Study Title: Estimating the Impact of an Integrated CCTV Program on Crime (which is associated with Outcome 1 and Outcome 2)
Research Design
Circo and McGarrell (2021) used a quasi-experimental design with propensity score matching to assess the effectiveness of Project Green Light on crime in Detroit, Michigan after 1 year of implementation. The first phase of Project Green Light, which included 88 businesses that were operational as of December 31, 2016, was evaluated.
Sample
Businesses included gas stations, liquor stores, bars and restaurants, retail and service stores, and other commercial and community outlets. Potential businesses were identified for participation by the Detroit Police Department and the city of Detroit based on a number of factors, such as the number of police calls for service, a history of crime reports, and noted community concerns about the business. Participating businesses were mandated to meet a series of requirements to enter into the program and were subject to inspection to ensure that cameras were operational and that adequate lighting and signage were present.
Data Collection
Crime data were obtained from the Detroit Police Department’s records management system for 1 year prior to Project Green Light (2015) through the implementation period (2016) and for 1 year following implementation (2017).
Outcome Measure
The primary outcomes of interest were the number of monthly crime incident reports. Crime incidents were categorized into two outcome variables: 1) violent crimes (serious violent crime, including aggravated assault, armed and unarmed robbery, and felony homicide) and 2) disorder crimes (misdemeanor assault, possession of drugs, open liquor citations, illegal gambling, public drunkenness, and disorderly conduct). To capture only crime incidents that occurred directly in and around the business, all crimes evaluated occurred within 200 feet of each business. Constructing the final observational units for analysis involved merging crimes with businesses, which was performed in several steps: 1) circular catchment buffers of 200 feet were drawn around each business, 2) all crimes falling within a buffer were retained, while all crimes outside the buffers were eliminated, and 3) crimes were merged with the nearest business within each buffer. Information about each crime included the time, date, location, crime type, and category, and the business where or near it occurred.
Statistical Analysis
To generate the propensity score, a logistic regression was used to estimate the probability of a given business receiving the Project Green Light treatment. This was based on a set of pretreatment covariates, including the annual number of violent crimes and disorder crimes reported at the business in 2015, and census tract-level variables relevant to crime outcomes that reflected the characteristics of the neighborhood in which the business operated: the percentage of the population that included Black individuals, male individuals, unemployed individuals, households with low income, households in poverty, households renting, households with rent greater than 30 percent of their income, households receiving Supplemental Nutrition Assistance Program, and vacant housing units. Optimal full matching was used to create matched sets with one treated unit and one or more control units that minimized the global distance between all matched pairs. Treated Project Green Light businesses could be matched with up to 5 untreated control businesses. After matching, there were 87 treated businesses and 201 untreated control businesses. There were no statistically significant differences between the treated and untreated businesses on any of the pretreatment covariates or outcomes of interest at baseline.
A hierarchical growth-curve modeling strategy was used to model the effect of the Project Green Light intervention on crime from the preintervention period to 1-year postimplementation. Specifically, separate Bayesian negative-binomial growth-curve models were fit, one for each crime type (disorder occurrences and violent crime). The estimates for the effect of Project Green Light were adjusted for residual pretreatment covariate differences, monthly seasonal effects (relative to January), and mean differences in crime reporting by business types (relative to gas stations).
Subgroup analysis was conducted by the number of calls for service at treated businesses, compared with matched untreated control businesses.
Citation:
Circo, Giovanni, and Edmund F. McGarrell. 2021. “Estimating the Impact of an Integrated CCTV Program on Crime.” Journal of Experimental Criminology 17:129–50.
Study Title: Assessing Causal Effects Under Treatment Heterogeneity: An Evaluation of a CCTV Program in Detroit. (which is associated with Outcome 1 and Outcome 2)
Research Design
Circo and colleagues (2023) used a quasi-experimental design with a difference-in-difference model to assess the effect of Project Green Light Detroit on crime at individual city parcel addresses in Detroit, Michigan.
Sample
The intervention group (n = 560) included all city parcels enrolled in the intervention between January 2017 and December 2019. Intervention addresses were divided into 11 distinct group-time categories based on the year-quarter in which they enrolled during the 2017 to 2019 intervention period. These intervention addresses included 318 retail establishments, 155 service-based businesses, and 87 residential properties. The comparison group (n = 1,172) consisted of addresses not currently enrolled in or never enrolled in the intervention that had reported at least one crime between 2017 and 2019 as determined by the Detroit Open Data parcel database. Vacant addresses were omitted from the study.
To establish equivalency, the covariates collected included the parcel size, the premise type, and block-level census variables (i.e., the block-level population, the proportion of households receiving Supplemental Nutrition Assistance Program benefits, the proportion in poverty, the proportion renting, the proportion of residential properties vacant, and the proportion of the residents who identified as Black and Hispanic). On average, the intervention and comparison city parcel addresses were similar; however, the intervention group included larger mean parcel sizes than comparison parcel sizes and had a larger proportion of residential properties enrolled in the later cohorts.
Data Collection/Outcome Measures
Crime data were obtained from the Detroit Police Department’s record management system, which contained information on crime incident types, addresses, dates, and geographic coordinates. Owing to a switch in providers of the police department’s record management system, historical data were only available on crimes from December 2016 onward.
Statistical Analysis
A “did” package in the R statistical software was applied to allow the fitting of models able to estimate treatment effects that occur over multiple time periods or have the presence of treatment heterogeneity. Using this method, 2x2 comparisons were made between intervention parcels and comparison parcels for each time period, adjusting for differential group sizes and varying treatment effects. A “doubly robust” estimation procedure was used to estimate the propensity score in the first stage and weight least-squares regression in the second stage. Treatment effects were estimated by aggregating group-time treatment effects based on the time of group assignment, then averaging a treatment effect proportional to each group’s size.
The study did not conduct subgroup analyses.
Citation:
Circo, Giovanni, Edmund F. McGarrell, June Werdlow Rogers, Julie M. Krupa, and Alaina De Biasi. 2023. “Assessing Causal Effects Under Treatment Heterogeneity: An Evaluation of a CCTV Program in Detroit.” Journal of Experimental Criminology 19:1033–1051.