As I mentioned in my earlier post on the future of astrostatistics, there is a tendency for astronomers to (imperfectly!) reinvent ABC for likelihood-free inference at regular intervals. The latest is this paper by Shimizu & Mushotzky which extends a method attributed to Uttley et al. (2002) for power spectral density (PSD) fitting via simulated-observed data comparison in the face of complicated observational errors (aliasing, red noise, white noise, etc.). Like most such reinventions (e.g. Sana et al. 2012 in the field of massive star modelling) the authors focus directly on a classical “chi squared” (translation = “weighted-sum-of-squares”) significance test as discrepancy distance; and in this case use the same for both parameter constraint *and* model selection (the potential pitfalls of which, to be fair, are perhaps still known only to statisticians working in this field; Robert et al. 2011).

Interesting perhaps is the fact that most astronomers are not only unaware of the past ten years or so of approximate likelihood developments in statistical methodology but also of the equivalent reinventions of these methods in other astronomical sub-disciplines (here, massive star modelling vs. AGN variability modelling). A good demonstration of why we need “astrostatistics” as an umbrella field, perhaps.

UPDATE: These things come in threes; so after another reinvention today by Holman et al. we should be expecting one more tomorrow …

### Like this:

Like Loading...

*Related*