Further evidence about alcohol consumption and * the business cycle by Sergi Jiménez-Martín^** José M. Labeaga^*** Cristina Vilaplana Prieto^**** DOCUMENTO DE TRABAJO 2006-06

February 2006

This study was supported by an unrestricted educational grant from The Merck Foundation, the philanthropic arm of Merck & Co. Inc. Whitehouse Station, New Jersey, USA and project SEJ2005-08793-C04-01-04. The authors thank participant at the XXX Simposio of Análisis Económico for their comments. The usual disclaimer applies.

** Universitat Pompeu Fabra.

*** FEDEA and UNED.

**** Universidad Católica San Antonio de Murcia.

Los Documentos de Trabajo se distribuyen gratuitamente a las Universidades e Instituciones de Investigación que lo solicitan. No obstante están disponibles en texto completo a través de Internet: http://www.fedea.es.

Further evidence about alcohol consumption and the business cycle

Sergi JimÈnez-MartÌny JosÈ M. Labeaga z Cristina Vilaplana Prietox

Abstract

The main goal of this paper is to test whether macroeconomic conditions a§ect alcohol consumption using data from the Behavioral Risk Factor Surveillance System for the period 1987ñ2003. We try to control unobserved heterogeneity by relying on the construction of pseudopanel data from the di§erent cross-sections available. Our results indicate that when we do not take into account unobserved heterogeneity, the unemployment rate is signiÖcant and reduces the probability of becoming drinker and the number of alcoholic beverages consumed. However, once we estimate the model using cohort data, controlling for both observed and unobserved heterogeneity, the unemployment rate becomes non-signiÖcant. This implies that unobserved e§ects are important when explaining alcohol consumption. As a result, inferences obtained without controlling for them should be interpreted with caution.

JEL-CLASS: E23, I12

KEYWORDS: alcohol consumption, Behavioral Risk Factor Surveillance System, economic cycle, macroeconomic conditions

yUniversitat Pompeu Fabra

This study was supported by an unrestricted educational grant from The Merck Foundation, the philanthropic arm of Merck & Co. Inc. Whitehouse Station, New Jersey, USA and project SEJ2005-08793-C04-01-04. The authors thank participant at the XXX Simposio of An·lisis EconÛmico for their comments. The usual disclaimer applies.

zFEDEA and UNED

xUniversidad CatÛlica San Antonio de Murcia

1 Introduction

Concerns about the economic implications of the relationship between alcohol consumption and the labor market are well grounded. Most of the empirical literature has maintained the commonly held view that alcohol drinking is associated with lower earnings, lower employment rates, greater unemployment and productivity losses. Nevertheless, some authors (Forcier, 1988; Catalano et al., 1993) emphasize that the direction of the casuality between unemployment and alcohol consumption is not conclusive. While some studies have shown that unemployment is positively correlated with alcohol consumption (Kessler et al., 1987), with alcohol abuse (Crawford et al., 1987) and with diseases and psychological problems derived from alcohol abuse (Catalano et al., 1993), other analyses suggest that the correlation is inexistent or even negative (Ettner, 1997; Rhum and Black, 2002).

A common argument within the Örst set of studies is that unemployment originates a situation of Önancial strain, which induces the individual to canalize stress through consumption of alcohol (Peirce et al., 1994). While some authors support the existence of a positive relation between Önancial strain and depression, understanding chronic Önancial strain as a situation in which is di¢cult to satisfy basic needs (Kessler et al., 1987; Hamilton et al., 1990), others have found a positive relation between depression and alcohol consumption (Hartka et al., 1991). Analyses within the second set argue that unemployment usually implies lower consumption through an income e§ect. This reduction could not happen when unemployment is transitory and the unemployed receive beneÖts or family support (see Bentolila and Ichino, 2003).

Most of the previous research is based on the existence of a representative consumer. This is a non-realistic approach because there could coexist economic and sociological factors (intrinsic to individuals) that do not allow to generalize results about participation and alcohol consumption. Recent works use individual data and relax the representative consumer assumption (Dee, 2001 and Ruhm and Black, 2002 are two good examples). In this work we propose a further step to explicitly consider unobserved heterogeneity among individuals. Since panel data is not available we rely on cohorts built from independent cross-sections taken from the Behavioral Risk Factor Surveillance System (BRFSS from now on).

The objectives of the paper are twofold. The Örst and fundamental one is to show the e§ects of misspeciÖcation caused by missing unobserved heterogeneity, which even uncorrelated with the regressors could bias the parameter estimates or their standard errors. As a previous step we also want to replicate results obtained by previous authors, especially by Dee (2001) and Ruhm and Black (2002) in samples of di§erent time dimension, in order to avoid the critique of obtaining di§erent results because of using di§erent sample periods. Finally, we also like to show that when there is a high percentage of zero observations in demand equations using individual data, it is sometimes necessary to consider the underlying reasons generating them when adjusting the model

The structure of the paper is the following one: in section 2 we describe the methodological aspects of the models studying the relation between alcohol consumption and the cycle and we propose alternative speciÖcations. Section 3 describes the dataset. Section 4 is devoted to comment on the results using individual and cohort data. In section 5, we propose econometric and economic interpretation of the results. Finally, section 6 summarizes the main conclusions.

2 Model and relationship to previous literature

Suppose that Y is an indicator of whether an individual is or not a drinker or the number of drinks consumed by him, whose latent variable Y is a linear function of some explanatory variables.

We observe as result of comparing the utility of consuming a number of drinks including zero consumption. So, the observability rule is for the binary choice being a drinker or not or for the number of drinks consumed. 1(A) is the indicator of event A.

Consider a general linear model for the latent variable:

\[Y _ {i s m t} ^ {*} = X _ {i s m t} \beta + U R _ {s m t} \gamma + \alpha_ {s} + \delta_ {m} + \lambda_ {t} + \eta_ {i} + \varepsilon_ {i s m t}\tag{1}\]

where the observed counterpart of denotes alcohol consumption (number of drinks) or the decision to drink of individual i interviewed in state s in month of year t; X is a vector of explanatory variables, UR refers to unemployment rate, " is an error term and ; ; and are state, month, year and individual unobserved factors.

Let suppose that X gathers all the determinants of the probability of being drinker or of the number of drinks consumed. Then, this model is equivalent to the one proposed by Ruhm and Black (2002), in which we allow the possibility that the dependent variable be limited or qualitative (binary or a count).¹

The lack of panel data requires some assumptions to identify the parameters, or in other words, in their absence we cannot control the . If we assume for all Ordinary Least Squares (in the case Y were a continuous variable) would provide consistent estimates of the parameters. On the other hand, we only require absence of correlation among the and the regressors for the consistency of the parameters with individual random e§ects. In any case, from an economic point of view a model that does not allow correlation between individual e§ects and explanatory variables does not seem very interesting. For example, if individual tastes were correlated with professional occupation, then the coe¢cients corresponding to occupation would be biased when unobserved e§ects are not controlled for. If unemployment rates were di§erent across occupations, then correlation with unobserved heterogeneity moves to the variables that proxy the economic situation. When panel data is available, this problem can be solved by treating as Öxed e§ects, using a transformation of the model or parameterizing the conditional expectation of the individual e§ects as a function of the explanatory variables. Obviously, it is not possible to apply these strategies if we do not have repeated observations for the same individuals. This last situation is analyzed by Dee (2001) and Ruhm and Black (2002) using information from the BRFSS for periods 1984-1995 and 1987-1999, respectively, with the aim of testing the relation between unemployment and consumption of alcoholic drinks. Our Örst priority will then be to reply their exercises before presenting results controlling for Öxed e§ects.

