Looking Back and Moving Forward: The Evolution and Potential Opportunities for the Future of Alcohol Outlet Density Measurement

Abstract

The literature consistently finds that areas with greater density of alcohol outlets (places that sell alcohol) tend to have higher levels of public health harms. However, conflicting findings arise when researchers drill down to identify the type(s) of alcohol outlets with the strongest associations with harms and the mechanisms that explain these associations. These disagreements could be a result of the outdated methods commonly used to quantify the alcohol environment: counts of the number of outlets in an area. This manuscript reviews the events and ideas that shaped the literature on the physical alcohol environment. It then defines the three main methods used to measure alcohol outlet density, conducts an exploratory factor analysis to explore the constructs underlying each method, and presents a novel conceptual framework that summarizes the three methods, their respective underlying constructs, and the setting(s) in which each may be most appropriate. The framework proposes that counts of alcohol outlets measure availability, proximity to the nearest outlet measures accessibility, and spatial access measures measure access, which comprises both availability and accessibility. We argue that researchers should consider using proximity and spatial access measures when possible and outline how doing so may present opportunities to advance theory and the design and implementation of alcohol outlet zoning regulations. Finally, this manuscript draws on research from other areas of the built environment to suggest opportunities to use novel methods to overcome common hurdles (e.g., separating subtypes of outlets, ecologic designs) and a new challenge on the horizon: home delivery.

Overview

Policymakers are rolling back regulations on the physical availability of alcohol around the world (World Health Organization 2018). For example, from 2010 to 2015, 16 countries reduced regulations on hours of alcohol retail sales, and seven countries loosened regulations on alcohol outlet density (Jernigan and Trangenstein 2017). At the same time, the accumulating evidence base - including a recent systematic review examining time series designs - outlines a more compelling case for establishing that increased alcohol outlet availability may cause a range of adverse outcomes (Sherk et al. 2018). One way to reduce the disconnect between the literature base and real-world events is to conduct research that more accurately reflects the spatial dynamics of the alcohol environment (i.e., the physical availability of alcohol in a location) in order to inform real-world policy decisions.

This manuscript first reviews the conceptual and methodological origins of alcohol outlet availability research to better understand the approach used in the current literature. The Contemporary Alcohol Outlet Density Research Methods section then defines three main methods of measuring alcohol outlet availability - counts, proximity, and spatial access – and reviews the strengths and weaknesses of each approach. To provide greater clarity on differences between these methods, an exploratory factor analysis to identify the constructs that each measurement method measures is conducted, and a novel conceptual framework that summarizes the concept(s) measured by each method and the accompanying strengths and limitations is presented. Reflecting on the challenges encountered in alcohol outlet availability research over the years and methodological limitations of commonly used methods, the Future Directions for Research section proposes a range of potential solutions that utilize technological advances and/or draw on other areas of built environment research to fill gaps in the current alcohol outlet density research.

Early alcohol outlet availability research

Theoretical underpinnings

After the repeal of prohibition, Jellinek’s medical model of alcoholism strongly influenced concepts of alcohol use and measurement (Jellinek 1960). Jellinek distinguished types of alcoholics, which focused on medicalizing individual-level “drunkenness” and consequently limited alcohol-related problems to alcohol dependence (Jellinek 1960). Researchers adopted this narrow focus of intentionally minimizing alcohol’s social and health risks to distance themselves from the temperance movement (Room 1984).

As alcohol research progressed, the “sociocultural” or “integrationist” model, which first introduced population-level alcohol use (Whitehead 1975), assumed that the United States had an unusually high prevalence of alcoholism because its society was ambivalent toward alcohol (Room 1984). The idea underlying this model was that areas with lower rates of alcoholism tend to have cultural norms that discourage heavy drinking and encourage moderate drinking (Chafetz 1974). Proponents of this perspective therefore fought to implement moral education and reshape cultural norms to combat alcohol-related problems (Chafetz 1974; Room 1984).

In the 1970’s, Kettil Bruun’s seminal book, Alcohol Control Policies in Public Health Perspective, first suggested evidence-based methods of reducing alcohol-related problems without focusing on cultural norms or education (Bruun et al. 1975). The World Health Organization (WHO) also published an expert committee report concluding that alcohol dependence comprises only a small portion of alcohol-related harms (WHO Expert Committee on Problems Related to Alcohol Consumption 1980). This argument summarized the prevention paradox: while dependent drinkers experience the highest rates of alcohol-related harms, the majority of alcohol-related harms at the population level stem from non-dependent drinkers, who far outnumber dependent drinkers (Room 1984).

