Alcohol outlet density and alcohol consumption in Los Angeles County and Southern Louisiana

Abstract

Objective

To assess the relationship between alcohol availability, as measured by the density of off-premise alcohol outlets, and alcohol consumption in Los Angeles county and southern Louisiana.

Method

Consumption information was collected through a telephone survey of 2881 households in Los Angeles County and pre-Katrina southern Louisiana nested within 220 census tracts. Respondents’ addresses were geocoded and both neighborhood (census tracts and buffers of varying sizes) and individual (network distance to the closest alcohol outlet) estimates of off-sale alcohol outlet density were computed.

Results

Alcohol outlet density was not associated with the percentage of people who were drinkers in either site. Alcohol outlet density was associated with the quantity of consumption among drinkers in Louisiana but not in Los Angeles. Outlet density within a one-mile buffer of the respondent’s home was more strongly associated with alcohol consumption than outlet density in the respondent’s census tract.

Conclusions

The relationship between neighborhood alcohol outlet density and alcohol consumption is complex and may vary due to differences in neighborhood design and travel patterns.

Introduction

Alcohol use and abuse and their consequences are significant problems in the United States. The CDC estimates 61% of adults drink alcohol (http://www.cdc.gov/nchs/fastats/alcohol.htm) and 32% of current drinkers had five or more drinks on at least one day in the past year. The availability of alcohol has been associated with drinking and driving (Treno et al., 2003), injury (Treno et al., 2001), motor vehicle crashes (Scribner et al., 1994), homicide (Scribner et al., 1999), and greater rates of assault (Lipton and Gruenwald, 2002; Scribner et al., 1995). There are over 20,000 alcohol-attributed deaths and additional 12,000 deaths due to alcoholic liver disease yearly (http://www.cdc.gov/nchs/fastats/alcohol.htm).

The availability of alcohol has also been linked with increased alcohol consumption. Surveying employees in a manufacturing plant, Ames and Grube 1999 found that alcohol availability at work was associated with greater consumption. Abbey et al. 1990 found that distance to the closest alcohol outlet was indirectly associated with alcohol consumption through perceived convenience of buying alcohol and subjective measures. In a sample of college students covering a wide range of outlet densities Weitzman et al. 2003 found that alcohol outlet density correlates with heavy drinking, frequent drinking and drinking related problems. A truncated range of outlet densities might make it harder to observe effects related to outlet density.

The majority of previous studies have utilized aggregate measures of alcohol consumption and alcohol availability exposure, for example, at the ZIP code or census tract level. Aggregate or ecological analyses do not control for individual demographic differences (Greenland, 2001)and inferences from aggregate data about individual behaviors can be misleading. This is known as the ecological fallacy (Robinson, 1950). Some recent studies of the relationship between alcohol availability and either consumption or alcohol-related health outcomes have combined characteristics at the individual and aggregate level in multilevel or hierarchical analyses.

The present study examined the relationship between alcohol availability and alcohol consumption across more than 200 neighborhoods in two distinct geographic areas. This study was designed to add to the existing literature in the following ways. First, we investigated measures of the quantity and frequency of ethanol consumption separately in two steps: we first investigated the association of availability and drinking status (drinker vs. nondrinker)and then estimated the association with ethanol consumption among drinkers only. Second, we further explored whether the quantity and frequency of drinking was more strongly associated with neighborhood exposures measured using census tract boundaries compared to buffers of varying distances from an individual’s home.

Materials and Methods

Data for these analyses come from the cross sectional study of alcohol availability, marketing, promotion, and consumption conducted in 220 census tracts in Los Angeles County and pre-Katrina Southern Louisiana (106 in Louisiana and 114 in Los Angeles). Data from the cross sectional a household level telephone survey were combined with data on the number of alcohol outlets to examine the hypothesized relationships. The phone survey was conducted from October 4, 2004 to August 28, 2005 in Louisiana and from October 4, 2004 to October 19, 2005 in Los Angeles County.

