Introduction to Meta-Analysis

ISBN-13: 9781119558354 / Angielski / Twarda / 2021 / 544 str.

Michael Borenstein; Larry V. Hedges; Julian P. T. Higgins

Spis treści

List of Tables xvList of Figures xixAcknowledgements xxvPreface xxviiPreface to the Second Edition xxxvWebsite xxxviiPart 1: Introduction1 How a Meta-Analysis Works 3Introduction 3Individual studies 3The summary effect 5Heterogeneity of effect sizes 6Summary points 72 Why Perform a Meta-Analysis 9Introduction 9The streptokinase meta-analysis 10Statistical significance 11Clinical importance of the effect 11Consistency of effects 12Summary points 13Part 2: Effect Size and Precision3 Overview 17Treatment effects and effect sizes 17Parameters and estimates 18Outline of effect size computations 194 Effect Sizes Based On Means 21Introduction 21Raw (unstandardized) mean difference D 21Standardized mean difference, d and g 25Response ratios 30Summary points 315 Effect Sizes Based On Binary Data (2 × 2 Tables) 33Introduction 33Risk ratio 33Odds ratio 35Risk difference 37Choosing an effect size index 38Summary points 386 Effect Sizes Based On Correlations 39Introduction 39Computing r 39Other approaches 40Summary points 417 Converting Among Effect Sizes 43Introduction 43Converting from the log odds ratio to d 44Converting from d to the log odds ratio 45Converting from r to d 45Converting from d to r 46Summary points 478 Factors That Affect Precision 49Introduction 49Factors that affect precision 50Sample size 50Study design 51Summary points 539 Concluding Remarks 55Part 3: Fixed-Effect Versus Random-Effects Models10 Overview 59Introduction 59Nomenclature 6011 Fixed-Effect Model 61Introduction 61The true effect size 61Impact of sampling error 61Performing a fixed-effect meta-analysis 63Summary points 6412 Random-Effects Model 65Introduction 65The true effect sizes 65Impact of sampling error 66Performing a random-effects meta-analysis 68Summary points 7013 Fixed-Effect Versus Random-Effects Models 71Introduction 71Definition of a summary effect 71Estimating the summary effect 72Extreme effect size in a large study or a small study 73Confidence interval 73The null hypothesis 76Which model should we use? 76Model should not be based on the test for heterogeneity 78Concluding remarks 79Summary points 7914 Worked Examples (Part 1) 81Introduction 81Worked example for continuous data (Part 1) 81Worked example for binary data (Part 1) 85Worked example for correlational data (Part 1) 90Summary points 94Part 4: Heterogeneity15 Overview 97Introduction 97Nomenclature 98Worked examples 9816 Identifying and Quantifying Heterogeneity 99Introduction 99Isolating the variation in true effects 99Computing Q 101Estimating tau² 106The I² statistic 109Comparing the measures of heterogeneity 111Confidence intervals for tau² 114Confidence intervals (or uncertainty intervals) for I² 115Summary points 11617 Prediction Intervals 119Introduction 119Prediction intervals in primary studies 119Prediction intervals in meta-analysis 121Confidence intervals and prediction intervals 123Comparing the confidence interval with the prediction interval 123Summary points 12518 Worked Examples (Part 2) 127Introduction 127Worked example for continuous data (Part 2) 127Worked example for binary data (Part 2) 131Worked example for correlational data (Part 2) 134Summary points 13819 An Intuitive Look At Heterogeneity 139Introduction 139Motivating example 140The Q-value and the p-value do not tell us howmuch the effect size varies 141The confidence interval does not tell us how much the effect size varies 142The I² statistic does not tell us how much the effect size varies 142What I² tells us 142The I² index vs. the prediction interval 145The prediction interval 145Prediction interval is clear, concise, and relevant 147Computing the prediction interval 147How to use I² 149How to explain heterogeneity 149How much does the effect size vary across studies? 150Caveats 150Conclusion 150Further reading 151Summary points 151The meaning of I² in Figure 19.2 15120 Classifying Heterogeneity As Low, Moderate, Or High 155Introduction 155Interest should generally focus on an index of absolute heterogeneity 155The classifications lead themselves to mistakes of interpretation 158Classifications focus attention in the wrong direction 158Summary points 158Part 5: Explaining Heterogeneity21 Subgroup Analyses 161Introduction 161Fixed-effect model within subgroups 163Computational models 172Random effects with separate estimates of tau² 174Random effects with pooled estimate of tau² 181The proportion of variance explained 189Mixed-effects model 192Obtaining an overall effect in the presence of subgroups 193Summary points 19522 Meta-Regression 197Introduction 197Fixed-effect model 198Fixed or random effects for unexplained heterogeneity 203Random-effects model 206Summary points 21223 Notes On Subgroup Analyses and Meta-Regression 213Introduction 213Computational model 213Multiple comparisons 216Software 216Analyses of subgroups and regression analyses are observational 217Statistical power for subgroup analyses and meta-regression 218Summary points 219Part 6: Putting It All In Context24 Looking At the Whole Picture 223Introduction 223Methylphenidate for adults with ADHD 226Impact of GLP-1 mimetics on blood pressure 228Augmenting clozapine with a second antipsychotic 228Conclusions 231Caveats 231Summary points 23225 Limitations of the Random-Effects Model 233Introduction 233Assumptions of the random-effects model 234A textbook case 234When studies are pulled from the literature 235A useful fiction 237Transparency 238A narrowly defined universe 