One Design to Rule Them All: The Clone-Censor-Weight Design as a General Framework for Epidemiologic Studies

Background: With modern advances in pharmacoepidemiology have come a plethora of study designs and analytic frameworks aimed at producing valid and interpretable of results. The clone-censor-weight approach, specifically intended to mimic a randomized controlled trial, provides a useful paradigm to unify the theory underlying most pharmacoepidemiologic study frameworks.

Objectives: To compare the results of a simulated study using two distinct study frameworks: a new user design and a clone-censor-weight design.

Methods: A cohort of 10000 individuals was generated. Data for each individual included a continuous covariate, a binary treatment, a censoring time that depended on treatment and the covariate (to create dependent censoring), and two potential event times, one for each level of treatment, that depended on the covariate. For the new user design, a propensity score model was fit with logistic regression to address confounding and, separately for each level of treatment, a censoring model was fit to address dependent censoring using a Cox proportional hazards regression. For the clone-censor-weight design, two copies, or clones, of each individual were created, one for each treatment level. For a given treatment level, individuals receiving the other level of treatment were artificially censored at a randomly determined miniscule time occurring earlier than any of the observed outcome or censoring times. To account for dependent censoring, a censoring model was fit, separately for each level of treatment, using a Cox proportional hazards regression that included an interaction term between the covariate and a time-varying indicator variable taking value 1 if the time was during the miniscule early period. Inverse probability of censoring and (for the new user design only) treatment weighted estimators were used to estimate the counterfactual risk and risk difference functions.

Results: The true risk difference in the simulated cohort at 6 months was 10.6%. The unadjusted 6-month risk difference was 14.2% (95% Confidence Interval [CI]: 11.0%, 17.5%). Using the new user design that accounted for confounding and dependent censoring, the causal risk difference was estimated to be 10.4% (95% CI: 7.0%, 13.8%). Using the clone-censor-weight design, the causal risk difference was estimated to be 10.5% (95% CI: 7.1%, 13.9%).

Conclusions: Using simulation, we have shown that the clone-censor-weight approach can replicate the results of a new user design approach. As the clone-censor-weight approach can additionally be used for studies involving time-varying treatments, dynamic treatment regimes, and other more complicated situations, it may provide a general framework encompassing many of the approaches often used in pharmacoepidemiology.