I'm comparing Bayesian (generalized) linear methods vs Frequentist ones in the case when $p>n$ ($p,n$ being respectively number of regressors, number of samples).
In the frequentist context when $p$ > $n$ I cannot invert the design matrix $X$ since it is not full rank. Of course there are methods like Ridge Regression to deal with this problem, but I don't want to consider them here. So in that case there is an identifiability issue.
My question is this:
Is there any case in which the same problem (impossibility of finding regression coefficients when $p>n$) arises in a Bayesian context? For instance, considering the simple linear model, when I specify the priors as Normal-Invgamma (like here) I have no need to invert the design matrix to obtain the posterior parameters, solving in that way the problem which occurs in the frequentist case.
I'm dealing with pretty complex multilevel models and I let STAN do the computations, so matrix algebra (which is the problem in the linear context) is not something I care about, may be wrongly. Thanks to @Tim I read Michael Betancourt answer in stanusersgroup look here and what I understood is that the prior imposition constraints the posterior sampling to a convex set, not always solving the identifiability problem,but at least allowing to spot high correlations between the regressors.
Thank you in advance