Estimating the Dark Figure of Crime Using Bayesian Additive Regression Trees Plus Poststratification (BARP)

Isabel Laterzo (University of North Carolina, Chapel Hill)

Abstract: Studies of both crime victimization and violence often suffer from demonstrably unreliable crime figures. Consequently, researchers typically use homicide rates as an indicator to reflect all types of violence, despite this figure’s biases. The problem this research community faces is what many refer to as the “dark figure,” or the difference in official crimes rates and true victimization levels. This project seeks to explore the presence of the dark figure in the context of Latin America, a region known to suffer from its presence. Using the case of Chile, I demonstrate a novel way to approximate the dark figure. I create a measurement of the dark figure at a subnational level that has not previously been estimated before: that of the comuna, a community-level administrative division used across Chile. Using the new BARP methodology (Bayesian Additive Regression Trees plus Poststratification) developed by Bisbee (2019) I successfully estimate the dark figure for 92 comunas, or the majority of urban communities across the country. This constitutes about 60% of the country’s total population. The BARP methodology allows for the estimation at this level as it combines both multilevel regression and post-stratification (MRP) and Bayesian additive regression trees (BART) to create a more accurate method of extrapolation of opinion-data from surveys to geographic units below their level of representativity. I rely on two main data sources to complete this process, the 2017 round of the National Urban Survey of Citizen Security (Encuesta Nacional Urbana de Seguridad Ciudadana, ENUSC) and Chile’s 2017 census, to create predictions of most-likely responses at the comuna level. I then compare these estimates to official crime reporting rates using data from Chile’s Ministry of Interior. The underreporting rates I calculate reveal a level of variance at the comuna level that has previously not been observed. I provide an initial exploration of this variation and how it can be explained by differences in community characteristics, such as income and quantity of police officers. This new application of the BARP methodology is valuable in a handful of ways. First, it extends the method beyond its original use of modeling public opinion across subnational units and demonstrates its utility in approximating previously poorly estimated “official” statistics. Further, it contributes a new estimator that can be used in research focusing on crime globally, as victimization surveys similar to the one I utilize are administered in many country contexts. Finally, it demonstrates that further research should be conducted at the community-level to understand why individuals do or do not call the police following victimization, as analyses that rely on regional level reporting rates do not reveal the full picture.

View Poster in a New Tab