Biostat Methods STAT 5820/6910 Handout #9a: Intro. to Meta-Analysis Methods

Similar documents
EPSE 594: Meta-Analysis: Quantitative Research Synthesis

Answer Key for STAT 200B HW No. 8

Regression #3: Properties of OLS Estimator

Central Limit Theorem ( 5.3)

Poisson regression: Further topics

STATS 200: Introduction to Statistical Inference. Lecture 29: Course review

A new strategy for meta-analysis of continuous covariates in observational studies with IPD. Willi Sauerbrei & Patrick Royston

COMPLETELY RANDOMIZED DESIGNS (CRD) For now, t unstructured treatments (e.g. no factorial structure)

Master s Written Examination

Biostat Methods STAT 5500/6500 Handout #12: Methods and Issues in (Binary Response) Logistic Regression

What is a meta-analysis? How is a meta-analysis conducted? Model Selection Approaches to Inference. Meta-analysis. Combining Data

Analysis of Variance

Statistics and Econometrics I

Math 494: Mathematical Statistics

Florida State University Libraries

Previous lecture. Single variant association. Use genome-wide SNPs to account for confounding (population substructure)

Biostat Methods STAT 5820/6910 Handout #5a: Misc. Issues in Logistic Regression

STAT 5200 Handout #7a Contrasts & Post hoc Means Comparisons (Ch. 4-5)

Probability Theory and Statistics. Peter Jochumzen

Fall 2017 STAT 532 Homework Peter Hoff. 1. Let P be a probability measure on a collection of sets A.

Stat 5102 Final Exam May 14, 2015

BIO5312 Biostatistics Lecture 13: Maximum Likelihood Estimation

STAT 135 Lab 3 Asymptotic MLE and the Method of Moments

STAT 135 Lab 5 Bootstrapping and Hypothesis Testing

UNIVERSITY OF TORONTO Faculty of Arts and Science

Some Curiosities Arising in Objective Bayesian Analysis

STAT 525 Fall Final exam. Tuesday December 14, 2010

MISCELLANEOUS TOPICS RELATED TO LIKELIHOOD. Copyright c 2012 (Iowa State University) Statistics / 30

Practical Meta-Analysis -- Lipsey & Wilson

STAT 430 (Fall 2017): Tutorial 8

Minimum Message Length Analysis of the Behrens Fisher Problem

Statistical Data Analysis Stat 3: p-values, parameter estimation

Module 22: Bayesian Methods Lecture 9 A: Default prior selection

Statistics 135 Fall 2007 Midterm Exam

Lectures on Simple Linear Regression Stat 431, Summer 2012

BTRY 4090: Spring 2009 Theory of Statistics

F & B Approaches to a simple model

Replicability and meta-analysis in systematic reviews for medical research

Inference Conditional on Model Selection with a Focus on Procedures Characterized by Quadratic Inequalities

Regression Estimation - Least Squares and Maximum Likelihood. Dr. Frank Wood

Meta-analysis of binary outcomes via generalized linear mixed models: a simulation study

The outline for Unit 3

Ph.D. Qualifying Exam Friday Saturday, January 6 7, 2017

Much of the material we will be covering for a while has to do with designing an experimental study that concerns some phenomenon of interest.

STAT215: Solutions for Homework 2

One-way ANOVA (Single-Factor CRD)

Chapter 2: Simple Random Sampling and a Brief Review of Probability

1 One-way analysis of variance

Workshop on Statistical Applications in Meta-Analysis

Lecture 2: Basic Concepts and Simple Comparative Experiments Montgomery: Chapter 2

Review. Timothy Hanson. Department of Statistics, University of South Carolina. Stat 770: Categorical Data Analysis

Theory of Statistics.