The sample frame consisted of urban census tracts in twenty-six contiguous parishes in southeastern Louisiana and within 20 miles of Charles R. Drew University of Medicine and Science in Los Angeles County covering most of the census tracts in Los Angeles, chosen so we could visit each tract. An urban census tract was defined to contain at least 2,000 persons per square mile. The U.S. census bureau considers a census block group urban if they have a population density of at least 1,000 persons per square mile and are adjacent to another census block group with a population density of at least 500. We drew a random sample of urban census tracts stratified by site.

Consumption survey

Within each sampled census tract we conducted a survey of households with listed telephone numbers to ask questions about various health behaviors, with particular emphasis on alcohol consumption. Households were sampled using a systematic sample of a household list provided by a commercial vendor to yield approximately 10 households per census tract. A list-based sample was chosen because, in addition to the phone numbers, it provided access to household addresses needed for geocoding. The respondent within each sampled household was selected using the most-recent-birthday method. Up to 25 contact attempts were made before a household was removed from the sample. Procedures were approved by the RAND IRB. The questionnaire contained 84 questions but due to skip patterns took only 15–20 minutes to complete. Advance letters were sent to all unique household addresses and a toll free phone number was provided if the respondent wanted to initiate the survey. Participants were informed that they would be sent a check of $15 upon completion of the survey.

We computed two consumption measures based on the survey—the average daily ethanol consumption based on the previous year and previous 90 days. The average daily ethanol consumption based on the previous 12 months was defined as:

$${\text{ethanol}}_{12\text{mth}}=\text{drinkdays}\ast \text{drinks}\ast 0.6/365$$

Where drink days is the number of drinking days per year, drinks the number of drinks per occasion and 0.6 the amount of ethanol in an average drink (Kerr et al., 2005). The number of drink days was computed from a question about drinking frequency with the levels “at least once a day”, “nearly every day”, “three or four times a week”, “once or twice a week”, “two or three times a month”, “about once a month” and “less once a month but at least once a year”, “none in the last 12 months.” The respondents was then asked how many drinks he/she usually had on those days when he/she drank alcohol in the last 12 months. This 12-month measure can also be interpreted as a quantity/frequency measure. The only difference is the multiplicative constant (0.6/365).

The average daily ethanol consumption based on the previous 90 days was computed as:

$${\text{ethanol}}_{90\text{days}}=\text{ethanol}\ast \text{ounces}\ast \text{drinks}\ast \text{drinkdays}90/90$$

where ethanol is the alcohol content of the respondent’s most common drink in the last 90 days, ounces the size of that drink in ounces, drinks the number of drinks per occasion and drink days 90 the number of days drinking within the last 90 days. The number of drinking days was computed from a question about drinking frequency with the categories, “Never”, “Less than once a month”, “About once a month”, “A few times a month”, “About once a week”, “Several times a week”, “Once a day”, “Several times a day”. We asked about the brand name of the beverage the respondent drank most often in the previous 90 days, the type of beverage (beer, wine, and 8 other categories), the specific type of container the respondent consumed this drink in, and how many ounces this type of container holds. The type of beverage was used to compute ethanol per ounce.

Alcohol Availability Measurement

Addresses of alcohol outlets were obtained from the California Department of Alcohol Beverage Control (ABC) and the Louisiana Department of Alcohol and Tobacco Control (ATC). The addresses were geocoded. We included in the analysis all outlets where alcohol purchased was for consumption off-premise. We restricted the analysis top off-premise outlets because of their association with problem drinking and community level morbidity and mortality ( ). We measured alcohol availability through the number of nearby alcohol outlets in several ways:

1. Network distance to the nearest outlet.

2. Number of outlets in the census tract in which the respondent lived.

3. Number of outlets in buffers around respondents’ homes.

We geo-coded the respondent’s home and computed the number of alcohol outlets within buffers of 0.1, 0.25, 0.5 and 1.0 mile radii. We chose these distances near half a mile because half a mile represents a 10–15 minute walk which is considered a maximum walking distance for some people (Lee et al., 2003; Truong and Sturm, 2003) ; we chose not to normalize the number of outlets by dividing by the number of roadway miles, square miles, or population size. In densely populated areas such a definition may result in a low density even when the individual lives very close to an alcohol outlet.

