A Bayesian view on causal inference under mis-specified models
Widemberg Nobre (IM-UFRJ)

In this talk, I will present recent advances in Bayesian analysis for causal inference under mis-specied models. A model is said to be correctly specified if the posterior distribution converges to the correct distribution of the data generating process as the sample size increases. Conversely, a model is considered mis-specified if it does not meet this criterion. I will focus on propensity score regression and discuss doubly robust estimation in the contexts of single-level and multi-level models. Synthetic analyses and an application to Tuberculosis data in Brazil will be discussed in order to illustrate the proposed method.