[Gretl-users] Re: The new ParMA package

Hi everybody, and thank you Sven for the suggestion: I’ll try to explain the main differences as clear as possible. The BMA package by Marcin Błażejowski and Jacek Kwiatkowski performs Bayesian Model Averaging (BMA) in linear models using the original approach by Madigan, York and Allard (1995), that is using a Metropolis-Hastings MCMC on the model space, which actually works nicely in linear models where analytical formulae are available. However, in non-linear alternatives the same cannot be applied because of the lack of analytical formulae, so the ParMA package comes into play at this point: it allows the use of BMA for not only linear models, but also for some other families such as binary or count models using the Reversible Jump MCMC framework by Green (1995). Since the Reversible Jump MCMC is quite CPU demanding with respect to the canonical MCMC ( since it explores models and model coefficients simultaneously), we enable parallelisation via MPI and convergence monitoring via additional statistics. Therefore, in order to fully exploit the ParMa package, MPI should be enabled. As concerns linear models, the ParMA pkg can be still applied, and it actually gets similar results wrt the BMA pkg, however it is less efficient in terms of elapsed time (the Reversible Jump MCMC is particularly suited for non linear cases) even though it compensates for this drawbacks by directly providing information on model coefficient marginal distributions. On the other hand, however, the BMA pkg includes jointness analysis. I’m sorry if it looks a bit technical :(