On 28 November 2014 at 19:40, Tim Gruene <t...@shelx.uni-ac.gwdg.de> wrote:
> Where freedom is given, liberties are > taken' by Kleywegt and Jones > Tim First to summarise the correct procedure: One performs optimisations with one of more starting models. By 'model' I mean here the mathematical model which includes not only the values of the model parameters and weights but also the choice of which parameters are to be varied in the optimisation. Each optimisation is terminated when the respective local, or more ideally global, minimum of the target function (e.g. negative log-likelihood) is reached, or as close to it as is numerically practical. Finally, model selection is by cross-validation. This is the fallacy with your argument: At the completion of each optimisation one selects the best model by cross-validation using the Rfree values (or better yet the free-likelihoods). It's not sensible to compare the values of the likelihoods (or Rwork) at convergence, so one is never 'comparing global minima': one is comparing the Rfrees or LLfrees for the parameter sets _at_ those global minima. Cheers -- Ian