Since the BRFSS is a combination of independent cross-sections, we cannot control for unobservable characteristics a§ecting consumption decisions (i.e., preferences for working, di§erent tastes, religious beliefs, genetics, etc.). Moreover, unobserved variables could be correlated with regressors in (1) and so, the e§ect of unemployment on consumption would not be properly identiÖed. We can deal with this problem by constructing pseudo-panels. Deaton (1985) suggests to divide the population in homogeneous groups (cohorts) according to one or several characteristics. At the population level, groups have to contain the same individuals along time. The basic idea of this procedure is to construct population means of the cohorts, in order to form a panel structure for the data. While it is true that cohort population means are not observable, we can use their sample analogs to proxy them, being aware that we end up with an errors in variables model. The advantage with respect to usual errors in variables models is that we can estimate the variances of the measurement errors using individual data. Moreover, if the size of the cohort is large enough (Deaton, 1985, establishes 150 observations per cell), we can forget measurement errors because sample means approximate well enough their population counterparts.

¹ From the speciÖcation above we can also generate simultaneous models if we established a double hurdle decision for consuming alcoholic drinks (Tobit type II, for example if Y were continuous for a part of the sample or Hurdle-Poisson or negative binomial for the decision and the counts). In case that variables a§ecting participation and consumption were the same and had identical e§ects over both decisions, we would be in the case of a standard Tobit (Poisson or negative binomial) model.

From (1), we derive the cohort speciÖcation by adding up in i (for all individuals who satisfy the aggregation criterion deÖned) and dividing by the sample size of the group. Hereby we have:

\[\bar {Y} _ {c m t} = \bar {X} _ {c m t} \beta + \overline {{U R}} _ {c m t} \gamma + \bar {\alpha} _ {s} + \bar {\delta} _ {m} + \bar {\lambda} _ {t} + \bar {\eta} _ {c} + \bar {\varepsilon} _ {c m t} c = 1, \dots , C\tag{2}\]

We deÖne as the size of cohort c in month m of year t. Every element of , for example the unemployment rate, is the average of the unemployment indicators observed for individuals belonging to cohort c in month m of year and analogously for other variables in the model. The main estimation problem is that is unobservable and probably still correlated with some variables in . Therefore, (2) does not constitute an appropriate base for obtaining consistent estimates unless the size of the cohorts is large enough. In this case, is a good approximation to and we can replace by a set of binary variables (Öxed e§ects) one for each cohort.

Then a natural estimator is the covariance or within groups estimator based on the weighted means of the cohorts, introducing weights to take into account potential heteroskedasticity between cohorts. Let be the average of the observed means for cohort and deÖne analogously. Then:

\[\hat {\beta} _ {W G} = [ n _ {c m t} (\bar {X} _ {c m t} - \bar {X} _ {c}) ^ {\prime} (\bar {X} _ {c m t} - \bar {X} _ {c}) ] ^ {- 1} [ n _ {c m t} (\bar {X} _ {c m t} - \bar {X} _ {c}) ^ {\prime} (\bar {Y} _ {c m t} - \bar {Y} _ {c}) ]\]

will be biased in small samples but it will be consistent as tends to inÖnity if standard assumptions about second order moments are met. There exists a trade-o§ between accuracy and number of pseudo-panel observations. The bigger is the number of cohorts (C), the smaller is their size , which implies a trade-o§ between bias and variance of the estimator.

The identiÖcation of the e§ects of all determinants of participation and consumption becomes more complicated when we rely on time series or cross-section data. In the Örst case, we have to claim for the existence of a representative consumer and it is only possible to identify and . For example, Brenner (1975, 1979) argues that during recessions, mortality rates increased, although there was a lag between the growth of unemployment rate and the increase in mortality rates. Nevertheless, Brennerís work has been very criticized by other authors as Gravelle et al. (1981), Stern (1983) or Wagsta§ (1985) arguing absence of rationality in the election of the unemployment rate lag structure, high collinearity and variability of the results conditioned on time period, country or proxies for health status chosen.

When using a single cross-section instead of time series data, a similar problem appears. In terms of (1) we only can estimate separately and (also and if we have information for month and state of the observation) . If cross-section data correspond to countries, states or regions, the problems of infra-speciÖcation and existence of representative consumer are equivalent to the previous ones (Jahoda, 1991). In a photograph of individual alcohol consumption, the advantage is that X will contain a wide range of demand determinants (income and socioeconomic characteristics). The disadvantage is that we are not going to be able to establish any result relating alcohol consumption and the economic cycle. An additional problem is the possible endogeneity of the unemployment rate since poor health may be the cause rather than the consequence of unemployment (Janlert et al., 1991). Some authors (Hammarstrˆm et al., 1988, for instance), have tried to test health status of employed and unemployed workers but only have managed to capture part of the impact of changes in economic conditions, since recessions do not a§ect only the unemployed workers.

Another alternative is to use Öxed e§ects panel data models for states or countries:

\[Y _ {s m t} = X _ {s m t} \beta + U R _ {s m t} \gamma + \alpha_ {s} + \delta_ {m} + \lambda_ {t} + \varepsilon_ {s m t}\tag{3}\]

where is the dependent variable for state (country or region) s, month m and year t; refers to the unemployment rate, is a vector of other explanatory variables and is an error term. The terms and reáect monthly and annual shocks common to all states and controls those factors that are constant across time, but di§erent among states. In this setting, DuMouchel et al. (1987) evaluated the relationship between the age at the onset of alcohol consumption and mortal accidents, and Sa§er and Chaloupka (1989) studied the impact of breath alcohol detection tests with respect to mortal tra¢c accidentes. On the other hand, Ruhm (1995, 2000) studied the relationship between consumption of alcoholic drinks and the main death causes. It is true that a speciÖcation like (3) allows to mitigate some of the previous problems, but some authors (Freedom, 1999, for instance) have pointed out econometric problems such as unit roots or omission of relevant variables (i.e., personal attitudes towards alcohol, legislation, advertising or dynamics in consumption).

3 Data

The main dataset is the BRFSS for the 1987-2003 period in which each wave constitutes an independent cross-section. This survey is a joint project of the Center for Disease and Control Prevention (CDC) and the US states and territories. The survey is a program designed by the CDCís Behavioral Surveillance Branch (BSB) to measure the behavioral risks of the population 18+ living in family households.

The BRFSS is a phone survey designed to give state uniform and speciÖc information of the prevalence of health habits, including alcohol consumption². Uniform data collection procedures ensures the comparability of the data from one point in time to another, as well as over a given period of time, across selected populations and geographic areas. The results are used by public head o¢cials to determine the problematic areas in their states, to develop prevention policies and intervention strategies, and to evaluate success in reducing the prevalence of behaviors that a§ect public health³.

In the Örst survey (1984) information for only 15 states is available. However, since 1995 all the states and Columbia district have been participating continuously. The questions referring to alcohol consumption are located in the main module and are made to all individuals in the sample, except for the 1994, 1996, 1998 and 2000 that they were located in the optional module. The sample size for the period 1987-2003 is 1730792 observations⁴.

3.1 Description of the variables

The survey reports several questions on alcohol consumption. First, respondents are asked whether they consumed at least one drink of any alcohol beverage (a can/bottle of beer, a glass of wine, one cocktail, a shot of liquor) in the last month⁵. Those answering a¢rmatively are questioned about the number of drinks, the number of days of the week with positive consumption, the number of times they have consumed more than Öve drinks and whether they drove under the e§ects of alcohol.

² Researchers who have approached the issue about the validity of self-reports of alcohol consumption, have concentrated their e§orts in the direction of under-reporting, and have tended to discount the possibility of over-reporting behaviors by attributing false positives to measurement errors (Midanik, 1989)

³ More information about the survey can be found at http://www.cdc.gov/nccdphp/brfss.

