Bowling Alone, Drinking Together

Transcription

1 Bowling Alone, Drinking Together Paolo Buonanno University of Bergamo Paolo Vanin University of Bologna February 1, 2012 Forthcoming in Empirical Economics Abstract This paper investigates to what extent youth alcohol consumption depends on consumption patterns by other household members and on social interactions outside the household. Exploiting the richness of the data we explore the possibility of asymmetric social influences by gender and by age, the differences between use and abuse and among different types of alcohol consumption (beer, wine and spirits). Moreover, we control for contextual effects, such as variables related to neighbourhood and family background. We find that both the drinking intensity by other household members and a richer social life outside home are positively related to alcohol consumption. We also find that siblings are more influential than parents, that the mother is more influential for females and the father for males. JEL-Classification: C21, D12, I12, Z13 Key-words: Social relations; Social interaction; Family; Alcohol consumption; Binge drinking; Youth We are indebted with Giacomo Pasini and Marco Piovesan for the title and for initial encouragement. To Giacomo Pasini we are also grateful for useful suggestions and advices throughout the work. We also thank two anonymous referees and conference and seminar participants at the 2007 AAWE meeting in Trier, at the 2008 EEA-ESEM conference in Milan, at the 2009 conference on Happiness and Relational Goods in Venice and at the Universities of Padova and Verona. We gratefully acknowledge financial support from the University of Padua (research grant CPDA071899). All responsibility and errors are ours. Department of Economics, University of Bergamo, Via dei Caniana 2, Bergamo, Italy. paolo.buonanno@unibg.it. Corresponding Author. Department of Economics, University of Bologna, Piazza Scaravilli 2, Bologna, Italy. paolo.vanin@unibo.it 1

2 1 Introduction Alcohol consumption, and especially alcohol abuse, has significant individual and social costs, since it is related to a number of risky or harmful behaviours, ranging from drunken driving to violence, risky sex, lower school and work performance and health diseases. 1 Like use and abuse of other substances, such as cigarettes or drugs, individual alcohol consumption, especially among adolescents and youths, may significantly depend on consumption patterns by other household members and on social interactions outside the household. 2 The objective of this paper is to investigate these effects. Starting from social interactions outside the household, both drinking and alcohol abuse may increase as a reaction to the lack of significant social relations, for instance to compensate for a sense of loneliness and isolation. In this case, we would say that alcohol is a substitute for social relations. Tragically, systematic abuse may in turn further disrupt an individual s social network, creating a loop from which it is hard to escape, and whose cost is split between the individual and his or her social network. At the same time, alcohol consumption is often a key ingredient of many social moments, from meals to parties, to socially enjoyed leisure. Thus the amount of drinking may be increasing in the number of friends and, more generally, in the richness of an individual s social life. Drinking together may make social interaction more pleasant and in this case we would say that alcohol consumption and social relations are complements. 3 Coming to family influence, having a father, a mother or siblings who are heavy drinkers might raise individual alcohol consumption, for instance because of a tendency to drink together at home. Yet a tendency to abuse by other household members might also generate a countervailing reaction, inducing an individual to reduce this risky behaviour. Moreover, parents might have a stronger influence on children of the same sex if they act as role models, but asymmetric influences might also be due to parental differences in time allocation between home and work. Finally, siblings may influence each other in a significant way. We focus on adolescents and young adults living with parents and ask two sets of questions. The first is whether alcohol consumption and social relations are complements or substitutes, and in particular whether the answer changes for different kinds 1 See Dee (1999); DeSimone and Chatterji (2006a,b); Dills and Miron (2003); Duarte and Escario (2006); Grossman and Markowitz (2005); Markowitz et al. (2005); Markowitz (2005). 2 See Clark and Lohéac (2007) and Loureiro et al. (2006) for related investigations on US and British adolescents, respectively, and DeSimone (2007, 2009) for US college students. 3 This paper s title refers to the prominent study by Putnam (2000), according to which the Americans appeared to be bowling alone, before a revival in their social life. 2

