A comparison of multiple-imputation methods for handling missing data in repeated measurements observational studies

Multiple-imputation (MI) methods for imputing missing data in observational health studies with repeated measurements were evaluated with particular focus on incomplete time varying explanatory variables. Standard and random-effects imputation by chained equations, multivariate normal imputation and Bayesian MI were compared regarding bias and efficiency of regression coefficient estimates by using simulation studies. Flexibility of the methods in handling different types of variables (binary, categorical, skewed and normally distributed) and correlations between the repeated measurements of the incomplete variables were also compared. Multivariate normal imputation produced the least bias in most situations, is theoretically well justified and allows flexible correlation for the repeated measurements. It can be recommended for imputing continuous variables. Bayesian MI is efficient and may be preferable in the presence of categorical and non-normally distributed continuous variables. Imputation by chained equations approaches were sensitive to the correlation between the repeated measurements. The moving time window approach may be used for normally distributed continuous variables with auto-regressive correlation.