Election Forensics and Strategic Buzz: Assessment Using Agent-Based Models

• In Event: Methodological Advances in the Study of Elections

Abstract

Election forensics tools are statistical models designed to help determine whether the results of an election accurately reflect the intentions of the electors. Usually estimates are based on vote and registration count data. Some methods identify anomalies in votes and turnout distributions, others use positive parametric models for what "frauds" are. Application of a wide range of methods to data from many elections strongly suggests that electors' strategic behavior can cause methods to signal frauds that do not exist. Elector behavior is strategic when electors act not merely based on their own preferences but also on their expectations about what other electors will do. Strategic behavior is of course the heart of normal politics. Electors' strategic actions on whether to vote and for whom can differ from what they would have done if acting based solely on sincere preferences, and strategic actions in or near equilibrium make different electors' actions mutually dependent. Many types of election forensics methods interpret multimodalities in election data as indicators of frauds, but both frauds and strategic behaviors can produce multiple modes. The aliased---false positive---estimates such features of normal politics can produce in election forensics tools might be called strategic buzz.

In order to start creating an infrastructure for studying such buzz, we use Agent-Based Models (ABMs) of simple-majority elections to generate simulated election data, with which we study whether and how election forensic tools are sensitive to electors’ strategic behavior. Our ABMs are designed to represent electors who decide how to act in elections based both on their own tastes and on their beliefs about what the other electors will do, and they match equilibria derived in models of wasted vote logic (Cox, 1994) and of strategic abstention (Palfrey and Rosenthal, 1983, 1985; Demichelis and Dhillon, 2010). To facilitate computing pivotal vote probabilities, the ABMs use a Poisson games (Myerson 1998, 2000) formulation of the models instead of the fixed-electorate assumptions the formal models originally used. Additional strategies of interest include bandwagon dynamics and coalitions. Simulations can be initialized with a diverse combination of parameters---e.g. different number of electors and of candidates, different percentage of electors who always vote vote or who always behave sincerely, etc. We use ABMs to impose strategic equilibrium constraints. We simulate at least three types of frauds: ballot-box stuffing, vote stealing and voter suppression. Frauds may or may not be treated as results of equilibrium actions by strategic malefactors.

We compare what happens when all electors (agents) act sincerely to what happens when some or all electors act strategically. We apply several election forensics methods for detecting or measuring frauds in election results to simulated clean and manipulated election results. We investigate how well different electoral forensic tools can differentiate between varying levels of frauds and varying levels of strategic behavior, exploring which combinations of parameters and characteristics of our models fool each fraud detection technique. We also explore the extent to which the original distribution of sincere preferences in our simulated electorate may trigger false negatives and false positive fraud detection. The results will be improved understanding of when different tools should be mobilized and, perhaps more important, the beginnings of infrastructure to expand the range of election rules and political conditions that can be studied.

Authors

• Walter R. Mebane, University of Michigan, Ann Arbor

• Samuel Baltz, Massachusetts Institute of Technology

• Fabricio Vasselai, University of Michigan