Abstract:
|
Continuous outcomes, such as length of stay, from stepped wedge cluster randomized trials (SWCRT), are traditionally compared using linear mixed-effects models (LMM) or quantile regression (QR). For many diseases, patients die during trial follow-up. A common approach to account for death is to give patients who die a fixed value, such as the longest length of stay, or zero-free days when free-day composite outcomes are used. While this solves a conceptual problem, it engenders a spike in the continuous distribution. Motivated by an ICU-based SWCRT, we propose using simultaneous quantile regression (SQR) to compare composite outcome distributions. SQR permits (1) a joint test of treatment effects at >1 quantiles of interest, (2) inference on the whole distribution profile using bootstrap confidence intervals, and (3) comparison of treatment effects between different quantiles. We compared SQR to LMM and QR via simulations and demonstrated that SQR maintained valid type I error rates and had comparable power, suggesting its promise for analyzing spiked distributions in SWCRTs.
|