Blog 10.29.14
Incorporating Bretschneider et al
Dr. Stein’s comments, 10.27.14:
This is an excellent start and I only suggest looking the Bretschneider et al article on the owlspace reading tab. Stark’s risk-limiting post-election audit is a simpler but blunter version of the same goal that Mebane/Benford and Kilmek are trying to achieve. Actually I think both approaches or rather all three might be appropriate. The question is where to get good data for testing the efficacy and correlates of each approach. My suggestion is to use precinct level data within Texas counties. These data should be available at the Legislature Research Council’s website. Choose a sample of counties i.e., by size, diversity and partisan competitiveness (Harris definitely should be included). The limitation of a single state study is that you might not get a whole lot of variation on your independent variables, e.g., method of voting (early, election day, mail), devices (paper, electronic) and polling place location, poll workers etc. But there should be enough variation for testing your hypotheses.
Summary: “Risk-Limiting Post-Election Audits,” Bretschneider et al 2012:
Bretschneider et al. propose a very simple thing: we actually look critically at the election results that are often certified. One anecdote was particularly startling: two city council races were exposed as certified with the wrong outcomes in a “routine audit” conducted in 2012 (Bretschneider et al. 6).
Recognizing that certification and auditing can range from very simple to very complicated, they advance the notion of a “risk-limiting audit,” which is basically a first-step detector of whether the outcome of an election was right or wrong. If there’s reason to believe the outcome is wrong--after all, the only charge of this audit is to determine whether the winner is correct--then a full-blown hand count will be mandated. The added benefit of these audits are an opportunity to not only confirm results, but explore the relative performance of each type of voting method.
There is strong reason to believe that fatigue is possible in hand-counting ballots and that machines can produce errors because of input problems, tampering, or other system failure. The proposal of auditing combines machine counting and hand-counting, which could produce more accurate and reliable results.
The important thing to keep in mind is the “risk-limiting” nature of this proposed audit. Sometimes the threshold for audits is too low to expose evidence of possible error. Sometimes the threshold is excessively high.
One note: systems that do not “produce voter-verifiable paper records, such as paperless touch-screen voting systems, cannot be audited this way” (Stark 1).
The risk limit is not the chance that an election is wrong, but rather the highest possible chance that the audit will not correct a wrong election result (Stark 1).
Remaining questions, Bretschneider et al:
Response: What kind of data do we have from the Texas Legislative Council?
The Texas Legislative Council has very specific election returns from every general election conducted in the state from 1996-2012. Another potential area for study we see is the party-specific runoffs from 2002 onwards. Every cycle has at least one party that had a runoff in that time frame, including some smaller elections. Studying smaller election returns in highly competitive races might allow us to read into voting irregularities across time, space, and level of contest competitiveness/clout. For example, the 2012 election results on the website include the Democratic Primary results for a State Board of Elections race. There might be a higher likelihood of fraud or irregularity in a down-ballot race that is both highly contested and not closely watched.
The results on the website includes VTD, which are “voting tabulation districts,” units that closely mimic county election precincts. As the website describes, however, there is a slight difference: “Created for the purpose of relating 2010 Census population data to election precinct data...VTDs can differ from actual election precincts because precincts do not always follow census geography. The VTDs in the redistricting database closely correspond to the 8,834 precincts in effect for the 2012 general election.” Prudently, all data reported on the website is adjusted for 2012 VTDs, making our analysis much easier.
Resources: