Learning from Recounts

Abstract

We compare the results of two recent statewide recounts in Wisconsin-the 2011 Supreme Court election and the 2016 presidential election. Using the measure of absolute differences between the original tally and the recount, we find an error rate at the reporting unit level of 0.21% in 2011 and 0.59% in 2016. The 2016 error rate drops to 0.17% when write-in votes are removed from the analysis. We also find that paper ballots originally counted with optical scanners were counted more accurately than ballots originally counted by hand. To reach these conclusions, we address the methodology of measuring differences in election night and recounted vote tallies. The most commonly used measure to compare election-night and recounted tallies, the net difference, significantly understates the magnitude of errors in the original tally. We also develop a regression-based technique that estimates what the error rate should be if a ballot-by-ballot recount were possible. We conclude by discussing the implications for requiring post-election audits.