Two of the most fundamental principles that underpin (different) strands of statistical practice are in apparent conflict: 1. Bayesian inference provides a uniquely self-consistent way of performing model comparison, but requires that there are (at least) two models under consideration. 2. Pure hypothesis tests (of the sort most commonly based on a p-value) are routinely used to question or reject “null” models even when there is no plausible alternative under consideration. This talk describes an attempt to reconcile these two perspectives, with the natural outcome being a Bayesian hypothesis test. Some aspects of the resultant formalism work well, but some quite fundamental difficulties arise as well.
Biography:
I did an undergraduate degree and a PhD in physics at the University of Melbourne before going to Cambridge in 1999 to work on the Planck cosmic microwave background satellite. During subsequent post-doctoral positions and Cambridge and Imperial I became increasingly interested in the data analysis aspects of astronomy, and Bayesian inference in particular. In 2011 I started a joint physics-mathematics lectureship at Imperial College London and in 2016 I also took up a part-time post as guest professor in astronomy at Stockholm University.