Three systematic reviews (1, 2, 18) identify very few individual-level studies that relate alcohol outlets to assault risk. Two studies (13, 14) describe cross-sectional analyses of survey data in which physical exposure to alcohol outlets was calculated by combining the spatial locations of participants' homes with spatial locations of alcohol outlets. Using data from a sample of residents aged 18 to 65 from Los Angeles and Southern Louisiana, Theall et al. (13) calculated the count of outlets within a 1-mile buffer of respondents' homes and the distance to the nearest outlet. Experiencing violence (defined as hearing or witnessing violence or being assaulted) was positively related to off-premise outlet counts, but all cause injury risk was not. Similarly, using data from a telephone survey of residents aged ≥ 18 years in two states, Treno et al. (14) found counts of both on-premise and off-premise outlets within a 2km buffer around respondents' residences were related to all-cause injury risk.

Other studies have used similar approaches to relate alcohol outlet density to outcomes other than assault, including alcohol-related aggression (19), driving after drinking (20), alcohol consumption (e.g. 21-23), and mortality (24). For example, in a multi-level cross-sectional analysis of students from eight colleges, Weitzman et al. (25) found that on-campus residence at a college with greater proximity to alcohol outlets was associated with more problems related to students' drinking. Nevertheless, these studies of individuals may be affected by measurement error. Because people routinely move through space over time, static measures such as the number of alcohol outlets around a person's home may not reflect their actual physical exposure to outlets (26, 27).

In recent years, there have been substantial advances in methods used to assess physical exposure to neighborhood features in the areas where study participants spend time (known as “activity spaces”) (28). Using survey methods (29-35) or GPS tracking (36, 37) to approximate participants' routine movements, physical exposures are measured within polygons that capture each persons' unique activity space. This approach provides greater spatial specificity compared to measures based on home location alone (36, 37).

Two individual-level studies from our group have used a variant of activity spaces (“activity paths”) to examine relationships between assault risks and physical exposure to alcohol outlets with high spatial and temporal specificity. The first (12) was a case-control study of adults comparing the neighborhoods in which cases were assaulted with a gun to the neighborhoods in which controls were located at the same time of day. Aggregated over the course of the day, risks for gun violence were higher when participants were located in areas with more off-premise outlets. The second (15) combined interviewer prompts with live electronic data collection using a Geographic Information System (GIS) to capture the activity paths of adolescents and young adults over a single day. The study was both a case-control and a case-crossover study of non-gun and gun assaults, comparing both momentary proximity to alcohol outlets and other neighborhood features between cases and controls (case-control) and within cases (case-crossover). Aggregated over the course of the day, there was no detectable association between assault risk and momentary proximity to alcohol outlets (i.e., combining on-premise, off-premise outlets, as well as other neighborhood features in a factor scale).

The aim of this study was to assess whether and how physical exposure to alcohol outlets (time spent nearby alcohol outlets) is related to assault risk for individuals, and whether relationships differed across times of day. To do so we conducted a secondary analysis of data from our GIS-assisted case-control study (15). Guided by routine activities theory, we hypothesized that individuals' risk of being assaulted would be greatest when they were in greater proximity to alcohol outlets at times when the outlets were most heavily patronized and were would be most likely to attract or generate motivated offenders (e.g. at night for bars and restaurants) (38).

The Space-Time Adolescent Risk Study (STARS) was a case-control study conducted among young people aged 10 to 24 in Philadelphia, Pennsylvania.

Eligible cases were patients aged 10 to 24 presenting to the Emergency Departments of the Hospital of the University of Pennsylvania or the Children's Hospital of Philadelphia for treatment of a non-gun assault (n = 194) or gun assault (n = 135) injury. Eligible non-gun assaults were patients admitted for treatment of a traumatic injury which they self-reported was intentionally inflicted by another person with or without a non-gun weapon; gun assaults were patients admitted for treatment of a traumatic injury which they self-reported was inflicted by another person with a gun. Age-matched controls (n = 274) were selected using random digit dialing among residents of the 12 ZIP code catchment area for the hospitals (response rate: 57.1%). The required sample size for this study was calculated for analyses examining aggregate relationships between assault risk and physical exposure to all alcohol outlets over participants' entire days. Because this secondary analysis disaggregates by time of day and alcohol outlet type, results may be biased towards null.

The materials and methods for STARS have been described in detail previously (15). Briefly, a trained interviewer administered a GIS-assisted survey to cases and controls in the hospital, at the research office, or in participants' homes. The survey assessed demographics, general health, and perceptions of their residential neighborhood. Participants chronologically described their minute-by-minute locations and activities for the day of the assault (cases) or a randomly selected day from among the three days prior to the interview (controls). Interviewers sat side-by-side with participants while looking at a shared computer screen. Data were entered into a custom GIS-based software package that collected the latitude and longitude coordinates and time of participants' movements. For each point in time at which their activities or location changed, participants described their current activities (free text), whether an adult family member was present (dichotomous), whether a peer was present (dichotomous), mode of transport (on foot, in vehicle, or other), and whether they were in possession of alcohol (dichotomous). We also identified weekend hours (from sunset on Friday to 11:59pm Sunday) and times when participants were at home.

