Does the response to alcohol taxes differ across racial/ethnic groups? Some evidence from 1984-2009 Behavioral Risk Factor Surveillance System

Abstract

Background

Excessive alcohol use remains an important lifestyle-related contributor to morbidity and mortality in the U.S. and worldwide. It is well documented that drinking patterns differ across racial/ethnic groups, but not how those different consumption patterns would respond to tax changes. Therefore, policy makers are not informed on whether the effects of tax increases on alcohol abuse are shared equally by the whole population, or policies in addition to taxation should be pursued to reach certain sociodemographic groups.

Aims of the Study

To estimate differential demand responses to alcohol excise taxes across racial/ethnic groups in the U.S.

Methods

Individual data from the Behavioral Risk Factor Surveillance System 1984-2009 waves (N= 3,921,943, 39.3% male; 81.3% White, 7.8% African American, 5.8% Hispanic, 1.9% Asian or Pacific Islander, 1.4% Native American, and 1.8% other race/multi-race) are merged with tax data by residential state and interview month. Dependent variables include consumption of any alcohol and number of drinks consumed per month. Demand responses to alcohol taxes are estimated for each race/ethnicity in separate regressions conditional on individual characteristics, state and time fixed effects, and state-specific secular trends.

Results

The null hypothesis on the identical tax effects among all races/ethnicities is strongly rejected (P < 0.0001), although pairwise comparisons using t-test are often not statistically significant due to a lack of precision. Our point estimates suggest that the tax effect on any alcohol consumption is largest among White and smallest among Hispanic. Among existing drinkers, Native American and other race/multi-race are most responsive to tax effects while Hispanic least. For all races/ethnicities, the estimated tax effects on consumption are large and significant among light drinkers (1-40 drinks per month), but shrink substantially for moderate (41-99) and heavy drinkers (≥ 100).

Discussion

Extensive research has been conducted on overall demand responses to alcohol excise taxes, but not on heterogeneity across various racial/ethnic groups. Only one similar prior study exists, but used a much smaller dataset. The authors did not identify differential effects. With this much larger dataset, we found some evidence for different responses across races/ethnicities to alcohol taxes, although we lack precision for individual group estimates. Limitations of our study include the absence of intrastate tax variations, no information on what type of alcohol is consumed, lack of controls for subgroup baseline alcohol consumption rates, and measurement error in self-reported alcohol use data.

Implications for Health Policies

Tax policies aimed to reduce alcohol-related health and social problems should consider whether they target the most harmful drinking behaviors, affect subgroups in unintended ways, or influence some groups disproportionately. This requires information on heterogeneity across subpopulations. Our results are a first step in this direction and suggest that there exists a differential impact across races/ethnicities, which may further increase health disparities. Tax increases also appear to be less effective among the heaviest consumers who are associated with highest risk.

Implications for Further Research

More research, including replications in different settings, is required to obtain better estimates on differential responses to alcohol tax across races/ethnicities. Population heterogeneity is also more complex than our first cut by race/ethnicity and needs more fine-grained analyses and model structures.

Introduction

Excessive alcohol use remains an important lifestyle-related contributor to morbidity and mortality in the U.S. and worldwide (World Health Organization, 2004). Both pricing and access policies have been used to reduce consumption. Internationally, current alcohol pricing policies include targeted taxation, inflation-linked taxation, taxation based on alcohol-by-volume, minimum pricing policies, bans of below-cost selling, and restricting price-based promotions (Meier et al., 2010). Recent meta-analysis shows significant effects of pricing policies on reducing alcohol-related disease and injury rates (Wagenaar et al., 2010). Access policies have expanded to regulating the physical availability of alcohol, modifying the drinking context, education and persuasion strategies, early intervention services, etc (Babor et al., 2010).

There is substantial literature on the relationship between alcohol consumption and taxes or prices at the overall population level, but recent systematic reviews (Chaloupka et al., 2002; Fogarty, 2006; Gallet, 2007; Wagenaar et al., 2009) find that the detailed underlying effects of price changes on different population subgroups remain understudied. Thus, policy makers are not informed whether the tax policies to reduce alcohol-related health and social problems target the most harmful drinking patterns, affect subgroups in unintended ways, or influence some groups disproportionately (Meier et al., 2010).

This study look at the heterogeneous effect of alcohol taxes on consumption across various races/ethnicities among U.S. adults. If the effects of tax increases on alcohol consumption differ systematically, additional or different policies (which may include alternative taxation schemes or access policies) could be pursued to reach specific subpopulation groups (Chaloupka et al. 2002). To our knowledge, the only study of heterogeneity across sociodemographic groups was conducted by Saffer and Chaloupka (1999). They reported that price effects were similar between racial/ethnic subgroups, although this null finding could be a consequence of a relatively small sample size (about 49,000 observations) for this type of question. This study revisits the question with a sample size that is almost two orders of magnitudes larger (about 4 million observations).