⁴ For period 1987-1999, we have 1032985 observations. We have excluded observations for Guam, Puerto Rico and Virgin Islands.

⁵ The survey does not distinguish among types of drinks (except for 1987, 1988 and 2003), so it is not possible to introduce any weighting that refers to their di§erent ethylic content.

We use in this study Öve di§erent proxies of alcohol consumption, in addition to the indicator:

Drinker: binary variable which takes the value one for respondents with some consumption during the last 30 days.

Conditional consumption: number of drinks for drinkers in 30 days (in logs).

Conditional mean consumption: number of drinks to drinkers ratio (in logs). We calculate the drinkers ratio as the proportion of drinkers in each state in each time period.

Chronic consumption: binary indicator which takes one for male (female) having more than 60 (30) drinks during the last month⁶.

Binge drinking: binary indicator that takes one if the respondent has imbibed Öve or more beverages on a single occasion.

Drive drinking: binary indicator that takes one if respondent has driven under the e§ects of alcohol⁷.

All these measures have been frequently used in the literature. For example, Manning et al. (1995) used the Örst two; Dee (1999) tried to capture the implications of alcohol abuse and used two measures very similar to the fourth and Öfth. Finally, Ruhm and Black (2002) and Dee (2001) used all indicators but the third. In our opinion, the third measure (conditional mean consumption) could have a lot of potential as an indicator as long as the probability of being a drinker and the consumption of drinks were a§ected by di§erent determinants (or the same determinants with di§erent e§ects).

We also use control of the socioeconomic characteristics of the respondent: race (white, black, hispanic), marital status (married, divorced, separated, widowed, single) and level of schooling (high school dropouts, some college, college). In addition to these variables we use the unemploymen rate (Bureau of Labor Statistics), real per capita income (Bureau of Economic Analysis) and alcohol state taxes⁸.

Some individuals do not provide information about age, race, level of studies or marital status. We deÖne missing-value dummies in order to keep the observations. This concerns 0.75% of the sample⁹. To avoid the ináuence of outliers we have established a maximum of 450 drinks consumed in the last month (an average of 15 per day). This upper limit a§ects 0.018% of the sample and information about drinking participation is unavailable for 0.21% of the sample.

⁶ The literature suggests that a moderate consumption of alcohol may have beneÖcial e§ects on health. Nevertheless, di§erences exist in the consumption depending on the sex (Baum-Baker, 1995). Women have lower probability of being alcoholic, it is more probable that they are abstemious and, on average, they consume less alcoholic drinks than men (Mullahy and Sindelar, 1991; Wilsnack and Wilsnack, 1992; Caetano, 1994; Wilsnack et al., 1994). There is also evidence that women answer in a di§erent way to alcohol consumption. With the same consumption, women experience more serious hepatic damage than men. Federal recommendations advise women not to consume more than one alcoholic drink a day, and for men not to consume more than two (US Department of Health and Human Services, 2000).

⁷ There is only information about this indicator until 2000. So we will include it in estimations for period 1987-1999, but not for period 1987-2003.

⁸ There are three types of taxes: beer, wine and spirits. Provided that Ruhm and Black (2002) use the state taxes on the beer we have done the same. http://www.taxfoundation.org

⁹ We do not present both sets of results, though they do not substantially di§er.

Table B.1 contains descriptive statistics. For the period 1987ñ2003, 49% of the sample has consumed at least one alcoholic drink in the last month. The average number of drinks consumed by the drinkers is 34.12. Nevertheless, 51.3 % has consumed less than 10 drinks, 76.5% less than 25 and 4.9% more than 80. Besides that, 16% declares at least 5 drinks in the same occasion and 5% has consumed more than 60 (30) drinks if he is a man (woman) in the last month. Finally, weights indicate that men, hispanics or other ethnic minorities and young people are underrepresented in the survey.

3.2 Alcoholic drinks and unemployment: a Örst look

In Table B.2 we compare our descriptive statistics with those of Dee (2001) and Ruhm and Black (2002). We can observe that unemployment, age, sex, composition of the population by race, percentage of drinkers and consumption are very alike for the three samples. The main di§erences appear in high school and divorced, which are higher for Dee (2001) and us than for Ruhm and Black (2002). We also present in Figure 1 of Appendix A, the standard deviation with respect to the mean for the unemployment rate and all alcohol consumption indicators. The pattern of the relationship between unemployment and alcohol consumption indicators follow a procyclical pattern for most of the Ögures.

Although when considering that the only source of heterogeneity in the decisions of becoming drinker and the number of drinks consumed is sex we get a clear procyclical proÖle, the situation completely changes when there are another sources of heterogeneity. We have done Ögures for men and women grouped by age cohorts in ten year intervals from 21 to 50 years, and a last one for those aged 50 to 65. The relationship between mean consumption and the rate of unemployment is not as clear as before¹⁰. According to Figures 2 and 3, it seems that average consumption is procyclical only for men and women from 21 to 30. Although we cannot establish any causal relationship based on these correlations, economic conditions could have some e§ect on consumption at the intensive margin for some group of the population.

4 Empirical results

4.1 Estimates using individual data

As a Örst step to cover one of the objectives of this research, we present in Table B.3 the same set of regressions than those provided by Dee (2001) and Ruhm and Black (2002). For each of the six alcohol consumption indicators we have estimated equation (1) by weighted LS, using BRFSS Önal weights. We have done two kind of regressions with and without state e§ects and including in both controls for month, year, age and its square, gender, race/ethnicity, level of schooling, marital status and real per capita income. When state e§ects are not included, the unemployment rate is signiÖcant in all equations except in the binge drinking one. In fact, all alcohol consumption indicators are procyclical. Alternatively, when state Öxed e§ects are introduced unemployment rate is non-signiÖcant for two cases: the percentage of drinkers and the percentage of binge drinkers. Real per capita income is positive and signiÖcant except for the participation decision.

Table B.4 presents a comparison of our estimates with those obtained by Dee (2001) and Ruhm and Black (2002). For the period 1987-1999 we have replicated the same results than Ruhm and Black (2002), and those obtained for period 1987-2003 are very similar. However, Dee (2001) Önds that real per capita income is non-signiÖcant for consumption, binge drinking and chronic drinking and the unemployment rate is positive and signiÖcant for binge drinking. On the other hand, our is greater than the one reported in Dee (2001). The reasons for this disparity among results may be that Dee (2001) neither introduces Önal weights in his estimations nor includes any measure for alcohol prices and considers Örst waves of the survey, which are less reliable due to the small number of states interviewed.

^{1 0} It also happens for the rest of alcohol consumption indicators. We omit these graphs for reason of space, but they, as well as additional graphs for men and women at di§erent age brackets, are available upon request.

Summing up the results on pooled cross-section data, all exercises Önd procyclical e§ects of unemployment at least on consumption and binge drinking. This result as well as the income coe¢cient and a number of coe¢cients corresponding to sociodemographic variables imply di§erent determinants (or e§ects) on the probability of drinking and on the consumption equations. As a consequence, a problem of endogenous sample selection could arise when modelling separately both decisions. Moreover, we show that the e§ects on the estimates of using di§erent sample periods are very small, at least during the time spans 1987-1999 and 1987-2003.

