I’m using Stan to fit clines of allele frequency across a hybrid zone. There are multiple cline models which can be used to fit this sort of data. The models share the same core parameters (the center and width of the hybrid zone), but vary in their inclusion of other parameters to describe the tails of the clines.
I’d like to include uncertainty with regard to model choice and get the model-averaged parameter estimates of the parameters I care most about (cline width and cline center). But, I’m unsure how best to go about this. In searching this forum and the Stan google groups, it seems using mixture models is the preferred way of doing model averaging within Stan, but I’m baffled about how to apply this to my models. Attached are two of the .stan models I’m using and a dataset from which to estimate clines. Any advice on how to begin making mixture models from these would be much appreciated.
Finally, a related but non-Stan specific question: my (probably naive) expectation when I began thinking about model averaging was that I could construct a sort of “model-averaged posterior” for each parameter by randomly drawing samples from the posterior distributions of each model, with the chance of drawing from a particular model being equal to that model’s Akaike weight (calculated from WAIC). This idea came from Richard McElreath’s Rethinking Statistics book, where he uses the same technique to make an ensemble of posterior predictive checks from multiple models. I wondered if that could work for parameter estimates as well and I could just avoid the mixture modeling process.