dL decrease in fasting blood glucose. This estimation may vary in the reference population within (0.10, 0.35) mg/ dL with a 95% level of confidence (i.e., the absolute value of the 95% CI for the b coefficient). The corresponding decrease for a 10-g/day increase in leafy vegetable consumption is 2 mg/dL (i.e., bcoefficient for 10 units increase ¼ 10  (0.2) ¼ 2.0) and the related 95% CI is (0.10  10 ¼ 1.0, 0.35  10 ¼ 3.5) mg/dL. Leafy vegetable consumption seems to account for 2.5% (R2 ¼ 0.025) of variability in fasting blood glucose levels. In addition, mean predicted fasting blood glucose for nonconsumers of leafy vegetables (i.e., leafy vegetable consumption ¼ 00Xi ¼ 0) is 110.1 mg/dL (i.e., be0 ); this estimation may vary in the reference population within (98.3, 121.2) mg/dL with a 95% confidence level (i.e., the 95% CI for be0 ). (b) Suppose that the study participants are grouped into three categories according to their consumption of leafy vegetables: First category: 0e10 g/day. Second category: 10e30 g/day. Third category: 30 g/day.

130

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

The researchers are again interested in modeling the association between the dietary factor and the health outcome. In this example, and to facilitate the calculations, we need to specify for the statistical software which of the three categories should be considered the reference category. Suppose the first category is set as the reference category; the results are given in the following table. bb Groups of leafy vegetable consumption 0e10 g/day 10e30 g/day 30 g/day Constant (be0 )

P Value

95% Confidence Interval

R2 0.025

Reference category 2.55 .001 4.80 .003 112.4 .002

3.81 6.78 99.5

1.72 3.51 124.5

Note: The P values are derived from the corresponding t tests. For simplicity, suppose the assumptions for applying the simple linear regression are fulfilled.

Similar to (a), analysis of the data showed that leafy vegetables were significantly negatively associated with the fasting blood glucose levels in all categories of consumption (b coefficients < 0; P values < .05). In fact, participants at 10e30 g/day leafy vegetable consumption had 2.55 mg/dL (b coefficient ¼ 2.55; P value ¼ .001) lower levels of fasting blood glucose compared with those consuming 0e10 g/day (i.e., reference category). This estimation may vary in the reference population within (1.72, 3.81) mg/dL with 95% level of confidence (i.e., the absolute value of the 95% CI for the b coefficient). In addition, participants in the 30-g/day consumption group had 4.80 mg/dL (b coefficient ¼ 4.80; P value ¼ .003) lower levels of fasting blood glucose compared with those consuming 0e10 g/day and the corresponding 95% CI was 3.51e6.78. A doseeresponse association was observed between the dietary exposure and the health outcome.

5.10.2 MULTIPLE LINEAR REGRESSION ANALYSIS 5.10.2.1 Definition Multiple linear regression analysis is the extended version of the simple form of analysis [7,16,54], in which the dependent variable Y is tested against a set of two or more independent variables, Xj, j ¼ 1,2,.,k. The theoretic form of the function is: Y ¼ b 0 þ b 1 X 1 þ b 2 X2 þ . þ b k X k þ ε where b0and bj, j ¼ 1,2,.,k are values estimated from applying multiple linear regression analysis to the sample data; and ε is random error or noise. Now, consider the observed Yi and Xij values in an n sample population for the random variables Y and Xj of the reference population. The model estimated from multiple linear regression analysis is: Yei ¼ be0 þ be1 Xi1 þ be2 Xi2 þ . þ bek Xik where Yei are the predicted values of Yi, i ¼ 1,2,3,.,n calculated by multiple linear regression analysis. be0 and bej , j ¼ 1,2,.,k are the estimated values of b0 and bj calculated by multiple linear regression analysis.

5.10 LINEAR REGRESSION ANALYSIS

131

Similar to simple linear regression analysis, an extended version of the least-squares approach (see Section 5.10.1.2) is used to estimate the be0 and bej coefficients by the sample data. However, because the description of this approach is advanced for a nonmathematical audience, we will limit its presentation by pointing out that most available statistical software delivers all of the required calculations with excellent precision.

5.10.2.2 Interpretation of be0 and bej

Interpretation of bb coefficients calculated by simple and multiple linear regression analysis presents similarities (see Section 5.10.1.3). The be0 expresses the expected value of Y for X1 ¼ 0, X2 ¼ 0, ., Xj ¼ 0. The be1 coefficient of the X1 continuous variable expresses an increase or decrease in the mean of Y for a one-unit increase in X1 given that X2, X3, ., Xj remain constant values. The sign of be1 defines whether there is an increase or decrease (i.e., an increase for “” and a decrease for ““). Similarly, the be2 coefficient shows an increase or decrease in the mean of Y for a one-unit increase in X2 given that X1, X3,., Xj remain constant values. Interpretation of all bej coefficients of every continuous variable Xj follows the same logic. If a X predictor is a categorical variable with l categories, suppose for X3 that l1 number of be3 terms are estimated from multiple linear regression analysis. One of the l categories is set as the reference category (usually the first or the last) by the researcher. In this analysis format, each of the l1estimated be3 terms expresses an increase or the decrease in the mean of Y between each category of X3 and the reference category given that the X1, X3,., Xj remain constant values. Each of the bej coefficients is also a measure of the strength and the significance of the association between each of the Xj predictors and the Y. If a bej z0, there is no evident association between the corresponding Xj predictor and Y. For each of the bej coefficients, the following hypotheses are tested: H0 : bej ¼ 0 HA : bej s0 A t test statistic is calculated along with a related pvalue [7,16,54]. When pvalue < a (i.e., the statistical significance level), there is statistical evidence to reject H0 and support the concept that bej s0. In other words, there is statistical evidence that there is a significant association between the corresponding Xj predictor and Y, considering all other Xj predictors. The sign of bej informs us about whether there is a positive or negative association.

5.10.2.3 Assumptions Before interpreting the results, it is essential to verify whether all assumptions for applying the multiple linear regression analysis are covered. This requires us to ask the statistical software to calculate the residuals and verify their normality, homoscedasticity, and independence using the same method described in simple linear regression analysis (see Section 5.10.1.4). In the case of multiple linear regression analysis, it is also essential to the check for the existence of multicollinearity among the predictors of the model (i.e., the set of Xj independent variables). The phenomenon of multicollinearity (known also as collinearity) frequently appears in nutrition research; it can be characterized by a high level of association between predictors [30,58e60]. More precisely, multicollinearity appears when one predictor in a multiple regression model can be linearly predicted by the others with a substantial degree of accuracy.

132

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

When the research hypothesis requires a set of two or more dietary predictors to be tested for their association with a health outcome, frequently these dietary factors are highly correlated [30,59,60]. For instance, when different types of polyphenols are tested for their association with a health outcome, it is true that dietary intakes in a sample population are highly correlated, because they derive from the consumption of common food sources (i.e., fruits, vegetables) [60,61]. To overcome this problem, various approaches are available. First, dietary predictors can be modeled separately for their association with the outcome and to extract the related conclusions. A second approach is to summarize their values mathematically when the nature of the dietary factors allows this to be performed. A third, modern approach is to derive a dietary pattern that may describe the set of dietary factors that are tested [30,59,60,62e64]. This last type is thoroughly presented in Chapter 4. To evaluate the presence of multicollinearity in a set of variables, a first method is to check their bivariate correlations. If a high number of Pearson correlation coefficients (see Section 5.8) are greater than 0.7e0.8, it indicates the presence of multicollinearity [17]. This can also be verified using another statistical measure called variance inflation factor (VIF) [17]. Most statistical software calculates the value of VIF in a multiple linear regression model, and when its value is greater than 10, it is a second indication for the presence of multicollinearity. It is suggested that both approaches (i.e., bivariate correlations and VIF) are used to determine the existence of this phenomenon.

5.10.2.4 Choice of the Best Model A serious challenge in applying multiple linear regression analysis in nutrition research is the choice of the best model that will comprehensively evaluate a certain research hypothesis. This situation occurs when, in studying the association between a dietary factor and a health outcome, we cannot be sure about the number and type of confounding factors. Most the health outcomes are multifactorial, and usually the research team measures various potential confounders (i.e., participant characteristics) during the study’s sample recruitment. However, the choice of appropriate variables to be inserted into the final multiple regression model is always in question. For instance, suppose that the association between the consumption of citrus fruits and systolic blood pressure is evaluated in a sample of an adult population. Then, a number of various characteristics such as age, gender, socioeconomic status, physical activity level, BMI, genetic variation, quality of diet, etc., can be considered potential confounding factors. In this case, the researcher is responsible for selecting the variables to be inserted into the final model. This can be performed based on various approaches described in the literature; however, there are several considerations. First, the definition of a confounder requires that it to be associated with both the dietary factor and the health outcome (see Chapter 1). Before the insertion of a potential confounder into the model, it is highly recommended that these two associations have been evaluated, first. Discussion can be raised regarding whether the degree of association may be a factor to be acknowledged in the final decisions. Another important consideration that sometimes leads the choice of confounding factors in a multiple regression model is the literature data available on the topic. Previous robust research evidence supporting the role of a characteristic as a confounding factor in a studied association can in itself support the need for related adjustments. In the previous example, age and gender were already considered to be major confounders in the study of diet and systolic blood pressure, so in this case, they should be prioritized for the final selection. Furthermore, most statistical software includes modules allowing the choice of predictive variables automatically. They frequently appear under the name stepwise regression [17,55e57]. In each step, a variable is considered for addition to or subtraction from the multiple regression model based on some

5.10 LINEAR REGRESSION ANALYSIS

133

prespecified criterion. The criterion used is often the P value of the bb coefficient of each variable. Stepwise regression has two major types: forward and backward selection. In forward selection, the software starts with no variable in the model and considers whether adding the potential confounders to the model, one by one, fulfils the criterion that has been set and improves the model. Usually, a P value between 0.05 and 0.10 is considered the criterion for inserting a variable into the model. The backward method does the opposite, starting from the full model that includes all potential confounders and deleting them one by one to provide the best final model. In this case, a P value of the bb coefficient of each variable from 0.05 to 0.10 is considered the criterion for deleting or keeping a variable. In the decision regarding the best linear multiple regression model, some useful statistical measures are also available. Among them, the most popular are R2 (see Section 5.10.1.4) and the Akaike information criterion (AIC) [65]. The R2 can provide us with information about the proportion of the variance of the dependent variable that can be explained by independent variables. Thus, the model with a greater R2 may be promoted in our final decisions. However, comparisons between models with different numbers of independent variables should be avoided because an increase in the number of predictors mathematically increases the R2. It has been suggested that this issue may be overcome by an alternative form of the R2: the adjustedR2. However, its use should also be performed with caution. At this point, it should be pointed out that most of the time, the investigation of a research hypothesis in nutrition science requires a study of health outcomes that are multifactorial. Factors that explain a considerable proportion of the outcome’s variability remain unknown and unmeasured. For this reason, it is often observed that even a multiple linear regression model with a lot of predictors cannot establish a relatively high R2, i.e., greater than 0.3e0.6. In other words, a considerable percentage of the health outcome’s variance remains unexplained. All related conclusions from the use and reporting of the R2 should be performed with this in mind. The AIC is a statistical measure that has been designed with main aim of quantifying the relative quality of statistical models for a given set of data in terms of goodness-of-fit [65]. For a certain dependent variable, lower values of the AIC propose a preferred multiple linear regression model, which is the one with the fewest parameters that still provides an adequate fit to the data. Most statistical software calculates this criterion, which is a ranking score with no other special meaning on its own. Altogether, the selection of the best multiple linear regression model is a challenging task requiring a good background in biostatistics and epidemiology, excellent knowledge of the literature on the topic of the tested research hypothesis, and experience in similar work. The researcher should consider all possible methodological approaches and criteria, whereas the method for the final choice of confounders (i.e., adjustment scheme) should be always properly reported and justified.

5.10.2.5 Assessment of Interactions Linear regression analysis is a useful tool for investigating the degree of the association between a health outcome and a dietary factor, which can be adjusted for the effects of potential confounding factors. However, it can be hypothesized that the degree or direction of the association between these two variables may vary significantly among different sample groups, which are defined by a third factor [5,66,67]. For instance, in Example 5.10, the negative association between leafy vegetable consumption and fasting blood glucose levels may have been different among sample groups defined

134

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

by the presence of a gene mutation related to the pathophysiology of glucose metabolism. The effect of the gene mutation in the studied association is called an interaction effect or effect modification. Thus, the multivariate analysis of an association between a health outcome and a dietary factor should evaluate all potential interaction effects by other studied sample characteristics. Most software allows an assessment of the significance, direction, and degree of the interaction effects in the multiple linear regression analysis modeling approach. When a sample characteristic introduces a significant interaction effect, the results of linear regression analysis should be reported separately for the sample groups defined by this characteristic. For instance, in the previous example, linear regression analysis should be performed stratified by the gene mutation, and the corresponding results (i.e., b coefficients, P values, CI, R2) should be reported separately for the sample groups defined by the gene mutation. Often, factors that introduce interaction effects are gender, age group, social status level, genetic profile, etc. However, an assessment of effects modifications and an interpretation of the results are challenging tasks that require excellent knowledge of the literature on the topic, a good background in biostatistics and epidemiology, and experience in similar studies.

5.10.2.6 Application to Dietary Data and Examples Multiple linear regression analysis can be used to model the relationship between a random continuous dependent variable and a number of independent variables (i.e., predictors), simultaneously. This statistical technique is commonly used to investigate more complex real-life questions in nutrition research. It can provide statistical evidence about how well a set of variables is able to predict an outcome. More important, it can address whether the association of a dietary exposure with a health outcome pointed out by simple linear regression is still significant after controlling (i.e., adjusting) for the effects of other variables (i.e., participant characteristics, confounders) on the same outcome. The estimated model is also able to forecast an unobserved value of Y for a given set of values of Xj. Example 5.11. In the study setting of Example 5.10, the researchers are interested in evaluating the association of fasting blood glucose levels after considering the confounding effects of age and BMI. In this case, multiple linear regression analysis can be used, considering the age and BMI to be the set of the confounding factors (adjustments). The results of the analysis are given in the following table.

Age (years) BMI (kg/m2) Leafy vegetables Constant (be0 )

bb

P Value

2.11 1.13 0.18 70.1

.001 .024 .004 .002

95% Confidence Interval 1.88 0.90 0.40 60.3

Note: P values were derived from the corresponding t tests.

The estimated model is: Yei ¼ 70:1 þ 2:11Xi1 þ 1:13Xi2  0:18Xi3

2.91 1.36 0.09 110.5

R2 0.08

5.11 LOGISTIC REGRESSION ANALYSIS

135

where Xi1 is age (years); Xi2 is BMI (kg/m2); Xi3is leafy vegetable consumption (g/day) evaluated in the sample population i ¼ 1,2,3,.,1000 (i.e., n ¼ 1000); and Yei are the corresponding predicted values of fasting blood glucose levels (mg/dL). Before interpreting the results, it is essential to examine whether all assumptions about applying simple linear regression analysis are fulfilled. For simplicity, let us assume that they are fully covered. The analysis showed that leafy vegetable consumption remains significant negatively associated with the fasting blood glucose levels (b coefficient ¼ 0.18 < 0; P ¼ .004 < 0.05) after considering the age and the BMI of participants. More precisely, a 1-g/day increase in leafy vegetable consumption was associated with a 0.18-mg/dL decrease in fasting blood glucose. This estimation may vary in the reference population within (0.09, 0.40) mg/dL with a 95% level of confidence (i.e., absolute value of the 95% CI for the b coefficient). The corresponding decrease for a 10-g/day increase in leafy vegetable consumption is 1.8 mg/dL (i.e., bcoefficient for 10 units increase ¼ 10  (0.18) ¼ 1.8) and the related 95% CI is (0.09  10 ¼ 9.0, 0.40  10 ¼ 4.0) mg/dL. As expected, the insertion of additional factors to the simple regression model of the example increased the R2. In fact, age, BMI, and leafy vegetable consumption together account for 8% (R2 ¼ 0.08) of fasting blood glucose level variability. Interpretation of the constant be0 does not have actual meaning because conditions of age and BMI equal to zero do not really exist. To conclude the example, let us assume that an assessment of interaction effects for age group or BMI categories in the studied association did not indicate significant results.

5.11 LOGISTIC REGRESSION ANALYSIS 5.11.1 ODDS RATIO The OR is a statistical measure applied mainly in retrospective data analysis, when the health outcome under investigation is not a random continuous variable as in the case of linear regression analysis, but it has a binary form [5,66,67]. This always arises in caseecontrol studies in which data for the two groups of diseased and disease-free participants are studied. The OR is also used in cross-sectional settings when the health outcome or parameter under investigation is dichotomous. Suppose that in a caseecontrol study, diseased and disease-free participants can be categorized according to their past exposure to a dietary factor in groups of exposed and nonexposed participants. Then, the sample population is presented in the following 2  2 table.

Exposed Nonexposed

Diseased (Cases)

Disease-Free (Controls)

a c

b d

where a,b,c and d are the corresponding number of participants in each group. In this case, the OR is defined as follows: Odds ratio ¼

odds of the disease in the exposed group ad ¼ odds of the disease in the non  exposed group b  c

The OR expresses the odds that an outcome occurred given the particular past exposure to a dietary factor, compared with the odds that the outcome occurred in the absence of this exposure.

136

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

Interpretation of the OR requires its comparison with one that is the null value in this case; the following situations may occur: • • •

OR ¼ 1; exposure to the dietary factor does not affect the odds of the outcome. OR > 1; exposure to the dietary factor is associated with higher odds of disease. OR < 1; exposure to the dietary factor is associated with lower odds of disease or health outcome/ parameter.

Example 5.12. In a caseecontrol study of premenopausal breast cancer, the consumption of citrus fruits in diseased and disease-free population was evaluated. The data are shown in the following table.

Citrus fruit consumers Nonconsumers of citrus fruits

Participants With Premenopausal Breast Cancer

Disease-Free Participants

160 220

200 150

The OR can be calculated as: OR ¼

odds of the disease in the exposed group a  d 160  150 24; 000 ¼ ¼ ¼ ¼ 0:8 odds of the disease in the non  exposed group b  c 200  220 30; 000

Thus, the odds of premenopausal breast cancer in the citrus fruit consumers is 0.8 times (or (10.8) 100% ¼ 20%) lower than in nonconsumers. It seems that the consumption of citrus fruit may be associated with a lower presence of the disease. The concept of the CI presented earlier can also be applied to the OR as a point estimator derived from sample data. Most statistical software calculates the CI for various confidence levels; the 95% CI is most frequently reported. The CI is also used as a proxy of statistical significance of the OR. This can be performed by checking whether the null value (i.e., OR ¼ 1) is included in the CI. If the one does not fall within the CI, there is evidence that the OR is statistically significant. The opposite occurs when the CI overlaps with the null value.

5.11.2 SIMPLE BINARY LOGISTIC REGRESSION ANALYSIS Binary logistic regression analysis is a statistical method that can be applied mainly in retrospective data to explore and model the relationship between a random dichotomous variable and one or more random independent variables (continuous or categorical) [68e70]. Simple binary logistic regression analysis is the basic form of this analysis format in which there is only one independent variable [71]. Similar to the linear regression, analysis is based on adjusting a linear function to the sample data. Here, the dependent variable Y as it was measured in the sample population is dichotomous (for instance, the presence of hypertension is 0 and 1) and does have not a continuous form. For this reason, the proposed formula for the simple binary logistic regression model is: pðxÞ ¼

eaþbx 1 þ eaþbx

5.11 LOGISTIC REGRESSION ANALYSIS

137

where p(x) is the probability that the health outcome occurred (i.e., Y has the value of 1) for a given value of independent variable x (i.e., dietary exposure). It can be proved that:   pðxÞ gðxÞ ¼ ln ¼ a þ bx 1  pðxÞ where g(x) is called the logit transformation. It is popular in biostatistics because it is conceptually connected with the OR. For example, when the dietary exposure is dichotomous, too (i.e., has values of 0 and 1), it is proved that: gð1Þ  gð0Þ ¼ OR The basic advantage of applying simple binary logistic regression analysis is that b coefficients are easily expressed using the form of the ORs [72]. In fact, on the previous occasion (where the dietary exposure was dichotomous), the odds that the health outcome occurred in the exposed group compared to the odds that the health outcome occurred in the unexposed group, in other words the OR, were equal to eb. Statistical software provides the values of the b coefficients by applying the maximum likelihood estimation approach [68e70,72]. This statistical methodology has an advanced level of complexity and its presentation is not relevant to the aims of this chapter. The software also calculates the related ORs along with the corresponding CIs. The hypothesis testing e H0 : e b¼0 e e HA : bs0 is also evaluated by the Wald statistic and the related pvalue is calculated. When pvalue < a (i.e., the statistical significance level), there is statistical evidence to reject H0 and support the concept e e that bs0. In other words, there is statistical evidence that the OR is statistically significant. When the x dietary exposure is a continuous variable, the eb expresses the OR for a one-unit increase in x showing how much greater (when OR > 1) or lower (when OR < 1) the odds are of the health outcome occurring when x increases one unit. Again, the statistical software provides the desired CIs and performs similar hypothesis testing for the significance of the b coefficients.

5.11.3 MULTIPLE BINARY LOGISTIC REGRESSION ANALYSIS Multiple binary logistic regression analysis is the extended version of the simple form of analysis [68e70,72], where the dependent variable Y is tested against a set of two or more independent variables, xj, j ¼ 1,2,.,k. Multiple binary logistic regression is given by: pðfxgÞ ¼ where {x} ¼ (x1, x2,.,xk).

eaþb1 x1 þb2 x2 þ.þbk xk 1 þ eaþb1 x1 þb2 x2 þ.þbk xk

138

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

The estimation and interpretation of b coefficients are similar to the simple form of analysis except that the estimated OR for a certain independent variable has been adjusted for the other x factors. The advantage of multiple binary logistic regression analysis is that it can address whether the association between a dietary exposure and the odds of the presence of a health outcome that it observed in the simple form of analysis remains significant after controlling for the effects of other variables (i.e., participants characteristics, confounding effects) [73e78]. The phenomenon of multicollinearity presented in the linear regression analysis should also be tested in this analysis format using similar approaches. The presence of a strong correlation between explanatory variables may result in biased estimations. In addition, the HosmereLemeshow criterion can be used to assess the goodness-of-fit for the binary logistic regression model [17,55e57]. This statistical measure contrasts the estimated with the observed occurrence of the health outcome given the sample values and the logistic regression model. The statistical software generates the value of the HosmereLemeshow criterion and a related P value. When the value of the criterion is small and the P value falls within the H0 rejection area, there is evidence to support a poor fit of the model to the sample data. Similar to linear regression analysis, the choice of confounders to be included in the final multiple logistic regression model is challenging. The researcher should consider all possible methodological approaches and criteria whereas the method for the final selection of confounders (i.e., adjustment scheme) should always be properly reported and justified. As in the case of linear regression analysis, when a sample characteristic introduces a significant interaction, the results of logistic regression analysis should be reported separately for the sample groups defined by this characteristic (see Section 5.10.2.5).

5.11.4 EXAMPLES Example 5.13. In a caseecontrol study of premenopausal breast cancer (450 cases and 490 controls), the consumption of red meat was assessed in both diseased and disease-free participants. (a) The researchers are interested in evaluating the association of dietary exposure with the studied health outcome. Because the health outcome is a dichotomous variable (i.e., the presence or absence of disease), simple binary logistic regression analysis can be used. The results of the analysis are given in the following table.

Red meat consumption (1: yes/0: no)

b b

P Value

0.47

.003

95% Confidence Interval 0.37

0.56

Note: The P value is derived from the corresponding Wald test. For simplicity, the estimated constant of the model is not reported. Also, suppose that the model has a good fit to the sample data according to the HosmereLemeshow criterion.

The dietary exposure is dichotomous, so the OR can be calculated as: OR ¼ eb ¼ e0:47 z 1:60

5.11 LOGISTIC REGRESSION ANALYSIS

139

The correspondent limits for the 95% CI of the OR are: Lower limit ¼ e0:37 z 1:45 Upper limit ¼ e0:56 z 1:75 Thus, the results indicate that past exposure to the dietary factor, i.e., red meat consumption, is significantly associated with higher odds of premenopausal breast cancer (OR ¼ 1.60 > 1; P value ¼ .003 < 0.05). In more detail, the odds of premenopausal breast cancer in red meat consumers is 1.60 times (or (1.601)100% ¼ 60%) greater than in nonconsumers. This estimation may vary in the reference population within the (1.45, 1.75) with a 95% level of confidence (i.e., the 95% CI for the OR). (b) Now, suppose that the assessment of red meat has been performed on a continuous scale, i.e., g/day. Results from the binary logistic regression are given in the following table.

Red meat consumption g/day

b b

P Value

0.01

.002

95% Confidence Interval 0.005

0.02

Note: The P value is derived from the corresponding Wald test. For simplicity, the estimated constant of the model is not reported. Also, suppose that the model has a good fit to the sample data according to the HosmereLemeshow criterion.

The dietary exposure is a continuous variable, so the OR for A ONE-unit (i.e., 1-g/day) increase in red meat consumption can be calculated as: OR ¼ eb ¼ e0:01 z 1:01 The corresponding limits for the 95% CI of the OR are: Lower limit ¼ e0:005 z 1:005 Upper limit ¼ e0:02 z 1:02 Again, the results indicate that past exposure to the dietary factor, i.e., red meat consumption was significantly associated with higher odds of premenopausal breast cancer (OR ¼ 1.01 > 1; P value ¼ .002 < 0.05). In more detail, a 1-g/day increase in red meat consumption was associated with 1.01 times greater odds of the disease (or (1.011)100% ¼ 1%). This estimation may vary in the reference population within the (1.005, 1.02) with a 95% level of confidence (i.e., the 95% CI for the OR). However, this interpretation has no actual meaning for clinical nutritionists because a 1-g/day increase in red meat consumption is hard to measure. For this reason, the OR may be expressed in a more understandable increment in red meat consumption: for instance, 50 g/day. In this case, the OR and 95% CI are calculated as: OR ¼ eb50 ¼ e0:0150 z 1:65 The corresponding limits for the 95% CI of the OR are: Lower limit ¼ e0:00550 z 1:28 Upper limit ¼ e0:0250 z 2:72

140

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

Thus, a 50-g/day increase in red meat consumption was associated with 1.65 times greater odds of premenopausal breast cancer (or (1.651)100% ¼ 65%). This estimation may vary in the reference population within the (1.28, 2.72) with a 95% level of confidence (i.e., the 95% CI for the OR). (c) Suppose further that the researchers are interested in evaluating the association presented in (a) after controlling for the confounding effects of age and BMI. In this case, the multiple form of analysis was used and the results are shown in the following table.

Age (years) Body mass index (kg/m2) Red meat consumption (1: yes/0: no)

b b

P Value

0.07 0.10 0.41

.001 .02 .01

95% Confidence Interval 0.06 0.08 0.33

0.08 0.12 0.58

Note: P values are derived from the corresponding Wald tests. For simplicity, the estimated constant of the model is not reported. Also, suppose that the model has a good fit to the sample data according to the HosmereLemeshow criterion and no multicollinearity is present.

The results indicate that red meat consumption remains significantly associated with higher odds of premenopausal breast cancer after adjusting age and BMI (OR ¼ e0.41z1.51 > 1; P value ¼ .01 < 0.05). In fact, the odds of premenopausal breast cancer in red meat consumers is 1.51 times (or (1.511)100% ¼ 51%) greater than in nonconsumers. This estimation may vary in the reference population within (e0.33 z 1.40, e0.58 z 1.79) with a 95% level of confidence (i.e., the 95% CI for the OR). In the same model, it was observed that both age (OR for 5 years increase¼e0.075 z 1.42 > 1; P ¼ .001 < 0.05) and BMI (OR for 1 kg/m2 increase ¼ e0.10 z 1.11 > 1; P ¼ .02 < 0.05) were also significantly associated with higher odds of the disease after adjusting for two other variables in each case. The corresponding 95% CIs for these estimations were (e0.065 z 1.35, e0.085 z 1.49) and (e0.08 z 1.08, e0.12 z 1.13).

REFERENCES [1] Dawson GF. Chapter 10dhypothesis testing. In: Easy interpretation of biostatistics; 2008. p. 87e96. https:// doi.org/10.1016/B978-1-4160-3142-0.50018-3. [2] Hoffman JIE. Chapter 10dhypothesis testing: the null hypothesis, significance, and type I error. In: Biostatistics for medical and biomedical practitioners; 2015. p. 153e8. https://doi.org/10.1016/B978-0-12802387-7.00010-X. [3] Riffenburgh RH. Chapter 8dhypothesis testing: concept and practice. In: Statistics in medicine; 2012. p. 157e74. https://doi.org/10.1016/B978-0-12-384864-2.00008-1. [4] Lovegrove J, Sharma S, Hodson L, Lanham-New S. Nutrition research methodologies. Wiley-Blackwell; 2014. [5] Willett W. Nutritional epidemiology. Oxford University Press; 2012. https://doi.org/10.1093/acprof:oso/ 9780199754038.001.0001. [6] Hoffman JIE. Chapter 11dhypothesis testing: sample size, effect size, power, and type II errors. In: Biostatistics for medical and biomedical practitioners; 2015. p. 159e71. https://doi.org/10.1016/B978-0-12802387-7.00011-1.

REFERENCES

141

[7] Woodward M. Epidemiology: study design and data analysis. Chapman & Hall texts in Statistical Science Series; 1999. [8] Dawson GF. Appendix Adflowchart of types of statistical tests. In: Easy interpretation of biostatistics; 2008. p. 175e8. https://doi.org/10.1016/B978-1-4160-3142-0.50033-X. [9] Lachat C, Hawwash D, Ocke´ MC, Berg C, Forsum E, Ho¨rnell A, Larsson CL, Sonestedt E, Wirfa¨lt E, ˚ kesson A, Kolsteren P, Byrnes G, De Keyzer W, Van Camp J, Cade JE, Slimani N, Cevallos M, Egger M, A Huybrechts I. Strengthening the reporting of observational studies in epidemiologydnutritional epidemiology (STROBE-nut): an extension of the STROBE statement. Nutr Bull 2016. https://doi.org/10.1111/nbu. 12217. [10] Dawson GF. Chapter 8ddistributions. In: Easy interpretation of biostatistics; 2008. p. 63e74. https://doi. org/10.1016/B978-1-4160-3142-0.50014-6. [11] Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 5dprobability distributions. In: Biostatistics; 2007. p. 103e33. https://doi.org/10.1016/B978-0-12-369492-8.50010-8. [12] Riffenburgh RH. Chapter 4ddistributions. In: Statistics in medicine; 2012. p. 77e94. https://doi.org/10. 1016/B978-0-12-384864-2.00004-4. [13] Carlson KA, Winquist JR. In: An introduction to statistics : an active learning approach. 2nd ed. SAGE; 2017. [14] Dawson GF. Chapter 9dthe normal distribution. In: Easy interpretation of biostatistics; 2008. p. 75e83. https://doi.org/10.1016/B978-1-4160-3142-0.50015-8. [15] Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 3ddescriptive methods. In: Biostatistics; 2007. p. 21e69. https://doi.org/10.1016/B978-0-12-369492-8.50008-X. [16] Wassertheil-Smoller S, Smoller J. Biostatistics and epidemiology. New York, NY: Springer; 2015. https:// doi.org/10.1007/978-1-4939-2134-8. [17] Pallant J. In: SPSS survival manual. 5th ed. Allen & Unwin; 2013. [18] Hoffman JIE. Chapter 6dnormal distribution. In: Biostatistics for medical and biomedical practitioners; 2015. p. 101e19. https://doi.org/10.1016/B978-0-12-802387-7.00006-8. [19] Lyon A. Why are normal distributions normal? Br J Philos Sci 2014;65:621e49. https://doi.org/10.1093/ bjps/axs046. [20] Tsokos C, Wooten R, Tsokos C, Wooten R. Normal probability. In: The joy of finite mathematics. Elsevier; 2016. p. 231e63. https://doi.org/10.1016/B978-0-12-802967-1.00007-3. [21] Rahman MM, Govindarajulu Z. A modification of the test of Shapiro and Wilk for normality. J Appl Stat 1997;24:219e36. https://doi.org/10.1080/02664769723828. [22] Royston P. Approximating the Shapiro-Wilk W-test for non-normality. Stat Comput 1992;2:117e9. https:// doi.org/10.1007/BF01891203. [23] Shapiro SS, Wilk MB. An analysis of variance test for normality (complete samples). Biometrika 1965;52: 591e611. https://doi.org/10.1093/biomet/52.3-4.591. [24] Smirnov N. Table for estimating the goodness of fit of empirical distributions. Ann Math Stat 1948;19: 279e81. https://doi.org/10.1214/aoms/1177730256. [25] Stephens MA. EDF statistics for goodness of fit and some comparisons. J Am Stat Assoc 1974;69:730. https://doi.org/10.2307/2286009. [26] Wilk MB, Gnanadesikan R. Probability plotting methods for the analysis of data. Biometrika 1968;55:1e17. https://doi.org/10.1093/biomet/55.1.1. [27] Mushkudiani N, Einmahl J. Generalized probabilityeprobability plots. J Stat Plan Inference 2007;137: 738e52. https://doi.org/10.1016/J.JSPI.2006.06.006. [28] Pounis G, Bonanni A, Ruggiero E, Di Castelnuovo A, Costanzo S, Persichillo M, Bonaccio M, Cerletti C, Riccardi G, Donati MB, de Gaetano G, Iacoviello L. Food group consumption in an Italian population using

142

[29]

[30]

[31]

[32] [33] [34] [35]

[36]

[37] [38] [39]

[40]

[41] [42] [43] [44] [45] [46] [47]

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

the updated food classification system FoodEx2: results from the Italian Nutrition & HEalth Survey (INHES) study. Nutr Metab Cardiovasc Dis 2017;27. https://doi.org/10.1016/j.numecd.2017.01.004. Pounis G, Costanzo S, Persichillo M, de Curtis A, Sieri S, Vinceti M, Zito F, Di Castelnuovo AF, Donati MB, de Gaetano G, Iacoviello L. Mushroom and dietary selenium intakes in relation to fasting glucose levels in a free-living Italian adult population: the Moli-sani Project. Diabetes Metab 2014a;40. https://doi.org/10.1016/ j.diabet.2013.09.007. Pounis G, de Lorgeril M, Salen P, Laporte F, Krogh V, Siani A, Arnout J, Cappuccio FP, van Dongen M, Donati MB, de Gaetano G, Iacoviello L. Dietary patterns and fatty acids levels of three European populations. Results from the IMMIDIET study. Nutr Metab Cardiovasc Dis 2014b;24. https://doi.org/10.1016/j. numecd.2014.01.012. Pounis G, Di Castelnuovo AF, de Lorgeril M, Krogh V, Siani A, Arnout J, Cappuccio FP, van Dongen M, Zappacosta B, Donati MB, de Gaetano G, Iacoviello L, Arnout J, Buntinx F, Dagnelie PC, Lorgeril M, Sian A, Dirckxc C, Castelnuovo AD, Dongen M, Bonanni A, Rink P, Vohnout B, Zito F. Folate intake and folate serum levels in men and women from two European populations: the IMMIDIET project. Nutrition 2014c;30. https://doi.org/10.1016/j.nut.2013.11.014. Riffenburgh RH. Chapter 7dconfidence intervals. In: Statistics in medicine; 2012. p. 137e55. https://doi. org/10.1016/B978-0-12-384864-2.00007-X. Riffenburgh RH. Tables of probability distributions. In: Statistics in medicine; 2012. p. 639e56. https://doi. org/10.1016/B978-0-12-384864-2.00038-X. Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 10danalysis of categorical data. In: Biostatistics; 2007. p. 269e96. https://doi.org/10.1016/B978-0-12-369492-8.50015-7. Hoffman JIE. Chapter 14dcategorical and cross-classified data: goodness of fit and association. In: Biostatistics for medical and biomedical practitioners; 2015. 183ee219. https://doi.org/10.1016/B978-012-802387-7.00014-7. Pearson K. On the criterion that a given system of deviations from the probable in the case of a correlated system of variables is such that it can be reasonably supposed to have arisen from random sampling. Philos Mag Ser 1900;5(50):157e75. https://doi.org/10.1080/14786440009463897. Plackett RL. Karl Pearson and the chi-squared test. Int Stat Rev/Rev Int Stat 1983;51:59. https://doi.org/10. 2307/1402731. Riffenburgh RH. Chapter 9dtests on categorical data. In: Statistics in medicine; 2012. p. 175e202. https:// doi.org/10.1016/B978-0-12-384864-2.00009-3. Hoffman JIE. Chapter 22dcomparison of two groups: t-tests and nonparametric tests. In: Biostatistics for medical and biomedical practitioners; 2015. p. 337e62. https://doi.org/10.1016/B978-0-12-802387-7. 00022-6. Lalanne C, Mesbah M, Lalanne C, Mesbah M. 2dmeasures of association, comparisons of means and proportions for two samples or more. In: Biostatistics and computer-based analysis of health data using Stata; 2016. p. 25e57. https://doi.org/10.1016/B978-1-78548-142-0.50002-X. Student. The probable error of a mean. Biometrika 1908;6:1e25. https://doi.org/10.1093/biomet/6.1.1. Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 12danalysis of variance. In: Biostatistics; 2007. p. 323e48. https://doi.org/10.1016/B978-0-12-369492-8.50017-0. Hoffman JIE. Chapter 25danalysis of variance I. One-way. In: Biostatistics for medical and biomedical practitioners; 2015. p. 391e420. https://doi.org/10.1016/B978-0-12-802387-7.00025-1. Pearson K. Notes on regression and inheritance in the case of two parents. Proc R Soc London 1895;58: 240e2. https://doi.org/10.1098/rspl.1895.0041. Suchmacher M, Geller M, Suchmacher M, Geller M. Chapter 14dcorrelation and regression. In: Practical biostatistics; 2012. p. 167e86. https://doi.org/10.1016/B978-0-12-415794-1.00014-8. Bagdonavicius V, Julius K, Nikulin MS. Non-parametric tests for complete data. ISTE/Wiley; 2011. Corder GW, Foreman DI. Nonparametric statistics: a step-by-step approach. Wiley; 2014.

REFERENCES

143

[48] Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 9dnonparametric tests. In: Biostatistics; 2007. p. 249e68. https://doi.org/10.1016/B978-0-12-369492-8.50014-5. [49] Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 13dlinear regression. In: Biostatistics; 2007. p. 349e86. https://doi.org/10.1016/B978-0-12-369492-8.50018-2. [50] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 3dlinear regression. In: Biostatistics and computer-based analysis of health data using Stata; 2016. p. 59e77. https://doi.org/10.1016/B978-1-78548-142-0.50003-1. [51] Riffenburgh RH. Chapter 21dregression and correlation. In: Statistics in medicine; 2012. p. 443e72. https://doi.org/10.1016/B978-0-12-384864-2.00021-4. [52] Vittinghoff E, McCulloch CE, Glidden DV, Shiboski SC. 3dlinear and non-linear regression methods in epidemiology and biostatistics. In: Essential statistical methods for medical statistics; 2011. p. 66e103. https://doi.org/10.1016/B978-0-444-53737-9.50006-2. [53] Hoffman JIE. Chapter 27dlinear regression. In: Biostatistics for medical and biomedical practitioners; 2015. p. 451e500. https://doi.org/10.1016/B978-0-12-802387-7.00027-5. [54] Sullivan LM, Lisa M. Essentials of biostatistics in public health. Jones & Bartlett Learning; 2012. [55] Lalanne C, Mesbah M. Biostatistics and computer-based analysis of health data using Stata. 2017. [56] Lalanne C, Mesbah M. Biostatistics and computer-based analysis of health data using SAS. Elsevier Science; 2017b. [57] Lalanne C, Mesbah M. Biostatistics and computer-based analysis of health data using R. ISTE Press Ltd; 2016. [58] Chatterjee S, Hadi AS, Price B. In: Regression analysis by example. 3rd ed. Wiley; 2000. [59] Pounis G, Costanzo S, Di Giuseppe R, De Lucia F, Santimone I, Sciarretta A, Barisciano P, Persichillo M, De Curtis A, Zito F, Di Castelnuovo AF, Sieri S, Benedetta Donati M, De Gaetano G, Iacoviello L. Consumption of healthy foods at different content of antioxidant vitamins and phytochemicals and metabolic risk factors for cardiovascular disease in men and women of the Moli-sani study. Eur J Clin Nutr 2013;67. https://doi.org/ 10.1038/ejcn.2012.201. [60] Pounis G, Di Castelnuovo A, Bonaccio M, Costanzo S, Persichillo M, Krogh V, Donati MB, De Gaetano G, Iacoviello L. Flavonoid and lignan intake in a Mediterranean population: proposal for a holistic approach in polyphenol dietary analysis, the Moli-sani Study. Eur J Clin Nutr 2016b;70. https://doi.org/10.1038/ejcn. 2015.178. [61] Pounis G, Bonaccio M, Di Castelnuovo A, Costanzo S, De Curtis A, Persichillo M, Sieri S, Donati MB, Cerletti C, De Gaetano G, Iacoviello L. Polyphenol intake is associated with low-grade inflammation, using a novel data analysis from the Moli-sani study. Thromb Haemostasis 2016a;115. https://doi.org/10.1160/ TH15-06-0487. [62] Agnoli C, Krogh V, Grioni S, Sieri S, Palli D, Masala G, Sacerdote C, Vineis P, Tumino R, Frasca G, Pala V, Berrino F, Chiodini P, Mattiello A, Panico S. A priori-defined dietary patterns are associated with reduced risk of stroke in a large Italian cohort. J Nutr 2011;141:1552e8. https://doi.org/10.3945/jn.111.140061. [63] Hoffmann K, Schulze MB, Schienkiewitz A, No¨thlings U, Boeing H. Application of a new statistical method to derive dietary patterns in nutritional epidemiology. Am J Epidemiol 2004;159:935e44. [64] Trichopoulou A, Costacou T, Bamia C, Trichopoulos D. Adherence to a mediterranean diet and survival in a Greek population. N Engl J Med 2003;348:2599e608. https://doi.org/10.1056/NEJMoa025039. [65] Akaike H. Information theory as an extension of the maximum likelihood principle. In: Petrov BN, Csaki F, editors. Second international symposium on information theory. Budapest: Akademiai Kiado; 1973. p. 267e81. [66] Gerstman BB. Epidemiology kept simple: an introduction to traditional and modern epidemiology. John Wiley & Sons; 2013. [67] Roht LH, Selwyn BJ, Holguin AH. Principles of epidemiology: a self-teaching guide. Academic Press; 1982.

144

CHAPTER 5 STATISTICAL ANALYSIS OF RETROSPECTIVE HEALTH

[68] Hoffman JIE. Chapter 33dlogistic regression. In: Biostatistics for medical and biomedical practitioners; 2015. p. 601e11. https://doi.org/10.1016/B978-0-12-802387-7.00033-0. [69] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 4dlogistic regression and epidemiological analyses. In: Biostatistics and computer-based analysis of health data using Stata; 2016. p. 79e99. https://doi.org/10. 1016/B978-1-78548-142-0.50004-3. [70] Spitznagel EL. 6 logistic regression. Handb Stat 2007;27:187e209. https://doi.org/10.1016/S0169-7161(07) 27006-3. [71] Cox D. The regression analysis of binary sequences (with discussion). J Roy Stat Soc 1958;B:215e42. [72] Suarez Perez EL, Perez Cardona C, Rivera R, Martinez MN. Applications of regression models in epidemiology. 2017. [73] Polychronopoulos E, Pounis G, Bountziouka V, Zeimbekis A, Tsiligianni I, Qira B-E, Gotsis E, Metallinos G, Lionis C, Panagiotakos D. Dietary meat fats and burden of cardiovascular disease risk factors, in the elderly: a report from the MEDIS study. Lipids Health Dis 2010;9. https://doi.org/10.1186/1476-511X9-30. [74] Pounis GD, Makri S, Gougias L, Makris H, Papakonstantinou M, Panagiotakos DB, Kapsokefalou M. Consumer perception and use of iron fortified foods is associated with their knowledge and understanding of nutritional issues. Food Qual Prefer 2011;22. https://doi.org/10.1016/j.foodqual.2011.05.004. [75] Pounis GD, Panagiotakos DB, Chrysohoou C, Aggelopoulos P, Tsiamis E, Pitsavos C, Stefanadis C. Longterm fish consumption is associated with lower risk of 30-day cardiovascular disease events in survivors from an acute coronary syndrome. Int J Cardiol 2009;136. https://doi.org/10.1016/j.ijcard.2008.04.063. [76] Pounis GD, Tyrovolas S, Antonopoulou M, Zeimbekis A, Anastasiou F, Bountztiouka V, Metallinos G, Gotsis E, Lioliou E, Polychronopoulos E, Lionis C, Panagiotakos DB. Long-term animal-protein consumption is associated with an increased prevalence of diabetes among the elderly: the Mediterranean islands (MEDIS) study. Diabetes Metab 2010;36. https://doi.org/10.1016/j.diabet.2010.06.007. [77] Tyrovolas S, Psaltopoulou T, Pounis G, Papairakleous N, Bountziouka V, Zeimbekis A, Gotsis E, Antonopoulou M, Metallinos G, Polychronopoulos E, Lionis C, Panagiotakos DB. Nutrient intake in relation to central and overall obesity status among elderly people living in the Mediterranean islands: the MEDIS study. Nutr Metab Cardiovasc Dis 2011;21. https://doi.org/10.1016/j.numecd.2009.10.012. [78] Yannakoulia M, Tyrovolas S, Pounis G, Zeimbekis A, Anastasiou F, Bountziouka V, Voutsa K, Gotsis E, Metallinos G, Lionis C, Polychronopoulos E, Panagiotakos D. Correlates of low dietary energy reporting in free-living elderly: the MEDIS study. Maturitas 2011;69. https://doi.org/10.1016/j.maturitas.2011.01.016.

