Alcohol demand and risk preference

Abstract

Both economists and psychologists have studied the concept of risk preference. Economists categorize individuals as more or less risk-tolerant based on the marginal utility of income. Psychologists categorize individuals' propensity towards risk based on harm avoidance, novelty seeking and reward dependence traits. The two concepts of risk are related, although the instruments used for empirical measurement are quite different. Psychologists have found risk preference to be an important determinant of alcohol consumption; however economists have not included risk preference in studies of alcohol demand. This is the first study to examine the effect of risk preference on alcohol consumption in the context of a demand function. The specifications employ multiple waves from the Panel Study of Income Dynamics (PSID) and the Health and Retirement Study (HRS), which permit the estimation of age-specific models based on nationally representative samples. Both of these data sets include a unique and consistent survey instrument designed to directly measure risk preference in accordance with the economist's definition. This study estimates the direct impact of risk preference on alcohol demand and also explores how risk preference affects the price elasticity of demand. The empirical results indicate that risk preference has a significant negative effect on alcohol consumption, with the prevalence and consumption among risk-tolerant individuals being 6–8% higher. Furthermore, the tax elasticity is similar across both risk-averse and risk-tolerant individuals. This suggests that tax policies are as equally effective in deterring alcohol consumption among those who have a higher versus a lower propensity for alcohol use.

1. Introduction

Although most individuals consume alcohol safely, excessive consumption by some individuals is associated with considerable costs. According to the 2006 Behavioral Risk Factor Surveillance System, over 55% of the US adult population are current drinkers. Data from the National Epidemiologic Survey on Alcohol and Related Conditions indicate that the number of adults with an alcohol abuse or dependence disorder has increased by 14% between 1991/1992 and 2001/2002, adjusting for population growth (Grant et al., 2004). There is extensive literature documenting the association between excessive drinking and lost productivity, crime and violence, injuries, and premature mortality.² In 2000, there were a total of 140,000 alcohol-attributable deaths, making alcohol consumption the third leading cause of mortality behind smoking and poor diet or inactivity (Mokdad, Marks, Stroup, & Gerberding, 2000). Alcohol-related morbidity also imposes considerable strain on the public healthcare system. Among the uninsured, hospital stays related to alcohol abuse comprise the fourth most common reason for hospitalization, and a quarter of all alcohol-related stays involve Medicaid patients (Owens, Myers, Elixhauser, & Brach, 2007). Estimates place the overall economic costs of alcohol abuse at $228 billion annually.³ As a consequence of these concerns, the public sector has regulated the sale and consumption of alcoholic beverages through taxation and controls over distribution and availability.

The need to understand the effects of alcohol regulation has motivated a number of studies of the demand for alcohol. These economics studies do not consider the confounding effect of risk preference. This is the first study to examine the effect of risk preference on alcohol consumption in the context of a demand function. The significance of this is underscored by the fact that unobserved heterogeneity across individuals accounts for most of the variance in alcohol consumption (Cutler & Glaeser, 2005). This heterogeneity is generally ignored due to lack of data. The purpose of this study is to explicitly incorporate the effect of risk preference, which is an element of this person-specific heterogeneity, in the alcohol demand analysis. This study utilizes unique data on risk preference from two large-scale population surveys: the Panel Study of Income Dynamics (PSID) and the Health and Retirement Study (HRS). Information on the individual's preference towards risk was collected from an identical survey instrument in both data sets. Utilizing both data sets also allows for separate analysis by age groups and provides information on the consistency and robustness of the estimates.

The economist's concept of risk preference involves a situation in which the outcome of a choice or an exogenous event is not known with certainty. A key assumption is that while the outcome is uncertain, a set of potential outcomes and their probabilities of occurrence can be subjectively estimated. Economists classify individuals into varying propensities towards risk-taking based on how income affects utility. Utility increases with income, though it is subject to diminishing marginal utility for risk avoiders. Among risk-avoiding individuals, those who are relatively more tolerant of risk have a slower rate of diminishing marginal utility, whereas those who are more risk averse have a faster rate of decline in marginal utility.⁴ This notion of risk preference has generally been studied by economists in relation to financial decisions and insurance modeling.

Psychologists view risk preference as a personality trait and their work on risk preference has linked it to a larger theoretical and empirical literature on personality. Psychologists, such as Cloninger (1987), developed a psychometric model of personality which includes personality traits closely resembling the economist's notion of risk preference.⁵ Individuals who score high on novelty seeking and low on harm-avoidance are defined by Cloninger as danger-seeking and confident individuals and correspond closely to the economist's concept of risk-tolerant or risk-preferring individuals. Each of these traits is considered moderately heritable, developmentally and situationally stable, and associated with specific neural systems that mediate different types of stimulus–response relationships (Menza, Golbe, Cody, & Forman, 1993). Thus, these components of risk preference reflect an underlying biogenic structure of personality that may interact with environmental stimuli. Howard, Kivlahan, and Walker (1997) applied Cloninger's psychometric model to substance-abusing populations. They found that individuals who score high on “novelty seeking” and low on “harm avoidance” are more likely to have early onset of alcohol abuse.

This paper builds on the work of psychologists by using an instrument developed by economists to measure risk preference. An important distinction in this work and that of psychologists is that risk preference is included as a demand parameter. This adds important price and income variables and links this work to the larger work by economists. The risk instrument used in this paper has undergone considerable testing in order to minimize misunderstandings and additional complications in interpretation and to ensure consistency with the economist's concept of risk preference. Barsky, Juster, Kimball, and Shapiro (1997) provide a detailed analysis of this survey instrument.

Accounting for risk preference in alcohol demand will provide information on any differential effects of prices and policy across heterogeneous sub-populations. Risk preference may affect the price elasticity of alcohol and affect the response to other alcohol control policies. This is an important question since the public policy recommendations of higher alcohol taxes and more stringent drunk driving laws may not each be equally effective with all segments of the population. An estimation of the average price or policy response for the overall population may mask important differences across groups. While many individuals regularly consume alcohol without imposing internal or external harm, the distribution of alcohol consumption in the US is far from uniform. Among current drinkers, the top quintile is responsible for over 65% of total alcohol consumption, and the top decile is responsible for 55% of total consumption.⁶ Since it is immoderate or problem drinking that leads to high social costs, it is important to determine whether policies are affecting the behavior of target groups most likely to over-consume alcohol. Data from the HRS indicate that individuals with the highest level of risk tolerance consume almost 30% more alcohol and are 32% more likely to binge, relative to the most risk-averse individuals.⁷ To evaluate the effectiveness of public policy it is important to have Information on whether the impact of price-based policies differs across groups that significantly differ in their alcohol consumption.

2. Prior studies

There is an empirical literature in economics which has examined the impact of price on consumption of alcoholic beverages. A review of older studies, based on individual-level data, suggests that the demand for alcohol is responsive to shifts in prices though there is considerable variation in estimates of the price elasticity (Leung & Phelps, 1993). More recent studies confirm this negative price response, and further point to important differences among certain demographic groups. Manning, Blumberg, and Moulton (1995), based on data from a supplement to the 1983 National Health Interview Survey, find that the median drinker is price responsive, with an estimated elasticity of 1.19; however, the elasticity significantly decreases in magnitude for the heavy drinkers. In fact, at the 95th percentile of drinkers, they cannot reject the hypothesis that demand is perfectly inelastic. Kenkel (1996), using the same data, further shows that the price responsiveness varies considerably and positively with respect to drinking-related health information. Heavy drinking by the most-informed consumers is much more price elastic than moderate drinking, while the estimated price elasticities of heavy drinking for the least-informed consumers are not statistically significant. Chaloupka and colleagues (2002) provide a good review of the literature looking at the effects of alcohol prices on consumption and indicators of alcohol abuse such as motor vehicle fatalities, adverse health effects, and violence and crime.

Studies have also considered specific demographic sub-populations and found differences between groups, although not always in a consistent manner with respect to the price sensitivity. Laixuthai and Chaloupka (1993) employ data from the 1982 and 1989 Monitoring the Future (MTF) surveys of high school seniors. They find that, for both years, higher beer excise taxes significantly reduced the frequency of drinking as well as the probability of heavy drinking. Using longitudinal data on youths ages 17–29 from the MTF surveys, Grossman, Chaloupka, and Sirtalan (1998) apply the rational addiction paradigm of Becker and Murphy (1988) to estimate a long-run price elasticity of −0.65. Chaloupka and Wechsler (1996) investigate college drinking patterns and find significantly negative price effects for underage drinking and binge drinking among female students. However, no effects are found for males. Saffer and Chaloupka (1999) utilize data from the National Household Surveys of Drug Abuse (1988, 1990 and 1991) to estimate differential price responses for various demographic groups. They find similar price elasticities for the frequency of past month alcohol use among males and females; however, blacks were found to be less sensitive relative to other races and especially whites. Cook and Moore (2001) study youth drinking patterns from the National Longitudinal Surveys of Youth (NLSY) 1979 cohort. Their results indicate that among youths between the ages of 17 and 32, the excise tax on beer has a significant deterrent effect on past month participation though not on binge drinking. For some specifications, elasticity estimates suggest that females may be more price-sensitive. Saffer and Dave (2006) study alcohol consumption among adolescents based on the MTF and the NLSY-97 data. They find that measures of alcohol participation and binge drinking are responsive to the weighted price of alcohol. Analyses stratified on various demographic characteristics suggest that the elasticity estimates are larger for females and for whites. While the literature has focused on the overall population and youths, there have been no studies that have specifically investigated the price response of older adults. Utilizing the HRS data, the present study addresses this gap in the literature by providing the first estimates of the price sensitivity of alcohol use among older adults, incorporating the effects of risk tolerance.