Missing values

The number of missing values was under 10% for all variables except for the brand name of the alcohol drink (45.8%). Therefore we imputed missing values multiple times to capture the variation between imputations. We imputed missing values 5 times using a Markov Chain Monte Carlo algorithm (Schafer, 1997). For indicator variables the imputed values were stochastically assigned 0 or 1 using Bernoulli sampling. The imputations rely on the usual missing at random assumption (MAR). The use of the MAR assumption is unverifiable but standard practice; we have found no reason to discredit this assumption. We used Rubin’s formula( Rubin, 1987)to combine the regression results from different imputations.

Statistical Analyses

Because the study was conducted in two very different geographic locations, we conducted all analyses separately by study location. To study how availability as measured by the number of outlets relates to alcohol consumption we first explored descriptive statistics on the number of outlets within census tracts as well as within radii of 0.1, 0.25, 0.5 and 1 mile around survey respondents ’ homes and identified which radius had the largest association with consumption.

We divided alcohol consumption into two components: 1) whether or not the respondent was a drinker and 2) if the respondent was a drinker, how much alcohol the respondent consumed. We adjusted for gender, age, race/ethnicity and income as they may be associated with consumption. Because education was correlated with income it was omitted as a covariate.

For the first component, we regressed an indicator variable of drinking status (drinker/non drinker) on covariates, for the second component we regressed each of the two measures of ethanol consumption on covariates. We used the log-transformed ethanol measurements. Because log values are harder to interpret, we used the model to compute sample-averaged predictions of daily ethanol consumption of two slightly altered data sets for each variable: for indicator variables we computed average predictions with the indicator 1) set to zero and 2) set to one in the entire sample. For continuous variables we set the variable to the 25^th and 75^th percentiles, respectively.

To account for possible correlation within tracts(e.g. due to overlapping buffers) we also explored a random intercept hierarchical model and a spatial regression using a power exponential correlation function within tract. Because tracts were not necessarily contiguous due to random sampling, our spatial regression also assumed that measurements for respondents in different tracts were uncorrelated.

For the hierarchical regression of ethanol consumption we computed the intraclass cluster correlation (ICC) to assess what fraction of the variation in consumption occurs at the census tract level. The ICC for a hierarchic allogistic regression is difficult to interpret because the distribution at the individual level is binomial and the distribution and the census tract level is normal. The variance at the individual level is a function of prevalence. Instead, we translated the census tract level variance into a median odds ratio (MOR) (Merlo et al., 2006). The MOR is a measure of variability and can be interpreted as the median increased risk in becoming a drinker by moving from one census tract to another. (The odds ratio is always formed as an increase, i.e. as a value greater than 1. For differences, this is analogous to considering absolute differences rather than differences). In the absence of a census tract-level effect the risk would be the same in all census tracts.

We used the Hosmer Lemeshow goodness of fit test and Pregibon’s link test (Pregibon, 1980)to assess model fit. We also examined residual plots and variance inflation factor to check for multicollinearity among independent variables.

We used SAS Version 9.1.3 for imputations, spatial autocorrelation models, and for hierarchical logistic regressions. All other analyses were conducted in Stata Version 9.

Results

Survey results

Data were collected in 220 census tracts – 114 from Los Angeles County, California and 106 southeastern Louisiana. Data collection in Louisiana was halted early due to Hurricane Katrina, resulting in a loss of data from 8 additional census tracts. The cooperation rate, the proportion of all cases interviewed of all eligible respondents ever contacted, was 76.2% in Los Angeles and 79.8% in Louisiana for the phone survey. The response rate, RR3 (AAPOR, 2006), which also takes into account people we were unable to reach, was 34.4% in Los Angeles and 37.9% in southern Louisiana. (Response rates of this magnitude have become common even among government surveys. Our results can be compared to the 2005 Behavioral Risk Factor Surveillance System (BRFSS) survey’s response rates in California which was 29.2% and to Louisiana which was 36.5%. [http://aspe.hhs.gov/hsp/06/Catalog-AI-AN-NA/BRFSS.htm])The number of phone survey respondents was 1578 in Los Angeles and 1303 in Louisiana. In Louisiana there was an average of 13.8 respondents per census tract. In Los Angeles there was an average of 12.3 respondents per census tract.

