The final project for this class had my group explore a past Frequentist Analysis in a Bayesian light. Since I worked with the same person for my Survival Analysis final project, we decided to use this analysis.
We implement a Bayesian Weibull proportional hazards model. The Weibull distribution was chosen based on the parametric survival analysis from our Stat 417 report, which identified it as the best-fitting distribution (Anderson-Darling test statistic = 16986.795). We will use the same predictors in the Bayesian model as we did in our frequentist model to enable a direct comparison between the two.
The Bayesian and frequentist approaches yielded similar estimates for key effects:
Contract effects differ by less than 2%
Payment method effects are consistent in direction and similar in magnitude
Both identify the same key predictors of churn
The main differences are:
The Bayesian approach doesn’t suffer from proportional hazards violations
We get full posterior distributions rather than point estimates