Average Response Curves for Treatment Time in the Emergency Department

Abstract

We estimate average responses curves for treatment time in the Emergency Department (ED). Extending treatment time is considered a promising solution for improving admission decisions. Providing empirical support for this solution, however, is difficult because this intervention (treatment) is a continuous time-to-event; is strongly influenced by unmeasured patient health needs and is jointly determined with the admission decision (admit vs. discharge); and may be only modifiable up to a shift in the realized time. We formalize the admission process as a directed acyclic graph and show that average responses curves for treatment time cannot be identified nonparametrically due to unmeasured confounding from patient health needs. We thus use a parametric model that includes a latent variable for health needs and a threshold regression model for the admission process. We fit this model to observational data (n = 28,862) from abdominal pain patients at a large tertiary teaching hospital. We estimate that fixing ED treatment time to 2 hours rather than 1 hour decreases admission rates from 41.6% (95% CI: [40.6, 42.7]) to 32.7% (95% CI: [32.2, 33.2]) and that increasing the realized treatment time by 30 minutes can reduce admission rates by 1.1% (95% CI: [-1.1, 3.2]), with little change to 30-day revisit and readmission rates.