Parameter Identification tests _Bayesian


I got a question,
I looked at those graphs coming out of the parameter identification tests (REF: Pfeifer 2013, Interpretation of Graphs in Dynare)

It seems it is impossible to pass all those tests as described. My model will fail in some of them no matter what I change and no matter how few parameters I decide to estimate.

This is related to a more serious problem. When I estimate the mode in (trying different algorithms) I get a certain set MODES for my parameters.
In a separate model with only minor changes I get completely opposite results. I.e in my first model shock

explains my the variable of interest 90 %. In my second model shock

explains 95 % of my variable of interest. Could this be related to parameter identification issues in both my models?

Thanks a lot

One issue is identification strength (which is a matter of degree). The other is local identifiability (whether a parameter even affects the observed moments at all, regardless how weakly). Particularly the latter is a conceptual problem. Priors are always going to give you curvature - unless they are uniform.

When you keep the model fixed and get different results in mode-finding, there are several local modes. You need to select the global one. This is often challenging and requires starting at different points and using different mode-finders and then select the one with the highest posterior density.

When you change the model, even small changes can have big effects on the model dynamics. In this case it is harder to say whether changes in results are due to problems in finding the correct mode (which presumable should be close to the one in the previous slightly different model) or due to the actual changes having big consequences.