(126) Use of a Potential Outcomes Framework to Simulate Bias Due to Unidirectional Treatment Switching in Real World Studies

Publication Number: 457

Background: Pseudo-randomization techniques such as propensity score matching are commonly used for confounding control in comparative pharmacoepidemiology research. Applications of these methods tend to assume initial treatment compliance, which may be problematic in real world data especially when treatment guidelines are evolving.

Intention-to-treat (ITT) effect estimates are popular, because, under the assumption of confounding control at baseline, bias due to treatment switching is commonly thought to be conservative (i.e., toward the null).

Objectives: Using a potential outcomes framework for disease causality, simulate the direction and magnitude of bias in the ITT hazard ratio (HR) in the presence of unidirectional treatment switching when baseline confounding control is achieved.

Methods: We simulated a cohort study with 1,000 pairs of exposed and unexposed patients and their potential outcomes: 1) doomed experience the outcome regardless of exposure; 2) harmed experience the outcome if exposed; 3) protected do not experience the outcome if exposed; 4) immune are not at risk of the outcome.

To simulate confounding control at baseline, the sampling distribution of potential outcomes was equivalent among the exposed and unexposed. To simulate a monotonic treatment effect, the proportions of doomed, harmed, protected, and immune patients were set to 0.27, 0, 0.06, and 0.67, respectively. For a non-monotonic effect, harmed patients were included in the population, and the proportions of potential outcomes were 0.18, 0.14, 0.21, and 0.46.

In both scenarios, only unexposed patients were at risk of treatment switching. Doomed and protected types had a 0, 50, 70, and 95% chance of treatment switching. Immune and harmed types did not switch.

Distributions of ITT HRs were obtained over 1,000 repetitions.

Results: When the effect was monotonic, the mean [standard deviation (SD)] ITT HR was 0.81 [0.07] when crossover did not occur and converged on the null (HR = 1) as the rate of crossover increased to 95% (mean [SD] 0.89 [0.08], 0.94 [0.08], 0.99 [0.08]). When the effect was non-monotonic, the ITT HR under treatment switching was directionally different (mean [SD] 1.12 [0.09], 1.32 [0.12], 1.69 [0.15]) from the compliance HR (mean [SD] 0.82 [0.06]).

Conclusions: Even when patients’ potential outcomes are balanced at baseline, the ITT HR in the presence of treatment switching may be directionally different from the compliance ITT HR when the effect is non-monotonic and the mechanism that drives crossover is related to the outcome. This may be the case when clinical practice is evolving and new treatments are given to the patients that are most likely to benefit.