Statistics in Practice

Presented Tuesday, 12 July

Time Room Statistics in Practice Session 1
11:00 – 12:30 Salon A – Carson Hall Session 1 (Part I, II)
Time Room Statistics in Practice Session 2
14:00 – 15:30 Salon A – Carson Hall Session 2 (Part III, IV)

Meta-analysis of Individual Participant Data

Simon G. Thompson, Professor of Biostatistics
Department of Public Health and Primary Care, University of Cambridge, UK
Mark Simmonds, Research Fellow
Centre for Reviews and Dissemination, University of York, UK

Presentation Slides available for download.

Meta-analyses of multiple studies, for which individual participant data (IPD) are available, are becoming common. The aim of these sessions is to update participants on statistical methods that can be used for such analyses, and the pitfalls to be avoided. Application to both trials and observational studies will be addressed, together with examples and reference to available software. Familiarity with the concept of meta-analysis is assumed. The two sessions are organised as four 30-minute presentations, each then allowing 10 minutes for discussion and questions. Specifically, the presentations will cover the following topics:

  1. IPD meta-analysis of clinical trials

We first introduce IPD meta-analysis, and analysis methods in the context of two-arm randomized trials:

  • Definition of IPD
  • Advantages and practicalities of IPD meta-analysis
  • Analysis methods: naïve lumping, two-stage, one-stage
  • Two-stage meta-analysis: binary and continuous outcomes
  • Fixed and random treatment effects


  1. IPD meta-analysis of observational studies

We here discuss the use of covariates, in the context of survival data in observational epidemiological studies:

  • Summarising data in each study in a consistent way
  • Adjusting for covariates
  • Analysing interactions
  • Separating within- and between-study information
  • Handling confounders that are completely missing in some studies


  1. Advanced IPD meta-analysis methods for clinical trials

Here we explore some more challenging topics in the context of IPD meta-analysis of randomized trials:

  • One-stage analysis: a general hierarchical model structure
  • Covariates and interactions: one-stage vs. two-stage
  • Survival data in trials
  • Combining IPD with aggregate data from non-IPD studies
  • Addressing missing outcome data


  1. Advanced IPD meta-analysis methods for observational studies

The usual measures of association derived from epidemiological studies, such as hazard ratios, do not have a direct interpretation in terms of the implications for public health.  Here we discuss how such measures can be derived from IPD meta-analysis:

  • Adjusting for measurement error and within-person variability
  • Assessing the usefulness of novel markers for medical screening
  • Deriving measures of public health impact
  • Estimating life expectancy
  • Estimating causal relationships using Mendelian randomization