The main obstacle to estimate heterogeneous price/tax effects is typically sample size, because cells quickly become too small for precise enough estimates for minority groups. Lack of variation in the explanatory variables can be another limitation. Many surveys on individuals’ drinking behavior and prices (Bluthenthal et al., 2008; Casswell and Gilmore, 1989; Gyimah-Brempong, 2001; Harwood et al., 2003; Herttua et al., 2008; Holder and Blose, 1987; Lange et al., 2002; Treno et al., 1993; Williams et al., 2005) were geographically limited (i.e., within several census tracts, a city, or county), which reduces price/tax variations. Studies using aggregate-level alcohol consumption data (Duffy, 1983; Levy and Sheflin, 1983; McGuinness, 1980; Ornstein and Levy, 1983; Uri, 1986) do not observe the amount consumed by subgroups.

Established in 1984 by the Centers for Disease Control and Prevention, the Behavioral Risk Factor Surveillance System (BRFSS) is a state-based system of random-digit-dial surveys that collects information on health risk behaviors, preventive health practices, and health care access, for adults aged 18 or older. Its large sample size (about 4 million for our analysis) over a long period (26 years from 1984 to 2009), consistent survey questions on alcohol use and individual characterises, and nationwide coverage, meet the requirement of our subgroup analysis across various races/ethnicities.

Two previous BRFSS analyses are particularly relevant to our work (Decker and Schwartz, 2000; Ruhm and Black, 2002). Using 1985-1993 waves, Decker and Schwartz (2000) applied two-part model to estimate both own- and cross-price elasticities for demand in cigarette and alcohol. They found that higher alcohol prices decreased both alcohol consumption and smoking participation, while higher cigarette prices seemed to decrease smoking participation but increase drinking. Only gender-based heterogeneity in price responsiveness was considered. We consider heterogeneous effect across races/ethnicities in this paper. Additional extension is that we use tax rather than price to avoid endogeneity, and control for state fixed effects and state specific secular trends.

Ruhm and Black (2002) examined the relationship between macroeconomic conditions and drinking using individual-level data from the BRFSS 1987-1999 waves. Federal and state beer excise taxes were used as a key control to account for the responses of alcohol consumption to changes in tax rates. They found a procyclical variation in overall drinking largely resulting from decrease in consumption among existing drinkers during recessions. The effect of state unemployment rate on alcohol use was estimated for subgroups (White, African American, Hispanic, and other race). Non-Whites were more likely to stop drinking during economic downturns, and Hispanics exhibited particularly large cyclical variations. Our study follows Ruhm and Black’s (2002) analytic framework, although with a different scientific goal, namely to investigate the differential tax effects on drinking across races/ethnicities. We use individual-level measures of employment and income as we are not studying the effect of business cycles.

Two studies using other data sources addressed related questions to ours. Using data from the Health and Retirement Study and finite mixture model (FMM), Ayyagari et al. (2009) recovered two latent groups which differed in their response to price. The more responsive group was more likely to be non-White, female, married and older; and to consume less alcohol. The FMM can identify multi-dimensional latent groups. However, it is much less clear how such latent groups relate to actual population subgroups of interest to policy makers. The number of groups in FMM has to be specified and this may require judgment decisions as it is typically difficult to identify more than a small number of all latent groups (Deb, 2008). The opposite approach is to start investigation by stratifying the data by demographic characteristics. Meier et al. (2010) stratified by age, gender, and level of drinking in data from the U.K. and found that age, sex, and level of drinking affected beverage preferences, drinking location, prices paid, price sensitivity, and tendency to substitute for other beverage types. Light drinkers were found to be more price-sensitive compared to heavy drinkers. In this paper, we estimate the heterogeneous tax effects across subpopulations stratified by race/ethnicity in the U.S.

Methods

Participants

Individual-level data come from the BRFSS 1984-2009 waves. The dependent variables are consumption of any alcoholic beverages (i.e., drinking participation) and number of standard drinks (one standard drink is any drink that contains 14 grams of pure alcohol, such as 12-oz of beer or 5-oz of wine) consumed per month calculated from the responses to the following three questions: “One drink is equivalent to a 12-ounce beer, a 5-ounce glass of wine, or a drink with one shot of liquor. During the past 30 days, have you had at least one drink of any alcoholic beverage such as beer, wine, a malt beverage or liquor?”, “During the past 30 days, how many days per week or per month did you have at least one drink of any alcoholic beverage?”, and “During the past 30 days, on the days when you drank, about how many drinks did you drink on the average?”

Individual characteristics from the survey used in the regression analysis as covariates are: gender, age (age in years and age squared), races/ethnicities (indicator variables for White, African American, Hispanic, Asian or Pacific Islander, Native American, and other race/multi-race), annual household income (indicator variables for annual income at the lowest quartile, the mid-low quartile, the mid-high quartile, and the highest quartile), education (indicator variables for education lower than high school, high school graduate, education higher than high school but lower than college, and college graduate or higher), employment status (indicator variables for employed, unemployed for less than one year, unemployed for more than one year, retired, disabled, student, and homemaker), and marital status (indicator variables for married, divorced/widowed/separated, and never married).

