OBJECTIVES: To illustrate the application of an extended Cox regression model with time-dependent treatment effects to real-world observational data. METHODS: A retrospective US hospital database analysis was conducted among adult invasive aspergillosis (IA) patients receiving their first antifungal therapy during ICU stay. To assess the initial antifungal treatment effect on survival/LOS, a survival analysis was conducted with event defined as “discharged alive”, censored being “expired in hospital”, and time variable being treatment-initiation-to-discharge (DITD). Key demographic, clinical and treatment variables were included. Regimen A was the reference group and compared to regimen B (inactive against IA, possibly indicating a treatment delay). Proportional hazard assumptions for the treatment variable were assessed by the Schoenfeld residuals significance test. When significant, treatment-by-time interaction together with its main effects were included within an extended Cox model to account for the proportional hazard violation. Multiple functional forms for time were considered in treatment-by-time interactions. Due to the lack of direct SAS output for specific time points, time-specific hazard ratios (HR) between treatments were manually calculated incorporating both main effects and time covariates. RESULTS: A continuous linear treatment-by-time interaction was constructed for Drug B since the null hypothesis of Schoenfeld test was rejected. Consistent with an increasing time-interaction effect (HR=1.032, p<0.0001), HR of drug B increased over time: Drug B patients were 57% less likely (HR=0.43, p=.0001) to be discharged alive compared to Drug A at mean switch time to other IA-active drugs (9 days), and 27% less likely (HR=0.73, p=0.0297) to be discharged alive at mean DITD (25 days).  Hazards would be equivalent (HR=1) at day 35, although by then most patients on regimen A (81%) had switched/been discharged. CONCLUSIONS: The extended Cox model can be implemented as an adjustment for the proportional hazard violation in observational data analysis to obtain unbiased survival results.