Alongside these developments, researchers adopted the distribution of consumption model, which integrated a more population-based approach than the previous sociocultural model. This model was based on the observation that, in the absence of individual-level controls, alcohol use is unimodal, positively skewed, and log-normal, meaning there are fewer drinkers at each successive level of per capita consumption after the modal value (Blane 1976; Room 1984). This idea set the stage for population-level alcohol policies because of the idea that the public would be more likely to believe that such policies could reduce heavy drinking if dependent drinkers existed on the same continuum (Room 1984).

These theoretical perspectives were the backdrop of the early research on the alcohol environment. The sociocultural model was evident in the frequently-used “social area analysis” method (Donnelly 1978; Rabow and Watts 1982; Watts and Rabow 1983). In these studies, researchers argued that changes could be traced back to differences in a) social rank (“the objective factors of social class like occupation, education, and income”); b) urbanization or lifestyle (“the way of life chosen by the population whether it is the ‘family-committed life’ in the single-dwelling-unit area or the life of the working couple in the apartment house area”), c) and racial/ethnic cultural backgrounds (Greer 1962) (p. 31-32). Overall, these early studies concentrated on cultural dimensions, and the alcohol environment was often relegated as an ad hoc sub-analysis of studies trying to unpack the sociocultural effects of outlets on alcohol dependence or per capita consumption (Colón et al. 1982; Rabow and Watts 1982; Watts and Rabow 1983).

Methods used in early alcohol outlet availability research

Early studies of physical availability of alcohol were largely located in the United States and were an exercise in trial and error. Smart (1977) was among the first to quantify alcohol availability nationally using an eight-item scale based on a self-proclaimed “not scientific” method (Medicine in the Public Interest 1976). This scale included several different types of physical and temporal availability, including minimal legal purchase ages, the number of alcohol outlets in an area, and days and hours of sales (Smart 1977). Subsequent work focused on one or two of Smart’s original eight categories (Parker D et al. 1978; Harford et al. 1979) and determined that the association between the number of alcohol outlets in an area and related harms depended on the type of outlet (e.g., on- vs. off-premise outlet) (Rabow and Watts 1982).

At the same time, there were several natural experiments in the Nordic countries that lent support to the causal association between alcohol outlet density and related harms even though the researchers were not yet quantifying alcohol availability. Sweden introduced medium-strength beer (4.5% alcohol by volume [ABV]) to grocery stores and state-run alcohol stores in 1965 and removed it from shelves 12 years later. These changes in beer availability were associated with fluctuations in total alcohol sales (Noval and Nilsson 1984) and hospitalizations among adolescents in the expected directions (Ramstedt 2002). In 1969, Finland permitted alcohol outlets to open in rural locations and medium-strength beer (4.7% ABV) to be sold in grocery stores and cafes, and consumption and alcohol-attributable hospitalizations rose notably after this policy change (Poikolainen 1980).

As researchers developed methods for evaluating the associations between alcohol availability and potential harms, it became clear that the spatial units of analysis that researchers used could affect results. Many early studies in the United States were national (Smart 1977; Parker D et al. 1978; Harford et al. 1979; Colón et al. 1982; Lester 1995) or multi-state in scope (Colon 1981), using states as the unit of analysis. Some of these studies detected a significant association despite small sample sizes (Parker D et al. 1978; Harford et al. 1979; Colon 1981) while others did not (Popham et al. 1975; Smart 1977). Consequently, researchers questioned whether heterogeneity across and within states’ sociodemographic factors, drinking patterns, and outlet distributions caused these conflicting findings (Rabow and Watts 1982; Gruenewald P 1993). To test these possibilities, subsequent studies used smaller units of analysis such as counties (Donnelly 1978; Rabow and Watts 1982; Rush et al. 1986; Kelleher et al. 1996) or cities (Watts and Rabow 1983; Scribner RA et al. 1994, 1995).

Ultimately, these early findings laid the foundation to include the alcohol environment as a determinant of public health outcomes (Parker D et al. 1978). Once enough studies accumulated significant findings, authors hypothesized that alcohol outlets may play a role in generating alcohol-related problems (Watts and Rabow 1983) and limiting the number of alcohol outlets may reduce consumption and related harms (Colón et al. 1982). Notably, this paradigm of evidence preceding advances in theory was common in epidemiologic studies in the latter half of the twentieth century. Although criticized as a “black box” approach to describing statistical associations without attempting to engage with and understand social mechanisms (Pearce 1996; Susser 2004), the approach has some utility for exploring under-researched phenomena (Greenland et al. 2004; Weiss 2004).

