Some might say it’s bad form to blog about one’s failed job applications, you know, keep a stiff upper lip and all that. On the other hand, to do so is deeply cathartic and may give others a smile when thinking back on their own rejections. So here goes.
As I note on the ‘About Me’ page of this blog, I’m very happy in my work post-astronomy but I do hold some interest in the possibility of returning should the right opportunity to make a difference in the field of astrostatistics present itself. By ‘right’ I mean two things: that the department has a genuine enthusiasm for astrostatistics (whether in terms of research, teaching, or both), and that the salary is sufficient to provide a reasonable standard of living. With this in mind I decided to send in an application for an advertised position as lecturer in astrostatistics at the University of Cambridge.
To be fair, the salary attached to this faculty position (£38,896-£49,230) was woeful which might signal to a more perceptive potential applicant than myself that the sought candidate would be a junior researcher. That said, one might also find it hard to imagine a junior researcher coping well with the responsibility of lecturing students at a high level in a complex discipline encompassing elements of both astronomy and statistics. Anyway, since I’m happy with my current work at the University of Oxford in the MAP team—and in particular since I’m rather curious about the opportunities at our new ‘Big Data’ institute which we’ll move into in February of this year—I felt I could afford to state up front in my application letter that I couldn’t accept an offer below £55,000. If that’s what sunk my application then I have no regrets about asking, but I would then be sad for potential students that the university would have such a miserly attitude towards the staff they employ to provide tuition.
Regarding my other (not-entirely-unrelated) condition—that the department has a genuine enthusiasm for the teaching of astrostatistics—I do also wonder. On the one hand I’ve seen their past astrostatistics course plan and it’s pretty damn softball from a statistical point of view. That is, there’s no mention in the syllabus of (and therefore I assume no time given in the course itself to) any substantive foundations for the probability theory upon which modern statistics is built: no measure theory, no Kolmogorov’s axioms, no probability triplets and random variables, no concepts of distributional convergence, no characteristic functions, etc. etc.—and then nothing for modern computational statistics beyond MCMC samplers: no importance sampling, no particle filters / SMC / population Monte Carlo, no samplers for complex stochastic processes (GPs, Dirichlet process, Markov Jump process, etc.) etc. Instead it’s just the quick set of basics for getting familiar with a couple of statistical software packages; the same kind of course I see in fields that astronomers might otherwise view as being for the less mathematically gifted: zoology, social science, psychology. While they do have some dudes at Cambridge who do strong work in astrostatistics, like Mike Hobson and Lindley Lentati, I might guess that they are not the ones steering the ship here.
Anyway, whatever, their loss …
Yeah, that felt good.
ps. I’m only assuming that I’m rejected because the job ad said they’d let successful candidates know by late December; if in fact I’m not rejected that would be rather amusing.
If anyone’s still reading here’s a little piece from my application regarding my thoughts on the importance of a strong astrostatistics program.
The development and application of complex statistical methodology has a long history in astronomical research from Le Verrier’s predictions via perturbing functions for the position of Neptune to the theory & simulation of random fields in cosmology (e.g. Bardeen et al. 1986; Kamionkowski et al. 1997)—and more recently, the probabilistic detection of transiting exoplanets amongst noisy stellar light curves (e.g. Ford et al. 2007), and the search for gravitational waves via pulsar timing arrays (Cornish & Sampson 2016). Astronomical problems have also served as inspiration for a host of important statistical investigations; see, for instance, the role of Postman, Huchra & Geller’s (1986) ‘galaxy dataset’ in studies of Normal mixture models (Roeder 1990; Richardson & Green 1997), or the role of Efron & Petrosian’s (1998) ‘quasar dataset’ in the study of restricted permutations (Diaconis, Graham & Holmes 2001; Liu 2001). Despite this historical potential, the cross-over of statistical methodology between astronomy and statistics (and vice versa) has been disappointingly limited to-date.
Innovations from astronomy, such as the method for conditional simulation of Gaussian processes identified by Hoffman & Ribak (1991), are left to be rediscovered independently by statisticians (see Doucet 2010). While algorithms first applied to, and largely developed on, astronomical problems, such as nested sampling (Skilling 2005, Mukherjee et al. 2006; subsequently MULTINEST and POLYCHORD; Feroz & Hobson 2008, Handley et al 2015), are often ignored by the statistics community (modulo Chopin & Robert 2010 in the case of nested sampling). On the other hand, many modern statistical techniques—for example, particle filtering & sequential Monte Carlo schemes (Del Moral et al. 2006), semi-parametric Bayesian algorithms based on the Dirichlet process (Escobar & West 1995; Roy & Teh 2009), and SPDE-based approximations for mapping continuous GPs to discrete GMRFs with sparse precision matrices (Lindgren et al. 2010, Rue et al. 2008)—are yet to be adopted by the astronomical community (with few exceptions: e.g. basic SMC samplers for marginal likelihood estimation along the thermodynamic path; Kilbinger et al. 2010).
Behind this paucity of knowledge exchange between the two communities is a lack of training: few astronomers receive any statistical tuition during either their undergraduate or post-graduate degrees. Instead they will typically learn on the job while re-coding the old algorithms from a supervisor’s paper, or while flipping through NUMERICAL RECIPES upon facing a problem that seems not to be solvable by a ‘chi-squared fit’. More recently, they may receive some further instruction from a self-styled ‘astro-statician’, a well-meaning post-doc who has managed to digest a portion of Gelman et al.’s BAYESIAN DATA ANALYSIS or Jaynes’ PROBABILITY THEORY: THE LOGIC OF SCIENCE but has very likely not found the motivation to delve deeper or to take an active interest in the statistical literature. Indeed, usually the latter two will go hand-in-hand: how is one to understand, say, Rao & Teh’s (2013) algorithm for posterior simulation from Markov Jump processes (perhaps with which to model state switching in variable X-ray sources) without the measure theory to tackle their description of the Borel sigma-algebra over paths? Or to phrase the description of the astronomers’ ‘path-integral evidence’ approach (Kitching & Taylor 2015) in a manner readily parsed by statisticians?
My vision is for the establishment of an ‘astro-statistics’ program (used here in the best sense to mean a genuine fusion between the astronomy and statistics departments) of teaching and research, in which astronomical problems motivate new statistical inquiries and new statistical methodologies reveal new possibilities for astronomical study.