Hi, I’m trying to find solutions for the bimodal posterior distribution issue I encountered recently. My model converges well and does not have warnings related to divergent transition or ESS, however, the parameter recovery indices (e.g., correlation and biases) look bad that indicates the estimates are not accurate. I then plotted the posterior distributions and there are some posterior distributions bimodal or there is a ridge, which I think is the cause of the bad estimates. I tried some possible solutions: (1), fix one of my parameter priors to N(0,1) to identify the model, (2) reparameterize the model to have hierarchical priors, (3) set vague priors (and this one has divergent transitions problem), but none of them works.
I also saw other suggested ways: (1) set stronger priors or change initial values to be close to the true value, however, N(0,1) is already tight to me and is commonly used in my area, and change initial values to be close to the true value is not useful if I don’t know the true values in real applications, so do you have any other suggestions on how to deal with the bimodal? Thanks!
You’ll need to include your model code and some pairs plots for folks to really be able to advise here.