Transition to contemporary alcohol outlet density research

Several key methodological developments propelled research on the alcohol environment forward. Building on earlier insights and recognizing that neighborhood-level factors also played a role—and leveraging advances in geographic information systems making computer mapping more accessible (see Gruenewald PJ et al. 1996, for an early example)—studies eventually began decreasing the size of units of analysis even further (Scribner RA et al. 1998; Speer et al. 1998). These efforts demonstrated that smaller units of analysis yielded estimates that were more responsive to local-level variations. Gruenewald (1996) and Scribner (1999) articulated that the benefit of using smaller units of analysis is that the populations they capture are relatively homogenous; the larger the units, the more likely they are to average across diverse populations. This renders larger units susceptible to aggregation biases, meaning that small area effects were lost because parameter estimates measured average associations across larger geographic units, while standard errors were also inflated as the number of units decreased.

The simultaneous advent of analytical approaches within geographic information systems (GIS) enabled researchers to detect and adjust for positive spatial dependence (called “autocorrelation,” which is the degree to which objects like alcohol outlets that are closer in space are more/less similar) (Gorman Dennis M et al. 2001; Treno et al. 2001; Lipton and Gruenewald 2002). Researchers learned about temporal autocorrelation with earlier privatization studies, but they were unaware of the significance of spatial autocorrelation (Wagenaar and Holder 1991; Scribner R et al. 1999). Like temporal dependence, spatial autocorrelation violates the ordinary least squares regression assumption that units are independent, and it can lead to misestimated standard errors, which (in the presence of positive autocorrelation) can result in false positives (Waller and Gotway 2004).

Contemporary alcohol outlet density research methods

While the initial alcohol outlet density measurement methods were developed in the United States, the field has expanded geographically and has benefitted greatly from researchers around the globe. Still, there has been little variation in the way that alcohol outlet density in quantified around the globe: there are three primary methods to measure the alcohol environment: counts, proximity, and spatial access (Table 1). Counts are based on a set of “containers” using existing (e.g., ZIP code or postcode) or user-defined boundaries (e.g., a buffer zone around a point of interest). Researchers may stop with a raw count, or divide their count by a measure of space or reach. Proximity methods summarize the distance from a fixed location to the closest alcohol outlet. Lastly, spatial access methods sum the raw or inverse distances from a geographic location to a subset of X alcohol outlets defined by either a set number or a radius. The overwhelming majority of studies on alcohol outlet access use count methods (Holmes et al. 2014).

Table 1.

Overview of Methods to Calculate Alcohol Outlet Access

Method	Definition	Measures	Strengths	Weaknesses	Geographic Appropriateness
Count	“Containers” are existing geopolitical boundaries or a user-defined area; count-based methods use either the raw count of alcohol outlets in the container or a simple fraction, where the numerator is a count of alcohol outlets and the denominator is a measure of space: $$\frac{C}{w}$$ Where:  C = count of outlets  w = weight	Availability	Easiest to calculate (Centers for Disease Control and Prevention 2017)	They do not measure any dimension of accessibility
			Permit comparisons across communities (Centers for Disease Control and Prevention 2017)	They assume alcohol outlets are uniformly distributed across the container, which means they are ill equipped to study outcomes where clustering is important (Grubesic T and Pridemore W 2011) or study heterogeneous areas (Scribner R et al. 1999; Van Meter et al. 2011; Cameron M. P. et al. 2016)	N/A – performance is similar in urban, suburban, and rural areas
			Do not rely on a container centroid, which can introduce an additional source of uncertainty (Waller and Gotway 2004)
				They can overestimate effects (Geronimus Arline T et al. 1996; Geronimus A. T. 2006)
				They are prone to edge effects
				The containers themselves may be poorly designed for spatial analyses (e.g., awkward shape or impractical scale) (Branas et al. 2009), which can lead to the modifiable areal unit problem (Guagliardo 2004; Gmel et al. 2016)
Proximity	Travel impedance (e.g., distance, time) from a fixed location: $$\min \left(\text{d}\right)$$ Where:  d = distance to the closest outlet	Accessibility	Avoid edge effects within the study area borders	Studies of the closest alcohol outlet in urban areas are insensitive to the reality that there are often many alcohol outlets at similar distances (Guagliardo 2004)	May perform best in rural areas
			Proximity to the nearest outlet is anticipated to be a better measure of access for rural areas, where residents are more likely to patronize the closest outlet (Guagliardo 2004)	May be biased to weight distance differently depending on whether people live near the centroid or the border (Halonen et al. 2013)
				Unable to adjust for outlet size or detect clustering of outlets, which can increase the effect of the outlet on related harms (Grubesic et al. 2016)
Spatial Access	Sum of the inverse distances from a geographic locale on (e.g., CT centroids) to a subset of X alcohol outlets, often calculated as: $$\sum _1^n\frac{s}{d}$$ Where:  N = the number of outlets in the choice set  D = distance to the outlet  S = Supply in the area	Availability & accessibility	Avoid edge effects if use a choice set instead of a boundary	Most difficult to calculate (Centers for Disease Control and Prevention 2017)	Container-based spatial access methods may perform best in urban areas Choice set spatial access methods may perform best in homogeneous areas
				Will encounter edge effects if use a radius/buffer Will combine micro, meso, and macro measures if use a choice set approach over heterogeneous area Approach will double count outlets if the containers/choice sets overlap units of analysis, which is particularly problematic when used to measure outcomes
			Research in other fields concludes that spatial access measures are the most robust (Fryer et al. 1999; Guagliardo 2004)	Researchers may encounter aggregation bias if the radius selected is too large, because the measure will average across heterogeneous areas (Yang et al. 2006)
			Inverse distance weighting is a common approach in spatial statistics, and researchers no longer need to aggregate information	If the measure does not use inverse distance weighting (e.g., uses the mean or median distance), it will overweight outlets toward the periphery of the container and will invite edge effects [110]

