Bayesian model selection is premised on the assumption that the data are \ngenerated from one of the postulated models, however, in many applications, all of \nthese models are incorrect. When two or more models provide a nearly equally good \n t to the data, Bayesian model selection can be highly unstable, potentially leading \nto self-contradictory ndings. In this paper, we explore using bagging on the posterior \ndistribution (\\BayesBag") when performing model selection { that is, averaging the \nposterior model probabilities over many bootstrapped datasets. We provide theoreti- \ncal results characterizing the asymptotic behavior of the standard posterior and the \nBayesBag posterior under misspeci cation, in the model selection setting. We empir- \nically assess the BayesBag approach on synthetic and real-world data in (i) feature \nselection for linear regression and (ii) phylogenetic tree reconstruction. Our theory \nand experiments show that in the presence of misspeci cation, BayesBag provides \n(a) greater reproducibility and (b) greater accuracy in selecting the correct model, \ncompared to the standard Bayesian posterior; on the other hand, under correct speci- \n cation, BayesBag is slightly more conservative than the standard posterior. Overall, \nour results demonstrate that BayesBag provides an easy-to-use and widely applicable \napproach that improves upon standard Bayesian model selection by making it more \nstable and reproducible.