(187) Identifying the average treatment effect in the treated: Illustration using the Northwest Germany Stroke Registry.

Background: The average treatment effect (ATE) and average treatment effect in the treated (ATT) are typical parameters estimated in clinical treatment studies. The ATE and ATT may differ in certain scenarios. When the ATT is preferred, the conditions sufficient for identification are a subset of the conditions for the ATE.

Objectives: To provide an example where the ATE and ATT differ and conduct a simulation study to evaluate the identification of the ATE and ATT under varying settings.

Methods: The Stroke Registry of Northwestern Germany (QSNWD) has documented treatment characteristics of stroke patients since 1999. Tissue-type plasminogen activator (t-PA) is a mainstay of ischemic stroke treatment but has strict eligibility requirements. Using registry data from 2020-2021, we updated an analysis published in 2006 that investigated the impact of t-PA on preventing in-hospital mortality under different estimation approaches. We estimated the probability of t-PA receipt using logistic regression and used the resulting propensity scores to estimate inverse probability of treatment weighted (IPTW) effect estimates to target the ATE and standardized mortality ratio (SMR) weighted effect estimates to target the ATT.
We conducted 5,000 simulations of 6,000 patients from three populations with differing prevalence of a treatment indication. We assigned treatment under two causal identification scenarios: 1. treatment assignment is independent of the treatment indication (ATE identification), and 2. patients without the indication cannot receive treatment (ATT identification). We generated outcomes assuming a treatment effectiveness risk ratio of 0.8 or 1. We applied the same statistical analysis as above.

Results: Among 6,000 randomly sampled patients eligible for analysis, 20% received t-PA, and 5% died. The propensity score distributions by treatment had limited overlap. The IPTW risk ratio targeting the ATE was 2.27 (95% CI: 0.89, 5.76). The SMR-weighted risk ratio targeting the ATT was 0.82 (95% CI: 0.55, 1.21). The simulations demonstrated that the ATE and ATT estimates were unbiased in scenarios where treatment assignment does not depend on the indication. In scenarios where patients not indicated for treatment do not receive treatment, only the ATT was unbiased. The magnitude of the bias in the ATE estimate increased as the prevalence of the indication increased.

Conclusions: When targeting the ATE, the effect estimate of t-PA on in-hospital mortality shows harm, whereas the effect estimate targeting the ATT indicates benefit. The simulations show that if causal identification conditions are only met in the treated, epidemiologists can leverage the weaker conditions to consistently estimate the ATT, even when estimates targeting the ATE are biased.