CHAPTER

STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH AND NUTRITION DATA

6 George Pounis Alimos, Athens, Greece

CHAPTER OUTLINE 6.1 Introduction .................................................................................................................................145 6.2 Descriptive Statistics....................................................................................................................146 6.3 Measures to Calculate the Occurrence of a Health Outcome ........................................................... 146 6.3.1 Incidence ................................................................................................................146 6.3.2 Relative Risk............................................................................................................148 6.4 Survival Analysis..........................................................................................................................149 6.4.1 Basic Concepts ........................................................................................................149 6.4.2 KaplaneMeier Analysis .............................................................................................150 6.4.3 Log-Rank Test..........................................................................................................152 6.4.4 Cox Regression Analysis............................................................................................155 6.4.4.1 Cox Proportional Hazards Model.......................................................................... 156 6.4.4.2 Application and Examples ................................................................................... 157 References ..........................................................................................................................................159

6.1 INTRODUCTION The statistical analysis of data collected through prospective studies is an important and challenging task in nutrition research. It gains additional gravity when the importance of prospective studies is considered in the context of nutrition research (see Chapter 1), especially their ability to provide some initial evidence for the existence of causality in the association between a dietary exposure and a health outcome. A comprehension of statistical techniques benefits researchers and may be characterized as essential. However, as mentioned in Chapter 5, the mathematical background required to study statistical methodologies and their wide variety be an unfamiliar environment for nutrition and medical researchers. This chapter provides an overview of statistical methods that can be applied to prospective data analysis along with a proper interpretation of their results. An effort was made to simplify the related Analysis in Nutrition Research. https://doi.org/10.1016/B978-0-12-814556-2.00006-3 Copyright © 2019 Elsevier Inc. All rights reserved.

145

146

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

presentations with no intention to substitute for handbooks of biostatistics but to make the reader aware of the basics of statistical analysis in nutrition science. The chapter starts with an introduction to descriptive statistics and the measures used to calculate the occurrence of a health outcome in prospective studies in nutrition research. Then, it emphasizes survival analysis by presenting the basics of KaplaneMeier methodology, the log-rank test, and Cox regression modeling.

6.2 DESCRIPTIVE STATISTICS A descriptive statistical analysis of prospective health and nutrition data is the first to be performed and reported. It aims to illustrate the main characteristics of the sample, which are relevant to the scope of the study, and to provide an overview of them. Previous to this, the researcher should check the quality of the collected data and the completeness of the related databases. Asin an analysis of retrospective data (see Chapter 5), descriptive statistics are reported at the beginning of the results section of a scientific report [1,2]. Basic elements include text, tables, and figures; usually data are illustrated stratified by the health outcome under investigation or a confounding factor or effect modifier [3,4]. Baseline measurements, meaning those at the time of the recruitment of participants, have a retrospective character that can be analyzed using all of the descriptive statistics and related methodology mentioned in Chapter 5. Additional measures to calculate the occurrence of a health outcome in prospective studies are introduced in the following sections along with examples.

6.3 MEASURES TO CALCULATE THE OCCURRENCE OF A HEALTH OUTCOME Prospective studies employ specific tools to quantify the occurrence of a health outcome of interest in a sample population. There is a variety of them;in this section, we will emphasize measures that can estimate and quantify how many individuals were newly diagnosed with the event of interest (i.e., disease or other health outcome).

6.3.1 INCIDENCE The incidence of a disease measures how quickly people present with the disease or health condition in the sample population; it mainly shows new patients in a specific time period [5]. Estimating the incidence is crucial to ensure that all participants during study sample recruitment do not present the event of interest. This eliminates the possibility of considering and quantifying old and new patients together. The validity of estimations for the incidence is related to the accuracy of the tools used to diagnose the disease or generally the measurement of the outcome of interest [6]. As a measure, incidence appears in the literature under various names and types, depending on the formula to calculate it [7,8]. A generic definition of the incidence: Incidence ¼

Number of participants who developed a disease or a health condition in specific period of time Total number of the participants

6.3 MEASURES TO CALCULATE THE OCCURRENCE OF A HEALTH OUTCOME

147

Example 6.1. A sample population of 1000 participants was observed for 5 years and the outcome of interest was the development of obesity. During this period, 280 people presented with obesity. Consequently, the incidence of obesity in this sample can be calculated as: 280 ¼ 0:28 or 28%; 1000 This means that 28% of the sample population was newly diagnosed with obesity over 5 years. Incidence ¼

When the denominator of the fraction is the number of participants who do not have the outcome of interest at the time of recruitment (i.e., and so were at risk for developing it),the incidence is named cumulative incidence or incidence proportion and takes the following form: CI ¼

Number of participants who developed a disease or a health condition in specific period of time Number of participants at risk of developing the disease or the health condition at the start of the time period

Example 6.2. In the scenario of Example 6.1, let assume that 200 individuals who were recruited into the study had incomplete or invalid measures for their body weight at the time of recruitment. In this case, we are unsure whether these individuals were obese at the start of the study. As a result, we cannot include them in the calculation of the CI: 280 ¼ 0:35 or 35% 1000  200 Note: None of the 280 new cases of obesity include any of the 200 participants with invalid or missing measures. CI ¼

However, an important parameter in prospective studies is the time. Often in these study settings we cannot be sure whether a participant who was initially free of the outcome of interest developed it during the follow-up period. We may also be unsure about the exact time when this occurred. For this reason, the concept of the person-time is introduced in calculating the incidence. It is defined as the time (i.e., years, months, hours) when a study participant was at risk for being diagnosed with the event of interest [5,6,9]. The incidence rate (IR) is the formation of incidence that incorporates the persontime. It can be defined as: IR ¼

Number of participants who developed a disease or a health condition in specific period of time Total person time

Calculation of total person-time is challenging. First, we need to specify the time when participants who developed the event of interest were at risk (i.e., free of it). Second, we have to consider the time when a participant was part of the study before being lost to follow-up. In this case, we do not know whether this person developed the outcome of interest. Finally, we should include the time when participants participated in the study until their end date without being diagnosed with the outcome. Example 6.3. Let us again use the scenario of Examples 6.1 and 6.2 and now consider person-time. In following table, the total person-years of the participants can be found as they have been categorized to (1) those who developed obesity, (2) those who were lost to follow up, and (3) those who completed the whole study period without developing obesity.

148

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

Person-Time of Participants Measured in Years Group of Participants

Person-Years

Developed obesity during study period (n ¼ 280) Lost to follow-up (n ¼ 220) Completed study with no obesity (300)

560 740 1500 2800

The IR is calculated as: 280 ¼ 0:10 or 2800 This means that 0.10 cases occurred per 1 person-year or 10 cases per 100 person-years. We may also state that if we observe 100 participants for 1 year (i.e., 100 person-years), 10 will develop obesity. IR ¼

When the studied outcome is death caused by a certain disease,various incidence types form measures of mortality. These are extremely helpful in prospective settings because they can be applied to studying dietary factors that are associated with the longevity of sample populations.

6.3.2 RELATIVE RISK Relative risk (RR) is a statistical measure applied in prospective data analysis when the incidence of a disease is assessed in association with a dietary exposure [10,11]. Suppose that in a cohort study, participants who are exposed or not exposed to a dietary factor can be categorized according to the development of a disease during the follow-up period in those who developed the disease and diseasefree participants. Then the sample population is presented in the following 2  2 table, where a; b; c and d are the corresponding number of participants in each group.

Exposed Not exposed

Participants Who Developed the Disease

Disease-Free

A C

b d

In this case, the RR is defined as: a a þ Relative risk ¼ c b cþd The RR expresses the risk that the participant presents with the disease within the follow-up period given the particular exposure to a dietary factor, compared with the risk of developing the disease in the absence of this exposure.

6.4 SURVIVAL ANALYSIS

149

Interpretation of the RR requires its comparison with onethat is the null value on this occasion; the following situations may occur: • • •

The RR ¼ 1; then exposure to the dietary factor does not affect the risk for the disease. The RR > 1; then exposure to the dietary factor is associated with higher risk for the disease The RR < 1; then exposure to the dietary factor is associated with lower risk for the disease

Example 6.4. In a cohort study of healthy young women, the consumption of citrus fruit was evaluated at the start of the follow-up period. Participants were observed for 10 years and the occurrence of premenopausal breast cancer was measured. The data are shown in the following table.

Citrus fruit consumers Nonconsumers of citrus fruit

Participants Who Developed Premenopausal Breast Cancer

Disease-Free Participants

10 20

190 80

The RR can be calculated as: 10 0:05 ¼ 0:25 RR ¼ 10 þ 190 ¼ 20 0:2 20 þ 80 Thus, the risk for premenopausal breast cancer in citrus fruit consumers is 0.25 times (or ð1  0:25Þ100% ¼ 75%) lower than in nonconsumers. It seems that the consumption of citrus fruit may be associated with lower risk for the disease. The concept of the confidence interval (CI) presented in Chapter 5 can also be applied to the RR as a point estimator derived from sample data. Most statistical software calculates the CI for various confidence levels although the 95%CI is the most frequently reported. The CI is also used as a proxy of the statistical significance of the RR. This can be performed by checking whether the null value (i.e., RR ¼ 1) is included in the CI. If it does not fall within the CI, there is evidence to support that the RR is statistically significant. The opposite occurs when the CI overlaps with the null value.

6.4 SURVIVAL ANALYSIS 6.4.1 BASIC CONCEPTS Survival analysis, as a term, describes a branch of statistical methodologies used in the analysis of time-to-event data [8,12e15]. Prospective study designs collect these data, evaluating the length of time from an origin to an endpoint of interest. Various types of “positive” or “negative” events might be of interest in a health and nutrition study, such as myocardial infarction, a diagnosis of diabetes, the development of obesity, the discharge from a hospital, recovery from a disease, or death from various causes. Survival analysis aims to address the proportion of participants who “survive” or do not present with the event of interest until the end of the follow-up time of the study. It may also evaluate whether a certain dietary or other characteristic of participants may affect their “survival.” In this way, conclusions can be drawn about the associations between dietary exposures and health outcomes.

150

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

At this point, it is helpful to introduce some basic concepts related to the time-to-event data that we meet later. The time from the beginning of the follow-up period to an event or end of the study, loss of contact, or withdrawal from the study is called survival time t. Often at the end of the follow-up period, the event will probably not have occurred for all participants and is described as censored. There are also occasions when participants may be uncooperative and it is hard to contact them or they may have changed their address. Others might no longer be interested in participating or may have a different event, such as death owing to an accident, which is irrelevant to the endpoint of interest. For these sample members, we have partial information. We know that the event has not occurred until the time we lost contact, but we are not sure if it will happen sometime after the date of the last follow-up. These situations are also described as censored. This approach is called right censoring. Right censoring generally happens when an individual is observed from a starting time point t0 up to t1 and has not had the event of interest, such that all we know is that the event has not occurred up to censoring time t1 . Other types of censoring are available that are less common in the literature; we will not present them here. The analysis of survival data is mainly based on an estimation of two probability functions: survival and hazard. Survival SðtÞ is defined as the probability of surviving (or not presenting with the event of interest) until time t. On the other hand, hazard hðtÞ is the conditional probability of dying at time t (or presenting with the event) having survived to that time. The figure illustrating SðtÞ against t is called the survival curve. The survival curve is an informative illustration of prospective research data that also visually allowcomparisons of survival between groups of participants exposed to a dietary factor (i.e., low, moderate, or high consumption of vegetables). There are several methods for estimating survival and drawing a survival curve. In this chapter, we will focus on two frequently used statistical approaches: KaplaneMeier and Cox regression analysis.

6.4.2 KAPLANeMEIER ANALYSIS KaplaneMeier analysis [16,17] is a popular nonparametric statistical approach to estimating survival and drawing a survival curve from the sample data of nutrition studies with a prospective character (i.e., cohort studies and clinicaltrials). The method assumes that the probability of surviving j or more periods of time from the start of the study can be calculated by j survival rates for each period: Sð jÞ ¼ P1  P2  .  Pj where P1 is the probability of surviving the first time period, P2 is the probability of surviving beyond the second period of time having survived up to the second period, etc. This probability can be generally calculated as: P¼

Number of participants alive ðor wihout the eventÞ at the start of the time period  number of deaths ðor eventsÞ Number of participants alive ðor wihout the eventÞ at the start of the time period

To comprehend KaplaneMeier analysis better, all further descriptions will be made in the context of Example 6.5.

6.4 SURVIVAL ANALYSIS

151

Example 6.5. A prospective study aimed to examine the development of obesity in a sample population of 100 normal-weight adults over 24 months (2 years). A total of 20 new cases of obesity occurred within the 2 years and seven participants were lost to follow-up. The researchers were interested in applying the KaplaneMeier methodology and drawing a survival curve for this sample population. The first step to apply the nonparametric KaplaneMeier approach is to draw a table that will assist in calculating survival, which in this case is also called the cumulative proportion of surviving, SðtÞ. The table is known as a life table and includes the survival time intervals, the number of participants with the event of interest, the number of censored participants, the number of participants at risk at the start of the time interval, the proportion (probability) of surviving P, and the cumulative proportion of surviving, SðtÞ. The life table for the data in this example was drawn as follows.

Censored, n

Participants at Risk (Those Who Have Not Developed Obesity at Start of Time Period), n

Proportion of Surviving P (i.e., Not Developing Obesity)

1

0

100

5

2

0

99

10

1

0

97

12

4

0

96

13

0

2

92

14

0

1

90

15

2

0

89

16

5

0

87

17

0

2

82

19

0

2

80

20

1

0

78

23

2

0

77

24

2

0

75

1001/ 100 ¼ 0.990 992/ 99 ¼ 0.980 971/ 97 ¼ 0.990 964/ 96 ¼ 0.958 920/ 92 ¼ 1 900/ 90 ¼ 1 892/ 89 ¼ 0.978 875/ 87 ¼ 0.943 820/ 82 ¼ 1 800/ 80 ¼ 1 781/ 78 ¼ 0.987 772/ 77 ¼ 0.974 752/ 75 ¼ 0.973

Survival Time t (months)

Participants Who Developed Obesity, n

0 1

Cumulative Proportion of Surviving, SðtÞ 1 1  0.990 ¼ 0.99 0.99  0.980 ¼ 0.97 0.97  0.990 ¼ 0.96 0.96  0.958 ¼ 0.92 0.92  1 ¼ 0.92 0.92  1 ¼ 0.92 0.92  0.978 ¼ 0.90 0.90  0.943 ¼ 0.85 0.85  1 ¼ 0.85 0.85  1 ¼ 0.85 0.85  0.987 ¼ 0.84 0.84  0.974 ¼ 0.82 0.82  0.973 ¼ 0.79

152

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

The proportion of surviving members in each time interval is calculated as: P¼

Number of participants at risk  number of participants who developed obesity Number of participants at risk

The cumulative proportion of surviving SðtÞ according to the KaplaneMeier assumption is calculated every time as the product of the cumulative proportion of surviving SðtÞ of the previous step (time interval) and the current proportion of surviving P. More precisely, in the current example, at the start of the study (time 0), no case of obesity had been observed and the cumulative proportion of surviving SðtÞ was 1 (i.e., 100%). In the first month, one case of obesity occurred and the cumulative proportion of surviving SðtÞ dropped from 1 to 1 x P2 ¼ 1  1001 100 ¼ 1  0:99 ¼ 0:99. Later, in the fifth month, two new cases of obesity occurred and the cumulative proportion of surviving SðtÞ dropped from 0.99 to 0:99  P5 ¼ 0:99  992 99 ¼ 0:99  0:980 ¼ 0:97. After 24 months, 20 news cases developed and the cumulative proportion of surviving SðtÞ dropped from 1 to 0.79. Using the data from the life table, we may easily draw the survival curve; the figure illustrates the SðtÞ against t. During the follow-up period of the study, two participants were lost to follow-up in the 13th month, one in the 14th, two in the 17th and two in the 19th. At those time intervals, no drop occurred in the cumulative proportion of surviving SðtÞ; however, the number of participants who were at risk was reduced (see the life table and Fig. 6.1). Despite the importance of the KaplaneMeier methodology, its application is based on three major assumptions. First, it is assumed that at any time, censored participants have the same prospects for survival as those who continue in the study. Second, the probability of survival does not alter for participants who were recruited early or late in the study. Third, the event occurs at the time specified [16]. KaplaneMeier analysis can be performed only if these assumptions are covered, and a related statement should be provided when the results are presented in a scientific report.

6.4.3 LOG-RANK TEST Another aim of survival analysis when it is applied in nutrition data is to compare survival among groups of participants exposed to a dietary factor. For instance, researchers of a cohort study might be interested in comparing the development of diabetes in consumers or nonconsumers of vegetables; others might be interest in comparing the effectiveness of two dietary interventions applied to different groups of participants in the management of obesity. Estimating survival using the KaplaneMeier approach separately for groups of participants exposed or not to the dietary factor and drawing survival curves together in one figure is the first step. This can help to visualize the difference in survivals. A vertical distance between them means that at a specific time point, one group had a greater fraction of subjects surviving. On the other hand, a horizontal distance means that it took longer for one group to experience a certain fraction of events. Visual distances cannot be considered proofs of a significant difference in the overall survival of groups of exposed and nonexposed participants. However, this can be confirmed by applying a logrank test, which assesses whether survival in the whole follow-up period of individuals in two or more groups is significantly different. This nonparametric statistic involves a concept similar to that of the chi-square test for comparing categorical data (see Chapter 5); it tests the null hypothesis that there is no difference in survival

6.4 SURVIVAL ANALYSIS

153

FIGURE 6.1 KaplaneMeier estimation of survival curve.

between groups, compared with the alternative hypothesis that there is a difference. It follows the same assumptions as KaplaneMeier analysis and compares the observed number of events against the expected ones. For a comparison of the survival of two groups, the test statistic is [18,19]: Log  rank test ¼

ðO1  E1 Þ2 ðO2  E2 Þ2 þ ; E1 E2

where O1 and O2 are the observed number of events in Groups 1 and 2 of the participants and E1 and E2 are the corresponding expected number of events in the same groups. Similar statistics are used to compare the survival of more than two groups of participants. Most statistical software [20e22] calculates the value of the log  rank test and the related P  value. When P  value < a (i.e., the statistical significance level), there is statistical evidence to

154

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

reject H0 and support for the concept that the survival rates of the studied groups are statistically different. Example 6.6. In a prospective study, Tevik et al. analyzed data from 131 Norwegian patients with chronic heart failure who were observed for 3 year after hospitalization [23]. The main dietary exposure was the nutritional status of the patients as measured using the Nutritional Risk Screening (NRS-2002) score. The total index ranged from 0 to 7; if the patient received a score of 3 or more, he or she was classified as being at nutritional risk. In this manner, two groups of participants were generated according to NRS-2002: those who were at nutritional risk at the time of recruitment (NRS-2002  3) and those who were not (NRS-2002 < 3). The main outcome of the study was the death from any cause. The researchers applied the KaplaneMeier approach to estimate the survival rates for the two groups of patients; the survival curves are illustrated in Fig. 6.2. The visual difference in the two curves indicates that the overall survival of patients not at nutritional risk (i.e., NRS-2002 < 3) was greater than that of those at nutritional risk. To prove the statistical significance of this visual difference, the researchers calculated the related log-rank test. The analysis revealed P < .001, which provided statistical evidence to reject H0 and supported the concept that the survival rates for the two groups of patients were statistically different.

FIGURE 6.2 Comparison of the survival functions in two groups of Nutritional Risk Screening (NRS)-2002. Cum, cumulative.

6.4 SURVIVAL ANALYSIS

155

Example 6.7. In another prospective study So¨derstro¨m et al. analyzed data from 1767 people aged 65 years, who were admitted to the hospital and observed for 50 months [24]. The main dietary exposure similar to Example 6.6 was the nutritional status at the time of recruitment, this time measured with the Mini Nutritional Assessment score. In this analysis, participants were categorized into three rather than two groups of well-nourished, malnourished patients who were at risk for malnutrition. The aim of the researchers was to examine whether nutritional status was an independent predictor of preterm death in this sample population. KaplaneMeier analysis was applied and the survival rates of the three groups of patients were estimated; the survival curves are illustrated in Fig. 6.3. The visual difference in the three curves indicates that the overall survival of well-nourished participants was greater than that of the other two groups. To prove the statistical significance of this visual difference, the researchers calculated the related log-rank test. The analysis revealed P 1) or lower (when RR < 1) the risk was for the occurrence of an event when the x increased 1 unit. When the Cox regression model has a multivariate form, we may add to the interpretation of the RRs that their estimation was performed after taking into consideration other explanatory variables.

6.4.4.2 Application and Examples The Cox proportional hazards model is semiparametric because there are no assumptions about the shape of the baseline hazard but there is aproportional hazard assumption. This should be confirmed when we apply the modeling procedure to real data. There are several more or less complicated ways to test this assumption. For the purposes of this chapter, we will refer to the KaplaneMeier curves, the ln-ln graph, and Schoenfeld global test. Generation of KaplaneMeier curves (presented earlier in this chapter) can be informative for aviolation of the proportional hazards assumption. In fact, when the estimated curves for groups of the dietary exposure do not cross over time, this indicates proportionality. This can also be verified by an ln-ln graph. It is another visual technique in which the -ln(-ln[survival probability]) is plotted against the ln(survival time) separately for the groups of dietary exposures. The proportional hazards assumption is not violated when the curves are parallel. The Schoenfeld global test is a statistic which under the null hypothesis confirms the assumption. When p  value > a (i.e., the statistical significance level) is indicated, there is statistical evidence to assume that the proportional hazards assumption is confirmed by the sample data. Both graphs and the Schoenfeld global test are generated bymost statistical software; all ofthese methods can be used in combination to provide more robust evidence to confirm the assumption. One of the great advantages of applying a Cox regression analysis to health and nutrition data is the potential of evaluating for confounding or interaction effects in an observed association. These concepts have been previously discussed for linear and binary logistic regression analysis and they have similar approaches for Cox modeling (see Chapter 5). Together with the choice of the best model, these evaluations are challenging tasks that require excellent knowledge of the literature on the topic, a good background in biostatistics and epidemiology, and experience in similar studies. Example 6.8. In a prospective study of type 2 diabetes, a cohort of 1000 male and female participants aged 40e50 years was observed for 10 years. The newly diagnosed cases of type 2 diabetes were recorded through a validated and accurate process. At the beginning of the study, the consumption of red meat was also assessed. a. The researchers were interested in evaluating the association of dietary exposure with the studied health outcome. During the follow-up period, 120 newly diagnosed cases of type 2 diabetes were recorded and the full survival data (i.e., follow-up status, time, censoring) were input into the statistical software.

158

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

Red meat consumption (1 ¼ yes/0 ¼ no)

b b

P Value

95% Confidence Interval

0.53

0.002

0.49

0.56

Notice: The survival data were formed as right censored. The proportional hazards assumption was tested graphically and by Schoenfeld global test.

The Cox regression analysis indicated the following results: The dietary exposure was dichotomous, so the RR can be calculated as: RR ¼ eb ¼ e0:47 z1:70 The correspondent limits for the 95%CI of the RR are: Lower limit ¼ e0:49 z 1:63 Upper limit ¼ e0:56 z 1:75 Thus, the results indicated that exposure to the dietary factor (i.e., red meat consumption) was significantly associated with a higher risk for type 2 diabetes (RR ¼ 1.70 > 1, P ¼ .002 < .05). In more detail, the risk for type 2 diabetes for red meat consumers was 1.70 times (or ð1:70  1Þ100% ¼ 70%) higher than fornonconsumers. This estimation may vary in the reference population within the (1.63, 1.75) with a 95% level of confidence (i.e., the 95%CI for the RR). b. Suppose that an assessment of red meat was performed on a continuous scale (i.e., grams per day). The results from the Cox regression analysis are given in the following table:

Red meat consumption, g/day

b b

P Value

0.01

0.002

95% Confidence Interval 0.005

0.02

Notice: The survival data were formed as right censored. The proportional hazards assumption was tested graphically and by Schoenfeld global test.

Dietary exposure is a continuous variable, so the RR for a 1-unit (i.e., 1 g/day) increase in red meat consumption can be calculated as: RR ¼ eb ¼ e0:01 z 1:01 The corresponding limits for the 95%CI of the RR are: Lower limit ¼ e0:005 z 1:005. Upper limit ¼ e0:02 z 1:02 Similar to (a), the results indicate that exposure to the dietary factor (i.e., red meat consumption) was significantly associated with a higher risk for type 2 diabetes (RR ¼ 1.01 > 1, P ¼ .004 < .05). In more detail, a 1-g/day increase in red meat consumption was associated with 1.02 times higher risk for the disease (or ð1:01  1Þ100% ¼ 1%). This estimation may vary in the reference population within the (1.005, 1.02) with a 95% level of confidence (i.e., the 95%CI for the RR). However, this interpretation has no actual meaning for clinical nutritionists because it is hard to measure a 1-g/day increase in red meat consumption. For this reason, the RR may be expressed as a more understandable increase in red meat consumption (for instance, 50 g/day).

REFERENCES

159

In this case, the RR and the 95% CI are calculated as: OR ¼ eb50 ¼ e0:0150 z 1:65 The corresponding limits for the 95%CI of the odds ratio are: Lower limit ¼ e0:00550 z 1:28. Upper limit ¼ e0:0250 z 2:72 Thus, a 50-g/day increase in red meat consumption is associated with a 1.65 times higher risk for type 2 diabetes (or ð1:65  1Þ100% ¼ 65%). This estimation may vary in the reference population within the (1.28, 2.72) with a 95% level of confidence (i.e., the 95%CI for the RR). c. Suppose further that the researchers are interested in evaluating the association presented in (a) after controlling for the confounding effects of age and body mass index (BMI). In this case, the multiple form of Cox regression analysis is used; the results are shown in the table.

Age, years Body mass index, kg/m2 Red meat consumption (1 ¼ yes/0 ¼ no)

b b

P Value

95% Confidence Interval

0.07 0.10 0.41

0.001 0.02 0.01

0.06 0.08 0.33

0.08 0.12 0.58

Notice: The survival data were formed as right censored. The proportional hazards assumption was tested graphically and by Schoenfeld global test.

The results indicate that red meat consumption remains significantly associated with a higher risk for type 2 diabetes after adjustments for age and BMI (RR ¼ e0:41 z1:51 > 1, P-value ¼ .01 < .05). In fact, the risk for type2 diabetes in red meat consumers is 1.51 times (or ð1:51  1Þ100% ¼ 51%) higher This estimation may vary in the reference population within  0:33 than for0:58nonconsumers.  e z1:40; e z1:79 with a 95% level of confidence (i.e., the 95%CI for the RR). In the same model, it was observedthat both age (RR for 5 years increase ¼ e0:075 z1:42 > 1, P ¼ .001 < .05) and BMI (RR for 1 Kg m2 increase ¼ e0:10 z1:11 > 1, P ¼ .02 < .05) were significantly associated with a higher odds of the disease after adjusting for two other variables in each case. corresponding 95%CIs for these estimations were e0:065 z1:35; e0:085 z1:49 and The  e0:08 z1:08; e0:12 z1:13 .

REFERENCES [1] Lachat C, Hawwash D, Ocke´ MC, Berg C, Forsum E, Ho¨rnell A, et al. Strengthening the reporting of observational studies in epidemiology e nutritional epidemiology (STROBE-nut): an extension of the STROBE statement. Nutr Bull 2016. https://doi.org/10.1111/nbu.12217. [2] Patience, G.S., Boffito, D.C., Patience, P.A., n.d. Communicate science papers, presentations, and posters effectively : papers, posters, and presentations. [3] Pounis G, Costanzo S, Bonaccio M, Di Castelnuovo A, de Curtis A, Ruggiero E, et al. Reduced mortality risk by a polyphenol-rich diet: an analysis from the Moli-sani study. Nutrition 2018;48:87e95. https://doi.org/ 10.1016/j.nut.2017.11.012.

160

CHAPTER 6 STATISTICAL ANALYSIS OF PROSPECTIVE HEALTH

[4] Pounis G, Tabolacci C, Costanzo S, Cordella M, Bonaccio M, Rago L, et al. Reduction by coffee consumption of prostate cancer risk: evidence from the Moli-sani cohort and cellular models. Int J Canc 2017; 141(1). https://doi.org/10.1002/ijc.30720. [5] Webb, P., Bain, C., Page, A., n.d. Essential epidemiology: an introduction for students and health professionals. Retrieved from: http://admin.cambridge.org/academic/subjects/medicine/epidemiology-publichealth-and-medical-statistics/essential-epidemiology-introduction-students-and-health-professionals-3rdedition#h0CZYvMgg2uhJLms.97. [6] Rothman KJ. Epidemiology : an introduction. Oxford University Press; 2012. [7] Greenberg RS, Daniels S, Flanders W, Eley J, Boring J. Medical epidemiology. Lange Medical Books/ McGraw-Hill; 2005. [8] Riffenburgh RH, (Robert H. Statistics in medicine. Elsevier/Academic Press; 2012. [9] Gerstman BB. Epidemiology kept simple : an introduction to traditional and modern epidemiology. John Wiley & Sons; 2013. [10] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 3 e Measures and tests of association between two variables. In: Biostatistics and computer-based analysis of health data using R; 2016. p. 41e63. https://doi.org/ 10.1016/B978-1-78548-088-1.50003-8. [11] Willett W. Nutritional epidemiology. Oxford University Press; 2012. https://doi.org/10.1093/acprof:oso/ 9780199754038.001.0001. [12] Forthofer RN, Lee ES, Hernandez M, Forthofer RN, Lee ES, Hernandez M. 11 e Analysis of survival data. In: Biostatistics; 2007. p. 297e321. https://doi.org/10.1016/B978-0-12-369492-8.50016-9. [13] Hoffman JIE, Hoffman JIE. Chapter 35 eSurvival analysis. In: Biostatistics for medical and biomedical practitioners; 2015. p. 621e43. https://doi.org/10.1016/B978-0-12-802387-7.00035-4. [14] Klein JP, Zhang M-J. 9 Survival analysis. Handb Stat 2007;27:281e320. https://doi.org/10.1016/S01697161(07)27009-9. [15] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 5 e Survival data analysis. In: Biostatistics and computerbased analysis of health data using stata; 2016. p. 101e15. https://doi.org/10.1016/B978-1-78548-1420.50005-5. [16] Bland JM, Altman DG. Survival probabilities (the Kaplan-Meier method). Br Med J 1998;317(7172):1572. Retrieved from: http://www.ncbi.nlm.nih.gov/pubmed/9836663. [17] Kaplan EL, Meier P. Nonparametric estimation from incomplete observations. J Am Stat Assoc 1958; 53(282):457e81. https://doi.org/10.1080/01621459.1958.10501452. [18] Bland JM, Altman DG. The logrank test. BMJ 2004;328(7447):1073. https://doi.org/10.1136/ bmj.328.7447.1073. [19] Kleinbaum DG, Klein M. Survival analysis : a self-learning text. Springer; 2005. [20] Lalanne C, Mesbah M. Biostatistics and computer-based analysis of health data using SAS. Elsevier Science; 2017. Retrieved from: http://www.sciencedirect.com/science/book/9781785481116. [21] Lalanne C, Mesbah M. Biostatistics and computer-based analysis of health data using Stata. 2017. [22] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 1 e Elements of the language. In: Biostatistics and computerbased analysis of health data using R; 2016. p. 1e22. https://doi.org/10.1016/B978-1-78548-088-1.50001-4. [23] Tevik K, Thu¨rmer H, Husby MI, de Soysa AK, Helvik A-S. Nutritional risk is associated with long term mortality in hospitalized patients with chronic heart failure. Clin Nutr ESPEN 2016;12:e20e9. https:// doi.org/10.1016/J.CLNESP.2016.02.095. [24] So¨derstro¨m L, Rosenblad A, Adolfsson ET, Saletti A, Bergkvist L. Nutritional status predicts preterm death in older people: a prospective cohort study. Clin Nutr 2014;33(2):354e9. https://doi.org/10.1016/ J.CLNU.2013.06.004. [25] Cox DR. Analysis of survival data. Routledge; 2018. https://doi.org/10.1201/9781315137438.

REFERENCES

161

[26] Collett D. Modelling survival data in medical research. Chapman & Hall/CRC; 2003. Retrieved from: http://go.galegroup.com/ps/anonymous?id¼GALE%7CA123086147&sid¼googleScholar&v¼2.1&it¼ r&linkaccess¼abs&issn¼01621459&p¼AONE&sw¼w. [27] Lalanne C, Mesbah M, Lalanne C, Mesbah M. 6 e Survival curves, Cox regression. In: Biostatistics and computer-based analysis of health data using SAS; 2017. p. 115e39. https://doi.org/10.1016/B978-1-78548111-6.50006-X. [28] Riffenburgh RH, Riffenburgh RH. Chapter 23 e Survival, logistic regression, and Cox regression. In: Statistics in medicine; 2012. p. 491e508. https://doi.org/10.1016/B978-0-12-384864-2.00023-8. [29] Wahba G. 23 Statistical learning in medical data analysis. Handb Stat 2007;27:679e711. https://doi.org/ 10.1016/S0169-7161(07)27023-3.

This page intentionally left blank

CHAPTER

META-ANALYSIS OF NUTRITION STUDIES

7

Emmanouil Bouras1, Konstantinos K. Tsilidis2, George Pounis3, Anna-Bettina Haidich1 Department of Medicine, Aristotle University of Thessaloniki, Thessaloniki, Greece1; University of Ioannina School of Medicine, Ioannina, Greece2; Alimos, Athens, Greece3

CHAPTER OUTLINE 7.1 Introduction .................................................................................................................................164 7.2 Methodology of Meta-Analysis in Nutrition Research ...................................................................... 165 7.2.1 Defining the Search Strategy .....................................................................................166 7.2.2 Study Selection Procedure ........................................................................................169 7.2.3 Quality Assessment ..................................................................................................169 7.2.4 Data Extraction ........................................................................................................170 7.3 Statistical Methodologies Applied in Meta-Analysis of Nutrition Studies ..........................................170 7.3.1 Statistical Measures of Effect Included in Meta-Analysis .............................................171 7.3.2 Choice of Meta-Analytical Method..............................................................................173 7.3.3 Statistical Heterogeneity ...........................................................................................175 7.3.3.1 Subgroup Analysis .............................................................................................. 176 7.3.3.2 Meta-Regression ................................................................................................. 176 7.3.3.3 Sensitivity Analysis .............................................................................................. 176 7.3.3.4 Prediction Intervals ............................................................................................. 177 7.3.4 Small-Study Effects..................................................................................................177 7.3.5 Software for Meta-Analysis ........................................................................................178 7.4 Presentation and Interpretation of Results...................................................................................... 178 7.4.1 Study Selection........................................................................................................179 7.4.2 Study Characteristics................................................................................................179 7.4.3 Forest Plot ...............................................................................................................179 7.4.4 Assessing Heterogeneity............................................................................................182 7.4.4.1 Subgroup Analysis .............................................................................................. 182 7.4.4.2 Meta-Regression ................................................................................................. 182 7.4.4.3 Sensitivity Analysis .............................................................................................. 182 7.4.5 Risk for Bias ............................................................................................................184 7.4.6 Funnel Plot..............................................................................................................184 7.5 Limitations and Biases..................................................................................................................187 7.5.1 Challenges...............................................................................................................189 References ..........................................................................................................................................191 Analysis in Nutrition Research. https://doi.org/10.1016/B978-0-12-814556-2.00007-5 Copyright © 2019 Elsevier Inc. All rights reserved.

163

164

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

7.1 INTRODUCTION The great advance in medical research that has taken place in past decades has been accompanied by an increase in the volume of studies produced yearly, which makes it challenging for health professionals to keep up with the literature. As a result, reviews that summarized the key points of a set of studies were an efficient approach to providing the bottom line on a research topic. In addition, it has been common for issues of great significance to be addressed in multiple studies, usually in a different context, and hence the idea of synthesizing data from a series of studies was invented. The journey of research synthesis existed for well over a century before it took the form it currently has [1]. One of the first remarks on the need to synthesize results from multiple studies dates back to 1885, when physicist Lord Rayleigh emphasized the need not only for “the reception of new material” but “the digestion and assimilation of the old” [2]. In 1904, Karl Pearson, who was asked to review evidence on the effects of a vaccine against typhoid, gathered data from 11 relevant studies of immunity and mortality among soldiers serving in various parts of the British Empire. He calculated correlation coefficients for each of the 11 studies and then synthesized the coefficients within two subgroups, thus producing average correlations; this was an early example of the use of metaanalytical techniques in medical research [3]. A few years later, Fisher, in his work Statistical Methods for Research Workers, proposed an easy method for combining Pvalues from several different studies through chi-square distribution [4]. A major contribution came by Cochran in 1954, who suggested a way to compute a test of association that controls for different sources of information [5]. This was followed by the development of an equivalent approach by Mantel and Haenszel in 1959, which included a method for obtaining a combined estimate of the odds ratio [6]. Meta-analysis was first introduced as a technique by Gene Glass in 1976, who defined metaanalysis as “the statistical analysis of a large collection of analysis results from individual studies for the purpose of integrating the findings” [7]. According to another definition, by Huque, metaanalysis is “a statistical analysis which combines or integrates the results of several independent clinical trials, considered by the analyst to be combinable”[8]. As the final component of a systematic review, meta-analysis can be considered a quantitative, formal, epidemiological study design used to assess the results of previous research systematically to derive conclusions about that body of research [9]. Most statistical techniques used today in metaanalysis have their origins in Gauss’s and Laplace’s work [10]. The first evidence from meta-analysis was published in the mid-1970s but it gained popularity in medicine only a decade later when researchers began to incorporate the idea systematically. To date, tens of thousands of meta-analyses have been published and their production is increasing exponentially (Fig. 7.1). Among the different forms of clinical research, it has been shown that meta-analyses have the highest relative citation impact [11]. Despite the criticism that they have received, like common allegorical expressions such as “mixing apples and oranges,”“Garbage in, garbage out,” and “One number cannot summarize a research field” [12], meta-analyses have been well-established at the highest level in proposed theoretical hierarchies of evidence-based medicine.

7.2 METHODOLOGY OF META-ANALYSIS IN NUTRITION RESEARCH

165

FIGURE 7.1 Approximate number of meta-analyses overall (blue) and in nutrition research (red) identified through a rough search in PubMed.

Meta-analysis in nutrition research intends to combine results from multiple studies on a single nutrient ora combination of nutrients or dietary supplements, or a particular dietary pattern in relation to an outcome and to provide an overall estimate of the effect on a defined parameter. One of the first meta-analyses in nutrition science was published in the mid-1980s and investigated whether diet modification was efficacious for hyperactivity [13]. Another study by Axelson examined the relationship between dietary intake and nutrition knowledge or nutrition-related attitudes [14]. Other meta-analyses at that time focused on the effectiveness of parenteral nutrition [15]. Considering the relative importance of meta-analysis in nutrition research, this chapter aims to provide a framework that enables researchers to understand the rationale of performing a metaanalysis, as well as how to implement and interpret meta-analytic procedures properly. The reader will have the opportunity to gain insight into the main aspects of a systematic review and metaanalysis, such as framing the research question, searching for relevant studies, and using appropriate meta-analytical techniques to combine the data and presenting them in a comprehensive way. All descriptions and presentations of the methodologies will be accompanied by examples assisting the reader in comprehending the main aspects of a meta-analysis.

7.2 METHODOLOGY OF META-ANALYSIS IN NUTRITION RESEARCH Over the years, protocol and guideline papers have been published providing guidelines for the sound production and reporting of the results of a systematic review and meta-analysis. Cochrane is an

166

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

organization specifically developed to promote, support, and disseminate systematic reviews and meta-analyses on the efficacy of interventions in the health care field; it serves as a remarkable example of the quality and methodology that can be employed in systematic reviews [16]. Since the launch of the Cochrane Database of Systematic Reviews in 1995, research syntheses published by this organization have had a considerable effect on the content of international guidelines and policies in health care. A sibling organization, the Campbell Collaboration, was inaugurated in the beginning of the 21st century to prepare, maintain, and disseminate systematic reviews of the effects of social and educational policies and practices [17]. Different meta-analytical approaches have been reported depending on the nature and availability of studies. Most published meta-analyses are produced using aggregate data extracted from a series of primary studies designed to address a particular research question. Another common meta-analytical approach includes the use of individual participant data (IPD), study-specific datasets containing a series of variables related to the study population. Although it may be more strenuous because it typically requires a lot of effort to collect raw data from multiple studies, such an approach has advantages over meta-analyses for aggregate data, such as the ability to standardize a statistical analysis, validate the quality of the data, and explore differences in effect in more detail across population subgroups [18]. However, most basic principles of systematic reviews and meta-analyses apply to both patient-level and summary-level analytical approaches. Potential peculiarities of IPD in the process of a meta-analysis are discussed in the relevant sections that follow. A clear definition of hypotheses to be investigated is the basis for planning and conducting a systematic review and meta-analysis. All steps to be followed must be described in transparently and reproducibly in the study protocol, which should be defined a priori and (ideally) be publicly available. A widely known registry of systematic reviews is PROSPERO, which includes protocol details for systematic reviews that focus on a health-related subject [19]. These steps include the search strategy, study selection process, assessment of risk for bias in the included primary studies, qualitative synthesis, and finally quantitative synthesis, which is the meta-analysis component. Comparisons to be made, the nature of the data needed for such comparisons, and the way in which the data will be combined all need to be considered when conducting a meta-analysis. The following sections describe the first steps in performing a systematic review and meta-analysis, such as the search strategy, inclusion and exclusion criteria, study selection and appraisal, and data extraction.

7.2.1 DEFINING THE SEARCH STRATEGY The first step, and, for many, the most challenging part in conducting a systematic review and metaanalysis, is to formulate the clinical question to be tested. According to the Preferred Reporting Items for Systematic Reviews and Meta-analyses (PRISMA) statement, an evidence-based minimum set of items for transparent reporting in systematic reviews and meta-analyses published in 2009, an explicit statement of questions being addressed should refer to participants (P), interventions (I), comparisons (C), outcomes (O), and study design (S), the so-called PICO(S) [20]. For instance, in the meta-analysis by Dibaba et al., which studied the effect of magnesium supplementation on blood pressure [21], the PICO can be formed as shown in Table 7.1.

7.2 METHODOLOGY OF META-ANALYSIS IN NUTRITION RESEARCH

167

Table 7.1 Formulating the Research Question P

I

C

O

(S)

Y Individuals with insulin resistance, prediabetes, or other noncommunicable chronic diseases

Y Magnesium supplementation

Y Placebo

Y Difference in systolic and diastolic blood pressure from baseline

Y Randomized controlled trials

A research question addressed by a meta-analysis is usually broader than those addressed by the original studies [22]. However, setting a broad question may limit comparability and hence the ability to combine results from such studies. On the other hand, a narrow research question may limit the number of eligible studies [23]. It is up to the researcher to decide where to set that limit to find the right balance. Key points of a well-framed research question can set up the basis for a comprehensive literature search. Typically, several components of a research question are incorporated into the search strategy through the use of relevant terminology in the form of controlled vocabulary (such as index terms) and free text (keywords) combined by boolean operators (such as “AND,”“OR,” and on certain occasions “NOT”). In the study by Liao et al., for instance, which aimed to identify the effect of protein supplementation combined with resistance exercise training on the body composition and physical function of overweight and obese elderly people [24], several general terms were used related to the agegroup of participants, exercise motive, and type of intervention. The researchers specifically stated that Regarding participant conditions, we used the following search terms: “older/aging/aged/elderly/seniors.” We used the following search terms to find intervention studies: “progressive resistance training, resistance exercise, strength training, weight training, and/or weight lifting,” and “protein/amino acid/nutrient supplement.”

