Alcohol outlets and firearm violence: a place-based case-control study using satellite imagery and machine learning

Area of Science:

• Environmental Epidemiology
• Geospatial Analysis
• Machine Learning Applications

Background:

• Observational injury studies can be confounded by the built environment.
• Alcohol outlets (AOs) are potential risk factors for firearm violence.
• Previous studies faced challenges in controlling for environmental confounders.

Purpose of the Study:

• To propose and apply a novel method for visually matching places using satellite imagery and machine learning.
• To investigate the spatial relationship between alcohol outlets and firearm violence in Philadelphia.
• To strengthen causal inference in injury research by mitigating confounding by the built environment.

Main Methods:

• Developed a place-matching technique using high-resolution satellite imagery and a pretrained convolutional neural network.
• Employed a case-control framework matching city blocks with shootings to visually similar blocks without shootings.
• Utilized logistic regression to estimate the odds ratio (OR) of firearm shootings near alcohol outlets, controlling for covariates.

Main Results:

• Visually matched case-control blocks were similar in appearance and balanced on covariates.
• Locations with beer stores within one block showed an increased shootings risk (OR=1.5, p=0.02).
• Locations with bar/restaurants on the same block had an increased shootings risk (OR=1.6, p=0.02).

Conclusions:

• The novel visual matching method improves observational injury studies by addressing built environment confounders.
• Findings provide stronger evidence for a causal effect of alcohol outlets on firearm violence.
• This approach enhances the reliability of place-based risk assessments in public health research.