Article type
Year
Abstract
Background: Systematic reviewers including all randomized participants in their meta-analyses need to make assumptions about the outcomes of those with missing data.
Objectives: To provide systematic review authors with guidance on dealing with participants with missing data for dichotomous outcomes.
Methods: The authors conducted a systematic survey of the methodological literature regarding ‘intention to treat’ analysis and used an iterative process of suggesting guidance and obtaining feedback to arrive at a proposed approach.
Results: We consider here participants excluded from the trial analysis for ‘non-adherence’ but for whom data are available, and participants with missing data (Fig. 1). Non-adherent participants excluded from the trial analysis but for whom data are available should in most instances be included in the meta-analysis, and in the arm to which they were randomized. For participants with missing data, systematic reviewers can use a range of plausible assumptions in the intervention and control arms (Fig. 2). Extreme assumptions include ‘all’ or ‘none’ of the participants had an event, but these assumptions are not plausible. Less extreme assumptions may draw on the incidence rates within the trial (e.g., same incidence in the trial control arm) or in all trials included in the meta-analysis (e.g., highest incidence among control arms of all included trials). The primary meta-analysis may use either a complete case analysis or a plausible assumption. Sensitivity meta-analyses to test the robustness of the primary meta-analysis results should include extreme plausible assumptions (Fig. 2). When the meta-analysis results are robust to extreme plausible assumptions, inferences are strengthened. Vulnerability to extreme plausible assumptions suggests rating down confidence in estimates of effect for risk of bias.
Conclusions: This guide proposes an approach to establishing confidence in estimates of effect when systematic reviewers are faced with missing participant data in randomized trials.
Objectives: To provide systematic review authors with guidance on dealing with participants with missing data for dichotomous outcomes.
Methods: The authors conducted a systematic survey of the methodological literature regarding ‘intention to treat’ analysis and used an iterative process of suggesting guidance and obtaining feedback to arrive at a proposed approach.
Results: We consider here participants excluded from the trial analysis for ‘non-adherence’ but for whom data are available, and participants with missing data (Fig. 1). Non-adherent participants excluded from the trial analysis but for whom data are available should in most instances be included in the meta-analysis, and in the arm to which they were randomized. For participants with missing data, systematic reviewers can use a range of plausible assumptions in the intervention and control arms (Fig. 2). Extreme assumptions include ‘all’ or ‘none’ of the participants had an event, but these assumptions are not plausible. Less extreme assumptions may draw on the incidence rates within the trial (e.g., same incidence in the trial control arm) or in all trials included in the meta-analysis (e.g., highest incidence among control arms of all included trials). The primary meta-analysis may use either a complete case analysis or a plausible assumption. Sensitivity meta-analyses to test the robustness of the primary meta-analysis results should include extreme plausible assumptions (Fig. 2). When the meta-analysis results are robust to extreme plausible assumptions, inferences are strengthened. Vulnerability to extreme plausible assumptions suggests rating down confidence in estimates of effect for risk of bias.
Conclusions: This guide proposes an approach to establishing confidence in estimates of effect when systematic reviewers are faced with missing participant data in randomized trials.
Images
