Missing data in meta-analysis - a practical guide

Tags: Workshop
Carpenter J

Abstract: Missing data are ubiquitous in individual trials,

and hence in meta-analysis. Yet the key issues and their implications are ofton obscured by the technical statistical nature of many research articles.

This workshop aims to give an accessible presentation of the issues and methodology available.

The objectives are:

a) to introduce participants to the key issues raised by missing data;

b) to review established and recent methods for including studies with missing data in contributing trials;

c) to work through examples illustrating these issues in small groups.

More details:

Missing data are ubiquitous in clinical trials. They introduce a level of ambiguity into the analysis and conclusions beyond conventional sampling imprecision. This is because additional assumptions, for example about the reason for the missing observations, must be made to justify any analysis and hence the conclusions. For a broad introduction, see www.missingdata.org.uk

We focus on how to handle the issues caused by missing data in studies contributing to a meta-analysis.

Building on the approach described by Carpenter and Kenward (2007), we argue that a principled approach is preferable to an ad-hoc approach, and describe how such an approach might work in practice. We show how different methods are needed, in general, to address the 'Intention To Treat' (ITT) and 'Per-Protocol' hypotheses.

Suppose, we are faced with a trial report where missing data is an issue. The initial focus is on understanding whether the methods adopted by the trialists are appropriate for an ITT analysis. If they are not, we consider the likely effects on the presented results and conclusions. The next issue is how to formally include studies with missing data in a meta-analysis. We discuss and critique two recently proposed methods for doing this.