Depending on the research question, it may be more appropriate to seek certain types of study design than others when performing a literature search [25]. Generally, randomized clinical trials are considered the reference standard when answering a question about interventional effects, such as therapy or a screening question. On the other hand, randomized clinical trials may not last long enough to provide a valid estimate for a question regarding harm, a prognosis, or an incidence,so in such cases observational studies may be included. The type of study design may be incorporated into the search strategy and/or set as an inclusion criterion for the original studies to be included in the meta-analysis. Sensitive search strategy algorithms for specific types of study design, such as randomized clinical trials, or observational studies have been proposed [26,27]. A source to identify published papers and abstracts is PubMed, an online MEDLINE database that includes up-to-date citations. Another popular source is Elsevier’s Scopus, which along with other utilities provides an overview of the citations a published article has received. Other databases likely to be relevant include the CENTRAL, ISI Science Citation Index, Embase, and BIOSIS.

168

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

Searching only one database may not be adequate because usually no single database can cover the entire set of articles available in a research field. Depending on the topic, the degree of overlap among different electronic databases varies considerably. For instance, it has been reported that the overlap was only 59% in English language “diabetes journals” indexed by both MEDLINE and Embase [28]. In another study, the relative overlap of nursing and allied health literature between CINAHL and Scopus was reported to be 58.5% [29]. Furthermore, going beyond a single database is important not only to ensure that all relevant studies are identified but also to minimize selection bias for those that are found. The researcher should keep in mind that the search strategy should be properly adopted in order to match each database’s standards (such as controlled vocabulary, field descriptor and others). The use of topic-specific databases, if available, may provide certain advantages. Examples of nutrition-related databases include the AGRIS [30], CABI [31], AGRICOLA [32] and Food Science and Technology Abstracts [33]. For example, examining the effect of oat b-glycans on blood lipids [34], Whitehead et al. used AGRICOLA, an online database developed and updated by the US National Agricultural Library of the US Department of Agriculture. Nevertheless, it is important to obtain multiple relevant databases because the loss of studies can lead to selection bias in the meta-analysis. Database searches can be augmented by manual searches of library resources for papers, books, abstracts, and conference proceedings or by searching the bibliographies of related systematic reviews or individual studies [35]. Searching the content of selected journals is also mandatory for a Cochrane systematic review. Time-consuming though it may be, it has been shown that manual searching is an effective technique for identifying reports of randomized controlled trials (RCTs) in the systematic review process. In a Cochrane review, it was reported that manual searching identified 92%e100% of the total number of reports of randomized trials, whereas searching MEDLINE retrieved 55%, Embase 49%, and PsycINFO 67% [35]. Cross-checking references or citations in review papers and communicating with scientists who work in the fieldare also helpful. A number of studies on the topic of interest may not be published or might not be indexed in computer-searchable databases. Evidence suggested that there was a difference in the effect size estimation between meta-analyses that include both published and unpublished studies versus metaanalyses that included only published data, which has driven a move toward the need to search for unpublished studies when performing a meta-analysis [36]. Useful sources of unpublished studies can be registries of clinical trials such as the National Library of Medicine, government sites such as the US Food and Drug Administration, organizations and foundations such as the World Health Organization (WHO), and dissertation or industry report repositories such as ProQuest and OpenGrey. For instance, in the previous meta-analysis example of Whitehead et al., the unpublished reports of a local industry were also searched. The set of search limits should be justified in the context of the research topic; otherwise, it may introduce bias or uncertainty in the review. For example, a language limit has been found to influence the summary effect [37]. Date limits are justified only in case an intervention was introduced at a specific timepoint or a particular disease emerged at a particular time. However, it is important for limitations that are set on each occasion to be well-presented and discussed. In addition, the reviewer should not forget that a number of duplications usually need to be removed before screening processes begin.

7.2 METHODOLOGY OF META-ANALYSIS IN NUTRITION RESEARCH

169

7.2.2 STUDY SELECTION PROCEDURE Once the literature search is complete, the study selection process follows, based on the inclusion and exclusion criteria that have been set by the researchers, considering the various components of the research hypothesis. The inclusion criteria may include the study design; the disease or health condition under investigation; the participants’ characteristics; the demographics; the characteristics of treatment or exposure; and the outcome elements, such as the domain, the metrics and measurements, and the methods of aggregation [38]. In the previous example by Liao et al., the use of broader terms in the search strategy is followed by a set of well-defined criteria, to support the research question. The criteria specify, among others, the participants’ characteristics (such as age and body composition at baseline), factors related to the intervention (in the form of supplements that are considered eligible), the comparator characteristics (such as the form of placebo), the elements of the primary and secondary outcomes (change in specific measures, such as the change in fat-free mass from baseline), and the type of study design. Inclusion and exclusion criteria are applied during the screening process, usually in two stages. First, title and abstract screening is used to filter out a significant number of studies. In the second stage, the remaining studies pass through a full text screening process, where they are scrutinized to determine suitability for inclusion. When performing a meta-analysis of IPD, the selection of studies can be more strenuous, depending on the individual study datasets that are offered by the principal investigators. It often depends on the investigators’ willingness to be involved in the meta-analysis as a project. On the other hand, obtaining the IPD enables the researcher to analyze the data from each study consistently, given that the methodology (e.g., definition of outcomes, adverse events, adjustments for potential covariates such as age, body weight, smoking) may differ somewhat from study to study.

7.2.3 QUALITY ASSESSMENT An assessment of the quality of the studies included in a meta-analysis is an essential process and ensures the validity of the review. A researcher needs to obtain all relevant studies on a research topic if possible and at the same time reduce the number of studies of low quality. However, restricting the meta-analysis to only high-quality studies may result in less available data for the final analysis. Ascertaining the quality of the studies and the meta-analysis itself is a challenging topic that attracts the attention of reviewers. As a result, comprehensive guidelines on the reporting of randomized trials, such as the Consolidated Standards of Reporting Trials statement [39] and the reporting of observational studies such as the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) statement (which provides an extension for nutritional epidemiology, the STROBE-nut) [40] and other articles providing recommendations for improving quality in studies have been published, most of which can be found in the Enhancing the Quality and Transparency of Health Research Network [41]. Tools have also been developed that provide checklists to help reviewers assess the quality of studies, such as the Cochrane Risk of Bias Tool [16], the NewcastleeOttawa Scale [42], and others [43,44]. These tools assist reviewers in identifying components of the study that have significance in terms of quality or the presence of potential bias. The components may differ according to the type of the study, but generally the terms pertain to the internal or external validity of each study. In RCTs, for instance, the

170

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

randomization process, allocation concealment, participant eligibility, the blinding feature, the similarity of interventions, and the number of withdrawals are among the several focal components. On the other hand, the comparability of cases and controls, the exposure and outcome assessment, and adjustment for confounders are crucial in a quality evaluation of observational studies. In a similar fashion, tools also exist to assess the quality of meta-analyses, such as A Measurement Tool to Assess Systematic Reviews and ROBIS [45,46]. In a previously described example, Whitehead et al. assessed the quality of the included studies using a specific set of criteria described by the European Food Safety Authority, pertinent to human intervention studies [47]. The authors specifically stated that The quality criteria were based on Appendix H of the European Food Safety Authority guidance for the preparation and presentation of the application for authorization of a health claim.

The quality assessment should be extended to clinical and methodological characteristics of the participating studies in a meta-analysis, such as the size of the sample, the inclusion and exclusion criteria for certain subgroups, the times used as study endpoints or the duration of the intervention, the strengths and limitations, the potential uncertainties in the design or execution of the studies, and the statistical methods, which may bias the results.

7.2.4 DATA EXTRACTION All available data necessary to answer the research hypothesis should be obtained from all participating studies. In performing a meta-analysis of aggregate data, components that are commonly extracted from the original studies include, among others, study design characteristics, several descriptives of the study groups, such as the number of participants, age and gender, relevant diagnostic criteria, treatment characteristics or levels of exposure, length of follow-up, information on the background diet, and other baseline or endpoint measures. A data extraction form such asa table or spreadsheet is usually constructed to record the data. Application software is available to assist a reviewer in study identification, selection, and data extraction, enabling online storage of the review data [48].

7.3 STATISTICAL METHODOLOGIES APPLIED IN META-ANALYSIS OF NUTRITION STUDIES Meta-analysis as a major research design in nutrition science is based on the statistical analysis of pooled data. Depending on the nature of the studies that are assessed and the aims of the meta-analysis, the researcher should consider applying various statistical measures and methods. The setting of the statistical methodology is a challenging point requiring a certain mathematical background. This might result an unfamiliar environment for nutrition and medical researchers. For this reason, and similar to information provided in Chapters 5 and 6, an effort was made to simplify the presentation of statistical techniques. This does not limit the attention that should be given when these methodologies are applied to real data; consultation with experienced biostatisticians is always recommended. Statistical terms commonly used in the field of meta-analyses are presented in Table 7.2 with the aim of helping the reader comprehend particular concepts that are described in the next sections.

7.3 STATISTICAL METHODOLOGIES APPLIED IN META-ANALYSIS

171

Table 7.2 Terminology Commonly Used in the Field of Meta-analysis

· Relative Risk: General term used to refer to relative measures of the magnitude of effect of the intervention or risk factor on the outcome, such as hazard ratio, odds ratio (OR), risk ratio (RR), or rate ratio. for Binary Outcomes: · Metrics Represents the odds that an outcome (e.g., disease or disorder) will occur given a particular exposure to the · OR: variable of interest (e.g., health characteristic, aspect of medical history) compared with the odds of the outcome occurring in the absence of that exposure. ORs are most commonly used in caseecontrol studies; however, they can also be used in cross-sectional studies. RR: Represents the ratio of the probability of an event occurring (e.g., developing a disease, being injured) in one group (e.g., the exposed ones) to the probability of the event occurring in a comparison, nonexposed group. Relative risk is used in randomized controlled trials and cohort studies. Risk Difference: The difference between two proportions. Mean Difference (MD): Measures the absolute difference between the mean values in two different groups. In a clinical trial, it provides an estimate of the extent of the difference between the averages of the experimental group and control groups. Metrics for Continuous Outcomes: Weighted MD: Measures the difference between the mean values in two different groups weighted by a precision measure. Standardized MD: Expresses the size of the intervention effect in each study relative to the variability observed in that study. It is used when the studies all assess the same outcome but measure it in various ways. Summary Effect Size: Pooled effect size generated by combining individual effect sizes in a meta-analysis. Fixed-Effects Model: Mathematical model that combines the results of studies that assume the effect of the intervention is constant in all subject populations studied. Only within-study variation is included when assessing the uncertainty of results. Random Effects Model: Mathematical model for combining the results of studies that allow for variation in the effect of the intervention among the subject populations studied. Both within-study variation and between-study variation are included when assessing the uncertainty of results.

·

· · · · ·

· · ·

7.3.1 STATISTICAL MEASURES OF EFFECT INCLUDED IN META-ANALYSIS A wide variety of statistical measures can be used in a meta-analysis. The selection of an appropriate one depends on the design of the original studies that are assessed and the type of data. Most metaanalyses on nutrition research are syntheses of effect sizes such as the standardized mean difference, risk ratio, and odds ratio. The most popular measures are included in Table 7.2. Quantitative measures are typically summarized using means and standard deviations. In a clinical trial, for instance, in which a continuous variable such as body weight is examined as the outcome, one might be interested in the mean change in the variable at a specified timepoint in a group that received a particular intervention, in contrast with the mean change in another group that received a different intervention or placebo. In this case, the effectiveness of the intervention can be defined as the difference in the change or percent change from baseline in the outcome (e.g., body weight) between an experimental intervention and a control group, which can be expressed overall as themean difference (MD). When conducting a meta-analysis, relevant studies are collected, each trial’s MD is given a relative weight depending on the study’s precision, and then the MDs are synthesized by metaanalytical methods to produce the summary effect size, which is termed the weighted MD(WMD) between the two groups. For example, in the study by Pan et al.[49], the estimate of the primary outcome was defined as the MD (net change in mmol/L) in lipid concentrations between subjects

172

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

assigned to consume flaxseed or its derivatives and those assigned to the control regimen. By combining results from 11 RCTs, the researchers specifically concluded that A significant reduction in total cholesterol was found in studies using whole flaxseed, with a net change of e0.19 mmol/L (95% confidence interval [CI]: 0.29,
0.09 mmol/L)

For some types of continuous measures, multiple assessment tools may be available. In this case, a standardized MD(SMD), expressed in standard deviation units, may assist in synthesizing the results of different studies. A popular formulation used to calculate the SMD from studies with two independent groups is the Cohen’s d, which can be estimated as: d¼

X1
X2 ; Swithin

where X1 and X2 are the sample means in the two groups and Swithin is the standard deviation pooled across groups. Alternative approaches are the Hedge’s G and Glass D methods [50e53]. To increase interpretability, SMD can be supplemented by reporting the MD of the most common measure [54]. Again, in a meta-analysis, SMDs may be pooled to produce the summary effect size. For instance, in the study by Abbott et al., in a population aged >18 years, a meta-analysis was performed to determine whether dietary or supplement-based ne3 polyunsaturated fatty acid (PUFA) interventions affected measures of insulin resistance (IR) and insulin sensitivity (IS) in a sexdependent manner [55]. To account for the variety of methods used to measure IR and IS, the principle summary measure was the SMD. The researchers specifically stated that With all studies pooled together, the ne3 PUFA intervention had no significant effect on measures of IR with an SMD of 0.089 (95% CI:
0.105, 0.283; P ¼ 0.367).

A continuous variable may be also dichotomized and analyzed as a binary outcome. For example, in a study of a cholesterol-lowering nutritional intervention, the low-density lipoprotein (LDL)-C measurement at the study end can be dichotomized into “normal” or “pathological,” depending on whether it is below a certain target level (such as 130 mg/dL). However, this procedure would require access to the raw data. Binary outcomes are classified into two categories, such as event versus nonevent. Results are typically summarized as proportions (p), which is the number of participants, with the event divided by the number at risk for the event, whereas measures of treatment effect are typically given as a risk difference (RD), risk ratio (RR), or odds ratio (OR). RD constitutes an absolute measure whereasOR and RR are relative measures (seeChapters 5 and 6). ORs are often pooled to produce the summary effect size. For instance,a study by Wu et al. examined the relationship of milk consumption to the risk for age-related or vascular cognitive disorders and calculated ORs (95% CIs) for the highest level of milk intake compared with the lowest to measure the effect sizes [56]. By pooling results from six studies, no significant association was found between milk consumption and cognitive impairment (OR ¼ 0.76; 95%CI, 0.50e1.17). Another relative measure used is the RR, which more difficult to work with than the OR because it is not symmetrical and consistent with regard to the outcome definition. On the other hand, the RR is easier to interpretbecauseof the ratio change in average risk owing to exposure among the exposed

7.3 STATISTICAL METHODOLOGIES APPLIED IN META-ANALYSIS

173

population [57]. In a similar fashion, RRs extracted from the original studies are weighted and pooled with the appropriate meta-analytical methodology to produce the summary effect size. One would typically work with the event proportions from each intervention group within each study; however, statistical methods are available for working directly with the relative measures [58]. Such direct use of these measures also requires estimates of their variability (CIs or standard errors) within each study. Depending on the availability of the data and the willingness to assume consistency of the effect size across trials, one might conduct the analysis using more than one metric (e.g., ORs and RDs). Localio et al. proposed a method for converting from ORs to other metrics [59]. Count data are sometimes summarized as incidence rates, which is events per person-time. In this case, measures of the treatment effect will typically be either a rate difference or a rate ratio, analogous to RDs and RRs. Survival data may be analyzed as the time-to-event or as an OR if the duration of all studies is similar. Otherwise, the hazard ratio usually serves as the primary effect measure for group comparisons with time-to-event data. Often, aggregate data from the original studies can be obtained in the form of simple proportions, and timeperiods may differ across studies. Using IPD for time-to-event outcomes allows the use of a common method of analysis (consistent timeperiods) for each primary study before the meta-analysis. Therefore, calculating hazard ratios sometimes depends on the availability of individual subject-level data. If IPD are not available, a variety of transformations and approximations have been proposed to assist reviewers in dealing with other summary statistics that are often reported in published studies and in extracting information from survival curves [60].

7.3.2 CHOICE OF META-ANALYTICAL METHOD The summary effect produced through meta-analytical methods is essentially a weighted average of the individual effects. However, the mechanism used to assign the weights and therefore the meaning of the summary effect depends on the assumptions about the distribution of effect sizes from which the studies were sampled. Two basic approaches are used when pooling the effect sizes from a set of primary studies: fixedeffects and random-effects statistical methods. With the fixed-effect model, we assume that all studies in the analysis share the same true effect size, and the summary effect is our estimate of this common effect size. With the random-effects model, we assume that the true effect size varies from study to study, and hence the summary effect is our estimate of the mean of the distribution of effect sizes. The fixed-effects or common effects model starts with the assumption that the true effect size is the same in all studies. In other words, a fixed-effects approach implies that all factors that could influence the effect size are the same in all studies, and therefore the true effect size is the same (hence the term “fixed”). Given that all studies share the same true effect, it follows that the dispersion of effects sizes observed from one study to the next reflects the random error inherent in each study and the different study weights are assigned with the goal of minimizing this within-study error. By contrast, the random-effects model assumes that each study provides information about a different effect size, all of which should be represented in the summary estimate, and since the goal is to estimate the mean of the distribution, two sources of variance should be considered. First is the within-study variation in estimating the true effect in each study, and second is the variation in the true effects across studies; the different study weights are assigned with the goal of minimizing both of these sources of variance.

174

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

For both fixed- and random¼effects approaches, a common method is the inverse-variance weighted average that can be applied to diverse types of data (binary or continuous), as long as the point estimate (effect size from individual studies) and its variance can be obtained. Apart from inverse variance, there are two widely used methods of fixed-effects meta-analysis for dichotomous outcomes (the ManteleHaenszel and Peto methods) and one random-effects method (DerSimonian and Laird). By viewing meta-analysis from the perspective of weighing the individual study results, in the 1 , and the weighted fixed-effect model the weight is the inverse of the within-variance estimate w ¼ Vi k P mean (M) is then computed as M ¼

W i Yi

i¼1 k

P

. In the random-effects model, it is the inverse of the Wi

i¼1

1 , where v is the within-study variance for study i within- and the between-variance estimate w ¼ ViþT 2 i

and T2 is the between-study variance. One method for estimating the T2 is the DerSimonian and Laird  2 k P Wi Y i k P i¼1 2
method: T2 ¼ Q
df , where Q ¼ W Y , df ¼ k
1 (where k is the number of the i i k C P i¼1

Wi

i¼1 P 2 P Wi P studies) and C ¼ Wi
. The variance in the summary effect is estimated as the reciprocal of

Wi

1 , and the estimated standard error of the summary effects is then the sum of the weights: VM ¼ P k Wi

pffiffiffiffiffiffiffi the square root of the variance: SEM ¼ VM ; then, 95% lower and upper limits for the summary effect are estimated as LLM ¼ M
1:96  SEM and ULM ¼ M þ 1:96  SEM . Study weights are more balanced under the random-effects model than under the fixed-effect model. Large studies are assigned less relative weight and small studies are assigned more relative weight compared with the fixed-effect model. In general, the random-effects model will result in a larger standard error for the estimate of the overall effect estimate and hence wider confidence intervals and harder-to-reach statistical significance; it is considered a more conservative approach. The choice of model should be informed by the inference that is to be drawn by the data. If the studies are functionally identical and the goal is to compute the common effect size for the identified studies (the defined population), a fixed-effects model is a plausible fit. By contrast, if the researcher is convinced that there are factors with the potential of introducing variations from study to study owing to the nature of the studies that produced the data (for example,owing to differences in participants or the implementation of interventions), the random-effects model is more easily justified than the fixedeffect one [61e63]. In the presence of high variability between studies, one should consider the appropriateness of greater equality of weights for large and small studies as a consequence of a random-effects model, because the greater equality of weights may be undesirable. In case there is no between-study variability (T2 ¼ 0), the estimates of the two models will coincide. However, the choice of the model should be defined apriori, in the context of prespecified goals and not starting with a fixed-effect model and then moving to a random-effects model if the test for heterogeneity is significant. i¼1

7.3 STATISTICAL METHODOLOGIES APPLIED IN META-ANALYSIS

175

An IPD meta-analysis can be performed using a one- or two-stage approach [64]. In general terms, a one-way approach requires all participant data to be modeled together and main effects to be estimated simultaneously. A two-stage approach implies that individual studies are initially analyzed on an equal basis but each one separately, and at the second stage the summary statistics of each of these individual studies are combined to provide a pooled estimate of effect in the same way as for a typical systematic review [65].

7.3.3 STATISTICAL HETEROGENEITY A meta-analysis attempts to broaden research questions by including available evidence from a set of primary studies. Because of this broader base, the combined studies may encompass study-specific characteristics such as a variety of treatment regimens, types of patients, and outcomes, resulting in differences among individual estimates. The scope of a meta-analysis is not simply to compute a summary effect, but rather to consider the pattern of effects. Regardless of whether a fixed-effects or random-effects meta-analysis is conducted, it is critical for it to be accompanied by an evaluation of the consistency of effects across studies. Any kind of variability among studies in a systematic review may be termed heterogeneity. In its broadest sense, heterogeneity refers to differences between studies and/or study results. Heterogeneity can generally be classified into three types: clinical, methodological, and statistical [66]. Clinical heterogeneity refers to differences between trials or observational studies in their participant selection, interventions, and outcomes. Methodological heterogeneity refers to differences in study design and conduct. Statistical heterogeneity represents the notion that individual studies may have results that may be numerically inconsistent with each other and that the variation is more than what is expected on the basis of sampling variability (chance) alone. Statistical heterogeneity may result from clinical or methodological heterogeneity or it may be due to chance. A common way to test for between-study heterogeneity is to use Cochran’s Q statistic (Cochran’s   k P c2 test), which is defined as Q ¼ Wi ðYi
MÞ2 , where Wi is the study weight V1i , Yi is the study i¼1

effect size, and M is the summary effect. Special attention should be given to interpreting this chi-square testbecause it has low power in a situation in which studies in the meta-analysis have a small sample size or are few in number. As such, a significance level of 0.10 is used to infer whether there is substantial statistical heterogeneity, instead of 0.05. A useful descriptive measure that quantifies heterogeneity is the inconsistency index (I2), proposed by Higgins et al. I2 ¼ 100  ðQ
kþ1Þ , where Q is Cochran’s Q statistic and k is the number of studies Q [5,67,68]. I2 is interpreted to be the percentage of total variation in the treatment effect between studies owing to heterogeneity. As a relative measure, it is subject to some potential idiosyncrasies. For example, in a situation in which there are very large studies, average within-study variability will likely be small, which implies that even a small degree of between-study variability in effect sizes could produce a large value of I2. Thus, the degree of heterogeneity should be interpreted within the context of substantive clinical implications. The choice of fixed-effects or random-effects model should not depend on heterogeneity or a lack of it. Nor does the use of a random-effects model obviate the need for an analysis of potential sources

176

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

of heterogeneity. Strategies for exploring heterogeneity in meta-analyses using aggregate data usually includes subgroup analyses, meta-regression, or sensitivity analysis. In case individual participant data are available, a more flexible type of analysis can be done that allows patient-level heterogeneity in the treatment effect to be separated from trial-level heterogeneity and investigated directly [64].

7.3.3.1 Subgroup Analysis Subgroup analyses involve separating studies into subgroups so that they can be compared. Subgroups may be composed of particular subsets of studies that share particular characteristics postulated to alter the treatment effect. Such examples may be different geographical locations, study durations, or doses. A meta-analysis is then carried out based on the subset of studies and the results can be used to compare mean effect sizes between the different subgroups and contrast them with the meta-analytical results produced from the entire collection of studies [69]. Although comparisons between subgroups are made informally, several methods could be used to compare the different subgroups. Some of these amount to study-level versions of z-tests for two subgroups, and analysis of variance if there are more than two subgroups for comparison [70]. However, the reviewer should keep in mind that observed differences in either the effect size or statistical heterogeneity in one subgroup versus the other does not indicate that the subgroup factor explains heterogeneity. Because different subgroups are likely to contain different amounts of information, and thus have different abilities to detect effects, it is misleading simply to compare the statistical significance of the results [71].

7.3.3.2 Meta-Regression Meta-regression is a form of regression analysis in which the unit of analysis is a study, not individual data within a single study. It is an extension of subgroup analyses that allows the effect of continuous as well as categorical characteristics to be investigated; in principle, it allows the effects of multiple factors to be investigated simultaneously [72]. In a meta-regression, the outcome is the effect estimate (for example, an MD, an RD, a log OR, or a log RR) and the explanatory variables include characteristics that might influence the size of intervention effect, such as the population characteristics, factors describing the study setting, or the intervention protocol. Because meta-regression uses study-level summary statistics as response and explanatory variables, it is important to weight each study in the regression by selecting the appropriate model(fixed-effects or random-effects), although most of the time a random-effects approach is used. Several methods for meta-regression are available; however, if there are few studies even if there are many patients, meta-regression is unlikely to be scientifically useful. Commonly, many characteristics may explain heterogeneity and multiple post hoc analyses lead to data dredging and increase the possibility of false-positive findings [73e75]. Therefore, it is not recommended to use meta-regression, especially with multiple explanatory variables, when the number of studies is relatively small (empirically when there are fewer than 10 studies for each explanatory variable in a meta-analysis).

7.3.3.3 Sensitivity Analysis Sensitivity analysis is a way to investigate the robustness of meta-analysis results according to a variety of decisions and assumptions. The investigator should identifyidiosyncrasies of individual studies that may affect the summary estimate as early as possible in the review process. Assumptions for sensitivity

7.3 STATISTICAL METHODOLOGIES APPLIED IN META-ANALYSIS

177

analysis may be the exclusion of studies with a high risk for bias, separating studies with different effect measures, or studies with missing information. The meta-analysis is then carried out twice, once including all studies and once excluding studies with identified peculiarities. A common approach is to“leaveoneout,”in which the summary effect is calculated multiple times,excluding a single study each time. The summary effects are then visually examined to determine the relative influence of each study in the meta-analytical results. When sensitivity analyses show that the overall results are unaffected by the different decisions that could be made during the review process, the results of the review can be regarded with a higher degree of certainty. Where sensitivity analyses identify key points that greatly influence the findings of the review, more resources can be deployed to attempt to resolve uncertainties by obtaining extra information. Generally, sensitivity analyses are concerned with the robustness of the primary results from methodological decisions, whereas subgroup analyses are concerned with exploring the treatment effect across specific study characteristics.

7.3.3.4 Prediction Intervals Often, prediction intervals (PIs) are reported to describe the distribution of true effect sizes regarding the summary effect for random-effects analyses. PIs may be defined as the interval within which a new trial’s estimate would fall if participants were selected at random from the same population [76]. Whereas the CI quantifies the accuracy of the mean, the PI addresses the actual dispersion of effect sizes and the two measures are not interchangeable. A PI is centered on the summary estimate; its width accounts for the uncertainty of the summary estimate, the estimate of between-study standard deviation in the true treatment effects, and the uncertainty in the between-study standard deviation estimate itself [77]. Formulas have been proposed to compute PI for meta-analysis [77,78]. Calculation is possible when at least three primary studies are available. A PIis most suitable when studies included in the meta-analysis have a low risk for bias. Otherwise, it will encompass heterogeneity in treatment effects caused by these biases, in addition to that caused by true clinical differences. The inclusion of a PI, which indicates the possible treatment effect in an individual setting, can often make interpretation of meta-analytical results more useful in clinical practice and decision making [79]. If we look at the study by Senftleber et al., for instance, which examined the effect of marine oil supplementation on arthritis pain, we may notice that the overall effect estimate corresponded to an SMD of
0.24 (95%CI
0.42 to
0.07); however, substantial heterogeneity was found among the primary studies (I2 ¼ 63%), resulting in much wider prediction intervals (
1.05e0.57) [80]. The interval overlaps zero, indicating that in some settings the treatment may actually be ineffective, a finding not apparent when only the summary effect and its CI are evaluated.

7.3.4 SMALL-STUDY EFFECTS In several meta-analyses, small studies systematically present with different effects compared with large studies, a phenomenon often called small-study effect. Small-study effects may occur for a number of reasons such as publication or other selection biases, poor methodological quality, true differences, or chance [81]. Because published studies are more likely to find their way into a metaanalysis, any bias in the literature is likely to be reflected in the meta-analysis as well, an issue

178

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

generally known as publication bias (see Section 7.5). Publication bias represents a common issue in meta-analysis; hence, a series of methods have been developed to assess the likely impact of bias in any given meta-analysis. To measure the impact of publication bias, several assumptions are made, such as that larger studies (which have less sampling error variation) are more likely to be published regardless of statistical significance because they involve large commitments of time and resources. Funnel plots can be visually inspected to check for asymmetry in the distribution of study results in a meta-analysis that may indicate publication bias. A funnel plot is usually plotted with the effect size on the X-axis and the standard error on the Y-axis. Because larger studies have less sampling error variation in effect sizes, they appear toward the top of the graph and cluster around the mean effect size, whereas smaller studies appear toward the bottom (see Section 7.4.3). This pattern resembles a funnel; hence, the plot’s name. In the absence of publication bias, studies will be distributed symmetrically about the mean effect size because sampling error is random. In the presence of publication bias, studies are expected to follow the model, with symmetry at the top, a few studies missing in the middle, and more studies missing near the bottom. Because the interpretation of a funnel plot could be subjective, several statistical tests have been proposed to quantify or test the relationship between the sample size and effect size [81e84]. The most common test is Egger’s test, but when it is applied to binary outcome data, it can give false-positive results in some situations, such as for large treatment effects, when there are few events per trial, or when all trials have similar sizes. Hence, for dichotomous outcomes, Harbord’s test or the arcsineThomson test often can be more appropriate [83,85]. However, the funnel plot should be seen as a generic way to display small-study effects because there is a tendency for intervention effects estimated in smaller studies to differ from those estimated in larger studies [86].

7.3.5 SOFTWARE FOR META-ANALYSIS Various statistical programs are available to perform meta-analyses. The most popular are REVMAN and R, which are freeware; STATA, SAS, and Comprehensive Meta-analysis are commercial.

7.4 PRESENTATION AND INTERPRETATION OF RESULTS A description of the results of a meta-analysis is visually enhanced with tables and figures. In this section, the main components of the results of a meta-analysis are discussed and examples are provided. Guidelines are available for the proper scientific reporting of a meta-analysis. These guidelines include “Meta-analysis of Observational Studies in Epidemiology: A Proposal for Reporting,” which was published in 2000 [87], and PRISMA, which was published in 2009 [20]. They provide checklists of recommendations for reporting meta-analyses of observational studies and RCTs, respectively. The checklists include recommendations regarding the background, search strategy, inclusion/exclusion criteria, methods, results, discussion, and conclusions, aspects useful to consider when planning a meta-analysis.

7.4 PRESENTATION AND INTERPRETATION OF RESULTS

179

7.4.1 STUDY SELECTION The flow of information through the different phases of a systematic review, such as the time that the search lasts; the databases; other methods used for study identification; the exact number of studies found, along with the number of duplicates retained through various sources; the number of studies excluded during the title and abstract screening process; and the number of studies excluded, along with the reasons for exclusion during the full text screening process, all should be recorded during the systematic review process and can be synopsized and presented in a flowchart, as proposed by the PRISMAstatement [88]. For example, in the flowchart in the study by Mocellin et al. (Fig. 7.2), [89] we notice that a database search yielded a total of 1011 unique records, and after title and abstract screening 22 full-text articles were assessed for eligibility, nine of which were finally included in the analysis.

7.4.2 STUDY CHARACTERISTICS Basic characteristics of the primary studies included in the analysis should be provided and are usually summarized in tabular format. These may include sample sizes and demographics by treatment group, intervention characteristics such as dosing and time to follow-up, outcome measures, or other factors specific to the research question being addressed.

7.4.3 FOREST PLOT The graphical display of data from meta-analyses is typically made using forest plots, in which each study is shown with its effect size and the corresponding 95% CI[90]. Forest plots are useful tools for showing the main features of meta-analysis results and comprehensively visualizing the direction and magnitude of the overall effect as well as the results and potential heterogeneity of individual studies. For instance, by looking at the forest plot in the study by Dewansingh et al (Fig. 7.3), which examined the effects of vitamin D supplementation on leg strength in older adults [91], we notice that seven studies were included. Each line and the square in the center represent the results from one study. The size of the square is proportional to the weight that each study had in the meta-analysis. The larger the square, the more weight the study has. The two horizontal lines around the square show the CI for each estimate of the study. At the bottom of the plot is a diamond depicting where the meta-analytical effect lies and where the point estimate is, along with the CI for the meta-analysis. Depending on the measure of association used, axis XX0 may have a different scale. In this example, an SMD was used because of the different methods used to measure the outcome (i.e., the leg strength). A difference of zero, denoted by the vertical line, corresponds to the null effect. The forest plot is labeled such that if the diamond lies to the right of the vertical line of no effect (as in the example demonstrated in Fig. 7.3), it favors the intervention arm (i.e., vitamin D supplementation). As the authors specifically stated: Six trials that investigated vitamin D supplementation were included in the meta-analyses which showed no significant effect of vitamin D on leg strength (n ¼ 735; SMD 0.09; 95% CI,
0.05 to 0.24; I2 ¼ 0%; P ¼ 0.22.

180

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

FIGURE 7.2 Flowchart of the literature search, screening, and selection process for controlled clinical trials. PUFA, polyunsaturated fatty acid. Adopted from Mocellin et al.

Fixed-effects meta-analysis for the effects of vitamin D supplementation on leg strength. “Total” denotes the cumulative sample size from all included studies. CI, confidence interval; IV, inverse variance; SD, standard deviation; Std., standard. Adopted from Dewansingh et al.

7.4 PRESENTATION AND INTERPRETATION OF RESULTS

FIGURE 7.3

181

182

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

7.4.4 ASSESSING HETEROGENEITY It is important in meta-analyses to investigate potential sources of heterogeneity (see Sections 7.3.3 and 7.5). Strategies for investigating heterogeneity typically include subgroup analysis or metaregression, which are usually applied when the effect of specific study characteristics on the overall estimates are examined, or sensitivity analysis, which is usually applied when the effect of individual studies on the overall estimates is investigated.

7.4.4.1 Subgroup Analysis Results of the subgroup analyses are usually summarized in a table, although it is common to present them in a single forest plot along with the overall effect size for better visualization. An example is provided from the study by Langlois et al., which investigated the efficacy of n-3 PUFA supplementation, alone or as combination therapy, compared with placebo on clinical outcomes in adult patients undergoing cardiac surgery [92]. To pool the ORs, a fixed-effects meta-analysis was performed using the method of ManteleHaenszel. An overall estimate of 0.76 (95%CI, 0.66e0.88) in the forest plot indicated a 24% decreased risk for postoperative atrial fibrillation in the group that received n-3 PUFA supplements enterally or parenterally (P < .001). Furthermore, subgroup analyses were conducted according to the route of n-3 administration: oral/enteral versus parenteral. In the forest plot presented in Fig. 7.4, the null effect equals 1 because the meta-analysis is based on ORs (a proportion-based metric) as measures of effects.

7.4.4.2 Meta-Regression Meta-regression results are typically presented in tabular format when more than one study characteristic is examined. When there is only one characteristic, a regression plot may be used. In the study by Serban et al., which aimed to assess the effect of spirulina supplementation on plasma lipid concentrations, meta-regression was used to explore the association between plasma lipid concentrations (total cholesterol, high-density lipoprotein-C, LDL-C, and triglycerides) and spirulina dose [93]. Fig. 7.5 represents the meta-regression plot, in which the solid line represents the linear prediction for the effect size as the function of an increase in the dose of the supplement in grams per day, circles represent the effect size of each study, and the size is inversely proportional to the variance. As the authors specifically stated: The impact of spirulina on plasma concentrations of triglycerides was independent of administered dose (slope ¼
1.39, 95%CI:4.26, 1.48, P ¼ 0.342).

7.4.4.3 Sensitivity Analysis Sensitivity analysis involves repeating the meta-analysis more than once, excluding studies that are likely to affect the results, such as studies with a high risk for bias or those with an unclear design. Often, meta-analysis is performed multiple times, each time excluding a single study, and results are plotted together to provide a visual presentation of the effect. For instance, in the study by Xiao et al. [94], which examined the effect of nut consumption on vascular endothelial function, defined as flow mediated dilation, the overall WMD was 0.41% (95%CI, 0.18e0.63%; P ¼ .001). The authors performed a sensitivity analysis and the results are presented in Fig. 7.6. Each small circle represents the meta-analysis result achieved by omitting one study (named

Fixed-effects meta-analysis for the effects of omega-3 polyunsaturated fatty acids on postoperative atrial fibrillation, including subgroup analysis for oral/enteral versus parenteral supplementation. CI, confidence interval; df, degrees of freedom; M-H, ManteleHaenszel.

183

Adopted from Langlois et al.

7.4 PRESENTATION AND INTERPRETATION OF RESULTS

FIGURE 7.4

184

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

FIGURE 7.5 Meta-regression plot of the association between mean changes in plasma triglyceride (TG) concentrations and administered spirulina dose (g/day). The size of each circle is inversely proportional to the variance in change. Adopted from Serban et al.

study on the left) each time; dotted lines that extend from the circle represent the 95% CI. Given the sensitivity analysis plot presented in Fig. 7.6, no study seemed to influence the summary effect size substantially.

7.4.5 RISK FOR BIAS Another crucial feature of a meta-analysis is the quality assessment of the included primary studies. Several tools exist to support the assessment (see Section 7.2.3); hence, the graphical display of the results is variable. A simple summary is often incorporated into the table with the general study characteristics and presented in detail in a table or graph. A frequently used tool to assess the risk for bias in clinical trials is the risk for bias tool of Cochrane, which was also used in the study by Ding et al. [95]. Results of this assessment are presented in Fig. 7.7. The risk for bias regarding six different components was assessed in each study. For example, in Fig. 7.7 the study by "Garg, 1990" included in the meta-analysis was judged to have unclear risk for bias in terms of random sequence generation and blinding of participants/ personnel, high risk for bias in terms of allocation concealment and blinding of outcome assessors, and low risk for bias in terms of incomplete outcome data and selective reporting.

7.4.6 FUNNEL PLOT As discussed in the previous section, funnel plots can be used to check for asymmetry in the distribution of study results in a meta-analysis, which may indicate publication bias [81]. Large studies appear toward the top of the graph and generally cluster around the mean effect size. Smaller studies appear toward the bottom of the graph, and (because smaller studies have more sampling error variation in effect sizes) tend to be spread across a broad range of values. This pattern resembles a funnel; hence, the plot’s name [96].

7.4 PRESENTATION AND INTERPRETATION OF RESULTS

185

FIGURE 7.6 Sensitivity analysis with the exclusion of one study at a time from the meta-analyses. Estimated effect size of weighted mean differences of flow-mediated dilation is plotted on the horizontal axis and studies (first author, year) on the vertical axis.CI, confidence interval. Adopted from Xiao et al.

Fig. 7.8 shows an example of the funnel plot presented in the study by Li et al. [97]. The plot shows that studies are evenly distributed in the funnel. This indicates the absence of publication bias in the analysis, which is confirmed by a statistical analysis, Egger’s test. It was specifically stated by the researchers that: Begg’s funnel plot and Egger’s test (P ¼ 0.795) showed no evidence of significant publication bias between fruit intake and glioma risk.

Another example is presented in Fig. 7.9, which indicates asymmetry, in contrast to the previous example [127]. As the authors specifically stated: “Publication bias was observed from the funnel plot and Egger’s test (P ¼ .005).” The reader should keep in mind, however,that funnel plot asymmetry should not be equated with publication bias, because it may have a number of other possible causes [98].

186

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

FIGURE 7.7 Risk for bias assessment of included reports in the meta-analysis studies, using a risk for bias tool proposed by Cochrane. Studies are presented vertically on the left; each component of the tool is shown horizontally at the top.
, highrisk; þ,lowrisk; ?, unclear risk. Adopted from Ding et al.

7.5 LIMITATIONS AND BIASES

187

FIGURE 7.8 Funnel plot for the analysis of vegetable intake and glioma risk. “X” axis represents the hazard ratios (HR) of the primary studies in log scale; “Y” axis represents the standard error (SE) of the logarithm of the hazard ratios (loghr). Adopted from Li et al.

FIGURE 7.9 Funnel plot for the analysis of tea consumption and lung cancer risk. “Y” axis represents the risk ratios of the primary studies in log scale;“X” axis represents the standard error (SE) of the logarithm of the risk ratios (log[RR]). Adopted from Wang et al.

7.5 LIMITATIONS AND BIASES In a broader sense, bias can be introduced in the meta-analysis through biases that are in individual studies included in the review and the meta-analytical procedure itself. Bias in the meta-analysis or meta-bias may stem from various stages of the review process. During study selection, reporting biases may occur most often in the form of publication bias and selective outcome reporting. Publication rates have been diverse among disciplines, but overall, a significant

188

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

proportion of studies remain unpublished and as research has shown, studies with positive results are more likely to be published than are those with negative results [99,100]. When studies are missing, this may threaten the validity of meta-analysis. Another important issue that has been a subject of investigation is selective outcome reporting, which is of great concern because it often manifests as the selective reporting of certain outcomes (including harms) in a trial but not others, depending on the nature and direction of the results [101]. It may be challenging to unveil such a slide in what was initially set in the protocol as a trial’s primary outcome because one may have to search deep in the study protocol registries. Generally, studies that are easiest to find might differ in their results from thosethat are more difficult to find, and searching only one or two easily accessible databases may drive the reviewer toward what is often referred to as ascertainment bias. A comprehensive and thorough literature search is imperative to avoid such selection biases. Another critical issue arises when setting inclusion criteria for a specific research topic. Particular differences in the way in which inclusion criteria are set may have a direct impact on studies that are to be included in the analysis, which might indirectly affect the meta-analytical results. Thus, all specifications given to the inclusion criteria should be in the context of the research question and any restrictions that are set should be justified accordingly. Another source of bias in the meta-analysis can be found in the extraction process. Information biases may occur as a result of errors during the abstraction of primary studies, or even selective reporting. Research has shown that when data extraction is carried out by two reviewers independently, the probability of obtaining such extraction errors is minimized irrespective of the reviewers’ experience [102,103]. The review protocol should be reported in a transparent way so that the reader can evaluate whether predefined analytical procedures were used in the analysis. It is critical to try to minimize selection and information or analytical bias in the study. However, potential limitations that derive from the individual studies should be discussed. Variations in individual study characteristics might increase uncertainty about the results of the meta-analysis and often introduce heterogeneity in the overall estimates. Nevertheless, the goal of a meta-analysis is only rarely to synthesize data from a set of identical studies. Almost invariably, the goals are to broaden the base of studies to a certain extent, expand the question, and study the pattern of answers. The questions of whether it makes sense to perform a meta-analysis and what studies to include must be asked and answered in the context of specific goals. However, when there are differences in participant demographics and study methods, combining studies increases variability in findings and makes it more difficult to identify real effects. From a statistical perspective, there is no restriction on the similarity of studies based on the types of participants, interventions, or exposures. However, for the analysis to be meaningful, attention should be paid to consider the diversity of studies carefully in these respects. An important parameter that should be thoroughly investigated is the so-called study quality, which actually refers to the methods used to conduct the study. Emphasis should be given to both internal and external validity along with the relevance to the review question. Several tools exist, with type or discipline-specific orientation, to guide reviewers through the quality assessment process (see Section 7.2.3). Primary studies of limited quality can be a source of heterogeneity as revealed by sensitivity analysis. Heterogeneity can be often explained by variations in study protocols that can be described as clinical or methodological heterogeneity. Differences in the way in which certain definitions are set may result in variability in the final estimates. For instance, studies in whichbody mass index(BMI) is

7.5 LIMITATIONS AND BIASES

189

used instead of percent body fat to specify overweight individuals could lead to a misclassification of trained individuals with a high BMI and a different metabolic profile as being overweight. Alterations in the protocol for dietary intakes may also lead to heterogeneous estimates. For example, if a low-carbohydrate diet is tested in a meta-analysis, a carbohydrate intake that ranges from 5% to 20% of daily energy from carbohydrates may directly lead to high heterogeneity and affect the summarized results. In a higher versus lower quartile of intake approach to examine the effect of a specific food group on a defined condition, alterations in the effect estimate may be expected when there are significant differences in the range of intakes between studies. For example, in a study in which n-3 interventions are examined, fish intake may have been considered in only a number of trials. Furthermore, several factors may indirectly affect the bioavailability of food ingredients and hence introduce variability to the effect size. Such factors may be the different cooking methods used, the quality of raw food products, and nondietary or dietary factors that may affect bioavailability, some of which may have been allowed for in a number of trials during the study design. Another factor that may introduce variability in the analysis is the different intervention periods and different times of follow-up in the trials. Some studies show a higher effect in the short term, which could be attributed to better short-term dietary compliance or a metabolic response. There is also high variability in the certainty with which specific dietary components that could be implicated in the research question are reported. For instance, coffee consumption can be reliably recalled in terms of quantity and frequency because it is a regular habit for many individuals, whereas salt consumption may be less reliably recalled. With sufficient data, diversity across studies can be explored through subgroup or sensitivity analyses. If the effect is similar among groups, the reviewers can report that the effect is robust. However, for many meta-analyses, there is insufficient power to achieve this reliably and there may also be problems of confounding.

