Handling drop-outs which imply a change of treatment: an evaluation of reference-based imputation

Excerpt

In some randomised trials, loss to follow-up is associated with end of treatment. This applies particularly, but not exclusively, to pharmaceutical trials. In this setting Carpenter et al (Journal of Biopharmaceutical Statistics 2013; 23: 1352-71) distinguished the “de facto” estimand, the mean of the observed and unobserved outcomes in the actual treatment circumstances of the trial, from the “de jure” estimand, the mean of the observed and unobserved outcomes if treatment were maintained after loss to follow-up. Standard missing at random analysis estimates the de jure estimand. Carpenter et al (2013) proposed “reference-based imputation” methods to estimate the de facto estimand. These involve various ways to use data from a reference arm (usually the control arm) to multiply impute post-dropout outcomes. However, the properties of these methods have been little explored. …