Abstract
Background: Complete case analysis (CCA) is the most common method used to handle missing outcome data in trials but may often lead to biased estimates. Newer methods that address missing not at random data are infrequently used.
Objective: We evaluated six missing data methods in a simulation study based on a melanoma surveillance trial.
Methods: We used the MEL-SELF pilot of patient-led surveillance for patients with early-stage melanoma as the empirical basis of our simulated study. We evaluated three commonly used methods (CCA, mixed models for repeated measurements (MMRM), multiple imputation (MI)), and three recently developed methods (retrieved dropout (RD) imputation, jump to reference (J2R) imputation and trimmed means (TM)), under 48 scenarios where treatment effect size, missingness percentage and missingness assumptions were varied.
Results: Under scenarios with small or small-moderate treatment effects and missing not at random outcome data, all methods produced some bias, with TM and CCA the most biased towards and away from the null, respectively. Both also had low precision and power. J2R performed best of methods that were biased towards the null (JR, TM), with small bias for small and small-moderate treatment effects, high precision and high coverage. RD performed best of methods that were biased away from the null (RD, CCA, MMRM, MI) with small bias, good precision and good coverage.
Conclusion: In this simulation of a melanoma surveillance trial with non-random missing outcomes, RD produced the least bias away from the null and J2R produced the least bias towards the null.