Background: Experimental and observational studies suggest that metformin holds promise in improving survival among pancreatic cancer patients. However, findings from prior observational studies have been questioned because most did not control for immortal time bias, which can overestimate the survival benefit of a drug. In addition, previous studies did not present data on African American patients. Thus, it is unknown if any survival advantage from metformin extends to African Americans. To address these limitations, we analyzed data from the U.S. Veterans Health Administration (VHA). Methods: A population-based retrospective cohort study of 3,811 (N = 773 are African Americans) pancreatic cancer patients with pre-existing diabetes mellitus diagnosed within the VHA between October 1, 1998 and December 30, 2010, and followed until December 2014. We calculated hazard ratios (HR) and 95% confidence intervals (CI) using both the time-varying Cox proportional hazards regression model, which controls for immortal time bias, and conventional Cox model. Analyses were adjusted for confounders. We also stratified analyses by race. Further, we performed analyses among patients who were metformin naïve (N = 1158) at the time of pancreatic cancer diagnosis (most representative of patients enrolled in clinical trials). Results: Median survival was 4.5 months among metformin users versus 3.7 months among non-users. Metformin use was not associated with pancreatic cancer survival in analysis using the time-varying Cox model: HR = 1.05 (95% CI 0.92-1.14, P-value = 0.28). Results were identical among non-Hispanic Whites and African Americans. In analysis using conventional Cox model, metformin use was associated with an artificial survival benefit: HR = 0.89 (95% CI 0.83-0.98, P-value = 0.01). Among patients who were metformin naïve at the time of pancreatic cancer diagnosis, metformin use was associated with improved survival in analysis using the time-varying Cox model: HR = 0.77 (95% CI 0.61-0.98, P-value = 0.03). The HRs were 0.78 (95% CI 0.61-0.99, P-value = 0.04) among non-Hispanic Whites and 1.20 (95% CI 0.75-1.93, P-value = 0.45) among African American patients. Conclusions: We observed no associations between metformin use and pancreatic cancer survival. Nevertheless, we noted improved survival (limited to non-Hispanic White patients) among patients who were metformin naïve at the time of pancreatic cancer diagnosis, which requires conformation in other studies.