7.5.1 CHALLENGES Systematic reviews and meta-analysis are valuable for summarizing existing evidence; moreover, they may be undertaken to inform specific decisions, often in the context of evidence-based policy making. They have a key role in informing recommendations on nutrition-related aspects to prevent disease and develop guidelines when internal and external validity issues are considered. Furthermore, they can be used when designing of novel research or contextualizing a primary study. Meta-analyses have been used by government agencies such as the European Food Safety Authority as a tool to inform the substantiation of nutritional or health claims. They provide an overview regarding the reproducibility and consistency of the effect across studies and study groups, about the doseeresponse relationship, and about the minimum effective dose of a food or food constituent required to obtain the claimed effect. As such, meta-analytical results have been used by the authorities’ research panel in published documents, to summarize overall evidence provided by individual human studies, and to establish conditions of use [104]. The World Cancer Research Fund Continuous Update Program collates the latest available evidence on the effect of nutrition on different cancers and uses meta-analyses as a tool to support recommendations. Systematic reviews and meta-analysis have been used to explore issues beyond intervention effectiveness and address a range of research questions in areas such as clinical testing and diagnostic accuracy, health economics, population screening, health equity, epidemiology, service quality and delivery, risk factor exposure, and government policies [105]. The WHO and several other parties

190

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

around the globe, such as departments of health, have established the use of systematic reviews and meta-analyses as a format to appraise and summarize evidence obtained though experimental and observational approaches as a tool to inform policies on public health promotion. Systematic reviews and meta-analyses can be used during the design process of new research projects. At a study’sconcept, they can determine whether the study is necessary in terms of covering a knowledge gap. It may be possible to obtain the required answer to a research question by synthesizing data from prior studies; in that case, further research may not be performed [106]. If a new study is needed, the meta-analysis may be useful in designing that study. For instance, a meta-analysis may show that in prior studies, particular study characteristics had been proven to produce results that were different from those of others or that a specific pattern of interventions had proven to be more effective than others, which may be used as a direction. Systematic reviews can also have a role in the publication of any new primary study, providing the context of what is already known and what it is anticipated to be learned from the new study. In the discussion section of the publication, a systematic review allows readers to address not only information provided by the new study but also the body of evidence as enhanced by the new study. On the other hand, meta-analyses do not consistently refer to and discuss findings of previously published meta-analyses on the same topic. Such neglect can lead to research waste and be confusing for readers [107]. Systematic reviews and meta-analyses that aim to identify the use of resources, health costs, and costs relative to benefits of alternative health interventions can guide policy makers toward committing to decisions that promote the efficient use of resources and best health outcomes [108]. However, the issue is under debate owingto the substantial diversity often observed among studies that come from different countries [109]. In the era of personalized medicine, a type of research methodology that has been gaining popularity is the development of prediction models for disease risk in population groups, thereby informing clinical diagnosis and prognosis. Such models are based on equations that are used to estimate an individual’s risk based on values of multiple predictors [110]. Known examples include the Framingham Coronary Heart Disease Risk Score [111] and QRISK2 [112]. In such models, a fundamental component is the external validation process, in which meta-analyses using IPD from multiple studies that involve a large number of patients from multiple practices, hospitals, or countries can have a crucial role [113]. A load of primary studies from multiple research fields are generated yearly, accompanied by a general increase in the production of systematic reviews and meta-analysis in the need to cover the totality of evidence. Furthermore, the ongoing shift toward implementing so-called individualized therapy has increased the number of available parameters that need to be incorporated in the review process in an often-limited timeline. As a result, the systematic review and meta-analysis process is constantly evolving to meet with challenges. Under time and resource constraints where urgent or emerging health-related issues need to be addressed by decisionmakers the concept of rapid reviews was introduced. Rapid reviews are a form of knowledge synthesis that can be characterized as an accelerated systematic review, in which components of the review process are simplified or omitted to provide information in a timely manner; however, they sacrifice some methodological rigor [114]. Meta-analysis methodologies are being developed for a complex evidence synthesis involving models incorporating evidence on multiple parameters that may model data from different study

REFERENCES

191

designs. An example is “field-wide meta-analyses,” which aim to assess an entire field covering all putative risk factors for a research question based on observational data [115]. With the high efflux of a number of primary studies in the literature, it can be challenging to keep up-to-date on a specific review question. As a result, in 2014, an approach was proposed to review updating, termed “living systematic reviews” (LSR). An LSR is a systematic review that is continually updated, incorporating relevant new evidence as it becomes available [116]. All forms of updating commonly require an updating of meta-analyses even though the potential for naively repeated metaanalysis might lead to an inflated rate of false-positive findings [117]. Another approach that has been become popular in the field of meta-analysis is the mixed treatment comparison, also known as “network meta-analysis” [118]. Network meta-analysis extends the standard pairwise meta-analysis framework to allow a simultaneous estimation of the comparative effectiveness of multiple interventions using an evidence base of trials that individually may not compare all intervention options, but form a connected network of comparisons [119]. A network meta-analysis hence combines direct and (when available) indirect evidence from multiple studies, through complex methodologies facilitating indirect comparisons of multiple interventions that have not been studied in head-to-head studies. This is considered an important advance, given that network meta-analyses may offer an overview of the entire set of a clinical condition including available treatments and comparisons of them, providing an evidence-based estimation of risks and benefits for each therapeutic option and guiding the conduct of new research. On the other hand, with the increase in the number of systematic reviews and meta-analyses available, a logical next step to provide all interested parties in healthcare with the evidence they require maybe toconduct reviews of existing systematic reviews [120]. Syntheses of existing systematic reviews and meta-analysis may be found using the term “umbrella review” or “overview of systematic reviews” [121e123]. A checklist was published with preferred reporting items in such studies [124]. Generally, an umbrella review allows the findings of systematic reviews and metaanalysis relevant to a review question to be compared and contrasted. Quantitative reviews of published epidemiologic studies of exposureeresponse relations typically include an assessment of the relation between exposure levels and risk for disease. Conclusions about differences in the effect owing to differences in dose can be examined through a “doseeresponse metaanalysis.” Although it is not new as a concept, it has been become popular and new techniques are becoming available in statistical software [125,126]. In addition, methodologies are being developed to cover the issue of heterogeneity and confront issues within a meta-analysis or potential bias, such as publication bias. Advances in the field of Bayesian methodologies often provide a basis for that direction.

REFERENCES [1] Chalmers I, Hedges LV, Cooper H. A brief history of research synthesis. Eval Health Prof 2002;25(1): 12e37. [2] Rayleigh TRHL. Presidential address at the 54th meeting of the British association for the Advancement of science, Montreal. London: John Murray; August/September 1884. [3] Pearson K. Report on certain enteric fever inoculation statistics. Br Med J 1904;2(2288):1243e6. [4] Fisher RA. Statistical methods for research workers. 4th ed. London: Oliver and Boyd; 1932. [5] Cochran WG. The combination of estimates from different experiments. Biometrics 1954;10(1):101e29.

192

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

[6] Mantel N, Haenszel W. Statistical aspects of the analysis of data from retrospective studies of disease. J Natl Cancer Inst 1959;22(4):719e48. [7] Glass GV. Primary, secondary, and meta-analysis of research. Educ Res 1976;5(10):3e8. [8] H MF. Experiences with meta-analysis in NDA submissions. In: Proceedings of the biopharmaceutical section of the American Statistical Association; 1988. p. 5 (2). [9] Haidich AB. Meta-analysis in medical research. Hippokratia 2010;14(Suppl 1):29e37. [10] Egger M, Smith GD, O’Rourke K. Introduction: rationale, potentials, and Promise of systematic reviews. In: Systematic reviews in health care. BMJ Publishing Group; 2008. p. 1e19. [11] Patsopoulos NA, Analatos AA, Ioannidis JP. Relative citation impact of various study designs in the health sciences. Jama 2005;293(19):2362e6. [12] Feinstein AR. Meta-analysis: statistical alchemy for the 21st century. J Clin Epidemiol 1995;48(1):71e9. [13] Kavale KA, Forness SR. Hyperactivity and diet treatment. J Learn Disabil 1983;16(6):324e30. [14] Axelson ML, Federline TL, Brinberg D. A meta-analysis of food- and nutrition-related research. J Nutr Ed 1985;17(2):51e4. [15] Detsky AS, et al. Perioperative parenteral nutrition: a meta-analysis. Ann Intern Med 1987;107(2): 195e203. [16] Cochrane. Available from:www.cochrane.org. [17] Campbell Collaboration. Available from:https://www.campbellcollaboration.org/. [18] Stewart LA, Parmar MK. Meta-analysis of the literature or of individual patient data: is there a difference? Lancet 1993;341(8842):418e22. [19] PROSPERO. International prospective register of systematic reviews. Available from:https://www.crd. york.ac.uk/prospero/. [20] Moher D, et al. Preferred reporting items for systematic reviews and meta-analyses: the PRISMA statement. J Clin Epidemiol 2009;62(10):1006e12. [21] Dibaba DT, et al. The effect of magnesium supplementation on blood pressure in individuals with insulin resistance, prediabetes, or noncommunicable chronic diseases: a meta-analysis of randomized controlled trials. Am J Clin Nutr 2017;106(3):921e9. [22] Borenstein M, et al. When does it make sense to perform a meta-analysis?, in introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 357e64. [23] Counsell C. Formulating questions and locating primary studies for inclusion in systematic reviews. Ann Intern Med 1997;127(5):380e7. [24] Liao CD, et al. Effects of protein supplementation combined with resistance exercise on body composition and physical function in older adults: a systematic review and meta-analysis. Am J Clin Nutr 2017. [25] Roberts J, Dicenso A. Identifying the best research design to fit the question. Part 1: quantitative designs. Evid Base Nurs 1999;2(1):4. [26] Lefebvre C, Clarke MJ. Identifyingrandomised trials. In: Systematic reviews in health care. BMJ Publishing Group; 2008. p. 69e86. [27] Marcano Belisario JS, et al. Search strategies to identify observational studies in MEDLINE and EMBASE. Cochrane Database Syst Rev 2013;(12). [28] Royle P, Bain L, Waugh N. Systematic reviews of epidemiology in diabetes: finding the evidence. BMC Med Res Meth 2005;5:2. [29] Hill B. Comparison of journal title coverage between CINAHL and Scopus. J Med Libr Assoc 2009;97(4): 313e4. [30] AGRIS. Available from: http://agris.fao.org/agris-search/home. [31] CABI. Available from: https://www.cabi.org/. [32] AGRICOLA. Available from: https://www.ebsco.com/products/research-databases/agricola. [33] FSTA. Available from: https://www.ebsco.com/products/research-databases/fsta.

REFERENCES

193

[34] Whitehead A, et al. Cholesterol-lowering effects of oat beta-glucan: a meta-analysis of randomized controlled trials. Am J Clin Nutr 2014;100(6):1413e21. [35] Hopewell S, et al. Handsearching versus electronic searching to identify reports of randomized trials. Cochrane Database Syst Rev 2007;(2). Mr000001. [36] Hart B, Lundh A, Bero L. Effect of reporting bias on meta-analyses of drug trials: reanalysis of metaanalyses. Bmj 2012;344:d7202. [37] Egger M, et al. Language bias in randomised controlled trials published in English and German. Lancet 1997;350(9074):326e9. [38] Cooper HM. The integrative research review: a systematic approach. Sage Publications; 1984. [39] Schulz KF, Altman DG, Moher D. CONSORT 2010 statement: updated guidelines for reporting parallel group randomized trials. Ann Intern Med 2010;152(11):726e32. [40] Lachat C, et al. Strengthening the reporting of observational studies in epidemiology-nutritional epidemiology (STROBE-nut): an extension of the STROBE statement. PLoS Med 2016;13(6):e1002036. [41] Enhancing the QUAlity and transparency of health research. Available from: http://www.equator-network.org/. [42] Wells G, et al. The NewcastleeOttawa scale (NOS) for assessing the quality of non-randomized studies in meta-analysis. 2000. [43] Higgins JP, et al. The Cochrane Collaboration’s tool for assessing risk of bias in randomised trials. Bmj 2011;343:d5928. [44] Jadad AR, et al. Assessing the quality of reports of randomized clinical trials: is blinding necessary? Control Clin Trials 1996;17(1):1e12. [45] Shea BJ, et al. AMSTAR 2: a critical appraisal tool for systematic reviews that include randomised or nonrandomised studies of healthcare interventions, or both. BMJ 2017;358. [46] Whiting P, et al. ROBIS: a new tool to assess risk of bias in systematic reviews was developed. J Clin Epidemiol 2016;69:225e34. [47] European Food Safety A. Opinion of the Panel on dietetic products, nutrition and allergies (NDA) on a request from the Commission related to scientific and technical guidance for the preparation and presentation of the application for authorisation of a health claim. EFSA J 2007;5(7). p. 530en/a. [48] Covidence. Available from: https://www.covidence.org/. [49] Pan A, et al. Meta-analysis of the effects of flaxseed interventions on blood lipids. Am J Clin Nutr 2009; 90(2):288e97. [50] Cohen J. The significance of a product moment rs. In: Statistical power analysis for the behavioral sciences (revised edition). Academic Press; 1977. p. 75e107. Chapter 3. [51] Cohen J. The t test for means. In: Statistical power analysis for the behavioral sciences (revised edition). Academic Press; 1977. p. 19e74. Chapter 2. [52] Hedges LV. Distribution theory for glass’s estimator of effect size and related estimators. J Ed Stat 1981; 6(2):107e28. [53] Hedges LV, Olkin I. Estimation of a single effect size: parametric and nonparametric methods. In: Statistical methods for meta-analysis. San Diego: Academic Press; 1985. p. 75e106. Chapter 5. [54] Guyatt GH, et al. GRADE guidelines: 13. Preparing summary of findings tables and evidence profilescontinuous outcomes. J Clin Epidemiol 2013;66(2):173e83. [55] Abbott KA, et al. Do omega-3 PUFAs affect insulin resistance in a sex-specific manner? A systematic review and meta-analysis of randomized controlled trials. Am J Clin Nutr 2016;104(5):1470e84. [56] Wu L, Sun D. Meta-analysis of milk consumption and the risk of cognitive disorders. Nutrients 2016;8(12). [57] Cummings P. The relative merits of risk ratios and odds ratios. Arch Pediatr Adolesc Med 2009;163(5): 438e45. [58] Deeks JJ, Altman DG. Effect measures for meta‒analysis of trials with binary outcomes. In: Systematic reviews in health care. BMJ Publishing Group; 2008. p. 313e35.

194

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

[59] Localio AR, Margolis DJ, Berlin JA. Relative risks and confidence intervals were easily computed indirectly from multivariable logistic regression. J Clin Epidemiol 2007;60(9):874e82. [60] Parmar MK, Torri V, Stewart L. Extracting summary statistics to perform meta-analyses of the published literature for survival endpoints. Stat Med 1998;17(24):2815e34. [61] Borenstein M, et al. Fixed-effect model. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 63e7. [62] Borenstein M, et al. Fixed-effect versus random-effects models. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 77e86. [63] Borenstein M, et al. Random-effects model. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 69e75. [64] Fisher DJ, et al. A critical review of methods for the assessment of patient-level interactions in individual participant data meta-analysis of randomized trials, and guidance for practitioners. J Clin Epidemiol 2011; 64(9):949e67. [65] Simmonds MC, et al. Meta-analysis of individual patient data from randomized trials: a review of methods used in practice. Clin Trials 2005;2(3):209e17. [66] Thompson SG. Why sources of heterogeneity in meta-analysis should be investigated. Bmj 1994; 309(6965):1351e5. [67] Higgins JP, et al. Measuring inconsistency in meta-analyses. Bmj 2003;327(7414):557e60. [68] Borenstein M, et al. Identifying and quantifying heterogeneity. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 107e25. [69] Deeks JJ, Altman DG, Bradburn MJ. Statistical methods for examining heterogeneity and combining results from several studies in meta-analysis. In: Systematic reviews in health care. BMJ Publishing Group; 2008. p. 285e312. [70] Borenstein M, et al. Subgroup analyses. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 149e86. [71] Edited by Julian, P.T.H. and G. Sally, Cochrane handbook for systematic reviews of interventions v. 5.1.0 [updated March 2011]. 2008: Chichester, West Sussex, Hoboken NJ: John Wiley & Sons; 2008 ©2008. [72] Thompson SG, Higgins JP. How should meta-regression analyses be undertaken and interpreted? Stat Med 2002;21(11):1559e73. [73] Higgins JP, Thompson SG. Controlling the risk of spurious findings from meta-regression. Stat Med 2004; 23(11):1663e82. [74] Thompson SG, Sharp SJ. Explaining heterogeneity in meta-analysis: a comparison of methods. Stat Med 1999;18(20):2693e708. [75] Knapp G, Hartung J. Improved tests for a random effects meta-regression with a single covariate. Stat Med 2003;22(17):2693e710. [76] Partlett C, Riley RD. Random effects meta-analysis: coverage performance of 95% confidence and prediction intervals following REML estimation. Stat Med 2017;36(2):301e17. [77] Higgins JPT, Thompson SG, Spiegelhalter DJ. A re-evaluation of random-effects meta-analysis. J R Stat Soc A 2009;172(1):137e59. [78] Borenstein M, et al. Prediction intervals. In: Introduction to meta-analysis. John Wiley & Sons, Ltd; 2009. p. 127e33. [79] Riley RD, Higgins JPT, Deeks JJ. Interpretation of random effects meta-analyses. BMJ 2011:342. [80] Senftleber NK, et al. Marine oil supplements for arthritis pain: asystematic review and meta-analysis of randomized trials. Nutrients 2017;9(1):42. [81] Egger M, et al. Bias in meta-analysis detected by a simple, graphical test. BMJ 1997;315(7109):629e34. [82] Begg CB, Mazumdar M. Operating characteristics of a rank correlation test for publication bias. Biometrics 1994;50(4):1088e101.

REFERENCES

195

[83] Harbord RM, Egger M, Sterne JA. A modified test for small-study effects in meta-analyses of controlled trials with binary endpoints. Stat Med 2006;25(20):3443e57. [84] Peters JL, et al. Comparison of two methods to detect publication bias in meta-analysis. Jama 2006;295(6): 676e80. [85] Rucker G, Schwarzer G, Carpenter J. Arcsine test for publication bias in meta-analyses with binary outcomes. Stat Med 2008;27(5):746e63. [86] Sterne JA, Gavaghan D, Egger M. Publication and related bias in meta-analysis: power of statistical tests and prevalence in the literature. J Clin Epidemiol 2000;53(11):1119e29. [87] Stroup DF, et al. Meta-analysis of observational studies in epidemiology: a proposal for reporting. Metaanalysis of Observational Studies in Epidemiology (MOOSE) group. Jama 2000;283(15):2008e12. [88] P.R.I.S.M.A. flow chart. Available from: http://prisma-statement.org/prismastatement/flowdiagram.aspx. [89] Mocellin MC, et al. A systematic review and meta-analysis of the n-3 polyunsaturated fatty acids effects on inflammatory markers in colorectal cancer. Clin Nutr 2016;35(2):359e69. [90] Lewis S, Clarke M. Forest plots: trying to see the wood and the trees. Bmj 2001;322(7300):1479e80. [91] Dewansingh P, et al. Supplemental protein from dairy products increases body weight and vitamin D improves physical performance in older adults: a systematic review and meta-analysis. Nutr Res 2018;49:1e22. [92] Langlois PL, Hardy G, Manzanares W. Omega-3 polyunsaturated fatty acids in cardiac surgery patients: an updated systematic review and meta-analysis. Clin Nutr 2017;36(3):737e46. [93] Serban M-C, et al. A systematic review and meta-analysis of the impact of Spirulina supplementation on plasma lipid concentrations. Clin Nutr 2016;35(4):842e51. [94] Xiao Y, et al. Effect of nut consumption on vascular endothelial function: a systematic review and metaanalysis of randomized controlled trials. Clin Nutr 2017. [95] Ding Y, Li Y, Wen A. Effect of niacin on lipids and glucose in patients with type 2 diabetes: a meta-analysis of randomized, controlled clinical trials. Clin Nutr 2015;34(5):838e44. [96] Altman DG, Summing up. The science of reviewing research, Richard J. Light, David B. Pillemer, Cambridge, Mass.: Harvard University Press; 1984. No. of pages: xþ191. Price: £6.25 (paperback); £14.95 (hardback). Statistics in Medicine 1986; 5(3): 289. [97] Li Y. Association between fruit and vegetable intake and risk for glioma: a meta-analysis. Nutrition 2014; 30(11):1272e8. [98] Sterne JAC, et al. Recommendations for examining and interpreting funnel plot asymmetry in metaanalyses of randomised controlled trials. BMJ 2011:343. [99] Song F, et al. Extent of publication bias in different categories of research cohorts: a meta-analysis of empirical studies. BMC Med Res Methodol 2009;9:79. [100] Dickersin K. How important is publication bias?A synthesis of available data. AIDS Ed Prev 1997;9(1 Suppl):15e21. [101] Chan AW, et al. Empirical evidence for selective reporting of outcomes in randomized trials: comparison of protocols to published articles. Jama 2004;291(20):2457e65. [102] Horton J, et al. Systematic review data extraction: cross-sectional study showed that experience did not increase accuracy. J Clin Epidemiol 2010;63(3):289e98. [103] Buscemi N, et al. Single data extraction generated more errors than double data extraction in systematic reviews. J Clin Epidemiol 2006;59(7):697e703. [104] EFSA Panel on Dietetic Products, N and Allergies. General scientific guidance for stakeholders on health claim applications. EFSA J 2016;14(1). p. 4367en/a. [105] Waters E. Evidence for public health decision-making: towards reliable synthesis. Bull World Health Organ 2009;87(3):164. [106] Chalmers I. Raising the standards of clinical trials and research. Bull World Health Organ 2007;85(9): 658e9.

196

CHAPTER 7 META-ANALYSIS OF NUTRITION STUDIES

[107] Helfer B, et al. Recent meta-analyses neglect previous systematic reviews and meta-analyses about the same topic: a systematic examination. BMC Med 2015;13:82. [108] Gomersall JS, et al. Conducting systematic reviews of economic evaluations. Int J Evid Based Healthc 2015;13(3):170e8. [109] Anderson R. Systematic reviews of economic evaluations: utility or futility? Health Econ 2010;19(3): 350e64. [110] Riley RD, et al. External validation of clinical prediction models using big datasets from e-health records or IPD meta-analysis: opportunities and challenges. BMJ 2016:353. [111] Framingham coronary heart disease risk Score. Available from: https://www.mdcalc.com/framinghamcoronary-heart-disease-risk-score. [112] QRISK2. Available from:https://qrisk.org/. [113] Bleeker SE, et al. External validation is necessary in prediction research: a clinical example. J Clin Epidemiol 2003;56(9):826e32. [114] Khangura S, et al. Rapid review: an emerging approach to evidence synthesis in health technology assessment. Int J Technol Assess Health Care 2014;30(1):20e7. [115] Serghiou S, et al. Field-wide meta-analyses of observational associations can map selective availability of risk factors and the impact of model specifications. J Clin Epidemiol 2016;71:58e67. [116] Elliott JH, et al. Living systematic review: 1. Introductiondthe why, what, when, and how. J Clin Epidemiol 2017;91:23e30. [117] Higgins JP, Whitehead A, Simmonds M. Sequential methods for random-effects meta-analysis. Stat Med 2011;30(9):903e21. [118] Bucher HC, et al. The results of direct and indirect treatment comparisons in meta-analysis of randomized controlled trials. J Clin Epidemiol 1997;50(6):683e91. [119] Lumley T. Network meta-analysis for indirect treatment comparisons. Stat Med 2002;21(16):2313e24. [120] Aromataris E, et al. Summarizing systematic reviews: methodological development, conduct and reporting of an umbrella review approach. Int J Evid Based Healthc 2015;13(3):132e40. [121] Kyrgiou M, et al. Adiposity and cancer at major anatomical sites: umbrella review of the literature. Bmj 2017;356:j477. [122] Markozannes G, et al. Diet, body size, physical activity and risk of prostate cancer: an umbrella review of the evidence. Eur J Cancer 2016;69:61e9. [123] Tsilidis KK, et al. Type 2 diabetes and cancer: umbrella review of meta-analyses of observational studies. Bmj 2015;350:g7607. [124] Bougioukas KI, et al. Preferred reporting items for overviews of systematic reviews including harms checklist: a pilot tool to be used for balanced reporting of benefits and harms. J Clin Epidemiol 2018;93: 9e24. [125] Berlin JA, Longnecker MP, Greenland S. Meta-analysis of epidemiologic dose-response data. Epidemiology 1993;4(3):218e28. [126] Orsini N, et al. Meta-analysis for linear and nonlinear dose-response relations: examples, an Evaluation of approximations, and software. Am J Epidemiol 2012;175(1):66e73. [127] Wang L, et al. Tea consumption and lung cancer risk: a meta-analysis of caseecontrol and cohort studies. Nutrition 2014;30(10):1122e7.

CHAPTER

PRINCIPLES OF RESEARCH PUBLICATION

8

Gregory S. Patience1, George Pounis2, Paul A. Patience3, Daria C. Boffito1 Department of Chemical Engineering, Polytechnique Montre´al, Montre´al, QC, Canada1; Alimos, Athens, Greece2; Department of Electrical Engineering, Polytechnique Montre´al, Montre´al, QC, Canada3

CHAPTER OUTLINE 8.1 Introduction .................................................................................................................................198 8.2 Citation Impact and Metrics ..........................................................................................................198 8.3 Article Elements...........................................................................................................................200 8.3.1 Title ........................................................................................................................200 8.3.2 Abstract ..................................................................................................................203 8.3.3 Introduction.............................................................................................................204 8.3.4 Methods and Results ................................................................................................205 8.3.5 Discussion and Conclusions ......................................................................................207 8.4 Web Tools for Writing...................................................................................................................207 8.4.1 Word Choices ...........................................................................................................208 8.4.1.1 Dictionaries......................................................................................................... 208 8.4.2 Thesauri ..................................................................................................................208 8.4.2.1 Word Repetition .................................................................................................. 208 8.4.3 Grammar .................................................................................................................210 8.4.4 Translation ..............................................................................................................211 8.5 Reporting Data and Analysis..........................................................................................................211 8.5.1 Dietary Evaluation and Analysis .................................................................................212 8.5.2 Statistical Analysis ...................................................................................................212 8.5.3 Graphs ....................................................................................................................213 8.5.3.1 Information Line Density...................................................................................... 213 8.5.3.2 Graphical Element Design ................................................................................... 213 8.5.3.3 Information Signal-to-Noise Ratio......................................................................... 215 8.5.4 Tables .....................................................................................................................217 8.6 Publishing Process.......................................................................................................................218 8.6.1 Selecting the Journal ................................................................................................218 8.6.2 A Winning Cover Letter .............................................................................................218

Analysis in Nutrition Research. https://doi.org/10.1016/B978-0-12-814556-2.00008-7 Copyright © 2019 Elsevier Inc. All rights reserved.

197

198

CHAPTER 8 PRINCIPLES OF RESEARCH PUBLICATION

8.7 Authorship Criteria and Acknowledgments ..................................................................................... 220 8.8 Strengthening the Reporting of Observational Studies in Epidemiology and Consolidated Standards of Reporting Trials Statements....................................................................................... 223 8.9 Conclusions .................................................................................................................................224 References ..........................................................................................................................................225

8.1 INTRODUCTION The world’s capacity to innovate, create, and publish continues unabated as the corpus of scientific research approaches 2.5 million articles a year [1] and with as many patents [2]. However, our capacity to communicate clearly and concisely suffers as we race to publish and flourish [3]. In his final editorial comment for Nature, John Maddox [4] wondered whether researchers write obscurely on purpose and whether the scientific language suffered because of the pressure to publish. This publish or perish attitude has prompted radical ideas such as granting researchers a lifetime limit on the number of words they can write [5]. Unfortunately, our best ideas and our writing may not come early in our career. Furthermore, if we applied this reasoning to other endeavors, Beethoven would not have composed the Ninth Symphony, Shakespeare would not have written Macbeth or King Lear, and people would not be listening to the Beatles’ Abbey Road. Gladwell [6] popularized the notion that people required 10,000 h of dedicated practice to become world-class experts, which was based on the research of Ericsson et al. [7] Meta-analysis of many studies, on the other hand, suggested that dedicated practice accounts for less than 25% of the differences in ultimate performance [8]. Regardless, complaining about poor writing dates to the Sumerians [9], and Caxton documented the earliest complaints about speaking English from the 13th century [10,11]. Much of the world’s researchers, like Caxton’s good wyfe’s patrons, speak and write languages other than English, so not only do manuscripts have a poor writing style, often the grammar is incorrect. Here, we limit our discussion to style and adopt Occam’s razor to writing [12,13]; we apply Tufts’ precept that maximizes ink efficiency to graphs, tables, and presentations. Strunk and White’s book The Elements of Style, which is among the top 100 English books [14], exhorted us to use the active voice (Rule 11), favor the positive form (Rule 12) (“incomplete” rather than “not complete,” for example) and remove unnecessary words (Rule 13) [15]. We recommend that text be concise and precise, verbs be active and vigorous, and sentences be straightforward [12,16]. In The Sense of Style, Pinker [9] exhorts scientists to shed the metadiscourse: Hedging and signposting are two forms that encumber text and make research articles tiresome to read. He endorses the classic style that assumes readers are competent and that you guide them through the work as a conversation. We adopted this approach but agree with Peramo that our text is assertive [11,17].

8.2 CITATION IMPACT AND METRICS The Web of Science Core Collection (WoS) indexed 560,000 articles in 1989 and tripled that to 1,740,000 in 2017 (Fig. 8.1). The number of articles it assigns to nutrition and dietetics rose 20% more and reached a maximum of 11,300 in 2016. The category now counts about 90 journal titles, of which Progress in Lipid Research, Annual Review of Nutrition, and the American Journal of Clinical Nutrition had the highest impact factors in 2016 (NIF ¼ 10.6, 9.1, and 6.9, respectively). As of Feb.

8.2 CITATION IMPACT AND METRICS

199

11.3

9.0

1.50

1.20

7.0

0.90

5.0

0.56 1989

Articles (103)/y

WoS articles (106)/y

1.74

3.0 1995

2000

2005

2010

2017

Year

FIGURE 8.1 Growth in the number of articles indexed by WoS (open circles) versus those assigned to nutrition and diet (triangles). Source: 1titlesperyear.

2018, the journals with the most citations, Ncit, were Food Chemistry (Ncit ¼ 75,700), the American Journal of Clinical Nutrition (Ncit ¼ 56,000), and the Journal of Nutrition (Ncit ¼ 36,500) and they were also among the top-cited journals in 2016 (Table 8.1). Food Chemistry published six times more articles than did the American Journal of Clinical Nutrition but was referenced about 50% less; it had the lowest citation rate with respect to the number of articles, published at 4 (Ncit/Nart). The Journal of Nutrition and the European Journal of Clinical Table 8.1 Top 10 Cited Journals in 2016 (Ncit) Journal

NIF

Ncit

Nart

Ncit/Nart

American Journal of Clinical Nutrition Food Chemistry Journal of Agricultural and Food Chemistrya Journal of Nutrition British Journal of Nutrition PLOS Onea Appetite European Journal of Clinical Nutrition Int Journal of Obesity Public Health Nutrition

6.9

13,264

319

42

4.5 3.2

8,163 8,029

1,884 1,067

4 8

4.1 3.7

7,527 5,271

315 444

24 12

2.8 3.4 3.1

4,605 4,134 3,682

21,695 447 217

5.5 2.3

3,646 3,498

237 324

0.2 9 17 15 11

WoS assigned 80 journals to nutrition and dietetics that published 11,300 articles. The average number of references per article was 44. WoS, Web of Science Core Collection. a Journals not assigned to nutrition and diet.

200

CHAPTER 8 PRINCIPLES OF RESEARCH PUBLICATION

Nutrition published the same number of articles as did the American Journal of Clinical Nutrition and had the second and third highest citations rates, at 24 and 19, respectively. We examined relationships between journals and key words with VoS Viewer software (Figs. 8.2 and 8.3) [18]. It groups journals into five categories; these journals referenced 42,000 sources in 2016: 72 of the top-cited journals were referenced 800 times (Fig. 8.2). This represents one-third of all references but only 1% of all journals. The size of the text and circles in Fig. 8.2 roughly correlate with the frequency.

8.3 ARTICLE ELEMENTS 8.3.1 TITLE Search engines match words from titles, so more words possibly mean more hits. However, TED Talk titles average five words and the top 1000 most cited article in WoS averages less than 10 words (Fig. 8.4) [12,13]. More than 99% of titles in Nature and Science average less than 16 words. The success of these articles suggests that shorter titles are better: less than 16 words. The titles of articles WoS assigned to nutrition and dietetics in 2016 averaged 16 words with a standard deviation of 5 words. The longest had 42 words [19] and three had two-word titles: “Docosahexaenoic Acid” [20], “Ketogenic Diets” [21], and “Intestinal Microbiome” [22]. We group titles into five types: descriptive, enumerative, comparative assertive, and interrogative (Table 8.2) [12,38]. Assertive titles are statements with vigorous verbs that state facts: “improves,” “reduce,” and “increases” [23e25]. We prefer sentences rather than labels because they help readers

FIGURE 8.2 Most frequently journals cited in the bibliography of 11,160 articles in nutrition and dietetics from 2016. Data from WoS and plotted with VOS Viewer [18]; Source: 2references.

8.3 ARTICLE ELEMENTS

201

FIGURE 8.3 Author key words compiled by VOS Viewer [18] from articles WoS assigned to nutrition and dietetics (2016). BMI, body mass index; CVD, cardiovascular disease. Source: 3kewords.

immediately identify the essential theme of the work. However, Rosner [39] suggests that they trivialize the scientific process and he rejects journalistic-type pronouncements; most researchers write descriptive titles. Activating mutilated verbs converts descriptive titles to assertive ones. In the case of the title by Garcia et al. [34], it saves two words, “by” and “of”: “Deep Eutectic Solvents Extract Phenolic Compounds From Virgin Olive Oil.” The title by Lustig et al. [33] is intriguing but would be more powerful if the authors were to state what was adapted. Enumerative titles are often the longest because they list applications, techniques, compounds, equipment, and processes. Twenty percent of titles in nutrition and dietetics have at least two commas. Colons divide titles into two (known as hanging titles) with the subject preceding the object or result. Thirty percent in that category are hanging titles. Only 3% are interrogative (315), 105 of which are

202

CHAPTER 8 PRINCIPLES OF RESEARCH PUBLICATION

Normalized frequency, %

45 Nutrition Diet, 2016 Science, Nature, 2013 TED WoS Top 1000

30

15

0

5

10

15

20

25

30

Words per title FIGURE 8.4 Number of words per title in TED talks, the top 1000 cited articles in WoS, Nature and Science, and nutrition and diet. Source: 4words per title.

both interrogative and hanging titles: for example, “Sleep Deprivation Makes You Fat: Myth or Reality?” [40] and “The GuteBrain Axis: A Reality?” [41]. The title by Rondanelli et al. [28] has 27 words and is both assertive and enumerative. It is so long because it lists four subjects and five objects. Words such as “design,” “analysis,” “interpretation,” and “study” are superfluous except when the subject of the article is developing or designing experimental methods and instruments [26] rather than applying them. The three assertive examples in Table 8.2 state an experimental method that adds words unnecessarily. These abridged versions emphasize the results and not the means: “Dietary Nitrate Improves Vascular Function in Patients With Hypercholesterolemia” “Lipid-Based Nutrient Supplements for Pregnant Bangladeshis Reduce Newborn Stunting” “Diets Low in Advanced Glycation End Products Increase Insulin Sensitivity in Healthy Overweight Individuals” Like those in most scientific categories, researchers in nutrition and dietetics frequently add zero words that make titles longer without adding substantive information. Only 0.4% of Nature and Science articles include the word “study(ies),” whereas it appears 20 times more often in nutrition journals. Published research is new and most often novel, so these adjectives are superfluous as are nouns ending in “-ion” or “-ment.” The most common zero word is the verb mutilation of affect: “effect(s).” Even “affect” is an anemic verb because it is ambiguous. State whether the effect increases or decreases, improves or deteriorates rather than saying there is an effect: “Effects of Grape Seed Extract Beverage on Blood Pressure and Metabolic Indices in Individuals With Pre-hypertension: A Randomised, Double-Blinded, Two-Arm, Parallel, Placebo-Controlled Trial” [42, p. 226]. “Grape-seed Beverages Improve Blood Pressure in Individuals With Prehypertension”

8.3 ARTICLE ELEMENTS

203

Table 8.2 Examples of Titles From Nutrition and Dietetics in 2016 Type

Example

Assertive

“Dietary Nitrate Improves Vascular Function in Patients With Hypercholesterolemia: A Randomized, Double-Blind, Placebo-Controlled Study” [23] “Lipid-Based Nutrient Supplements for Pregnant Women Reduce Newborn Stunting in a Cluster-Randomized Controlled Effectiveness Trial in Bangladesh” [24] “Diet Low in Advanced Glycation End Products Increases Insulin Sensitivity in Healthy Overweight Individuals: A Double-Blind, Randomized, Crossover Trial” [25] “Best (but Oft-Forgotten) Practices: The Design, Analysis, and Interpretation of Mendelian Randomization Studies” [26] “Conventional, Ultrasound-Assisted, and Accelerated-Solvent Extractions of Anthocyanins From Purple Sweet Potatoes” [27] “Whey Protein, Amino Acids, and Vitamin D Supplementation With Physical Activity Increases Fat-Free Mass and Strength, Functionality, and Quality of Life and Decreases Inflammation in Sarcopenic Elderly” [28] “Higher Compared With Lower Dietary Protein During an Energy Deficit Combined With Intense Exercise Promotes Greater Lean Mass Gain and Fat Mass Loss: A Randomized Trial” [29] “Effect of Alirocumab on Specific Lipoprotein Non-High-Density Lipoprotein Cholesterol and Subfractions as Measured by the Vertical Auto Profile Method: Analysis of 3 Randomized Trials Versus Placebo” [30] “A Systematic Comparison of Sugar Content in Low-Fat Versus Regular Versions of Food” [31] “Persistent Metabolic Adaptation 6 Years After ‘The Biggest Loser’ Competition” [32] “Isocaloric Fructose Restriction and Metabolic Improvement in Children With Obesity and Metabolic Syndrome” [33] “Extraction of Phenolic Compounds From Virgin Olive Oil by Deep Eutectic Solvents (DESs)” [34] “Vitamin D Deficiency in Europe: Pandemic?” [35] “What Does Cooking Mean to You? Perceptions of Cooking and Factors Related to Cooking Behavior” [36] “How Can the EU Climate Targets Be Met? A Combined Analysis of Technological and Demand-Side Changes in Food and Agriculture” [37]

Enumerative

Comparative

Descriptive

Interrogative

8.3.2 ABSTRACT Researchers are likely to read the abstract if the title is intriguing, but they are unlikely to finish reading those that ramble. It must be concise and precise. To make it concise, avoid lists of tests, instruments, and procedures. To make it precise, report data, interpret them, and state conclusions and limitations. Science and Nature require authors to follow a prescribed structure that first includes a background sentence followed by a problem statement and then a sentence or two about what the researchers have achieved (starting with “We”), results, and finally conclusions. This format is similar to the Strengthening the Reporting of Observational Studies in Epidemiology (STROBE) and Consolidated Standards of Reporting Trials (CONSORT) guidelines.

204

CHAPTER 8 PRINCIPLES OF RESEARCH PUBLICATION

Metadiscourse is a writing style that clutters; Pinker [9] refers to it as “verbiage about verbiage.” Hyland [43] describes it more favorably as a writer’s pursuit to guide the reader’s perception of the text and includes hedging (adjectives that express uncertainty: “it would appear,” “suggest,” “seem”), boosting (“very,” “extremely,” “barely”), signposting (“as discussed in the next section”), narcissism (“researchers are now examining”), and self-consciousness (“this work is difficult/controversial”) [44]. Abstracts starting with “In the present study,” “The aim of this study was to,” or “This work focuses on” are poorly focused. Abstracts that state the motivation first (context and problem) are more apt to retain readers. Substituting “This study aimed to assess” with “We assessed” makes the text more succinct. Many journals are adopting structured formats to write abstracts; 42% of the 11,000 articles in nutrition and dietetics have sections headings such as those in the New England Journal of Medicine. The first heading includes “Aim” (106 occurrences), “Background” (2696), “Objective” (225), or “Purpose” (117). The second heading is often “Design” or “Methods.” However, even for these prescribed headings, researchers return to flabby expressions such as “Background/Objectives: This study aims to develop.” Here is an abstract that takes advantage of the section heading and clearly states the context [45]: Background and aims: Malnutrition and frailty are frequent and serious conditions within the geriatric population. Both are of multifactorial origin and linked to adverse outcomes.

8.3.3 INTRODUCTION Whereas the abstract asserts facts and highlights significant results, the introduction establishes the authors’ ethos, their credibility. It introduces the subject, reviews the literature, mentions controversies, and highlights the importance and context of the work. Like most sections of the manuscript, the first word and the first sentence are often the hardest to compose. Resist starting with broad generalizations or truths that are universally known or accepted as fact unless you can add information such as [46]: “Obesity prevalence, which has increased in all age groups over the past 30 years [47], may be leveling off.” Everyone knows that obesity is increasing but not that it is leveling off. Sharing how much obesity has increased over 30 y would improve the statement. The following sentence is a broad, vague generalization that the authors presented better in the title (“Review of Nut Phytochemicals, FatSoluble Bioactives, Antioxidant Components and Health Effects”) [48]: “Nuts contain a number of bioactives and health-promoting components [49,50].” Wolf et al. [51] introduces facts and cites the best journals to substantiate the first sentence in their introduction: “MicroRNAs [mRNAs] are about 22 nucleotides long [52], hybridize with complementary sequences in the 3#-untranslated regions in mRNA [52], and silence genes through destabilizing mRNA or preventing translation of mRNA [53,54].” Citing recent literature confirms to readers that the work is up-to-date, but what is a good mixture between established research and recent work? High impactefactor journals publish articles with more recent articles compared with low impactefactor journals [12,55]. A Weibull distribution characterizes the age of the references and accounts for 99% of variance in many top journals:

8.3 ARTICLE ELEMENTS   ti ni ¼ N 1  b

205

(8.1)

where N is the total number of references, ni is the number of references that are ti years old, and b is scale factor, which is 8.2 y for Nature: 63% of the references are younger than b. Almost 20% of references in Nature are less than 3 y whereas in nutrition and diet, only about 6% are that recent. Articles or journals that have a larger fraction of articles over 8.2 y are cited less and have lower impact. The last couple of sentences in the introduction often introduce the methods and motivation of the study. Although we often resort to signposting, which Pinker [9] recommends against doing, a concise sentence or two helps readers grasp the research question and objective, such as this example from Nutrition [56]. “However, whether probiotics have direct benefits on depressive symptoms and metabolic status in patients with MDD [major depressive disorder] has to date not been assessed. The present study was therefore conducted to assess the favorable effects of probiotic supplementation on symptoms of depression, parameters of glucose homeostasis, lipid concentrations, biomarkers of inflammation, and oxidative stress in patients with MDD.”