In the 26 waves of BRFSS from 1984 to 2009, a total of 4,605,214 individuals were interviewed, among whom 103,819 (2.3%) have missing values in one or more of the covariates (e.g., education, marital status) and are excluded from our analysis. The remaining total sample size included in our models is 3,921,943 (39.3% are male). The racial/ethnic composition is White (81.3%), African American (7.8%), Hispanic (5.8%), Asian or Pacific Islander (1.9%), Native American (1.4%), and other race/multi-race (1.8%). These are the sample statistics which are unweighted. In the analysis, we follow Ruhm and Black (2002) and use the BRFSS final sampling weights for descriptive statistics and regression models. Information on alcohol use is available for all respondents except in 1994, 1996, 1998, and 2000, when these questions are in optional modules fielded in only a subset of the states (12, 17, 12, and 12 states respectively). Those four years’ data account for 3.2% of the sample size. We have estimated drinking participation and conditional drinking with and without the data in 1994, 1996, 1998, and 2000. Because the estimated tax elasticities are very similar, we use data of all years from 1984 to 2009 in our models shown in Table 3-5.

Table 3.

Predicted tax effects on drinking participation and conditional drinking across alternative model specifications

Outcome	Model 1	Model 2	Model 3	Model 4	Model 5	Model 6	Model 7	Model 8	Model 9	Model 10
Any drinka	−0.0338 (−9.18)	−0.0391 (−9.50)	−0.0648 (−11.43)	−0.0626 (−9.68)	0.0041 (0.19)	−0.0395 (−9.34)	−0.0462 (−9.80)	−0.0736 (−11.46)	−0.0711 (−9.72)	0.0060 (0.25)
Log of number of drinksb	−0.4855 (−37.08)	−0.5106 (−36.14)	−0.4169 (−17.58)	−0.4586 (−17.45)	0.1084 (1.62)
State fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Month fixed effects	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
National linear time trendc	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
State-specific linear time trendd	No	Yes	No	Yes	Yes	No	Yes	No	Yes	Yes
Sampling weightse	No	No	Yes	Yes	Yes	No	No	Yes	Yes	Yes
Year fixed effects	No	No	No	No	Yes	No	No	No	No	Yes
Model	OLS	OLS	OLS	OLS	OLS	Probit	Probit	Probit	Probit	Probit

Note: (1) All models use datafrom BRFSS 1984-2009 waves.

(2) The sample size for drinking participation (i.e., any drink) is 3,921,943 and for conditional drinking (i.e., log of number of drinks) 1,939,550.

(3) All models control for individual characteristics listed in Table 2, namely gender, age, race/ethnicity, annual income, education, employmentstatus, and marital status.

(4) In the models for drinking participation, the treatment variable is the sum of federal and state excise taxes pergallon of beer. In the models for conditional drinking, the treatment variable is the natural logarithm of the sum of federal and state excise taxes per gallon of beer.

(5) OLS and probit models are used to estimate drinking participation. Marginal effects at the mean of the independent variables are reported for the probit models to ease the interpretation of coefficients and make them comparable to their OLS counterparts.

(6) OLS models are used to estimate conditional drinking. Due to the logarithmic transformation for the dependent variable and the treatment variable (i.e., tax), the coefficients can be interpreted as elasticity, namely percentage change in alcohol consumption given 1% change in tax.

(7) Eicker-Huber-White sandwich estimator is used to calculate robust standard errors clustered at year-month-state cell because all BRFSS respondents in that cell face the same tax rate.

(8) T-statistic is in parenthesis.

^aA dichotomous variable which denotes drinking any alcoholic beverages in the last 30 days.

^bNatural logarithm of number of standard drinksconsumed per month. A standard drink is any drink that contains 14 grams of pure alcohol, such as 12-oz of beer or 5-oz of wine.

^cH (months elapsed since January 1st, 1984 till December 31st, 2009) + U (adjusted monthly national unemployment rate).

^dInteractions between state dichotomous variables and H.

^eBRFSS final sampling weights.

Table 5.

Predicted tax effects on light, moderate, and heavy drinking across races/ethnicities

