Some publicity for the first round of papers from the COIN collaboration: the COsmostatistics INitiative. Both are concerned with promoting Generalized Linear Modelling as a stand-alone tool for basic astronomical analysis (especially exploratory analysis) and as a launching point for more sophisticated statistical analyses (such as for building hierarchical models for errors-in-variables logistic regression and so on). Paper I (de Souza et al.) introduces GLMs and presents an application of logistic and probit regression to the quantification of trends in ‘observations’ from a cosmological simulation; while Paper II (Elliott et al.) presents an application of gamma regression to prediction of galaxy redshifts given broadband photometry only (i.e. photometric redshifting).
Hopefully these papers will be successful at bringing attention to this family of regression techniques, the application of which is already widespread in healthcare settings, finance, insurance, and many other such quantitative fields perhaps attractive to astronomy PhD graduates. For instance, one of my statistical colleagues from QUT noticed that when she went for her regular check ups during pregnancy her doctor was using the model from a pre-fitted logistic regression to estimate the probability of birth difficulties based on her measured weight, blood pressure, age, resting heart rate, etc. Likewise, a British-based statistician I know took a break from academia before his PhD during which he used to fit logistic regression models for the tax office to identify the characteristics of business groups with unusually high propensity to tax avoidance (beyond the obvious, like off-meter taxis, of course!). Perhaps more exciting is the role of negative binomial regression (also from the GLM family, with fixed over-dispersion parameter) in computing betting odds for sports scores (e.g. in premier league football). In my day job we use logistic and negative binomial link functions in the hierarchical likelihood computation for our malaria parasite prevalence maps.
Anyone interested in contributing to future COIN projects is warmly invited to contact Rafael de Souza (rafael dot 2706 at gmail dot com) to discuss ideas and gain access to the COIN group on the AWOB (astronomer’s workbench) collaborative tool. The COIN team is very much non-hierarchical with no “core team” of people grandfathered in to future papers; only those who collaborate on a particular project go on that author list. On the other hand, there is a wide range of expertise within the team and it may be that some ideas will benefit from (or perhaps only become possible by) drawing on the experiences of many team members. If the enthusiasm is there, some ideas for future COIN projects might include applications highlighting the utility of quantile regression, propensity score matching, and/or meta-analysis techniques for astronomy. Again, all of these have strong relevance to outside fields as well (esp. medical research / epidemiology / population health). Also worth noting is that, although the focus of Paper I and II is relatively shallow (meaning here, introductory) coverage of a wide ranging topic (GLMs), one could easily push future projects in a deeper statistical or application direction: Bayesian versions of quantile regression can take a difficult non-parametric form, for instance.