Post-Election Audit Efforts in Iowa Successes and Challenges Luke - PowerPoint PPT Presentation

Post-Election Audit Efforts in Iowa Successes and Challenges Luke Fostvedt* Iowa State University Survey Workgroup September 15, 2009 *Jon Hobbs did most of the work 1 / 25 Statistics in the Community

Outline Background Iowa Legislation Methodologies Horizon 2 / 25 Statistics in the Community

ASA Policy • 2008 ASA Board of Directors endorsements • March position on electoral integrity It is critical that the integrity of central vote tabulations be confirmed by audits of voter-verified hard-copy records in order to provide high – and clearly specified – levels of confidence in electoral outcomes. • September endorsement of election auditing principles 3 / 25 Statistics in the Community

Transparency • Ohio Joint Audit Working Group definition Transparency entails that the public should have the opportunity to observe the audit and to ensure that all phases have been conducted correctly. . . Everyone should understand what the procedure requires and why, with little room or need for subjective interpretation during the audit. • How is this interpreted? 4 / 25 Statistics in the Community

Audit Terminology • True result is a full hand recount • Risk-limiting audits reduce the risk of confirming an incorrect outcome • Risk - probability of certifying a result different than what a full recount would reveal • Methodologies vary in their efficiency 5 / 25 Statistics in the Community

IA Bill • HF682 introduced in 2009 • Based on a Minnesota law implemented in 2008 • Passed House 98-0 on March 24, 2009 • Did not leave Senate State Government Committee • Plans to introduce in 2010 6 / 25 Statistics in the Community

A Look at HF682 • Counties select precincts for audit by lot • “Tiered” audit protocol • One precinct if county has 7 or fewer precincts • Two precincts if county has 50,000 or fewer registered voters • Three precincts if county has 50,001-100,000 registered voters • Four precincts if county has over 100,000 registered voters • President and governor always audited • One additional race randomly selected 7 / 25 Statistics in the Community

A Look at HF682 • No computerized randomization • Escalation mandated when hand count reveals a discrepancy of at least 0.5% • Additional two precincts selected in second round • State commissioner of elections may mandate further escalation • Precinct requirements are minimums 8 / 25 Statistics in the Community

Registered Voters Iowa Registered Voters 7.7 5.6 8.5 4.7 13.4 7.6 7.3 6.4 10.3 14.8 12.1 19.7 10.5 12.4 6.9 8.3 32.2 9.2 11.4 12.2 11.8 7 18.2 12.4 5.7 9.4 7.5 10 17.5 8.8 64.4 87.3 14.1 12.4 26.4 11.1 12.2 9 62.2 5.3 8.1 7.3 Voters (1000) 14.9 50 13.9 12.2 18.2 145.9 64.3 27.3 100 6.8 10.2 14.8 7 19.1 33.1 150 12.7 200 13.9 12 26.5 9.3 9.5 4.7 8 43 279.5 100.5 118.9 250 28.6 7.6 14.8 31.9 24.3 15.2 61.4 10.7 5.6 11.3 7.5 23.6 12.9 3.3 8.7 6.2 6.3 5.3 14 10.5 7.8 29.1 5.2 9.5 5.3 6 10.8 4.5 3.2 6 4.1 23.7 9 / 25 Statistics in the Community

Assessing Iowa Bill Precincts Audited Under Current Bill Audit 1 2 3 4 10 / 25 Statistics in the Community

Assessing Iowa Bill Proportion Sampled Prop Audited 0 0.05 0.1 0.15 0.2 0.25 11 / 25 Statistics in the Community

StatCom Team Analysis • ISU StatCom team assessing proposed methodology • Using 2006 Iowa election data • Actual risk depends on apparent margin of victory • Method seems inefficient for large margins • Risk can be high for close races • Handling varying precinct sizes 12 / 25 Statistics in the Community

Precinct Sizes Precinct Sizes by Congressional District 1 2 3 80 60 40 20 0 count 4 5 80 60 40 20 0 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 0 500 1000 1500 2000 2500 Precinct Size 13 / 25 Statistics in the Community

How Good is the Tiered Method • when the apparent margin of victory is 0.5% but the outcome of the election was wrong, the method only detected a miscount around 80% of the time. • The loser was confirmed the winner 20% of the time 14 / 25 Statistics in the Community

Benefits of Risk-Limiting Procedures • Based on Power and Margin of Victory • X = number of miscounted precincts in sample • Power = P( X > 0 | B min miscounted precincts) • B min = minimum number of miscounted precincts to overturn election • Power set at 99% • Efficient • Samples as few Precincts as necessary 15 / 25 Statistics in the Community

Sample Randomly McCarthy et. al. 2008 • Method 1: Randomly Sample Precincts • Based on Margin of Victory and Desired Power • Assumes equal precinct sizes • Uses a Hypergeometric Distribution to classify miscounts 16 / 25 Statistics in the Community

Sample Randomly McCarthy et.al. 2008 • Within Precinct Miscount (WPM) is somewhat controversial (Stark 2009) • sets a maximum of a 40-pt shift in the percentage margin within that precinct (it seems rather arbitrary) �� �� m • B min = N · 2 WPM 17 / 25 Statistics in the Community

Weight Precincts by Size Aslam & Aslam 2007 • Method 2: Sample Proportional to Size • There is an ”adversary” who wants to tamper with as few precincts as necessary • Assigns each precinct a probability of being sampled proportional to its size • Assumes tampering would happen to larger precincts • requires the use of a computer 18 / 25 Statistics in the Community

Ballot Based Auditing • Method 3: Sample Ballots • Randomly sample individual ballots • Must have a way to cross examine ballots with the results • Would voting still be completely anonymous? • Logistical nightmare to execute 19 / 25 Statistics in the Community

Problems in Iowa • Precincts Sampled at County Level • The Size and Number of Precincts Varies heavily among Counties • This seems like ”Stratifying by County” • Does it make any sense to Stratify by County? • What are possible solutions for this problem? 20 / 25 Statistics in the Community

Current Ideas 1. Aggregate precincts (from entire state) into groups of equal size • How do you aggregate the Precincts? • Minimize L = � n � n k > j ( p k − p j ) 2 j =1 2. Randomly sample from these new ”Precincts” • Ideally the precincts being sampled would be spread across the state 21 / 25 Statistics in the Community

Escalation Procedures • Given a miscount is detected, what next? • Do Nothing? • Full recount? • Statistically how should we proceed? • Suggestions from the audience? 22 / 25 Statistics in the Community

Summary • Where is balance between ”Transparency” and ”Risk”? • Logistics of a Risk-Limiting Method must be simple • must be comparable to the Tiered method 23 / 25 Statistics in the Community

Iowa Statisticians • Participation from statisticians across Iowa • Faculty • Drake University: Rahul Parsa • Iowa State University: Alicia Carriquiry, Dianne Cook, Heike Hofmann • University of Iowa: Russell Lenth • Iowa State StatCom Team • Lisa Bramer, Luke Fostvedt, Randy Griffiths, Jonathan Hobbs, Eunice Kim, Adam Pintar, David Rockoff 24 / 25 Statistics in the Community

Suggestions • Questions? • Comments? • Suggestions? 25 / 25 Statistics in the Community