Article type
Year
Abstract
Background: Missing participant data (MPD) relates to trial participants for whom outcome data are not available for systematic review (SR) authors. A number of methods to assess the impact of MPD on the results of meta-analyses have been proposed. No study has compared the use of these different methods.
Objectives: The aim of this study is to compare the impact on the pooled effect estimates by different methods of accounting for MPD when conducting a meta-analysis.
Methods: We included one meta-analysis from each of 100 clinical interventional SRs published in 2012. Eligible SRs reported a group-level meta-analysis of a patient-important dichotomous efficacy outcome, with a statistically significant pooled effect estimate. Ten reviewers, working in pairs, independently extracted data from all included meta-analyses and from all trials that contributed data to those meta-analyses. We considered 19 categories of participants that could potentially have MPD (Table 1). We included in our analyses participants belonging to any of these categories and explicitly reported on by the trialists as not followed-up (i.e. missing data). We reran each meta-analysis by applying nine different assumptions about the outcomes of participants with missing data using the same statistical methods used by the SR authors (Table 2). We calculated for each assumption the percentages of meta-analyses that 1) lost statistical significance or 2) changed direction. We also calculated for each assumption the mean change in the pooled effect estimates across the 100 meta-analyses.
Results: We included 50 eligible Cochrane SRs and 50 non-Cochrane SRs, and a total of 653 trials. We have collected all data and we are in the process of analysing it. We will present the findings of the study at the Colloquium.
Conclusions: Our findings will inform recommendations regarding what assumptions systematic review authors should make when considering the extent to which MPD impacts risk of bias.
Objectives: The aim of this study is to compare the impact on the pooled effect estimates by different methods of accounting for MPD when conducting a meta-analysis.
Methods: We included one meta-analysis from each of 100 clinical interventional SRs published in 2012. Eligible SRs reported a group-level meta-analysis of a patient-important dichotomous efficacy outcome, with a statistically significant pooled effect estimate. Ten reviewers, working in pairs, independently extracted data from all included meta-analyses and from all trials that contributed data to those meta-analyses. We considered 19 categories of participants that could potentially have MPD (Table 1). We included in our analyses participants belonging to any of these categories and explicitly reported on by the trialists as not followed-up (i.e. missing data). We reran each meta-analysis by applying nine different assumptions about the outcomes of participants with missing data using the same statistical methods used by the SR authors (Table 2). We calculated for each assumption the percentages of meta-analyses that 1) lost statistical significance or 2) changed direction. We also calculated for each assumption the mean change in the pooled effect estimates across the 100 meta-analyses.
Results: We included 50 eligible Cochrane SRs and 50 non-Cochrane SRs, and a total of 653 trials. We have collected all data and we are in the process of analysing it. We will present the findings of the study at the Colloquium.
Conclusions: Our findings will inform recommendations regarding what assumptions systematic review authors should make when considering the extent to which MPD impacts risk of bias.