Background: Clone censoring weighting (CCW) is a powerful analytical method that can be used to estimate the expected counterfactual outcome for patients following a specified treatment protocol. However, it is a relatively new technique and there isn't a lot of guidance available for certain practical aspects of using the method to conduct epidemiological studies.
Objectives: Our goal is to provide some guidance for estimating one of the key quantities of CCW studies which is the probability of remaining adherent to a particular treatment protocol. Patients who aren't adherent are said to be artificially censored and adherent patients need to be reweighted to account for such patients. We discuss several types of CCW studies that have been performed in the literature and present techniques for estimating artificial censoring.
Methods: One appealing quality of CCW is that it provides a clean way of avoiding immortal time bias in certain studies. One type of study that CCW can help us with is measuring the effect of adherence to a vaccine regimes. We discuss approaches for estimating adherence to a given regime that includes a sequence of administrations. Other studies include adherence to recurring screenings and treatment initiation based on medical biomarkers.
Results: We highlight one simulation study in which we compared two methods for estimating the probability that patients in a study would remain on protocol by undergoing a treatment within a certain prespecified amount of time from the study start. Any patients that haven't undergone the treatment by the specified time are artificially censored at that time. One technique for estimating the probability of being artificially censored is to use logistic regression for the at-risk patients with the model outcome being whether each patient was artificially censored or not. We present a scenario in which there are time-varying covariates that effect the likelihood of undergoing the treatment and show that estimation using logistic regression had a nearly three times higher mean absolute error than estimation based on a Cox proportional hazards model (MAE of 0.0620 compared to 0.0212), even when including the time-varying covariates in the logistic regression model.
Conclusions: CCW can be a valuable tool for estimating quantities of interest in epidemiological studies. By applying the techniques in this presentation CCW can be effectively used on an increasing number of studies.