3 of alcoholic beverages, for different forms of social relations, for different consumption levels (in particular, use and abuse), and for individuals with different personal and social characteristics. Our main measure of social relations is the frequency of friends meeting; but we also consider going out to shows or to dance, participating in any kind of cultural and recreational associations and church attendance. The second is whether young people patterns of alcohol consumption and alcohol abuse are affected by analogous behaviours by their father, mother and siblings, and how the answer changes between adolescents and young adults, and between males and females. To tackle such questions, we take advantage of a rich dataset containing detailed individual information on young people s alcohol consumption patterns, on their social relations, on consumption patterns by other members of the same household, as well as on various personal characteristics of each household member. While most studies, as discussed below, either control for interaction within the household or for social interaction outside of it, our data allow to simultaneously control for both. Of course, since individuals choose their social life, correlations between alcohol consumption and social interaction outside the household need not reflect a causal impact, but might simply arise due to substitutability or complementarity between alcohol consumption and various forms of socially enjoyed leisure. By contrast, since adolescents cannot choose their household, and since it is more likely that adolescents are influenced by parents than the other way around, sorting is not an issue in establishing family influences and these are more likely to be causal. 4 Yet the identification of causal effects remains difficult. For instance, within household correlations in drinking behaviour might reflect common unobserved traits, such as a genetic predisposition or other unobserved environmental factors. We discuss such issues in more detail in Section 2 and exploit the richness of our dataset to tackle some of them. The main results are the following. First, controlling for a long list of covariates, alcohol consumption is clearly a socially enjoyed good, in the sense that it is positively and significantly correlated to the intensity of social relations. This holds for consumption of wine, beer and spirits, as well as for binge drinking, and it holds for almost any form of social life: friends meeting, going out to shows or to dance, and social and political involvement, all significantly raise alcohol use and abuse. The only exceptions are church attendance, which is negatively and significantly related to all forms of al- 4 Compared with adolescents, young adults obviously have more possibilities to choose their household, yet Manacorda and Moretti (2006) report that more than 80% of Italian men aged live with their parents. We exploit this fact to investigate how the influence by other household members on individual drinking behaviour changes between adolescents and young adults. 3

4 cohol consumption, and going out to shows, which is not significant for binge drinking and barely significant for wine consumption. Second, consumption by other household members is positively and significantly correlated to all forms of alcohol consumption, in all specifications of the model, with siblings being more influential than parents. Third, for water consumption the effect of friends and of many forms of social life is absent, confirming that our results are specific to alcohol and do not just reflect a general thirst effect. Yet the effect of consumption by other household members is still present for water, suggesting a general mechanism of drinking habit formation at the household level. Fourth, the father is more influential for males and the mother for females, except for wine consumption, possibly due to the higher frequency of meals with the mother. Fifth, these results are broadly confirmed when we split the sample between males and females, and between adolescents and young adults, with some indication of children s progressive convergence toward their parents drinking habits. Finally, our findings suggest that unobserved heterogeneity is unlikely to fully explain the correlation between drinking behaviour and social influences. In particular, concerning the role of unobserved heterogeneity, we first argue that common genetic traits are not likely to explain the asymmetry between father and mother s influence over sons and daughters, which seems rather consistent with differences in time allocation between home and work. Genetic traits are also unlikely to explain the changes in such influence between adolescents and young adults, which are rather suggestive of the progressive formation of family drinking habits. Next, unobserved heterogeneity is addressed by not only controlling for a great number of individual variables, including demographics, health, habits, education, labour market status, satisfaction, and family and environmental characteristics, but by also controlling in detail for each parent s characteristics and habits, including determinants of their own drinking habits. Further, we employ an approach that parallels DeSimone (2009) and explore the robustness of our findings to the progressive inclusion of different sets of covariates: we start with a very limited set of exogenous controls (i.e., age and gender) and then progressively add exogenous family characteristics, contextual variables, survey and region fixed effects, parental characteristics and additional individual controls. This exercise suggests that our results for parental influence (as indeed all our results) are unlikely to be driven by unobserved heterogeneity, so, unless reverse causality plays a major role (which does not seem very plausible, especially for adolescents), a causal interpretation (with all the necessary caveats) may be the most appropriate. For siblings influence reverse causation may be a more serious issue. Our data do not allow to resolve it, since, unlike parents, siblings are not univocally 4