Table 1 gives sample characteristics by state. Approximately two-thirds of respondents were female. The samples differed with respect to race and ethnicity, with the Louisiana sample including approximately equal numbers of whites and African-Americans, and the Los Angeles sample more than one-third Hispanic. Separate analyses were conducted by state and adjusted for the sample characteristics.

Table 1.

Sample Demographic Characteristics by state

Variable	Louisiana	Los Angeles	p-Value
Age, mean sd(age)	44.1 (12.96)	41.9 (13.33)	0.000
Gender	-	-	0.002
Male, %	33.0	38.6	-
Female, %	67.0	61.4	-
Respondent’s income, %	-	-	0.047
≤15,000	24.3	22.9	-
≤ 25,000	16.3	18.8	-
≤ 50,000	24.6	21.4	-
≤75,000	14.2	13.3	-
>75,000	20.6	23.6	-
Race/Ethnicity	-	-	0.000
Non-Hispanic White, %	50.4	35.1	-
Non-Hispanic Black, %	42.1	16.3	-
Other, %	1.9	4.1	-
Asian, %	1.0	5.5	-
Hispanic, %	4.6	38.9	-

The mean number of off-premise alcohol outlets in buffers of various radii is shown in Table 2. The mean number of alcohol outlets ranged from approximately 0.25 outlets in a 0.1 mile radius to 20 in a 1 mile radius. Census tracts contained on average 3.3 outlets, corresponding in size to a buffer of between a ¼ mile and a ½ mile radius. The mean numbers of outlets in various categories in Los Angeles and Louisiana were comparable and not significantly different. However, the mean distance to the closest outlet in Louisiana was somewhat larger and significantly different.

Table 2.

Descriptive statistics of the number of alcohol outlets in buffers of varying radii around respondents, in census tract and mean distance to closest alcohol outlet by study site.

	Louisiana					Los Angeles
	Min	1st quartile	Mean	3rd quartile	Max	Min	1st quartile	Mean	3rd quartile	Max
Number of outlets in .1 mile buffer	0	0	0.19	0	4	0	0	0.17	0	4
Number of outlets in .25 mile buffer	0	0	1.33	2	16	0	0	1.27	2	8
Number of outlets in .5 mile buffer	0	0	5.41	8	43	0	2	5.30	8	19
Number of outlets in 1.0 mile buffer	0	6	10.58	27	115	0	10	19.67	28	63
Number of outlets in census tract	1	2	3.30	5	14	1	2	3.36	5	9
Distance to closest outlet (miles)	0	0.36	0.67	0.62	30.84	0	0.17	0.48	0.49	4.50

Drinkers vs. non Non-drinkers

The MOR is a measure of how much drinking status varies by census tract. The smallest MOR, 1, implies no variability in drinking status among census tracts; a large MOR implies a large variability. The unadjusted MOR for drinking status were 1.77(p =.0018) for Louisiana and 1.74 (p<.0001) for Los Angeles. This means that the median risk in becoming a drinker by moving from one census tract to another increased by 77% in Louisiana and 74% in Los Angeles. The adjusted median odds ratios was 1.32 (p=0.24) for Louisiana. In Los Angeles, however, the tract-level variance was zero after adjusting for covariates. This means that there was no longer any variation in drinking status after adjusting for covariates.

The corresponding random intercept logistic regressions in Table 3 show that after adjustment for individual-level covariates there was not a strong relationship between drinking status and outlet density. Even though the odds ratios for the number of alcohol outlets in a one mile radius were marginally statistically significant, the odds ratios themselves were very small(1.01 in Louisiana and 0.99 in Los Angeles). The pseudo R squared values for the models were 14.0% and 10.7%, respectively.

Table 3.

