Bayesian analysis
Gist
General guide on how to do Bayesian analysis.
Workflows
For general data analysis
Martin et al. [1] suggested this Bayesian workflow for data analysis.
Directly copied from their book, but I like this list of questions one should ask when they are going to make a Bayesian model.
- (Likelihood) What distribution describes the observed data given X?
- (Covariates) What is the structure of the latent data generating process?
- (Ignorability) Do we need to model the data collection process?
- (Priors) Before seeing any data what is a reasonable set of parameters?
- Can I express my model in a Probabilistic Programming framework?
- Can we estimate the posterior distributions in a reasonable amount of time?
- Does the posterior computation show any deficiencies?
- Does the prior make sense in the context of math?
- Does the prior make sense in the context of the domain?
- Can our inference engine produce a posterior with the chosen prior?
For SIR modeling
This is the suggested workflow for modeling disease transmission using Model - SIR with Stan. But I think this workflow is generalizable to other analyses [2]
-
**Criticize model before looking at data
- Prior predictive check: We want priors that allow for every reasonable configuration of the data but exclude patently absurd scenarios, per our domain expertise.