8.3.4 METHODS AND RESULTS We group these two sections together because they represent factual information. The prominence of the methods section is declining because some journals now include it at the end of the manuscript. Nonetheless, we recommend reporting the experimental methodology thoroughly so others can repeat and verify the data. Describe analytical equipment and major pieces of equipment, list reagents and materials, and include professional drawings when necessary. The past tense is appropriate to report the methods section, but minimize the passive voice. Take responsibility for the work and write in the third person plural [57]. Here, the authors explicitly stated what they did [58]: “In this retrospective study, we analyzed 130 consecutive patients with liver cirrhosis (LC) (age 6 of 9) had a lower cardiovascular risk compared with those with the lowest adherence score (30 g/d in the study of Grosso et al. versus >100 g/d from previous literature). Importantly, fish and fish oil consumption also result in lower risk for other diseases. Two prospective studies showed an inverse association of fish intake with various types of cancers [64,65]. Also, in a randomized, double-blind, placebo-controlled trial, high serum concentration/consumption of DHA were shown to be inversely associated with the level of cerebral amyloidosis (a preclinical stage of Alzheimer disease) [66]. It appears clear that fish and fish oil consumption should be encouraged as long as there are no allergies and care is taken in selecting the fish (preserving the marine environment and avoiding contaminants). This includes sea fish, seafoods, and freshwater fish.

9.4.2 PLANT OMEGA-3 FATS The traditional MD is rich in plant n-3 fatty acids, the main one which is a-linolenic acid (ALA). This n-3 fatty acid is essential to humans, which means they cannot synthesize it independently. Major dietary sources of ALA are vegetable oils (e.g., rapeseed oil and soybean oil), seeds (e.g., flaxseed), and nuts (e.g., walnuts). It is also found in wild plants consumed by goats, sheep, chickens, and rabbits, products of which (eggs, fermented milk, and meat) are then consumed by humans, in particular in the Mediterranean area. ALA consumption is associated with health benefits [67e69]. ALA is in fact a precursor molecule that can be converted into longer-chain n-3 such as EPA and DHA, as discussed in the previous section [70,71]. This conversion is catalyzed by an enzymatic system acting through a series of elongation and desaturation steps. The conversion of ALA to EPA and DHA might be influenced by genetic variations, gender, hormonal status, and nutrient substrate

9.4 FOOD COMPONENTS OF MEDITERRANEAN DIET

241

competition [72e74]. This leads to a conversion efficiency that appears low in humans, thus leading to the need to consume certain amounts of marine n-3. Another interesting aspect lies in the consumption of polyphenols, which seem to increase the conversion of ALA to DHA and EPA, as shown in epidemiological and animal studies [75,76] that were, however, not confirmed in human trials [77,78]. This might have been because in those studies, polyphenol ingestion was increased through supplementation and/or pure substances and not through the consumption of a healthy diet.

9.4.3 OLIVE OIL There is no doubt that olive oil is highly connected with the notion of an MD, in which extravirgin oil has been the main source of fat for a long time. Extravirgin olive oil is obtained purely from the fruit of the olive tree and more than 98% of it consists of fatty acids. Its fatty acids are mainly monounsaturated, such as oleic acid, but it also contains polyunsaturated fatty acids such as linoleic acid, and saturated fatty acids such as stearic or palmitic acid. In terms of its minor components, olive oil contains alcohols, sterols, hydrocarbons, and phenolic compounds; oleuropein is the most abundant one. The minor components of olive oil have a potential cardioprotective effect because of their antiinflammatory and antioxidant activity. The content of the minor components in the olive oil depends on the cultivar, climate, ripeness of the olives at harvest, and processing system applied. Studies have found an association of olive oil consumption and lowered risk for CVDs, such as CHD, stroke, and peripheral artery disease, and a lower risk for other diseases, such as hypertension, cancer, diabetes, and neurodegenerative disorders [79e83]. One of the most important findings comes from the PREDIMED large trial study, which reported a 39% CVD risk reduction in the group that consumed extravirgin olive oil. For each 10-g/day increase in extravirgin olive oil consumption, CVD and mortality risks decreased by 10% and 7%, respectively [84]. The proposed mechanisms of the protective action of olive oil are an improvement in lipid profile (high-density lipoprotein [HDL] increase and decrease in low-density lipoprotein [LDL], total cholesterol, and triglycerides); improvement in insulin resistance, blood pressure, and endothelial function; decrease in inflammatory markers; reduction in LDL and HDL oxidation; and reduction in coagulation factors and platelet aggregation [85].

9.4.4 FRUITS AND VEGETABLES Most if not all dietary plans encourage fruit and vegetable intake, usually followed by suggestions about portion sizes. For example, the Healthy Eating Plate created by the Harvard School of Public Health [86] advises that half of our plate should consist of fruits and vegetables. The American Heart Association also recommends five or more servings per day for adults [87]. This advice is based on epidemiological studies, metaanalyses, and RCTs pointing to the beneficial health effects of increased fruit and vegetable intake. In 2003, WHO concluded that evidence on the preventive effect of fruits and vegetables against CHD is powerful and advised an intake of 400e500 g/day (five to six portions of 80 g each) [88]. An example of epidemiological studies is the European Prospective Investigation Into Cancer and Nutrition (EPIC)-Heart, which found that fruit and vegetable consumption was associated with a 22% lower risk of fatal ischemic heart disease for subjects consuming at least eight portions (80 g each) of

242

CHAPTER 9 MEDITERRANEAN DIET: A HEALTH-PROTECTIVE DIETARY

fruits and vegetables a day, compared with those consuming fewer than three portions a day [89]. Another example is the Health Survey of England data, which reported an association of fruit and vegetable intake with decreased all-cause cardiovascular and cancer mortality [90]. A possible role for the cardioprotective effects of fruits and vegetables could be their high content of antioxidant vitamins and phytochemicals [91e93]. Several metaanalyses have been conducted in this area of research. In 2014, Wang et al. analyzed 16 prospective cohort studies and found a significant association between consumption of fruits and vegetables and all-cause mortality, particularly cardiovascular mortality. The study found a threshold of around five servings of fruit and vegetable per day, after which the reduction in risk of all-cause mortality did not continue [94]. Moreover, the systematic review and metaanalysis of Aune et al. [95] included a higher number of studies (n ¼ 95). The researchers reported a reduced risk for CHD, stroke, CVD, cancer, and all-cause mortality for fruits and vegetables combined, as well as individually examined [95]. Results from RCTs are not as aligned regarding the health effects of fruits and vegetables. A small intervention study in which a group of individuals consumed a standardized meal of 500 g fruits and vegetables per day, as well as 200 mL fruit juice per day, was compared with a group that consumed 100 g fruits and vegetables per day, with an energy- and fat-controlled diet. Over a short time (4 weeks), serum lipids, blood pressure, and hemostatic parameters were not significantly affected [96]. Conversely, a 6-month RCT including 690 healthy participants showed that participants allocated to the intervention group, in which a brief negotiation method to encourage fruit and vegetable consumption was used, had higher plasma concentrations of a-carotene, b-carotene, lutein, b-cryptoxanthin, and ascorbic acid and a more important decrease in systolic and diastolic blood pressure compared with the control group [97]. In addition, an increasing number of studies have examined the effects of specific fruits or vegetables. For example, a study found that within a 6-week time frame, subjects with habitually low intake of fruits and vegetables who consumed a black currant juice drink high in vitamin C and polyphenols showed a decrease in oxidative stress and an improvement in vascular health compared with the placebo group that did not receive the juice [98]. Another study reported that highsulforaphane broccoli sprouts may reduce undesirable overproduction of nitric oxide metabolites in Helicobacter pylorieinfected patients; these metabolites are a known risk factor for gastritis and gastric cancer [99]. The full evidence supporting the protective role of fruits and vegetables is lengthy and cannot be described in detail in this part of the chapter. However, the benefits of these products are undeniable and well-established from a plethora of studies and recommendations from global health organizations.

9.4.5 NUTS AND SEEDS Nuts and seeds are nutrient-dense foods that are high in vitamins and minerals; they have been part of the human diet since early times. These products contain, among other components, beneficial fatty acids and proteins. Common nuts of the MD are almonds, hazelnuts, walnuts, and pistachios. The saturated fat content of these nuts is low (4%e6%) whereas the MUFA and polyunsaturated fatty acid content is high. Especially walnuts can provide linoleic acid and a-linolenic acid, which have been linked to health benefits. Nuts are also a great source of fiber, protein, and micronutrients such as

9.4 FOOD COMPONENTS OF MEDITERRANEAN DIET

243

potassium, calcium, magnesium, and folate, antioxidant vitamins (such as tocopherols), and polyphenols, which may protect against oxidative stress [100]. Several epidemiological studies have looked into the association between nut consumption and health outcomes. A summary of data from the US cohort studies Adventist Health Study, Iowa Women’s Health Study, Nurses’ Health Study, and Physician’s Health Study found a doseeresponse relation in all of them, estimating an 8.3% reduction on CHD death risk for each weekly serving of nuts [101]. The study also indicated that the average risk reduction for CHD death in these four cohorts was 37%. Moreover, the cross-sectional, Multi-Ethnic Study of Atherosclerosis found that frequent nut and seed consumption was associated with lower levels of inflammatory markers. This could explain the inverse association of nut consumption with CVD and diabetes risk [102]. Clinical trials are aligned with epidemiological studies, with the example of the PREDIMED trial. In that study, participants who were allocated to the intervention group with 30 g/day mixed nuts had a 1-year significantly lower 1-year prevalence of high waist circumference and elevated triglycerides level, and high blood pressure compared with the control group (advice on low-fat diet) [103]. In addition, a pooled analysis of 25 intervention trials concluded that nut consumption may improve blood lipid levels in a dose-related manner, especially in subjects with high LDL-C or lower body mass index [104]. From the scientific evidence on the beneficial effects of nuts and seeds, it can be concluded that they are integral components of the MD, and although the specific mechanisms of their protective effects remain unclear, nut consumption is highly recommended, especially in individuals with a high risk for CHD.

9.4.6 DIETARY FIBER Dietary fiber traditionally refers to the indigestible part of plants; it can be found in fruits, vegetables, legumes, nuts, and seeds, as well as whole grains and cereals. Examples are lignin and polysaccharides. Oligosaccharides such as inulin and resistant starches have been added to the definition of dietary fiber [105]. The recommended intake of fibers is 25e30 g/day. Most available data derive from epidemiological studies and some intervention studies. Although the effect of specific food sources of fiber has not yet been determined, a large body of evidence exists on the association of dietary fiber intake and cardiovascular health, diabetes prevention and management, weight management, and gastrointestinal health. Because knowledge on the effects of fiber on weight management and gastrointestinal health is more widespread than the benefits of fiber against CVD and diabetes, this section of the chapter will focus on the latter two areas. To begin with, a high intake of dietary fiber has been associated with a decrease in prevalence of CHD and stroke. First evidence that highlighted the association between fiber intake and heart disease came from the Iowa Women’s Health Study, in which postmenopausal women with higher intake of whole-grain products had a lower risk for ischemic heart disease, with a one-third risk decrease in women consuming one or more servings of a whole-grain product per day, compared with those who rarely consumed such products [106]. Results from the Nurses’ Health Study agreed, showing that women in the highest quintile of whole-grain consumption (2.5 servings/day) had a >30% lower risk of CHD than women in the lowest quintile (median, 0.13 servings/day) [107]. Similar results were reported in later years [108]. An association of fiber intake with decreased risk of stroke was also found in both epidemiological studies as well as metaanalyses [109,110].

244

CHAPTER 9 MEDITERRANEAN DIET: A HEALTH-PROTECTIVE DIETARY

An additional clinical area in which dietary fiber could benefit is diabetes prevention and management. Early [111] and later cohorts [112] showed that cereal fiber intake was associated with a decreased risk for type 2 diabetes. Another study found that in addition to cereal fiber, total grain, whole-grain, and total dietary fiber were inversely associated with the risk for type 2 diabetes [113]. As demonstrated by a randomized crossover study in which participants followed either a diet with moderate amounts of fiber (24 g) or a high-fiber diet (50 g), the beneficial effects of the high-fiber diet could be attributed to the improvement of glycemic control and decreased hyperinsulinemia and plasma lipid concentrations [114]. Finally, a metaanalysis of prospective studies found a dosee response relation in which the risk for type 2 diabetes decreased by 6% for 2-g/day increments in cereal fiber intake [115].

9.4.7 WINE Wine consumption, especially red wine, has been shown to reduce risk for CVD in populations throughout the world. The first report of the protective effect of wine was in 1979 [116]. The wellknown “French paradox” was a term to define the finding that despite their saturated fat intake, which is similar to that of other populations, French people have a low incidence of CHD [117]. Although the “French paradox” did not imply a causal relation between wine consumption and lowered risk for CHD, it spurred more in-depth research on the subject. Many studies have shown a cardioprotective effect of wine, including systematic reviews and metaanalyses such as the one by Di Castelnuovo et al., who analyzed 26 studies and found that a J-shaped inverse association exists between the amount of wine intake and vascular risk [118]. This association was confirmed later by Costanzo et al. [119]. Another systematic review and metaanalysis by Ronksley et al. included 84 prospective cohort studies and found that light to moderate alcohol consumption is linked to a reduced risk for several cardiovascular outcomes [120]. A review of controlled clinical trials found that the most consistent positive results of wine relate to lipid metabolism and are mainly attributed to ethanol. The microconstituents of wine, such as phenolic compounds, seem to have an important role in the hemostatic and inflammatory/endothelial system, but this needs to be confirmed by more intervention studies [121]. One of the main principles of the MD lies in the concept of moderation. Consuming higher amounts of alcohol than recommended or allowed results in not only a lack of protection but probably even an increase in the risk for CVD. This implies that whatever the type of alcohol consumed in the setting of an MD, it is fundamentally important to respect the Mediterranean “way of drinking.” That means in moderate quantities and most of the time combined with consuming solid foods, which leads to a delay in the passage of ethanol into the bloodstream.

9.5 MEDITERRANEAN DIET ADHERENCE IN MODERN TIMES 9.5.1 LEVEL OF ADHERENCE IN MODERN POPULATIONS The typical method for determining adherence to the MD has been to score the dietary habits of sample populations, in which in most cases a higher score stands for higher adherence [122]. For instance, the Greek EPIC cohort study used the widely accepted MDS and indicated that an increase of two units in the MDS was associated with a 25% reduction in all-cause mortality [5]. This index was calculated

9.5 MEDITERRANEAN DIET ADHERENCE IN MODERN TIMES

245

based on nine components of the MD: vegetables, legumes, fruits and nuts, cereals, fish and seafood, meat and meat products, dairy products, moderate alcohol intake, and MUFAeSFA ratio. A value of 0 or 1 was assigned to subjects whose consumption was below the median (value: 0) or at or above the median (value: 1). Therefore, the total MDS could range from 0 to 9. The Mediterranean Adequacy Index (MAI) is another score that characterizes a diet compared with the MD; it was generated by dividing the sum of the percentage of total energy from typical Mediterranean food groups by the sum of the percentage of total energy from nontypical Mediterranean food groups [123]. The reference MD was that from Nicotera (southern Italy) in 1960 and the data for calculating the MAI were derived from the Seven Countries study. Research found the median MAI value of Nicotera in 1960 to be 7.2 among men aged 40e59 years. This value differed significantly from Crevalcore, where it was 2.9 in 1965 and 2.2 in 1991. A similarly decreased score was calculated for other regions: for example, in Montegiorgio, the MAI score was 5.6 in 1965 and 3.9 in 1991. Thus, it was made clear that the dietary habits of Italian populations have changed over time, shifting away from the reference Italian MD. A large Italian cohort studied the association between stroke and adherence to four a priorie defined dietary patterns: Healthy Eating Index 2005, Dietary Approaches to Stop Hypertension, the MDS (designed to estimate adherence to the Greek variant of the MD), and the Italian Mediterranean Index (specifically developed to estimate better adherence to the Italian MD). For the Italian population, a diet with a high score on the Italian Mediterranean Index performed better than the other indexes; it had inverse associations with stroke and ischemic stroke and a tendency to be inversely associated with hemorrhagic stroke [124]. Another study focusing on the Molise region in Italy recorded adherence to the MD by the Greek MDS (Trichopoulou et al.) and an Italian MD index, which was based on the intake of 11 items and was calculated according to whether an individual’s consumption of the foods was in the third tertile of the distribution (score of 1) or not (score 0). Possible scores ranged between 0 and 11. The health-related quality of life was also measured by the Short Forme36 scale, which contained 36 items on domains such as physical functioning, bodily pain, and role limitations owing to emotional or mental health problems. The study concluded that there was a positive association between adherence to the MD and health-related quality of life, and that the association was stronger for mental health than physical health [125,126]. The MD 55 Score, developed by Panagiotakos et al., was based on the Greek ATTICA study and aimed to detect clinical characteristics associated with cardiovascular risk by using its cutoff point [127]. They used 11 main components of the MD (unrefined cereals, fruits, vegetables, potatoes, legumes, olive oil, fish, red meat, poultry, full-fat dairy products, and alcohol) and assigned scores from 0 to 5, depending on the frequency of consumption of each of these items (0 ¼ no consumption to 5 ¼ daily). A total score from 0 to 55 was then calculated. The study reported the mean diet score to be 25.46 for men and 27.18 for women. The score was 23.5 in hypertensive versus 26.8 in normotensive subjects, 24.7 in subjects with hypercholesterolemia versus 26.6 in subjects with normal levels of blood cholesterol, 22.2 in diabetic subjects versus 26.2 in nondiabetic subjects, and 22.2 in obese subjects versus 26.5 in normal/overweight subjects. The 14-point MD Adherence Screener is another index that was created to provide rapid control of compliance with the dietary intervention of the Spanish PREDIMED study. The assessment consisted of 14 food consumption frequency questions on several characteristics of the MD; each question was scored as 0 (no fulfillment) or 1 (criterion fulfillment). The study reported a mean

246

CHAPTER 9 MEDITERRANEAN DIET: A HEALTH-PROTECTIVE DIETARY

score ( standard deviation) of 8.5  2.0 for the 14-item screener. For men, this was 8.7  2.0 and for women, 8.5  2.0 [128]. A Spanish study developed a self-efficacy scale for adherence to the MD that measures the extent to which people are confident about their ability to adhere to the MD [129].

9.5.2 CHALLENGES OF ADHERING TO THE MEDITERRANEAN DIET Despite the plethora of health benefits associated with the MD, it seems that adhering to it can be challenging. Globalization, economic, urban, and technology-driven developments have led to a significant shift toward the homogenized Western diet, which is high in refined sugars, and saturated and trans fats and low in fresh fruits, vegetables, legumes, nuts, and seeds. Although the traditional MD used to be found in poor rural societies of the Mediterranean area in the 1950s, today, low-income groups show an increase in the prevalence of CVDs as a result of the shift toward the Western diet. Therefore, the MD is no longer the diet of the poorest. The higher price of healthy, fresh foods compared with the low price of unhealthier snacks and fast food could also reduce adherence to the MD, because it has been found in the Moli-Sani study, in which adherence to the MD was highly related to material resources [130]: the greater the income, the higher the adherence to the MD. Especially when the economic crisis in Italy became manifest between 2007 and 2010, adherence to the MD decreased dramatically (18.3%), especially in the elderly, the less affluent, and urban inhabitants. Moreover, it was seen that socioeconomic indicators were greatly importance in this period, although this had not been seen before the onset of the economic crisis [131]. Two additional factors that might be important for adhering to the MD are nutrition knowledge and exposure to mass media [125,126,132]. The higher these two parameters are, the higher conformity to the MD might be. However, the authors of the study that suggested these relations noted that the quality of information from mass media should be studied as well, for the previous conclusion to be established, because in that study the researchers asked about only the time participants spent using mass media and not the kind of information to which they were exposed. It is likely that material resources are the most important factor influencing adherence to the MD, with a positive relation between the two. Other factors such as nutritional information and mass media exposure might have a role as well.

9.6 SHIFTING TO THE MEDITERRANEAN DIET IN THE MODERN CONTEXT 9.6.1 UPDATED MEDITERRANEAN DIET RECOMMENDATIONS Despite the previously mentioned challenges, the MD remains a healthy lifestyle worth following. Although there is general consensus among the scientific community regarding the characteristics of the MD, there has been discussion the MD pyramid should be updated to cover the changing dietary and sociocultural lifestyle as well as the environmental and health challenges of the modern society. Therefore, the new MD pyramid [133] initiated by the Mediterranean Diet Foundation includes both quantitative and qualitative aspects of food selections indicating both the frequency of consumption and the serving amount of foods that can make people healthy or unhealthy. Main recommendations of the new MD pyramid are related to food. It is recommended to consume plant-based foods on a daily basis, such as fruits and vegetables; olive oil; bread, pasta, and rice;

9.6 SHIFTING TO THE MEDITERRANEAN DIET IN THE MODERN CONTEXT

247

grains; and cereals. To cover the intake of MUFA, foods such as extravirgin olive oil and avocado can be consumed. Extravirgin olive oil might be considered expensive compared with other oils, but owing to its high caloric content it does not need to be consumed in large amounts. Nuts are recommended because they are a rich source of proteins, unsaturated fats, fiber, and micronutrients. Mediterranean nuts include almonds, hazelnuts, pine nuts, pistachios, and walnuts. Walnuts in particular are a rich source of a-linolenic acid and have the highest level of phenolic compounds compared with other nuts [134]. Another important component of the MD is water. Hydration is important, and a daily intake of 1.5 to 2 L of water is strongly advised. This intake can be complemented with nonsugar herbal infusions such as tea, and broths. Dairy products can be consumed in moderate amounts (2 servings/day), primarily in the form of yogurt, cheese, and other fermented dairy products. Dairy products provide a range of nutrients such as proteins, calcium, and vitamins; however, products such as butters, creams, and dairy ice-creams are not included in this recommendation and should be consumed less often and in small amounts, owing to their high energy density. A moderate consumption of wine (1 glass/day for women and 2 glasses/day for men) during meals is recommended, but religious and social beliefs should be considered. On a weekly basis, it suggested to consume fish and seafood (
2 servings/week), which provide essential protein and lipids. The alteration of consumption of oily and lean fish and shellfish is recommended. Fish is a good nonvegetarian source of marine omega-3 fatty acids but until recently, vegetarians and vegans who wanted to supplement with omega-3 fatty acids had only a few natural options such as flaxseeds. These options do not provide the marine omega-3 fatty acids, such as EPA and DHA, which can be found only in fish oil, and the conversion of plant omega-3 to marine omega-3 is limited in humans. A solution to this could be algae, which is an emerging food source that is rich in marine omega-3 fatty acids, offering a great solution to those who avoid animal products. White meat, such as poultry, turkey, and rabbit (2 servings/week), and eggs (2e4 servings/week) can also be consumed on a weekly basis because they are good sources of protein and are lower in saturated fat than red meat. Red meat (35 y N ¼ 29,079

Soy, natto, ISO, natto-derived ISO

CPS-II [48]

70 y (mean) N ¼ 98,469

TOTFL, ANTHO, FLV3L, FLVNN, FLVN, FLVONL, ISO, PRO

Iowa Study [50]

55e69 y N ¼ 34,489 p.m.

NHS [13]

30e55 y N ¼ 66,360 women 32e81 y N ¼ 43,880 men 18 y N ¼ 5179 29e70 y N ¼ 41,438

TOTFL, ANTHO, FLVNN, FLVN, FLVONL, ISO, FLV3L, PRO FLVONL, FLVN

United States

HPFS [15] NHANES [53] Spain

EPIC-Spain [56]

Australia

CIFOARE [47]

Italy

[17]

Finland

Kuopio Ischemic Heart Study [51]

42e60 y N ¼ 1950 men

Kuopio Ischemic Heart Study [55] ATBC [22]

42e60 y N ¼ 1889

>75 y N ¼ 1063 p.m. 45e64 y N ¼ 1658

ANTHO, FLVNN ISO, lignan TOTFL, FLV3L, FLVONL, FLVNN, FLVN, ISO, ANTHO, PRO, lignan TOTFL TOTFL, FLV3L, FLVNN, FLVONL, FLVN, ISO, ANTHO, PRO TOTFL, FLVONL, FLVN, FLVNN, FLV3L, ANTHO Lignan (enterolactone) Lignan (enterolactone)

Significant Outcomes CVD: Soy () in women Stroke: n.s. CVD: Natto and Natto-derived ISO () Stroke: Soy and Natto () CVD: TOTFL, ANTHO, FLV3L, FLVN, FLVONL, PRO () CHD: FLVN () Stroke: TOT () in men CVD: ANTHO () CHD: ANTHO, FLVNN () Stroke: n.s. CHD: FLVONL ()

CHD: n.s. Stroke: n.s. CVD: ISO (þ), lignan () CVD: TOTFL, FLVNN, FLVONL, FLV3L, PRO ()

CVD: () CVD: n.s.

CVD: n.s. Stroke: FLVONL () CVD: () CHD: () CHD: n.s. Continued

270

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

Table 10.2 Polyphenols and Cardiovascular Disease (CVD) Incidence and Mortality: Evidence From Observational Studiesdcont’d Mortality of CVD, CHD, and Stroke Country

Cohort

Population

PP/Food

Zutphen Elderly Study [46]

50e69 y N ¼ 711 male smokers 65e84 y N ¼ 774 men

FLV3L (epicatechin)

Zutphen Elderly Study [49]

65e84 y N ¼ 570 men

Lignan

Singapore

Singapore Chinese Health Study [54]

45e74 y N ¼ 60,298

Soy, tofu equivalents, ISO

United Kingdom

EPIC-Norfolk [20]

40e75 y N ¼ 24,885

FLV3L

Netherlands

Significant Outcomes

CVD: () only in men with prevalent CVD CHD: () CVD: matairesinol () CHD: matairesinol () CVD: soy, tofu equivalents (þ) in men CHD: n.s. Stroke: n.s. CVD: n.s.

(), Inverse association or decrease; (þ), positive association or increase; ANTHO, anthocyanin; ATBC, Alpha-Tocopherol, BetaCarotene Cancer Prevention Study; CHD, coronary heart disease; CIFOARE, Calcium Intake Fracture Outcome Age-Related Extension; CPS, Cancer Prevention Study; EPIC, European Prospective Investigation of Cancer; FLV3L, flavanol; FLVN, flavone; FLVNN, flavanone; FLVONL, flavonol; HAPIEE, Healthy, Alcohol and Psychosocial Factors in Eastern Europe; HPFS, Health Professionals Follow-up Study; InCHIANTI, Invecchiare in Chianti (Aging in the Chianti Area); ISO, isoflavone; JPHCBC, Japan Public Health Center-Based Cohort; MEAL, Mediterranean Healthy Eating, Aging, and Lifestyle; n.s., not significant; NHANES, National Health and Nutrition Examination Survey; NHS, Nurses’ Health Study; PM, postmenopausal women; PP, polyphenol; PRO, proanthocyanidin; TOTFL, total flavonoids.

10.4.1.3 Stroke Incidence Five studies [12,14,15,19,20] assessed the relation between polyphenols and stroke incidence. Isoflavone was found to increase the risk of stroke incidence in a cohort of Shanghainese women [19] observed for an average 10 years but it decreased the risk in Japanese women, with no significant association in men [12]. Cassidy et al. similarly found a risk increase with flavanone but no association with anthocyanin in the Health Professionals Follow-up Study [15]. No association was found with flavanol in the EPIC-Norfolk cohort [20], and also with all flavonoid subclasses in the Nurses’ Health Study [14].

10.4.1.4 Cardiovascular Disease Mortality CVD mortality was the most frequently assessed outcome among the observational studies, totaling 14 studies across nine countries [12,17,20,46e56]. Total flavonoid intake was negatively associated with

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

271

CVD mortality in three studies [47,48,56], and three studies found no significant association [17,50,51]. In Japan, Kokubo et al. found that soy, but not isoflavone, was negatively associated with CVD mortality [12], whereas Nagata et al. found no significant relation between soy and CVD mortality [52]. In addition, natto and isoflavone derived strictly from natto, but not isoflavone derived from all foods, had a negative association with CVD mortality. Finally, both soy and tofu were positively associated with CVD mortality in Singaporean men but not women; however, isoflavone had no association [54]. Isoflavone was generally not associated with CVD mortality; eight studies [12,17,48,50e52,54,56] concluded that there was no significance and one study observed a positive association between isoflavone intake and CVD mortality in the Nurses’ Health Study [53]. Flavonol was associated with a decreased risk for CVD mortality in two studies [48,56] but had no association in three studies [17,50,51]. Flavone was negatively associated with CVD mortality in only one study [48] and was not significantly associated in four [17,50,51,56]. Similarly, flavanone was found to decrease CVD mortality risk in one study [56]; four studies found no association [17,48,50,51]. Flavanol was negatively associated in two studies [48,56] and not significantly associated in four [17,20,50,51]. In one study, epicatechin intake decreased CVD mortality risk but only in men with CVD [46]. Two studies [48,50] found negative associations between anthocyanin and CVD mortality and three studies [17,51,56] found no association. Lignan was associated with decreased CVD mortality risk in three studies [49,53,55] and was not significantly associated in one [56].

10.4.1.5 Coronary Heart Disease Mortality Nine studies [13,15,22,46,48e50,54,55] assessed polyphenols’ relation with CHD mortality risk. Three studies observed a negative association between lignan and CHD mortality [22,49,55]. CHD mortality risk was decreased in one of three studies with high intakes of flavonol [13,48,50], flavone [13,48,50], flavanone [15,48,50], flavanol [46,48,50], and anthocyanin [15,48,50]. Total flavonoid [48,50] and isoflavone [48,50,54] intake was not associated with CHD mortality risk.

10.4.1.6 Stroke Mortality Seven studies [12,15,48,50e52,54] assessed polyphenol intake and stroke mortality risk; most found no significant association. In the Cancer Prevention StudyeII, total flavonoid intake was associated with decreased stroke mortality risk only in men; two other studies found no association [50,51]. Five studies [12,48,50,52,54] reported no association between isoflavone and stroke mortality, although subjects in the Takayama study who consumed large amounts of soy and natto had a lower risk for stroke mortality [52]. In the Kuopio Ischemic Heart Study, which observed 1950 men over an average of 15 years, flavonol was the only flavonoid class associated with decreased risk for stroke mortality. There were no significant associations between flavone, flavanone, flavanol, and anthocyanin and stroke mortality [15,48,50,51].

10.4.1.7 Current Evidence at a Glance A total of 28 observational studies, mostly prospective cohorts (Table 10.2), representing North American, European, and Asian countries, assessed the relation between polyphenol intake and the incidence and mortality of various categories of CVD. Flavonoids were the most commonly studied polyphenol; lignans, stilbenes, and phenolic acids were assessed only rarely. Results were not entirely uniform but flavonoids showed potential in preventing CVD mortality. There is ample evidence to suggest that lignans may reduce CHD mortality. Polyphenols likely have either a beneficial or null

272

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

effect on CVD incidence and mortality, although a small minority of studies reported a possibility of increased risk. It is important to consider factors that were and were not controlled for in the statistical analyses in each study, which may have contributed to the inconsistent results. Age, body mass index (BMI), physical activity, smoking, history of CVD, medication use, and intakes of certain nutrients were fairly consistently considered in adjusted models across studies. Other factors such as education, blood pressure (BP), family history of CVD, and menopausal status were only sometimes included in adjustments. It is also likely that residual confounding may have been present in those studies owing to factors that could not be measured and/or controlled.

10.4.2 CARDIOVASCULAR DISEASE BIOMARKERS: EVIDENCE FROM EPIDEMIOLOGICAL STUDIES 10.4.2.1 Body Composition Three cross-sectional studies [57e59], each examining data from the National Health and Nutrition Examination Survey (NHANES) at different times, assessed the relations between polyphenol intake and weight, BMI, or waist circumference. Using 2007e12 data, Kim et al. found that BMI was inversely associated with anthocyanidin (aglycones of anthocyanin) intake, but no association was found with any other flavonoid class; waist circumference also was not associated with flavonoids [57]. Sebastian et al. used data from 2007 to 10 and found that among the flavonoid subclasses, there was a negative association among flavonoid, flavanol, anthocyanin, and flavanone intake and BMI and waist circumference [58]. Struja et al. assessed urinary phytoestrogen excretion from 1999 to 2004 data and reported that increased excretion of urinary isoflavone, indicative of increased intake, was associated with lower waist circumference [59]. One study investigating the relation between polyphenol intake and metabolic syndrome in Iranian adults found that total flavonoid, but not any other polyphenol class, was associated with lower waist circumference [60].

10.4.2.2 Lipid Profile Eleven studies [57,59e68] compared dietary polyphenols and cholesterol, lipoproteins, and triglycerides. Four studies [57,59,65,66] used NHANES cross-sectional data to assess triglycerides and highdensity lipoprotein (HDL). In the study by Kim et al, total flavonoid intake was favorably associated with both triglycerides and HDL and anthocyanin was negatively associated with triglycerides [57]. In the study by Struja et al., both isoflavone and lignan were associated with lower triglycerides and increased HDL [59]. The third study found that lignan intake was associated with increased HDL and decreased triglycerides [65]. The fourth study assessed the relation between urinary isoflavone concentration and CVD biomarkers in pregnant women and reported that increased isoflavone intake was associated with lower triglyceride levels [66]. Two studies [61,68] used data from the Spanish PREDIMED cohort and found that triglycerides were negatively associated with total polyphenol intake [61] and increased urinary resveratrol metabolites [68], a biomarker of red wine and resveratrol intake. In an Italian cohort of middle-aged type 2 diabetic individuals, high intakes of total polyphenols, total flavonoids, and phenolic acids were associated with improved low-density lipoprotein (LDL), HDL, and triglycerides [67]. Total flavonoid intake was similarly associated with improved triglycerides and HDL in a cross-sectional study of

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

273

2618 healthy Iranian adults; high lignan intake, however, was associated with increased triglyceride levels [60]. In a Chinese study assessing the relation between flavonoid, flavonoid subclass, and stilbene intake with various CVD biomarkers, anthocyanin intake was associated with increased HDL, and total flavonoid, flavone, and flavonol intake were associated with decreased triglyceride levels; these effects were observed only in women [62]. Flavanone intake was associated with lower levels of cholesterol and apolipoprotein B (apoB) in a cohort of Korean women with metabolic syndrome, but not in women without metabolic syndrome [64]. There was no association between moderate consumption of coffee, an abundant source of phenolic acids, and any lipid biomarkers [63].

10.4.2.3 Blood Pressure Ten studies [57,59e63,67,69e71] assessed BP as an outcome. Five studies [61,63,67,70,71] reported negative associations between various polyphenols and BP. Isoflavone was associated with lower BP in middle-aged Japanese [70] and young-adult American [71] cohorts. Total polyphenol intake was associated with lower BP in both healthy [61] and diabetic [67] cohorts. Moderate coffee consumption was also associated with lower BP in a Brazilian study [63]. Two studies [60,69] reported mixed results: In the British TwinsUK cohort, anthocyanin was the only flavonoid associated with lower BP [69]. In the Iranian Tehran Lipid and Glucose Study (TLGS), total flavonoid intake was associated with lower BP, stilbene was associated with increased BP, and both lignan and total polyphenol intake had no association [60]. Three studies [57,59,62] reported no significant associations with flavonoids [57,59,62], isoflavone and lignans [59], or stilbenes [62].

10.4.2.4 Glucose Tolerance Eight studies [57,59,60,62,63,66e68] assessed the relation between polyphenols and various biomarkers of glucose tolerance, including fasting blood glucose (FBG), hemoglobin A1C (HbA1c), serum insulin, and homeostatic model assessment of insulin resistance. In two [57,66] of the three NHANES studies [57,59,66], isoflavone intake was associated with improved glucose tolerance. Of all polyphenol classes, only total flavonoid was associated with improved FBG in the TLGS [60]. Lignan was associated with elevated FBG [60] and had no association in the study by Struja et al. [59]. In type 2 diabetic people, total polyphenol and total flavonoid but not phenolic acid intake were associated with lower HbA1c levels [67]. Zamora-Ros et al. [68] linked increased urinary resveratrol metabolites with lower FBG in the PREDIMED cohort. There were no significant associations between glucose tolerance and coffee [63], flavonoids, or stilbenes [62].

10.4.2.5 Inflammation, Endothelial, and Vascular Function Five studies [62,67,69,72e73] assessed outcomes directly related to arterial health. Of all flavonoid subclasses, flavonol was associated with lower soluble vascular cell adhesion molecule-1, and total flavonoid, anthocyanin, and flavone were associated with lower interleukin-18 in women in the Nurses’ Health Study [72]. Pounis et al. [73] collected dietary data and blood measurements from an Italian cohort from 2005 to 10 and found that total flavonoid and lignan intake were associated with a lower inflammatory score, an aggregated composite of four inflammatory biomarkers. In type 2 diabetic individuals, total polyphenols, total flavonoid, and phenolic acid were all negatively associated with C-reactive protein (CRP), a biomarker of inflammation [67]. In the TwinsUK cohort, flavone and

274

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

anthocyanin were associated with decreased pulse wave velocity, the reference standard biomarker of arterial stiffness [74]. All flavonoids and stilbenes were not significantly associated with intima-media thickness, an indicator of vascular health, in two studies [62,69].

10.4.2.6 Current Evidence at a Glance Seventeen observational studies, mostly cross-sectional (Table 10.3), examined the relation between polyphenols and biomarkers of CVD. Lipids and BP were the two most commonly measured outcomes, and flavonoids were the polyphenols most frequently assessed. Many of the studies included subjects with elevated risk for CVD. Metabolic syndrome was a disease of particular interest. Other studies measured components of metabolic syndrome as outcomes or recruited subjects with prevalent metabolic syndrome. Results suggest that polyphenols likely have a role in improving cardiovascular health by modifying key CVD biomarkers, at least in populations at high CVD risk, but the evidence is not uniform. The large variety of polyphenols and the wide array of biomarkers measured, with a different selection of biomarkers examined in each study, add to the difficulty in reconciling findings.

10.4.3 CARDIOVASCULAR DISEASE BIOMARKERS: EVIDENCE FROM INTERVENTIONAL STUDIES 10.4.3.1 Polyphenol-Rich Diets Four studies [75e78] assessed the relation between diets rich in polyphenols and CVD biomarkers in subjects at high risk for CVD. One study found that a polyphenol-rich diet decreased plasma triglycerides, very-LDL, and urinary 8-isoprostane, but also unexpectedly decreased HDL [76]. The polyphenol-rich diet also decreased FBG and increased serum insulin response in the same group of subjects [77]. In a 12-week polyphenol-rich diet intervention trial [79], which included six servings of fruits and vegetables and 50 g dark chocolate daily, subjects saw improved CVD biomarkers with decreased total cholesterol and increased forearm blood flow response to acetylcholine, an independent biomarker of CVD morbidity in hypertensive individuals [80]. Medina-Remon et al. compared the effects of a Mediterranean diet (MED) with extravirgin olive oil (EVOO) and a MED with nuts with a low-fat control diet for 1 year in elderly subjects in the Spanish PREDIMED cohort [78]. The MED with EVOO and MED with nuts diets, which were both higher in polyphenol content than the control diet, significantly improved BP, an effect that was shown to be mediated by nitric oxide (NO).

10.4.3.2 Cocoa/Chocolate Eight studies [81e88] applied cocoa or dark chocolate as interventions, four [81,82,86,87] using strictly cocoa beverages, two [85,88] using solid dark chocolate, and two [83,84] using both solid dark chocolate and cocoa beverages. Cocoa and dark chocolate are rich sources of flavanol, whereas white chocolate is a poor source of any polyphenol; thus, white chocolate is commonly used as a polyphenolpoor control for dark chocolate. The four studies employing cocoa beverages as the experimental intervention each observed the expected outcome: improved glucose sensitivity, lipid profile, BP, and 8-isoprostane in an elderly population [86]; improved vascular function in healthy young adults [87]; favorable improvements in BP and flow-mediated dilation (FMD) response to exercise in healthy overweight subjects [82]; and improved FMD in type 2 diabetic subjects [81].

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

275

Table 10.3 Cardiovascular Disease Biomarkers: Evidence From Observational Studies Country

Cohort

Population

PP/Food

End Point

Japan

JPHCBC II [70]

40e69 y N ¼ 41,651

BP

United States

WWEIA [58]

20 y N ¼ 10,538

CARDIA [71]

18e30 y N ¼ 3142 19 y N ¼ 4042

Soy, fermented soy, ISO, fermented soyderived ISO TOTFL, ANTHO, FLV3L, FLVNN, FLVN, FLVONL ISO

NHANES (Kim et al., 2016) [57]

Italy

BMI, WC

Fermented soy and fermented soy-derived ISO () BP TOTFL, FLV3L, ANTHO, FLVNN () BMI and WC

BP

ISO () BP

TOTFL, FLVONL, FLVN, FLVNN, FLV3L, ANTHO, ISO

BMI, WC, lipids, BP, GT

TOTFL () TG, (þ) HDL FLVN & ISO () insulin & HOMA-IR ANTHO () BMI & TG ISO () FBG, insulin, HOMA-IR & TG Lignan () TG, (þ) HDL, & () MetS incidence ISO () WC & TG, (þ) HDL Lignan (þ) HDL, () TG TOTFL, ANTHO, FLVN () IL-18 FLVONL () sVCAM-1

NHANES [66]

28 y (mean) N ¼ 299 pregnant

ISO

GT, lipids

NHANES [59]

20 y N ¼ 1748

ISO, lignan

MetS

NHANES [65]

20 y N ¼ 1492 43e70 y N ¼ 2115 women

Lignan

HDL, TG

TOTFL, FLVONL, FLVN, FLVNN, FLV3L, ANTHO, PRO TOTFL, lignan

Infla, AM

NHS [72]

Moli-Sani [73] TOSCA.IT [67]

35 y N ¼ 11,913 50e75 y N ¼ 2573 w/T2DM

Significant Outcomes

Total PP, TOTFL, phenolic acids

Infla Lipids, BP, CRP, HbA1c

TOTFL, lignan () infla score Total PP () LDL, TG, BP, HbA1c, CRP, (þ) HDL TOTFL () LDL, TG, HbA1c, BP, CRP, (þ) HDL Continued

276

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

Table 10.3 Cardiovascular Disease Biomarkers: Evidence From Observational Studiesdcont’d Country

Cohort

Population

PP/Food

End Point

Brazil

Health Survey of Sao Paulo [63]

12 y N ¼ 557

Coffee (phenolic acids)

Lipids, BP, FBG, HCY

Spain

PREDIMED [61]

55e80 y N ¼ 573 55e80 y N ¼ 1000 56 y (mean, w/o MetS), 60 y (mean, w/MetS) N ¼ 502 women w/T2DM 10e84 y N ¼ 2618

Total PP Resveratrol

Lipids, BP, FBG Lipids, FBG

FLVNN

Lipids

Total PP, TOTFL, phenolic acids, stilbene, lignan

MetS

TOTFL, ANTHO, FLVONL, FLVN, ISO, stilbene TOTFL, FLVNN, ANTHO, FLV3L, PRO, FLVONL, FLVN

BP, FBG, lipids, IMT

PREDIMED [68] Korea

[64]

Iran

TLGS [60]

China

[62]

35e70 y N ¼ 1393

United Kingdom

TwinsUK [69]

18e75 y N ¼ 1898

BP, AS, IMT

Significant Outcomes Phenolic acids () LDL, TG, BP, CRP, (þ) HDL Moderate coffee consumption () BP & HCY Total PP () FBG, DBP, TG Resveratrol () FBG & TG FLVNN () TC, LDL, apoB in patients with MetS

TOTFL () WC, TG, FBG, BP; (þ) HDL Stilbene (þ) BP Lignan (þ) TG, FBG ANTHO (þ) HDL in women TOTFL, FLVN, FLVONL () TG in women ANTHO () pSBP, cSBP, MAP, PWV FLVN () PWV

(), Inverse association or decrease; (þ), positive association or increase; AM, adhesion molecules; ANTHO, anthocyanin; ApoB, apolipoprotein B; AS, arterial stiffness; BMI, body mass index; BP, blood pressure; CARDIA, Coronary Artery Risk Development in Young Adults; CRP, C-reactive protein; cSBP, central systolic blood pressure; DBP, diastolic blood pressure; FBG, fasting blood glucose; FLV3L, flavanol; FLVN, flavone; FLVNN, flavanone; FLVONL, flavonol; GT, glucose tolerance; HbA1c, hemoglobin A1c; HCY, homocysteine; HDL, high-density lipoprotein; HOMA-IR, homeostatic model assessment-insulin resistance; IL-18, interleukin-18; IMT, intima-media thickness; infla, inflammation; ISO, isoflavone; JPHCBC II, Japan Public Health Center-Based Cohort II; LDL, low-density lipoprotein; MAP, mean arterial pressure; MetS, metabolic syndrome; NHANES, National Health and Nutrition Examination Survey; NHS, Nurses’ Health Study; PP, polyphenol; PREDIMED, Prevention with Mediterranean Diet; PRO, proanthocyanidin; pSBP, peripheral systolic blood pressure; PWV, pulse wave velocity; SBP, systolic blood pressure; sVCAM, soluble vascular cell adhesion molecule 1; T2DM, type 2 diabetes mellitus; TG, triglycerides; TLGS, Tehran Lipid and Glucose Study; TOSCA.IT, Thiazolidinediones or Sulfonylureas and Cardiovascular Accidents Intervention Trial; TOTFL, total flavonoid; TwinsUK, TwinsUK registry; w/, with; w/o, without; WC, waist circumference; WWEIA, What We Eat in America.

