Numerical Detection of Irregular Crime Statistics in the American Context

• In Event: Poster Session: Political Methodology

Abstract

Crime is usually a primary concern among voters in local political elections. The mayor often puts pressure on the police chief, who is usually mayor-appointed, to lower crime in the city. There are conceivably incentives for the police to "juke the stats," such as underreporting crime. Anecdotes and various investigative journalism pieces have confirmed that crime statistics fraud occurs. Yet, we cannot even begin exploring the relationship between crime and politics because there is no systematic empirical method for detecting fraudulent crime data. Moreover, social science findings and policy recommendations may be misaligned if they are based on fraudulent crime data. Cleaner crime data will lead to policies that better allocate public resources to those who need it the most and will enhance research on urban crime reduction.

This project seeks to develop the first set of methodologies for numerically detecting irregularities in crime statistics. I focus on detecting two types of crime statistics fraud: underreporting crime and downgrading crimes. Detecting underreporting is to detect missing data; detecting downgrading is to detect when an incident has been misclassified. The set of methodologies developed will offer evidence that either one or both types of frauds are occurring and will estimate the prevalence of each type of fraud.

Specifically, I focus on the empirical distribution of the frequency of a specific non-violent crime category against the frequency of violent crimes (defined as homicide, criminal sexual assault, aggravated assault, aggravated battery, and robbery), where counts are on a weekly police district basis. I call these distributions crime fingerprints. Because the distribution of each type of crime is different and is often shaped by city-specific factors, it will be difficult to develop general parametric models to detect fraud. Instead, I focus on non-parametric clustering techniques. Rates across crime types are correlated positively with each other. Thus, non-fraudulent crime fingerprints should not typically exhibit multimodality, especially among similar police districts. A highly multimodal crime fingerprint could mean that the police districts are heterogeneous, that is there is significant downgrading, and/or there is significant underreporting. A hypothesized approach to detecting downgrading is to simulate crime count switching. More specifically, pick with probability p a set of problematic police districts; then, some random proportion of the non-violent crime counts in those districts, multiplied by fraud intensity factor x, are switched into violent crime counts. The simulated distribution can then be re-examined for multimodality. Running through a range of values for p and x and finding the least multimodal distribution could be an approximation of the rate and intensity of crime downgrading. Theoretically, this new simulated distribution is free from downgrading. Therefore, numerical empirical distributions, such as comparing the second-digits to the second-digit Benford Distribution, can be used to detect further deviations that could be evidence of underreporting. Adding in a temporal dimension, such as taking into account audits of certain police districts, can help explain anomalies in crime fingerprints. An alternative approach, currently being developed, is to use the framework of intention-to-treat analysis to analyze the level of false crime statistics. In this alternative approach, instances of crime that are not reported are analogous to participants dropping out of the experiment, and instances of crimes that were supposed to be switched but were then reported truthfully are analogous to treatment participants not taking the treatment.

My current focus is on the city of Chicago from 2011 to 2015, but there are plans to extend this analysis to other cities such as Los Angeles, New York, and Philadelphia. Currently, the analysis with Chicago crime data shows that lower counts of violent crimes are associated with both low and high counts of larcenies; the high counts evaporate when violent crime counts increase. This pattern could have been created if violent incidents, such as robberies, were categorized as larcenies, which is a non-violent crime category. When I break down the larceny counts by the race-majority of the districts, I find that the bimodal pattern is driven primarily by White-majority districts, which has two clusters; Black-majority districts and Hispanic-majority districts exhibit only a single cluster. There is no single White-majority district that drives this bimodality. Preliminary statistical simulations that randomly recategorize some larceny cases into violent crime cases have shown that the data can be made to look like other non-violent crime fingerprints. Furthermore, I find that the second-digits of aggravated assault counts deviate significantly from the second-digit Benford Distribution starting in 2011.

Author

• Patrick Y. Wu, University of Michigan, Ann Arbor