238Two important caveats 239In context 239Extreme cases 240Summary points 24126 Knapp-Hartung Adjustment 243Introduction 243Adjustment is rarely employed in simple analyses 243Adjusting the standard error 244The Knapp-Hartung adjustment for other effect size indices 246t distribution vs. Z distribution 247Limitations of the Knapp-Hartung adjustment 248Summary points 249Part 7: Complex Data Structures27 Overview 25328 Independent Subgroups Within a Study 255Introduction 255Combining across subgroups 255Comparing subgroups 260Summary points 26029 Multiple Outcomes or Time-Points Within A Study 263Introduction 263Combining across outcomes or time-points 264Comparing outcomes or time-points within a study 270Summary points 27530 Multiple Comparisons Within a Study 277Introduction 277Combining across multiple comparisons within a study 277Differences between treatments 278Summary points 27931 Notes On Complex Data Structures 281Introduction 281Summary effect 281Differences in effect 282Part 8: Other Issues32 Overview 28733 Vote Counting - A New Name For An Old Problem 289Introduction 289Why vote counting is wrong 290Vote counting is a pervasive problem 291Summary points 29334 Power Analysis For Meta-Analysis 295Introduction 295A conceptual approach 295In context 299When to use power analysis 300Planning for precision rather than for power 301Power analysis in primary studies 301Power analysis for meta-analysis 304Power analysis for a test of homogeneity 309Summary points 31235 Publication Bias 313Introduction 313The problem of missing studies 314Methods for addressing bias 316Illustrative example 317The model 317Getting a sense of the data 318Is there evidence of any bias? 320How much of an impact might the bias have? 320Summary of the findings for the illustrative example 324Conflating bias with the small-study effect 325Using logic to disentangle bias from small-study effects 326These methods do not give us the 'correct' effect size 327Some important caveats 327Procedures do not apply to studies of prevalence 328The model for publication bias is simplistic 328Concluding remarks 329Putting it all together 330Summary points 330Part 9: Issues Related To Effect Size36 Overview 33537 Effect Sizes Rather Than P-Values 337Introduction 337Relationship between p-values and effect sizes 337The distinction is important 339The p-value is often misinterpreted 340Narrative reviews vs. meta-analyses 341Summary points 34238 Simpson's Paradox 343Introduction 343Circumcision and risk of HIV infection 343An example of the paradox 345Summary points 34839 Generality of the Basic Inverse-Variance Method 349Introduction 349Other effect sizes 350Other methods for estimating effect sizes 353Individual participant data meta-analyses 354Bayesian approaches 355Summary points 357Part 10: Further Methods40 Overview 36141 Meta-Analysis Methods Based On Direction and P-Values 363Introduction 363Vote counting 363The sign test 363Combining p-values 364Summary points 36842 Further Methods For Dichotomous Data 369Introduction 369Mantel-Haenszel method 369One-step (Peto) formula for odds ratio 373Summary points 37643 Psychometric Meta-Analysis 377Introduction 377The attenuating effects of artifacts 378Meta-analysis methods 380Example of psychometric meta-analysis 381Comparison of artifact correction with meta-regression 384Sources of information about artifact values 384How heterogeneity is assessed 385Reporting in psychometric meta-analysis 386Concluding remarks 386Summary points 387Part 11: Meta-Analysis In Context44 Overview 39145 When Does It Make Sense To Perform a Meta-Analysis? 393Introduction 393Are the studies similar enough to combine? 394Can I combine studies with different designs? 395How many studies are enough to carry out a meta-analysis? 399Summary points 40046 Reporting The Results of a Meta-Analysis 401Introduction 401The computational model 402Forest plots 402Sensitivity analysis 404Summary points 40547 Cumulative Meta-Analysis 407Introduction 407Why perform a cumulative meta-analysis? 409Summary points 41248 Criticisms of Meta-Analysis 413Introduction 413One number cannot summarize a research field 414The file drawer problem invalidates meta-analysis 414Mixing apples and oranges 415Garbage in, garbage out 416Important studies are ignored 417Meta-analysis can disagree with randomized trials 417Meta-analyses are performed poorly 420Is a narrative review better? 420Concluding remarks 422Summary points 42249 Comprehensive Meta-Analysis Software 425Introduction 425Features in CMA 426Teaching elements 427Documentation 427Availability 427Acknowledgments 427Motivating example 428Data entry 428Basic analysis 429What is the average effect size? 430How much does the effect size vary? 430Plot showing distribution of effects 431High-resolution plot 432Subgroup analysis 433Meta-regression 435Publication bias 438Explaining results 43950 How To Explain the Results of An Analysis 443Introduction 443The overview 444The mean effect size 444Variation in effect size 444Notations 444Impact of resistance exercise on pain 445Correlation between letter knowledge and word recognition 450Statins for prevention of cardiovascular events 455Bupropion for smoking cessation 460Mortality following mitral-valve procedures in elderly patients 465Part 12: Resources51 Software For Meta-Analysis 471Comprehensive meta-analysis 471Metafor 471Stata 472Revman 47252 Web Sites, Societies, Journals, and Books 473Web sites 473Professional societies 476Journals 476Special issues dedicated to meta-analysis 477Books on systematic review methods and meta-analysis 477References 479Index 491

