After confirming (for about the tenth time) that my own paper was indeed properly processed by the arXiv demons I had a more thorough scan down the new astro-ph submissions list for today and found three other papers presenting Bayesian methodologies and applications in astronomy. Lucy ( http://arxiv.org/pdf/1309.2868.pdf ) has an application of Bayesian inference to the challenge of mass estimation for visual binaries with limited data; Martinez ( http://arxiv.org/pdf/1309.2641.pdf ) has a hierarchical Bayesian modelling application to mass estimation for the Milky Way dwarf satellites; and Foley & Mandel ( http://arxiv.org/pdf/1309.2630.pdf ) have an application of Bayes theorem to a SNe type classification problem.
My first thought: there’s obviously no shortage of Bayesians in the astronomical community, so perhaps it’s time for organisers of astro-statistics themed conferences/sessions to start getting a little more adventurous in their choice of speakers? I have the impression that only a few names keep popping up at these things time and time again. But maybe that’s just me? (Having said that, I did see Mandel’s talk at ISBA 2012; so it’s not like we’re ignored entirely!)
My two cents on the above papers: (1) At first read they all seem to present sound, straightforwards Bayesian methodologies. BUT, I think Martinez unnecessarily complicates his presentation of the hierarchical modelling process (such that it might in fact even dissuade its potential audience of astro-Bayesian neophytes from trying it out), and I think his attempt to present the hierarchical approach as something more profound than a (modelling) device for shared learning across groups (here, galaxies) comes off as rather unconvincing. (2) I was surprised that Foley & Mandel opted for a binned structure in their analysis where a parametric form could have been used (for relating type fractions to predictor variables). There is a vast range of flexible such models within the GLM family (e.g. with the logistic or probit link functions) with pre-built routines for fitting them. One advantage of the parametric form is that one can then include observables for which you might not have complete coverage over the entire SNe dataset, while correcting for selection biases in the availability of these observable via propensity score matching (which I have something of a penchant for).