Handling missing participant data for dichotomous outcomes in randomized controlled trials: systematic survey of the methods literature

Article type
Authors
Darzi A1, Shawwa K1, el Asmar K1, Kahale L1, Agoritsas T2, Brignardello Petersen R2, W Busse J3, Carrasco A2, Ebrahim S2, C Johnston B4, Neumann I5, Sola I6, Sun X7, Vandvik P8, Coello PA9, Zhang Y8, H Guyatt G2, Akl E1
1American University of Beirut, Lebanon
2McMaster University, Canada
3Chinese Evidence-Based Medicine Center, China
4McMaster University , Canada
5Pontificia Universidad Católica de Chile, Chile
6Iberoamerican Cochrane Centre, Spain
7Centre for Asian Studies, China
8McMaster Univeristy, Canada
9Institute of Biomedical Research (IIB Sant Pau), Spain
Abstract
Background: Missing outcome data for participants in clinical trials could lead to biased effect estimates. Analytical approaches for handling missing participant data (MPD) are important for both estimating the effect estimates (main analysis), and for assessing risk of bias associated with those estimates (sensitivity analyses).
Objectives: To review the literature systematically for simulation studies testing different analytical approaches for dealing with dichotomous MPD in individual randomized clinical trials (RCTs).
Methods: We considered dichotomous outcome data only. We searched MEDLINE, the Cochrane Library, Web of Science, and Journal Storage (JSTOR) up to January 2015. Pairs of reviewers conducted screening and data abstraction independently and in duplicate. We are abstracting information about the identified methods and their performance in the simulation studies (in terms of bias, precision, coverage, accuracy, power, and type I error).
Results: Of 15,205 citations retrieved by the electronic search strategy, 28 studies proved eligible. Twenty-one studies were designed based one or more patterns of missingness: missing completely at random (MCAR: n = 5), missing at random (MAR: n = 16); and not missing at random (NMAR; n = 8). Seven studies did not account for missingness. The data abstracted thus far suggests that multiple imputation, Maximum Likelihood method, or a combination of the two will be the most appropriate methods under the MCAR and MAR assumptions. We will present the full results at the Colloquium.
Conclusion: The results of this study will inform trialists as well as systematic reviewers conducting individual participant data meta-analysis on how to handle MPD for categorical outcomes in both the main and sensitivity analyses.