With a little help from the crowd: Estimating election fraud with forensic methods

Election forensics are a widespread tool for diagnosing electoral manipulation out of statistical anomalies in publicly available election micro-data. Yet, in spite of their popularity, they are only rarely used to measure and compare variation in election fraud at the sub-national level. The typical challenges faced by researchers are the wide range of forensic indicators to choose from, the potential variation in manipulation methods across time and space and the difficulty in creating a measure of fraud intensity that is comparable across geographic units and elections. This paper outlines a procedure to overcome these issues by making use of directly observed instances of fraud and machine learning methods. I demonstrate the performance of this procedure for the case of post-2000 Russia and discuss advantages and pitfalls. The resulting estimates of fraud intensity are closely in line with quantitative and qualitative secondary data at the cross-sectional and time-series level.