Partial imputation approach to analysis of repeated measurements with dependent drop-outs

In clinical trials repeated measurements of a response variable are usually taken at prespecified time-points to compare the treatment effects. However, the comparison of treatment effects is often complicated by missing data caused by the withdrawal of some patients before the end of the study (that is, drop-outs). When the drop-out process depends on the response variable of interest, ignoring missing data may lead to biased comparison of the treatment effect. In this paper, conditions for ignoring the dependent missingness are investigated and a new approach using the usual testing procedure based on data with partial carrying-forward imputation is proposed. The proposed approach is conceptually and practically simple, and is motivated by making incremental improvement on the familiar 'all available data' (AAD) approach and the 'last value carrying forward' (LVCF) approach, which are commonly used in data analysis with drop-outs by practitioners. It is also compared favourably to the mixed-effect model approach with dependent drop-outs. Simulations and real data are used to evaluate and illustrate statistical properties of the proposed approach. The principle of the proposed approach can also be extended to using other imputation methods such as the multiple imputation.