For the current study, we divided each participant's one-day data into ten-minute segments beginning at 4:00am (e.g., 4:00 to 4:09am). The latitude and longitude at the start of each segment described participants' geographic locations for the following ten minutes (Figure 1a). In total, there were 10,808 segments for the non-gun cases, 8,989 segments for the gun assault cases, and 24,077 segments for the controls. We then estimated spatially and temporally specific proximity to alcohol outlets and other neighborhood features for each segment.

The main independent variable was proximity to alcohol outlets. The Pennsylvania Liquor Control Board issues three classes of retail alcohol license relevant to this study. Restaurant Licenses permit beer, wine and liquor sales for on-premise consumption (hereafter, “bars and restaurants”). Eating Place licenses permit the sale of beer for on-premise consumption with a meal and for off-premise consumption provided sales are not in single containers (“beer stores”). Off-premise sales of wine and liquor are limited to government monopoly off-premise outlets located throughout the state (“liquor stores”). To construct raster layers describing continuous densities of the city's 427 bars and restaurants, 1056 beer stores, and 634 liquor stores, we spatially smoothed these data using kernel density estimation (Figure 1b, 1c and 1d). We then spatially joined these raster data to the points representing the beginning of each ten-minute segment. Given the high spatial precision of our data, we considered this a better approach than taking values aggregated within arbitrary administrative units (e.g. Census block groups). The three resulting variables describing proximity to alcohol outlets were heavily positively skewed (0.89 ≤ skewness ≤ 6.25). We calculated their natural logarithm to reduce the likelihood that the extreme high values would inordinately influence results.

We collected key demographic characteristics and behavioral indicators that could be related both to participants' physical exposure to alcohol outlets and to risk for assault, and may therefore confound relationships (e.g. whether they had ever carried a gun).

Neighborhood characteristics may also be related both to assault risk and to alcohol outlet density. Data to describe 26 neighborhood characteristics were obtained from four sources: Census 2010 data described demographic characteristics for the local resident population within Census block groups; tax parcel data from Philadelphia described land use; the 2008 Southeastern Pennsylvania Household Survey (39) described local population characteristics; and participant responses described perceptions of local areas. Similar to the approach for alcohol outlets, these spatially referenced data were converted to raster layers using kernel density estimation for point data and inverse distance weighting based on the centroids of polygon data. In a factor analysis, 23 characteristics loaded onto five factors: (i) neighborhood connectedness, (ii) income, (iii) vacancy, vandalism, violence, (iv) emergency services, and (v) race/ethnicity. The composition of these scales is described in a previous paper (15). The three variables not captured in the factor scales but included as independent variables in the current analysis were commercial land use (to account for the possibility that alcohol outlets mark for physical exposure to retail areas), population density (possibly representing guardianship) and the density of middle/high schools (also representing guardianship).

The units of analysis were the ten-minute segments nested within eight uniformly defined three-hour time phases (e.g. 4am to 6:59am). Two conditional logistic regression models estimated the (i) overall odds of non-gun assault compared to controls, or (ii) the overall odds of gun assault compared to controls. Both models included a fixed effect for the time phase, such that exposures for cases were compared to controls within the same three-hour period. This approach partially controlled for the possibility that segments were autocorrelated within participants. All independent variables describing physical exposure to alcohol outlets, neighborhood conditions, and other behavioral and temporal characteristics were included in both models. Analyses included all segments for the controls, but only the final observation before the assault for cases. Gun assault cases were dropped from the non-gun assault model, and non-gun cases were dropped from the gun assault model. This procedure produced an overall estimate (i.e. aggregated over the whole day) of the odds ratio and 95% confidence interval for non-gun and gun assault per unit increase in physical exposure to alcohol outlets.

We then conducted separate logistic regression models within each time phase, dropping cases whose assault event occurred outside the three-hour window. We thus produced time-of-day specific estimates for the relationships between proximity to alcohol outlets and the odds of assault. Given that proximity to alcohol outlets was calculated from the kernel density estimate then log transformed, individual parameter estimates are not easily interpretable, but the direction of the association and comparison within outlet types across day phases is informative.

Spatial data management was performed using ArcGIS v10.1 (40). Parameters for the regression models were estimated using Stata v14 (41).

We conducted several specification tests to reduce the likelihood that observed relationships were artefacts of model construction. First, after comparing individual-level participant characteristics for cases vs. controls using Students t-test for continuous measures and Chi-square tests for categorical measures, we adjusted our main effects analyses for variables where the groups differed systematically. Second, a correlation matrix for the logged alcohol outlet variables demonstrates that these measures were moderately collinear (Table 1). After systematically removing one and then two of the alcohol outlet variables, we repeated the conditional logistic regression and the logistic regression models. Finally, we then tested alternate time-phases (e.g., two-hour periods, four-hour periods), and further stratified the analyses by age (< or ≥ 18).