Positive Empirical Models of Election Frauds

Regents Of The University Of Michigan - Ann Arbor, Ann Arbor MI

Investigators

Abstract

Election forensics describes a growing body of work devoted to using statistical methods to try to determine whether the results of an election are accurate: whether the results are the collective choice implied by citizens' intentions given the election rules. We propose to improve a model of election frauds that positively describes what a fraudulent election looks like and supports using eligible voter and vote counts to estimate the amount of fraud occurring in a particular election. Our goal is to be able to estimate the incidence and magnitudes of any frauds that occur, in a variety of election systems. We will use a Bayesian approach that provides a rigorous and explicit way to use and express evidence-based beliefs about the integrity of an election outcome. Ideally we would like to recover what the election results would have been in the absence of frauds. The principal challenge is whether it is possible to distinguish frauds from the effects of the kinds of strategic behavior that are intrinsic to political activity and are not frauds. To pin down whether the model is responding to strategies or frauds we will supplement the vote count data that are the primary basis for statistical analysis with information from postelection complaint processes and election observer reports. The public in general (and specifically election monitoring organizations) will benefit from having a reliable and flexible tool that can be used to help confirm the accuracy or inaccuracy of election results. The proposed research will improve the Election Forensics Toolkit currently being produced for the United States Agency for International Development. Software and data will be made available and will be usable independent of the Toolkit. We will generalize and improve the positive empirical model of election frauds introduced by Klimek, Yegorov, Hanel and Thurner (PNAS 2012). We will generalize the original model specification to suit a variety of election systems and to account for voting and registration suppression and spoiling ballots. We will replace the simulation methodology Klimek, et al. (2012) originally proposed with statistical estimation using Bayesian Markov Chain Monte Carlo methods (an Expectation-Maximization algorithm may also work in simple cases). The statistical models feature finite mixtures of irregular parametric distributions with, in general, an unknown number of components, hence methods such as reversible jump MCMC are of interest. Not only will a Bayesian approach provide sound estimates for the distributions of parameters that represent the configurations and magnitudes of frauds, but such an approach will allow the analysis of one election to learn in a systematic and principled way from the analysis of other elections: for example, posterior distributions from analysis of data from one election can be used to inform the prior distributions applied when analyzing a later election. We will study how well the models adapt to systems with various election rules, using polling station (or precinct) data from a diverse array of elections. We will examine whether the estimates respond to frauds and not to strategic behavior. We will study how the estimates relate to and perhaps can be systematically informed by auxiliary information such as officially filed postelection complaints.