MS thesis abstract:

This study uses a Bayesian approach and survival models to analyze a large observational data set obtained from electronic health records. We model the association between DPP4 and SGLT2 diabetes therapies and major adverse cardiovascular events. The Bayesian approach allows us to incorporate information from previous studies and obtain credible intervals. Credible intervals allow us to make probability statements when discussing the parameters of interest. To address the lack of randomization, we implement propensity score matching using the nearest-neighbor approach and a caliper. We compare the traditional Cox proportional hazards model to three Bayesian survival models: one with an uninformative prior, one with a prior derived from a metaanalysis of previous trials, and one with a prior having a small variance. We compare results by looking at common estimates of interest, including the survival function, hazard ratio, and restricted mean survival time. We found that a Bayesian model with an uninformative prior has similar results to the Cox proportional hazards model. Models with informative priors are an effective way to incorporate clinical knowledge but note that the variance of the prior should be considered carefully.