Subgroup	Number of standard drinksaconsumed in last 30 days
	1 - 20	1 - 30	1 - 40	21 - 59	31 - 79	41 - 99	≥60	≥80	≥100
Total population	−0.2512 (−18.48)	−0.3029 (−20.54)	−0.3204 (−20.33)	−0.0120 (−1.35)	−0.0616 (−6.76)	0.0058 (0.58)	−0.0194 (−0.69)	−0.0765 (−2.95)	−0.0148 (−0.49)
White (non-Hispanic)	−0.2508 (−17.01)	−0.3149 (−19.55)	−0.3381 (−19.61)	−0.0074 (−0.79)	−0.0541 (−5.79)	0.0115 (1.11)	−0.0328 (−1.18)	−0.0901 (−3.36)	−0.0325 (−1.00)
African American (non-Hispanic)	−0.2024 (−3.87)	−0.1966 (−3.57)	−0.1662 (−2.90)	−0.0554 (−1.68)	−0.1315 (−3.44)	−0.0353 (−0.88)	0.0022 (0.02)	0.0278 (0.24)	0.0826 (0.73)
Hispanic	−0.2174 (−4.27)	−0.2449 (−4.47)	−0.2780 (−4.80)	−0.0251 (−0.76)	−0.0715 (−2.04)	0.0025 (0.07)	0.1157 (1.51)	−0.0189 (−0.25)	0.0684 (0.77)
Asian or Pacific Islander (non-Hispanic)	−0.4697 (−5.04)	−0.4171 (−4.01)	−0.3217 (−2.99)	−0.0290 (−0.36)	−0.1910 (−2.37)	.01739 (0.17)	−0.0414 (−0.25)	0.0181 (0.11)	−0.0583 (−0.33)
Native American (non-Hispanic)	−0.4257 (−2.41)	−0.5254 (−2.95)	−0.5014 (−2.75)	−0.0382 (−0.45)	−0.1076 (−1.23)	−0.0607 (−0.59)	0.0255 (0.11)	−0.0992 (−0.39)	−0.3761 (−1.48)
Other race or multi-race (non-Hispanic)	−0.5048 (−2.92)	−0.4887 (−2.76)	−0.4426 (−2.41)	0.0852 (0.73)	−0.1952 (−2.18)	−0.1302 (−1.17)	−0.4074 (−1.65)	−0.6353 (−1.99)	−0.1224 (−0.35)

Note: (1) All models use datafrom BRFSS 1984-2009 waves.

(2) All models control for individual characteristics (i.e., gender, age, annual income, education, employment status, and marital status), state and month fixed effects, and national and state-specific time trends.

(3) All models are OLS weighted by BRFSS final sampling weights.

(4) Due to the logarithmic transformation for the dependent variable (i.e.,number of drinks consumed in last 30 days) and the treatment variable (i.e., the sum of federal and state taxes per gallon of beer), the coefficients can be interpreted as elasticity, namely percentage change in alcohol consumption given 1% change in tax.

(5) Eicker-Huber-White sandwich estimator is used to calculate robust standard errorsclustered at year-month-state cell because all BRFSS respondents in that cell face the same tax rate.

(6) T-statistic is in parenthesis.

^aA standard drink is any drink that contains14 grams of pure alcohol, such as 12-oz of beer or 5-oz of wine.

Alcohol excise tax

The primary explanatory variable is the sum of the monthly federal and state excise taxes per gallon of beer during the period 1984-2009, using data from the Federation of Tax Administrators and various issues of the U.S. Brewers’ Association Brewers’ Almanac. The taxes are converted to the December 2009 US$ using the all-items consumer price index from the U.S. Bureau of Labor Statistics.

Data analytic procedures

Our basic regression equation follows Ruhm and Black (2002) with a few modifications - a finer categorization of races/ethnicities and employment status, controlling for individual annual income and national unemployment rate, and inclusion of national and state-specific linear time trends. It is a two-part model with first part Probit and second part Ordinary Least Squares (OLS). The tax effects on drinking participation (i.e., any alcohol consumption) are estimated in the first part using probit models. Percentage change in drinking participation is calculated at the weighted sample mean. The tax effects on conditional drinking (i.e., alcohol consumption among drinkers) are estimated in the second part using OLS. Both the dependent variable (i.e., alcohol consumption) and the treatment variable (i.e., tax) are in logarithmic form so that the coefficient of tax can be interpreted as elasticity, namely percentage change in alcohol consumption given 1% change in tax. The two-part model takes the form:

$$\begin{array}{cc}\hfill & \text{Part}\mathrm{I}:\mathrm{Pr}(A_{\mathit{ismy}}>0\mid T,X,\alpha ,\lambda ,m,y)=\Phi (T_{\mathit{ismy}}\theta +X_{\mathit{ismy}}\beta +{\alpha }_s+{\lambda }_m+g(m,y\left)\right);\text{and}\hfill \\ \hfill & \text{Part}\mathrm{II}:\log \left(A_{\mathit{ismy}}\right)=\log \left(A_{\mathit{ismy}}\right)\theta +X_{\mathit{ismy}}\beta +{\alpha }_m+g(m,y)+{∊}_{\mathit{ismy}}\text{if}{A}_{\mathit{ismy}}>0\hfill \end{array}$$

where A is the alcohol outcome for individual i residing in state s interviewed in month m of year t, T the sum of federal and state beer excise taxes, X the vector of individual characteristics, α and λ the unobserved determinants of alcohol use associated with the state and survey month, g(m, y) the function of national and state-specific time trends, and ε the individual error term.

