Informing sample size and power calculations for clone-censor-weight analyses: a Monte Carlo simulation study of the variance inflation factor relative to the underlying target trial.

Abstract

Background: Clone-censor-weighting (CCW) is a causal inference method used in target trial emulation in settings in which the time of application of the treatment occurs during follow-up. An important component of study design is determining the sample size required for a specified level of statistical power or to determine the statistical power associated with a specific sample size. Determining statistical power is relatively straightforward for the underlying randomized controlled trial (RCT) that one is emulating. In contrast, conducting power and sample size calculations for observational studies that employ CCW is not straightforward because the analysis incorporates inverse probability of censoring weights, which are only known at the analysis phase, and not during the design phase.

Methods: Monte Carlo simulations (270 scenarios varying sample size, treatment prevalence, threshold for defining early treatment, and covariate-treatment odds ratios) were used to estimate the variance inflation factor (VIF) of a CCW analysis compared to the underlying target RCT, across two estimands: the risk difference and the relative risk.

Results: In settings with 10 baseline covariates, the use of a CCW design tended to result in more precise estimation of risk differences and relative risks compared to the underlying RCT when the magnitude of confounding was low to moderate. The magnitude of the VIF depended on treatment prevalence, sample size, the time horizon at which risk was estimated, and the strength of the baseline covariates' association with treatment initiation. In secondary simulations, it was observed that the VIF comparing a CCW design to the underlying RCT tended to increase as the number of baseline covariates increased.

Conclusions: Researchers designing CCW studies can use the simulation results and the accompanying R code, available in a public GitHub repository, to obtain study-specific VIF estimates for incorporation into a priori power calculations.