According to NIAAA (2000), alcohol use and abuse are best viewed as functions of a combination of genetic, psychological and social influences. Although the association between psychiatric disorders and alcohol has been widely researched by psychologists (Kessler et al., 1996), there has been little work in this area by economists. Saffer and Dave (2005) estimate the effect of mental illness on the demand for addictive substances, including alcohol participation. They show that individuals with a history of mental illness are 26% more likely to consume alcohol. Accounting for the endogenous selection of mental illness, individuals diagnosed with a mental disorder in the past year or in their lifetime are also found to be more sensitive to alcohol prices relative to healthy individuals. The price elasticity for individuals with a recent or lifetime mental illness is estimated at −0.49 to −0.63, versus −0.38 for individuals with no mental illness.

This study builds on the work done by psychologists which showed that risk preference affects alcohol consumption. There has been no prior research in economics that has integrated personality traits into the alcohol demand function nor estimated the differential response of demand to price variations across measures of risk tolerance.

3. Analytical framework

The objective of this study is to assess how risk preference affects an individual's alcohol use and their response to shifts in the price of alcohol. Since alcohol is ultimately a consumer good, this question can be framed within the context of utility theory.

$$U=f(A,Y;e)$$ (1)

Eq. (1) specifies an individual's utility as a function of alcohol consumption (A) and income Y that can be used to consume other goods, with exogenous preference parameter (e). The marginal utility of alcohol consumption is positive and diminishing.⁸ The positive outcomes associated with alcohol use include intoxication and stimulation of the dopamine receptors in the brain's pleasure center. Moderate alcohol use can also have positive health effects and improve social functioning. Short-run negative consequences of drinking include dehydration and gastrointestinal disorders, reduced productivity, an increased probability of accidental injury including automobile accidents, perpetrating or being a victim of violence or crime, and sexual abuse. Long-run effects can also include addiction, loss of employment, problems in interpersonal relationships, and more serious health consequences such as cirrhosis of the liver and obesity (NIAAA, 2000).

The role of an individual's tolerance towards risk in the alcohol demand function stems from the fact that these outcomes are probabilistic rather than known with certainty. The potential positive and negative effects of alcohol consumption map into a potential set of gains and losses in income or utility. The consumer maximizes an expected utility function that takes into account these subjective probabilities and the set of gains and losses.⁹ Concavity of the expected utility function with respect to income is equivalent to risk aversion, and the more concave (that is, the higher is the degree of diminishing marginal utility of income) the more risk averse is the consumer.¹⁰ Intuitively, a risk-averse individual places a higher weight on the loss in income than on an equal sized gain. As the expected utility function becomes less concave (more convex), the degree of risk aversion declines and the individual becomes more tolerant of risk. In this case, the weight attached to the potential loss from alcohol consumption declines, and conversely is shifted towards greater weight on the potential gain from alcohol consumption. Ceteris paribus, it follows from expected utility maximization that individuals who are more risk tolerant will have a higher demand for alcohol. Simple correlations generally confirm this prediction. Andrucci et al. (1989) and Gerra et al. (1999) show that individuals with high levels of novelty seeking and low levels of harm-avoidance are more likely to abuse substances. Similarly, Barsky et al. (1997) link increased tolerance towards risk with higher alcohol consumption. However, these studies do not estimate this link within a demand framework, and do not control for other factors that are correlated with alcohol use and may be confounding this relationship.

The degree of risk aversion may also affect an individual's response to alcohol price and policy. Since risk-tolerant individuals are more likely to consume alcohol, and may do so at immoderate levels, it is important to determine whether public policy prescriptions of higher alcohol taxes would be effective for this high-use group. In this respect, theory is ambiguous. Generally, price sensitivity depends inversely on the magnitude of the second derivative of the utility function with respect to alcohol consumption, the rate at which marginal utility diminishes. The faster the rate of decrease, the less price-sensitive is the consumer. Since, in its general form, income (and other goods) can interact with alcohol consumption in the utility function, risk aversion may therefore have an effect on the price elasticity of demand. If marginal utility of alcohol use diminishes rapidly for risk-tolerant individuals, then they may be less responsive to price. On the other hand, risk tolerant consumers may have a higher price elasticity if their rate of diminishing marginal utility is lower. Studies, based on the Cloninger tridimensional scale, have generally found that two different types of temperaments have a higher propensity for alcohol abuse (Howard et al., 1997). Type I temperaments have high harm-avoidance and reward-dependence characteristics. Type 2 temperaments have high novelty-seeking traits. Thus, it is difficult to say a priori how marginal utility diminishes for these consumers, and therefore how they would react to shifts in costs. Due to this theoretical ambiguity, the nature of the price response for risk-tolerant versus risk-averse individuals remains an empirical question.

The following specification will be estimated based on the above discussion.¹¹

$$A_{\text{ist}}=B_0+B_1{P}_{st}+B_2{R}_{is}+B_3{Y}_{\text{ist}}+B_4{X}_{\text{ist}}+B_5{H}_{\text{ist}}+{\mu }_s+{\nu }_t+{\epsilon }_{\text{ist}}$$ (2)

Eq. (2) represents the alcohol demand function for the ith individual residing in state s at year t. Alcohol consumption (A) depends on prices and policies (P) regulating the sale and consumption of alcohol, income (Y), and other socio-demographic factors (X) such as age, gender, race, and education, with ε representing a classical error term. An indicator (R) that dichotomizes individuals as risk averse or risk tolerant will also be included. Alternative models that contain measures of physical and mental health status (H) are also estimated. Studies have found that mental illness raises participation in substance use including alcohol, consistent with the self-medication hypothesis (Saffer & Dave, 2005). Certain physical ailments (diabetes, gastrointestinal disorders) or even general poor health may also reduce the individual's demand for alcohol. Since the key policy instrument (alcohol taxes) vary at the state-level over time, it is important to control for unobserved state-specific factors and time trends. Specifications are therefore estimated with a vector of state (μ) and year (ν) fixed effects.¹²

In order to allow for differential price responses across risk-tolerant and risk-averse individuals, the above specification can be expanded to include an interaction between the price measure (P) and the indicator of risk preference (R). In some specifications where limited sample size is an issue, differential price responses are estimated using such an interaction effect. These specifications restrict the effect of other factors to be the same across groups, while allowing the price effect to differ. However, in general, a more flexible formulation based on sample stratification is followed to allow for differences in all parameters across risk tolerance.

$$A_{\text{ist}|\text{risk averse}}=B_0+B_1{P}_{st}+B_2{Y}_{\text{ist}}+B_3{X}_{\text{ist}}+B_4{H}_{\text{ist}}+{\mu }_s+{\nu }_t+{\epsilon }_{\text{ist}}$$ (3)

$$A_{\text{ist}|\text{risk tolerant}}={\alpha }_0+{\alpha }_1{P}_{st}+{\alpha }_2{Y}_{\text{ist}}+{\alpha }_3{X}_{\text{ist}}+{\alpha }_4{H}_{\text{ist}}+{\gamma }_s+{\delta }_t+{\eta }_{\text{ist}}$$ (4)

Comparison of the parameters (B₁) and (α₁) shows whether risk-tolerant individuals are more or less sensitive to price changes, relative to risk-averse individuals.

4. Data

The Panel Study of Income Dynamics (PSID) is a longitudinal study of a representative sample of individuals and their family units, conducted by the Institute of Social Research. Originating in 1968, the PSID core sample combines the Survey Research Center sample, which is a cross-sectional national sample, and the Survey of Economic Opportunity sample, which is a national sample of low-income families. From 1968 to 1996, individuals from families in the core sample were interviewed every year, including adults as they have grown older and their respective family units. From 1997, data collection became biennial. A number of other changes were also made, including a reduction in the core sample and the introduction of a refresher sample of post-1968 immigrant families and their adult children, in order to keep the sample representative.