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

277

In the two studies strictly applying dark chocolate as an intervention, both used white chocolate as a control. Grassi et al. [85] used a crossover trial design in which hypertensive subjects with impaired glucose tolerance were randomized into groups consuming 100 g dark chocolate per day for 15 days or the white chocolate control with a 7-day washout. At the end of the trial, the subjects saw significantly improved endothelial function, lipid profile, and insulin sensitivity after consuming dark chocolate. Taubert et al. [88] used a randomized, placebo-controlled trial design in which each hypertensive subject consumed either 6.3 g dark chocolate per day for 18 weeks or white chocolate as a control. Neither study groups saw a significant change in lipid profile, FBG, or 8-isoprostane, but significantly improved BP and S-nitrosoglutathione (GSNO), a key regulator in BP [88].

10.4.3.3 Soy Five studies [89e93] assessed the role of soy products, an abundant source of isoflavone, in modulating CVD biomarkers, each study recruiting postmenopausal women as at least part of the study population. Four studies [89e91,93] intervened with soy products replacing non-soy protein in the diet. Jenkins et al. [91] evaluated the effects of isoflavone on various CVD biomarkers in 41 men and postmenopausal women with hypercholesterolemia in a randomized crossover design. Each subject followed three different diets, each for 1 month: a high-isoflavone soy diet (73 mg isoflavone/d), a lowisoflavone soy diet (10 mg isoflavone/d) to control for the effects of non-isoflavone soy constituents, and a low-fat dairy control diet. Both soy diets significantly decreased lipid profile and 10-year risk for coronary artery disease (CAD), calculated by using systolic BP and the ratio of total cholesterol to HDL in the Framingham coronary risk equation [94], compared with the control; the high-isoflavone soy diet further decreased LDL and apoB, suggesting that both soy and isoflavone provide beneficial effects for CVD health. Azadbakht et al. [90] randomly assigned 42 postmenopausal women to replace red meat in their diets with either soy protein (84 mg isoflavone/d) or soy nuts (102 mg isoflavone/d) for 2 months, and then switched groups in a crossover period to follow the Dietary Approaches to Stop Hypertension (DASH) diet as a control. It has been postulated that components of soy nuts, such as unsaturated fatty acids and isoflavone, provide more benefits than soy protein in preventing CVD [90]. Both soy diets improved lipid profile and glucose sensitivity biomarkers compared with the DASH diet; however, the soy nuts diet provided greater magnitudes of improvements than the soy protein diet [90]. In another study, when 30 g soy protein, equivalent to a standard portion of meat, replaced animal protein and was compared with a control diet for 12 weeks in a randomized, parallel-arm trial, subjects consuming the soy protein saw reductions in weight, cholesterol, and lipoproteins [93]. Notably, the subjects in that study were all overweight and met at least three criteria for metabolic syndrome at baseline. In the fourth study, Acharjee et al. [89] evaluated the effects of an intervention intended to reduce cholesterol, the Therapeutic Lifestyle Changes (TLC), which encourages low saturated fat and cholesterol consumption with high dietary fiber intake, paired with a daily 25-g (101 mg isoflavone) soy nut replacement of non-soy protein, in 60 postmenopausal women for 8 weeks. Results were stratified into equol-production status and prevalence of metabolic syndrome. Equol is an estrogen metabolite formed from the isoflavone daidzein by intestinal bacteria and has been postulated to inhibit CVD progression by inhibiting inflammation and oxidation [95]. It appears that only a portion (30%e50%) of individuals are capable of producing equol through this mechanism [96]. The

278

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

substitution of soy nuts appeared to improve BP, triglycerides, CRP, and soluble intercellular adhesion molecule-1 over the TLC-alone group among subjects with metabolic syndrome. In subjects without metabolic syndrome, the beneficial effects were not as pronounced. All of the beneficial effects observed, independent of metabolic syndrome prevalence, were significant in equol producers; non-equol producers observed no changes in measured biomarkers, which suggests that equol production is a critical mechanism in receiving cardiovascular benefits from soy in postmenopausal women [93]. One study [92] provided the intervention as a supplement rather than a replacement for protein already in the diet. Seventy grams of soy nuts (101 mg isoflavone) were provided as a snack for 4 weeks in hypertensive men and postmenopausal women at high risk for metabolic syndrome, and participants crossed over to a low-isoflavone control snack. Subjects consuming the soy nuts saw improvements in peripheral arterial tonometry, a measure of arterial stiffness, and FBG, but no changes in serum insulin or inflammatory biomarkers.

10.4.3.4 Red Wine Five studies [97e101] analyzed the effect of consuming red wine, a rich source of resveratrol, flavonoids, and phenolic acids, on a range of CVD biomarkers, with inconsistent results. Two studies [97,100] used a crossover design with red wine, dealcoholized red wine, and a control beverage as intervention arms. Naissides et al. [100] provided 45 postmenopausal, hypercholesterolemic subjects with 400 mL red wine (1 g total polyphenol), equivalent amounts of dealcoholized red wine with an identical polyphenol profile, and water for 6 weeks each. The red wine intervention saw improvements in LDL and HDL compared with the other two interventions, which suggests that alcohol may have a synergistic effect with the polyphenols present in red wine to induce beneficial lipid-altering effects in this population [100]. Chiva-Blanch et al. [97] carried out a similar study, except that they used 272 mL red wine (30 g ethanol) and dealcoholized red wine, with 100 mL gin (30 g ethanol) as the control, for 4 weeks each in 67 subjects at high CVD risk. Only the dealcoholized red wine intervention significantly improved BP and NO production; lipids were not measured. Because the gin intervention saw no changes in BP or NO, these findings suggest that alcohol by itself does not raise BP; rather, it may impair polyphenols’ vasodilatory capabilities in subjects with high CVD risk [97]. One study [98] applied a 2  2 factorial design with the presence and absence of lifestyle counseling (based on the MED with moderate physical activity) and red wine as variables in subjects with carotid atherosclerosis. The amount of red wine provided was less than in the previous two studies mentioned, giving 200 mL/d to men and 100 mL/d to women, which are closer to a typical serving size of wine than in the previous studies. When stratified by variable, the effect of red wine did not significantly alter lipid profile with the exception of favorably improving the LDLeHDL ratio [98]. Although this study showed no effect on lipid profile, a 2-week red wine intervention study (250 mL red wine twice daily versus water) performed in hospitalized postmyocardial infarction patients showed that red wine beneficially altered total cholesterol, LDL, and erythrocyte membrane fluidity [101]. The red wine group also saw improved antioxidant capacity; however, there were no significant differences in CRP, BP, body weight, or FBG between the groups. One study [99] found no significant difference in endothelial function between red wine (300 mL/d, 540 mg total polyphenol) and polyphenol-poor white wine (300 mL/d, 75 mg total polyphenol) in 20 healthy subjects after 15 days.

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

279

10.4.3.5 Fruit/Juice Eight studies [102e109] assessed the cardiometabolic effects of berries, grapes, apples, or fruit juices. Fruits, particularly berries, are rich sources of the flavonoid anthocyanin. Three randomized controlled studies [102,106,109] provided blueberries (BB) as an intervention in the form of a smoothie or beverage in subjects with elevated CVD risk, with mixed results. Two studies [102,106] found that BB lowered BP whereas one found no effect [109]. In the 50 -g BB/d, 8-week intervention of Basu et al. [10], the BB group saw significant improvements in antioxidant function in obese subjects with metabolic syndrome compared with control, whereas in the 22 -g BB/d, 8-week intervention of Johnson et al. [106], no significant changes in antioxidant function were observed, but measurements of arterial stiffness improved in hypertensive postmenopausal women. Insulin sensitivity improved in the BB intervention study performed in obese, insulin-resistant subjects receiving 45 -g BB smoothies daily for 6 weeks [109]. Lipid profile did not differ between the BB and control groups in the two trials that measured it [102,109]; likewise, no differences in inflammation biomarkers were noted in all three studies that assessed these markers [102,106,109]. One randomized, placebo-controlled study [105] recruited 72 middle-aged subjects with high risk for CVD and used a variety of berries and berry juices as an intervention, alternating daily between 100 g bilberries plus 50 g lingonberries with 100 g black currant or strawberry puree with 0.7 dL chokeberry and raspberry juice. After 8 weeks, the berry intervention group observed significantly more improvements in platelet function, HDL, and systolic BP compared with the control group, which received one of four energy-matched products on each day including sugar water, semolina porridge, sweet rice porridge, and marmalade sweets. Three crossover interventions [104,107,108] used fruit juice as an intervention vehicle. Antioxidant enzyme activity and capacity, but not inflammatory biomarkers, were significantly improved with 14 days of grape and bilberry consumption in 30 healthy women [107]. Hesperidin is a major polyphenol present in oranges and orange juice. Four weeks of orange juice consumption (500 mL daily) did not improve plasma antioxidant capacity or biomarkers of inflammation, endothelial function, or glucose tolerance, but it improved diastolic BP in 24 healthy overweight men [108]. This study used a crossover design with orange juice, a control juice with supplemental hesperidin, and control juice with a placebo (starch) as the intervention arms. Because diastolic BP improved in both the orange juice and control juice with hesperidin groups and did not change in the control juice with placebo group, the authors posited that hesperidin is likely the primary component in orange juice responsible for favorably altering BP [108]. The third crossover trial found that 4 weeks of daily cranberry juice consumption (480 mL/d) improved arterial stiffness and vascular function in subjects with prevalent CAD [104]. One crossover trial [103] assessed the relation between whole and peeled apples, a rich source of flavonoids and anthocyanins, and various CVD biomarkers, including BP, FMD, FBG, lipids, F2e isoprostane, NOe 3 , and NO2 . The whole apple, with the flavonoid-rich peel intact, improved FMD compared with peeled apple in subjects at high CVD risk and significantly improved the other markers examined.

10.4.3.6 Tea Two studies [110,111] assessed the relation between the consumption of tea, a rich source of the flavonoid catechin, and CVD biomarkers. Twelve weeks of goishi tea consumption (400 mL/d, 122 mg total polyphenol) in Japanese subjects with elevated LDL led to improved HDL, but only in subjects

280

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

with a BMI of less than 25; no change in BP was observed [111]. BP was significantly lowered, however, in response to daily black tea consumption (300 mg total polyphenol/d) for 8 days in hypertensive Italian subjects; in addition, arterial stiffness improved [110].

10.4.3.7 Current Evidence at a Glance A total of 32 intervention studies (Table 10.4) assessed the effect of polyphenol-rich foods or drinks on CVD biomarkers. Study lengths ranged from 1 day to 1 year. The crossover trial design was frequently employed because subjects served as their own control, reducing variability, and to allow for a smaller sample size. Chocolate and fruit or juice were the foods and drinks most frequently studied and most studies recruited subjects at high CVD risk. Polyphenol-rich diets, including the MED [73], consistently improved biomarkers such as lipids, FBG, and BP. Chocolate appears to improve BP and vascular function, and berries appear to increase plasma antioxidant capacity. Overall, the experimental diets produced more favorable results compared with control diets, although results were not entirely consistent.

10.4.4 BIOLOGICAL FUNCTIONS OF POLYPHENOLS IN CARDIOVASCULAR DISEASE PREVENTION 10.4.4.1 Atherosclerotic Pathophysiology Polyphenols primarily exert cardioprotective effects by interrupting the progression of atherosclerosis, a chronic inflammatory condition characterized by the buildup of fat deposits (plaque) in the arteries. Atherosclerosis develops in three main steps: endothelial dysfunction initiated by decreased NO bioavailability, oxidation of LDL and its uptake by macrophages forming an initial lesion, and the maturation of the lesion with plaque rupture [112]. First, NO bioavailability is inhibited. NO is a key signaling molecule that regulates the vascular tone by modulating vasoconstrictors such as acetylcholine, endothelin-1, and thromboxane A2 [112], and is used by the endothelium to relax surrounding smooth muscle cells, promoting vasorelaxation [7]. An imbalance between the oxidative environment and antioxidant defense systems promotes the production of reactive oxygen species (ROS), degrading NO. Enzymes promoting the production of ROS consequently decrease NO bioavailability. A deficiency in endothelial NO synthase (eNOS), the enzyme capable of synthesizing NO, can result in decreased NO availability [113]. Second, LDL becomes oxidized in prooxidative environments. The accumulation of oxidized LDL (LDLox) in the endothelium promotes proinflammatory cytokines and chemokines such as macrophage colony stimulating factor and macrophage chemoattractant protein-1 (MCP-1) and increases the expression of cell adhesion molecules such as intercellular adhesion molecule 1 (ICAM-1) and vascular cell adhesion protein 1 (VCAM-1) [114]. Monocytes migrate to the intima and differentiate into macrophages, taking up the LDLox and becoming foam cells, contributing to the development of an initial lesion [112]. Third, foam cells necrotize and undergo apoptosis, releasing their fatty contents and generating a large lipid core. This induces an inflammatory response, prompting vascular smooth muscle cells (VSMC) to proliferate and migrate to the intima, forming a fibrous cap around the lipid core and stabilizing the lesion [112]. When the collagen in the fibrous cap degrades, for example, by the action of matrix metalloproteinase (MMP), the plaque can rupture [115]. This activates the aggregation of platelets, potentially leading to clogging of the artery [112].

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

281

Table 10.4 Cardiovascular Disease (CVD) Biomarkers: Evidence From Intervention Studies Significant Outcomes

Author

PP/Food

Subjects

Intervention

Endpoint

[76]

PP-rich diet

N ¼ 86 High CVD risk, overweight/ obese, multiple MetS criteria

Lipids, 8-iso

Diets rich in PPs () TG, HDL, VLDL, 8-iso

[77]

PP-rich diet

N ¼ 86 High CVD risk, overweight/ obese, multiple MetS criteria

GT

Diets rich in PPs () FBG; (þ) insulin

[79]

PP-rich diet

N ¼ 93 Stage I/II HTN

Vaso, BP, lipids

[78]

PP-rich diet

N ¼ 200 High CVD risk

8 w 2  2 factorial: diets poor in LCn3 and PPs, rich in LCn3, rich in PPs, rich in LCn3 and PPs. 8w 2  2 factorial: diets poor in LCn3 and PPs, rich in LCn3, rich in PPs, rich in LCn3 and PPs, OGTT after 8w 12 w high PP diet (6 portions F and V, incl 1 portion berries) þ 50 g/ d (35.7 mg epicatechin) DC versus low PP diet 1y Med þ EVOO versus Med þ nuts versus low-fat control diet

[86]

Cocoa

N ¼ 90 Elderly, w/o cognitive dysfunction

8 w daily cocoa drink: high FLV3L (993 mg) versus intermediate FLV3L (520 mg) versus low FLV3L (48 mg)

GT, lipids, 8-iso, BP

[87]

Cocoa

N ¼ 10 Healthy, young adults

Vascular

[82]

Cocoa

N ¼ 21 Healthy, overweight

1d XO: high FLV3L cocoa drink (300 mL, 917 mg FLV3L), low FLV3L drink (37 mg FLV3L) 1d XO (repeated once): high FLV3L (701 mg) cocoa drink, low FLV3L (22 mg) drink

High PP diet () TC, (þ) max % response to acetylcholine Both diets () BP, n.s. between diets Med þ EVOO and MED þ nuts (þ) NO; () BP compared with control High and intermediate FLV3L groups (þ) HDL; () BP, FBG, insulin, HOMAIR, TC, LDL, TG, 8-iso compared with low FLV3L High FLV3L drink (þ) FMD, PAT and NO metabolites

NO, BP

BP, FMD

High FLV3L cocoa drink favorably improves BP and FMD Continued

282

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

Table 10.4 Cardiovascular Disease (CVD) Biomarkers: Evidence From Intervention Studiesdcont’d Author

PP/Food

Subjects

[81]

Cocoa

N ¼ 10 (feasibility study) T2DM, N ¼ 41 (efficacy study) T2DM

[85]

Dark chocolate

N ¼ 19 HTN and impaired GT

[84]

Dark chocolate

N ¼ 45 Healthy

[88]

Dark chocolate

N ¼ 44 Pre-HTN or stage I HTN

[83]

Dark chocolate

N ¼ 100 Healthy, older adults

[91]

Soy

N ¼ 41 Elevated LDL, Men and PM

Intervention

Feasibility (XO) study: cocoa drink (75, 317, 963 mg FLV3L) Efficacy (2-arm) study: cocoa drink 3 times/d (321 mg each) versus control 15 d XO: FLV3Lrich DC (100 g/d), FLV3L-free white chocolate Phase 1: 1 d XO: solid DC (821 mg TOTFL), cocoa-free placebo Phase 2: 1 d XO: sugar-free cocoa drink (805 mg TOTFL), sugared cocoa drink (805 mg TOTFL), placebo drink 18 w daily DC (6.3 g, 30 kcal, 30 mg PP) versus control (white chocolate) 6 w daily DC (37g, 397.3 mg total PRO/g) þ 237 mL cocoa beverage (357.4 mg total PRO/g) versus placebo 1 m XO: high ISOsoy food diet (50 g soy protein, 73 mg ISO/d), low ISOsoy food diet (52 g soy protein, 10 mg ISO/d), low-fat dairy control

Endpoint

FMD

BP, FMD, lipids, HCY, GT

BP, FMD

Significant Outcomes response to exercise Cocoa (þ) FMD in dosedependent response

DC () BP, TC, LDL, HOMAIR; (þ) FMD, ISI, QUICKI, CIR120 Solid DC, sugar-free cocoa drink, and sugared cocoa drink (þ) FMD Solid DC and sugar-free cocoa drink () BP

Lipids, BP, FBG, 8-iso, GSNO

DC () BP; (þ) GSNO

Lipids, CRP, BP

n.s.

Lipids, BP, HCY, 10y CAD risk

High ISO-soy food diet () TC, LDL, apoB, 10y CAD risk Low ISO-soy food diets () TC, 10 y CAD risk

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

283

Table 10.4 Cardiovascular Disease (CVD) Biomarkers: Evidence From Intervention Studiesdcont’d Author

PP/Food

Subjects

Intervention

Endpoint

[90]

Soy

N ¼ 42 PM w/o MetS

2 m XO: soy protein diet (84 mg ISO), soy nut diet (102 mg ISO), control (DASH diet)

MetS, GT

[93]

Soy

N ¼ 53 Overweight, MetS, elevated LDL, men and PM

MetS

[89]

Soy

N ¼ 60 PM

12 w whole soy food diet (30 g/ d soy protein substituting animal protein) versus control diet 8w XO: TLC diet þ 0.5 cup/d soy nuts (25g soy protein, 101 mg ISO) replacing 25 g non-soy protein, control diet (TLC diet w/o soy nuts)

[92]

Soy

4w XO: 70 g soy nuts (101 mg ISO), control snack

PAT, FBG, insulin, infla

[98]

Red wine

N ¼ 17 Elevated BP and  1 MetS criteria, Men and PM N ¼ 108 Carotid atherosclerosis

20w 2  2 factorial: no lifestyle counseling (1 glass RW versus no alcohol) versus lifestyle counseling (1 glass RW versus no alcohol)

Lipids

BP, lipids, infla, FBG, CRP, AM

Significant Outcomes Both soy diets () LDL, TC, insulin, HOMAIR; soy nut diet also () FBG, and had more dramatic effects than soy protein diet Soy diet () weight, BMI, TC, LDL, apoB, lipoprotein (a)

In equolproducers w/MetS, TLC diet þ daily soy nuts () DBP, TG, CRP, sICAM-1 versus TLC diet (n.s. in non-equol producers) In equolproducers w/o MetS, TLC diet þ daily soy nuts () SBP, DBP, CRP (n.s. in non-equol producers) Soy nuts () PAT, FBG

Lifestyle counseling () LDL, TC and TG RW () LDL: HDL ratio

Continued

284

CHAPTER 10 POLYPHENOL-RICH DIETS IN CARDIOVASCULAR DISEASE

Table 10.4 Cardiovascular Disease (CVD) Biomarkers: Evidence From Intervention Studiesdcont’d Author

PP/Food

Subjects

Intervention

Endpoint

[101]

Red wine

N ¼ 39 Hospitalized, post-MI

2w RW (250 mL 2 times/d) versus water

[99]

Red wine

N ¼ 20 Healthy

[100]

Red wine

N ¼ 45 High cholesterol, PM

[97]

Red wine

N ¼ 67 High CVD risk

[102]

Blueberries

N ¼ 48 MetS, obese

[106]

Blueberries

N ¼ 48 Pre- and stage I HTN, PM

[109]

Blueberries

N ¼ 32 Obese, IR, nondiabetic

[107]

Grape and bilberry

N ¼ 30 Healthy women

[105]

Berries

N ¼ 72 High CVD risk

15d RW (300 mL/d, 1.8 g/L total PP) versus WW (0.25 g/L total PP) 6w XO: 400 mL RW (1 g total PP), dealcoholized RW (identical PP profile), water 4 w XO: 272 mL RW (30g EtOH, 789 EGA/d), 272 mL dealcoholized RW, 100 mL gin (30 g EtOH) 8w BB beverage (50 g freeze-dried BB) versus control beverage (960 mL water) 8 w freeze-dried BB powder (22 g, 845 mg total PPs) versus control powder 6w BB smoothie (22.5 g powder 2 times/d, 1462 mg total PPs/d) versus control smoothie 14d XO: 330 mL juice (984 mg/L ANTHO), 330 mL smoothie (841 mg/ L ANTHO), 330 mL placebo (9 mg/L ANTHO) 8w 2 portions/ d berries versus energy-matched control

Lipids, CRP, AO, membrane erythrocyte properties, FBG, weight, BP Platelet e NOe 2 þ NO3

Lipids, GT

BP, NO

Significant Outcomes RW () TC, LDL; (þ) TEAC and erythrocyte membrane organization Platelet e NOe 2 þ NO3 (þ) in both groups RW () LDL; (þ) HDL compared with dealcoholized RW and water Dealcoholized RW () BP; (þ) NO

Lipids, GT, BP, weight, WC, CRP, AM, infla, AO

BB () BP, LDLox, MDA, HNE

BP, MAP, VF, CRP, NO, SOD

BB () BP, brachial-ankle PWV; (þ) NO

Weight, BMI, BF %, BP, FBG, serum insulin, lipids, infla

BB (þ) IS

Infla, AO

Juice and smoothie (þ) SOD, CAT, TEAC; () TBARS

Platelet, BP lipids

Berry group (þ) CADP-CT, HDL; () SBP

10.4 DIETARY POLYPHENOLS AND CARDIOVASCULAR DISEASE

285

Table 10.4 Cardiovascular Disease (CVD) Biomarkers: Evidence From Intervention Studiesdcont’d Author

PP/Food

Subjects

Intervention

Endpoint

Significant Outcomes

[108]

Orange Juice

N ¼ 24 Healthy overweight men

Infla, EF, BP, GT, lipids, AM, FRAP

OJ and hesperidin groups () DBP

[104]

Cranberry Juice

N ¼ 15 (acute pilot study) CAD

4 w XO: 500 mL OJ, 500 mL control drink þ hesperidin, 500 mL control drink þ placebo 4w XO: 480 mL/ d CJ (835 mg total PP, 94 mg ANTHO), placebo

VF, AS, lipids, GT

CJ () PWV; (þ) FMD, PAT

BP, FMD, AS, FBG, lipids, F2-iso, NO-3, NO-2

Apple w/skin (þ) FMD

Lipids, BP

Goishi tea (þ) HDL in subjects w/BMI 30 mL/min/1.73 m2). The data indicated a higher CKD risk in the lowetotal water intake (TWI) group (4.3 L/d; OR ¼ 2.52; CI, 0.91e6.96) [18]. Also, plain water intake was advantageous in protecting against CKD (OR ¼ 2.36; CI, 1.10e5.06) compared with water intake from beverages other than plain water (OR ¼ 0.87; CI, 0.30e2.50). Urinary output is also an important measure when evaluating kidney function because it indicates the proper dynamics of fluid in the body. Decreased GFR is correlated with urine osmolality over 300 mOsm/kg in rats [19]. Anastasio and colleagues examined renal function in humans with low (0.5 mL/kg) versus high (4 mL/kg) water intake every 30 min. They reported that high water intake was associated with higher GFR and natriuresis in both the fasting and postprandial conditions [20].

11.2.3 ARGININE VASOPRESSIN AND CHRONIC KIDNEY DISEASE Arginine vasopressin (AVP) also has an important role in kidney health. When the body experiences fluid loss, plasma osmolality increases, stimulating the secretion of AVP and water reabsorption in the kidney. As AVP levels increase and bind to vasopressin 2 receptors (V2R) in the principal cells of the thick ascending collecting duct, the permeability of water, sodium, and urea increases. This process specifically occurs via responses seen in aquaporin 2erich vesicles, luminal sodium channels, and urea transporters (UT) (UT-A1 and UT-A3) [21e23]. A combination of these responses can inhibit kidney function, leading to less efficient excretion of solutes and potentially resulting in hyperfiltration, albuminuria, or salt-sensitive high blood pressure [23,24]. Increased levels of copeptin, a surrogate marker of AVP, can also damage the kidney via increased vasoconstriction, water absorption,

11.2 HYDRATION AND KIDNEY HEALTH

303

renineangiotensinealdosterone system (RAAS) activity, glomerular pressure, and insulin resistance [25]. Roussel and colleagues [26] examined copeptin in people with CKD (n ¼ 83) using data from the Epidemiological Study on Insulin Resistance Syndrome and found that higher levels of either plasma copeptin or plasma osmolality were associated with decreasing eGFR.

11.2.4 MESOAMERICAN NEPHROPATHY Data indicated that sugarcane harvesters in Central America are at higher risk for heat stress because workloads exceed four times the amount recommended by the Occupational Safety and Health Administration (OSHA). Environmental conditions for this specific population are extremely stressful, with wet bulb globe temperatures easily exceeding the value OSHA considers safe [27]. This population may also have a high probability of low-grade hematuria and/or mild rhabdomyolysis, considering reported levels of chronically concentrated urine most likely owing to poor hydration habits [28]. Using renal biopsies, extensive glomerulosclerosis and signs of chronic glomerular ischemia along with tubular atrophy and interstitial fibrosis were observed in Mesoamerican populations [29]. These adverse events can be observed in the kidney with increasing AVP levels and hyperuricemia [30] caused by recurrent dehydration via a combination of strenuous work and insufficient rehydration with high environmental temperatures [31]. Mesoamerican nephropathy is exacerbated by ingesting concentrated juices for rehydration after work, which activates aldose reductase. Activation of this system leads to endogenous fructose production and metabolism, increasing uric acid production and ultimately oxidative stress. Combined with ingestion of beverages containing fructose, this process yields increased kidney stress. One animal study examined the effect of sweetened beverage intake as a rehydration technique on kidney health compared with plain water consumption [32]. Dehydration (36 C) was induced in Wistar rats for 1 h/day followed by 2 h rehydration with either tap water or 11% fructoseeglucose (FG) solution, a composition often seen in soft drinks. This study found that after 4 weeks of treatment, the 11% FG solution group resulted in greater dehydration associated renal damage accompanied by higher osmolality and copeptin levels and activated aldose reductase-fructokinase.

11.2.5 OTHER KIDNEY DISEASES The probability of kidney stone recurrence is approximately 50% [33,34] and is a contributory factor to CKD [35]. Urolithiasis is mostly driven by genetics [36]; however, low fluid intake also contributes to kidney stone development [37]. Water intake of 2 L/day has been reported to have a preventive effect on kidney stone formation [38]. The American Urological Association also recommends drinking more fluids to achieve a urine volume of at least 2.5 L/day to reduce the probability of recurrence [39]. A metaanalysis investigated the doseeresponse relationship of self-fluid management in preventing kidney stones and showed that the risk decreased significantly for each 500-mL increase in water intake (P < .01) [40]. Similar data from the Nurses’ Health Study reported that kidney stone risk was reduced with TFI exceeding 1850 mL/day (P < .001). Using receiver operating characteristics, researchers established an optimal urine osmolality cutoff of 525 mOsm/kg (area under the curve ¼ 0.92) [37,41]. Another study also found that higher urine osmolality (569 mOsm/kg versus 489 mOsm/kg; P ¼ .02) in participants was associated with stone formation [42]. The potentially

304

CHAPTER 11 HYDRATION AND HEALTH

preventive role of fluid intake on urinary tract infection was also investigated in women; however, its effectiveness remains controversial [43,44].

11.2.6 WATER INTAKE INTERVENTION IN CHRONIC KIDNEY DISEASE Owing to reported associations between TFI and CKD risk, recommendations have been made to reach a certain amount of urine output as a potential preventive method [39,45,46]. To investigate the effectiveness of water intake on CKD, Clark and colleagues published preliminary data on a 1-year randomized controlled trial in patients with kidney disease (stage 3 CKD and microalbuminuria) [13,47]. Participants in this study were asked to increase water intake by 1e1.5 L/day, depending on sex and weight. The researchers observed decreased 24-h urine osmolality and plasma copeptin as well as increased urinary volume. Considering the potential preventive effect of increased urinary output discussed earlier, this protocol could serve as a prospective intervention. These results were consistent with a previous 7-year follow-up study examining the effect of urine volume on kidney function. In this prospective cohort study of 2148 subjects, the decline in eGFR was significantly higher in the group with 24-h urine volume less than 1 L/day compared with the group with urine volume greater than 3 L/day (age  18; eGFR  60 mL/min/1.73 m2) [48]. As shown earlier, the body of literature suggests a preventive effect of water and/or fluid intake on kidney health. Water intake is a lifestyle behavior that is cost-efficient with no negative effects compared with other beverages [1] that can be improved with simple intervention [49]. Thus, water intake is recommended for proper hydration status and potentially to prevent kidney disease.

11.3 HYDRATION AND GLUCOSE REGULATION 11.3.1 INTRODUCTION Diabetes is the seventh leading cause of death in the United States; annually, it kills more Americans than HIV and breast cancer combined. In 2015, 30.3 million Americans (9.4% of the population) had diabetes, whereas 84.1 million qualified as prediabetic, presenting with elevated fasting blood glucose levels. Type 1 diabetes is caused by a malfunctioning pancreas from birth and manifests in children and adolescents, but it affects merely 3% of the US population. However, most diabetic patients today have type 2 diabetes mellitus (T2DM), a gradual deterioration of a formerly functional pancreas, caused by the development of insulin resistance at the level of the target cells [50]. Persisting increased glucose levels in the plasma, referred to as hyperglycemia, trigger heightened rates of insulin production and release from the beta cells, which ultimately undergo exhaustion and decay. It is therefore notable that this digression in glucose clearance is acquired and not a genetic defect. The onset of impaired glucose tolerance is the reference standard for the diagnosis of prediabetes.

11.3.2 IMPLICATIONS OF ARGININE VASOPRESSIN IN GLUCOSE REGULATION As the research on T2DM and its potential risk factors continues, epidemiological studies from Europe have linked low water intake levels with the incidence of insulin resistance and metabolic syndrome [51]. Interestingly, high water intake levels were strongly correlated with healthy blood glucose ranges in a study reporting on a British cross-sectional analysis [52,53] and classified elevated copeptin levels

11.3 HYDRATION AND GLUCOSE REGULATION

305

as an independent predictor variable for increased risk of developing T2DM [54]. Plasma copeptin is a stable surrogate marker of the unstable fluid regulatory hormone AVP. It is released along with AVP from the carrier protein neurophysin upon stimulus [55]. Considering the strong correlation with AVP release in the blood (r ¼ 0.8), copeptin has therefore been widely used to quantify AVP responses in cross-sectional and epidemiological studies. AVP, also known as antidiuretic hormone, is secreted by the posterior pituitary. The main stimulus for AVP secretion is a rise in plasma concentration (osmolality), which is often seen during periods of high salt intake as well as low water intake and acute water loss. However, AVP release also responds to changes in circulating blood volume and hypotension [56], as well as nausea, vomiting, pain, certain drugs, and insulin-induced hypoglycemia [57]. Once osmoreceptors are stimulated, AVP is released from the hypothalamus and signals water reabsorption in the renal filtration ducts. This results in water retention in the vasculature while fluid losses via the urinary tract are reduced. Interestingly, beyond its role of regulating fluid balance at the level of the kidneys, AVP has many effects on other parts of the body. More specifically, AVP has been shown to act upon three types of AVP-specific receptors: V1aR, which is located throughout the body; V1bR, which is located in the pituitary gland and pancreas; and V2R, which is highly expressed in the renal collecting ducts [58]. Mechanistic animal studies have shown V1a receptors to mediate vasoconstriction and glycogenolysis in the liver, whereas V1b receptors have been linked to secretion of insulin and glucagon in the pancreas, as well as adrenocorticotropin hormone (ACTH) release from the pituitary. V2R have been reported exclusively in the renal collecting ducts, where they regulate body fluid homeostasis via a cyclic adenosine monophosphateeaquaporin 2 channel pathway [58]. To investigate further the potency of AVP levels on glucose regulation, a rodent model compared the effect of three different chronic AVP concentrations on subsequent glucose tolerance in lean and obese Zucker rats, which are prone to developing metabolic dysfunction with age [59]. AVP levels were kept chronically high via infusion, low by increasing the daily water intake of animals via administration of aqueous gel pellets, or physiologically normal by not interfering with the animals’ daily water intake for control. Results showed that when chronically infused with high AVP levels, even lean animals showed an increase in fasting glycemia after 4 weeks of intervention. Interestingly, obese animals that had hyperinsulinemia and glucose intolerance showed improved glucose handling when V1aR were blocked by a specific antagonist. As a final result of this study, chronically low AVP levels had a protective effect on liver tissue in obese animals by decreasing hepatic steatosis. After these findings, the first human intervention study was conducted in 2016 by Johnson et al. To assess the acute effects of low water intake on glucose regulation in patients with diabetes, nine male T2DM patients were recruited to undergo an oral glucose tolerance test (OGTT) administered in either a wellhydrated or hypohydrated state in a randomized crossover design [60]. A markedly higher level of plasma glucose was measured during the water-restricted trial. Interestingly, cortisol was significantly elevated during the OGTT in the hypohydrated trial, whereas no differences in aldosterone, plasma renin activity, or angiotensin II were observed. Although not measured, AVP is involved in ACTH stimulation via V1bR [58]. This could be a mechanism explaining the higher cortisol levels in the hypohydrated trial when AVP levels are presumably high to induce water retention at the renal collecting ducts. In an epidemiological study, a French cohort of 5110 subjects from the general population was assessed at baseline for copeptin levels and variations in the AVP-neurophysin II gene with follow-up measurements for 9 years examining outcomes of impaired fasting glucose (IFG) and eventually

306

CHAPTER 11 HYDRATION AND HEALTH

stratifying the risk of new-onset T2DM. Findings showed a high association between high copeptin levels and the development of blunted insulin sensitivity indexed by homeostatic model assessment, as well as increased risk for developing IFG and T2DM. In addition, men inheriting certain allelic variations showed significant associations with hyperglycemia incidence [61]. To investigate whether increased water intake had a physiologically significant effect on human plasma copeptin levels, Lemetais et al. grouped 82 healthy adults as low, normal, or high drinkers based on reported fluid. Over the 6-week intervention, participants from the low and normal drinker group increased their water intake to that of the high drinkers. As a result, copeptin levels significantly decreased in both intervention groups compared with baseline. This was one of the first studies to show that a 6-week water intervention had the potency to reduce circulating copeptin levels in healthy humans [62]. In line with this finding, a study from the United Kingdom reported that high plain water intake was associated with lower glycated hemoglobin (HbA1c) in men [53]. Previous literature by Pan et al. [63] suggested that substitution of fruit juices and sugar-sweetened beverages (SSBs) with plain water was associated with lower risk for T2DM. However, Carroll et al. found that replacing juice and SSB with plain water resulted in no further significant improvements in HbA1c levels [53].

11.3.3 OTHER CONSIDERATIONS IN HYDRATION AND GLUCOSE REGULATION Whereas glucose regulation seems to be influenced by AVP via the V1a and V1b receptors, changes in cell hydration state might affect carbohydrate and protein metabolism significantly, especially at the level of the liver [64] and peripheral tissue uptake sites. During hyperosmolarity and cell shrinkage, cell cultures of adipocytes revealed inhibited insulin-dependent glucose uptake and hindered glucose transporter type 4 (GLUT4) translocation, resulting in peripheral insulin resistance. In rat hepatocytes, GLUT4-independent glucose uptake was decreased after an elevation in the osmotic concentration of the interstitial fluid by 10 mOsm/L. In addition, increased plasma osmolality induced glycogenolysis, rendering increased glucose-6-phosphate availability [65] in states of impaired peripheral glucose uptake.

11.4 IMPLICATIONS OF FLUID BALANCE AND OBESITY 11.4.1 INTRODUCTION Obesity has remained a hindering and costly problem in societies across the globe. According to data from NHANES, US obesity (body mass index >29.9) prevalence has reached 35% and 40.4% in men and women, respectively. In men, previous data showed that the 35% prevalence rate remained constant whereas a linearly increasing trend was observed in women [66]. Yearly health care costs attributable to obesity and related issues are projected to reach approximately $900 billion by 2030 [67]. Similar data have been recorded in Mexico from a sample of 2511 individuals with an obesity prevalence of 25% [68]. However, the problem is not isolated to the western hemisphere. As of 2014, the World Health Organization reported data categorizing 39% of women and 38% of men around the globe as overweight (body mass index >24.9) [69]. In 2015, the Global Burden of Disease Obesity Collaborators reported that a total of 107.7 million children and 603.7 million adults were obese, with a higher rate of increase seen in children compared with adults [70].

11.4 IMPLICATIONS OF FLUID BALANCE AND OBESITY

307

11.4.2 SUGAR-SWEETENED BEVERAGES Parallel to the rise in obesity and diabetes since the 1970s, SSB consumption has increased [71]. Although the problem of obesity is multifactorial, SSB consumption could serve as a significant contributing factor. High-fructose corn syrup has been adopted as a common beverage sweetener because of the lower production costs compared with other sweeteners. Considering that an increase in energy intake and any decrease in energy expenditure has the potential to contribute to obesity [72], a product as energy dense as high-fructose corn syrup is an important factor to consider in the daily caloric intake. This is especially important considering that the preprandial consumption of beverages containing extra calories may not reduce the number of calories consumed throughout the actual meal [71,73,74]. Several studies have indicated SSBs to have a reduced effect on satiety, resulting in greater total caloric intake [75]. From sugars alone, one can of soda supplies 150 kcal energy. Over a time span of 10 years, ingesting an additional 150 kcal/day can lead to a 110-lb increase in body mass [72,76,77] Although caloric surplus is a major factor in the development of obesity, fat accumulation may occur with fructose consumption independent of superfluous caloric intake. A study by Schwarz and colleagues indicated that when fructose was substituted for complex carbohydrates, de novo lipogenesis and hepatic fat accumulation decreased even when caloric consumption was held constant [78]. Considering the amount of high-fructose corn syrup in SSBs, it is important to elucidate fructose digestion and absorption and how these processes differ from other forms of sugar. Compared with glucose uptake, which occurs primarily through the protein transporter GLUT4, fructose is transported via GLUT5, which appears abundantly in the liver. Consequently, metabolism of fructose predominantly occurs in the liver. Once transportation is complete, ATP transfers a phosphate to create phosphorylated fructose ready to be broken down by the enzyme aldolase. As a result of this reaction, trioses are formed that contribute to triglyceride formation and potentially increased adiposity over time. In 2009, Stanhope and colleagues found a positive correlation between de novo lipogenesis and the consumption of fructose-containing beverages [79]. Several studies using large sample sizes indicated an association between SSB consumption and obesity-related outcomes, especially in younger populations. Malik and colleagues conducted a systematic review in 2013 exploring this association. After analyzing 32 different studies, the researchers concluded that increased SSB consumption is associated over time with obesity in children and adults. The review included randomized controlled trials measuring weight gain or loss with an increase or reduction of ingesting SSBs [80]. One such study published in the New England Journal of Medicine analyzed body weight in 224 participants before and after an intervention requiring a reduction in SSB consumption. Compared with controls, those who received the intervention had a slower rise in body weight throughout a 1-year period [81].

11.4.3 HYDRATION STATUS AND OBESITY Most data suggesting a link between hydration biomarkers and adiposity are associational [82]. Chang et al. employed body mass index and urine osmolality data from a sample of 9528 NHANES participants. Compared with hydrated individuals (urine osmolality, 40 cups/week) of coffee, tea, and alcohol has been indicated in numerous caseecontrol studies as a risk factor for bladder cancer [125,127,129]. Furthermore, the chronic ingestion of tap water containing potential carcinogens is recognized as a contributing factor [130]. In perhaps the most convincing study, Michaud and colleagues collected food frequency questionnaires from nearly 48,000 individuals over a 10-year span. In this cohort, there were 252 newly diagnosed cases of bladder cancer. Total daily fluid intake was inversely associated with the development of bladder cancer compared with individuals drinking >2.5 L/day and those drinking 97%) of calcium is stored in bone tissue in the form of insoluble hydroxyapatite, providing structural rigidity to the skeleton. The regulation of calcium balance is tightly controlled by PTH and vitamin D, as shown in Fig. 13.1. If serum calcium levels decrease, bone resorption is stimulated by PTH to release stored calcium into the blood. PTH also increases calcium reabsorption in the kidneys and decreases its urinary excretion. Positive effects of calcium supplementation on bone mass are attributed to a reduction in bone remodeling [7]. Because bone is constantly undergoing remodeling, a sufficient supply of calcium is needed throughout the life course to maintain bone integrity. Nutritional requirements vary widely depending on the stage of the life cycle. Requirement for calcium is highest during adolescence, due to accelerated bone growth. In the United Kingdom, the

340

CHAPTER 13 DIET AND BONE HEALTH

Sunlight

Skin

7-Dehydrocholesterol in skin

Vitamin D3 (cholecalciferol) Dietary sources of D2 and D3

Liver

25 (OH) Vitamin D3 formed in liver Parathyroid hormone

Kidney

1,25 (OH)2 Vitamin D3 (calcitriol) formed in kidney

Gut

Increase in absorption of calcium and phosphate from gut

Bone

Increase in renal calcium and phosphate retention

Increase in bone resorption

FIGURE 13.1 Overview of vitamin D metabolism and sites of action.

Reference Nutrient Intake (RNI) is 1000 mg for male adolescents and 800 mg for female adolescents [8], compared with a recommended intake of 1300 mg/day for those between 9 and 18 years of age in the United States [9]. Overall there is conflicting evidence regarding calcium and vitamin D supplements. As most supplementation trials include both micronutrients, it is difficult to isolate the effects of each nutrient singly. Similarly, studies that have investigated the effects of dairy products on bone health cannot be attributed to solely calcium, because of the number of bone-enhancing nutrients contained in dairy foods. Specific research findings relating to calcium at different life stages are described in other paragraphs later in this chapter.

13.3.2 VITAMIN D AND BONE, AN OVERVIEW The two major forms of vitamin D are vitamin D2 (also referred to as ergocalciferol) and vitamin D3 (also referred to as cholecalciferol). An overview of vitamin D metabolism is illustrated in Fig. 13.1. Vitamin D3 is synthesized in the skin from a cholesterol precursor, upon exposure to sufficient ultraviolet B radiation from sunlight. Vitamin D may also be derived from food and synthetic nutritional supplements. Vitamin D from sun exposure, food, and supplements is biologically inert and is activated by hydroxylation in the liver to 25(OH)D3 [10]. Also known as calcidiol, this is the main storage form of vitamin D3. Calcidiol is subsequently converted to the activated form of vitamin D, 1,25(OH)2D3, also known as calcitriol, in the kidney. Renal conversion is stimulated by PTH and suppressed by phosphate. Calcitriol, via its interaction with the vitamin D receptor, causes an increase in calcium and phosphorus absorption in the proximal small intestine and an increase in bone turnover. The net result

13.3 CALCIUM AND VITAMIN D

341

is an increase in plasma calcium and phosphate, leading to the mineralization of new bone. Without vitamin D, only 10%e15% of dietary calcium and about 60% of phosphorus is absorbed. Calcitriol has a strong negative feedback, which downregulates its own synthesis. Calcitriol is more potent than calcidiol. However, the concentration of calcidiol may be 1000 times more and therefore is an important determinant of skeletal health in those with renal impairment. A meta-analysis of studies that have compared the effectiveness of vitamins D2 and D3 in raising serum calcidiol concentration showed that the results have been inconsistent [11]. Serum concentration of 25(OH)D is the best indicator of vitamin D status. It accounts for both endogenous and dietary-derived vitamin D and has a circulating half-life of 15 days [12]. In contrast, circulating 1,25(OH)2D is not thought to be a good indicator of vitamin D status. Its serum concentrations are closely regulated by PTH, calcium, and phosphate and it has a shorter half-life of approximately 15 h. A limitation of the use of 25(OH)D is that it may decrease in response to acute inflammation and may also be influenced by body mass index (BMI) and genetic variation. Laboratory methods for measurement of serum 25(OH)D concentration can produce variable results. As of this writing there is no clear consensus on the threshold serum 25(OH)D concentration used to define vitamin D deficiency, and cutoffs vary across studies [10]. All of these factors should be considered when interpreting research concerning serum 25(OH)D concentration and bone outcomes [10]. Specific research findings relating to vitamin D and bone at different life stages are described in other paragraphs later in this chapter.

