When interviewing impressive applicants from a stunning variety of places and background for fellows in our Data Science for Social Good program (in Warwick and Kaiserslautern) this summer, we came through the common conundrum of comparing ranks while each of us only meeting a subset of the candidates. Over a free morning, I briefly thought of the problem (while swimming) and then wrote a short R code to infer about an aggregate ranking, ρ, based on a simple model, namely a Poisson distribution on the distance between an individual’s ranking and the aggregate
a uniform distribution on the missing ranks as well as on the aggregate, and a non-informative prior on λ. Leading to a three step Gibbs sampler for the completion and the simulation of ρ and λ.
I am aware that the problem has been tackled in many different ways, including Bayesian ones (as in Deng et al., 2014) and local ones, but this was a fun exercise. Albeit we did not use any model in the end!