Statistical modeling of days alive out-of-hospital: An illustration using the SSU trial

Days alive out-of-hospital (DAOOH) is a well-accepted outcome in clinical research, as it summarizes both hospitalization and mortality. However, its distribution is often skewed, making accurate modeling challenging in practice. We utilized data from a randomized trial to discuss estimands and statistical methods of interest for DAOOH.

We examined several estimands and statistical models targeting these estimands, including linear regression, gamma regression, ordinal regression, and the probabilistic index model. Comparison was based on interpretability, precision, and model fit. We adopted G-computation to assess causal effects leveraging different regression models to target a simple unified estimand.

For unadjusted analyses, the Mann-Whitney U-test is a simple and assumption-light approach, which can be paired with a probabilistic index to enhance interpretation. For adjusted analyses, while the mean difference and mean ratio are intuitive, the methods targeting them (linear and gamma regression) showed sub-optimal model fit. Conversely, ordinal regression showed a superior fit, but its odds ratio estimand is difficult to interpret. Pairing gamma or ordinal regression with g-computation facilitates estimation of a marginal causal mean difference estimand under certain assumptions, allowing for both a good model fit and interpretability.

Our findings illustrate trade-offs of estimands and model choices for DAOOH, providing guidance for method selection. Based on data from this randomized trial, the Mann-Whitney U-test is recommended for unadjusted analyses. Ordinal regression paired with g-computation provides the best balance of precision, model fit, and interpretability to report an adjusted effect.