test models in the domain of the data

Marina Spivak (SCDA) led a discussion of a paper about the RELION code for the reconstruction of molecular structures from cryo-EM imaging data. We discussed spatial smoothness prior pdfs (pdfs over three-dimensional spatial models, that is), which they construct in the Fourier domain. We also discussed the optimality of estimators built from marginalized likelihoods. I'm a bit unsure about the theory after Kyle Cranmer (NYU) sowed doubt in me last week. But in general we were disappointed with the RELION papers in justifying the improvement brought by their prior: They made no argument that their resulting structures are better in the domain of the data! I don't know any other way to check, test, or compare models.