need help to improve mixing of indicator variables in reversible jump MCMC

[Gabriel Pigeon]

Hello,

I'm trying to do variable selection using RJ MCMC. The model is written using indicator variables. As can be seen from the trace plots below, the mixing of the indicators is terrible. And why would the black chain in variable 1 never switch "off" when it's clearly not that far from 0 anyway?

what are things (in the RJ sampler or the specification of the model) I could try to improve the mixing?

- more informative prior for beta?

- Could constraints on the indicator variable (I have interactions in the model and saw a post here that suggested using constraints to avoid having the interaction present without the main effect) be the cause of the bad mixing?

thanks for the help and great package by the way

Gabriel Pigeon