Each of these methods has advantages and disadvantages. Counts are easy to calculate and understand. They do not require street-level data, so they are the only method that can be used if local jurisdictions summarize alcohol outlets with counts per unit. The primary disadvantage of counts is that they lose precious information about the configuration of outlets. Fundamentally, counts assume alcohol outlets are uniformly distributed across areas (Centers for Disease Control and Prevention 2017), and they do not include any measure of the distance to, from, or between outlets. In doing so, they put outlets into geographical ‘bins’, ignoring the complexity of the way the built environment is experienced. This is problematic, because clustering is associated with increased levels of harms (Grubesic TH and Pridemore WA 2011; Han and Gorman 2014; Zhang et al. 2015). This means that counts are not sensitive enough to detect all of the variation in the alcohol environment that will determine its association to related harms (Grubesic et al. 2016).

All counts and some spatial access measures (i.e., those defined by buffers) are container-based, and container-based methods have their own limitations. First, container-based methods run the risk of introducing aggregation biases if they average effects over large, heterogeneous populations (Holmes et al. 2014). In addition, they suffer from edge effects (Waller and Gotway 2004), which means that alcohol outlets across a container boundary may influence the outcome inside the container. Similarly, container-based methods are susceptible to the modifiable areal unit problem (MAUP), which is a statistical challenge where the size and shape of the containers used for container-based methods change the results (Sadler and Lafreniere 2017). Lastly, container-based variables are often multimodal, because they have one distribution for all of the containers with alcohol outlets and a separate distribution for the containers without alcohol outlets, which creates variables that are not log-normal.

Methods that incorporate street-level information into the measure of alcohol outlet density like proximity and spatial access measures – regardless of whether they use containers – are more sensitive (Grubesic et al. 2016) and may provide superior model fit than counts (Trangenstein, Curriero, et al. 2018a). Proximity methods only require two data points: the reference point and the nearest alcohol outlet. This is simultaneously a strength and a weakness. It is a strength because it means they are easy to calculate. Conversely, it is a weakness because relying on the alcohol outlet nearest to a reference point is a somewhat random process. Researchers who calculate proximity measures need to decide between using Euclidian (straight-line distance or “as the crow flies”) or network/Manhattan distance (road-based). Network/Manhattan distances are more reflective of real-world conditions and may generate more accurate estimates of access/exposure (Arbour et al. 2009; Larsen et al. 2015), but they are also more difficult to calculate for novice GIS users, requiring the use of network datasets. Finally, one remaining limitation is that they oversimplify the equation of alcohol access into the nearest store and fail to account for spatial polygamy, or the idea that personal exposure is not limited to only one store (Matthews 2011; Kestens et al. 2012).

In contrast to container-based methods, all proximity methods and some spatial access methods (i.e., those using a “choice set” approach) are smoothed estimates, meaning that they do not use containers and instead draw information from surrounding areas, which can yield more stable estimates in locations with few data points (Waller and Gotway 2004).

The third method of calculating alcohol outlet density is spatial access measures. These methods incorporate the configuration of specific outlets into the measure. These measures may either be smoothed (the choice set approach) or container-based (the buffer approach). The choice set approach sums the raw or inverse distances from a geographic location to a subset of X alcohol outlets. Spatial access measures provide the most sensitive measures (Grubesic et al. 2016) with potentially better model fit and lower error than count or proximity measures (Trangenstein, Curriero, et al. 2018a). Spatial access methods will, however, suffer from the limitations associated with either smoothed and/or container-based approaches. Finally, spatial access methods rely on a reference point, and this can introduce uncertainty (Waller and Gotway 2004) due to being located inside unrealistic areas (e.g., parks, colleges) or away from the bulk of the population.