A dichotomous indicator for current alcohol consumption is constructed from the PSID data. Among individuals between the ages of 21–54, 66% are current drinkers; the prevalence decreases with age as evidenced in the older cohort from the Health and Retirement Study. Detailed information on various health measures is also available. An index of functional difficulties associated with bathing, dressing, eating, walking, and getting outside is created and ranges from zero to five. Dichotomous measures are also constructed for lifetime diagnoses of hypertension, diabetes, heart disease, and stroke. A depression scale, ranging from zero to four, measures the number of depression-related symptoms experienced in the past month. Indicators for age, gender, race, ethnicity, marital status, and education are defined and included in the models. Measures of labor force attachment are also constructed, capturing whether the individual is currently working (part time or full time), unemployed, or retired. Additional variables are defined in Table 1.

Table 1.

Weighted sample means and distribution of risk aversion – PSID and HRS

		PSID (1999–2003)				HRS (1992–2004)
Variable	Definition	Ages 21-54			Ages 55+	Ages 55+
		All	Risk Averse	Risk Tolerant	All	All	Risk Averse	Risk Tolerant
Panel A Sample Means
Alcohol Participation	Dichotomous indicator for current drinker	0.6607 (0.4735)	0.6421*** (0.4795)	0.7528 (0.4315)	0.4893 (0.4999)	0.3508 (0.4772)	0.3590*** (0.4797)	0.4131 (0.4924)
Drinks	Average number of drinks consumed daily	–	–	–	–	0.3466 (0.7569)	0.3537*** (0.7713)	0.4304 (0.8321)
Real Beer Tax	State excise tax on beer, adjusted by the consumer price index, in dollars per gallon	0.1286 (0.0911)	0.1355*** (0.0966)	0.1276 (0.0925)	0.1346 (0.0944)	0.1301 (0.0843)	0.1315*** (0.0876)	0.1257 (0.0819)
Risk Averse	Dichotomous indicator for high degree of risk aversion	0.4293 (0.4950)	–	–	0.6233 (0.4847)	0.6388 (0.4803)	–	–
Age	Age of respondent	39.599 (9.274)	41.834*** (8.274)	40.706 (8.174)	67.3427 (9.096)	67.857 (9.874)	63.256*** (6.571)	62.5130 (6.468)
Male	Dichotomous indicator for male	0.4707 (0.4992)	0.4537*** (0.4979)	0.5647 (0.4959)	0.4190 (0.4934)	0.4453 (0.4970)	0.4442*** (0.4969)	0.4988 (0.5000)
Black	Dichotomous indicator for Black	0.1204 (0.3254)	0.1512*** (0.3583)	0.0977 (0.2970)	0.0836 (0.2768)	0.0912 (0.2879)	0.0990*** (0.2986)	0.0873 (0.2823)
Other Race	PSID: Dichotomous indicator for race other than Black, White or Hispanic HRS: Dichotomous indicator for race other than Black or White	0.0400 (0.1959)	0.0165* (0.1275)	0.0245 (0.1545)	0.0291 (0.1681)	0.0319 (0.1758)	0.0349 (0.1835)	0.0376 (0.1902)
Hispanic	Dichotomous indicator for Hispanic	0.0690 (0.2534)	0.0140 (0.1174)	0.0171 (0.1297)	0.0261 (0.1595)	0.0586 (0.2348)	0.0631 (0.2431)	0.0656 (0.2475)
High School	Dichotomous indicator for highest level of schooling completed being high school	0.2811 (0.4496)	0.3461* (0.4763)	0.2851 (0.4519)	0.4241 (0.4946)	0.3599 (0.4800)	0.3914*** (0.4881)	0.3299 (0.4702)
Some College	Dichotomous indicator for highest level of schooling completed being some college	0.2683 (0.4431)	0.2541 (0.4358)	0.2566 (0.4372)	0.1475 (0.3549)	0.1947 (0.3960)	0.1986*** (0.3989)	0.2315 (0.4218)
College	Dichotomous indicator for highest level of schooling completed being college	0.3704 (0.4830)	0.3223** (0.4679	0.4148 (0.4932)	0.2018 (0.4017)	0.1853 (0.3885)	0.1890*** (0.3915)	0.2555 (0.4361)
Education Missinga	PSID: Dichotomous indicator for respondents whose education level is missing	0.7321 (0.4429)	0.8657 (0.3410)	0.8691 (0.3373)	0.9122 (0.2830)	–	–	–
Real Income	Income from all sources, adjusted by the consumer price index, in thousands of dollars	40.549 (45.331)	42.831*** (48.736)	47.200 (44.732)	34.4105 (56.336)	13.068 (22.834)	14.242*** (22.853)	16.034 (24.279)
Married	Dichotomous indicator for married	0.6402 (0.4800)	0.6400 (0.4801)	0.6265 (0.4838)	0.6514 (0.4766)	0.6256 (0.4840)	0.6843*** (0.4648)	0.6724 (0.4693)
Working	Dichotomous indicator for full-time or part-time work	0.8149 (0.3884)	0.8927 (0.3096)	0.8916 (0.3110)	0.3449 (0.4754)	0.3811 (0.4857)	0.4756*** (0.4994)	0.5207 (0.4996)
Unemployed	Dichotomous indicator for unemployed	0.0400 (0.1961)	0.0272 (0.1627)	0.0346 (0.1827)	0.0058 (0.0761)	0.0080 (0.0889)	0.0099*** (0.0989)	0.0131 (0.1139)
Retired	Dichotomous indicator for complete retirement	0.0106 (0.1024)	0.0137** (0.1163)	0.0076 (0.0868)	0.4755 (0.4994)	0.4762 (0.4994)	0.3910*** (0.4880)	0.3477 (0.4762)
Household Size	Number of members residing in the household	2.9389 (1.4876)	2.8248 (1.4072)	2.7918 (1.3977)	1.9379 (0.8559)	2.1433 (1.1061)	2.2593 (1.1260)	2.2457 (1.1292)
Parental Education	PSID: Dichotomous indicator for whether mother or father is a college graduate HRS: Dichotomous indicator for whether mother and father have completed at least 8 years of schooling	0.2516 (0.4339)	0.1901*** (0.3925)	0.3214 (0.4671)	0.0904 (0.2868)	0.5153 (0.4998)	0.5566*** (0.4968)	0.6133 (0.4870)
Functional Difficulties	PSID: Count of the difficulties associated with: 1) bathing, 2) dressing, 3) eating, 4) walking, 5) getting outside HRS: Count of the difficulties associated with: 1) walking 1 block, 2) walking several blocks, 3) walking across a room, 4) climbing 1 flight of stairs, 5) climbing several flights of stairs	0.0404 (0.2749)	0.0202*** (0.1757)	0.0298 (0.2168)	0.3654 (0.8993)	0.9945 (1.4179)	0.8650*** (1.3234)	0.8047 (1.2751)
Hypertension	Dichotomous indicator for whether respondent has been diagnosed with high blood pressure	0.1517 (0.3587)	0.1667** (0.3728)	0.1420 (0.3491)	0.4464 (0.4972)	0.4745 (0.4994)	0.4598*** (0.4984)	0.4384 (0.4962)
Diabetes	Dichotomous indicator for whether respondent has been diagnosed with diabetes	0.0494 (0.2166)	0.0527 (0.2234)	0.0443 (0.2058)	0.1488 (0.3560)	0.1431 (0.3501)	0.1448** (0.3519)	0.1383 (0.3452)
Heart Disease	Dichotomous indicator for whether respondent has been diagnosed with heart disease	0.0289 (0.1676)	0.0269 (0.1617)	0.0205 (0.1417)	0.1764 (0.3812)	0.2315 (0.4218)	0.1909*** (0.3930)	0.1800 (0.3842)
Stroke	Dichotomous indicator for whether respondent has ever had a stroke	0.0097 (0.0981)	0.0117 (0.1074)	0.0076 (0.0868)	0.0784 (0.2687)	0.0738 (0.2614)	0.0527 (0.2234)	0.0499 (0.2176)
Depression	PSID: Number of depression-related symptoms in the past month: 1) sad, 2) nervous, 3) restless, 4) hopeless HRS: Center for Epidemiologic Studies Depression Scale; Sum of mental health symptoms in the past week: 1) depressed, 2) everything an effort, 3) restless sleep, 4) not happy, 5) lonely, 6) sad, 7) could not get going, 8) did not enjoy life	0.5247 (0.9900)	0.4572*** (0.9191)	0.5500 (0.9675)	0.4231 (0.9080)	1.4394 (1.9073)	1.3232*** (1.8755)	1.3770 (1.9079)
BAC 08	Dichotomous indicator for whether the state has an effective 0.08 BAC per se law in the given interview period	0.4812 (0.4997)	0.4379 (0.4962)	0.4525 (0.4978)	0.4862 (0.4999)	0.4778 (0.4995)	0.5000*** (0.5000)	0.5311 (0.4990)
Percent Dry	Percent of state population residing in dry counties in the given period	0.0429 (0.0934)	0.0578*** (0.1094)	0.0394 (0.0893)	0.0470 (0.1011)	0.0299 (0.0698)	0.0309*** (0.0692)	0.0284 (0.0685)
Monopoly State	Dichotomous indicator for whether the state controls the sale of distilled spirits	0.3152 (0.4646)	0.3706*** (0.4831)	0.3317 (0.4709)	0.3300 (0.4703)	0.2956 (0.4563)	0.3086*** (0.4619)	0.2909 (0.4542)
Planning Horizon 5-10 years	Dichotomous indicator for whether the respondent's relevant financial planning horizon is the next 5-10 years	–	–	–	–	0.2751 (0.4466)	0.2851*** (0.4515)	0.3141 (0.4642)
Planning Horizon 10+ years	Dichotomous indicator for whether the respondent's relevant financial planning horizon is longer than 10 years	–	–	–	–	0.0789 (0.2696)	0.0879*** (0.2832)	0.0971 (0.2961)
Observations		27,531	4,430	5,280	5,819	101,477	43,487	23,449
Panel B Distribution of Risk Aversion
Risk Classification
Category 1 (Most Risk-Averse)	42.9			62.3	63.9
Category 2	15.4			12.3	13.1
Category 3	18.6			9.0	10.3
Category 4 (Most Risk-Tolerant)	23.0			16.4	12.7
Observations	9,710			1,953	118,902

