A comparison of multiple imputation methods for incomplete longitudinal binary data

Longitudinal binary data are commonly encountered in clinical trials. Multiple imputation is an approach for getting a valid estimation of treatment effects under an assumption of missing at random mechanism. Although there are a variety of multiple imputation methods for the longitudinal binary data, a limited number of researches have reported on relative performances of the methods. Moreover, when focusing on the treatment effect throughout a period that has often been used in clinical evaluations of specific disease areas, no definite investigations comparing the methods have been available. We conducted an extensive simulation study to examine comparative performances of six multiple imputation methods available in the SAS MI procedure for longitudinal binary data, where two endpoints of responder rates at a specified time point and throughout a period were assessed. The simulation study suggested that results from naive approaches of a single imputation with non-responders and a complete case analysis could be very sensitive against missing data. The multiple imputation methods using a monotone method and a full conditional specification with a logistic regression imputation model were recommended for obtaining unbiased and robust estimations of the treatment effect. The methods were illustrated with data from a mental health research.