There has been relatively little work done to compare the statistical properties of these different measurement methods. One rare example is Grubesic’s (2016) comparison of count and spatial access methods, which concluded that these two methods measure separate constructs (Grubesic et al. 2016). Using a similar approach as Grubesic, we present correlations between different measurement methods in Figure 1, showing stronger correlations within each of these three categories of measurement methods.

Figure 1.

Scatterplot matrix for count, proximity, and spatial access measures with Pearson’s Correlation Coefficient

To go one step further and identify the potential dimensions underlying these methods, we drew on literature from healthcare, which argues that there are five dimensions of the environment: availability, accessibility, affordability, acceptability, and accommodation (Penchansky and Thomas 1981). The combination of these dimensions comprises access, and only two of these five dimensions (availability and accessibility) are spatial (Guagliardo 2004). Availability measures the number and type of options from which patrons may choose when accessing alcohol (Penchansky and Thomas 1981). Accessibility measures how easy or hard it is to access each option, which is often operationalized as travel distance (Penchansky and Thomas 1981; Guagliardo 2004). Availability and accessibility are related, though they measure separate dimensions of the alcohol environment.

To contextualize the concept of access, it may be helpful to consider that the word is both a noun that describes qualities of the built environment and a verb that describes how a person physically interacts with the features of the built environment. We focus on the former, which examines key features of the built environment that determine how people can interact with it.

We hypothesize that these three methods measure separate constructs: counts measure availability, proximity methods measure accessibility, and spatial access methods measure access. To test this hypothesis, we conducted an exploratory factor analysis to detect the dimensions that underlie these alcohol outlet density measurement methods (see Box 1, Table 2). The results suggest that counts, proximity, and spatial access methods measure different constructs. In light of these findings, we propose the model of alcohol outlet density measurement shown in Figure 2.

Box 1.

Objective:

This analysis aims to identify classify the constructs underlying different methods of measuring of alcohol outlet density.

Methods

Data Source.

Liquor license information, including license type and address, was obtained for 1,204 alcohol outlets from Baltimore City as of June 4, 2016.

Measures.

To compare methods for measuring alcohol outlet access, we calculated alcohol outlet access using three separate methods: counts, proximity, and spatial access, as described below.

Counts.

We summed the number of alcohol outlets located in each CBG. We then created four versions of this variable using different denominators: no denominator (i.e., a crude count), population denominator, area (measured in square miles) denominator, and roadway miles denominator.

Proximity.

We calculated two proximity measures as the distance from each CBG centroid to the closest alcohol outlet. The first variable used network (road-based) distance, and the second used Euclidian (straight line or “as the crow flies”) distance.

Spatial Access.

We created nine spatial access variables. The first set of three spatial access variables were spatial accessibility indices (SAI) using a choice set (smoothed) approach. The second set of three spatial access methods were SAIs using a container-based (buffer) approach. The final set of three spatial access methods were average distances between the CBG centroid and a choice set of alcohol outlets. For each set of spatial access variables, we calculated a micro-, meso-, and macro-level variable. For the SAIs with the choice set approach, we summed the inverse distance to the nearest 3 (variable 1), 7 (variable 2), or 10 (variable 3) alcohol outlets. For the second set of three SAIs, we summed the inverse distances to all alcohol outlets within 0.25- (variable 4), 0.5 (variable 5), and 1-mile buffers (variable 6). For the mean variables, we again used the nearest 3 (variable 7), 7 (variable 8), or 10 (variable 9) alcohol outlets.

Analysis.

We used a natural logarithm transformation for all alcohol outlet density variables. We first created a scatterplot matrix to understand the correlation between the different measures of alcohol outlet density. We also calculated Pearson’s correlation coefficient as well as covariates between each alcohol outlet density measure. Finally, we used exploratory factor analysis to assess dimensionality of the alcohol outlet density measures. We used eigenvalues, parallel analysis, and a scree plot to determine the number of factors to extract. We used iterative principal factor analysis, because the multivariate normal assumption was not met. We rotated our solution using promax rotation, because the factors were correlated.

Results

Two components had eigenvalues greater than 1 in the principal components analysis, and the parallel analysis suggested 2–7 relevant factors. The elbow on the scree plot was around the second or third component. The first two components explained 86.3% of the variation in the alcohol outlet density measures, and the first three measures explained 92.0%. In the two and three factor models, no measures had uniqueness >0.5 or loadings <0.4 (Table 2). In the four-factor model, no variables had a loading >0.4 for the fourth factor, so it was dropped from consideration. We ultimately selected the three-factor solution because the factors each contained at least two measures with a loading of at least 0.4 and no variables had a uniqueness larger than 0.5.