Notes: Means are weighted by the sampling weights. Standard deviations are in parentheses. Number of observations listed represents the maximum number. For some variables, the actual sample size is less due to missing information. Asterisks denote that the difference in means between the risk averse and risk tolerant samples is statistically significant as follows: ***Significant at the 1% level, ** significant at the 5% level, *Significant at the 10% level.

^aDue to a large number of missing observations for education in the PSID, a separate category for missing information is created.

The Health and Retirement Study (HRS) is conducted by the Institute for Social Research at the University of Michigan. It is an ongoing longitudinal study, which began in 1992 and is repeated biennially.¹³ Prior to 1998, the HRS cohort included individuals born between 1931 and 1941, and a separate Study of Assets and Health Dynamics among the “Oldest Old” (AHEAD) included individuals born before 1924. Since 1998, AHEAD respondents have been contacted as part of a joint data collection effort with the HRS, and the sample frame was also expanded by including cohorts born between 1924 and 1930 and those born between 1942 and 1947. As older adults are over-represented in the HRS, this is an ideal data-set, in terms of sample size and available information on correlates of alcohol use, to study alcohol demand for this segment of the population. The present analysis utilizes the first seven waves, spanning 1992–2005, and restricts the sample to older adults ages 55 and over. This yields a maximum sample size of about 107,000 person-wave observations.

Both dichotomous and continuous measures of alcohol use are constructed. The dichotomous indicator measures whether the individual currently participates in drinking. The individual is also asked about their frequency and intensity of alcohol consumption.¹⁴ Based on these questions, a measure of the average number of drinks consumed daily is obtained. Approximately 35% of older adults are current drinkers. Among those who drink, slightly less than one drink is consumed daily on average.¹⁵

The HRS is administered for the specific purpose of studying life-cycle changes in health and economic resources, and includes detailed information on various health outcomes. A composite index is defined to measure difficulties associated with mobility. It ranges from zero to five and indicates difficulties in walking one block, walking several blocks, walking across a room, climbing one flight of stairs, and climbing several flights of stairs. Additional indicators are defined separately for whether the respondent reports that he or she has been diagnosed with the following illnesses: diabetes, heart disease, stroke, and high blood pressure. The HRS contains a depression scale, as defined by the Center for Epidemiologic Studies (CES), which ranges from zero to eight. This CESD score measures the sum of adverse mental health symptoms for the past week (listed in Table 1). Studies have confirmed the validity and reliability of the CESD scale as a screening instrument for the identification of major depression in older adults (Irwin, Artin, & Oxman, 1999).

Measures are selected from the HRS to ensure consistency with variables constructed from the PSID. Indicators for age, gender, race, ethnicity, marital status, and education are defined and included in the models. Measures of labor force attachment are also constructed, capturing whether the individual is currently working (part time or full time), unemployed, or retired. Real income is calculated for each individual from all available sources including earnings, pension, supplemental security, social security retirement, and other government transfers deflated by the consumer price index.¹⁶ Description of these covariates is provided in Table 1.

Risk preference instrument

The PSID and the HRS contain the identical risk preference instrument. Individuals are classified into four ordinal categories of risk preference based on the answers to three questions probing the individual's willingness to gamble over lifetime income. The individual is asked to assume that their job guarantees their current income for the rest of their life and that their current and offered jobs all have equal and desirable non-pecuniary aspects. The questions are:

1. Would you take a new job with a 50–50 chance it will double your income and a 50–50 chance that it will cut your income by a third?

   If the answer to (1) is “Yes,” then the individual is asked:

2. Would you take a new job with 50–50 chance that it would double your income and a 50–50 chance that it would cut your income in half ?

   If the answer to (1) is “No,” then the individual is asked:

3. Would you take a new job with a 50–50 chance that it would double your income and 50–50 chance that it would cut your income by 20%?

Individuals are defined as most risk-averse if the answers to questions (1) and (3) are both no. The second most risk-averse group of individuals answers “No” to question (1) and “Yes” to question (3); they reject the income loss of one-third but accept the scenario with an income loss of one-fifth. The third group (with a lower degree of risk aversion) answers “Yes” to question (1), accepting the one-third income loss, but answers “No” to question (2), rejecting the one-half income loss. The fourth group comprising of the most risk-tolerant individuals answers “Yes” to both questions (1) and (2), accepting both the one-third and one-half income loss scenarios.¹⁷

Table 1B presents the data for these questions from the PSID and from the HRS. The data show that 43% of the PSID respondents (between ages 21 and 54) and 64% of the HRS respondents (over age 55) are in the most risk-averse category. When restricting the PSID sample to individuals 55 years of age and older, the prevalence of risk aversion is 62.3% which is approximately the same as found in the HRS. The consistency of these rates across two independent samples is mutually validating. The four risk categories are aggregated into two categories, the most risk-averse and everyone else.¹⁸ This maximizes the size of the subsamples and reduces sample variation as a source of bias. The dichotomous measure also minimizes errors from classifying an individual into the exact risk-aversion category due to any misreporting.

Barsky et al. (1997) studied these data in the HRS data set and using simple means showed that more risk-averse individuals participate less in risky behaviors. These behaviors include smoking and drinking. They also found that more risk-averse individuals are more likely to have health insurance coverage, less likely to be self-employed, less likely to be an immigrant, and less likely to reside in the western US. More risk-tolerant respondents were found to have a higher share of their portfolio in equities while less risk-tolerant individuals were found to have a higher share of their portfolio in relatively safe assets, such as Treasury bonds and savings accounts.

Barsky and colleagues also find variation in risk tolerance with age which is inconsistent with the psychologist's view that the risk preference is time-invariant. This inconsistency is due to the framing of the questions used in the risk instrument. The risk questions relate specifically to labor market income. Individuals who are more marketable will be more confident about their labor market opportunities. As individuals age, their marketability diminishes and they will become less secure about their labor market opportunities. To investigate this point more fully, a set of regressions were estimated to examine the effect of age on risk tolerance, conditioning on demographics. The results, presented in Table 2, indicate that each year of age increases the respondent's probability of classification in to the most risk-averse category by .004. This implies that a 65-year-old individual would have a 10% point higher probability of being classified as risk-averse relative to a 40-year-old individual. This is a potential source of bias in the risk measure derived from these questions. The bias will result in an underestimate of the effect of risk aversion on alcohol consumption. Thus, the results presented in section V should be interpreted as conservative estimates of the effect of risk on alcohol consumption.

Table 2.

Regressions of risk aversion on demographic variables

Variable	PSID Ages 21–54		HRS Ages 55 +
	1	2	3	4
Male	-0.0848*** (0.0182)	-0.0799*** (0.0186)	-0.0416*** (0.0086)	-0.0424*** (0.0087)
Black	0.0770*** (0.0201)	0.0562** (0.0231)	0.0158 (0.0122)	0.0081 (0.0127)
Other Race	-0.0248 (0.0628)	-0.0270 (0.0646)	-0.0100 (0.0229)	-0.0028 (0.0230)
Hispanic	-0.0445 (0.0932)	-0.0109 (0.1017)	-0.0423*** (0.0167)	-0.0300* (0.0178)
High School	0.0341 (0.0382)	0.0391 (0.0386)	0.0084 (0.0114)	0.0114 (0.0115)
Some College	0.0409 (0.0392)	0.0440 (0.0396)	-0.0516*** (0.0133)	-0.0438*** (0.0134)
College	-0.0461 (0.0377)	-0.0379 (0.0385)	-0.0816*** (0.0140)	-0.0747*** (0.0142)
Real Income	-0.0006** (0.0002)	-0.0004* (0.0002)	0.0002 (0.0001)	0.0002* (0.0001)
Married	0.0492*** (0.0177)	0.0482 (0.0179)***	0.0285*** (0.0088)	0.0266*** (0.0088)
Age 26-30	0.0168 (0.0433)	0.0235 (0.0440)	–	–
Age 31-35	0.0553 (0.0489)	0.0608 (0.0498)	–	–
Age 36-40	0.0601 (0.0484)	0.0635 (0.0493)	–	–
Age 41-45	0.0883* (0.0482)	0.0905 (0.0492)*	–	–
Age 46-50	0.1208** (0.0484)	0.1261 (0.0493)**	–	–
Age 51-55	0.1568*** (0.0493)	0.1615 (0.0500)***	–	–
Age 61-65	–	–	0.0399*** (0.0046)	0.0403*** (0.0046)
Age 66-70	–	–	0.0677*** (0.0079)	0.0694*** (0.0079)
Age 71-75	–	–	0.0756*** (0.0100)	0.0784*** (0.0101)
Age 76 +	–	–	0.0726*** (0.0151)	0.0760*** (0.0151)
Year Indicators	Yes	Yes	Yes	Yes
State Indicators	No	Yes	No	Yes
Percent Correctly Classified a	56.8	57.3	55.6	56.7
Pseudo R-Sq	0.017	0.030	0.009	0.016
Observations	9,602	9,552	70,215	70,078

Notes: Dependent variable is a dichotomous indicator representing the most risk-averse category. Marginal effects from Probit estimation are presented. Standard errors are clustered at the individual level, and presented in parentheses.