Bias Variance Trade-off

Master s Written Examination

CAMPBELL COLLABORATION

The Multilevel Logit Model for Binary Dependent Variables Marco R. Steenbergen

A re-appraisal of fixed effect(s) meta-analysis

Statistical Models with Uncertain Error Parameters (G. Cowan, arxiv: )

Categorical Predictor Variables

arxiv: v1 [stat.me] 16 Jun 2009

Bayesian linear regression

A Significance Test for the Lasso

Generating the Sample

arxiv: v2 [stat.me] 1 Aug 2009

Association studies and regression

Ling 289 Contingency Table Statistics

Statistics - Lecture One. Outline. Charlotte Wickham 1. Basic ideas about estimation

Meta-analysis. 21 May Per Kragh Andersen, Biostatistics, Dept. Public Health

Lecture 25. Ingo Ruczinski. November 24, Department of Biostatistics Johns Hopkins Bloomberg School of Public Health Johns Hopkins University

STAT 526 Spring Midterm 1. Wednesday February 2, 2011

Math 423/533: The Main Theoretical Topics

Introduction to Statistical Inference

Outline of GLMs. Definitions

Prediction intervals for random-effects meta-analysis: a confidence distribution approach

Multivariate Survival Analysis

Bayesian Linear Models

Cluster investigations using Disease mapping methods International workshop on Risk Factors for Childhood Leukemia Berlin May

EXAMINERS REPORT & SOLUTIONS STATISTICS 1 (MATH 11400) May-June 2009

Lecture 3 September 1

Fin285a:Computer Simulations and Risk Assessment Section 2.3.2:Hypothesis testing, and Confidence Intervals

First Year Examination Department of Statistics, University of Florida

Effect and shrinkage estimation in meta-analyses of two studies

Estimation, Inference, and Hypothesis Testing

More on nuisance parameters

18.05 Practice Final Exam

Small Area Confidence Bounds on Small Cell Proportions in Survey Populations

Statistics GIDP Ph.D. Qualifying Exam Theory Jan 11, 2016, 9:00am-1:00pm

STAT 135 Lab 11 Tests for Categorical Data (Fisher s Exact test, χ 2 tests for Homogeneity and Independence) and Linear Regression

3. (a) (8 points) There is more than one way to correctly express the null hypothesis in matrix form. One way to state the null hypothesis is

A simulation study comparing properties of heterogeneity measures in meta-analyses

Diagnostics can identify two possible areas of failure of assumptions when fitting linear models.

Physics 403. Segev BenZvi. Parameter Estimation, Correlations, and Error Bars. Department of Physics and Astronomy University of Rochester

META ANALYSIS OF BINARY OUTCOMES DATA IN CLINICAL TRIALS

Introduction to Survey Data Integration

STA 4504/5503 Sample Exam 1 Spring 2011 Categorical Data Analysis. 1. Indicate whether each of the following is true (T) or false (F).

Evidence synthesis for a single randomized controlled trial and observational data in small populations

Categorical Variables and Contingency Tables: Description and Inference

Simple and Multiple Linear Regression

STAT 135 Lab 13 (Review) Linear Regression, Multivariate Random Variables, Prediction, Logistic Regression and the δ-method.

Linear Regression Models

Transcription:

Biostat Methods STAT 5820/6910 Handout #9a: Intro. to Meta-Analysis Methods Meta-analysis describes statistical approach to systematically combine results from multiple studies [identified follong an exhaustive literature review] that have addressed the same research question. Why multiple studies? If question has been clearly settled, may be unethical to conduct more RCTs But sometimes RCTs: run concurrently have inadequate sample size to detect evidence of treatment effect don t get published due to non-significance get lost in the literature address different sub-populations An exhaustive literature review [non-trivial!] can often identify similar studies, and systematically combining their results can [meta-analysis objective]: RCTs have clear protocols, often requiring such a literature review and research synthesis (meta-analysis) to justify a new RCT. Meta-Analysis Methods (presented in this handout) 1. Combining p-values Fisher s method Stouffer s method 2. Combining effect sizes Fixed Effects Random Effects Hierarchical Bayes

Approach 1: simplest & oldest use only p-values Fisher s composite testing method p 1,..., p k from k independent studies th common H 0 ( 2 log p i ) χ 2 2k Fishers null: the null in each study is true Fishers alternative: the null is false in at least one study Fishers known to be highly sensitive to very small (or very large) p-values Another way: Stouffer s method (based on a marginal note in 1949 issue of The American Soldier) Transform p-value p i to a standard normal deviate Z i (assume one-tailed test) Z S = Z i k Z i k Focuses on consensus test of nulls from multiple studies If p-values p 1,..., p k all correspond to true nulls, their distribution (and average) ll be: If some (enough) p-values correspond to (sufficiently) false nulls, their distribution (and average) ll be: Not as sensitive to very small (or very large) p-values