The state fixed effects control for time-invariant determinants that differ across states, and the month fixed effects hold constant the seasonal variations in alcohol consumption. Function g controls for various national and state-specific secular trends which may confound with our estimates of tax effects. It is specified as the following linear form

$$g(m,y)=H_{my}+{\alpha }_s{H}_{my}+U_{my}$$

where H registers months elapsed since January 1^st, 1984 till December 31^st, 2009, αH controls for factors that vary over time within-state, and U denotes adjusted monthly national unemployment rate reported by the U.S. Bureau of Labor Statistics. Robustness check using quadratic form of function g (i.e., linear form of g plus H_my²+α_sH_my²) was performed. It provides very similar estimate for the tax effect on drinking participation as the linear trend but differs in predicting conditional drinking. Model using linear time trend predicts that 1% rise in excise tax reduces drinking participation and conditional drinking by 0.14% (t = −9.72) and 0.46% (t = −17.45), while model using quadratic time trend predicts 0.12% (t = −6.65) and 0.10% (t = −3.30). It seems that the inclusion of quadratic terms absorbs much of the remaining variations in tax rates, and thus the predicted tax effect on conditional drinking is substantially attenuated. Attenuation of similar magnitude takes place in all subgroup analyses. We adopt the linear time trend specification in our analysis because the approach is consistent with Ruhm and Black (2002). It also turns out that the elasticity estimate (−0.46) is within the range reported in recent meta-analyses (Gallet, 2007; Wagenaar et al., 2009).

Eicker-Huber-White sandwich estimator is used to calculate robust standard errors clustered at year-month-state cell because all BRFSS respondents in that cell face the same tax rate. Following Ruhm and Black (2002), all models use the BRFSS final sampling weights to account for the differential probabilities for individual observations to be selected, and thus we estimate the population means.

Results

Figure 1 illustrates the change in federal and state beer excise taxes and drinking participation from 1984 to 2009. The tax usually declines in real terms because its nominal value is fixed and there is inflation. The single increase in the federal tax over the 26 years took place in 1991, when the Congress doubled the federal beer excise tax to $18 per barrel ($0.53 per gallon). Until 2009, 36 out of 51 U.S. states had not changed their beer excise tax for the last 20 years. This limited variation in the primary explanatory variable is obviously a constraint in what can be estimated. The average drinking participation rate among the national population was 52% but exhibited some fluctuations. Although the correlation coefficient between alcohol consumption and tax is −0.25, alcohol consumption seemed to decline from 1984 to 1989 and from 2004 to 2009. Since both individual characteristics and secular trends in drinking (Nephew et al., 1999) may confound with the relationship between alcohol consumption and tax, this should be dealt with in the multivariate analysis by controlling for personal characteristics and national/state-specific time trends. Using state fixed effects, we explore the change in individual’s alcohol consumption driven by the change in tax rate within each state. For all states, the overtime variations of taxes resemble the national pattern illustrated in Figure 1. The average intrastate difference between the highest and lowest tax rate during 1984-2009 is 0.55 US$, with the maximum 0.94 in Alaska and minimum 0.46 in Nebraska.

Figure 1. Beer excise tax and drinking participation among U.S. adults 1984-2009.

Note: (1) Federal and state excise taxes per gallon of beer are converted to 2009 US$ using all-items consumer price index and weighted by state population of relevant years. (2) Drinking participation is weighted by BRFSS final sampling weights.

Table 1 provides descriptive statistics for drinking participation and alcohol consumption stratified by race/ethnicity in 1984-2009. There are large differences in the number of standard drinks consumed per month across subgroups. Table 2 shows the sample means for the beer excise taxes and individual characteristics. The statistics are adjusted by the BRFSS final sampling weights.

Table 1.

Alcohol consumption by race/ethnicity 1984-2009

Race/ethnicity	Number of observations	Drinking participationa	Number of drinksb consumed per month among drinkers	Heavy drinking participationc
Total population	3,921,943	0.523 (0.499)	22.59 (43.67)	0.017 (0.129)
White	3,192,830	0.551 (0.497)	22.64 (40.59)	0.017 (0.130)
Africa American	304,762	0.401 (0.490)	20.22 (44.20)	0.013 (0.112)
Hispanic	227,642	0.467 (0.499)	24.00 (56.39)	0.020 (0.139)
Asian and Pacific Islander	74,011	0.431 (0.495)	15.61 (32.99)	0.009 (0.093)
Native American	53,157	0.430 (0.495)	32.68 (70.59)	0.029 (0.169)
Other race/multi-race	69,541	0.498 (0.500)	27.65 (87.42)	0.023 (0.150)

Note: (1) Standard deviation is in parenthesis.

(2) Statistics are weighted by BRFSS final sampling weights.

^aDrinking participation is a dichotomous variable which denotes drinking any alcoholic beverages in the last 30 days.

^bA standard drink is any drink that contains 14 grams of pure alcohol, such as 12-oz of beer or 5-oz of wine.

^cHeavy drinking participation is a dichotomous variable which denotes having 100 or more drinks in the last 30 days.

Table 2.