^aPercent correctly classified is based on a cutoff corresponding to the mean of the dependent variable in each sample.

In the first wave of the HRS, the risk instrument was administered to 11,707 individuals and to 8125 entrants into the study in subsequent waves. The resulting sample size is over 66,000 person-wave observations. The PSID data is limited to three waves that contain information on the respondent's alcohol use (1999, 2001, and 2003) and limited to the individuals who were asked the risk questions. The resulting sample size is over 9700 person-wave observations.

Information on the respondent's state of residence is available in the PSID and made available to this project for the HRS. Policy measures affecting the sale and consumption of alcohol are merged to records in both datasets based on the interview period and the state of residence. As a proxy for the cost of alcohol, the state excise tax on beer is utilized for several reasons.¹⁹ First, focusing on the state excise tax bypasses the simultaneity between price and demand. Changes in the state-level excise tax are plausibly exogenous to the individual's alcohol demand, often changing in response to the state's budgetary needs. Excise tax rates on wine and liquor are poor proxies for the prices of wine and liquor in control (monopoly) states because such states derive most of their revenue from the sale of wine and liquor from the price markups rather than from the excise taxes. Beer, however, is sold privately in monopoly states. Changes in excise tax rates within states over time are also strongly correlated with changes in the respective alcohol beverage price. A one-cent increase in the excise tax has been shown to raise the price by at least as much (Kenkel, 2005; Young & Bielinska-Kwapisz, 2002). Excise tax rates on beer, wine and liquor are also highly correlated as are their prices (correlation > 0.5). Thus, the beer tax provides a good proxy for the cost of alcohol while bypassing colinearity issues with including multiple tax rates in the specification. While some uniformity has emerged in certain other alcohol-related regulations such as the minimum purchase age or blood alcohol concentration (BAC) limits, there remains substantial variation in alcohol excise taxes. States have enacted widely differing rates, leaving considerable room for policy manipulation. For instance, the beer tax currently ranges from $0.02 per gallon in Wyoming to $1.07 per gallon in Alaska. Thus, estimates of the tax elasticity also provide a direct estimate of the effect of an important public policy tool. Furthermore, the price elasticity of alcohol demand can be recovered from the tax elasticity based on the percent of the tax represented in the price and the tax pass-through rate. Data on the state-level excise tax for beer are obtained from the Brewers Almanac, published annually by the US Brewers' Association (various years).

In order to control for state-level sentiment related to alcohol regulation and other policy shifts concurrent to taxes, three additional measures are included in the specifications. The first is a dichotomous indicator for those states that control the wholesaling and/or retailing of distilled spirits.²⁰ The second measure represents the percent of the state population residing in dry counties where there is no sale of alcoholic beverages (available from the Brewers Almanac). A dichotomous indicator is also created for whether the given state had an effective 0.08 BAC per se law in the month and year that the respondent was interviewed. In these states, it is illegal to drive or operate a motor vehicle with a BAC at or above 0.08% in and of itself, and impairment does not need to be demonstrated. As of July 2004, all states plus D.C. had adopted these laws. However, since the sample period covers 1992–2005, there was considerable variation due to the timing of when the laws were enacted in each state.

Weighted means for the full HRS and PSID samples along with stratification by risk preference are presented in Table 1. Risk-tolerant individuals are found to have a significantly higher prevalence of drinking as well as higher daily consumption. There are also important differences in risk preference across demographic groups. For instance, males are relatively more risk tolerant as are Whites. Risk-tolerant individuals are also less likely to be married, and their schooling distribution is shifted more towards higher levels of education. Both the PSID and the HRS show that risk preference is similarly positively correlated with parental education. Among older risk-tolerant individuals, more are likely to be unemployed and fewer are currently working or retired. Income also seems to increase with risk tolerance, consistent with a higher average return for individuals who may be bearing greater risks through self-employment or working in riskier occupations. There is some evidence across both age groups that risk-tolerant individuals have fewer indications of physical illnesses, but a higher indication of depression. It is noteworthy that certain characteristics of the state or residence also vary across risk-averse and risk-tolerant groups. For instance, those who are risk-averse are more likely to reside in states with stricter alcohol regulations: higher alcohol taxes, more likely to have enacted BAC 0.08 laws, higher percentage of the population residing in dry counties, and state monopoly on retailing and wholesaling of alcoholic beverages. While these simple correlations may reflect other observed and unobserved factors, they also point to risk preference as a potential confounder in an individual's alcohol demand and its sensitivity to prices and policy. The multivariate models presented in the next section account for these differences.

5. Results

Table 3 presents estimates of the baseline specification in Eq. (2), for adults (ages 21–54) based on the PSID and older adults (ages 55 and up) based on the PSID and the HRS. The first column estimates a basic model for adults utilizing a sparse set of covariates. Beer tax has a significant and negative effect on current alcohol participation, with the tax participation elasticity estimated at −0.04. This specification controls for unobserved state sentiment towards alcohol regulation by including indicators for the BAC 0.08 per se law and for whether the state controls the retailing and/or wholesaling of alcoholic beverage. The percent of the state population residing in dry counties is also included. Enactment of the BAC 0.08 law significantly reduces current drinking as does residing in a state with more dry counties. For instance, a 10% increase in the population residing in dry counties is associated with a 4% point decline in the probability of drinking. These regulations raise the non-monetary cost of consuming alcohol, through higher penalties and search or time costs, and therefore reduce alcohol participation. Residing in a monopoly state that controls the distribution of spirits seems to increase the probability of drinking. This may reflect substitution from the consumption of spirits to the consumption of beer or wine (Nelson, 2003; Holder & Wagenaar, 1990). Even if total ethanol consumption is lower in monopoly states, higher alcohol participation may be reflecting higher beer consumption relative to liquor.

Table 3.

