PolBeRG Seminar - Latent Class Approach to Electoral Fraud Detection

This Friday (at 17.30 in FT308) we will have the final seminar of the season - Juraj Medzihorsky will present his ongoing work on electoral fraud detection. The seminar will be light on the participants, this time there will be no paper, but please do note the abstract for it below:

Latent Class Approach to Electoral Fraud Detection

A group of electoral fraud detection methods focuses on digit distributions in electoral returns. Such methods assume that for fraud-free elections digit distributions are known in advance, and evaluate whether the observed digit distributions were drawn from them using null hypothesis significance testing (NHST). This paper offers two new methods for digit-based electoral fraud detection based on the π*mixture index of fit and the Δ dissimilarity index. The methods rest on the decomposition of the returns into two latent classes, the fraudulent and the non-fraudulent one. In this framework the smallest possible size of the former class is a natural measure of fraudulence. A re-analysis of several electoral datasets shows that the proposed methods lead to new substantive conclusions while requiring considerably fewer and more realistic assumptions than the extant methods. The proposed methods are applicable in any similar setting, such as digit-based forensic accounting and scientific fraud detection.