Sample means for independent variables

Variable	Attribute	Weighted Mean	Standard deviation
Tax
Sum of federal and state excise taxes per gallon of beer	continuous	0.972	0.289
Natural logarithm of taxes	continuous	−0.067	0.270
Gender
Male	dichotomous	0.480	0.500
Age
Age in years	continuous	44.849	17.658
Age in years squared	continuous	2323.215	1749.424
Race/ethnicity
White (non-Hispanic)	dichotomous	0.753	0.431
African American (non-Hispanic)	dichotomous	0.094	0.292
Asian or Pacific Islander (non-Hispanic)	dichotomous	0.025	0.157
Native American (non-Hispanic)	dichotomous	0.009	0.093
Other race or multi-race (non-Hispanic)	dichotomous	0.014	0.117
Hispanic	dichotomous	0.105	0.306
Annual household income
Income in the lowest quartile	dichotomous	0.234	0.423
Income in the mid-low quartile	dichotomous	0.245	0.430
Income in the mid-high quartile	dichotomous	0.331	0.471
Income in the highest quartile	dichotomous	0.190	0.395
Education
Education lower than high school	dichotomous	0.138	0.345
High school graduate	dichotomous	0.317	0.465
Education higher than high school but lower than college	dichotomous	0.259	0.438
College graduate or higher	dichotomous	0.286	0.452
Employment status
Employed	dichotomous	0.619	0.486
Unemployed for less than one year	dichotomous	0.031	0.174
Unemployed for greater than one year	dichotomous	0.021	0.144
Retired	dichotomous	0.167	0.373
Disabled	dichotomous	0.030	0.171
Student	dichotomous	0.046	0.209
Homemaker	dichotomous	0.086	0.280
Marital status
married	dichotomous	0.634	0.482
divorced or widowed or separated	dichotomous	0.178	0.383
never married	dichotomous	0.188	0.391

Note: (1) Data are from BRFSS 1984-2009 waves.

(2) Sample means are weighted using BRFSS final sampling weights.

Table 3 shows the estimated tax effects on drinking participation and conditional drinking among the whole population using alternative econometric models. All models are performed with StataSE 11.0 software (StataCorp, College Station, TX), controlling for individual characteristics listed in Table 2, state and month fixed effects, and national linear time trend (months elapsed since January 1st, 1984 and adjusted monthly national unemployment rate).

In Table 3, both OLS and probit models are used in estimating the tax effects on the probability of drinking (i.e., drinking participation). Marginal effects at the mean of the independent variables are reported for the probit models to ease the interpretation of coefficients and make them comparable to their OLS counterparts. Three points are noteworthy. First, the OLS and probit models produce very similar results. Second, the estimated tax effects are reasonably similar across alternative model specifications, namely with or without sampling weights and/or with or without controlling for national and state-specific time trends. Third, after the inclusion of year fixed effects, both the OLS and the probit coefficient shrink dramatically, become insignificant from zero, and even reverse the sign (i.e., positive). This suggests that adding those year indicator variables absorbs much of the remaining variations in intrastate tax rates, and thus the predicted tax effects are considerably attenuated. As Figure 1 indicates, during the entire period under study, the overtime variations in excise tax are limited and primarily driven by deflation, which may be well captured by the general year effects. The fundamental issue is that, even with 26 years of data, variations in tax are largely accounted for by the inclusion of year, month, and state fixed effects, and national and state-specific secular trends.

In Table 3, OLS models are applied to estimate the tax effects on conditional drinking. With logarithmic transformation for the dependent variable (i.e., number of standard drinks consumed in the past 30 days) and the treatment variable (i.e., the sum of federal and state tax per gallon of beer), the coefficients of tax can be interpreted as elasticity. Analogous to the case of drinking participation, the year fixed effects (Model 5) substantially attenuate the tax effects and the coefficient becomes even positive but insignificant. For the remaining analysis, we use models without year fixed effects. More specifically, we use Model 9 and 4 to estimate the tax effects on drinking participation and conditional drinking, respectively, for each race/ethnicity in separate regressions, because they use sampling weights, control for national and state-specific secular trends, and Model 9 adopts our preferred probit probability function. We argue that leaving out the year fixed effects is the better model because they account for so much of the variation that we need to identify any tax effect. We recognize that an ideal data source would have sufficient local variations in taxes to relax this exclusion, but there is nothing we can do about the fact that there were very few changes in taxes. Model 9 and 4 predict that 1% rise in excise tax reduces drinking participation and conditional drinking by 0.14% and 0.46%.

To assess the overall difference in demand responsiveness to alcohol tax across races/ethnicities, we conduct F-test assuming identical tax effects among all subgroups. This null hypothesis is strongly rejected in both drinking participation and conditional drinking models at P < 0.0001. Table 4 shows the point estimates of the tax effects on drinking participation and conditional drinking for White, African American, Hispanic, Asian or Pacific Islander, Native American, and other race or multi-race in separate regressions. Two aspects deserve notice. First, the tax effects on drinking participation for Hispanic, Native American, and other race/multi-race are not significant from zero, but the difference in tax effect between Hispanic and White is significant at 5% level in two-sample t-test. Second, tax effects on conditional drinking are all statistically significant, with Native American (-0.75) and other race/multi-race (−0.80) the highest and Hispanic (−0.35) lowest. Pairwise comparisons among those coefficients using t-test are not statistically at the 5% level.

Table 4.