Adjusted odds ratios of drinking status (yes/no) for covariates separately for Louisiana and Los Angeles.

	Louisiana Coefficient	p	Los Angeles Coefficient	p
Outlets (1 mile buffer)	1.01	0.04	0.99	0.08
Age 18–20	0.39	0.03	0.50	0.31
Age 25–30	1.69	0.14	1.34	0.63
Age 30–40	1.38	0.25	0.90	0.84
Age 40–50	0.90	0.68	0.64	0.36
Age 50–60	0.58	0.04	0.51	0.16
Age 60–65	0.54	0.04	0.35	0.07
Male	1.84	0.00	1.79	0.00
Black (non-Hispanic)	0.45	0.00	0.64	0.01
Hispanic	0.60	0.10	0.46	0.00
Asian (non-Hispanic)	1.54	0.60	0.54	0.11
Other race (non-Hispanic)	0.92	0.86	0.48	0.09
Household income <15 K	1.15	0.49	1.32	0.39
Household income 25K–50K	1.71	0.00	2.85	0.00
Household income 50K–75K	2.63	0.00	3.33	0.00
Household income >75K	2.97	0.00	3.85	0.00
constant	1.41	0.24	1.65	0.22

Daily ethanol consumption among drinkers: unadjusted results

Table 4 shows the relationship between the number of alcohol outlets and the quantity of alcohol consumption among drinkers. Coefficients can be interpreted as the percent increase in ethanol consumption for one additional outlet and for a 10% increase in outlets. The increases in ethanol consumption for the 10% increases in outlets are comparable across buffers of different sizes. The increases in ethanol consumption for one additional outlet are not comparable because larger buffers contain far more outlets. In Louisiana, the density of outlets was associated with ethanol consumption. The largest effect size was seen with the 1-mile buffer; the coefficient here indicates that for each additional alcohol outlet within one mile a respondent’s residence the quantity of alcohol consumed over the previous 12 months increased by 0.9%. The number of alcohol outlets in the census tract and the distance to the closest outlet were not significantly associated with consumption. Based on these results we chose the 1-mile radius for the regressions below. No estimate was significant for Los Angeles, indicating no relationship between alcohol outlet density and the quantity of consumption.

Table 4.

Percent increase in daily ethanol consumption (90 day measure and 12 months measure) for (a) one additional alcohol outlet and (b) a 10% increase in the number of alcohol outlets based on unadjusted regressions for Louisiana and Los Angeles. (For the distance measure: 10% decrease in distance to the closest alcohol outlet).

	Louisiana						Los Angeles
	Ethanol (90 days)			Ethanol (12 months)			Ethanol (90 days)			Ethanol (12 months)
	increase per outlet	10% increase in # of outlets	p-value	increase per outlet	10% increase in # of outlets	p-value	increase per outlet	10% increase in # of outlets	p-value	increase per outlet	10% increase in # of outlets	p-value
Number of outlets in .1 mile buffer	2.7%	0.1%	0.841	3.4%	0.1%	0.719	13.7%	0.2%	0.256	9.6%	0.2%	0.319
Number of outlets in .25 mile buffer	8.6%	1.1%	0.022	7.1%	0.9%	0.003	2.5%	1.2%	0.576	−3.7%	−0.5%	0.284
Number of outlets in .5 mile buffer	2.6%	1.4%	0.019	2.5%	1.3%	0.001	−0.7%	1.3%	0.668	−2.2%	−1.2%	0.066
Number of outlets in 1.0 mile buffer	0.7%	0.8%	0.033	0.9%	1.0%	0.000	−0.1%	−7.2%	0.828	−0.4%	−0.8%	0.277
Number of outlets in census tract	1.8%	0.6%	0.549	−0.8%	−0.3%	0.757	−2.9%	−0.2%	0.425	5.4%	1.8%	0.067
Distance to closest outlet (miles)	3.4%	0.2%	0.337	1.3%	0.1%	0.651	−5.0%	−0.1%	0.615	−6.0%	−0.3%	0.453

Spatial Autocorrelation and intraclass cluster correlation

There was no evidence of spatial autocorrelation for any outcome for both Los Angeles and Louisiana. Even though different starting values were explored, the restricted maximum likelihood (REML) estimates of the range were estimated to be zero implying no spatial autocorrelation.

