I'm a data scientist working on a project where we need to quantify uncertainty in our model predictions for a clinical diagnostics tool, and my team is considering a shift from our frequentist methods to a Bayesian statistics framework to better incorporate prior knowledge from earlier studies and provide more intuitive probabilistic outputs. I understand the theory conceptually, but I'm struggling with the practical implementation, specifically choosing appropriate priors and selecting the right computational tools like Stan or PyMC3 for our large dataset. For practitioners who have made this transition, what were the biggest hurdles in adopting Bayesian statistics for real-world, production-level analytics? How do you effectively communicate the results, like credible intervals, to stakeholders accustomed to p-values, and what resources would you recommend for building competency beyond introductory textbooks?
Hurdles are usually threefold: choosing priors that reflect domain knowledge without bias, getting reliable computational performance on large data, and communicating uncertainty to clinicians and leadership who expect p-values. A pragmatic path is to start with a small, well-specified Bayesian model (e.g., logistic regression with a few key predictors) and build up. For priors, use weakly informative choices: normal priors on coefficients (for example, Normal(0, 2.5)) and appropriate priors on variance terms (half-Cauchy or half-Normal for scale). Run prior predictive checks to ensure your priors generate plausible data before you even look at the real data. Then add hierarchical structure to borrow strength across sites or subgroups if that makes sense for your data. On computation, Stan is reliable for MCMC with large datasets; PyMC3 (or PyMC) is very flexible if you want easier Python integration. If you’re dealing with truly massive data, consider VI as a stopgap while keeping a smaller, high-fidelity MCMC model for validation.