UC Santa Barbara

In practice, survival analyses appear in pharmaceutical testing, procedural recovery environments, and registry-based epidemiological studies, each reasonably assuming a known patient population. Less commonly discussed is the additional complexity introduced by non-registry and spatially-referenced data with time-dependent covariates in observational settings. In this short report we discuss residual diagnostics and interpretation from an extended Cox proportional hazard model intended to assess the effects of wildfire evacuation on risk of a secondary cardiovascular events for patients of a specific healthcare system on the California's central coast. We describe how traditional residuals obscure important spatial patterns indicative of true geographical variation, and their impacts on model parameter estimates. We briefly discuss alternative approaches to dealing with spatial correlation in the context of Bayesian hierarchical models. Our findings/experience suggest that careful attention is needed in observational healthcare data and survival analysis contexts, but also highlights potential applications for detecting observed hospital service areas.