O autorze

Michael Borenstein is the Director of Biostat, a leading developer of statistical software. He is the primary developer of Comprehensive Meta-Analysis (CMA), the world's most widely used program for meta-analysis. He is the recipient of numerous grants from the NIH to develop methods, software, and educational materials for meta-analysis. He has lectured widely on meta-analysis, including at the NIH, CDC, and FDA.Larry V. Hedges is Board of Trustees Professor of Statistics and Education and Social Policy, Professor of Psychology, Professor of Medical Social Sciences, and IPR Fellow, Northwestern University, USA. He is a national leader in the fields of educational statistics and evaluation and is an elected member of many leading associations.Julian P.T. Higgins is Professor of Evidence Synthesis at the University of Bristol, UK, and a National Institute for Health Research (NIHR) Senior Investigator. He has had numerous core roles in the Cochrane Collaboration, including editing its methodological Handbook since 2003. His many contributions to meta-analysis include the foundation of network meta-analysis, methods for describing and explaining heterogeneity and a general framework for individual participant data meta-analysis. He is a Highly Cited Researcher with over a quarter of a million citations to his work and has been a recipient of the Ingram Olkin Award for distinguished lifetime achievement in research synthesis methodology.Hannah R. Rothstein is Professor of Management at Baruch College and the Graduate Center of the City University of New York. She is a Fellow of the American Psychological Association and a past President of the Society for Research Synthesis Methodology. She is former Editor-in-Chief of Research Synthesis Methods and serves on the editorial boards of Psychological Bulletin, Psychological Methods, and Organizational Research Methods. Professor Rothstein is a co-developer of the Comprehensive Meta-Analysis software and has published numerous systematic reviews and meta-analyses.