Methods/Statistics
To Clone or not to Clone into Incompatible Regimens in Non-experimental Settings Michael Webster-Clark* Michael Webster-Clark Alexander Breskin Robert Platt
Cloning, censoring, and weighting (CCW) is a great tool to contrast regimens that cannot be distinguished at baseline. The origins of CCW involved emulating target trials contrasting regimens like “never treat with X” (Regimen 0) with multiple protocols like “with X for time ” (Regimens ) with people randomized to X=1 or X=0 at baseline. As t0 treatment was random, there is no need to clone patients randomized to X=0 to Regimens or clone patients randomized to X=1 to Regimen 0. Observational studies, however, may be confounded by non-random t0 treatment assignment. They have two options: A) clone all patients into all regimens, censor incompatible patients at time 0, and create censoring weights for them; and B) do not clone into incompatible regimens, create CCW in each regimen, and account for confounding between regimens as in nonexperimental studies (e.g., inverse probability of treatment weights). While both can generate unbiased estimates, approach 1 ensures consistent target populations across comparisons. To demonstrate, we simulated using CCW to compare risk of an outcome Y by time 3 under three regimens: “start X at t0 or t1,” “start X at t0,” and “never start X.” The probability of starting X at each time was associated with a variable W associated with Y. We calculated risk differences (RDs) between regimens using A and B in the presence and absence of risk difference-scale heterogeneity and compared them to true RDs (Figure 1). RDs using A estimated the full study population RDs. RDs using B, however, only estimated full population RDs absent treatment effect heterogeneity or if comparing “start treatment X at t0” to “never start X” because those compatible with multiple regimens at t0 appear multiple times in the IPTW pseudopopulation while those compatible with a single regimen appear once. If applying CCW to nonexperimental data, clone participants into all regimens at baseline unless regimens are exclusive and exhaustive.
