The impact of differential censoring and covariate relationships on propensity score performance in a time-to-event setting: a simulation study

The gold-standard for establishing causation in biomedical and epidemiologic studies is the randomized-controlled trial, in which all subjects are randomly selected to be in either the experimental arm (receiving the treatment under evaluation) or the control arm (receiving an alternative therapy). This methodology is ideal due to the fact that all subjects have an equal probability of being in each arm, regardless of their underlying personal factors. In this setting, if a sufficient number of subjects are enrolled in a study, these underlying factors should be balanced in each study group and thus should not unduly bias the results of the study.

In practice, however, there are many scenarios in which conducting a randomized-controlled trial is difficult. For this reason, statistical ways of mimicking this randomization in observational studies, where the investigator has no control over which subjects receive the treatment in question, have become increasingly important. One method involves estimating the probability that a subject would receive the treatment, termed their propensity score. These scores may then be incorporated into the analysis of the study results in various ways in an attempt to balance the underlying subject-specific factors across the treated and untreated study groups.

This simulation study aims to evaluate the extent to which propensity score methods accurately estimate the effect of a treatment while maintaining balance across underlying factors in the treated and non-treated groups in a setting where the outcome of interest is the amount of time lapsed prior to a particular event.