Alcohol participation – PSID and HRS

	Ages 21 - 54			Ages 55 and older
Sample	PSID	PSID	PSID	HRS	HRS	PSID
	1	2	3	4	5	6
Real Beer Tax	-0.1588*** (0.0445) [ε = -0.037]	-0.2198 (0.2200) [ε = -0.050]	-0.1877*** (0.0735) [ε = -0.042]	-0.5308*** (0.1247) [ε = -0.224]	-0.4525*** (0.1618) [ε = -0.173]	-0.5721*** (0.1826) [-0.130]
Risk Averse	–	–	-0.0583*** (0.0147)	–	-0.0329*** (0.0074)	-0.0574* (0.0340)
Male	0.1681*** (0.0084)	0.1692*** (0.0084)	0.1453*** (0.0152)	0.1594*** (0.0057)	0.1677*** (0.0075)	0.1720*** (0.0368)
Black	-0.1595*** (0.0108)	-0.1420*** (0.0119)	-0.1495*** (0.0184)	-0.1088*** (0.0072)	-0.0812*** (0.0104)	-0.1242** (0.0493)
Other Race	-0.1769** (0.0225)	-0.1767*** (0.0228)	-0.1340*** (0.0519)	-0.1298*** (0.0124)	-0.1302*** (0.0166)	-0.4996*** (0.0738)
Hispanic	-0.1982*** (0.0194)	-0.2019*** (0.0207)	-0.0330 (0.0687	-0.0746*** (0.0102)	-0.0417*** (0.0142)	0.0160 (0.2487)
High School	0.0124 (0.0137)	0.0129 (0.0138)	0.0134 (0.0312)	0.0909*** (0.0073)	0.0583*** (0.0102)	-0.1750* (0.1023)
Some College	0.0278** (0.0139)	0.0241* (0.0140)	0.0557* (0.0282)	0.1718*** (0.0091)	0.1201*** (0.0121)	0.0093 (0.1276)
College	0.0850*** (0.0140)	0.0808*** (0.0141)	0.0666** (0.0293)	0.2405*** (0.0098)	0.1806*** (0.0130)	-0.1612 (0.1195)
Real Income	0.0021*** (0.0002)	0.0019*** (0.0002)	0.0019*** (0.0003)	0.0010*** (0.0002)	0.0008*** (0.0002)	0.0022*** (0.0005)
Married	-0.1567*** (0.0093)	-0.1539*** (0.0093)	-0.1222*** (0.0171)	0.0090 (0.0056)	0.0068 (0.0080)	-0.0541 (0.0432)
Age 26-30	-0.0078 (0.0139)	-0.0116 (0.0140)	-0.0324 (0.0385)	–	–	–
Age 31-35	-0.0182 (0.0156)	-0.0253 (0.0157)	-0.0274 (0.0406)	–	–	–
Age 36-40	-0.0230 (0.0155)	-0.0321** (0.0156)	-0.0424 (0.0405)	–	–	–
Age 41-45	-0.0518*** (0.0160)	-0.0620*** (0.0160)	-0.0606 (0.0407)	–	–	–
Age 46-50	-0.0612*** (0.0167)	-0.0717*** (0.0168)	-0.0697* (0.0415)	–	–	–
Age 51-55	-0.0779*** (0.0183)	-0.0897*** (0.0184)	-0.1025** (0.0440)	–	–	–
Age 61-65	–	–	–	-0.0195*** (0.0045)	-0.0175*** (0.0054)	-0.0193 (0.0338)
Age 66-70	–	–	–	-0.0390*** (0.0062)	-0.0452*** (0.0082)	-0.0250 (0.0451)
Age 71-75	–	–	–	-0.0568*** (0.0066)	-0.0537*** (0.0102)	-0.0058 (0.0556)
Age 76-80	–	–	–	-0.1240*** (0.0063)	-0.0765*** (0.0146)	-0.1091 (0.0820)
Employed	–	–	0.1044*** (0.0247)	–	0.0416*** (0.0105)	0.1348** (0.0650)
Unemployed	–	–	0.1289*** (0.0304)	–	0.0358 (0.0235)	0.1057 (0.1105)
Retired	–	–	0.1423** (0.0544)	–	0.0565*** (0.0101)	0.1249* (0.0629)
Household Size	–	–	-0.0219*** (0.0057)	–	-0.0182*** (0.0029)	-0.0109 (0.0206)
Parental Education	–	–	0.0358* (0.0186)	–	0.0340*** (0.0078)	0.0890 (0.0534)
Functional Difficulties	–	–	-0.0399 (0.0248)	–	-0.0294*** (0.0026)	-0.0447** (0.0228)
Hypertension	–	–	0.0118 (0.0178)	–	-0.0094 (0.0069)	-0.0381 (0.0329)
Diabetes	–	–	-0.0229 (0.0330)	–	-0.1372*** (0.0084)	-0.1061** (0.0477)
Heart Disease	–	–	-0.0242 (0.0369)	–	-0.0371*** (0.0087)	-0.0056 (0.0432)
Stroke	–	–	0.0144 (0.0582)	–	-0.0672*** (0.0146)	-0.0863 (0.0795)
Depression	–	–	0.0108 (0.0074)	–	-0.0040** (0.0016)	0.0187 (0.0174)
BAC 08	-0.0218*** (0.0083)	–	-0.0143 (0.0141)	–	–	-0.0423 (0.0319)
Percent Dry	-0.4052*** (0.0466)	–	-0.3725*** (0.0744)	–	–	-0.0905 (0.1683)
Monopoly State	0.0443*** (0.0092)	–	0.0515*** (0.0150)	–	–	0.0146 (0.0334)
Year Indicators	Yes***	Yes**	Yes***	Yes***	Yes***	Yes
State Indicators	No	Yes***	No	Yes***	Yes***	No
Pseudo R2	0.077	0.089	0.089	0.135	0.145	0.117
Percent Correctly Classified	62.5	63.4	65.7	67.2	68.0	67.2
Observations	26,762	26,762	9,545	107,509	66,374	1,935

Notes: Specifications are estimated via Probit, and marginal effects are reported. Standard errors in parentheses are clustered robust. Tax elasticities, evaluated at the sample means, are reported in brackets. Significance is denoted as follows: ***Significant at the 1% level. **Significant at the 5% level. *Significant at the 10% level. Significance of the tax effect is based on a one-tailed test. Percent correctly classified is calculated using a cutoff based on the observed mean prevalence of alcohol participation for each corresponding sample.

Since the PSID analysis is based on three waves, it is not feasible to control for unobserved state sentiment through state-level fixed effects. There is not sufficient variation in the excise tax rate within each state over this period to allow for the fixed effects. For example, regressing the tax on state and year fixed effects yields an R-square of 0.99, suggesting that only about 1% of the variation in taxes represents within-state variation. Hence, specification 1 proxies for unobserved state sentiment through other measures of alcohol regulation. Specification 2 controls for state-level fixed effects for comparison. As expected, the lack of state-specific time-series variation results in imprecision, inflating the standard error for the tax effect. However, it is somewhat reassuring that the magnitude of the effect and the elasticity remains similar. Subsequent models utilizing the PSID omit the state fixed effects in favor of a more parsimonious set of controls for state sentiment.

Specification 3 expands on the basic model by incorporating risk aversion, labor market behavior, parental education, and health measures. Risk-averse individuals have a significantly lower probability of current alcohol participation relative to those who are more tolerant of risk. Unadjusted means showed a difference in participation by about 11% points. In the multivariate model, this effect diminishes, though still remaining sizeable at 5.8% points. The tax elasticity is significantly negative and remains stable at −0.04.

Demographic variations in alcohol demand are often taken for granted and attributed to differences in tastes or culture. However, incorporation of risk preference in the alcohol demand function can explain some of these differences. For instance, males participate more in drinking. However, after controlling for risk preference in specification 3, the marginal effect of being male on participation declines by over 2% points.²¹ This is consistent with males being more tolerant of risk, as indicated by the simple means. Similarly, it is true that non-Whites have been consistently found to have a lower drinking prevalence. Part of this effect may reflect differences in attitudes towards risk, especially for Blacks who are more likely to be risk averse. Similarly, while married individuals drink less, the magnitude of the effect is diminished by controlling for risk since married individuals tend to be relatively more risk averse.

Educated individuals are more likely to participate in drinking, though here also the magnitude of the effect diminishes for those who have completed college since there is a positive association between risk preference and education. There is a small significantly positive effect of income on drinking, with the elasticity estimated at 0.10.²² The coefficients on the age categories confirm that drinking prevalence declines with age. Relative to individuals not in the labor force (disabled, homemakers, students), those who are employed, unemployed, or retired also have higher alcohol participation. Drinking is negatively associated with household size. This may reflect lower drinking prevalence in households with children present or in family units, where the individual may be internalizing the external costs of their drinking on household members. For the PSID sample, the effects of health measures are generally insignificant. This may be due to the low prevalence of some of these illnesses in the adult population, or it may reflect that at least for younger adults (21–54), these health outcomes may not have consistent effects on drinking propensity. For older adults (below), these health conditions lead to significant reduction in alcohol participation.

Specifications 4 and 5 estimate the demand function for adults 55 years of age and older, based on the HRS. Since the HRS sample is observed for a maximum of seven waves (1992–2005), there is sufficient variation in excise taxes within states over this period to allow for state fixed effects. Increases in the beer tax significantly reduce current drinking among older adults. Furthermore, the older demographic appears to be far more price-sensitive relative to younger adults; the tax elasticity is about four to five times larger and is estimated at between −0.17 and −0.22. To ascertain that this increase in the tax response relative to the younger demographic is not an artifact of the sampling differences between the PSID and the HRS, a similar specification is also estimated for the PSID restricted to older adults. Specification 6 suggests that the estimated marginal effect and elasticity are robust across both the PSID and the HRS samples. These estimates indicate that drinking among older adults is more sensitive to prices, relative to the general population.²³

The simple means suggest that older risk-averse individuals have a lower prevalence of drinking by about 5.4% points. Specification 5 shows that while the effect diminishes somewhat after controlling for other confounders, it remains significant at 3.3% points. The effects of the other covariates are generally similar to those discussed above.

The next set of specifications reported in Table 4 explore whether response to taxes varies across risk-averse and risk-tolerant individuals across both age groups. These models estimate Eqs. (3) and (4), stratifying the samples based on risk tolerance. For both age groups, increases in the beer tax significantly reduce alcohol participation, with higher elasticities being estimated for the older demographic as before. Among younger adults from the PSID, the marginal effect of the tax is relatively similar for both risk-averse and risk-tolerant individuals.²⁴ A likelihood-ratio (LR) test was implemented to check if the marginal effect is significantly different across both risk-aversion categories;²⁵ there is no significant difference. For older adults in the HRS, the marginal effect and the tax elasticity suggest that risk-tolerant individuals may be more sensitive to prices relative to those who are more risk averse. However, this difference is not statistically significant based on the LR test. These estimates suggest that even among groups that are more likely to consume and abuse alcohol, demand is relatively sensitive to prices.