13.3.2.1 Recommended Vitamin D Intake Recommended vitamin D intakes in the United Kingdom were updated in 2016 by the Scientific Advisory Committee on Nutrition (SACN) [10] and are shown in Table 13.1, alongside recommended intakes from the US Food and Nutrition Board (FNB) of the Institute of Medicine [9]. As vitamin D is sparsely distributed across few foods, such as the flesh of oily fish and egg yolk, fortification of staple foods such as breakfast cereals, milk, infant formula, and spreading fats is common in some countries. Supplementation may be required to achieve the recommended intake and is recommended for those not exposed to sunlight. Derivations of dietary recommendations for vitamin D are not straightforward, due to variations in sun exposure, latitude, and skin pigmentation at different ages and different locations around the world. Therefore in the UK recommendations, the SACN did not take account of sunlight exposure, considering that evidence on musculoskeletal health outcomes (rickets, osteomalacia, falls, muscle strength and function) was sufficient to use as the basis for making dietary recommendations [10]. As there was deemed to be insufficient data for children under 4 years of age, “safe intakes” were set for this age group, rather than RNIs. Similarly the US FNB established a Recommended Daily Allowance for vitamin D representing a daily intake that is sufficient to maintain bone health and normal calcium metabolism in healthy people, on the basis of minimal sun exposure [9].

13.3.3 PHOSPHORUS Phosphorus is a key component of bone minerals and is required for the growth and maintenance of the skeleton. Phosphorus is widely distributed in foods, meaning deficiency is rare. As phosphorus is commonly found in the same foods as protein and calcium (e.g., meat and dairy), separating and disentangling any effect on the skeleton is difficult. It is also consumed in the form of phosphoric acid

342

CHAPTER 13 DIET AND BONE HEALTH

Table 13.1 A Comparison of Vitamin D Recommendations for Different Age Groups in the United Kingdom and United States UK (SACN) [10] Infants

Children

Adults

Pregnancy/lactation

Older adults

Population groups at increased risk of vitamin D deficiency

A recommended intake of 8.5 mg per day for infants up to 6 months A “safe intake” of 8.5e10 mg per day for all infants from birth to 1 year of age A “safe intake” of 10 mg per day for children ages 1e4 years 10 mg per day for everyone in the general population age 4 years and older 10 mg per day for pregnant and lactating women No specific recommendation

10 mg per day

USA (Food and Nutrition Board) [9] Adequate intake of 10 mg per day

15 mg per day for those ages 1e18 years 15 mg per day for those ages 18e50 years

15 mg per day for those ages 18e50 years 15 mg per day for those between 50 and 70 years 20 mg per day for those over 70 years d

SACN, Scientific Advisory Committee on Nutrition. 10 mg ¼ 40 IU.

in carbonated cola drinks and as an additive in convenience processed food, meaning consumption may be underestimated [13]. A high dietary intake of phosphorus may be harmful to bone health as it has an acidic effect, requiring buffering by calcium salts. Excess phosphorus intake may lead to secondary hyperparathyroidism, lowering urinary calcium excretion, although a systematic review found no deleterious effect of phosphate on osteoporosis risk [14]. Due to the differing absorption rates of phosphorus and calcium (70% and 30%, respectively), a dietary intake ratio of 1:1 is recommended.

13.3.4 MAGNESIUM Magnesium is a component of mineralized bone, forming approximately 1% of the bone matrix. Major dietary sources of magnesium are green vegetables, nuts, cereals, and shellfish; however, low intake is common in developed countries [15]. Because food sources of magnesium are frequently high in other nutrients that are beneficial to bone, such as potassium and calcium, it can be difficult to separate the

13.3 CALCIUM AND VITAMIN D

343

effects of individual nutrients [16]. Magnesium deficiency can result in changes to calcium homeostasis due to impaired PTH secretion and end-organ resistance to PTH. The evidence regarding magnesium intake and bone outcomes is unconvincing. A large prospective cohort study of postmenopausal women by Orchard et al. [16] reported that although lower magnesium intake is associated with lower BMD of the hip and whole body, this does not translate into increased risk of fractures [16]. A meta-analysis of 12 studies conducted the following year found similar results [17], reporting a positive marginally significant correlation between magnesium intake and BMD in the hip and femoral neck; however, magnesium intake was not associated with BMD in the lumbar spine. High intake of magnesium was not associated with increased risk of hip and total fractures. Because four different bone sites were used as the outcome measure, there were only a small number of studies investigating each site. In addition there were few studies conducted in men, limiting the generalizability of the findings.

13.3.5 POTASSIUM Acid-producing diets, characterized by low fruit and vegetable relative to protein intake, have been implicated as potential contributors to bone loss in the elderly population [18]. Potassium may prevent bone loss because of its neutralizing effect on dietary acid and by increasing renal calcium retention [18]. Treatment with alkaline salts of potassium has been proposed as a treatment option. A doubleblind, randomized, placebo-controlled study of 244 men and women age 50 years and older for 3 months found favorable effects on bone turnover and calcium excretion; however, bone mass was not assessed and supplementation was for only 3 months. Jehle et al. [19] conducted a randomized controlled trial (RCT) of potassium citrate supplementation in 201 men and women >65 years of age. Participants did not have osteoporosis and were also supplemented with calcium and vitamin D. Treatment for 24 months resulted in a significant increase in BMD and volumetric BMD at several sites tested, while also improving bone microarchitecture. Future research using a larger and more diverse multisite sample is required.

13.3.6 VITAMIN K Vitamin K may be protective against bone loss, due to its role in the organic matrix of bone, including collagen synthesis. It modifies osteocalcin, which is involved in bone remodeling, although the exact effect of vitamin K is not well defined. A meta-analysis performed by Fang et al. [20] reported that vitamin K supplementation was effective in increasing BMD at the lumbar spine, but not the femoral neck. Subgroup analysis indicated that ethnicity, gender, and vitamin K type were associated with variable effects on BMD at the lumbar spine. Overall the treatment effects were deemed to be modest, with a risk of bias, and should therefore be interpreted with caution [20].

13.3.7 OTHER NUTRIENTS This chapter aims to give a general overview of research related to nutrition and bone health, by highlighting the most important dietary components that are supported by clear evidence from welldesigned studies. However, it is not exhaustive, and it is not possible to outline the role of every nutrient in relation to bone. Research exists regarding other nutrients that act as cofactors in bone

344

CHAPTER 13 DIET AND BONE HEALTH

metabolism and collagen synthesis (such as zinc, copper, iron, and vitamin C). In addition, fluoride, sodium, and vitamin A have been implicated in fracture risk; plus, folate, vitamin B12, and vitamin B6 may improve bone health and decrease fracture risk by decreasing the homocysteine level. Of note, the European Food Safety Authority has assessed the scientific evidence regarding health claims related to bone health and has approved specific claims related to protein, calcium, vitamin D, magnesium, manganese, phosphorus, vitamin K, and zinc, with other nutrients not approved [21].

13.4 LIFE COURSE PERSPECTIVE ON NUTRITION AND BONE 13.4.1 OSTEOPOROSIS: BURDEN AND EPIDEMIOLOGY Osteoporosis is a skeletal disorder characterized by low bone mass and microarchitectural deterioration of bone tissue, with a consequent increase in the fragility of bone and risk of fracture [22]. It is a debilitating and usually silent disease, whereby individuals are usually asymptomatic until a fracture occurs. Osteoporosis represents a major public health problem through its association with fragility fractures. The most widely recognized sites of osteoporotic fractures are the hip, wrist, and vertebral body, and these have a higher than average percentage of trabecular bone. However, fractures at other sites such as the pelvis, proximal humerus, and proximal tibia are not uncommon. Individuals with osteoporosis are at risk of fractures at any site exposed to trauma and may suffer a fracture at a lower level of trauma than those without osteoporosis. These fractures lead to severe mortality and morbidity, a significant burden on society in general and a huge impact economically [23]. Worldwide, osteoporosis-related fracture is common in both genders in the developed world. In 2010, the annual mean incidence of nontraumatic fractures in those ages >50 years in North America, Europe, Australia, and Japan was approximately 6.7%, with women accounting for most of the total nontraumatic fracture burden (77%) [24]. In Europe, the disability due to osteoporosis is greater than that caused by cancers (with the exception of lung cancer) and is comparable to or greater than that lost to a variety of chronic noncommunicable diseases, such as rheumatoid arthritis, asthma, and high blood pressuree related heart disease [3]. In the United Kingdom, the estimated annual cost of fragility fractures to the National Health Service is £3.4 billion, which is projected to increase by 24% in 2025 [23]. Overall, the public health burden of osteoporotic fracture is likely to rise in future generations, due partly to an increase in life expectancy. Understanding the epidemiology of this disease and the preventative role nutrition can play is therefore essential in trying to develop strategies to help reduce this load.

13.4.2 ETIOLOGY OF OSTEOPOROSIS AND ROLE OF DIET Although genetic factors are estimated to account for up to 60%e80% of variability in BMD [25], there is evidence that modifiable nutritional factors (e.g., poor intake of vitamin D and calcium), alongside physical inactivity can adversely affect bone mineral accrual. Dietary intake plays a role directly and indirectly (through its effect on body stature) in both the primary and the secondary prevention of osteoporosis. The development of osteoporosis can be traced back many years before a fracture manifests, starting in utero and involving nutrition across the whole life course. The bone mass of an individual in later life is dependent upon PBM achieved and subsequent rate of loss [2]. Hence any factor that influences the achievement of PBM, or prevents the loss of bone in middle age, affects the risk of a fragility fracture later in life. This is shown in Fig. 13.2, adapted from Cooper and

13.4 LIFE COURSE PERSPECTIVE ON NUTRITION AND BONE

345

Melton [26]. Prevention of osteoporosis may therefore be targeted at optimization of PBM, reducing bone loss, or both.

13.4.3 MATERNAL NUTRITION AND BONE OUTCOMES Nutrition exposure in utero has a developmental effect on several metabolic systems, including the development of the skeletal system. This effect is known as programming, which describes persisting changes in structure and function caused by environmental stimuli acting at critical periods during early development. Intrauterine programming, in the form of maternal malnutrition, can contribute to the risk of osteoporosis in later life, as mothers need to be sufficiently well nourished to support an infant’s development in utero [7]. The finding from a systematic review [27], that higher birth weight is associated with greater BMC of the lumbar spine and hip in adulthood across a range of settings, provides convincing evidence for the intrauterine programming of skeletal development and tracking of skeletal size into adulthood.

13.4.3.1 Maternal Diet: Observational and Interventional Studies of Bone During pregnancy there are alterations in calcium and bone metabolism due to increasing demands made by the fetus on maternal skeleton stores. There is increased intestinal absorption from the diet, increased renal retention, and increased bone mobilization. Mineralization of the fetal skeleton occurs

FIGURE 13.2 Development of bone mass over the life course. Bone mass increases from the intrauterine period to a peak in early adulthood, with a decline thereafter. Modulation of the growth trajectory early in life may influence the magnitude of peak bone mass; intervention to reduce the rate of postpeak bone loss may be appropriate in older age. Reproduced with permission from Cooper and Melton [26].

346

CHAPTER 13 DIET AND BONE HEALTH

mainly during the last 10 weeks of pregnancy; by late pregnancy calcium accrual by the skeleton is >100 mg/day. Mothereoffspring studies have demonstrated the effect of maternal nutrition on childhood bone health. Vitamin D is the nutrient that has been most studied, although observational studies have identified a positive association between a number of nutrients and higher childhood bone mass, including maternal protein, calcium, and phosphorus intake and vitamin B12 [28]. Of note, research suggests that although pregnant women consume slightly higher amounts of nutrients required for bone health than their nonpregnant counterparts, they do not necessarily consume more of those nutrients that have an increased demand [29]. Traditionally, most research in this area was observational. Lower concentrations of gestational 25(OH)D were reported to be associated with lower whole-body and lumbar spine BMC and BMD in children at 9 years [30]. Conflicting findings have also been reported [31]; however, a subsequent systematic review found that 5 of 8 identified studies reported a significant positive relation between maternal vitamin D status and offspring bone outcomes [32]. More recently an interventional study, the MAVIDOS trial, took place in the United Kingdom [32]. The multicenter randomized trial of either oral vitamin D (1000 IU cholecalciferol/day) or placebo at 14 weeks gestation did not lead to increased offspring whole-body BMC at birth compared with placebo, but did find that 1000 IU of cholecalciferol daily is a safe dose and sufficient to ensure that most pregnant women are vitamin D replete. Secondary analyses found an interaction between vitamin D treatment and season, with the suggestion of a benefit for offspring born during winter months. Longer term follow-up of the offspring is currently ongoing.

13.4.3.2 Maternal Diet: Dietary Pattern Analysis Other nutrients and specific foods have also been implicated in maternaleoffspring bone studies, as well as research that has taken a dietary pattern approach. “Healthier” maternal diets have been associated with greater offspring bone mass [33,34]. In the Danish National Birth Cohort study, the association between dietary patterns in midpregnancy in >100,000 women and offspring forearm fractures between birth and 16 years of age was investigated. Seven different dietary patterns were identified, including a “Western” diet pattern, characterized by higher intakes of meat, potato, white bread, and margarine alongside lower consumption of fruit, vegetables, nuts, and water. Overall there was a positive association between the consumption of a Western dietary pattern during midpregnancy and the incidence of offspring forearm fractures, with a similar association found for consumption of artificially sweetened beverages. The other dietary patterns were not associated with offspring forearm fractures, nor were individual macro- and micronutrients or single food groups.

13.4.4 CHILDHOOD AND ADOLESCENCE Bone mass is accrued relatively slowly throughout childhood before increasing more rapidly with the onset of puberty [4]. During this time, bone elongates, but also undergoes modeling (resorption of bone in one area and deposition in another). PBM is the amount of bone tissue present at the end of skeletal maturation; thought to occur early in the third decade of life. Timing of PBM may vary slightly between skeletal sites. Because of different timings of puberty, it is estimated that maximal gain in bone mass occurs in Caucasian children between 11 and 14 years for girls and 13 and 17 years in boys. Childhood growth spurts could be considered a particularly sensitive or vulnerable period in relation to the later risk of fracture; therefore suboptimal nutritional intake at this time may be correlated to bone health in childhood [35]. In adolescents, bone disorders due to primary or secondary

13.4 LIFE COURSE PERSPECTIVE ON NUTRITION AND BONE

347

nutritional deficiencies can be divided into osteomalacia and osteopenia/osteoporosis, both of which can result in fractures. Rickets (before epiphyseal plate closure) and osteomalacia (after epiphyseal plate closure) are characterized by defective bone mineralization. Protein, calcium, and vitamin D are arguably the most important nutrients for bone health in the first two decades [7].

13.4.4.1 Energy and Protein in Childhood and Adolescence Calorie intake and bone mass are positively associated across all stages of the life course, although obesity and overweight may not be beneficial. Body composition in childhood and adolescence, in particular at extremes of the weight spectrum, exerts a significant effect on bone accrual, geometry, and subsequent fracture risk [36]. In childhood, a sufficient energy intake is required to support skeletal growth. Adolescents with malnutrition are at high risk for insufficient bone accrual, as demonstrated in studies of adolescents with anorexia nervosa [37,38]. At the other end of the weight spectrum, the effect of pediatric overweight and obesity on bone mass is a subject of debate. Although a high body mass increases mechanical loading on weight-bearing bones, the direct effect of fat mass on bone is not clear [36].A population-based, cross-sectional study of 913,178 patients ages 2e19 years demonstrated that overweight, moderately obese, and extremely obese young people had a stepwise increased risk of fractures of the foot, ankle, knee, and leg. The association was apparent in both sexes and strongest in the 6- to 11-year-old group; however, there was no association between increasing BMI and risk of fractures of the femur and hip [39]. The study did not assess activity level, mode of injury, differences in treatment settings, fracture severity, or differences in course of care, all of which may confound the results. As bone is composed largely of protein, an optimal intake of amino acids is required to stimulate bone synthesis. Protein also stimulates insulin-like growth factor-I. Historically animal and human studies reported that high dietary intakes of animal protein lead to increased urinary calcium excretion, thereby leading to reduced availability for bone deposition. Theoretically this might be due mobilization of calcium from bone to act as a buffer to the acid produced from protein catabolism. However, this theory is unproven. Much of what is known about protein intake and bone quality has been extrapolated from adult studies, with limited work in children and adolescents [2]. In a systematic review of studies published since the year 2000, the majority of prospective and cross-sectional studies support a positive relationship between protein intake and bone in childhood and adolescence, with only one prospective study showing a negative effect [2].

13.4.4.2 Calcium in Childhood and Adolescence The requirement for calcium peaks during adolescence, due to the high growth velocity during puberty. The age of peak calcium accretion is estimated to be 14 and 12.5 years for boys and girls, respectively [7]. The role of calcium in achieving PBM is an area of significant research, with numerous studies having been conducted. Studies have investigated the amount and type of calcium needed at different ages and stages and the relationship between calcium intake and physical activity. A meta-analysis assessing the effect of calcium intake on bone mass in children identified 16 RCTs published since 2000. Despite heterogeneity in study design, dose, and outcome variables, calcium supplementation consistently resulted in increases in bone mass and density in children and adolescents in the region of 1%e5%. This was the case for supplement pills, fortified foods, or dairy products. Specifically looking at the RCTs of supplement pills, 90% had a small, biologically and statistically significant positive effect on BMD or BMC accrual. Bone accrual was largest in children who had the lowest calcium intake at baseline, suggesting there is a threshold effect [2].

348

CHAPTER 13 DIET AND BONE HEALTH

13.4.4.3 Vitamin D in Childhood and Adolescence A recent review of the evidence for vitamin D and rickets reported that most evidence was derived from cross-sectional observational studies and case reports and may therefore have been influenced by confounding [10]. As is the case in studies of bone health at all ages, it was not clear whether the cause of rickets was vitamin D deficiency and/or calcium deficiency. A meta-analysis assessing the effect of vitamin D on bone in childhood and adolescence, also published in 2016, identified 8 RCTs, 1 prospective study, and 3 cross sectional studies published since the year 2000. The RCTs included studies with doses varying from 200 to 2000 IU/day and mostly recruited female adolescents. Overall, half of the RCTs were found to have a positive effect. However, as previously mentioned at the beginning of this chapter, different methodologies were used to assess serum 25(OH)D, making direct comparisons difficult. In addition, across all included RCTs, participants had a mean baseline serum 25(OH)D between 18 and 48 nmol (i.e., lower than the 50 nmol/L used to define vitamin D deficiency), with an absence of RCTs in children and adolescents who were vitamin D replete at baseline. This confirms the findings of a previous systematic review and meta-analysis conducted in 2010, which found that vitamin D supplementation of deficient children and adolescents could result in clinically useful improvements; however, it is unlikely that vitamin D supplements are beneficial in children and adolescents with normal vitamin D levels [40].

13.4.4.4 Dairy Foods and Dietary Patterns in Childhood and Adolescence Dairy products provide a combination of essential nutrients that are relevant for bone health and as such are part of many international nutritional recommendations. Numerous observational studies and RCTs have shown a favorable effect of dairy products on bone health during childhood and adolescence [41]. Specifically, a systematic review concluded that although the best evidence was specifically for calcium intake and physical activity [2], more generally, there was good evidence for a positive role of dairy consumption. Looking at other food groups and food items generally, a systematic review investigating dietary intake in childhood and risk of fracture identified 18 observational studies, which were primarily caseecontrol in design, but no RCTs [35]. Overall, fracture risk seemed to be associated with milk avoidance, high energy intake, high cheese intake, high intake of sugar-sweetened beverages, and no breastfeeding [35]. In a Dutch cohort study that analyzed dietary patterns rather than individual nutrients [42], an infant dietary pattern characterized by high intakes of dairy and cheese, whole grains, and eggs was positively associated with bone development at age 6 years. However, the association between the dairy and whole grains pattern and the bone outcomes was observed only in children who did not receive vitamin D supplementation. These two studies indicate the complexity of nutrition and bone research and the possibility that a commonly consumed food (i.e., cheese) may be associated with either positive or negative bone outcomes, depending on the study design used, dietary analysis method, and outcome measure.

13.4.5 BONE HEALTH IN THE OLDER ADULT: MENOPAUSE AND BEYOND Following on from achievement of PBM, there is a period of 2e3 decades wherein skeletal growth has ceased. During this time, physical activity and pregnancy/lactation are the main contributing factors to changes in BMD, until the gradual process of bone loss begins. The timing of loss of bone later in life differs between men and women. In women skeletal losses usually occur postmenopausally, from

13.4 LIFE COURSE PERSPECTIVE ON NUTRITION AND BONE

349

approximately 50 years onward. Postmenopause, bone remodeling increases in response to reduced ovarian production of estrogen, leading to increased bone fragility. In men, skeletal losses typically start at a later age and, as such, generally result in less overall bone loss compared with women. The combination of adequate calcium and vitamin D, physical activity, and bone-conserving medication may help slow the process of bone loss. As the burden of osteoporosis typically presents in the older adult, there is a concentration of nutrition research in this age group, which will now be outlined.

13.4.5.1 Energy and Protein in the Older Adult and Bone Outcomes In adulthood bone mass is positively associated with energy intake and, conversely, undernutrition can increase the risk of osteoporosis and fracture. Lower BMI is a significant risk factor for hip fracture [43]. Body weight is a predictor of osteoporosis, whereby a higher body weight is associated with fewer fractures. Individuals who are malnourished are more likely to develop muscle weakness and experience falls, meaning they are more at risk of fracture. Maintaining sufficient subcutaneous fat into later life has two important roles in terms of bone health. Fatty tissue contains the enzyme aromatase, which produces estradiol in men and postmenopausal women, which in turn plays a role in maintaining bone. Second, the presence of subcutaneous fat provides cushioning in the event of a fall, thus providing some protection on impact. In certain population groups, increases in protein intake may be indirectly beneficial in preventing fractures, by preventing sarcopenia and falls in elderly people with malnutrition. In 2013 an international study group (the PROT-AGE group) published a position paper identifying the specific needs of older people with regard to protein requirements [44]. They concluded that older people (>65 years) require more dietary protein than younger people, advising an intake of 1e1.2 g/kg body weight/day, which may increase to 2.0 g/kg body weight/day if there is considerable malnutrition, but not in the case of kidney disease. The importance of protein intake by older people hospitalized with hip fracture was also recognized in the position statement. Subsequently a systematic review identified positive trends of dietary protein on BMD at most bone sites, but only moderate evidence to support benefits of higher protein intake on the lumbar spine [45]. Due to heterogeneity and the risk of confounding, highquality, long-term studies are needed to clarify dietary protein’s role in bone health.

13.4.5.2 Calcium in the Older Adult and Bone Outcomes Typically, intestinal calcium absorption declines with age, as does calcium dietary intake as a result of declining appetite. If calcium intake and absorption are insufficient, bone is used as a source of calcium to maintain extracellular concentrations. Supplementation may be required if dietary intake is suboptimal. There is some evidence from RCTs in participants over 50 years of age that increasing calcium intake, either through the diet or in the form of supplements, leads to small increases in BMD [46]. Increases in BMD were similar in trials of calcium monotherapy versus coadministered calcium and vitamin D, in trials with calcium doses of 1000 versus 500 mg/day, and in trials in which the baseline dietary calcium intake was 67 years not demonstrating an effect [55]. The low number of included studies in their review may be due to inconsistent criteria used to define the Mediterranean diet. They concluded that there is a distinct lack of research to understand the relationship between the Mediterranean diet and musculoskeletal health in all ages. A meta-analysis that specifically examined the risk of fracture as its outcome measure identified six eligible studies [56] with promising results. Malmir et al. [56] reported that increased adherence to the Mediterranean diet was associated with a 21% reduced risk of hip fractures and positively associated with a higher BMD in the lumbar spine, whole body, total hip, and trochanter. However, as the included studies were cross-sectional, a cause and effect relationship cannot be established. It appears that further prospective epidemiological and clinical studies in this area are particularly important.

13.4.5.5 Fruit and Vegetables in the Older Adult and Bone Outcomes Observational cross-sectional studies have found an association between diets rich in fruit and vegetables and reduced bone loss. This could be attributable to specific micronutrients, e.g., potassium or magnesium, or the alkalinizing effect of fruit and vegetables. However, the evidence is inconsistent and may be confounded. Participants with higher intakes of fruit and vegetables may have an overall healthier diet and lifestyle, meaning factors such as physical activity and smoking need to be

REFERENCES

351

statistically controlled for. Byberg et al. [57] investigated whether there was a doseeresponse effect between fruit and vegetable intake and subsequent hip fracture in >40,000 men and >34,000 women, demonstrating a 54% increase in hip fractures in those who consumed less than one serving per day, compared with those who consumed 3-5 servings per day. This could be attributed to the antioxidant content or the presence of other nutrients such as potassium, magnesium, and vitamin K in fruit and vegetables.

13.5 CHALLENGES IN NUTRITION AND BONE RESEARCH As is the case with nutritional epidemiological research generally, there are several common factors such as dietary assessment and randomization to dietary treatment that present practical and ethical challenges. In the area of bone research there are additional specific issues [58], listed next. •

•

•

•

There are few, if any, studies of the cumulative effects of diet throughout the life course. As outlined in this chapter, deleterious bone outcomes (i.e., fragility factors) do not usually manifest until later in life; therefore longitudinal studies are required, which have resource implications and pose participant burden. Most studies focus on discrete time periods, or stages of life, with a lack of long-term follow-up of RCTs. A single dietary nutrient is a small part of total intake and dietary composition. As outlined using the example of calcium and dairy foods, food groups and nutrients in the diet are interrelated and there may be a cumulative or synergistic effect, which cannot be accounted for in studies of single nutrients. Taking a whole-diet rather than a single- or few-nutrient approach may be a superior way to analyze the effect of diet on bone outcomes. Diet reflects the socioeconomic status and lifestyle of the individual. In observational studies of bone health, there are likely to be confounding effects from both measured and unmeasured sociodemographic and lifestyle factors, particularly because lifestyle factors such as smoking and physical inactivity (þ/ caffeine and alcohol) are also risk factors for bone disease. Heterogeneous outcome measures and lack of consensus. As already outlined, there are numerous different outcome measures employed in bone research studies, namely imaging and fragility fractures at various body sites. There is also disagreement regarding the optimal criteria for the assessment of vitamin D sufficiency and deficiency, which makes identifying specific thresholds or treatment goals difficult.

In light of these constraints, the following gaps have been identified as meriting further investigation with regard to achievement of PBM [2]: differing effects of interventions depending on the life stage of growth, geneeenvironment interactions and how they may impact the development of PBM, the need to identify and utilize biomarkers of exposure and effects, and the interaction of bone with other tissues throughout the body. In addition, longitudinal studies are needed to document the relationship between growth and measures of bone fragility and fractures and to identify lifestyle interventions that may prevent fractures during this period of susceptibility.

352

CHAPTER 13 DIET AND BONE HEALTH

REFERENCES [1] Parfitt AM. The two faces of growth: benefits and risks to bone integrity. Osteoporos Int 1994;4(6):382e98. [2] Weaver CM, Gordon CM, Janz KF, Kalkwarf HJ, Lappe JM, Lewis R, et al. The National Osteoporosis Foundation’s position statement on peak bone mass development and lifestyle factors: a systematic review and implementation recommendations. Osteoporos Int 2016;27(4):1281e386. [3] Johnell O, Kanis JA. An estimate of the worldwide prevalence and disability associated with osteoporotic fractures. Osteoporos Int 2006;17(12):1726e33. [4] Heaney RP, Abrams S, Dawson-Hughes B, Looker A, Looker A, Marcus R, et al. Peak bone mass. Osteoporos Int 2000;11:985e1009. [5] Oftadeh R, Perez-Viloria M, Villa-Camacho JC, Vaziri A, Nazarian A. Biomechanics and mechanobiology of trabecular bone: a review. J Biomech Eng 2015;137(1):10802. Available from: http://biomechanical. asmedigitalcollection.asme.org/article.aspx?doi¼10.1115/1.4029176. [6] Kanis JA, Glu¨er CC. An update on the diagnosis and assessment of osteoporosis with densitometry. Committee of Scientific Advisors, International Osteoporosis Foundation. Osteoporos Int 2000;11(3): 192e202. [7] Mitchell PJ, Cooper C, Dawson-Hughes B, Gordon CM, Rizzoli R. Life-course approach to nutrition. Osteoporos Int 2015;26:2723e42. [8] Department of Health. Dietary reference values for food energy and nutrients for the United Kingdom. 1991. United Kingdom. [9] IOM I of M (US). In: Dietary reference intakes for calcium and vitamin D, vol. 130. IOM; 2011. [10] Scientific Advisory Committee on Nutrition. Vitamin D and health [Internet]. 2016. Available from:, https:// www.gov.uk/government/groups/scientific-advisory-committee-on-nutrition. [11] Tripkovic L, Lambert H, Hart K, Smith CP, Bucca G, Penson S, et al. Comparison of vitamin D2 and vitamin D3 supplementation in raising serum 25-hydroxyvitamin D status: a systematic review 2012;3:1357e64. [12] Jones G. Pharmacokinetics of vitamin D toxicity. Am J Clin Nutr 2008;88. [13] Calvo MS, Uribarri J. Narrative Review. Public health impact of dietary phosphorus excess on bone and cardiovascular health in the general population. Am J Clin Nutr 2013;(1e3):6e15. [14] Fenton TR, Tough SC, Lyon AW, Eliasziw M, Hanley DA. Causal assessment of dietary acid load and bone disease : a systematic review & meta-analysis applying Hill’s epidemiologic criteria for causality. Nutr J 2011;10. [15] Whitton C, Nicholson SK, Roberts C, Prynne CJ, Pot GK, Olson A, et al. National Diet and Nutrition Survey: UK food consumption and nutrient intakes from the first year of the rolling programme and comparisons with previous surveys. Br J Nutr 2011;106(12):1899e914. Available from: http://www.journals.cambridge. org/abstract_S0007114511002340. [16] Orchard TS, Larson JC, Alghothani N, Bout-tabaku S, Cauley JA, Chen Z, et al. Magnesium intake, bone mineral density, and fractures : results from the Women’s Health Initiative Observational Study. Am J Clin Nutr 2014;8:926e33. [17] Farsinejad-Marj M, Saneei P, Esmaillzadeh A. Dietary magnesium intake, bone mineral density and risk of fracture: a systematic review and meta-analysis. Osteoporos Int 2016;27:1389e99. [18] Dawson-hughes B., Harris S.S., Palermo N.J., Gilhooly C.H., Shea M.K., Fielding R.A., et al. No title. [19] Jehle S, Hulter HN, Krapf R. Effect of potassium citrate on bone density, microarchitecture, and fracture risk in healthy older adults without osteoporosis: a randomized 2017;98(January 2013):207e17. [20] Fang Y, Hu C, Tao X, Wan Y, Tao F. Effect of vitamin K on bone mineral density: a meta-analysis of randomized controlled trials. J Bone Miner Metabol 2012;30:60e8. [21] European Commission. EU register on nutrition and health claims on foods. 2015. Available from: https://ec. europa.eu/food/safety/labelling_nutrition/claims/health_claims_en.

REFERENCES

353

[22] Bouillon R, Burckhardt P, Christiansen C, Fleisch H, Fujita T, Gennari C, et al. Consensus development conference: prophylaxis and treatment of osteoporosis. Osteoporos Int 1991;1:114e7. Available from: https://lirias.kuleuven.be/handle/123456789/173726. [23] Svedbom A, Hernlund E, Iverga˚rd M, Compston J, Cooper C, Stenmark J, et al. Osteoporosis in the European Union: a compendium of country-specific reports. Arch Osteoporos 2013;8(1e2). [24] Wade SW, Strader C, Fitzpatrick LA, Anthony MS. Sex- and age-specific incidence of non-traumatic fractures in selected industrialized countries. Arch Osteoporos 2012;7(1e2):219e27. [25] Morrison N, QI J, Tokita A, Kelly P, Crofts L, Nguyen T, et al. Predictors of bone density from vitamin D receptor alleles. Nature 1994;367(6460):284e7. [26] Cooper C, Melton L. Epidemiology of osteoporosis. Trends Endocrinol Metabol 1992;3(6):224e9. [27] Baird J, Kurshid MA, Kim M, Harvey N. Does birthweight predict bone mass in adulthood ? A systematic review and meta-analysis. Osteoporos Int 2011:1323e34. [28] Heppe DHM, Medina-gomez C, Hofman A, Franco OH, Rivadeneira F, Jaddoe VWV. Maternal firsttrimester diet and childhood bone mass: the Generation R Study. Am J Clin Nutr 2013;98:224e32. [29] Brennan-olsen NKHSL, Bennett K, Moloney DJ, Pasco JA. Maternal nutrition during Pregnancy: intake of nutrients important for bone health. Matern Child Health J 2017;21(4):845e51. Available from: https://doi. org/10.1007/s10995-016-2178-7. [30] Javaid MK, Crozier SR, Harvey NC, Gale CR, Dennison EM, Boucher BJ, et al. Maternal vitamin D status during pregnancy and childhood bone mass at age 9 years: a longitudinal study. Lancet (London, England) 2006;367(9504):36e43. Available from: http://www.sciencedirect.com/science/article/pii/ S0140673606679221. [31] Lawlor DA, Wills AK, Fraser A, Sayers A, Fraser WD, Tobias JH. Association of maternal vitamin D status during pregnancy with bone-mineral content in off spring : a prospective cohort study. Lancet 2013; 381(9884):2176e83. Available from: https://doi.org/10.1016/S0140-6736(12)62203-X. [32] Cooper C, Harvey NC, Bishop NJ, Kennedy S, Papageorghiou AT, Schoenmakers I, et al. Maternal gestational vitamin D supplementation and off spring bone health (MAVIDOS): a multicentre, double-blind, randomised placebo-controlled trial. Lancet Diab Endocrinol 2016;4(5):393e402. Available from: https:// doi.org/10.1016/S2213-8587(16)00044-9. [33] Cole ZA, Gale CR, Javaid MK, Robinson SM, Law C, Boucher BJ, et al. Maternal dietary patterns during pregnancy and childhood bone mass: a longitudinal study. J Bone Miner Res 2009;24(4):663e8. [34] Petersen SB, Rasmussen MA, Olsen SF, Vestergaard P, Mølgaard C, Halldorsson TI, et al. Maternal dietary patterns during pregnancy in relation to offspring forearm fractures: prospective study from the Danish National Birth Cohort. Nutrients 2015:2382e400. [35] Handel MN, Heitmann BL, Abrahamsen B. Nutrient and food intakes in early life and risk of childhood fractures . Am J Clin Nutr 2015:1182e95. [36] Bialo SR, Gordon CM. Underweight, overweight, and pediatric bone fragility: impact and management. Curr Osteoporos Rep 2014;12(3):319e28. [37] Misra M, Katzman DK, Clarke H, Snelgrove D, Brigham K, Miller KK, et al. Hip structural analysis in adolescent boys with anorexia nervosa and controls. J Clin Endocrinol Metab 2013;98(7):2952e8. [38] Faje AT, Fazeli PK, Miller KK, Katzman DK, Ebrahimi S, Lee H, et al. Fracture risk and areal bone mineral density in adolescent females with anorexia nervosa. Int J Eat Disord 2014;47(5):458e66. Available from: http://www.pubmedcentral.nih.gov/articlerender.fcgi? artid¼4053520&tool¼pmcentrez&rendertype¼abstract. [39] Kessler J, Koebnick C, Smith N, Adams A. Childhood obesity is associated with increased risk of most lower extremity fractures. Clin Orthop Relat Res 2013;471(4):1199e207. Available from: http://link.springer.com/ 10.1007/s11999-012-2621-z.

354

CHAPTER 13 DIET AND BONE HEALTH

[40] Winzenberg T, Powell S, Shaw KA, Jones G. Effects of vitamin D supplementation on bone density in healthy children: systematic review and meta-analysis. BMJ 2011;342:c7254. [41] Rozenberg S, Body J, Bruyere O, Bergmann P, Brandi M, C C. Effects of dairy products consumption on Health : benefits and beliefs d a commentary from the Belgian Bone Club and the European Society for clinical and Economic aspects of osteoporosis, osteoarthritis and musculoskeletal diseases. Calcif Tissue Int 2016:1e17. [42] van den Hooven EH, Heppe DHM, Jong JCK. Infant dietary patterns and bone mass in childhood : the Generation R Study. Osteoporos Int 2015:1595e604. [43] De Laet C, Kanis JA, Ode´n A, Johanson H, Johnell O, Delmas P, et al. Body mass index as a predictor of fracture risk: a meta-analysis. Osteoporos Int 2005;16(11):1330e8. [44] Bauer J, Biolo G, Cederholm T, Cesari M, Cruz-Jentoft AJ, Morley JE, et al. Evidence-based recommendations for optimal dietary protein intake in older people: a position paper from the prot-age study group. J Am Med Dir Assoc 2013;14(8):542e59. [45] Shams-white MM, Chung M, Du M, Fu Z, Insogna KL, Karlsen MC, et al. Dietary protein and bone health: a systematic review and meta-analysis from the National Osteoporosis Foundation. Am J Clin Nutr 2017;2(C). [46] Tai V, Leung W, Grey A, Reid IR, Bolland MJ. Calcium intake and bone mineral density: systematic review and meta-analysis. BMJ 2015:h4183. Available from: http://www.bmj.com/lookup/doi/10.1136/bmj.h4183. [47] Bolland MJ, Leung W, Tai V, Bastin S, Gamble GD, Grey A, et al. Calcium intake and risk of fracture: systematic review. BMJ 2015;351(July):h4580. Available from: http://www.bmj.com/lookup/doi/10.1136/ bmj.h4580%5Cnpapers3://publication/doi/10.1136/bmj.h4580. [48] Compston J, Cooper A, Cooper C, Gittoes N, Gregson C, Harvey N. UK clinical guideline for the prevention and treatment of osteoporosis. Arch Osteoporos 2017. [49] Avenall A, Mak J, O’Connell D. Vitamin D and vitamin D analogues for preventing fractures in postmenopausal women and older men. Cochrane Database Syst Rev 2014:CD000227. [50] Bischoff-Ferrari H, Willett W, Wong J, Stuck A, Staehelin H, Orav E, et al. Prevention of nonvertebral fractures with oral vitamin D and dose dependency: a meta-analysis of randomized controlled trials. Arch Intern Med 2009;169:551e61. [51] Bach-Faig A, Berry EM, Lairon D, Reguant J, Trichopoulou A, Dernini S, et al. Mediterranean diet pyramid today. Science and cultural updates. Publ Health Nutr 2011;14(12A):2274e84. Available from: http://www. journals.cambridge.org/abstract_S1368980011002515. [52] Romero Pe´rez A, Rivas Velasco A. Adherence to Mediterranean diet and bone health. Nutr Hosp 2014;29(5): 989e96. Available from: http://www.ncbi.nlm.nih.gov/pubmed/24951976. [53] Craig JV, Bunn DK, Hayhoe RP, Appleyard WO, Lenaghan EA, Welch AA. Relationship between the Mediterranean dietary pattern and musculoskeletal health in children, adolescents, and adults: systematic review and evidence map. Nutr Rev 2017;75(October):830e57. Available from: http://academic.oup.com/ nutritionreviews/article/doi/10.1093/nutrit/nux042/4103051/Relationship-between-the-Mediterraneandietary. [54] Benetou V, Orfanos P, Pettersson-Kymmer U, Bergstro¨m U, Svensson O, Johansson I, et al. Mediterranean diet and incidence of hip fractures in a European cohort. Osteoporos Int 2013;24(5):1587e98. [55] Feart C, Lorrain S, Ginder Coupez V, Samieri C, Letenneur L, Paineau D, et al. Adherence to a Mediterranean diet and risk of fractures in French older persons. Osteoporos Int 2013;24(12):3031e41. [56] Malmir H, Saneei P, Larijani B, Esmaillzadeh A. Adherence to Mediterranean diet in relation to bone mineral density and risk of fracture: a systematic review and meta-analysis of observational studies. Eur J Nutr 2017:1e14. [57] Byberg L, Bellavia A, Orsini N, Wolk A, Michae¨lsson K. Fruit and vegetable intake and risk of hip fracture: a cohort study of Swedish men and women. J Bone Miner Res 2015;30(6):976e84. [58] Ward KA, Prentice A, Kuh DL, Adams JE, Ambrosini GL. Life course dietary patterns and bone health in later life in a British Birth Cohort Study 2016;31(6):1167e76.

CHAPTER

DIET AND LUNG HEALTH

14

Foteini Malli1, 2,, Themis Koutsioukis1, , George Pounis3, Konstantinos I. Gourgoulianis1 Respiratory Medicine Department, University of Thessaly, School of Medicine, Larissa Greece1; Technological Institute of Thessaly, Nursing Department, Larissa, Greece2; Alimos, Athens, Greece3

CHAPTER OUTLINE 14.1 Introduction .............................................................................................................................356 14.2 Diet and Pulmonary Function.....................................................................................................357 14.3 Diet and Asthma .......................................................................................................................357 14.3.1 Dietary Patterns................................................................................................358 14.3.2 Vitamins ..........................................................................................................358 14.3.3 Minerals ..........................................................................................................359 14.3.4 Fatty Acids.......................................................................................................359 14.3.5 Probiotics ........................................................................................................359 14.3.6 Phytochemicals ................................................................................................360 14.4 Diet and Chronic Obstructive Pulmonary Disease ........................................................................ 360 14.4.1 Dietary Patterns................................................................................................360 14.4.2 Vitamins ..........................................................................................................360 14.4.3 Minerals ..........................................................................................................361 14.4.4 Fatty Acids.......................................................................................................361 14.4.5 Probiotics and Dietary Fibers .............................................................................361 14.4.6 Phytochemicals ................................................................................................362 14.5 Diet and Lower Respiratory Tract Infections ............................................................................... 362 14.5.1 Vitamins ..........................................................................................................363 14.5.2 Minerals ..........................................................................................................364 14.5.3 Fatty Acids.......................................................................................................364 14.5.4 ProbioticsePrebiotics .......................................................................................364 14.5.5 Phytochemicals ................................................................................................365 14.6 Diet and Tuberculosis ...............................................................................................................365 14.6.1 Dietary Patterns................................................................................................365 14.6.2 Vitamins ..........................................................................................................365 14.6.3 Minerals ..........................................................................................................366 14.6.4 Fatty Acids.......................................................................................................366



Equally contributed to the preparation of the manuscript.

Analysis in Nutrition Research. https://doi.org/10.1016/B978-0-12-814556-2.00014-2 Copyright © 2019 Elsevier Inc. All rights reserved.

