A systematic survey of the methods literature on the reporting quality and optimal methods of handling participants with missing outcome data for continuous outcomes in randomized controlled trials

Abstract

Objective To conduct (1) a systematic survey of the reporting quality of simulation studies dealing with how to handle missing participant data (MPD) in randomized control trials and (2) summarize the findings of these studies. Study Design and Setting We included simulation studies comparing statistical methods dealing with continuous MPD in randomized controlled trials addressing bias, precision, coverage, accuracy, power, type-I error, and overall ranking. For the reporting of simulation studies, we adapted previously developed criteria for reporting quality and applied them to eligible studies. Results Of 16,446 identified citations, the 60 eligible generally had important limitations in reporting, particularly in reporting simulation procedures. Of the 60 studies, 47 addressed ignorable and 32 addressed nonignorable data. For ignorable missing data, mixed model was most frequently the best on overall ranking (9 times best, 34.6% of times tested) and bias (10, 55.6%). Multiple imputation was also performed well. For nonignorable data, mixed model was most frequently the best on overall ranking (7, 46.7%) and bias (8, 57.1%). Mixed model performance varied on other criteria. Last observation carried forward (LOCF) was very seldom the best performing, and for nonignorable MPD frequently the worst. Conclusion Simulation studies addressing methods to deal with MPD suffered from serious limitations. The mixed model approach was superior to other methods in terms of overall performance and bias. LOCF performed worst. © 2017 Elsevier Inc.