I blogged last year about claims that fraud in the 2009 Iranian election could be detected by studying irregularities in the distribution of terminal digits. Eric A. Brill just e-mailed me an article of his which argues against this methodology, pointing out that the provincial vote totals (the ones with the fishy last digits) agree with the sums of the county totals, which in turn agree with the sums of the district totals. In order for the provincial totals to have been made up, you’d have to change a lot of county totals too (changing the total in just one county by a believable amount presumably wouldn’t make a big enough difference in the provincial totals.) But if you add Ahmadinejad votes to a county here and a county there, the provincial total would be the sum of a bunch of human-chosen numbers, and there’s no reason to expect such a sum to have non-uniformly distributed last digits. The Beber-Scacco model requires that the culprits start with a target number at the provincial level and then carefully modify county and district level numbers to make the sums match. But why would they?