The trends in the factor loadings mirror our hypothesized framework. All of the spatial access measures load on factor 1 (access), the counts load on factor 2 (availability), and the proximity measures load on factor 3 (accessibility). The inter-factor correlations also supported our classification. Factor 1 (access) correlated with both factor 2 (availability) and factor 3 (accessibility) (factors 1 and 2 r=−0.52; factors 1 and 3 r=0.37). This is consistent with our hypothesis that availability and accessibility were components of access. Further, we would not expect factors 2 (availability) and 3 (accessibility) to be correlated because they measure separate features of access (r=0.04).

Limitations

Our approach of using exploratory factor analysis was designed to describe correlation structures between variables. While the labels for factors can often be arbitrary, the factors we identified in this analysis conformed to a seminal conceptual framework. Another common criticism of exploratory factor analysis is that methodological decisions will all shape the results, and our process involved several junctures that required decisions. However, our decisions were guided by output from the analyses and clearly stated.

Table 2.

Factor loadings in the exploratory factor analysis

	Factor 1	Factor 2	Factor 3	Uniqueness
Counta	>−0.01	1.00	0.02	0.01
Count over populationb	−0.02	1.00	0.01	0.03
Count over areac	−0.04	0.97	−0.01	0.01
Count over roadway milesd	0.03	0.97	−0.01	0.02
Proximity with network distancee	0.36	0.03	0.72	0.01
Proximity with Euclidian distancef	0.37	−0.03	0.65	0.07
SAI to 3 nearest outletsg	0.61	0.47	<0.01	0.02
SAI to 7 nearest outletsh	−0.60	0.02	−0.26	0.34
SAI to 10 nearest outletsi	−0.83	>−0.01	−0.21	0.02
SAI with 0.25-mile bufferj	−0.47	0.09	−0.40	0.27
SAI with 0.5-mile bufferk	−0.50	0.19	−0.23	0.34
SAI with 1.0-mile bufferl	−0.97	0.01	0.04	0.11
Mean distance to 3 nearest outletsm	0.81	−0.04	0.18	0.04
Mean distance to 7 nearest outletsn	1.02	−0.03	−0.06	0.02
Mean distance to 10 nearest outletso	1.08	−0.01	−0.13	0.01

SAI spatial accessibility index

^aCalculated as the number of alcohol outlets in a census block group.

^bCaclualted as the number of alcohol outlets in a census block group divided by the population of the census block group.

^cCaclualted as the number of alcohol outlets in a census block group divided by the area of the census block group.

^dCaclualted as the number of alcohol outlets in a census block group divided by the total roadway miles of the census block group.

^eCalculated as the minimum network (road-based) distance from the census block group centroid to the closest alcohol outlet.

^fCalculated as the minimum Euclidian (as the crow flies) distance from the census block group centroid to the closest alcohol outlet.

^gCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to the three closest alcohol outlet.

^hCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to the seven closest alcohol outlet.

ⁱCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to the 10 closest alcohol outlet.

^jCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to all outlets within a 0.25-mile buffer.

^kCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to all outlets within a 0.5-mile buffer.

^lCalculated as the sum of the inverse network (road-based) distances from the census block group centroid to all outlets within a 1.0-mile buffer.

^mCalculated as the average of the network (road-based) distances from the census block group centroid to the three closest alcohol outlet.

ⁿCalculated as the average of the network (road-based) distances from the census block group centroid to the seven closest alcohol outlet.

^oCalculated as the average of the network (road-based) distances from the census block group centroid to the 10 closest alcohol outlet.

Figure 2.

Conceptual framework of dimensions of spatial access and alcohol outlet density measurement methods

Future directions for research

Moving forward, measures that incorporate street-level data have potential to advance our understanding of the role of alcohol outlets in related harms. Still, we are aware of only 12 analyses to date that use these methods (Branas et al. 2009; Branas et al. 2011; De Boni et al. 2013; Groff ER 2013; Richardson et al. 2015; Zhang et al. 2015; Cameron Michael P et al. 2016; Fone et al. 2016; Morrison et al. 2017; Trangenstein, Curriero, et al. 2018b). Evidence that documents the statistical advantages of proximity and spatial access methods over counts is mounting (Grubesic et al. 2016; Trangenstein, Curriero, et al. 2018a), and the field is at a pivotal juncture where researchers should consider the benefits of retaining street-level data. Like reducing the size of the units of analysis in the late 1990’s, the shift from polygons to points will remove unnecessary error from models and increase the defensibility of data used in policy debates. This shift is already afoot in work on healthy food access (Kestens et al. 2012; Hillier et al. 2015; Sadler et al. 2016), and research on alcohol environments would benefit from following suit.

