Is there such a thing as equivalence between statistical models (in my particular case, Bayesian hierarchical models) ? If so, how to prove it ?

Let me explain myself with an example. Consider a Gaussian linear regression model $y\sim \mathcal{N}(X\beta^T, \sigma^2)$. I parametrized the normal distribution with the variance $\sigma^2$ but I could use the standard deviation, the precision, even the log variance... but the models would all be equivalent Gaussian models. From this example, I guess that one has to show that the density from the first model is equal to the density of the second model, using a change of variables ( https://en.wikipedia.org/wiki/Probability_density_function#Vector_to_vector ). Is this a correct approach ? Some references for a principled definition of model equivalence would be welcome.

However, my original problem is not exactly this one. I am working on Bayesian hierarchical models, and I have two models : a ''basic'' one, and a ''richer'' one with more parameters. I would like to show that the second model is an extension of the first model. My approach is :

  1. Make some assumptions to harmonize the hyperpriors of the complex and simple model
  2. Condition the complex model by an event $E$ that "freezes" the additional variables of the complex model with respect to the simple model
  3. Show that, conditionally on $E$, the densities are the same

What happens in step 3 is that I do not find strict equality of the distributions but equality within a very simple bijective transformation of the parameters. This is my first question (I could tinker either model to make the problem vanish, but it would be a crooked solution imo). Additionaly, any comments on my approach ?


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.