4.2 Results using cohort data

4.2.1 DeÖnition of cohorts

Once we have covered our Örst aim of comparing the results with previous ones in the literature, we move on to estimates using cohort data. We deÖne three types of cohorts: by date of birth, by date of birth and gender, and by date of birth, gender and educational level (some studies using these methodology are among others Attanasio and Weber, 1993 or Blundell et al., 1994). In the Örst case, we deÖne 10 groups. From age 21, we take Öve years brackets to deÖne a group, until age 64. The last group includes those aged 65 and more. This procedure generates a sample with 2040 observations. In the second case, we group ten year age brackets and sex. The resulting sample is formed by 4080 observations. In the case of age-gender-education cohorts, we employ ten years age brackets (23-30, 31-40, 41-50, 51-60, 61+) and two education groups (some college+ and the rest). As a result we also dispose of 4080 observations. Since the sample size in the 1987-2003 period is 1730792, we have, in the case of age cohorts, an average of 848 observations per cell, while in the two other cases, we have on average 424 observations. Given this sample size, we can neglect the errors in variables problem.

4.2.2 Homogeneous results

Before estimating the model, we test for the exogeneity of the unemployment rate and real per capita income. The unemployment rate could be endogenous because although we observe that unemployed individuals consume more alcoholic beverages, we do not know a priori the direction of the causality (Ettner, 1997). Income is potentially endogenous Örst because alcohol consumption is a component of total income, and second due to the good ináuence over e¢ciency at work that moderate consumption of alcohol may produce. It c

onsequently could a§ect earnings (Hamilton and Hamilton, 1997; French and Zarkin, 1995). As instruments for the unemployment rate and real per capita income we propose the corresponding to the same month of the previous year. Since consumption exhibits seasonality the correlation between the regressor and the instrument is high, while we do not think it exhibits correlation with the error term. We use Hausman tests for comparing LS and IV estimates. For all the cohorts and the six indicators it is not possible to reject the null of absence of systematic di§erences in the coe¢cients. Then, it seems both variables are exogenous under the identifying assumption of exogeneity of the other variables in the regression.

In Tables B.5 and B.6 we present LS estimates for age and age-sex cohorts. All regressions include month, year and state Öxed e§ects, age and its square, gender, race/ethnicity, the level of schooling, marital status and real per capita income. We present two di§erent sets of results with and without cohort e§ects. The results show some common traits for all beverage consumption indicators. When cohort e§ects are omitted the unemployment rate appears to be signiÖcantly negative, except in the binge drinking equation. The magnitude of the coe¢cient is very similar to that Önd in the pooled cross-section sample as it is the magnitude of the coe¢cient of income. When cohort e§ects are introduced the unemployment rate becomes non-signiÖcant. Real per capita income is always signiÖcant (even for the participation decision), regardless the inclusion of cohort e§ects. Cohort dummies are highly signiÖcant in all regressions. The magnitude of the coe¢cient of income changes signiÖcantly in models with Öxed e§ects. Income seems to be negatively correlated with the preference for drinking, except for chronic drinkers, as expected. A seemingly unexpected result is that income is positively correlated with unobserved heterogeneity in the consumption equation. However, notice that income is a signiÖcant determinant of participation and the unobserved e§ects are capturing participation in this equation. Since income positively ináuences participation, it should not constitute a surprise that positive correlation arises.

It would be possible to argue that cohort dummies and the rate of unemployment show a high level of collinearity, but we test this is not the reason for loosing signiÖcance. We just run a regression of the unemployment rate on cohort e§ects in the sample of cohorts by age and we obtain an of 0.33. On the other hand, we might also think that when including month, year, state and cohort Öxed e§ects the variation of the unemployment rate is not su¢cient to properly identify its e§ects separately from other micro and macroeconomic determinants. In order to check it we have re-estimated all the models excluding individually each of the subsets of monthly, annual and geographical dummy variables. The result are conclusive: we get negative and signiÖcant e§ects of unemployment on the demand for alcoholic drinks only when cohort e§ects are excluded from the speciÖcations, independently of other set of dummies being excluded or not. These results conÖrm our hypothesis that unobserved e§ects seem to be important determinants of alcohol consumption. We have re-estimated the models based on cohort data excluding income. We observe that unemployment rate is signiÖcant without cohort Öxed e§ects, but is not when we include them¹¹. Although the magnitude of the coe¢cient experiments small variations (ranging from 1 to 10 per cent), it seems to be su¢cient to loose its signiÖcance. These changes could be related to negative correlation among unobserved e§ects capturing preference for working, for instance, and the unemployment rate.

Finally, Table B.7 presents estimates for the sample of cohorts formed using age, sex and education. Since we doubt about the exogeneity of education for building cohorts, we have checked that its distribution remains almost unchanged during the sample period. Thus, we rule out the possibility of taking simultaneous decisions. The results for the unemployment rate and real per capita income remain unchanged. All this evidence implies that unemployment could gather factors di§erent from the relation with the economic activity in models estimated from individual data without controlling for unobserved heterogeneity among individuals. An example could be preferences for working which change with age or tastes for alcohol of di§erent quality potentially correlated with income and education.

4.2.3 Heterogeneous results

In the previous section we have assumed that and , where is the unemployment rate coe¢cient, is unobserved heterogeneity and c is and index referred to each group. Next, we have relaxed the equality of average consumption levels (speciÖc constants), that is, we have estimated the model imposing c; s but we have allowed and have deÖned cohorts by age , age and sex and age, sex and education level Nevertheless, alcohol consumption may be a§ected on a di§erent way by the unemployment rate for di§erent individuals (see Figures 2 and 3). Young people in an extended family going outside may decide to drink independently of their labor status. Individuals at older ages su§ering an unemployment spell could face less uncertainty concerning their expected future income áows because this spell could be used as a path to retirement. Thus, the next step consists in analyzing the impact of relaxing the assumption that the e§ect of the unemployment rate on alcohol consumption is equal across cohorts. So, we will assume and and analyze whether the relationship is cyclical, procyclical or inexistent at di§erent ages.

^{1 1} All these results are available upon request.

In Table B.8 we show results using cohorts by age¹². We observe that men have a higher probability of being drinkers and having chronic consumption for every age group. Men also present higher consumption and average consumption until they are 60 years old. This may be explained because there is a high correlation between alcohol consumption and activity. Finally, men from 21 to 45 years old are more prone to declare binge drinking than the rest. It is probably related to better health status.

Unemployment is neither signiÖcant for the participation decision nor for chronic consumption. On the other hand, for the indicators of consumption, mean consumption and binge drinking, it is signiÖcant with negative sign for men being 21 to 35 years old. When looking at cohorts based on age and sex, average consumption has a procyclical shape for young people. A plausible explanation is that in periods of economic expansion, when unemployment decreases, young men have higher probability of Önding new jobs. Then, the income e§ects prevails. This result must not be interpreted as contradiction but as conÖrmation of previous evidence. When we take into account unobserved heterogeneity among consumers, we observe that although the relationship between consumption of alcoholic drinks and unemployment is not procyclical for the whole population, there are reasons for both cyclical and procyclical shapes at di§erent ages. However, we must be aware that these reasons mainly arise from unobservables such as preferences for working or di§erent consumption tastes.

4.2.4 Impact of other variables

In the Table B.9 we show the impact of other socioeconomic variables on the probability of being drinker and the consumption of alcoholic drinks. We have estimated a model on cohorts formed by age and sex, including as explanatory variables month, year, state and cohort Öxed e§ects, unemployment rate, real per capita income, age and its square, race/ethnicity, marital status (married/cohabiting, divorced/separated, single), level of schooling (high school, some college, college) and beer state tax rate.

Unemployment has no e§ect on consumption in any of the regressions whereas income is always signiÖcant. A 5 % increase in real per capita income increases 8.92% the number of drinks consumed, 8.77% mean consumption, 4.14% binge drinking and 0.54% chronic drinking, with respect to its mean value. This implies that very few individuals start drinking as a result of an increase in their income levels, and on the other hand, that for the case of chronic consumption the addictive component exerts an important ináuence.

An increase in beer state tax rate has a negative and signiÖcant impact on consumption and average consumption. In particular, a 5% increase in the beer state tax rate reduces 2.75% consumption and 3.50% average consumption, with respect to their mean values. The price elasticity is thus very small and it points towards habits as one of the determinants of alcohol consumption. Since we are not able to distinguish between consumption dynamics or unobserved heterogeneity, we are attributing to the later all the e§ects.

^{1 2} We have 2040 observations, that is 204 observations for each cohort. We keep enough degrees of freedom even including several sets of dummy variables (16 for year, 11 for month and 50 for state) as well as other explanatory variables.

White men have a higher probability of being drinkers and consuming more. However, estimated coe¢cients for black or hispanic individuals are negative and signiÖcant. In fact, race is one of the few signiÖcant variables when explaining chronic drinking. This can be linked to certain religious beliefs or cultural behaviors. Race is also a good proxy for the level of household income, since the di§erence on average income between white people and people belonging to other races is more than 4000 dollars.

Education may exhibit a double ináuence on consumption of alcoholic drinks. On one hand, it acts as a proxy of income. On the other hand, it may reáect a higher degree of awareness of the harmful e§ects of alcohol consumption on health. We observe that individuals with university degree (some college, college) have a higher probability of becoming drinkers and in case they are, their consumption is higher. So it seems that the income e§ect dominates the substitution one due to health reasons.

Concerning marital status, being separated/divorced or single has a positive signiÖcant ináuence on all indicators including binge drinking and chronic consumption. On the other hand, married/cohabiting couples do not show signiÖcant di§erences neither in the participation nor in the consumption decision. It may be explained because having familiar responsibilities limits the possibilities of going out freely and stimulates the educating behavior (57.89% of the households have children living at home with their parents).

Finally, we can compare the coe¢cients of demographic variables corresponding to individual and cohort data. We do not observe great di§erences neither in their magnitude nor in their signiÖcance. In our opinion, these results conÖrm that as soon as unobservable heterogeneity is controlled for, unemployment becomes non-signiÖcant in explaining alcohol consumption while the rest of determinants maintain both their impact and signiÖcance.

5 Interpretation of the results

Our hypotheses about the relationship among alcohol consumption and the business cycle are: i) unobserved factors drive decisions about both being drinker or consuming more than the cycle do; ii) the presence of habits in the consumption of alcoholic drinks could generate spurious correlations which are not properly detected in cross-sectional studies. We are also aware that only the availability of panel or pseudo-panel data can help in distinguishing among true and spurious state dependence and also among true state dependence and unobserved heterogeneity. In this section we want to o§er econometric and economic interpretations of the results obtained.

In estimations with independent cross-sections we face a problem of omission of relevant vari ables. The least squares estimate of in (1) is biased and the magnitude of the bias depends on the covariance between UR and and on the variance of UR. As an intuitive more than a formal check we can calculate the covariance between unemployment rate and each of the cohort Öxed e§ects, or obtain the percentage of the di§erence between unemployment rate coe¢cients with and without cohort Öxed e§ects. The percentages of the di§erences explained by each cohort, from the youngest (21-25 years) to the oldest (61-65 years) are, 13.16, 13.89, 11.90, 11.87, 11.54, 10.89, 9.56, 8.75, 8.44, respectively. We observe that all covariances are negative and cohorts corresponding to age intervals 21-25, 26-30 and 31-35 explain a higher proportion of the discrepancy in the unemployment coe¢cient.

If the covariance is zero, is an unbiased estimator of independently of considering the presence of . However, even when UR and are orthogonal, the omission of causes biased estimates of the standard errors of . One possible explanation for some results obtaining signiÖcant correlations between the unemployment rate and the consumption of alcohol is that the results are based upon biased coe¢cients (if unobserved heterogeneity is correlated with unemployment rate) and/or biased standard errors.

A second problem is that, under certain circumstances, it is necessary to consider the decisions of participation and consumption together. If zero demands were due to abstention and if participation and consumption decisions were a§ected equally by the same determinants, the problem would not be so serious. Moreover, if we tried to perform valid inferences at population level (i.e., aggregate demand, evaluation of costs derived from the consumption of alcohol), the fact of having positive and zero observations raises a problem since both groups of consumers belong to di§erent demand regimes and have heterogeneous preferences (Fry and Pashardes, 1994). In fact, preference heterogeneity among those with positive and zero consumption may be problematic due to changes in the income proÖle of drinkers. Changes in the percentage of drinkers for period 1987-2003 can introduce important estimation problems, in terms of instability of parameters (see graphs in Appendix A).

Zero consumption may be due to a problem of abstention, infrequency of purchase or corner solutions. Interview respondents are asked about alcohol consumption in the last month. Given that the observation period is wide enough, and given that individuals may choose whatever alcoholic beverage is attainable to his budget, we can ignore zeros arising from infrequency and corner solutions, possibly with the caution that corner solutions may a§ect individuals in the lowest centiles of the income distribution.

We are going to rewrite the equations for participation , per capita consumption of alcoholic drinks and mean consumption including as explanatory variables the unemployment rate (UR), other socioeconomic or demographic determinants including income and taxes and an error term for each equation, respectively . We also compute elasticities with respec to the unemployment rate

\[\begin{array}{r c l} P & = & \alpha_ {0} + \alpha_ {1} U R + \alpha_ {2} X + u \Rightarrow \varepsilon_ {U R} (P) = \alpha_ {1} \frac {U R}{P} \\ \log (C _ {c}) & = & \beta_ {0} + \beta_ {1} U R + \beta_ {2} X + v \Rightarrow \varepsilon_ {U R} (C _ {C}) = \beta_ {1} U R \\ \log (C _ {M}) & = & \gamma_ {0} + \gamma_ {1} U R + \gamma_ {2} X + w \Rightarrow \varepsilon_ {U R} (C _ {M}) = \gamma_ {1} U R \end{array}\]

Per capita consumption can be expressed as the product of participation rate and average consumption, where D is the number of drinks consumed, N is the population size and is the number of drinkers (Jones, 1989).

\[C _ {c} = \frac {D}{N} = \frac {N _ {D}}{N} * \frac {D}{N _ {D}} = P * C _ {M}\tag{4}\]

Equation (4) implies a relationship among elasticities:

\[\varepsilon_ {U R} (C _ {C}) = \varepsilon_ {U R} (P) + \varepsilon_ {U R} (C _ {M})\tag{5}\]

If we measure alcohol consumption using per capita consumption we will be underestimating the real value of the consumption of drinkers. Thus pooling positive and zero observations is going to a§ect parameter estimates and elasticities. The estimates of the per capita consumption equation are going to be larger (in absolute value) than those of the mean consumption equation, whenever the variables a§ect both decisions. This is true except in the case of changes of regime, when participation and consumption are a§ected by di§erent variables, or even if they are the same ones but their ináuence over both decisions is di§erent. In these three cases, we will have to consider separately and :We can compare these results using the coe¢cients of equations previously estimated as shown in Table 1. We provide participation and consumption elasticities with respect to unemployment (standard deviations between brackets) computed for the three types of cohorts. We have used estimates from Tables B.6 to B.8 including cohort Öxed e§ects and per capita real income¹³.

Table 1. Participation and consumption elasticity with respect to unemployment rate

For period 1987-2003, the three elasticities are negative but non-signiÖcant, although that corresponding to per capita consumption is in the border of signiÖcance for the three types of cohorts¹⁴. An important implication from these results is that we can Önd signiÖcant e§ects of the unemployment rate on alcohol consumption when pooling positive and zero observations due to, for instance, changes of regime which exaggerates the estimates.

