A Comparison of Methods for Estimating an Average Treatment Effect on Hometime

Oregon Convention Center 

Hometime, the number of days a patient is home within a specified time period following a clinical event, is an important patient-centered outcome. Modeling hometime is challenging due to its zero-inflated, right-skewed, and right-censored distribution. There is a lack of consensus in clinical literature on how to best make inference on the average treatment effect on hometime. This work aims to provide practical recommendations on methods selection for this inference. As hometime's U-shaped distribution does not fit a known distribution, we simulate data by generating patients' daily status (home vs. not home) over two periods (90 days and 365 days) to obtain realistic data. Due to the unique data-generating process, any non-zero treatment effect is unknown. Using the simulated data, we compare popular methods of estimating a treatment effect on hometime under different settings, including various types of censoring. In order to compare a range of methods that estimate the treatment effect on different scales, we focus on performance with respect to type I error under the null and power under the alternative. Our findings are applied to a large cardiovascular disease registry.

Hometime

Days Alive and Out of Hospital (DAOH)

Epidemiologic Methods

Clinical Research 

Biometrics Section