The intraclass correlations in Los Angeles for the 12-month ethanol consumption was 0.008 (p=0.30)and for the 90-day ethanol consumption was 0.049 (p=0.03). In Louisiana the intraclass correlation for the 12-month ethanol consumption was 0.029 (p=0.06)and for the 90-day ethanol consumption 0.0 (p=0.88). Intra class correlation here refers to correlation of ethanol consumption within clusters defined by census tracts. The largest and only significant correlation was for the 90 day measure in Los Angeles. We constructed a random intercept model for this outcome adjusted for the covariates as below, and found that the random intercept model is statistically insignificant from a linear regression model based on the corresponding likelihood ratio test(p=0.24). We therefore simply ran linear regression models on the log transformed responses going forward.

Daily ethanol consumption among drinkers: Adjusted results

The average predicted values based on the Gaussian regressions of ethanol consumption of drinkers are shown in Table 5. For each categorical variable we computed average predicted values for each level of that variable. For the number of outlets variable we computed average predictions for 6 outlets and for 25 outlets corresponding to the 25^th and 75^th percentiles of its distribution.

Table 5.

Model-based predictions of daily ethanol consumption (ounces) of respondents who drink for both measures of consumption (12-month measure and 90 day measure) by site (Louisiana and Los Angeles). The 25^th (75^th) percentile of the number of alcohol outlets correspond to 6 (27) outlets in Louisiana and 10 (28) outlets in Los Angeles.

	Louisiana				Los Angeles
	90-day question		12-month question		90 -day question		12-month question
	Daily ethanol (oz)	p-value	Daily ethanol (oz)	p-value	Daily ethanol (oz)	p-value	Daily ethanol (oz)	p-value
Outlets (1 mile buffer)
6 outlets (25th %tile)	0.072	0.035	0.133	0.001	0.061	0.290	0.131	0.960
25 outlets (75th%tile)	0.084		0.156		0.055		0.131
Age
age 18–20	0.039	0.140	0.105	0.230	0.054	0.780	0.128	0.570
age 20–25	0.098	NA	0.171	NA	0.063	NA	0.162	NA
age 25–30	0.083	0.690	0.198	0.590	0.047	0.360	0.140	0.500
age 30–40	0.086	0.730	0.154	0.640	0.056	0.700	0.142	0.480
age 40–50	0.091	0.840	0.185	0.720	0.070	0.710	0.147	0.610
age 50–60	0.070	0.350	0.115	0.077	0.054	0.560	0.104	0.025
age 60–65	0.074	0.490	0.119	0.150	0.056	0.730	0.113	0.110
Race/Ethnicity
White (non-Hisp)	0.106	NA	0.186	NA	0.082	NA	0.170	NA
Black (non-Hisp)	0.056	0.002	0.105	0.000	0.063	0.250	0.116	0.012
Hispanic	0.048	0.050	0.111	0.038	0.035	0.000	0.101	0.000
Asian (non-Hisp)	0.022	0.029	0.088	0.110	0.045	0.078	0.099	0.019
Other (non-Hisp)	0.049	0.180	0.135	0.390	0.044	0.180	0.120	0.200
Gender
female	0.051	NA	0.103	NA	0.038	NA	0.094	NA
male	0.172	0.000	0.270	0.000	0.096	0.000	0.201	0.000
Household Income
<15 K	0.073	NA	0.160	NA	0.035	NA	0.121	NA
15K–25K	0.101	0.320	0.187	0.420	0.037	0.830	0.099	0.280
25K–50K	0.073	1.000	0.130	0.230	0.062	0.022	0.136	0.490
50K–75K	0.079	0.790	0.150	0.740	0.060	0.067	0.107	0.550
>75K	0.087	0.510	0.141	0.490	0.083	0.001	0.171	0.064

In Louisiana, the number of outlets was significantly associated (p<0.001) with the 12-month ethanol consumption measure and for the 90-day ethanol consumption measure (p=0.04). The counterfactual simulation shows that increasing the number of outlets from the 25^th percentile to the 75^th percentile is associated with an increase of ethanol consumption of about 17% for both the 12 months (from 0.133 to 0.156) and the 90 day measurement(from 0.072 to 0.084). This increase in exposure corresponds to an additional 14 drinks per year for each drinker in Louisiana. The percent increase from the 25^th to the 75^th percentile is consistent with the increase per additional outlet from the unadjusted results in Table 4.