In Table 2 we show participation and consumption elasticities with respect to real per capita income. The coe¢cient and t-Student of the income variable in the per capita consumption regressions are 0.05460 (2.83) for age cohorts, 0.04987 (3.27) for cohorts by age and sex, and 0.05056 (3.10) for cohorts by age, sex and education.

Table 2. Participation and consumption elasticity with respect real per capita income (thousands $)

If income increases 1%, the percentage of drinkers increases 0.12% and mean consumption 1.4%. If we look at per capita consumption instead of mean consumption, we would say that consumption increases 1.1%. Ruhm and Black (2002) conclude that almost all procyclical variation is due to consumption of existing drinkers with very little variations in the participation rate. But in absence of changes of regime, the indicator that condenses all information about participation and consumption is mean consumption, which only requires as hypothesis that the determinants of consumption corresponding to both equations are the same. We can provide evidence about this kind of results with an example. Let us assume a situation without starters and quitters. In a per capita consumption equation we are attributing to the unemployment rate the same e§ect for drinkers and non-drinkers. A change in the unemployment rate implies some e§ects on the quantity consumed but even if this is true, it only a§ects actually drinkers but it does not a§ect non-participants. So, results obtained pooling observations from the two regimes are capturing inadequately the whole e§ect. Moreover, if we assume that a variable has only e§ects at the intensive margin (or that the e§ects at both the intensive and extensive margins are the same), and this assumption is not true, we get inadequate results in an equation like per capita consumption combining participatans and non-participants. This is the case for the income e§ects for two reasons: i) income is important for the decision to start or quit drinking; ii) the magnitude of the e§ects at both decisions is not the same.

²1 ^{= 1:29 (p} value ^{= 0:2561)}

"TP(C_C) = "TP(P) + "TP(C_M)

(p value = 0:3349)

(p − value = 0.2965)

^{1 3} We have performed an OLS regression for per capita consumption. The coe¢cient and t-Student for the unemployment rate is -0.0433 (-1.20) for cohorts by age,-0.0455 (-0.92) for cohorts by age and sex, and -0.0459 (-1.14) for cohorts by age, sex and education.

^{1 4} We have performed a test with the following null hypothesis: . The chi-squared statistic is for age cohorts, 1:09 (p value = 0:2965) for age and sex cohorts and 0:93 for age, sex and education cohorts.

Another source of omitted variable bias comes from the exclusion of dynamics in consumption. There are nowadays a lot of papers analysing the existence of habits (myopic or rational) in consumption of several goods as tobacco and alcohol (see Moore and Cook, 1995, Grossman, Chaloupka and Sirtalan, 1998, Bentzen, Ericksson and Smith, 1999 or Baltagi and Gri¢n, 2002). Even in those models which control unobserved heterogeneity, the omission of dynamics will produce biased estimates. We just conduct a simple exercise of estimating the model in a rational addiction framework, whose results are reported in Table B.10. This can only be done when repeated observations for the same individuals are available. We estimate again equations using cohort data for consumption, mean consumption, binge drinking and chronic drinking where we add the lag and lead of the respective measure of consumption. We report results for equations with cohort, month, year and state Öxed e§ects and we also present results without cohort e§ects in order to avoid potential problems of multicollinearity between the e§ects and the unemployment rate. Although we only present results using as instruments for the lag and lead of the dependent variable some previous lags, we have done both OLS and IV regressions with alternative sets of instruments. Our results are robust to di§erent instrument sets once we avoid potential correlation among the lag and lead of the dependent variable and the error term. The evidence reported seems to suggest both that individual heterogeneity and dynamics are important determinants of alcohol consumption. We do not Önd evidence about any e§ects of the cycle on alcohol consumption. This result is robust to any speciÖcation estimated either including or not cohort e§ects, using OLS or IV or using di§erent instruments. We should also emphasize that the e§ects of the rest of determinants of consumption remain una§ected.

6 Conclusions

In this paper we have analyzed the ináuence of macroeconomic conditions captured by unem ployment on the decisions of participation and consumption of alcoholic drinks. We have used cross-section data for the period 1987-2003 from the BRFSS. Opposite to previous studies (Dee, 2001; Ruhm and Black, 2002) that did not controlled for unobservable heterogeneity, we have considered it explicitly. Since genuine panel data is not available to us, we have constructed cohorts combining the cross-sections through time. We have estimated cohort models with Öxed e§ects by LS and IV and the results conÖrm that unemployment is not a signiÖcant determinant of the decisions of becoming drinker and consuming alcohol. It is particularly important to conÖrm the robustness of most of the results to alternative speciÖcations of cross-section, homogeneous and heterogeneous, static and dynamic cohort models.

We have also tried to o§er explanations for the di§erent results obtained with respect to previous studies. Our opinion is that there is some scope for the unemployment to a§ect decisions of becoming drinker and the amount of alcoholic drinks consumed. However, there is also the possi bility for di§erent e§ects, a positive one (income e§ect) and a negative one (stress or health issues)

may be compensated. In addition, we believe that unobservables such as preferences for working or drinking, environmental variables or genetic characteristics of the individuals may explain decisions about participating and consuming alcoholic drinks more than the cycle do.

There are some important for health policies from these results. If alcohol consumption is independent from the business cycle as estimated in this paper, the health expenditure associated to alcohol abuse is not going to be a§ected by the phase of the cycle. Whether the authorities are interested on preserving the e¢ciency of public expenditure, it is necessary to identify di§erent groups of individuals to carry out speciÖc policies, since any attempt to perform universal and homogeneous actions is going to be fruitless.

