Blue Data Lab

Welcome! This is the basic workspace for the Blue Data Lab. We study police violence, surveillance, and tech. Here is an incomplete list of open potential projects. Below is a list of ongoing or completed projects:

1. Causality

  1. Backtracking counterfactual fairness (with S. Madhu and S. Babel)

    Building off of recent developments formalizing backtracking counterfactuals, we introduce the notion of backtracking counterfactual fairness extending existing frameworks of interventional counterfactual fairness. We benchmark this new fairness measure by evaluating it on the COMPAS dataset.

  2. How to Use Causal Inference to Study Use of Force. (J. Bourgeois, A. Haensch, S. Kher, D. Knox, G. Lanzalotto, T.A. Wong)
    CHANCE 37 (2024), no. 4. journal

  3. We discuss the role causal inference can play in the study of police use of force.

2. Social contagion of police misconduct

This project studies the spread of police misconduct as simplicial contagion on Chicago police department misconduct data. More to come!

3. Shotspotter in Detroit

Dashboard and Github on Detroit's ShotSpotter. (Also preliminary study website)

  1. Analysis of a gunshot detection system: Null effects of ShotSpotter in Detroit code (D. Kinsman, H. Chaaban, M. Majumder, D. Ramjee, A. Koumpias, T.A. Wong)

    ShotSpotter, owned by SoundThinking Inc., is the most widely used automated acoustic gunshot detection system in the United States with contracts in 85 cities, including Chicago, IL, Columbus, OH, and Oakland, CA. We evaluate for the first time the impact of ShotSpotter technology in Detroit, Michigan, on shooting-related 911 calls for service and non-fatal shooting and homicide arrests. We combine ShotSpotter geolocation information with data on 911 calls reporting gunshots as an indicator of gunshot incidents from 2018 to 2022 in weekly frequency. We similarly combine ShotSpotter geolocation data with non-fatal shooting and homicide arrest records from the Detroit Police Department's record management system as indicators of gunshot and violent crimes leading to arrests on a weekly and monthly frequency, respectively. Using a difference-in-difference (DiD) design, we estimate the effect of ShotSpotter by comparing scout car areas (SCAs)--subdivisions of precincts to which Detroit Police Department officers can be assigned for patrol--where ShotSpotter was deployed to SCAs with no ShotSpotter deployment before and after its implementation in 2021. To confirm the validity of the DiD design, we conduct event-study falsification tests that indicate the absence of any pre-treatment non-parallel trends in the evolution of reported gunshot incidents in SCAs with and without ShotSpotter, thereby, permitting a causal interpretation of our study estimates. We determined that implementation of ShotSpotter technology generated 0.97 (SE=0.26) fewer gunshot-related alerts relative to non-ShotSpotter SCAs, which is equivalent to a 42.1% decrease relative to the pre-ShotSpotter baseline number of 2.30 weekly gunshot-related 911 calls. Placebo tests of gunshot-related 911 calls illustrate no divergent trends between SCAs where ShotSpotter was deployed or not, further supporting the reliability of our study estimates. Importantly, our findings indicate that the deployment of ShotSpotter had no measurable effect with the rates of non-fatal shootings and homicide arrests, suggesting its limited influence in preventing these specific types of violent crimes. Viewing this as an algorithmic impact assessment of the downstream effects of this surveillance technology, our analysis suggests that while ShotSpotter is caused a decrease in the number of civilian reports of gunshot incidents to the Detroit Police Department, the system has no measureable effect on the number of gun-violence and homicide arrests made by the Detroit Police Department.

  2. Meta-analysis of acoustic gunshot detections system evaluations in the USA (with D. Caro)

    More soon!

4. Fatal Police Violence

This project studies the the Mapping Police Violence database. Outputs include:

  1. The homological persistence of police violence: analysis and limitations (with D. Kinsman) preprint code
    PLOS One (to appear)

    We use topological data analysis, namely persistent homology and persistence landscapes, to study trends in fatal police violence in the US between 2016 and 2022. The results of our analysis do not reveal any novel trends or characterizations about our data, but we are able to determine clusters where police violence occurs and de- termine the stability of these clusters using persistence landscapes. In other words, our findings on the homological persistence of police violence confirms general expectations on where police violence is concentrated. We also discuss the inherent methodological limitations of analyzing police data, also called Blue Data, and consequently the need from broader analyses involving other data sets in order to account from the full scale of the policing apparatus and the consideration of critical questions surrounding the function of policing and carcerality at large.

  2. A Mixed Methods Assessment of Off-Duty Police Shootings in a Media-Curated Database (E. Asabor, E. Lett, B. Mosely, C. Boone, S. Sundaresan, T.A. Wong, M. Majumder)
    Health Services Research. (2023) journal

    The aim of this study was to examine rates of killings perpetrated by off-duty police and news coverage of those killings, by victim race and gender, and to qualitatively evaluate the contexts in which those killings occur. There were threefold higher odds of news reporting of a police-perpetrated killing and the off-duty status of the officer for incidents with Black and Hispanic victims. Qualitative analysis revealed that off-duty officers intervened violently within their own social networks; their presence escalated situations; they intentionally obscured information about their lethal violence; they intervened while impaired; their victims were often in crisis; and their intervention posed harm and potential secondary traumatization to witnesses. We conclude that police perpetrate lethal violence while off duty, compromising public health and safety. Additionally, off-duty police-perpetrated killings are reported differentially by the news media depending on the race of the victim.

5. Predictive Policing

Bibliography of predictive policing.

  1. Predictive policing: A mathematical primer (with T. McKenzie and J. Johnson)
    Notices of the American Mathematical Society 71 (2024), no.7, 929–937. journal

    There are many surveys and explainers on predictive policing from various points of view, such as surveillance technology, algorithmic fairness, and criminal justice, but few specifically address its mathematical aspects. The purpose of this article is to present a selective overview of the mathematics underlying the model used by the company PredPol, followed by critiques and further developments of the model. We focus on PredPol as it is a well-known method of predictive policing and the methodology is arguably less oblique than other predictive policing algorithms that rely more heavily on black-boxed machine learning algorithms.

  2. Proactive policing as reinforcement learning (with D. Kinsman)
    ICLR Tiny Papers (2023) OpenReview

    Recent analyses of predictive policing have shown the inherent biases in such systems. We show that the models considered in fact apply to proactive policing in general, which can be also viewed as a reinforcement learning system, and thus may also lead to over-policing.