Predicted tax effects on drinking participation and conditional drinking across races/ethnicities

Subgroup	Any drinka	Sample size	Log of number of drinksb	Sample size
Total population	−0.0711 (−9.72)	3,921,943	−0.4586 (−17.45)	1,939,550
White (non-Hispanic)	−0.0863 (−11.28)	3,192,830	−0.4786 (−18.62)	1,652,560
African American (non-Hispanic)	−0.0637 (−3.46)	304,762	−0.3541 (−5.26)	108,161
Hispanic	0.0138 (0.64)	227,642	−0.3512 (−4.97)	96,774
Asian or Pacific Islander (non-Hispanic)	−0.0792 (−2.32)	74,011	−0.4253 (−3.58)	30,575
Native American (non-Hispanic)	−0.0504 (−0.96)	53,157	−0.7454 (−3.51)	19,789
Other race or multi-race (non-Hispanic)	0.0119 (0.18)	69,541	−0.7993 (−3.57)	31,691

Note: (1) All models use data from BRFSS 1984-2009 waves.

(2) All models control for individual characteristics (i.e., gender, age, annual income, education, employment status, and marital status), state and month fixed effects, and national and state-specific time trends.

(3) All models are weighted by BRFSS final sampling weights.

(4) In the models for drinking participation (i.e., any drink), the treatment variable is the sum of federal and state excise taxes per gallon of beer. In the models for conditional drinking (i.e., log of number of drinks), the treatment variable is the natural logarithm of the sum of federal and state excise taxes per gallon of beer.

(5) Probit models are used to estimate drinking participation. Marginal effects at the mean of the independent variables are reported for the probit models to ease the interpretation of coefficients.

(6) OLS models are used to estimate conditional drinking. Due to the logarithmic transformation for the dependent variable and the treatment variable (i.e., tax), the coefficients can be interpreted as elasticity, namely percentage change in alcohol consumption given 1% change in tax.

(7) Eicker-Huber-White sandwich estimator is used to calculate robust standard errors clustered at year-month-state cell because all BRFSS respondents in that cell face the same tax rate.

(8) T-statistic is in parenthesis.

^aA dichotomous variable which denotes drinking any alcoholic beverages in the last 30 days.

^bNatural logarithm of number of standard drinks consumed per month. A standard drink is any drink that contains 14 grams of pure alcohol, such as 12-oz of beer or 5-oz of wine.

Drinkers of different types might respond differently to tax or price change (Manning et al., 1995). We next examine whether the tax effects are heterogeneous among light, moderate, and heavy drinkers across races/ethnicities. Although widely used, those terms associated with alcohol consumption level are not standardized (Abel et al., 1998). For robustness check, we consider several alternative definitions and define light drinkers as having 1-20, 1-30, or 1-40 standard drinks per month, moderate drinkers as 21-59, 31-79, or 41-99, and heavy drinkers as more than 60, 80, or 100. We use OLS to estimate the tax effects for each drinker type and race/ethnicity in separate regressions. Both the dependent variable (i.e., number of drinks consumed per month) and the treatment variable (i.e., the sum of federal and state excise taxes) are in logarithmic scale so that the coefficients can be interpreted as elasticity. Results are shown in Table 5.

The estimated tax effects among the whole population, displayed in Table 5, are relatively large and significant among light drinkers but shrink substantially for moderate and heavy drinkers. We can not reject the alcohol consumption of the latter types is unresponsive to tax changes (except for drinking category 31-79 and ≥ 80). All races/ethnicities share similar trends as the national population, although to a minor extent, heavy drinkers among other race/multi-race seem to become responsive again to tax change (although not all coefficients are statistically significant). Among light drinkers, the estimated tax effects are relatively large for Native American, Asian, and other race/multi-race, compared to their White, African American, and Hispanic counterparts. One important caveat is that the estimated elasticity may only be indicative when tax fluctuates within a small range. In fact, during the entire study period from 1984 to 2009, the mean of the intrastate maximal differences in tax rates (per gallon) is merely $0.55. Therefore, failure to reject that a type of drinker (in our case, moderate and heavy drinker) has tax inelastic demands does not predict its responsiveness under more extreme tax fluctuations.

Regarding other covariates in the overall drinking participation model (Model 9 in Table 3), being male, having college or higher education (compared to education lower than high school), and earning income among the highest quartile (compared to the lowest quartile), are associated with 0.13 (t = 113.89), 0.21 (t = 99.31), and 0.14 (t = 74.17) higher probability of drinking. Among the heaviest drinkers (≥100 drinks per month), being male (t = 8.05) predicts higher alcohol consumption while being highly educated (t = −7.29) and having highest earning (t = −2.99) predict lower consumption.

