Missing crime records could skew predictive policing against Black residents in Chicago

When crime records go missing, the damage may not fall evenly. A new modelling study of Chicago argues that gaps in recorded crime data can propagate through predictive policing systems in ways that systematically shift which ethnic groups are exposed to predicted over-policing and under-policing — and that the direction of the shift depends on where in the city the data disappear.

The study, published in the journal Crime Science, was conducted by Mengqin K. Jiang, Federico Liberatore and Alun Preece, who set out to quantify something previous fairness research had largely left unmeasured: the impact of event-level data incompleteness in spatio-temporal crime prediction on each ethnic group. Using Chicago crime records from 2022 to 2024, census-derived demographic data, and a state-of-the-art predictive policing model called HAGEN, the researchers deliberately removed crime records at varying rates and tracked how the resulting prediction errors redistributed across ethnic communities.

Predictive policing uses historical crime records to forecast where and when future events may occur, allowing police agencies to deploy resources accordingly. The authors focused on event-level spatio-temporal prediction — forecasting specific crime occurrences in space and time — rather than hotspot ranking, which only flags high-risk areas. A key problem in practice is that crime records are often incomplete: events go missing because of under-reporting, uneven detection, and administrative recording choices, and this missingness is unlikely to be random, varying across neighbourhoods and ethnic groups.

To measure the consequences, the team introduced a neighbourhood-level method that allocates prediction errors to ethnic groups in proportion to each group’s share of a neighbourhood’s population. False positives — crimes predicted but not recorded — are treated as potential over-policing, while false negatives — recorded crimes the model misses — are treated as potential under-policing. The authors are careful to stress that these are model-based proxies of policing exposure driven by prediction errors, not measurements of actual police behaviour.

The experiments were structured as a controlled stress test rather than a simulation of real-world missingness. The researchers trained HAGEN on the original Chicago data, using January 2022 through May 2024 for training, June through August 2024 for validation, and September through December 2024 as separate test periods. They then removed records from two kinds of communities: nine neighbourhoods with low racial diversity, measured by Simpson’s Diversity Index, where reductions of 25, 50, 75 and 100 percent were applied; and communities dominated by a single ethnicity — Hispanic, Non-Hispanic Black, Non-Hispanic White, or Non-Hispanic Asian — where 50 and 100 percent reductions were applied. Within targeted communities, removal was by uniform random sampling, so the only non-random element of the design was the choice of which communities to target.

Across eight crime types — theft, criminal damage, narcotics, robbery, assault, deceptive practices, burglary and battery — a consistent pattern emerged. When records were removed from low-diversity neighbourhoods, greater missingness amplified prediction bias against Non-Hispanic Black residents specifically: predicted over-policing fell while predicted under-policing rose, with the opposite trend holding for other groups. Across all eight crime types, data incompleteness in these communities produced the most negative policing outcomes for Black residents.

Removing records from a single-ethnicity-majority community generally reduced predicted over-policing and increased predicted under-policing for that ethnicity across the whole study area, while reversing these effects for other groups. One consistent exception was the Asian-major scenario, where changes across all groups remained small. The largest overall shifts occurred when data were removed from Black-major communities: across narcotics, criminal damage, theft, robbery, deceptive practices and assault, enforcement bias for Black residents changed the most under that scenario. By contrast, residents classified as Non-Hispanic Other or Multiple Races consistently showed the smallest changes, typically below 1.0 percent.

Crime type also mattered. Battery stood out as the clearest outlier: removing battery records from Hispanic-major and White-major communities produced large changes in predicted over-policing for Black and Hispanic residents — groups that did not dominate those communities. This underscores one of the study’s more striking implications: missing data in one demographic context can amplify disparities for entirely different groups, depending on the type of crime involved.

The authors frame their results as diagnostic rather than definitive. They note that the Chicago records used as the baseline are themselves known to be incomplete and should not be taken as ground truth, so all reported changes are shifts relative to an already incomplete starting point — and likely understate the total distortion affecting a deployed system. They also caution that their controlled missingness does not model real missingness mechanisms, which vary with crime type, victim–offender relationship and willingness to report, and that the direction of induced bias depends on the missingness pattern. The scenarios are directional evidence on how predictive policing outputs react to incompleteness, not estimates of the disparities produced by real-world missingness. The study is further limited to one city, one period, a fixed set of crime types and one model, and the chronological train-test split spans seasons, so results might differ if tested on other months.

Despite those caveats, the conclusions carry practical weight. The authors recommend that before any operational use, predictive policing systems should include systematic audits of missingness and of its demographic distribution, and should report over- and under-prediction disparities by ethnic group, since these are the quantities that translate into potential policing exposure. When data quality is uncertain, they argue, predictions should be treated as sensitive to missingness and the data collection and recording process should be documented. In a field where the inputs are quietly shaped by who reports crime, who gets detected, and who gets recorded, the study suggests that the first fairness question a predictive policing system must answer is not about its algorithm — but about what its data never contained.

Subject of Research: How incomplete crime data shapes ethnic disparities in predictive policing predictions in Chicago

Article Title: How crime data incompleteness shapes ethnic disparities in predictive policing: evidence from Chicago

Article References: Jiang, M. K., Liberatore, F., & Preece, A. (2026). How crime data incompleteness shapes ethnic disparities in predictive policing: evidence from Chicago. Crime Science, 15(1), Article 40. https://doi.org/10.1186/s40163-026-00306-9

Image Credits: AI Generated

DOI: 10.1186/s40163-026-00306-9

Keywords: predictive policing, crime data, missing data, algorithmic bias, Chicago, ethnic disparities, over-policing, under-policing, HAGEN, machine learning, data auditing, false positives