The influence of AOs on interpersonal violence

Numerous cross-sectional studies have found a positive association between the density of AOs and rates of violence at the neighbourhood level, using units such as census tracts, even after controlling for other neighbourhood-level differences.5 6 These results, however, reveal relatively little about how or why AOs increase violence6 and have left persistent questions about whether AOs increase risk independent of other neighbourhood-level factors.7 8 Other study designs, examining the effects of changes in alcohol availability5 and the individual-level risk of violent victimisation from spending time in higher-AO locations,9–11 have yielded mixed results.

Another common approach has been to examine whether violent incidents cluster around AOs. Such an approach has deep roots in spatial epidemiology: Snow found the source of an 1854 London cholera outbreak by mapping the precise locations of cholera cases and showing that they clustered around a particular drinking water source.12 Similarly, a number of criminologists have examined the spatial influence of AOs (see ref 13 for a review.) These studies have typically found that violence is higher in the immediate vicinity of AOs and up to a few blocks away, after which no effect is observed, bolstering the conclusion that AOs play a causal role in nearby violence.6

AOs, however, may be located in places that are systematically different from other nearby places, making place a potential confounder. A few recent crime studies have accounted statistically for differences such as neighbourhood demographics and socioeconomic status13 and for block-level features (eg, street trees, vacant lots and public facilities).14 While adding block-level features aims to account for differences in the built environment, the selection of covariates is limited by data availability and relies on questionable modelling assumptions, discussed further below.

The present study extends this literature by (A) focusing on firearm violence, a form of violence that causes a large proportion of injury-related deaths, particularly among young men and boys of colour,15 and (B) making place-based comparisons more rigorous and less model dependent.

Data used

The dependent variable was fatal and non-fatal shootings in Philadelphia in 2017–2018 that did not involve police officers. For privacy reasons, shooting locations were aggregated to the 100-block level: for example, ‘123 Main Street’ would become ‘100 Block Main Street’. A small proportion (6%) coded to intersections was removed, as prior studies focusing on street segments have done.16 Philadelphia street segments were combined when they shared the same 100-level address, forming a dataset of city blocks (n=24 408). Shootings were joined to city blocks based on addresses, with a match rate of 94%. For this analysis, shooting outcomes on a given block were treated as binary, with 7% of blocks experiencing at least one shooting in 2017–2018. Repeating the analysis with shootings as a count variable did not substantially alter the results.

The analysis tested on-premises AOs (‘bar/restaurants’) (n=1723) and off-premises AOs that sell beer (‘beer stores”) (n=272) based on alcohol licencing data. In Pennsylvania, only state-run outlets may sell hard liquor for off-site consumption. Correspondingly, Philadelphia has many fewer liquor stores than similar-sized cities in other states, and real or perceived increases in guardianship at these outlets due to their official designation may influence nearby violence rates.11 To focus the analysis on locations where results were more readily interpretable, state-run liquor stores were omitted.

To assign each AO to a corresponding city block, Thiessen polygons were calculated using block midpoint locations, following Wheeler.14 For each block midpoint, the corresponding Thiessen polygon delineated the space that was closer to that midpoint than any other. For city blocks that comprised multiple street segments, the corresponding polygon combined multiple Thiessen polygons, generated from the midpoint of each constituent street segment and the midpoint of the combined unit. This approach was intended to prioritise the influence of large blocks over the smaller side streets whose midpoint might have been closer to a shared intersection. Each AO was assigned to the block whose corresponding polygon contained the AO coordinates.

Block-level exposure to nearby AOs was measured as binary variables based on whether one or more AO was located on an adjacent block (‘within one block’) or on a block adjacent to those blocks (‘within two blocks’) based on Thiessen polygon contiguity. This measure was used by Wheeler14 and was more appropriate than a metric based on fixed distance units, since shooting location was only known at the block level. Following Wheeler,14 polynomial splines were used to account for the disproportionately sized polygons assigned to blocks that lacked nearby neighbours. These adjustments for polygon size were included in every regression model; for readability, these terms are not included in the results below.

