Abstract / Summary
Abstract Background: Because clinicians selectively request cardiac testing for emergency department (ED) patients judged to be at higher risk, a naive comparison of tested and untested patients confounds the effect of testing with its indication. We conducted an observational target-trial emulation to quantify how far adjustment for recorded covariates can resolve this confounding.
Methods: We linked ED encounter, patient, and ECG metadata and emulated a two-strategy trial: cardiac testing (stress testing or catheterisation) within 10 days of an ED visit versus none, evaluating the 30-day risk of a composite of positive cardiac troponin or death. The estimand was the marginal risk difference, estimated by five-fold cross-fitted augmented inverse-probability weighting (AIPW) with logistic and boosted-tree nuisance models. We also report patient-clustered inference, a one-encounter-per-patient restriction, a grace-period sensitivity analysis, a check on a covariate of doubtful baseline status, and an E-value for unmeasured confounding.
Results: The analytic cohort comprised 51,158 of 71,460 encounters (20,302 excluded), of whom 4,609 (9.0%) received testing and 4,747 (9.3%) had the outcome. The crude risk difference was 22.1 percentage points (pp). Cross-fitted AIPW gave 5.9 pp (95% CI 3.9-7.9) with logistic nuisance models and 5.1 pp (3.6-6.6) with boosted trees; patient clustering left the interval essentially unchanged. The estimate was less stable under two checks: it fell to 0.2 pp (95% CI -1.3-1.7) when the exposure-ascertainment window shortened from 10 to 1 day, rising steadily back to 5.9 pp as the window widened, and rose to 11.1 pp (9.2-13.0) when a same-day AMI-diagnosis covariate (present in 42.8% of tested but 0.0% of untested encounters) was dropped from the adjustment set. The E-value for the main estimate was 2.66 (2.19 at the confidence limit closer to the null).
Conclusions: Adjustment substantially reduces, but does not eliminate, the crude contrast, and the residual estimate is not robust to the length of the exposure-ascertainment window or to the baseline status of covariates measured close to the treatment decision. The project illustrates both a causal machine-learning workflow and why a well-specified estimand does not, by itself, convert observational care data into treatment-effect evidence.