Table 4.

Alcohol participation stratified by risk preference

	PSID: Ages 21 - 54		HRS: Ages 55 and older
Sample	Risk Averse	Risk Tolerant	Risk Averse	Risk Tolerant
	1	2	3	4
Real Beer Tax	-0.2157** (0.1094) [ε = -0.056]	-0.1822** (0.0977) [ε = -0.037]	-0.3863** (0.1942) [ε = -0.159]	-0.5987** (0.2902) [ε = -0.205]
Male	0.1988*** (0.0227)	0.0875*** (0.0203)	0.1641*** (0.0094)	0.1729*** (0.0126)
Black	-0.1056*** (0.0264)	-0.1918*** (0.0257)	-0.0750*** (0.0126)	-0.0904*** (0.0181)
Other Race	-0.1006 (0.0826)	-0.1775*** (0.0671)	-0.1211*** (0.0205)	-0.1367*** (0.0286)
Hispanic	-0.0417 (0.1095)	-0.0349 (0.0843)	-0.0613*** (0.0169)	0.0012 (0.0257)
High School	0.0294 (0.0428)	-0.0054 (0.0455)	0.0586*** (0.0123)	0.0575*** (0.0181)
Some College	0.1176*** (0.0402)	-0.0107 (0.0401)	0.1252*** (0.0151)	0.1159*** (0.0203)
College	0.1401*** (0.0411)	-0.0029 (0.0404)	0.1573*** (0.0164)	0.2200*** (0.0213)
Real Income	0.0013*** (0.0004)	0.0026*** (0.0004)	0.0007*** (0.0002)	0.0007** (0.0003)
Married	-0.1390*** (0.0260)	-0.1139*** (0.0219)	0.0094 (0.0099)	0.0046 (0.0135)
Employed	0.1146*** (0.0361)	0.0881*** (0.0323)	0.0443*** (0.0129)	0.0365** (0.0178)
Unemployed	0.1549*** (0.0483)	0.1108*** (0.0372)	0.0447 (0.0311)	0.0220 (0.0362)
Retired	0.1927** (0.0696)	0.0769 (0.0942)	0.0562*** (0.0124)	0.0577*** (0.0173)
Household Size	-0.0167** (0.0085)	-0.0262*** (0.0075)	-0.0173*** (0.0036)	-0.0193*** (0.0049)
Parental Education	0.0561* (0.0302)	0.0157 (0.0231)	0.0316*** (0.0095)	0.0404*** (0.0135)
Functional Difficulties	-0.0466 (0.0343)	-0.0373 (0.0360)	-0.0280*** (0.0032)	-0.0329*** (0.0045)
Hypertension	-0.0008 (0.0266)	0.0210 (0.0235)	-0.0097 (0.0084)	-0.0089 (0.0119)
Diabetes	-0.0498 (0.0465)	-0.0002 (0.0461)	-0.1401*** (0.0101)	-0.1278*** (0.0149)
Heart Disease	-0.0423 (0.0532)	0.0002 (0.0491)	-0.0258** (0.0107)	-0.0620*** (0.0149)
Stroke	0.0949 (0.0787)	-0.0924 (0.0836)	-0.0503*** (0.0181)	-0.0990*** (0.0246)
Depression	0.0023 (0.0114)	0.0180* (0.0096)	-0.0022 (0.0020)	-0.0066*** (0.0028)
BAC 08	-0.0127 (0.0211)	-0.0130 (0.0186)	–	–
Percent Dry	-0.3604*** (0.1066)	-0.3651*** (0.1055)	–	–
Monopoly State	0.0585** (0.0228)	0.0412** (0.0196)	–	–
Age Indicators	Yes	Yes**	Yes***	Yes***
Year Indicators	Yes**	Yes	Yes***	Yes***
State Indicators	No	No	Yes***	Yes***
Pseudo R2	0.077	0.100	0.143	0.153
Percent Correctly Classified	63.6	66.4	67.5	68.6
Observations	4,355	5,190	43,181	23,183

Notes: See Table 3.

The models thus far have considered the response of taxes at the extensive margin, on the decision to currently consume alcohol. Policies may also affect the intensity of consumption, conditional on being a drinker. Data on average daily number of drinks consumed from the HRS allow estimation of the alcohol demand function at the intensive margin. Specification 1 in Table 5 suggests that taxes can also reduce the level of alcohol consumption, though the consumption elasticity (−0.086) is smaller in magnitude compared to the participation elasticity for older adults. This is consistent with demand studies for the general population, which have also found larger elasticities at the extensive margin; that is, most of the price effect operates through the decision to drink rather than the level of drinking conditional on participation (Chaloupka, Grossman, & Saffer, 2002; NIAAA, 2000; Manning et al., 1995). Risk aversion continues to be a significant predictor of alcohol consumption. Across models 1 and 2, risk-averse drinkers consume between 5.5% and 8.6% less alcohol.

Table 5.

HRS: older adults – specification checks

Sample	Intensive Margin				Ever Drinkers		Time Preference
					Current Price Effect	Current & Lagged Price Effects	All	Risk Averse	Risk Tolerant
Dependent Variable	Log Drinks	Log Drinks	Chronic Consumption	Chronic Consumption	Alcohol Participation	Alcohol Participation	Alcohol Participation	Alcohol Participation	Alcohol Participation
Specification	1	2	3	4	5	6	7	8	9
Real Beer Tax	-0.6243* (0.4064) [ε = -0.086]	-0.8929** (0.4281) [ε =-0.109]	-0.2023* (0.1330) [ε = -0.278]	-0.1909* (0.1374) [ε = -0.265]	-0.3921** (0.2040) [ε = -0.100]	Pt: -0.3431 (0.4908) Pt-1: -0.0496 (0.4656)	-0.5717*** (0.1739) [ε = -0.216]	-0.5355*** (0.2085) [ε =-0.215]	-0.6031** (0.3094) [ε = -0.208]
Risk Averse	-0.0549*** (0.0193)	-0.0864*** (0.0337)	-0.0063 (0.0057)	-0.0041 (0.0100)	-0.0212** (0.0085)	-0.0212** (0.0085)	-0.0322*** (0.0086)	–	–
Beer Tax * Risk Averse	–	0.2298 (0.1987)	–	-0.0156 (0.0575)	–	–	–	–	–
Planning Horizon 5-10 years	–	–	–	–	–	–	0.0402** (0.0094)	0.0386*** (0.0117)	0.0425*** (0.0159)
Planning Horizon 10+ years	–	–	–	–	–	–	0.0571** (0.0156)	0.0406** (0.0194)	0.0822*** (0.0263)
Year Indicators	Yes**	Yes**	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***
State Indicators	Yes	Yes	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***
R2/Pseudo R2	0.080	0.080	0.083	0.083	0.141	0.141	0.130	0.131	0.135
Percent Correctly Classified	–	–	61.0	60.9	67.4	67.4	67.5	67.3	67.0
Observations	24,635	24,635	24,612	24,612	46,212	46,212	54,237	35,423	18,796

Notes: See Table 3. Each column represents a separate regression model. All specifications include the extended set of covariates listed in Table 3.

Model 2 also allows for the tax effect at the intensive margin to differ between the risk categories. Since alcohol participation among older adults is about 35 percent, estimating the conditional demand significantly restricts the sample size. Thus, a parsimonious specification, which includes an interaction term between the beer tax and risk aversion, is employed to allow for any difference in the tax effect. The coefficient on the interaction term is positive, suggesting that risk-tolerant individuals may be more sensitive to taxes. However, the large standard error on the interaction effect does not allow the rejection of the null that the tax response is similar across both groups. This is consistent with the alcohol participation models which also followed the same pattern.

While these results suggest that excise taxes do have an impact on the level of alcohol use, across both risk groups, it is specifically problem drinking that imposes externalities and is the target of public policy. Specifications 3 and 4 therefore examine the tax response on chronic drinking, conditional on being a current drinker. Chronic drinking is defined by a dichotomous indicator for whether the respondent consumes more than two drinks daily, on average.²⁶ Approximately 10% of drinkers in the HRS fall in this category. The estimated tax elasticity is significantly negative and of a similar magnitude to the participation elasticity among all individuals. The interaction effect in model 4 is insignificant, suggesting that taxes can reduce the propensity of chronic alcohol use even among risk-tolerant individuals. Risk aversion has a negative effect on problem consumption, though the coefficients are imprecisely estimated. Unadjusted means show that the prevalence of chronic drinking is significantly higher among risk-tolerant individuals by about 10%.

