5B-05 - The Dangers of Stabilizing Censoring Weights

Background: Weighting is a popular tool to address confounding and selection bias. When creating inverse probability of treatment weights (IPTW) to address confounding, one very large group results in large weights for smaller groups. “Stabilizing” IPTW, e.g., by multiplying by marginal probabilities of treatment, centers weights around 1 and can reduce variability. The benefits and drawbacks of stabilizing weights used to address informative censoring (often referred to as “inverse probability of censoring weights” or IPCW) is less well established.

Objectives: Examine how stabilizing IPCW impacts estimates of the incidence rate ratio (IRR) and hazard ratio (HR) in simulated studies using cloning, censoring, and weighting (CCW) to contrast the effect of two treatment regimens.

Methods: We simulated 1,000 replicates of 25,000 patients with a binary outcome Y, a binary exposure X, and a confounding variable W associated with Y and the probability of starting X. Within each replicate, we used CCW to estimate IRRs and HRs comparing “start X by time 1” and “never start X” treatment regimens with and without stabilizing the CCW using the marginal probability of censoring at each time point. We examined four different scenarios: 1) no effect, no association between time and outcome conditional on W; 2) constant HR for X; 3) no effect of X with a time-varying hazard conditional on X and W; and 4) constant HR with time-varying hazard. We calculated relative bias vs the true IRR and HR obtained from Monte Carlo simulations of 10,000,000 patients.

Results: Stabilization improved the precision of treatment effect estimates, regardless of measure or scenario, with confidence limit widths shrinking by 25% to 13%. The unstabilized approach was unbiased for the IR and HR regardless of scenario (relative bias = 1.00). The stabilized approach was unbiased when estimating the HR in scenario 1 and scenario 3 but biased when estimating the HR in scenario 2 (bias = 0.98) and 4 (bias = 0.95). Stabilization was always biased when estimating the IRR (scenario 1 bias = 0.90, scenario 2 bias = 0.88, scenario 3 bias = 0.69, scenario 4 bias = 0.66).

Conclusions: Stabilizing IPCW improves precision of IRRs and HRs at the cost of introducing bias, especially when the probability of experiencing the outcome varies over time. This bias is the result of stabilization under-representing later sections of person-time relative to what would be expected if we were to actually assign the entire population to the treatment regimen of interest. IRRs are particularly susceptible to bias from stabilization. The combination of increased confidence in biased estimates makes IPCW stabilization a risky choice.