Discussion

This paper estimates demand responses to federal and state alcohol excise taxes for various racial/ethnic groups in the U.S. We could not find other recent data on this topic, even though there is a large literature on tax/price effects (but which does not distinguish racial/ethnic groups) and a large literature on drinking behaviors (which distinguishes racial/ethnic groups, but does not analyze tax/price effects). The only study of this topic we could identify was published more than a decade ago and it concluded that price effects did not differ between demographic groups, which would imply that taxes had a similar effect on all demographic groups (Saffer and Chaloupka, 1999). Our sample size is more than 80 times larger that that study, which should improve statistical precision. We do not confirm this null finding, but instead find some evidence on the differential tax effect on drinking across racial/ethnic groups. The null hypothesis on the identical tax effects among all races/ethnicities is strongly rejected (P < 0.0001), although most pairwise comparisons using t-test are not statistically significant at P < 0.05. Our point estimates suggest that the tax effect on drinking participation (i.e., drink or not) is largest among White and smallest among Hispanic. Among existing drinkers, Native American and other race/multi-race are most responsive to tax effects while Hispanic least, although our estimates are not precise enough to statistically distinguish these effect sizes. For all races/ethnicities, the estimated tax effects on consumption are large and significant among light drinkers (1-40 drinks per month), but shrink substantially for moderate (41-99) and heavy drinkers (≥ 100).

Due to several important weaknesses in our study, the results are tentative and require replications in different data and settings, ideally in settings with larger and more frequent tax/price variations. The limited variation of taxes is arguably the biggest limitation for this study as it constrains our modelling choices (such as the inclusion of a time trend instead of year fixed effects). An ideal data source would have sufficient local variations in taxes to make this exclusion unnecessary. Intrastate taxes vary in a small range during the study period, and therefore, the estimated elasticities may not be indicative under more extreme tax fluctuations. We tested replacing state fixed effects with annual federal and state tobacco tax (a proxy for state’s attitude towards addictive commodity) and adjusted monthly state unemployment rate (following Ruhm and Black, 2002). Year fixed effects, together with month fixed effects, and national and state specific linear time trends were included in the model. For drinking participation, the model without state fixed effects predicts that a 1% rise in excise tax reduces drinking participation by 0.25% (t = −19.33), which is substantially larger than 0.14% (t = −9.72) estimated by the model using state fixed effects but without year fixed effects. For conditional drinking, model without state fixed effects gives a significantly positive elasticity estimate of 0.08 (t = 4.41), which is rather odd and therefore suggests misspecification bias. In contrast, a model using state fixed effects but without year fixed effects predicts the elasticity to be −0.46 (t = −17.45), which is within the common range found in recent meta-analyses (Gallet, 2007; Wagenaar et al., 2009).

A second major limitation is that the alcohol use data in BRFSS does not distinguish beverage type so that we are not able to match the tax rate of beer, wine, and liquor to their corresponding consumption. We use beer excise tax to proxy for the “average” tax for all types of alcoholic beverages and estimate tax elasticities. The assumption is that beer taxes adequately represent, or are correlated with total alcoholic beverage taxes in a given state, month, and year. This treatment, although necessary, is not innocuous. It is possible that different racial/ethnic groups have different beverage preferences (Graves and Kaskutas, 2002; Kerr et al., 2009), and these appear as differential responses to excise taxes. Moreover, alcohol taxes, especially for wine and liquor, are likely to operate differently between license and control states. Because government monopolies in control states possess much more direct control in setting prices, prices could be changed without using the tax mechanism. To at least partially examining the differential effect of beer tax on drinking between license and control states, we included both a dummy variable for control states and its interaction with beer tax in predicting drinking participation and conditional drinking. The model predicts that a 1% rise in tax reduces drinking participation by 0.11% (t = −7.00) in license states and 0.21% (t = −9.98) in control states. For conditional drinking, the model predicts a 0.42% (t = −13.73) reduction in license states and a 0.51% (t = −15.12) in control states. Our pooled estimates (i.e., estimates for all states without distinguishing license and control status) of 0.14% (t = −9.72) and 0.46% (t = −17.45) lie between.

Other limitations of this study include measurement error in self-reported alcohol use data, sensitivity in modeling conditional drinking using alternative parameterizations of state specific secular trends, and not being able to control subgroup baseline alcohol use rate. Differences in tax elasticity across races/ethnicities might simply be due to their different baseline alcohol use rates. We considered addressing this issue by controlling for national annual average alcohol consumption of each race/ethnicity in the model, but it also was too collinear with taxes to allow an estimation of tax effects.

The estimated tax elasticity of beer for the whole population (−0.46) is comparable to the results of recent meta-analyses. Gallet (2007) conducted meta-analysis on 132 studies during 1945-2003 and reported median price elasticity for beer −0.36. Wagenaar et al. (2009) reviewed 112 studies during 1968-2007 and estimated the mean elasticity of beer −0.46. It may also be interesting to compare the qualitative results of our work to the findings in Ayyagari et al (2009). Among the two latent groups recovered in their study, one is significantly responsive to price while the other not. The responsive group consists of light drinkers (number of drinks per day less than 1) while the moderate and heavy drinkers occupy the unresponsive group. This coincides with our finding on the inelastic demand among moderate and heavy alcohol users across all races/ethnicities.

Population heterogeneity in tax responses is likely to be multi-dimensional (Meier et al., 2010; Ayyagari et al., 2009). No single policy may address efficiency and equity in an optimal way, while combinations of certain pricing and access policies could work best. Designing optimal policy scenarios will require a deeper understanding of heterogeneity in population responses to alternative alcohol policies. While far from providing a conclusive answer here, this analysis provides a first step toward this direction.