September 07, 2010

36-835, Paper of the Week

As threatened, I'll post links to the paper being discussed each week in the statistical modeling seminar. This will happen after the discussion, and my own brief comments here will also not be shared with the students beforehand. This should be an RSS feed for this page.

  1. Leo Breiman, "Statistical Modeling: The Two Cultures", Statistical Science 16 (2001): 199--231
    Comment: I remember being very excited by this paper when it came out. The students were less taken with it — "Of course you use cross-validation to check predictive performance, why does he feel like he has to say that?" In retrospect, I would say that what Breiman calls "data models" are very rarely serious scientific models of the data-generating mechanism, but more "algorithmic models" of a pre-computer age...
  2. Sarat C. Dass and Mingfei Li, "Hierarchical mixture models for assessing fingerprint individuality", Annals of Applied Statistics 3 (2009): 1448--1466
    Comment: This is interesting, but the big problem is that they did absolutely nothing to convince me that their model works. (Cf.) Consequently, why should I think that their estimates of false-identification probabilities are even roughly right? (Also, why not model a spatial point process as a spatial point process?)

Corrupting the Young; Enigmas of Chance

Posted at September 07, 2010 16:49 | permanent link

Three-Toed Sloth