7 References

1. 1. Attanasio, O. P., Weber, G., 1993. Consumption growth, the interest rate and aggregation. Review of Economic Studies 60, 631-649.
2. 2. Baltagi, B., Gri¢ n, J., 2002. Rational addiction to alcohol: Panel data analysis of liquor consumption. Health Economics 11, 485-491.
3. 3. Baum-Barker, C., 1985. The psychological beneÖts of moderate alcohol consumption: a review of the literature. Drug and Alcohol Dependence 15, 305-322.
4. 4. Bentolila, S., Ichino, A., 2003. Unemployment and consumption: Why are job losses less painful than expected in the Mediterranean?. CEPR Discussion Paper 2539.
5. 5. Bentzen J., Ericksson T., Smith V., 1999. Rational addiction. and alcohol consumption: evidence from the Nordic. countries. Journal of Consumer Policy 22, 257-279
6. 6. Blundell, R., Browning, M., Meghir, C., 1994. Consumer demand and the life-cycle allocation of household expenditures. Review of Economic Studies 61, 57-80.
7. 7. Brenner, M. H., 1975. Trends in alcohol consumption and associated illness. American Journal of Public Health 65, 1279-1292.
8. 8. Brenner, M. H., 1979. Ináuence of social environment on psychopathology: the historic perspective, James E. Barret et al. (eds.) ìStress and mental disorderî. New York, 167-177.
9. 9. Caetano, R., 1994. Drinking and alcohol related problems among minority women. Alcohol Health and Research World 18, 233-242.
10. 10. Catalano, R., Cooley, D., Wilson, G., Hough, R., 1993. Job loss and alcohol abuse: a test using data from the epidemiologic catchment area project. Journal of Health and Social Behaviour 34, 215-225.
11. 11. Crawford, A., Plant, M. A., Kreitman, N., Latcham, R. W., 1987. Unemployment and drinking behaviour: some data from a general population survey of alcohol use. British Journal of Addiction 82, 1007-1016.
12. 12. Deaton, A. 1985. Panel data from time series cross-sections. Journal of Econometrics 30, 109-126.
13. 13. Dee, T. S., 2001. Alcohol abuse an economic conditions: evidence from repeated cross-sections of individual-level data. Health Economics 10, 257-270.
14. 14. DuMouchel, W., Williams, A. F., Zador, P., 1987. Raising the alcohol purchase age: its e§ects on fatal motor vehicle crashes in twenty-six states. Journal of Legal Studies 16, 249-266.
15. 15. Ettner, S. L., 1997. Measuring the human cost of a weak economy: does unemployment lead to alcohol abuse?. Social Science Medicine 44, 251-260.
16. 16. Forcier, M. W., 1988. Unemployment and alcohol abuse: a review. Journal of Occupational Medicine 30, 246-251.
17. 17. Freedom, D. G., 1999. A note on ìEconomic conditions an alcohol problemsî. Journal of Health Economics 18, 661-670.
18. 18. French, M., Zarkin, G., 1995. Is moderate alcohol use related to wages? Evidence from our worksites. Journal of Health Economics 14, 263-400.
19. 19. Fry, V., Pashardes, P., 1994. Abstention and aggregation in consumer demand: zero tobacco expenditures. Oxford Economic Papers 46, 502-518.
20. 20. Gravelle, H. S., Hutchinson, G., Stern, J., 1981. Mortality and unemployment: a critique of Brennerís time series analysis. The Lancet, September 26.
21. 21. Grossman, M., Chaloupka, F. J., Sirtalan, I., 1998. An Empirical Analysis of Alcohol Addiction: Results from the Monitoring the Future Panels. Economic Inquiry, Oxford University Press 36, 39-48.
22. 22. Hamilton, V., Broman, C., Ho§man, W., Renner, D., 1990. Hard times and vulnerable people: initial e§ects of plant closing and autoworkersímental health. Journal of Health and Social Behavior 31, 123-140.
23. 23. Hamilton, V., Hamilton, B. H., 1997. Alcohol and earnings. Does drinking yield a wage premium? Canadian Journal of Economics 30, 135-151.
24. 24. Hammarstrˆm, A., Janlert, U., Theorell, T., 1988. Youth unemployment and ill health: results from a 2-year follow-up study. Social Science Medicine 26, 1025-1033.
25. 25. Hartka, E., Johnstone, B., Leino, E., Motoyoshi, M., Temple, M., Middleton, K., 1991. A meta-analysis of depressive symptomatology and alcohol consumption over time. British Journal of Addiction 86, 1283-1298.
26. 26. Jahoda, M., 1979. The impact of unemployment in the 1930s an the 1970s. Bulletin of the British Psychological Society 32, 568-573.
27. 27. Janlert, U., Asplund, K., Weineball, L., 1991. Unemployment and cardiovascular risk indicators. Scandinavian Journal of Social Medicine 20, March, 14-18.
28. 28. JimÈnez-MartÌn, S., Labeaga, J. M., 1994. Is it possible to reduce tobacco consumption via alcohol taxation?. Health Economics 2, 231-241.
29. 29. Jones, A., 1989. The UK demand for cigarettes 1954-1986, a double-hurdle approach. Journal of Health Economics 8, 133-141.
30. 30. Junankar, P. N., 1991. Unemployment and mortality in England: a preliminary analysis. Oxford Economic Papers 43, 305-320.
31. 31. Kessler, R., Turner, J. B., House, J. S., 1987. Intervening processes in the relationship between unemployment and health. Psychological Medicine 17, 949-961.
32. 32. Manning, W. G., Blumberg, L., Moulton, L. H., 1995. The demand for alcohol: the di§erential response to price. Journal of Health Economics 14, 123-148.
33. 33. Midanik, L., 1989. Perspectives on the validity of self-reported alcohol use. British Journal of Addiction 84, 1419-1423.
34. 34. Michael, J. M., Cook, P. J., 1995. Habit and heterogeneity in the youthful demand for alcohol,îNBER Working Papers 5152, National Bureau of Economic Research.
35. 35. Mullahy, J., Sindelar, J. L., 1991. Gender di§erences in labor market e§ects of alcoholism. American Economic Review (Papers and Proceedings) 81, 161-165.
36. 36. Peirce, R. S., Frone, M. R., Russell, M., Cooper, M. L., 1994. Relationship of Önancial strain and psychosocial resources to alcohol use and abuse: the mediating role of negative a§ect and drinking motives. Journal of Health and Social Behaviour 35, 291-308.
37. 37. Remington, P. L., Smith, M. Y., Williamson, D. F., Anda, R. F., Gentry, E. M., Hogelin, G. C., 1988. Design, characteristics and usefulness of state-based behavioral risk factor surveillance: 1981-1987. Public Health Reports 103, 366-375.
38. 38. Ruhm, D. J., 1995. Economic conditions and alcohol problems. Journal of Health Economics 14, 583-603.
39. 39. Ruhm, D. J., 2000. Are recessions good for your health? Quarterly Journal of Economics 115, 617-650.
40. 40. Ruhm, D. J., Black, W. E., 2002. Does drinking really decrease in bad times? Journal of Health Economics 21, 659-678.
41. 41. Sa§er, H., Chaloupka, F., 1989. Breath testing and highway fatality rates. Applied Economics 21, 901-912.
42. 42. Stern, J., 1983. The relationship between unemployment, morbidity and mortality in Britain. Population Studies 37, 61-74.
43. 43. US Department of Health and Human Services, 2000. E§ect of changes in alcohol taxes and prices in 10th special report to Congress on Alcohol and Health. NIH Publication n^o 00-1583, National Institutes of Health, Bethesda, MD.
44. 44. Wagsta§, A., 1985. Time series analysis of the relationship between unemployment and mortality: a survey of econometric critiques and replications of Brennerís studies. Social Science Medicine 21, 985-996.
45. 45. Wilsnack, R. W., Wilsnack, S. C., 1992. Women, work and alcohol: failures of simple theories. Alcoholism: clinical and experimental research 16, 172-179.
46. 46. Wilsnack, S. C., Wilsnack, R, W., Hiller-Sturmhofel, 1994. How women drink: epidemiology of womenís drinking and problem drinking. Alcohol Health and Research World 18, 173-184.

A Figures

Figure 1. Population over 21 years

F Unemeloytenlrak —Elrue Drirkng

Uremelomenl rak —Chrai: DrirkInal

Uremploytenlrak-Trhenbergdrurk

Figure 2. Age cohorts for men (mean consumption)

Figure 3. Age cohorts for women (mean consumption)

UremplgymenlraleMeanCoreumpon(logs)

lneorloymerlrat-Mean Comunlmoas

B Tables

Table B.1. Descriptive statistics. 1987-2003^a

^a Data are from 1987 to 2003 period of BRFSS. Information of all-items Consumer Price Index used to deáate income comes from Bureau of Economic Analysis. The Örst column of the table shows unweighted means; the second weights the observations using BRFSS Önal sampling weights.

Table B.2. Comparison of descriptive statistics^a

^aDee (2001) doesnít indicate if descriptive statistics are weighted or not. To build this table we have used Ruhm and Black (2002) descriptive statistics and ours without using Önal weights. ^bFor Ruhm and Black and us (2002) per capita real income is measured in 1999$. Dee (2001) doesnít indicate which is the base year, but there is a great disparity among his Ögures, ours and Ruhm and Black (2002). El crecimiento de la renta pc (24.9 para Ruhm y Black y 26.8 para nosotros) se corresponde aproximadamente con la tasa de variaciÛn del Consumer Price Index para el periodo 99-03

^cRuhm and Black (2002) use beer tax per case ($ de 1999) from Federation of Tax Administrators (www.taxadmin.org) Dee (2001) doesnít introduce any price of alcoholic drinks. We have used information from Tax Foundation web page (www.taxfoundation.org) referred to state tax rate per gallon of beer.

