Methods for handling missing participant data in systematic reviews of dichotomous outcome data: a case example of plausible and extreme plausible assumptions

Article type
Authors
Johnston B1, Akl E2, Goldenberg J3, Ma S4, Ebrahim S5, Guyatt G5
1SickKids Research Institute & Department of Anaesthesia & Pain Medicine, The Hospital for Sick Children, Toronto, Canada
2Department of Medicine, State University of New York at Buffalo, NY, USA
3Bastyr University, Seattle, WA, USA
4Division of Plastic and Reconstructive Surgery, McMaster University, Hamilton, Ontario, Canada
5Department of Clinical Epidemiology and Biostatistics, McMaster University, Hamilton, Ontario, Canada
Abstract
Background: Missing participant data is common in systematic reviews of randomized trials. To include all randomized participants, reviewers typically must make assumptions about the outcomes of participants with missing data. Current methods (e.g. all patients had event, no patients had the event) are limited in that their assumptions are often implausible.

Objectives: To apply a recently developed approach to missing data to a meta-analysis assessing the effects of probiotics for the prevention of clostridium difficile-associated diarrhea (CDAD).

Methods: Our conceptual approach begins with a complete case analysis and then applies a series of increasingly stringent assumptions regarding events in those with missing data. We assumed that the event rate in those in the control group who had missing data was the same as the event rate for those successfully followed. For the probiotic group we calculated effects using assumed ratios of event rate in those with missing data in comparison to those successfully followed: 1.5:1, 2:1, 3:1, and 5:1.

Results: A complete case analysis suggested a very large reduction in the incidence of CDAD with a narrow 95% confidence interval (RR 0.34; CI 0.24–0.49). Of 20 eligible trials, 13 had missing CDAD data (5–45% of patients). Imputing ratios for the missing responses of 1.5 and 2.0 to 1 resulted in little change to the magnitude of effect (RR 0.36; CI 0.26–0.50; RR 0.38; CI 0.27–0.53). Larger ratios, 3.0 and 5.0 to 1, diminished the magnitude of effect, but still suggested a large effect with precise confidence intervals (RR 0.43; CI 0.30–0.62; RR 0.50; CI 0.34–0.76).

Conclusions: Application of our approach demonstrated that results were robust to ratio-based plausible and extreme plausible assumptions, allowing us to dismiss missing participant data as a significant threat to validity.