Specification 5 exploits the longitudinal waves of the HRS and restricts the sample to ever-drinkers. This results in a slight difference in the control group that is used for identification of the tax response. That is, among those individuals who drink, some may shift in and out of current drinking status. This model checks whether this response is related to changes in the excise tax. Thus, individuals who never drink and therefore do not change their drinking status over the sample period are excluded. The tax elasticity declines somewhat in magnitude from −0.17 to −0.10. This is presumably because current drinking propensity for ever-drinkers is less responsive to prices. Also, in the prior specification, never-drinkers served as part of the control group used to compare the responses of those who do drink. By excluding the never-drinkers, ever-drinkers who change their alcohol participation are being compared to those who do not. Employing this alternative control group for identification, the tax elasticity of older adults continues to be larger in magnitude relative to younger adults.

Economic models of addiction typically predict that current consumption depends on past consumption (as a proxy for the accumulated addictive stock) due to the reinforcement effect. Thus, current addictive consumption would also be a function of past prices.²⁷ Specification 6 includes both the current and the 1-year lagged excise tax. Since excise taxes within a state are highly correlated over time due to infrequent changes, the standard errors are substantially inflated. However, judging from the magnitudes, the current and lagged tax effects in specification 6 almost add up to the current tax effect in specification 5 that excludes the lagged excise tax. This suggests that the contemporaneous tax response may also be picking up the effect of past taxes, and may in some sense be more indicative of the long-run elasticity that takes account of changes in past and present prices on current consumption.

While the estimates and differences have been interpreted with respect to the economist's concept of risk aversion, one potential concern is that the risk instrument may be reflecting differences in time preference rather than attitudes towards risk per se. Theoretically, some of the effects of risk preference are difficult to disentangle from the effects of differential discount rates. For instance, more present-oriented individuals with a high discount rate would also be predicted to participate more in risky activities such as alcohol and cigarette use since they are likely to discount the future harmful consequences. From a policy perspective, it is immaterial whether the risk-tolerant group is reflecting a more dismissive attitude towards risk or a higher discount rate. Since these individuals, regardless of whether they are risk-tolerant or present-oriented or both, are heavier consumers of alcohol, it is important to explore the extent to which their demand is sensitive to variations in taxes. Nevertheless, for interpretation purposes, it may be helpful to determine that the risk preference instrument is indeed picking up variation in attitudes towards risk.

The HRS administered a module on preferred consumption paths to 198 respondents in order to elicit estimates of time preference parameters. Barsky et al. (1997) show that for these individuals, their degree of risk aversion is uncorrelated with time preference. Alternately, specifications 7–9 in Table 5 also confirm that the risk preference categories are reflecting variations beyond discount rates. Individuals in the HRS are also asked about their relevant financial planning horizon. While certainly prone to measurement error and noise, variations in the planning horizon, conditional on age, would partially reflect differences in time preference. Accounting for age, more future-oriented individuals with lower discount rates should take greater account of future events and therefore have more distant planning horizons. The marginal effects and elasticity estimates remain robust to controlling for the individual's reported planning horizon. The tax elasticity for older adults continues to be larger than younger adults, and there are no significant differences in the tax response across risk groups. Combined with earlier evidence that the risk preference instrument correctly predicts behaviors that it would be expected to predict a priori (self-employment, insurance status, etc.), there is some evidence that the measure is indeed indicative of an individual's propensity to undertake risk.

Selective sample attrition is not a concern with the PSID sample since the analysis employs only three waves (1999–2003) and the demographic of interest comprises young adults in relatively good health. In the HRS, however, selective attrition may be relevant due to the longer period of study and the older demographic. The average mortality rate between waves is 2.3%. Thus, about 14% of the individuals who were surveyed in the first wave (1992) have died by the seventh wave (2004). The mortality rate for the HRS sample is consistent with the Social Security Administration life table mortality rates (Kapteyn, Michaud, Smith, & Van Soest, 2006).

The specific concern is that since alcohol-related illnesses are a significant cause of premature death, mortality among the heavy drinkers may lead to a progressively selective sample in later waves that consumes less alcohol. Results thus far consistently indicate that older adults are far more responsive to tax policies than younger adults. It is important to determine if this effect is being driven by attrition bias. If the heavier drinkers, who may be less responsive to price, are being progressively excluded from the sample, then the remaining price response may be biased upwards. Controlling for physical and mental health status alleviates some of this concern. Table 6 also presents two additional strategies which provide information on potential bias due to such attrition.

Table 6.

HRS: older adults – sample attrition

Sample	Sample Attrition
	Balanced Sample			Inverse Probability Weighting a
	All	Risk Averse	Risk Tolerant	All	Risk Averse	Risk Tolerant
Dependent Variable	Alcohol Participation	Alcohol Participation	Alcohol Participation	Alcohol Participation	Alcohol Participation	Alcohol Participation
Specification	1	2	3	4	5	6
Real Beer Tax	-0.5871*** (0.1998) [ε = -0.219]	-0.5446** (0.2398) [ε = -0.217]	-0.6751** (0.3621) [ε = -0.224]	-0.3671** (0.1797) [ε = -0.147]	-0.2902* (0.2115) [ε = -0.125]	-0.5451* (0.3331) [ε = -0.185]
Year Indicators	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***
State Indicators	Yes***	Yes***	Yes***	Yes***	Yes***	Yes***
R2/Pseudo R2	0.151	0.153	0.130	0.137	0.134	0.146
Percent Correctly Classified	68.4	68.4	68.6	67.6	67.6	67.7
Observations	40,599	26,544	14,036	55,137	35,740	19378

Notes: See Table 3. Each column represents a separate regression model. All models include the extended set of covariates listed in Table 3.

^aInverse probability weights are predicted using baseline characteristics (gender, race, ethnicity, education, parental education, religion, and native-born) along with other time-varying factors (age indicators, wave indicators, census division indicators), lagged covariates (income, marital status, health insurance), and health status and alcohol consumption in the prior wave.

First, specifications 1–3 utilize a balanced sample that only includes individuals who are observed in all seven waves. If selective attrition is severe, then results from the unbalanced panel (Tables 3 and 4) versus the balanced panel would be expected to be different. Comparing the marginal effects and tax elasticities from the balanced panel to those reported earlier, there are no material differences.

The second approach employs inverse probability weights (IPW) to adjust for selection bias due to observable characteristics (Fitzgerald, Gottschalk, & Moffit, 1998). This involves using baseline characteristics (gender, race, ethnicity, education, parental education, religion, and native-born) along with other time-varying factors (age indicators, wave indicators, census division indicators) and lagged covariates (income, marital status, and health insurance) to predict survival status. Most importantly, observed illness conditions in the previous wave and the number of drinks consumed in the previous wave are also included to predict survival. Since past health status is observed, this model is able to correctly predict about 77% of the attritors, based on a very conservative cutoff of 0.9 for the predicted probability; with the standard cutoff of 0.5, the prediction rate is 92%. The IPW correction involves weighting observations by 1/p_i, where p_i represents the probability of survival, therefore giving more weight in the regression to those individuals whose observable characteristics predict higher attrition rates. The results in specifications 4–6 show that the elasticity magnitudes decline somewhat as expected, since attrition bias may have inflated the price response. However, the general pattern of results and conclusions remains unaffected.

6. Conclusions

Past estimates of the alcohol demand functions and the price elasticity have not accounted for variations in individuals' preference towards risk. While some of the prior studies have included individual-level fixed effects to control for unobserved heterogeneity, they have not been able to directly estimate the effects of risk preference and also fail to consider heterogeneous responses across measures of risk tolerance. Accounting for risk can partially explain persistent demographic variations in alcohol consumption. For instance, drinking prevalence is higher among males due, in part, to a higher tolerance towards risk. Similarly, a higher degree of risk aversion among Blacks and married individuals can also partially explain why these individuals are less likely to consume alcohol.

Risk-tolerant individuals are found to participate more in drinking and have a higher level of consumption. This group, through their higher consumption, also bears a greater share of the responsibility of the societal costs of alcohol abuse. Thus, it is important to determine whether their alcohol use responds to tax policy. There are some indications that while overall drinking prevalence remained stable over the past decade, problem drinking may be trending upwards. Data from the Behavioral Risk Factor Surveillance System (BRFSS) show that the median prevalence of binge drinking has increased from 14.4% (1992) to 16.1% (2002). Similarly, chronic drinking has also increased from 3.0% to 5.9% over this period.²⁸ Thus, it is important to determine whether their alcohol use responds to tax policy. Since the majority of drinkers consume alcohol safely with little external harm, the rationale for higher taxes falls on whether high-participation groups curb their consumption in response to higher costs. There is a tradeoff in terms of the social gains from making heavy users face higher prices that reflect the full social costs of their drinking versus the burden of higher prices on moderate drinkers (Manning et al., 1995). If, for instance, risk-tolerant individuals are not price-sensitive then increasing taxes may have a greater proportional effect on light or moderate drinkers and would not be effective in reducing alcohol-related harms. Estimates from this study, however, suggest that this is not the case; risk-tolerant individuals are equally responsive to excise taxes as are risk-averse individuals.²⁹ These results therefore strengthen the rationale for raising alcohol excise taxes as a policy tool for deterring use among groups likely to over-consume alcohol.