355

356

CHAPTER 14 DIET AND LUNG HEALTH

14.6.5 Probiotics ........................................................................................................366 14.6.6 Phytochemicals ................................................................................................366 14.7 Diet and Lung Cancer................................................................................................................367 14.7.1 Dietary Patterns................................................................................................367 14.7.2 Vitamins ..........................................................................................................367 14.7.3 Minerals ..........................................................................................................368 14.7.4 Fatty Acids.......................................................................................................368 14.7.5 Phytochemicals ................................................................................................368 14.8 Diet and Cystic Fibrosis ............................................................................................................369 14.8.1 Vitamins ..........................................................................................................369 14.8.2 Minerals ..........................................................................................................369 14.8.3 Fatty Acids.......................................................................................................370 14.8.4 Probiotics ........................................................................................................370 14.8.5 Phytochemicals ................................................................................................370 14.9 Diet and Interstitial Lung Diseases.............................................................................................370 14.10 Maternal Diet in Early Life and Lung Health ............................................................................... 371 14.10.1 Dietary Patterns................................................................................................371 14.10.2 Vitamins ..........................................................................................................371 14.10.3 Fatty Acids.......................................................................................................372 14.10.4 Probiotics ........................................................................................................372 14.11 Challenges in Diet and Lung Health Research ............................................................................ 372 References ..........................................................................................................................................373 Further Reading ...................................................................................................................................381

14.1 INTRODUCTION Diet may act as a significant contributor to lung health. It has been hypothesized that various dietary factors may be implicated in the pathogenesis, clinical presentation, severity, and treatment of lung diseases, mainly by manipulating oxidative stress and lung inflammation. This chapter aims at presenting recent research evidence on the effects of dietary habits on the most prevalent respiratory diseases and highlighting the importance of a healthy diet in lung health. Dietary modifications associated with body weight and their possible association with respiratory health will not be addressed here. The chapter starts by discussing how a healthy diet can contribute to pulmonary function and primary and secondary prevention of obstructive lung diseases, mainly asthma and chronic obstructive pulmonary disease (COPD), which are characterized by chronic airway inflammation. It continues by addressing diet as an independent risk factor and adjuvant therapy in respiratory infections and tuberculosis (TB). Research data on the association of various dietary components with the development of lung cancer (LC) are presented later. A special focus on cystic fibrosis (CF) and interstitial lung disease (ILD) and the possible role of dietary supplements in disease modification is given here, too.

14.3 DIET AND ASTHMA

357

The chapter finishes by addressing how maternal diet during pregnancy and diet in early life may contribute to the development of respiratory diseases and closes with a short suggestion of future goals and research challenges in the field of diet and lung health.

14.2 DIET AND PULMONARY FUNCTION Pulmonary function testing is commonly used to diagnose respiratory diseases and/or to identify the severity of respiratory illnesses. Spirometry is the most commonly used tool to assess pulmonary function and the most useful parameters it encompasses are forced expiratory volume in the first second (FEV1), forced vital capacity (FVC), and their ratio (FEV1/FVC). Spirometry can differentiate between obstructive (low FEV1/FVC) and restrictive (normal or increased FEV1/FVC) diseases. A healthy overall diet, as assessed by the Healthy Eating Index 2005, is associated with better lung function as determined by FEV1% predicted and FEV1/FVC. Consumption of animal protein and polyunsaturated fatty acids (PUFAs) shows a positive association with pulmonary function, whereas there is a negative association between pulmonary function and total calories and dietary saturated fatty acids [1]. Total dietary antioxidant intake has been linked to lung function results [2,3]. A dietary pattern based on higher-antioxidant food consumption may be associated with improvement in lung function, at least in COPD patients [4]. Epidemiological studies have indicated that vitamin C consumption is beneficially related to lung function, regardless of smoking intensity, and prevents the development of emphysema in chronic smokers. Vitamin A is necessary for lung development and pulmonary cell differentiation and its deficiency could lead to lung dysfunction due to the disordered composition of collagen IV and laminin and reduction in matrix metalloproteinase concentrations [5]. b-Carotene is positively associated with FEV1 and is related to slower decline of FEV1 in an 8-year follow-up study [6]. In accordance with these findings, serum carotenoid concentrations are inversely associated with a decline in lung function [7]. Higher vitamin E intake has a positive association with FEV1 and FVC, with a dose-dependent response, while others have questioned the relationship between vitamin E and lung function. An increase in serum selenium was associated with an increase in FEV1 among smokers [7a]. According to a 2017 epidemiological study, dietary intake of magnesium, folate (vitamin B9), niacin (vitamin B3), vitamins A and D, long-chain unsaturated fatty acids (eicosanoic fatty acid and PUFAs), and dietary fiber is associated with better FEV1 in chronic smokers [8]. It has been also suggested that keeping sufficient vitamin D levels is necessary for optimal lung health and may regulate the lung microbiome in a sex-specific fashion [9]. Low fiber intake is related to decreased measures of lung function. Data suggest that a higher overall polyphenol content in the human diet may result in better pulmonary function [10]. Interestingly, intake of anthocyanins (flavonoids found in fruits, mainly in berries) is associated with slower age-related rates of FEV1 and FVC decline [11].

14.3 DIET AND ASTHMA Asthma is a heterogeneous chronic disease characterized by airway inflammation and airway hyperresponsiveness to direct or indirect stimuli [12]. Patients with asthma exhibit variable expiratory

358

CHAPTER 14 DIET AND LUNG HEALTH

airflow limitation associated with episodic flare-ups (exacerbations) that may be life threatening. In the same context, asthma symptoms, including wheeze, shortness of breath, chest tightness, and cough, typically vary over time and in intensity [12]. The pathogenesis of asthma is complex and involves airway inflammation and enhanced T-helper cell-2 (Th2) responses. In an oversimplified model, airway inflammation results in the two central components at asthma, i.e., airway hyperresponsiveness and airway obstruction. Asthma prevalence has increased in recent years, mainly in industrialized countries, possibly due to environmental and lifestyle changes, including diet [13]. Airway inflammation associated with asthma may be partially triggered by aggravated oxidative stress and proinflammatory properties associated with a Western diet.

14.3.1 DIETARY PATTERNS Epidemiological data linking diet and asthma are fluent. Different components found in foods have antioxidant, antiallergic, and antiinflammatory properties, which may protect against asthma development. The majority of epidemiological studies suggest that greater adherence to Western dietary patterns (i.e., high consumption of meat, dairy products, processed food, and alcohol) and reduced consumption of fruits, vegetables, whole grain cereals, and fish may result in greater asthma risk [14,15]. Association of asthma with dietary habits has found support in various studies suggesting that higher intake of soft drinks, energy drinks, sugar-added drinks (i.e., tonic), fast food, and butter are associated with increased asthma risk. Experimental evidence suggests that dietary habits may enhance inflammatory responses and alter the Th1/Th2 balance, thus underlying the association with asthma [16]. The effects of dietary pattern on asthma control and acute exacerbations of the disease are well studied and the results are straightforward; adherence to a Western diet has a negative impact on asthma control, while high fresh fruit and vegetables consumption and high fish consumption are associated with fewer exacerbations and better asthma control [17e19]. Current guidelines encourage patients with asthma to consume a diet high in fruit and vegetables. Similar results are obtained in children. High fruit intake, and especially apples, citrus fruit, and tomatoes, is negatively associated with wheezing and asthma risk, suggesting a possible role for flavonoids or other polyphenols [20].

14.3.2 VITAMINS Extensive data have examined the possible association of asthma risk and vitamin consumption based on the hypothesis that a decline in dietary antioxidant intake may increase asthma susceptibility. Experimental data support the aforementioned hypothesis. Vitamin A deficiency worsens ovalbumininduced lung inflammation through enhanced Th2 responses [21], while vitamin E reduces reactive oxygen species (ROS) production and improves antioxidant defense in animal asthma [22]. In the same context, observational studies have shown reduced levels of antioxidant vitamins in asthmatics. Low vitamin A has been reported in severe asthma and may lead to increased airway hypersensitivity. Similarly, vitamins C and E have been inversely correlated with asthma risk. Interventional studies examining the effects of vitamin supplementation in asthmatics are inconclusive. While some studies have shown significant improvements in asthma symptoms or pulmonary function following vitamin E supplementation, others have shown no significant clinical benefit.

14.3 DIET AND ASTHMA

359

Interestingly, multiple nutrient supplementation, including vitamins A, C, and E, may benefit asthmatics. Vitamin D is under examination as a potential modulator of the development of respiratory diseases characterized by chronic lung inflammation. Epidemiological data showed that vitamin D deficiency is common among asthmatics. Serum vitamin D is inversely correlated with airway hyperresponsiveness to methacholine and exercise-induced bronchoconstriction, and vitamin D deficiency has been associated with airway remodeling and asthma severity [23]. Vitamin D supplementation improves asthma symptoms, while there seems to be no clinical evidence for the reduction of asthma exacerbation. However, the relationship of vitamin D with pulmonary function and control of asthma is not consistent among studies. To summarize, the data provide strong evidence for a beneficial effect of vitamin D in asthma, but the clinical benefit of vitamin D supplementation is uncertain.

14.3.3 MINERALS Little is known concerning the association of minerals with asthma. Asthma patients present reduced selenium concentrations, although selenium supplementation has no clinical benefit in asthmatics [23a].

14.3.4 FATTY ACIDS Population data have shown that groups with increased n-6 PUFA consumption have greater asthma predominance compared with those consuming increased n-3 fatty acids [24]. One possible explanation is that n-3 fatty acids produce eicosanoids that are less proinflammatory (prostaglandin E3, leukotriene B5) than those derived from n-6 fatty acids. In addition, metabolites of n-3 fatty acids have the capability to dissolve inflammation. A 20-year longitudinal study revealed that consumption of PUFAs was significantly related to a low rate of incidents of asthma [24a], while a high n-3/n-6 ratio was associated with reduced airway hyperresponsiveness (as suggested by reduced methacholineinduced provocative dose) and leukotriene-4 excretion. However, the beneficial effects of PUFAs are not consistent among studies, and a meta-analysis has suggested that n-3 may reduce the risk of asthma in children but have no effects in adults. Although experimental data suggest a beneficial role of n-3 in asthma, their benefits in real life remain to be clarified. In the same context, PUFA supplementation studies have inconsistent results, possibly due to methodological discrepancies.

14.3.5 PROBIOTICS The term “probiotics” refers to microorganisms that, when administered in adequate amounts, confer a health benefit for the host, while prebiotics are food ingredients that enhance the proliferation or activity of beneficial microorganisms. Probiotics such as lactobacilli and bifidobacteria may reduce asthmatic symptoms by enhancing T-regulatory cell development and rebalancing Th1/Th2 responses toward a Th1-dominant state. Experimental evidence demonstrated that higher intake of the probiotic Lactobacillus paracasei L9 prevents particulate matter 2.5einduced enhancement of the lung inflammatory response [25], while in a murine house dust miteeinduced asthma model, dietary prebiotics prevented and reduced symptoms of asthmatic disease [26]. Two meta-analyses indicated that probiotics may not be efficient at decreasing the risk of asthma and wheezing [26a] [27]).

360

CHAPTER 14 DIET AND LUNG HEALTH

However, a double-blind randomized controlled trial (RCT) showed that synbiotics may decrease episodes of viral respiratory infection in asthmatic children [28]. Further research with long-term clinical trials needs to be conducted to assess the effectiveness of specific strains of probiotics and prebiotics in asthma.

14.3.6 PHYTOCHEMICALS Phytochemicals as bioactive compounds of plant foods have been linked with risk reduction of major chronic diseases. Studies mainly in animal models provide encouraging results for asthma. Kaempferol seems to be efficacious in improving epithelial thickening and airway smooth muscle hypertrophy [28a]. Curcuma longa extract and its component curcumin have antioxidant and antiinflammatory properties, suggesting a therapeutic prospective [29]. Propolis extracts containing flavonoids and phenolics exert antiinflammatory properties by free radical scavenging in a mouse conalbumin-induced asthma model [29a]. Similarly, aerosolized honey reduces airway inflammation. Ginger suspends lung inflammation by suppressing Th2-mediated immune responses and airway eosinophilia [30].

14.4 DIET AND CHRONIC OBSTRUCTIVE PULMONARY DISEASE Currently, COPD is the fourth leading cause of death and a major cause of chronic morbidity [31]. The disease is characterized by persistent respiratory symptoms and airway obstruction due to airway abnormalities usually caused by exposure to noxious particles or gases. The main risk factor is smoking and the disease results from a complex interaction between genetic and environmental factors. Nutritional support is indicated for malnourished patients since low body mass index is associated with worse outcomes in COPD and poor diet contributes to skeletal muscle dysfunctions. Current guidelines suggest that all COPD patients should have general advice on healthy living including diet [31]. In the following paragraphs, we will address the possible effects of dietary patterns and nutrients on the prevention of COPD.

14.4.1 DIETARY PATTERNS Both retrospective and prospective epidemiological studies have proven that dietary patterns including increased consumption of fruit, vegetables, fish, and whole grains may decrease the risk of COPD in smokers and nonsmokers, improve pulmonary function, and decrease long-term COPD mortality [32]. In addition, the Mediterranean diet has been related to a 50% reduction in the risk of COPD [33] and higher fresh fruit intake was associated with a greater fall in FEV1 [34]. Similarly, a randomized intervention study of COPD patients showed that an increase in fruit and vegetable consumption over a period of 3 years might preserve pulmonary function [4].

14.4.2 VITAMINS The systemic inflammation related to COPD may be triggered by the impaired oxidative stress that is prevalent in the disease. The level of airway inflammation in COPD correlates with disease severity

14.4 DIET AND CHRONIC OBSTRUCTIVE PULMONARY DISEASE

361

and is involved in the disease pathogenesis. Thus, reduction of oxidative stress through diet may protect against the development of COPD [35]. Research data have almost unequivocally suggested that vitamin C is related to better lung function regardless of smoking history (for more details see Section 14.2). Serum vitamin A is negatively related to the presence of COPD. Antioxidant supplementation with vitamins A and C and a-lipoic acid reduces oxidative stress in COPD. Low a-tocopherol concentrations in lung tissue are related to more severe cases of COPD [36]. Few interventional studies exist, however. Vitamin E supplementation in almost 39,000 females demonstrated a reduction in the diagnosis of COPD [37], while others have challenged this finding [38]. Taking these findings into account, it seems that the role of antioxidant supplementation in COPD is not conclusive. Exploration of their role should be examined with caution because of the safety concerns associated with increased lung cancer prevalence [39]. Interestingly, vitamin D deficiency accelerated emphysema in animal models through increased protease/antiprotease ratio [40]. Vitamin D deficiency has been consistently reported in the COPD population and has been associated with recurrent exacerbations and hospitalization. Vitamin D supplementation has no effect on spirometry parameters or exacerbation frequency or time to first exacerbation except in patients with dramatically low vitamin D levels. However, vitamin D intake has been associated with improvement in inspiratory muscle strength and oxygen uptake [41].

14.4.3 MINERALS A number of experimental studies suggested a possible link of minerals with COPD, but the quality of evidence is low. Experimental studies demonstrated that copper deficiency results in emphysematous destruction of the lungs and an increase in the mean alveolar airspace areas and mean linear intercept [42]. Mg-deficient patients with COPD reported notably worse COPD-related quality of life, although Mg levels are not associated with spirometry values. Beetroot juice, which is rich in nitrate (NO3 ), reduces oxygen consumption during exercise in COPD patients. COPD patients present lower calcium, phosphorus, and iron levels, while studies have suggested that low calcium intake is associated with increased COPD risk.

14.4.4 FATTY ACIDS PUFA intake may protect against COPD development through a reduction in COPD-related inflammation. Varraso et al. [43] reported no significant relationship between PUFA intake and risk of COPD in two cohort studies. However, high dietary intake of n-3 is inversely associated with COPD risk in a dose-dependent manner. The associations between PUFAs and COPD are inconsistent. PUFA supplementation improves exercise capacity in COPD and may reduce the rate of lung function deterioration. Based on limited proof there is a weak support for the role of n-3 PUFAs in COPD, with some evidence for the improvement of functional ability [44].

14.4.5 PROBIOTICS AND DIETARY FIBERS Probiotics may act as immunomodulatory agents and may regulate immune responses responsible for COPD. The aforementioned hypothesis has little literature confirmation. Mortaz et al. [45] suggested that the probiotic Lactobacillus casei strain Shirota (LcS) may be useful in COPD patients, especially

362

CHAPTER 14 DIET AND LUNG HEALTH

those with recurrent viral infections, since daily consumption of LcS increases natural killer activity in smokers . These evidence suggest a pathway that may interfer with COPD exacerbation pathogenesis [46,47]. Increase in dietary fiber consumption has been associated with reduced COPD risk, improvement in pulmonary function, and decreased respiratory symptoms [48].

14.4.6 PHYTOCHEMICALS Phytochemicals are chemicals produced by plants with possible health effects. For example, polyphenols are associated with low-grade systemic inflammation [49]. A link has been suggested between phytochemicals and COPD. Delphinidin and cyanidin decrease the production of interleukin (IL)-8 in cells after cigarette smoke extract treatment [50]. Lycopene, proanthocyanidin, and quercetin reduce inflammatory mediators and oxidative stress indicators in THP-1 macrophages exposed to cigarette smoke [51,52]. Green tea limits oxidative stress and protease/antiprotease imbalance in the airways after exposure to cigarette smoke [53]. Spirulina consumption decreases oxidative markers and increases antioxidants in COPD patients [53a]. Sulforaphane stimulates nuclear factor erythroid-2related factor-2 (Nrf2) activity in vitro and in vivo, thereby possibly decreasing oxidative stress and ameliorating bacterial clearance in lung macrophages [54]. Dietary intervention with phytochemicals has also grown as a treatment for COPD but the level of evidence is low and most of the studies are not encouraging. For instance, pomegranate juice supplementation adds no benefit to the current standard therapy of patients with COPD [54a]. However, treatment with a Pingchuan Guben decoction reduces symptoms, improves pulmonary function, and reduces exacerbation of COPD, and sulforaphane does not affect antioxidant levels or inflammatory markers in COPD [55]. Clearly much is yet to be discovered concerning the implication of phytochemicals in COPD.

14.5 DIET AND LOWER RESPIRATORY TRACT INFECTIONS Lower respiratory tract infections (LRTIs) are a leading cause of morbidity and mortality affecting both adults and children worldwide. Although not uniformly defined, the term LRTI usually refers to infections of the airway such as acute bronchitis and acute bronchiolitis, as well as infection of the lung alveoli (pneumonia). Community-acquired pneumonia (CAP) is one of the most common infectious diseases worldwide and is defined as pneumonia in patients who do not reside in a longterm care facility or pneumonia that occurs within 48 h after hospital admission. Pneumonia that was not incubating at the time of hospital admission and that occurs more than 2 days after hospitalization is termed hospital-acquired pneumonia. Health-care-associated pneumonia mainly affects patients that were hospitalized for more than 2 days in the past 90 days or residents of longterm care facilities. The increase in the incidence of acute LRTIs (mainly pneumonia), as well as their severity and mortality, may be associated with malnutrition, and impaired immune response may be one of the underlying mechanisms. Several nutrients have immunomodulatory roles and studies have suggested that they may be implicated in the development as well as the outcome of LRTIs.

14.5 DIET AND LOWER RESPIRATORY TRACT INFECTIONS

363

14.5.1 VITAMINS Vitamins may be involved in the development and outcome of respiratory infections. Nutrients with antioxidant properties may enhance immune response and natural defenses and therefore one may speculate that they may protect against LRTIs. However, literature data are not straightforward concerning this hypothesis. In addition to their antioxidant capacity, vitamins may exert direct effects in the immune system. For example, vitamin C enhances the function of phagocytes and the proliferation of lymphocytes. In a population-based study, vitamin C relieved oxidative stress and proinflammatory mediators associated with pneumonia [55a]. Vitamin A and C deficiency may be related to a higher burden of respiratory infections, possibly through the negative impact on immune function. However, these results have not been fully replicated by other groups that have failed to demonstrate an association of vitamins C and E and b-carotene with milder forms of respiratory infections, such as common cold [56,57]. Low levels of vitamins A and D are related to severe outcomes of LRTIs such as admission to the intensive care unit or need for mechanical ventilation. A meta-analysis concluded that vitamin A did not confer a significant benefit in preventing LRTIs [58]. In the same context, studies have examined the possible benefit of antioxidant vitamin supplementation in LRTIs, with conflicting results. Vitamin A supplementation does not affect the clinical course or reduce the severity of CAP [59]. On the other hand, vitamin C intake reduces the duration of mechanical ventilation in patients with pneumonia and decreases the pneumonia risk in patients with vitamin C shortage. A study showed reduced hospitalization duration for pneumonia in patients with vitamin C supplementation [60]. Vitamin E intake may reduce the incidence of pneumonia in elderly. However, vitamin E and b-carotene supplementation does not affect the risk of severe hospital-treated pneumonia, even though some have stated that vitamin E might augment the risk of pneumonia in patients with increased vitamin C intake. Multivitamin and mineral supplementation (including vitamins A, B1, B2, B3, B5, B6, B7, B12, C, and E and the minerals selenium, zinc, copper, and manganese) seemed to decrease the incidence of clinically diagnosed acute respiratory infections, which include common cold, flu, pharyngitis, sinusitis, laryngitis, bronchitis, and pneumonia [61]. The aforementioned results are conflicting and inconclusive concerning vitamin E and b-carotene supplementation and this may be due to the heterogeneity of the studies’ populations, since some have speculated that the beneficial effects of vitamin supplementation depend on the prevalence of vitamin deficiency in the patients. Moreover, vitamin D has attracted a lot of attention since it possesses significant immunomodulatory properties. It suppresses Th1 and enhances Th2 immune responses and induces the production of antimicrobial peptides such as cathelicidin. Observational studies have linked vitamin D deficiency with increased risk of LRTIs in both adults and children [62] and greater CAP severity in adults. The role of vitamin D supplementation in LRTI prevention and treatment has been studied as well. Data from RCTs suggest that vitamin D may have a protective role against influenza infection [63], although it is not associated with the duration of the resolution of pneumonia or the incidence of LRTIs [63a]. Interestingly, a 2017 meta-analysis suggested that vitamin D supplementation may prevent LRTIs [64]. Although data suggest that vitamin D has important implications in LRTIs, the beneficial effect of vitamin D supplementation remains unclear.

364

CHAPTER 14 DIET AND LUNG HEALTH

14.5.2 MINERALS Proofs of clinical benefits of mineral supplements like zinc in LRTIs have been provided. A recent meta-analysis concluded that zinc supplementation for more than 3 months decreases the risk of LRTIs [65]. Zinc supplementation may reduce pneumonia mortality in children and may serve as an adjuvant therapy [65a]. Zinc also reduces the incidence of severe pneumonia. In the same context, selenium reduces Ventilstor associated pneumonia (VAP) prevalence. Selenium, zinc, and copper administration is linked to reduced incidence of nosocomial pneumonia [66].

14.5.3 FATTY ACIDS PUFAs have great immunomodulatory and inflammatory effects. Experimental evidence suggests that PUFAs promote phagocytosis of Pseudomonas aeruginosa [67] as well as the survival of mice following infections of P. aeruginosa and Klebsiella pneumoniae [68]. Similarly, n-3 PUFAs amplify the phagocytic ability of mouse alveolar macrophages and reduce alveolar macrophage apoptosis by Streptococcus pneumoniae [69]. Higher consumption of a-linolenic acid and linoleic acid is related to a reduced pneumonia risk. In contrast, docosahexaenoic acid (DHA) and eicosapentaenoic acid (EPA) have been associated with increased CAP risk, while oleic acid is inversely associated with pneumonia prevalence. Dietary supplementation with DHA and arachidonic acid decreases viral infections and therefore LRTIs. n-3 intake may be beneficial in acute respiratory distress syndrome (ARDS) patients [70]. However, a meta-analysis suggests that there are no benefits of n-3 supplementation in patients with ARDS [70a]. The data concerning the role of PUFAs in respiratory tract infections (RTIs) are inconsistent.

14.5.4 PROBIOTICSePREBIOTICS Data support the use of probiotics for the prevention and treatment of gastroenteritis. The possible role of probiotics in respiratory infections has attracted attention as well. For example, probiotics like lactic acid bacteria have proven effects in innate and humoral immune responses against S. pneumoniaee associated LRTIs. Meta-analyses support that the use of probiotics is related to a statistically notable decrease in the prevalence of hospital-induced pneumonia [71]. Similarly, consumption of probiotics is associated with lower occurrence of VAP. Probiotic introduction in the diet reduces cases of LRTIs as well as RTIs in the elderly. Bernard et al. [72] demonstrated that prebiotics, such as pectin-derived acidic oligosaccharides taken from citrus, improved the outcomes of P. aeruginosa lung infection in mice by regulating the intestinal microbiota and the inflammatory and immune responses. As there was a decrease in bacterial load, pectin-derived acidic oligosaccharides could be suggested as an adjuvant therapy to antibiotics. A systematic review and network meta-analysis demonstrated that synbiotic therapy (combination of probiotics and prebiotics) was the best regulated course in decreasing pneumonia in adult surgical patients in comparison with probiotics and prebiotics alone [72a]. Although the research data are encouraging, further studies are warranted before any definite suggestions are made concerning the role of probiotics and prebiotics in LRTIs.

14.6 DIET AND TUBERCULOSIS

365

14.5.5 PHYTOCHEMICALS Data suggest that phytochemicals may have a preventive and therapeutic role in RTIs; however, the quality of evidence is low [72b] [73]. Green tea consumption is related to a lower risk of death from pneumonia. Pleuran decreases the number of pneumonia and bronchitis cases in children. Manuka honey has antimicrobial activities against P. aeruginosa. Extract from North American ginseng (Panax quinquefolius) suspends P. aeruginosa development. Interestingly, combination therapy of curcumin with antibiotics in mouse models reduces inflammation, bacterial proliferation, and lung tissue injury.

14.6 DIET AND TUBERCULOSIS TB represents an infectious disease usually caused by Mycobacterium tuberculosis (Mtb). One-third of the world population is infected with Mtb, while TB was responsible for 1.3 million deaths in 2016 [74]. About 90% of patients do not present symptoms and suffer from latent TB, while active cases present with cough, hemoptysis, fever, night sweats, and weight loss. TB can affect any part of the body although it most commonly affects the lungs. The disease pathogenesis commonly involves the formation of caseous granulomas in infiltrated tissues involving macrophages, T lymphocytes, B lymphocytes, and fibroblasts. Treatment of TB involves the combination of anti-TB drugs for long periods of time.

14.6.1 DIETARY PATTERNS Several factors make people more susceptible to TB. Malnutrition among others has been recognized as a risk factor for pulmonary TB [75]. Data on the role of dietary patterns in TB are sparse. Poor fruit and vegetable consumption is related to an increased susceptibility to infection with Mtb but not active TB. In an animal experiment it was found that alcohol consumption dulled the development of the adaptive immune response to BCG vaccination [76].

14.6.2 VITAMINS Although studies have shown deficiency of several antioxidant vitamins in TB, the available data cannot confirm causality. Experimental data suggest that vitamins like vitamins C and E have bactericidal activity against Mtb and thus may exert potential benefits in anti-TB treatment. TB patients present low levels of vitamin A, suggesting it as a risk factor of TB infection and clinical severity. Similar results have been reported for vitamins A, C, and E. Taking these findings into consideration, nutritional supplements could help people recover from TB because of their effects on the immune system. However, there are no clear data that routine supplementation provides clinically important benefits. Although vitamin supplementation may help in the treatment of these patients in terms of earlier sputum smear conversion or reduced hepatotoxicity, no definite conclusions can be drawn yet. Observational studies have shown a strong positive association between low levels of vitamin D and TB prevalence. The association of vitamin D deficiency and TB may be due to the enhanced production of cathelicidin, an antimicrobial peptide, which is implicated in the innate immune response to Mtb infection [77]. In addition, vitamin D hypovitaminosis is a risk factor for worse

366

CHAPTER 14 DIET AND LUNG HEALTH

treatment outcomes, such as delayed sputum conversion. Vitamin D deficiency has been widely accepted as a risk factor for active TB. The effect of vitamin D supplementation in TB remains controversial. Adjunct therapy with vitamin D has some beneficial effects toward clinical recovery. The results are inconsistent among the studies. At this point the beneficial effect of vitamin D supplementation remains unclear, and highquality studies are necessary to establish the role of vitamin D manipulation in TB.

14.6.3 MINERALS Patients with TB present low serum concentrations of zinc and selenium, and high levels of copper and cobalt [78]. Low plasma selenium is a risk factor related to anemia in these patients [79]. High copper is related to worse outcomes and low zinc is associated with clinical severity. Experimental evidence suggests that copper is essential for the control of Mtb infection [79a]. Further studies in this field are clearly needed.

14.6.4 FATTY ACIDS The impact of PUFAs on TB is not well known. PUFA supplementation might have a harmful effect on immunity against Mtb [80]. Experimental data have linked n-3 PUFAs with reduced skin test positivity and reduced inflammatory response to Mtb. Although observational data suggest that n-3 and n-6 consumption may be related to a reduced risk of active TB, overall the available data suggest that PUFAs may increase susceptibility to Mtb infection [81]. Clearly well-designed studies are warranted to elucidate the role of PUFAs in TB in vivo.

14.6.5 PROBIOTICS Strains of lactic acid bacteria, such as Lactobacillus rhamnosus GG and Bifidobacterium bifidum MF 20/5, increased the autophagic capacity of mononuclear phagocytes in response to Mtb antigen in a vivo study [82]. Unfortunately, there are no studies about the impact of probiotics on TB.

14.6.6 PHYTOCHEMICALS Many studies have been conducted on the potential effects of medicinal plants as alternative and adjunctive therapies for the treatment of TB. It has been reported that ursolic acid and hydroquinone may have chemotherapeutic potency against Mtb and immunoregulatory properties against TB in mice [83]. C. longa provides defense against Mtb infection in alveolar macrophages, via repression of nuclear factor-kB activation [84]. In addition, Aristolochia brevipes, garlic and garlic-derived fatty acids, and Ambrosia confertiflora have antibacterial activity against Mtb. Excoecaria agallocha may protect against multidrug-resistant TB [85]. Pomegranate fruit inhibits Mtb and Manuka honey may be efficient as a hepatoprotective agent. Although the aforementioned data are encouraging, the lack of well-designed clinical trials limits their importance.

14.7 DIET AND LUNG CANCER

367

14.7 DIET AND LUNG CANCER LC is the leading cause of cancer death, and smoking causes the vast majority of cases, with the remaining often due to a combination of genetic factors and exposure to radon or asbestos [86]. Most patients are not curable, thus avoidance of risk factors, mainly smoking, is of special importance to prevent the disease. An issue of ongoing research is the effects of diet on LC development and its therapeutic potential.

14.7.1 DIETARY PATTERNS Emerging data have indicated that dietary habits play a crucial role in LC development and as a result they have attracted increased attention. Processed meat and meat mutagens (found in fried, barbecued, and processed meat), red meat, white bread, fat and high dietary glycemic index foods, and polycyclic aromatic hydrocarbons (found in charcoal-broiled, fried, and smoked meat) have all been associated with increased LC risk mainly among smokers. On the opposite side, consumption of fish, fruits, and vegetables in an overall healthy dietary pattern has been consistently associated with lower LC risk. Of great importance seems to be adherence to the Mediterranean diet, which has strong evidence in favor of a reduced LC risk [87]. According to published data, the Mediterranean diet exerts protective effects against various forms of cancer, including LC.

14.7.2 VITAMINS Oxidative stress has been implicated in LC development and thus one may hypothesize that dietary antioxidants may prevent carcinogenesis. Several epidemiological studies have associated antioxidant intake with lower LC risk. Intake of vitamins A, C, and E may provide protection against LC development but with a possible modification of this effect by smoking habit [87a]. Similar data have been reported for riboflavin (vitamin B2) and folate (vitamin B9). Interventional studies have not confirmed these results, however, while the use of antioxidant supplements may be harmful. In a double-blind RCT that involved almost 30,000 male smokers, there was an association of b-carotene supplementation with higher incidence of LC as early as 18 months after the initiation of the study [39]. Similarly, a study involving 18,000 men and women at high risk for LC supplemented with b-carotene and vitamin A was stopped early because of a higher death rate in the antioxidant groups [88]. The mechanism underlying the potential increase in LC incidence is not known, but studies have suggested that antioxidants may promote tumor growth and metastasis. The effect of antioxidant supplements during LC treatment is not well studied. Data have shown that vitamin C addition combined with chemotherapy may inhibit cell proliferation in cell lines [88a]. However, the results have been inconsistent and some have reported worse outcomes [89]. The role of vitamin D in LC seems beneficial. Vitamin D may exhibit anticarcinogenic effects via the inhibition of cell proliferation and angiogenesis, as well as the enhancement of apoptosis and cell differentiation. Experiments have shown that vitamin D intake might provide protection against LC development and metastatic ability by giving off of E-cadherin and catenin, which help the adherence of cancer cells and decrease the possibility of metastasis [90]. In the same context, high levels of 25hydroxyvitamin D may be related to a decreased risk of LC, particularly in people with vitamin D

368

CHAPTER 14 DIET AND LUNG HEALTH

deficiency. Some have reported that the protective nature of vitamin D against LC is limited to women and nonsmokers. Data have connected positively vitamin D deficiency with LC mortality. Unfortunately, the literature lacks data from well-designed studies and the causal direction of these results cannot be established.

14.7.3 MINERALS Dietary mineral intake may influence LC development but the direction of the association depends on the type of mineral. A prospective cohort study among 482,875 participants suggested that total calcium consumption was protective for smokers and individuals with adenocarcinoma, while total magnesium intake increased risk in men and smokers and total iron consumption was inversely related to risk in women. Mineral intake from supplements did not influence LC risk [91]. A diet rich in zinc and iron may be related to reduced risk but no relationship was found between selenium, calcium, magnesium, and copper consumption [92]. Others have associated low selenium with high risk of LC [93], while selenium supplementation is not beneficial in terms of prevention of second primary tumors in patients with resected non-small-cell LC (NSCLC).

14.7.4 FATTY ACIDS Fatty acids, especially PUFAs, have been associated with a lower predominance of various types of cancer. Experimental studies support the protective role of PUFAs in LC possibly through increased oxidative stress and apoptosis of cancer cells and modulation of cyclooxygenase activity and cell surface receptors [94]. The aforementioned data are consistent with observational data suggesting that high consumption of fish, which is rich in PUFAs, is related to a reduction of LC risk even after adjustment for smoking status. Supplementation of EPA increased lean body mass, energy, and protein intake and reduced fatigue and neuropathy in NSCLC patients undergoing chemotherapy [95]. Adjunctive n-3 supplementation in LC therapy may prevent cachexia and improve performance status and physical activity in LC patients [96]. Due to the limited available data that are generally based on small cohorts, no definitive conclusions can be drawn.

14.7.5 PHYTOCHEMICALS Many phytochemicals have been studied in the context of LC because of their possible role in LC cell apoptosis, chromatin remodeling, and DNA methylation, among other pathways. Quercetin-rich food is inversely linked with LC risk, while quercetin enhances the activity of anticancer drugs. (6)-Shogaol suppresses the growth of NSCLC cells and enhances apoptosis in cancer cells. Silibinin in combination with epigenetic drugs (histone deacetylase or DNA methyltransferase inhibitor) inhibits both aggression and migration of NSCLC cells and could be a potent treatment for more advanced stages of NSCLC, while intake of the combination of indole-3-carbinol and silibinin may protect against LC. Both green and black tea consumption is related to lower risk for LC. Triptolide, which has been studied for its antirheumatic effects, shows anticancer properties against NSCLC. Chlorella sorokiniana causes mitochondrial-mediated apoptosis in NSCLC cells, and Chlorella vulgaris inhibits LC cell growth and migration.

14.8 DIET AND CYSTIC FIBROSIS

369

Studies also assessed the inhibitory activity of curcumin on LC cell line growth and metastasis in vitro and in vivo as well as its reinforcement of anticancer drugs. A few data proved the inhibitory effect of Panax ginseng on tumor growth and lung metastasis in vitro and in animals. Unfortunately, RCTs are lacking and the aforementioned findings are mostly based on experimental studies.

14.8 DIET AND CYSTIC FIBROSIS CF is a genetic disorder caused by mutations of the gene for the CF transmembrane conductance regulator (CFTR) protein. CFTR is involved in sweat, digestive fluid, and mucus production and when affected it results in thicker secretions. The disease affects the lungs via mucus clogging of the airways, resulting in inflammation and infection. The thick mucus blocks the digestive system, causing meconium ileus, and in later life the thickened secretions of the pancreas block the exocrine part of the organ and result in irreversible damage. Due to malabsorption, patients with CF are often malnourished and have poor growth. Replacement of digestive enzymes is indicated in CF, and the patients present malabsorption of fat-soluble vitamins (A, D, E, and K). Also, they exhibit signs of oxidative stress, due to inflammation and the malabsorption of vitamins A and E. CF patients should consume 120%e150% of the recommended daily allowance (RDA) for energy expenditure, with 40% coming from fat. Nutritional supplementation may include n-3 PUFAs; vitamins A, D, E, and K; and proteins 120% of RDA [97]. Concerning the lung, in the early stages, inflammation and decreased mucociliary clearance result in copious phlegm production and insufficient coughing. In later life patients develop structural changes in the lungs, mainly bronchiectasis. Here we will report the main findings on the role of dietary compounds in the disease.

14.8.1 VITAMINS Vitamin C reduces with age in people with CF and supplementation with antioxidant vitamins may retard the deterioration of pulmonary function. Experimental data suggest that vitamin C activates the CFTR channel in the respiratory tract [98]. Clinical studies have shown that vitamin C concentration in CF patients is associated with indexes of lung inflammation. High levels of vitamin A are related to improved lung function in CF patients, and b-carotene supplementation could benefit CF patients. Vitamin E levels have been reported low in CF. Little is known about vitamin A. High a-tocopherol levels have no beneficial effects on pulmonary function in CF patients. RCTs in the field are lacking. Data on vitamin D are rather encouraging. Vitamin D supplementation has antimicrobial and antiinflammatory properties in CF patients and is related to a reduction in inflammatory cytokines, i.e., IL-6 and tumor necrosis factor-a [99,100]. Pulmonary exacerbations have been associated with vitamin D deficiency. In the same context, vitamin D consumption may positively regulate inflammation in CF by decreasing serum total IgG levels and serum haptoglobin [101].

14.8.2 MINERALS Serum levels of calcium, copper, and iron in CF patients are reduced during exacerbation of the disease. Zinc deficiency is common in CF patients and low zinc levels are associated with worse

370

CHAPTER 14 DIET AND LUNG HEALTH

pulmonary function. However, zinc supplementation does not improve pulmonary function and neither decreases pulmonary infections [102]. CF patients have lower levels of selenium. Iron deficiency is usual in adults with CF and associated with disease severity. Iron administration does not reduce respiratory symptoms or change the sputum microbiome in CF patients [103]. CF patients have increased lung iron levels that enhance ROS production and encourage bacterial growth [104]. High iron levels in the airways of CF patients may be conducive to the sensitivity to chronic bacterial infections that are related to CF. Administration of iron chelators might isolate host iron and prevent the approachability of iron to bacteria [105].

14.8.3 FATTY ACIDS Low-dose supplementation with n-3 and n-6 fatty acids improves lung function, respiratory exacerbation, antibiotic consumption, lean body mass, inflammation, and oxidative markers in patients with CF. An RCT also showed that n-3 PUFAs may decrease the leukotriene LTB4/LTB5 ratio, demonstrating antiinflammatory effects [106].

14.8.4 PROBIOTICS A systematic review revealed that the effectiveness of probiotics in children with CF is restricted [107].

14.8.5 PHYTOCHEMICALS Data on phytochemicals and CF are mostly based on experimental studies. Curcumin stimulates CFTR Cl channels [108]. Garlic improves lung function in CF patients having chronic P. aeruginosa infection. Rhodiola kirilowii (Regel) Maxim. and cocoa flavanols have antidiarrheal activity in CF patients via inhibition of CFTR Cl channel activity [109]. Also, bergamot (Citrus bergamia Risso) extract inhibited IL-8 expression in an in vitro study [110].

14.9 DIET AND INTERSTITIAL LUNG DISEASES ILDs are a group of heterogeneous diseases that affect the lung interstitium. The group includes many different conditions. Here we will emphasize two of the most common ILDs, idiopathic pulmonary fibrosis (IPF) and sarcoidosis. Reduced antioxidant defense, causing oxidative stress, and inflammation are two key triggers in the pathophysiology of IPF. A diet rich in antioxidants might have an advantageous effect on ILD patients and some argue that it might protect against the development of IPF. In addition, fruit consumption may prevent the development of IPF. A study found a reduction in the risk of IPF, related to high consumption of green tea and vegetables and higher intake of fish [110a]. High consumption of vitamin A has been related to decreased rate of progression in asbestosrelated lung fibrosis [111]. However, the role of antioxidants is probably of limited importance since the administration of N-acetylcysteine in IPF patients offers no significant benefits [111a,b]. Vitamin D deficiency is known in patients with ILD. Sarcoidosis patients should be supplemented with vitamin D with caution and only when diagnosed with vitamin D deficiency due to the high levels of 1-a-hydroxylase expressed in sarcoid granulomas.

14.10 MATERNAL DIET IN EARLY LIFE AND LUNG HEALTH

371

Experimental studies have suggested the potential therapeutic properties of phytochemicals in ILDs. Fenugreek (Trigonella foenum-graecum) seed extract and berberine exhibit antifibrotic results through induction of Nrf2 (an antioxidant balance regulation factor) and inhibit profibrogenic molecules in bleomycin-induced rat models [112]. Citrus reticulata (commonly known as mandarin orange) extract inhibits the proliferation of human lung fibroblasts and inflammation [113]. Flaxseed oil, which is rich in n-3 and n-6 PUFAs, attenuates bleomycin-induced pulmonary fibrosis in rats and reduces pulmonary oxidative stress [114]. Intraperitoneal curcumin administration exhibited antifibrotic effects on bleomycin-induced pulmonary fibrosis in mice, and curcumin administration protects against the development of carbon tetrachloride-induced pulmonary fibrosis. Hesperidin reduces the severity of pulmonary fibrosis in rats. Nigella sativa has antiinflammatory and antifibrotic effects on bleomycin-induced pulmonary fibrosis in rats. Administration of quercetin could benefit IPF patients by reducing inflammation and oxidative stress and retarding disease progression [115]. Similarly, Boots et al. [116,117] indicated that quercetin might be useful for sarcoidosis patients by decreasing oxidative stress and inflammation.

14.10 MATERNAL DIET IN EARLY LIFE AND LUNG HEALTH Maternal diet in pregnancy is considered to be one of the most relevant prenatal and early postnatal risk factors for the development of respiratory diseases. The development of the lungs is almost completed prenatally and, following birth, airway development is mostly restricted to size growth. Thus, fetal or early life exposure may have disproportional effects on the development of respiratory diseases [118]. Research has focused on the obstructive lung diseases such as asthma and COPD.

14.10.1 DIETARY PATTERNS Various dietary patterns may contribute to lung development. Experiments in animals showed that maternal high-fat diet and hypercaloric diet were associated with deteriorated fetal lung development, airway hyperresponsiveness, and chronic airway inflammation in the offspring [119]. Epidemiological data suggest that high maternal intake of peanuts, milk, wheat, leafy vegetables, malaceous fruits, and chocolate may be related to a decreased likelihood of childhood wheeze and/or asthma, but not consistently. Better adherence to the Mediterranean diet during pregnancy seems to be preventive for wheeze in the offspring. Subsequently, early import of solid food has been associated with reduced probability of developing asthma. The effect of breastfeeding for 4e6 months or longer is related to reduced risk of developing asthma and LRTI occurrence in the offspring [120].

14.10.2 VITAMINS Maternal dietary antioxidant intake may contribute to increases in asthma risk. Low maternal vitamin E and zinc intake is associated with increased asthmalike symptoms, while vitamin E levels during pregnancy are positively related to postbronchodilator FEV1 in children [121]. Vitamin E affects positively neonatal airway epithelial cell secretory function [122]. Increased intake of vitamin C in pregnancy is associated with wheezing. Interestingly, maternal antioxidant intake may decrease the risk of RTIs in the offspring. The data support the association of antioxidants with childhood asthma

372

CHAPTER 14 DIET AND LUNG HEALTH

but must be interpreted with caution, since the observational nature of the studies cannot establish causality. The data concerning vitamin D are rather conflicting. Experimental evidence suggests that prenatal vitamin D is associated with lung development and surfactant synthesis. Low maternal and low neonatal vitamin D levels have been linked to bronchopulmonary dysplasia, childhood asthma, and wheeze. Others have challenged these results, reporting no relation of maternal levels of vitamin D with asthma and wheeze in the offspring [123]. Low levels of vitamin D in newborns and low vitamin D blood levels have been linked to increased risk of LRTIs and higher airway resistance. High maternal selenium and copper levels may prevent wheeze, an effect that is limited to the first years of life [124]. The data support the association of antioxidants with childhood asthma but must be interpreted with caution, since the observational nature of the studies cannot establish causality.

14.10.3 FATTY ACIDS PUFA intake in pregnancy seems to have a significant role in respiratory symptoms in the offspring. Increased content of n-3 fatty acids in the maternal diet has protective effects against allergic diseases (eczema, rhinoconjunctivitis, asthma) in the offspring. In the same context, fish oil supplementation during pregnancy may prevent childhood asthma, although the relationship may be limited to asthmatic mothers [125]. Supplementation with n-3 fatty acids in infants may influence later respiratory health. Infants who consumed PUFAs had reduced odds of wheezing at 18 months, a result that is dulled at 3 and 5 years of age [126]. Also, early introduction of fish into the diet (at