First introduced by Daniela Pauger and Helena Wagner in this 2019 paper[1], effect fusion priors are designed to induce sparsity on the differences between the coefficients that represent the effects for each level of a categorical variable.
My colleague Javier Aguilar and I are currently trying to expand on the core idea in ways which we hope will be useful to applied researchers. An overview of these research directions can be found on this poster that I presented at Bayesian Methods for the Social Sciences II.
A basic R implementation of the density (sampling only) that I used for the plots in the poster can be found here.
| [1] | Actually, the earliest record I can find of their idea is their 2014 contributed paper for the 29th International Workshop on Statistical Modelling, followed by a 2017 arXiv preprint (which I prefer over the 2019 version as it includes the very informative appendices in the same pdf), but the "official publication" somehow feels like the thing I should try to bring the most attention to, so here we are. |