Bayesian Nonparametric Meta-Analysis Model George Karabatsos University of Illinois-Chicago (UIC)

Similar documents
A Workshop on Bayesian Nonparametric Regression Analysis

Bayesian Nonparametric Rasch Modeling: Methods and Software

Bayesian Methods for Testing Axioms of Measurement

Bayesian nonparametric predictive approaches for causal inference: Regression Discontinuity Methods

Bayesian Nonparametrics: some contributions to construction and properties of prior distributions

Plausible Values for Latent Variables Using Mplus

Ronald Christensen. University of New Mexico. Albuquerque, New Mexico. Wesley Johnson. University of California, Irvine. Irvine, California

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

A marginal sampler for σ-stable Poisson-Kingman mixture models

DNA polymorphisms such as SNP and familial effects (additive genetic, common environment) to

Metropolis-Hastings Algorithm

Estimation of Optimally-Combined-Biomarker Accuracy in the Absence of a Gold-Standard Reference Test

A Fully Nonparametric Modeling Approach to. BNP Binary Regression

Bayesian inference for sample surveys. Roderick Little Module 2: Bayesian models for simple random samples

Contents. Part I: Fundamentals of Bayesian Inference 1

Review: Probabilistic Matrix Factorization. Probabilistic Matrix Factorization (PMF)

Multilevel Statistical Models: 3 rd edition, 2003 Contents

Using R in Undergraduate and Graduate Probability and Mathematical Statistics Courses*

Semiparametric Generalized Linear Models

Flexible Regression Modeling using Bayesian Nonparametric Mixtures

Luke B Smith and Brian J Reich North Carolina State University May 21, 2013

STA 216, GLM, Lecture 16. October 29, 2007

Bayesian Inference for Regression Parameters

Introduction to Multivariate Genetic Analysis. Meike Bartels, Hermine Maes, Elizabeth Prom-Wormley and Michel Nivard

Stat 5101 Lecture Notes

Dirichlet process Bayesian clustering with the R package PReMiuM

Measurement Error and Linear Regression of Astronomical Data. Brandon Kelly Penn State Summer School in Astrostatistics, June 2007

STA 4273H: Statistical Machine Learning

DEPARTMENT OF COMPUTER SCIENCE Autumn Semester MACHINE LEARNING AND ADAPTIVE INTELLIGENCE

Mixture modelling of recurrent event times with long-term survivors: Analysis of Hutterite birth intervals. John W. Mac McDonald & Alessandro Rosina

Lecture 2: Linear Models. Bruce Walsh lecture notes Seattle SISG -Mixed Model Course version 23 June 2011

Gaussian Processes. Le Song. Machine Learning II: Advanced Topics CSE 8803ML, Spring 2012

The STS Surgeon Composite Technical Appendix

Low-Level Analysis of High- Density Oligonucleotide Microarray Data

Fractional Imputation in Survey Sampling: A Comparative Review

Bayesian non-parametric model to longitudinally predict churn

Longitudinal Data Analysis Using Stata Paul D. Allison, Ph.D. Upcoming Seminar: May 18-19, 2017, Chicago, Illinois

Bayesian Modeling of Conditional Distributions

Gentle Introduction to Infinite Gaussian Mixture Modeling

Truncation error of a superposed gamma process in a decreasing order representation

A Bayesian Nonparametric Model for Predicting Disease Status Using Longitudinal Profiles

On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models

On the Behavior of Marginal and Conditional Akaike Information Criteria in Linear Mixed Models

General Regression Model

3 Joint Distributions 71

Accounting for Complex Sample Designs via Mixture Models

Bayesian nonparametric estimation of finite population quantities in absence of design information on nonsampled units

Part 8: GLMs and Hierarchical LMs and GLMs

... x. Variance NORMAL DISTRIBUTIONS OF PHENOTYPES. Mice. Fruit Flies CHARACTERIZING A NORMAL DISTRIBUTION MEAN VARIANCE

Statistics & Data Sciences: First Year Prelim Exam May 2018

Approximating high-dimensional posteriors with nuisance parameters via integrated rotated Gaussian approximation (IRGA)

BAYESIAN METHODS FOR VARIABLE SELECTION WITH APPLICATIONS TO HIGH-DIMENSIONAL DATA

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

(Where does Ch. 7 on comparing 2 means or 2 proportions fit into this?)

Nonparametric Bayes tensor factorizations for big data

Chapter 4 Multi-factor Treatment Designs with Multiple Error Terms 93

Lecture 16: Mixtures of Generalized Linear Models

Basic concepts in estimation

Part 6: Multivariate Normal and Linear Models

Estimation Tasks. Short Course on Image Quality. Matthew A. Kupinski. Introduction

On some distributional properties of Gibbs-type priors

Resemblance among relatives

Non-Parametric Bayes

Bayesian construction of perceptrons to predict phenotypes from 584K SNP data.

A Bayesian multi-dimensional couple-based latent risk model for infertility

Density Estimation. Seungjin Choi

Nonparametric Bayes Uncertainty Quantification

Bayesian Nonparametrics

A Complete Spatial Downscaler

Discussion of On simulation and properties of the stable law by L. Devroye and L. James

Spatial Bayesian Nonparametrics for Natural Image Segmentation

Bayes methods for categorical data. April 25, 2017

Prerequisite: STATS 7 or STATS 8 or AP90 or (STATS 120A and STATS 120B and STATS 120C). AP90 with a minimum score of 3

MAD-Bayes: MAP-based Asymptotic Derivations from Bayes

Truncation error of a superposed gamma process in a decreasing order representation

STAT Advanced Bayesian Inference

Biost 518 Applied Biostatistics II. Purpose of Statistics. First Stage of Scientific Investigation. Further Stages of Scientific Investigation

Asymptotic Multivariate Kriging Using Estimated Parameters with Bayesian Prediction Methods for Non-linear Predictands

Bayesian nonparametric models for bipartite graphs

PART I INTRODUCTION The meaning of probability Basic definitions for frequentist statistics and Bayesian inference Bayesian inference Combinatorics

The concept of breeding value. Gene251/351 Lecture 5

Bayesian semiparametric inference for the accelerated failure time model using hierarchical mixture modeling with N-IG priors

Wavelet-Based Nonparametric Modeling of Hierarchical Functions in Colon Carcinogenesis

Probe-Level Analysis of Affymetrix GeneChip Microarray Data

Meta-analysis of case-control studies

Slice sampling σ stable Poisson Kingman mixture models

Slice Sampling Mixture Models

Passing-Bablok Regression for Method Comparison

Advanced Statistical Modelling

Recent advances in statistical methods for DNA-based prediction of complex traits

Bayesian Mixture Modeling of Significant P Values: A Meta-Analytic Method to Estimate the Degree of Contamination from H 0 : Supplemental Material

Bayesian Nonparametric Inference Methods for Mean Residual Life Functions

On the Truncation Error of a Superposed Gamma Process

Latent Variable Models Probabilistic Models in the Study of Language Day 4

Department of Statistical Science FIRST YEAR EXAM - SPRING 2017

Review of Statistics 101

COPYRIGHTED MATERIAL CONTENTS. Preface Preface to the First Edition

An Alternative Prior Process for Nonparametric Bayesian Clustering

Basic Statistical Analysis

Multiple Linear Regression for the Supervisor Data

Transcription:

Bayesian Nonparametric Meta-Analysis Model George Karabatsos University of Illinois-Chicago (UIC) Collaborators: Elizabeth Talbott, UIC. Stephen Walker, UT-Austin. August 9, 5, 4:5-4:45pm JSM 5 Meeting, Seattle Session on Multivariate Meta-Analysis 4: PM - 5:5 PM, CC-3 Organizer: Simina M. Boca, Georgetown University Chair: Valerie Langberg, Brown University Supported by NSF Research Grant SES-5637.

I. Aims of meta analysis Outline A. Meta analysis data framework. Examples of effect sizes. B. Aims of meta analysis. (Main aim: infer overall effect size from a universe of studies) C. Publication bias assessment. D. Conventional normal models for meta-analysis. E. Potential issues: Normality assumptions about errors (and random effects), when not supported by the effect size data, can cause misleading meta-analytic conclusions. II. Proposed solution: A Bayesian nonparametric (BNP) meta-analysis model, which allows the entire effect-size distribution (density) to vary flexibly as a function of covariates. A covariate-dependent infinite mixture model. III. Illustration of the BNP model on real meta-analytic data involving behavioral genetics research of antisocial behavior. IV. Conclusions. Free menu-based software for BNP analysis.

Meta Analysis Data Framework Data: D n = {y i,, x i } i=:n i n' ( n) studies provide data on n study reports y i of a common effect size variable (Y). Each effect size report y i has sample size n i, sampling variance, i and covariates x i = (, x i,, x ip ) describing study characteristics. 3

Meta Analysis Data Framework Examples of effect sizes: Effect-size Description Effect-size (y i ) Variance ( i ) Unbiased standardized mean difference, two independent groups (Hedges, 98). i i n i i n i i n i n i c n i n i n i n i y i n i n i c ; c 3 4 n i n i Fisher z transformation of the correlation i. log i i n i 3 Log odds ratio for two binary (-) variables. log n i/n i n i /n i n i n i n i n i More examples: E.g., see Cooper, Hedges, Valentine (9). 4

Aims of Meta Analysis Given a set of meta analytic data, D n = {y i,, x i } i=:n, infer the overall effect size, after accounting for the covariates x i and the observation weights /. Basic (regression) Parameters: = (,,, p ) T. Mean overall effect size: i i Publication bias (e.g., file drawer effect) may affect meta-analytic conclusions. Publication bias may be assessed from the data by / including ˆ as one of the p covariates in x. This provides regression analysis for the funnel plot (Egger et al., 997; Thompson & Sharp, 999). Significant regression slope coefficient for the covariate suggests a presence of publication bias. 5 ˆ /

Conventional Meta Analytic Models General f y i x i, i ; n yi x i i t i, i, i,,n; Normal x Model: i x i p x ip ;,, n n n, I n M n ; t n,, t,,t. Meta-analysis model: (,, p ) =. Meta-regression-analysis model: (,, p ) can be non-zero. Fixed effects model: = = =. -level random effects model: = =. 3-level random effects model: =. Model with correl. between study reports (Stevens-Ta.9). Each model can be fit by restricted ML (Harville,977) or by Bayesian MCMC methods (Spiegelhalter et.al. 9). 6

Conventional Meta Analytic Models General f y i x i, i ; n yi x i i t i, i, i,,n; Normal x Model: i x i p x ip ;,, n n n, I n M n ; t n,, t,,t. However, the normality assumptions (above) may not be exactly true for real meta-analytic data. This may cast doubt on the adequacy of the meta-analysis. Also, the normal model focuses inferences on how the mean effect size changes as a function of x, instead of how additional features of the effect size distribution (e.g., variance, quantiles, entire distribution/density, etc.) changes as a function of x. 7

Proposed BNP Meta-Analytic Model f y i x i, i ; n y i x i, i dgx j n y i j x i, i j x i,, i,,n, j x i, j x i / j x i / j n,, j,,, n,v k, k n, v k v k.5 k.5 k, k,,p ga a /,a / un,b, n p, 5 I p ga,. 8

Density f(y x) BNP Meta-Analytic Model (behavior) = / = / = = Mixture weight j.5.5.5.5 - Index j - Index j - Index j - Index j........ - y - y - y - y 9

Proposed BNP Meta-Analytic Model f y i x i, i ; n y i x i, i dgx n y i j x i, i j x i,, i,,n, j j x i, j x i / j x i / j n,, j,,, n,v k, k n, v k v k.5 k.5 k, k,,p ga a /,a / un,b, n p, 5 I p ga,. The posterior distribution of the model can be estimated by standard MCMC Gibbs sampling methods. Slice sampler for. See Karabatsos & Walker (, Appendix, Elec. J Stat.).