Four types of covariates were included as potential confounders: sociodemographic characteristics, built environment, drug-related activity and land use. Sociodemographic characteristics were obtained from the 2016 American Community Survey. Following Branas et al,9 these variables were the population per square mile of black, Hispanic and unemployed residents, as well as the total area income in USD, estimated using inverse distance weighting with a maximum bandwidth of 2 miles. Built environment variables were vacant lots and street trees, which multiple studies have found to be independently associated with place-based firearm violence risk, including specifically in Philadelphia.17–20 Drug-related activity, which Lipton et al 8 adopted as a control variable,8 was represented as the density of reported incidents related to illegal drugs, both for distribution and possession, calculated using kernel density estimation (bandwidth=1000 feet) from 2012 to 2016 police data. To separate the effects of AOs from other establishments with high foot traffic, we included commercial land use, that is, the number of parcels zoned for commercial use within the Thiessen polygon, following Morrison et al.11

For each city block, an aerial image was downloaded using the Google Maps Application Programmer Interface (API). Images were centred on the block midpoint and displayed a ground surface area of approximately 886 by 886 feet. This resolution was sufficient to show the full length of almost every block, since 886 feet was approximately the 90th percentile for block length. Typically, a few neighbouring blocks were visible as well. Most images were taken in mid-2017.

Matching cases and controls on visual appearance

Each aerial image was first converted to a simpler numeric representation using a pretrained convolutional neural network (CNN) trained on the ImageNet dataset. This representation, a 256-long vector, was extracted from the third convolutional layer of the VGG16 algorithm.21 Each value in this vector corresponds to an abstract attribute, consisting of colours, edges and textures, that is generically relevant to image interpretation. These values function like latent variables describing the physical environment, effectively distinguishing among scenes, but do not correspond directly to real-world features such as the number of trees or buildings observed in the scene. In prior studies, researchers have used CNN-derived features from high-resolution satellite imagery to model the built environment’s association with wealth22 and obesity.23

Next, these 256 latent variables were reduced to two latent dimensions using t-distributed stochastic neighbour embedding (t-SNE). Like principal component analysis (PCA), t-SNE is used to summarise a complex dataset in lower-dimensional space.24 Whereas PCA transforms observations according to the dimensions of maximum variance, t-SNE uses machine learning techniques to situate similar observations near to each other, making it a more suitable technique for this analysis of visual similarity.

Finally, one matched control location was selected for every case location (n=1409) based on proximity in the t-SNE-generated two-dimensional space. The nearest neighbour was assigned as a match, without replacement, with locations removed from the potential matching pool as soon as they were assigned to a case (‘greedy’ matching). There were two additional conditions imposed on these matches. First, controls were only eligible if they were located at least 0.5 mi away from the case site to avoid potential contamination. Second, controls were required to belong to the same roadway type as the index case block. This step was included because preliminary testing showed that the matching algorithm would sometimes select a control that was a suitable match for a block depicted on the margins of the image, but not the index case block, which was centred in the image.

For this study, no additional covariates were used for matching, making it possible to assess the level of balance on typical control variables that was achieved using imagery alone. Compared with the full pool of potential matches, matching on imagery was expected to improve case–control balance substantially for the visible environmental variables (street trees, vacant lots and commercial land use) and for physical infrastructure characteristics that were not otherwise measured.

Estimating AO effects on shootings

After matching, the study dataset consisted of observations from street segments where a 2017–2018 shooting occurred (n=1609) and from matched street segments where no shooting occurred (n=1609). Following the conventions of case–control analysis, logistic regression was used to estimate the ORs associated with the AO exposures: a bar/restaurant or beer store: (A) on the same block, (B) within one block and (C) within two blocks. For rare outcomes such as shootings, ORs are commonly used to measure relative risk.25

The simplest model contained only the binary shooting outcome (dependent variable) and the AO variables (independent variables). Additional models were specified with different combinations of the potential confounders discussed above. To enable direct comparisons of regression coefficients for both continuous and binary inputs, all continuous inputs were mean centred and scaled by two SD prior to modelling.26

These models all displayed strong spatial dependence in the deviance residuals. In other words, the observations violated the assumption of independence and identical distribution with respect to underlying spatial trends in firearm violence. The next step was to address this problem. A thin-plate spline was fit to each model’s residuals then added back to the original model as a covariate, adapting from Goldstick et al.27 The corrected models, which are reported here, displayed no residual spatial dependence in Moran’s I tests.

The convolutional neural network was implemented in Python using the Keras API. t-SNE was implemented in Python using the scikit-learn package. Sample Python code to conduct the image analysis is included in the online supplementary data. All subsequent analyses were conducted in R. Institutional review board review was waived as non-human subjects research.