Table B.3. Alcohol and economic conditions.1987-1999. Individual 1 WITHOUT STATE FIXED EFFECTS

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted . All regressions include month and year Öxed e§ects and the individual variables representing age, age squared, race/ethnicity, sex, marital status, educational level, beer tax rates and interactions between age, sex and race/ethnicity. Total number of observations is 1032695 and for the sample of drinkers 490653. We have estimated by weighted-OLS using BRFSS Önal weights.

Table B.4. Comparison of estimations with individual data^a

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted :We have included state dummy variables and state speciÖc linear time trends and explanatory variables representing age, age squared, race, sex, educational level, marital status, per capita real income and beer tax rate. Ruhm and Black (2002) include month, year and state Öxed e§ects, individual characteristics and beer tax rate is not reported). Dee (2001) include month, year and state Öxed e§ects and the same explanatory variables. Total sample size is 1032965 for Ruhm and Black (2002), 742821 for Dee (2001) and 1730792 for 1987-2003. The size of the drinkers subsample is 490653 for Ruhm and Black (2002), 359069 for Dee (2001) and 837075 for 1987-2003. Our estimations and Ruhm and Black (2002) use BRFSS Önal weights, but Dee (2001) doesnít.

Table B.5. Cohorts by age. 1987-2003 (N=2040)^a

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted : All regressions include month, year and state Öxed e§ects and the individual variables representing age, age squared, race/ethnicity, sex, marital status, educational level and beer tax rates. (N=12 months x 17 years x 10 age groups = 2040)

Table B.6. Cohorts by age and sex. 1987-2003

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted : All regressions include month, year and state Öxed e§ects and the individual variables representing age, age squared, race/ethnicity, sex, marital status, educational level and beer tax rates. (N=12 months x 17 years x 10 age groups x 2 sex =4080)

Table B.7. Cohorts by age, sex and educational level. 1987-2003

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted : All regressions include month, year and state Öxed e§ects and the individual variables representing age, age squared, race/ethnicity, sex, marital status, educational level and beer tax rates. (N=12 months x 17 years x 5 age groups x 2 sex x 2 educational levels =4080)

Table B.8. Heterogenous estimations with cohort data. 1987-2003

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Regressions are performed over cohorts by age and sex with month, year and state Öxed e§ects and the same variables used in Table B.7.

Table B.9. Impact of social and economic conditions over alcohol consumption. 1987-2003^a

^a(I.D.= Individual data; C.= Cohort data). Heteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted . Regressions with cohort data are performed over cohorts by age and sex with the same variables included in Table B 7. Regressions with individual data include month, vear and state fixed effects and the same explanatory variables than with cohort data. Data are weighted using BRFSS Önal weights.

Table B.10. Rational addiction models estimated on cohort data by

^aHeteroscedastic-consistent standard errors. t-Student statistics reported in parenthesis. Adjusted of the OLS model: Unemployment rate, lag and lead of dependent variable are instrumented using: lag 2 and12 of unemployment rate, and lags 3, 4, 12 and 13 of dependent variable. All regressions include month, year and state Öxed e§ects and the individual variables representing age, age squared, race/ethnicity, sex, marital status, educational level and beer tax rates. N^o of observations: cohorts by age (2040), cohorts by age and sex (4080), cohorts by age, sex and level of education (4080).

DOCUMENTOS DE TRABAJO

References

1. 2006-06: “Further evidence about alcohol consumption and the business cycle”, Sergi Jiménez-Martín , José M. Labeaga, y Cristina Vilaplana Prieto.

References

1. 2006-05: “Las dimensiones del bienestar en las Comunidades Autónomas Españolas. Un análisis de Sigma y Gamma – Convergencia”, Andrés J. Marchante Mera y Bienvenido Ortega Aguaza y José Sánchez Maldonado.

References

1. 2006-04: “Efectos del Programa Operativo Integrado de Canarias, 2000-2006: Un análisis basado en el Modelo Hermin”, Simón Sosvilla-Rivero y Emma García.

References

1. 2006-03: “Understanding and Forecasting Stock Price Changes”, Pedro N. Rodríguez y Simón Sosvilla-Rivero.

References

1. 2006-02: “A Macro and Microeconomic Integrated Approach to Assessing the Effects of Public Policies”, Xavier Labandeira, José M. Labeaga y Miguel Rodríguez.

References

1. 2006-01: “The dynamics of regional inequalities”, Salvador Barrios y Eric Strobl.

References

1. 2005-26: “It's a Small Small Welfare Cost of Fluctuations”, Franck Portier y Luis A. Puch.

References

1. 2005-25: “Obsolescence and Productivity”, Fernando del Rio y Antonio R. Sampayo.

References

1. 2005-24: “EU Structural Funds and Spain’s Objective 1 Regions: An Analysis Based on the Hermin Model”, Simón Sosvilla-Rivero.

References

1. 2005-23: “A sequential model for older workers’ labor transitions after a health shock”, Sergi Jiménez-Martín, José M. Labeaga y Cristina Vilaplana Prieto.

References

1. 2005-22: “Price Convergence in the European Car Market”, Salvador Gil-Pareja y Simón Sosvilla-Rivero.

References

1. 2005-21: “Implicit regimes for the Spanish Peseta/Deutschmark exchange rate”, Francisco Ledesma-Rodríguez, Manuel Navarro-Ibáñez, Jorge Pérez-Rodríguez y Simón Sosvilla-Rivero.

References

1. 2005-20: “A Projection of Spanish Pension System under Demographic Uncertainty”, Namkee Ahn, Javier Alonso-Meseguer y Juan Ramón García.

References

1. 2005-19: “The Dynamic of temporary jobs: Theory and Some Evidence for Spain (The Role of Skill)”, Elena Casquel y Antoni Cunyat.

References

1. 2005-18: “The Welfare Cost of Business Cycles in an Economy with Nonclearing Markets”, Franck Portier y Luis A. Puch

References

1. 2005-17: “Life Satisfaction among Spanish Workers: Importance of Intangible Job Characteristics”, Namkee Ahn.

References

1. 2005-16: “Persistence and ability in the innovation decisions”, José M. Labeaga y Ester Martínez-Ros.

References

1. 2004-15: “Measuring Changes in Health Capital”, Néboa Zozaya, Juan Oliva y Rubén Osuna.

References

1. 2005-14: “Discrete choice models of labour Supply, behavioural microsimulation and the Spanish tax reforms”, José M. Labeaga, Xisco Oliver y Amedeo Spadaro.

References

1. 2005-13: “A Closer Look at the Comparative Statics in Competitive Markets”, J. R. Ruiz-Tamarit y Manuel Sánchez-Moreno.

References

1. 2005-12: “Wellbeing and dependency among European elderly: The role of social integration”, Corinne Mette.

References

1. 2005-11: “Demand for life annuities from married couples with a bequest motive”, Carlos Vidal-Meliá y Ana Lejárraga-García.

References

1. 2005-10: “Air Pollution and the Macroeconomy across European Countries”, Francisco Álvarez, Gustavo A. Marrero y Luis A. Puch.

References

1. 2005-09: “The excess burden associated to characteristics of the goods: application to housing demand”, Amelia Bilbao, Celia Bilbao y José M. Labeaga.

References

1. 2005-08: “La situación laboral de los inmigrantes en España: Un análisis descriptivo”, Ana Carolina Ortega Masagué

References

1. 2005-07: “Demographic Uncertainty and Health Care Expenditure in Spain”, Namkee Ahn, Juan Ramón García y José A. Herce.