Manuel Gomes FDA Presentation May 20, 2014

Handling missing data in meta-analysis of individual participant data with mixed outcomes

M Gomes, LA Hatfield, S-LT Normand


Meta-analysis of individual participant data (IPD) is increasingly utilized in combining data from both randomized trials and observational studies. IPD meta-analysis has several advantages over meta-analysis based on aggregate study-level data, such as standardization of analysis across studies, analysis of multiple outcomes, and adjustment for additional potential confounders. IPD meta-analysis hosts, however, particular features making inferences challenging; the hierarchical structure of individuals within studies, the multiplicity of outcomes – often consisting of a mixture of continuous and discrete variables, and the missing data. Missing data typically arise because outcomes may be partially observed across studies (sporadically missing), or may not be collected in a specific study (systematically missing). We illustrate a multivariate Bayesian approach that accounts for the correlation among the (mixed) outcomes, between-study heterogeneity, and missing data.

Multiple imputation (MI) offers an alternative approach for addressing missing data as this approach also recognizes the hierarchical nature of the data and the correlation among the multiple outcomes. An important advantage of MI involves its relative ease of implementation through the availability of standard software. In this talk, we illustrate approaches in a meta-analysis of randomized controlled trials comparing the effectiveness of implantable cardiac devices (ICD vs CRT-D) to treat heart failure, with both sporadically and systematically missing mixed outcomes (mortality, 6-minute walk, NYHA class and quality of life). A simulation study characterizes the relative performance of alternative methods across different settings in IPD meta-analysis.
This work is supported by the MDEpiNet Methodology Center (Drs. Hatfield and Normand) and by funding from UK MRC Early Career Fellowship (Dr. Gomes).

Learning objectives

  1. Understanding the specific challenges posed by missing data in meta-analysis of individual-participant data.
  2. Illustrating different methods for meta-analysis of mixed outcomes with missing data.
  3. Assessing the relative merits of alternative methods for addressing the missing data in IPD meta-analysis, across a wide range of circumstances.