Effect Sizes: focus on magnitude of treatment effect Let θ i be true effect size (a standardized treatment effect) in study i, estimated by ˆθ i. Example: Two-sample mean comparison H 0 : µ 2 = µ 1 Define: θ i = µ 2,i µ 1,i σ i ˆθi = c i Ȳ2,i Ȳ1,i S p,i c i a bias correction factor such that E[ˆθ i ] = (exact form involves Γ function) c i 1 3 4(n 1 + n 2 ) 9 This ˆθ i is often referred to as d (or adjusted Hedges g); not the same as Cohen s d Example: Difference of proportions H 0 : p 1 = p 0 Let p j = P {Y = 1 T rt = j} Need a useful standardized treatment effect Y 0 1 Trt 1 a b Trt 0 c d Consider treatment effect in terms of odds ratio OR = Estimate this OR: ( ) p1 1 p ( 1 ) p0 1 p 0

But what if a or d are 0? Or b or c if we stch to odds of Y = 0? Could add 1/2 to allow for this and reduce bias: ˆθ i = log ( ) (bi + 1/2)(c i + 1/2) (a i + 1/2)(d i + 1/2) Other approaches exist, such as the Peto Method (later) Use ˆθ = log of ÔR (possibly adjusted for zero counts) do odds ratio on log scale so distribution of ˆθ is closer to normal Approach 2: Combine effect sizes Simplest way: Fixed Effects Model (weighted averages) ˆθ = ˆθi V ar[ˆθ] = 1 Choose weights to minimize V ar[ˆθ]: If ˆθ i are iid normal, then θ V ar[ˆθ] N(0, 1) and approximate 95% CI for θ is Example: Two-sample mean comparison ˆθ i = c i Ȳ2,i Ȳ1,i S p,i V ar[ˆθ i ] c 2 i (Derivation of variance involves noncentral t distribution) ( 1 + 1 ˆθ 2 ) i + n 1 n 2 2(n 1 + n 2 ) 3.94 Example: Difference of proportions (odds ratio comparison) ˆθ i = log (ÔR i ) V ar[ˆθ i ] 1 a i + 1/2 + 1 b i + 1/2 + 1 c i + 1/2 + 1 d i + 1/2 (Derivation of variance involves delta method: V ar[g(x)] (f (X)) 2 V ar[x])

Example (Steroid therapy): Look at this fixed effects model: ˆθ i = θ i + ɛ i = θ + ɛ i ɛ i N(0, σi 2 ) This is the homogeneity assumption: w i = 1/ˆσ 2 i = ( V ar[ˆθ i ] ) 1 All studies examined and provided estimates of the same parameter θ, and any differences between estimates are attributable to sample error ɛ alone. Test H 0 : θ 1 = = θ k Q = w i (ˆθi ˆθ ) 2 χ 2 k 1 In practice, this test has low power, so even if it s not significant, maybe can t safely assume homogeneity Instead, allow for slight (& unaccountable) differences among study results random effects model The Random Effects Model: ˆθ i = θ i + ɛ i = θ + δ i + ɛ i ɛ i N(0, σi 2 ) δ i N(0, τ 2 ) δ i is the between-study random effect Test of heterogeneity (above) equivalent to H 0 : Estimate τ 2 and proceed as before: w i = 1/ (ˆσ 2 i + ˆτ 2) = ( V ar[ˆθ i ] ) 1

DerSimonian-Laird approach to estimate τ 2 : the method of moments (uses quantity Q above) Q = w i (ˆθi ˆθ ) 2 χ 2 k 1 E[Q] = τ 2 ( w 2 i ) + (k 1) (get this from expected value of a quadratic form) τ 2 E[Q] (k 1) = w 2 i ˆτ 2 Q (k 1) = max w 2, 0 i A third model class is becoming more common: Hierarchical Bayes Model ˆθ i = θ i + ɛ i = θ + δ i + ɛ i ɛ i N(0, σi 2 ) δ i N(0, τ 2 ) τ π(τ) This model is particularly powerful when also accounting for dependence among study results (R package metahdep) In all three models (Fixed, Random, Hierarchical Bayes), can also account for covariates (fundamental differences between studies), coded as numeric predictor variables X i,l = predictor variable l in study i, l = 1,..., j θ i = β 0 + β 1 X i,1 + β 2 X i,2 +... + β j X i,j