MZ-DZ Twin Comparison Sample Probability Density BNP Meta-Analysis Data Illustration 7 6 5 4 3 Heritability (Effect Size) Data 6 3 3 9 9 4 4 4 9 9 9 8 8 7 7 6 9 7 4 9 8 3 7 7 8 6 6 5 5 4 3 3 3 7 8 89 5 5 4.5 Heritability and Variance (+) 8 6 3 5 3 5 9 Heritability Distribution Over Studies 3.5 3.5.5.5 Mean=.5 Med=.5 Var= Skew=-. Kurt=8.3.5 Heritability Falconer (& Mackay 996) (antisocial) heritability: Sampling Variance: 4 MZ /n MZ DZ /n DZ h MZ DZ

BNP Meta-Analysis Data Illustration 4 covariates in data: Publication year; Square root heritability variance SE(ES) to assess for publication bias; Indicators (-) of female status versus male; Ten indicators of antisocial behavior ratings done by mother, father, teacher, self, independent observer, and ratings done on conduct disorder, aggression, delinquency, and externalizing antisocial behavior; Indicator of whether weighted aver of heritability measures was taken within study over different groups of raters who rated the same twins; Mean of the study subjects in months; Indicators of hi-majority ( 6%) white twins in study, zygosity obtained by questionnaire or through DNA samples, study inclusion of low socioeconomic (SES) status subjects versus mid-to-high SES subjects, missing SES information, representative sample, longitudinal sample; Latitude and longitude of study.

BNP Meta-Analysis Data Illustration Model D m Model D m BNP-ss. 6 DL-x 5. 5 DL- 4. 8 3L-x 5. 5 L-, by MZ-DZ 4. 8 L-, by Study 5. 8 3L- 4. 8 FE- 5. 9 DL-ss 5. 4 L-ss, by Study 6. L-ss, by MZ-DZ 5. 4 FE-ss 6. 3L-ss 5. 4 L-x, by Study 6. L-x, by MZ-DZ 5. 5 FE-x 6. Model comparisons: D(m) is posterior predictive mean-square error criterion (Laud & Ibrahim, 995). 3

Heritability Heritability BNP Meta-Analysis Data Illustration.8.8 Mom.8 Dad.6.6.6.4.4.4.....4.6.8 SE(ES) 5 5 5 5.8 Teacher.8 Self.8 Observer.6.6.6.4.4.4... 5 5 5 5 5 5 Posterior predictive median and IQR of heritability, as a function of rater type and child. Overall mean heritability ES estimate (of β₀) was.5. 4

Heritability Heritability BNP Meta-Analysis Data Illustration.8.8 Mom.8 Dad.6.6.6.4.4.4.....4.6.8 SE(ES) 5 5 5 5.8 Teacher.8 Self.8 Observer.6.6.6.4.4.4... 5 5 5 5 5 5 The SE(ES) covariate was not significant according to the spike-and-slab prior indicators. Thus, no significant publication bias in data. 5

Conclusions We proposed a useful and flexible BNP model for meta-analysis. Illustrated the model on real data. Better predictive utility vs. normal meta-analytic models. This leads to more reliable inferences. Provides richer meta-analysis compared to normal models. E.g., provides quantile regression of effect size. Article on BNP meta-analytic model: Karabatsos, G., Talbott, E., & Walker, S.G. (4). A Bayesian nonparametric meta-analysis model. Research Synthesis Methods, 6(), 8-44. 6

Conclusions Free User-friendly menu-driven software for BNP model (user point and click): http://www.uic.edu/~georgek/homep/bayessoftware.html Paper/software user s manual: http://arxiv.org/abs/56.5435 Software currently offers 83 Bayesian mixture models, including normal mixture models, several versions of the BNP infinite-mixture model of this talk, as well as other BNP infinite-mixture models, with mixture distribution assigned a prior defined by either the Dirichlet process (Ferguson 973); Pitman-Yor (997) process, the normalized stable process (Kingman, 975); geometric weights process (Fuentes-Garcia, Mena, Walker ); or the normalized inverse-gaussian process (Lijoi et al. 5). 7

8