A Logistic-Binomial Mixture Model for Fraud Detection in Elections


This work presents a Bayesian finite mixture model of logistic regression models to estimate the probability of fraud in elections. The features of the electorate and polling places are built into the structure of the model, which allows researchers to estimate of voters’ support for the winning party using election data even if fraud had occurred. Simultaneously, the model estimates fraud probabilities. The inclusion of the voters’ covariates eliminates, under certain conditions, misleading estimates of fraud probabilities due to the spatial concentration of votes with similar characteristics. The model is implemented in a Bayesian fashion, and MCMC samples are used for inference. A Monte Carlo exercise is provided to evaluate the model properties. I use the model to estimate fraud probabilities in the second round of the 2010 presidential election in Brazil. To obtain voters covariates, I adopt geostatistic models to combine census data, geolocation of pooling places, and ballot box-level election data.

Polmeth XXXIII