Election Forensics: Statistics, Recounts and Fraud∗

I use data from the 2006 federal election in Mexico to compare the impression conveyed by tests based on the second digits of reported vote counts to the impression conveyed by a manual recount done for a nonrandom sample of the ballots cast for president. The patterns identified by the 2BL tests match classical ideas about how local political machines operate: in municipalities whose mayors are affiliated with two of the three major party coalitions, the party’s candidates do better in voting for president, senate and deputy than expected according to a natural voting baseline. For the three parties that were not competitive in the presidential election, the second-digit tests strongly suggest vote counts were affected either by massive intimidation or by widespread strategic voting. The manual recount did not detect any such patterns, and the changes produced by the recount are unrelated to the vote counts’ second digits. Second-digit tests can detect election phenomena that have nothing to do with tabulation errors