1: \begin{abstract}

2: Parametric Bayesian modeling offers a powerful and flexible toolbox for

3:  scientific data analysis. Yet the

4:  model, however detailed, may still be wrong, and this can make

5:  inferences untrustworthy. In this paper we study nonparametrically

6:  perturbed parametric (NPP) Bayesian models, in which a parametric

7:  Bayesian model is relaxed via a distortion of its likelihood. 

8:  We analyze the properties of NPP models when the target

9:  of inference is the true data distribution or some functional of it,

10:  such as in causal inference. We show that NPP models can offer the

11:  robustness of nonparametric models while retaining the data

12:  efficiency of parametric models, achieving fast convergence when the

13:  parametric model is close to true. To efficiently analyze data with

14:  an NPP model, we develop a generalized Bayes procedure to approximate

15:  its posterior. We demonstrate our method by estimating causal effects of gene expression from single cell RNA sequencing data. NPP modeling offers an efficient approach to robust Bayesian inference and can be used to robustify any parametric Bayesian model.

16:  \end{abstract}

17: