Restoring voter confidence with data science

July 16, 2018

BIDS Data Science Fellow Kellie Ottoboni develops efficient statistical methods that help reveal errors or fraud through post-election audits, providing voters with confidence that their vote was counted correctly.

ballot boxRisk-limiting audits (RLAs) offer a statistical guarantee: if a full manual tally of the paper ballots would show that the reported election outcome is wrong, an RLA has a known minimum chance of leading to a full manual tally. The risk limit is the maximum chance the audit will not, in that case, lead to a full manual tally. RLAs generally rely on random samples of ballots, but the logistics of drawing such a random sample can be difficult for contests that cross jurisdictional boundaries or contain multiple kinds of ballots that arrive at different times (e.g. in-person, mail, and provisional). We have derived a statistical method for auditing stratified random samples - samples drawn by dividing ballots into distinct groups (or strata) - and then sampling from each group independently. This approach entails auditing strata independently, then combining the results into an overall risk.

This method has an immediate application in the state of Colorado. Colorado Revised Statute 1-7-515 required the State of Colorado to conduct risk-limiting audits beginning in 2017. The first risk-limiting election audits under this statute were conducted in November, 2017; the second were conducted in July, 2018. Counties cannot audit cross-jurisdictional contests on their own: margins and risk limits apply to entire contests, not to the portion of a contest included in a county. Colorado has not yet conducted a risk-limiting auditĀ of a cross-jurisdictional contest. As of this writing, about 98% of active Colorado voters are in counties that have voting equipment that allows them to perform ballot-level comparison audits, the most efficient auditing method currently available. The remaining counties can only perform less efficient auditing methods such as ballot-polling audits, which require handling more ballots.

Colorado could simply revert to ballot-polling audits for cross-jurisdictional contests that include votes in counties with legacy equipment, but that would entail handling more ballots in the other counties. Instead, Colorado could implement the stratified RLA we have developed by dividing ballots into two distinct strata: ballots cast in counties with modern voting equipment and ballots cast in legacy counties, and applying the most efficient method available to each stratum. In simulations representative of contests that might occur in Colorado, this stratified "hybrid" audit reduced the workload by up to 90% relative to statewide ballot-polling audits.

Kellie is a graduate student in the Department of Statistics, and social impact has always been the center of her research interests. The theme of her dissertation is developing and applying hypothesis tests to real world data, and her first research paper explored gender bias in student evaluations of teaching. She has also studied the mechanics of the pseudorandom number generators and their relationship to random sampling, the basis of sound statistical results. Most recently, she has been developing statistical tests to carry out risk-limiting audits of elections, which helps give voters confidence that their ballot was counted, with lower cost and effort than a full recount. In addition to developing new statistical methods and studying their theoretical properties, Kellie writes open source software implementing nonparametric methods in R and Python.



Featured Fellows

Kellie Ottoboni

Statistics