The introduction of GIS technologies allowed researchers to begin adjusting for spatial autocorrelation, but researchers are not yet using other potentially beneficial analytic methods and tools. Alcohol outlet clusters are one particular configuration of alcohol outlets associated with greater levels of harm (Gorman Dennis M et al. 2005; Zhu et al. 2006; Grubesic TH and Pridemore WA 2011; Han and Gorman 2014; Zhang et al. 2015), and they have clear policy implications: studying clusters could help establish evidence-based standards for minimum distances between alcohol outlets. The challenge is that cluster detection requires specialist methods such as local indicators of spatial association (LISA) (Anselin 1995), Getis-Ord Gi statistics (Getis and Ord 1992), and spatial scan statistics (Kulldorff 1997) that may be inaccessible to researcher teams with little GIS expertise. Still, the potential gains from this line of inquiry are high, as they may shed new light on the types of alcohol environments associated with the most harms. Opportunity for methodological advancement in this area is strong because the current literature lacks guidance on how to tailor cluster detection methods for urban, suburban, and rural settings. We encourage exploration in this area and suggest that rurality corrections for spatial scan statistics (Gangnon 2010)—particularly among HIV (Namosha et al. 2013) and obesity researchers (Penney et al. 2014)—may provide inspiration.

One limitation of alcohol outlet density research is that commonly-used data sources do not provide details necessary to disaggregate outlets more granularly than on- or off-premise. It is likely that some alcohol outlets bear larger burdens than others, and researchers cannot isolate those features without more detailed data sources. A few researchers have used observational data to supplement administrative data (Parker RN et al. 2011). We are only aware of a few observation instruments for alcohol outlets in general (Parker RN et al. 2011; Graham and Homel 2012; Milam et al. 2013; Morrison Christopher et al. 2016; Morrison C. et al. 2016; Alcohol Prevention Enhancement Site nd), but further examples of such instruments can be found in observation instruments designed to scan outlets located near college campuses (Ryan et al. 2009; Martin et al. 2012). However, food environorment reserachers have demonstrated that observations can be used to assign scores to outlets denoting healthy food (Glanz et al. 2007), which can also be used to understand spatial patterning in access to specific types of food (Shaver et al. 2018). A more modern and potentially cost-effective solution is to use Google StreetView, which has been used successfully used in the physical activity literature (Hara et al. 2013; Li et al. 2015; Gebru et al. 2017). Google StreetView is a source of existing data that can help identify sub-types of alcohol outlets (e.g., street frontage, advertising, adjacency to other known risk factors) and potentially refine measures of neighborhood characteristics. Applying these methods to the alcohol environment could allow researchers to begin to understand the spatial distribution of potentially problematic products (e.g., single-serving containers, alcopops), cheap alcohol, and alcohol marketing, and the association of different distribution patterns with related harms.

Another hurdle that may require creativity to overcome is that most of the alcohol outlet literature is ecologic, meaning findings do not apply at the individual outlet level. Like container-based methods of alcohol outlet density measurement, larger units of analysis often contain a more heterogeneous mix of outlets and surrounding areas, which can invite aggregation bias and MAUP (Waller and Gotway 2004). Researchers should consider quasi-experimental or case-control designs (e.g., Zhang et al.) to complement ecologic studies where possible. We advocate using identical methods in multiple cities to determine how alcohol environments may yield disparate outcomes in different locales based on factors such as urbanicity, street design, population density, local cultural factors, policy differences, and policy changes. Only by directly comparing two or more cities or regions will research be able to more definitively understand not only which methods are most effective in measuring exposure to alcohol environments, but how these methods can help guide policymakers interested in creating conditions for healthier environments.

An emerging challenge is the expansion of alcohol home delivery services, which may require researchers to separate harms associated with home deliveries from those associated with brick and mortar outlets. One potential tool that could help overcome this hurdle from the spatial criminology literature is Risk Terrain Modeling (RTM), which is a software that first creates individual layers of risk surfaces using what is known about the spatial extent of harms associated with each type of outlet (i.e., the distance that harms from a given location extend in space) (Caplan et al. 2011; Caplan and Kennedy 2016). For example, preliminary data estimate the harms from off-premise outlets extend approximately one-quarter mile or 1,320 feet (Groff E 2011; Furr-Holden et al. 2014; Groff ER 2014), and the spatial extent of the home delivery services could be assigned as the delivery area. RTM would then weight and add the individual risk surfaces to create one continuous surface over a study area that summarizes the total risk from all outlets and home delivery services.

The use of point-level data proposed here may also help advance Routine Activity Theory as applied to alcohol outlets. Briefly, Routine Activity Theory hypothesizes that violence arises from the combination of the types of people in a space and the constraining or facilitating components of those spaces (Cohen and Felson 1979). Place managers are responsible for maintaining order in a given location (e.g., bartenders, wait staff), and guardians are responsible for protecting individuals (e.g., police officers, bystanders). Place managers and guardians work together to alter how outlets bring people together and whether violence sparks (Cunradi et al. 2011; Livingston 2011; Snowden 2016). These key concepts suggest outlet characteristics that could be associated with violence; for example, place managers are most effective at preventing violence when they have a clear, unobstructed view of patrons, so outlets with plexiglass barriers or crowded clubs may inhibit place managers’ ability to maintain order and increase the changes for violence. The notion of guardianship also reveals that the location of alcohol outlets is critical, because the areas around outlets (e.g., parking lots, vacant buildings) can act as de facto bars/taverns (Grubesic T and Pridemore W 2011).

Currently, routine activity theory offers a conceptualization of the spatial relationships between alcohol outlets and violence but it stops short of clarifying the mechanisms through which alcohol outlet availability may cause related harms and whether those mechanisms differ based on the neighborhood context in which the outlet is located. Madensen and Eck (2008) present a conceptual framework for outlet characteristics and place manager decision making that could provide a useful tool for researchers seeking to use more careful measurement to isolate potential mediators of the association between alcohol outlets and public health harms, particularly when used in combination with other innovations presented in this section such as identifying sub-types of outlets.

We are currently facing a “translation gap” where the results of alcohol outlet studies cannot directly inform policy recommendations (Holmes et al. 2014), and the use of point-level data and a better understanding of how alcohol outlets cause harms may help bridge this gap. In addition, innovative translational study designs can help make complex scientific information more accessible to diverse audiences as part of the policy change process. To date, we are aware of only a few peer-reviewed analyses that prospectively modeled the effects of potential alcohol outlet zoning policies (Grubesic et al. 2012; Ahern et al. 2013; Trangenstein, Jennings, et al. 2018). In particular, simulation modeling is an innovative tool that has been applied to hours of retail alcohol sales (Atkinson et al. 2018) and could add greater nuance to these prospective estimates through feedback loops. Future research should continue to expand this body of research to help policymakers utilize this evidence to prevent health harms, including modeling the prospective use of other availability restrictions such as cumulative impact zones (de Vocht et al. 2015; de Vocht et al. 2016; De Vocht et al. 2017) and potential utility of a risk-based licensing approach (Mathews and Legrand 2013; Curtis et al. 2019; Nepal et al. 2019). Still other areas ripe for translational research are interpretations of spatial access methods. Researchers should consider ways to transform results in ways that are more approachable for lay populations.

Conclusions

Alcohol availability research has progressed remarkably since Smart’s 1977 “not-scientific” analyses of eight dimensions of alcohol availability. Researchers have successfully expanded a strict individual-level frame of alcohol-related harms to include the alcohol environment, which then opened the field to a range of new tools and techniques. The primary challenge facing alcohol outlet access researchers today is much the same as it was 30 years ago: how do we eliminate sources of measurement error? Incorporating street-level data when quantifying alcohol outlet access and defining alcohol outlet clusters are two important steps that can reduce measurement error and support more defensible conclusions. The alcohol policy climate today is different from the climate shortly after the repeal of prohibition. Those researchers often wanted to distance themselves from the temperance movement, and often aimed to produce results that did not have implications for how populations drink. Today, there are several examples of jurisdictions “modernizing” their alcohol outlet regulations in ways that are likely to (in the absence of other controls) increase related harms (Zhao et al. 2013; Tabb et al. 2016; Gorman D. M. et al. 2017). This growing trend should provide motivation for researchers to consider how to make their models more relevant to real-world debates.

Biographies

Biographical Notes

Dr. Trangenstein is an Assistant Professor at the University of North Carolina at Chapel Hill. Her research interests include measurement, alcohol policy, and alcohol-related harms to others.

Dr. Sadler is an urban geographer at Michigan State University with expertise in environmental science, GIS, food systems planning, and land use policy in legacy cities.

Dr. Morrison is a social epidemiologist with expertise in spatial analytic methods. His research seeks to understand how social and physical environmental conditions affect population health, particularly injuries and alcohol-related harms.

Dr. Jernigan is a leading expert in alcohol policy. He is best known for his action-research approach to the issue of alcohol advertising, marketing, and promotion and its influence on young people.