Been a while since my last update here on account of having a tonne of work to get finished before a fast approaching deadline at the beginning of August. But for now I have all 64 + 12 available linux cpu cores plus my macbook air crunching out calculations, so I can spare a few minutes to talk about recent astro ph postings. The first one that caught my eye was “A statistical reconstruction of the planet population around Kepler solar-type stars” by Silburt, Gaidos & Wu. A key statistical innovation the authors claim is the introduction of Gelman & Rubin’s Iterative Simulation (IS) algorithm for an improved treatment of selection bias. However, I remain deeply skeptical that the algorithm presented in the astronomical paper (Section 2.5) bears any relation to the original IS algorithm. The former is a scheme for iteratively resampling the properties (e.g. stellar types, planet radii) of a master planet population until the output from feeding them into the simulated detection process matches the observed planet radius distribution: giving a debiased distribution somewhat like what would be obtained by weighting each detected planet by the inverse of its detectability. While Gelman & Rubin’s IS is a scheme for seeding (initializing) and then combining the output of multiple MCMC chains with a focus on the assessment of convergence towards stationary. If any readers can offer insight into the relationship between these two schemes then I’d be happy to hear from you …

Another paper worth talking about is that by Licquia & Newman (here) presenting a meta-analysis of recent estimates of the galactocentric radius of the Sun. Meta-analysis (i.e. the statistical combination of multiple past studies to produce an ensemble estimate of some unknown quantity) is a key field of work in clinical research and epidemiology, but is as yet only rarely done in astronomy. Hence, for this reason I find the Licquia & Newman paper quite encouraging. My only criticisms are that the meta-analysis models used are a bit limited in that they suppose a single common rescaling of the error bars of the component studies or a single common additive factor on the component variances; whereas in clinical meta-analysis the a more flexible model is typically used in which each study takes a unique random effect term suppose a priori to be drawn from a shared distribution. For reference, I discuss these models and their semi-parameteric Dirichlet process extension in my Fine Structure II paper.