For Los Angeles the coefficients for the number of outlets are not significant and the coefficients are much smaller. The regressions explain between 10% and 15% of the variation.

Discussion

We found an association between alcohol outlet density and consumption after controlling for demographics among drinkers in Louisiana. However, we found no such association in Los Angeles. Density could be associated with consumption through the frequency of exposure to cues relating to alcohol, including the presence of alcohol itself. The design of Louisiana neighborhoods may differ from Los Angeles neighborhoods significantly, and given the same alcohol outlets densities, could still yield considerably different exposures to outlets.

The models for Los Angeles also tended to explain less variation. One possible explanation is that the association between availability and consumption is more complex in Los Angeles. Residents of Los Angeles may drive more than residents in Louisiana and access to alcohol may be less affected by the number of outlets in their neighborhood.

Our findings add to the inconsistencies in the literature on the association between alcohol availability and consumption. Using a random intercept model for neighborhoods, Pollack et al. 2005 did not find any association between alcohol availability and heavy drinking. The analysis might have had reduced power as both the outcome and the availability measured (outlet density, nearest distance to outlet, 0.5 mile buffer) were dichotomized. Using individual level and census tract data in California, Truong and Sturm 2007 found intraclass cluster correlations of less than 5% for dichotomous measures of heavy drinking and decided to use logistic regression rather than hierarchical models. They found that off-sale outlets were not related to problem drinking. Using individual and county level data in California, Gruenewald et al. 2002 found that restaurant densities were directly related to greater drinking frequencies and drinking after driving, whereas bar densities were inversely related to drinking after driving.

We found that the association between alcohol density and consumption is stronger when using buffers around individuals than either distance to the closest outlet or the number of outlets in the census tract. Specifically, results suggest that the association is larger for 1 mile buffers than for smaller buffers. This is consistent with the findings in Scribner et al., 2000. Using individual level and census tract data in New Orleans, Scribner et al. 2000 found that neighborhood alcohol availability predicted alcohol consumption better than individual availability. This suggests that density of alcohol outlets is a better measure of alcohol availability than distance to the nearest alcohol outlet.

This study is not without limitations. First, this is a cross-sectional study and while associations found provide important clues, they are not necessarily causal. It is possible that there are more alcohol outlets near persons who drink more to meet the increased demand by those drinkers. Second, ethanol consumption is self-reported rather than objectively measured. While not perfect, there is some evidence (Gruenwald and Johns on, 2006), however, that a self-reported quantity-frequency measure based on the previous month of drinking exhibits good reliability (0.76). Third, survey non-response potentially introduced bias. The danger of non-response bias is somewhat mitigated because non-response here is mostly due to non-contact rather than due to non-cooperation which can be more selective. For establishing associations (rather than prevalences) the bias is also mitigated through the presence of regression covariates.

In summary, findings illustrate important limitations in determining exposures to alcohol based upon geographic proximity and outlet density. The main purpose was to assess the magnitude of the association between exposure and drinking, so that policy-relevant solutions can be identified that can reduce the negative consequences of alcohol use. Varying geographies and travel patterns make this investigation very complex, and ultimately, more sophisticated technologies that objectively measures exposures based upon actual interactions will be needed. Furthermore, other exposures that potentially influence drinking including print ads, billboards, and television viewing where the alcohol industry expends a significant amount of resources need to be considered. Developing a model that would take the full scope of exposures into account will be helpful in isolating the most significant exposures that lead to alcohol consumption and alcohol-related morbidity and mortality.