5 identified in the dataset. Still the correlation is informative. The remainder of the paper is organised as follows. Section 2 discusses some problems in the estimation of social influence, Section 3 presents the data and our empirical strategy, Section 4 presents results and Section 5 concludes. 2 Estimating social influence Over the last two decades, more and more attention has been devoted to the estimation of social influence and peer effects in risky behaviours, especially by adolescents and youths. Many studies argue that alcohol use or abuse is substantially affected by peer behaviour, even after controlling for individual characteristics and other factors. However, it is difficult to empirically assess whether the correlation between individual and peer behaviour is due to peer effects, to self-selection or to contextual characteristics. Several recent papers, such as Sacerdote (2001) and Kremer and Levy (2008) exploit random housing assignment to determine the peer effects on several outcomes such as GPA or alcohol or substance use among college students. The quasi natural experimental setting allows to adequately deal with identification issue in determining a causal effect of peers on individual behaviours. For instance, Kremer and Levy (2008) find that students assigned to drinking roommates have one quarter-point lower GPA than those assigned to non-drinking roommates. In a similar vein, other studies exploit school-class based dataset to assess the effect of peers on individual behaviours. Under the assumption that students are randomly allocated to classes and cannot sort themselves into school or classes, it is possible to identify peer effects on individual outcome. Gaviria and Raphael (2001) find strong evidence of peer-group effects at the school level for drug use, alcohol consumption and smoking. Similarly, Lundborg (2006) using Swedish school-class based data finds significant and positive peer effects for binge drinking, smoking and illicit-drug use. Some other scholars have exploited changes in law. Cutler and Glaeser (2007) use workplace smoking bans as an instrument in order to identify peer-group effect on smoking. In many cases, quasi natural experimental settings and exogenous variations cannot be exploited and thus scholars need to rely on IV techniques. Koch and McGeary (2005) find that alcohol initiation before age 14 significantly reduces the probability of completing high school between 7% and 22%. Their identification relies on appropriately excluded variables related to alcohol consumption but not to high school completion: log beer tax rate and minimum legal drinking age. Markowitz et al. (2005) investigate the effects of 5

6 alcohol consumption and alcohol policies on youth risky sexual behaviours using the Altonji et al. (2005) approach based on estimation of a bivariate probit model without any identifying restrictions, but with a constrained correlation coefficient. Altonji et al. (2005) argue that selection on the unobservables is the same as selection on the observables. In these studies, scholars have alternatively defined the peer group as the family (Case and Katz, 1991), the neighborhood (Norton et al., 1998) and the school (Gaviria and Raphael, 2001; Lundborg, 2006; Clark and Lohéac, 2007; Kooreman, 2007). Strong social interaction effects are usually found, with important asymmetries in influences within and across gender. A number of papers, notably those by Manski (1993, 2000) describe potential econometric and methodological problems in identifying social interactions. Manski (1993) mainly distinguishes between two types of social interaction effects: i) endogenous effects, whereby individual behaviour is influenced by the prevalent behaviour of others in the group and ii) exogenous effects, whereby individual behaviour is affected by the exogenous characteristics or socioeconomic composition of the reference group. 5 The standard econometric specification of social interaction just includes as regressor the average participation in the considered behaviour by other members of the reference group (excluding the individual). Many of the works cited above are related to the role of social influences in determining risky behaviours and in particular alcohol or smoking habits, mainly by adolescents and youths. In a vein closer to our approach, DeSimone (2007, 2009) estimates the effect of social fraternity and sorority membership on different drinking outcomes and in particular on binge drinking. As we do here, he takes advantage of the richness of the dataset used to control for a large set of proxies for individual preferences in order to account for unobserved heterogeneity. His findings suggest that fraternity and sorority membership are significantly and strongly related to alcohol consumption. 3 Data and empirical strategy Our data originate from the 2002, 2005, 2006 and 2007 waves of the Survey on aspects of daily life run by the Italian national statistical office (ISTAT). The survey collects a wide range of information on alcohol consumption habits, social relations, personal characteristics and family background for a representative sample of Italian individuals. 5 Manski (1993) also notes that individual and reference group s behaviour might co-jointly move because they are both exposed to some unobserved effect, which creates a statistical correlation (what he calls correlated effects), but has nothing to do with social interaction. 6

7 By merging different years, we obtain a large pooled cross-section. To assess peer and family influence on youths, we restrict attention to individuals aged 14 to 30 and living with both parents. 6 We thus have a sample of 22,880 youths, for whom we also have detailed information on each parent s characteristics and habits. 3.1 Dependent variables and regressors of interest In our empirical analysis we exploit the following information contained in the surveys. Our dependent variables are Wine, Beer, Spirits, Water and Binge. The first four reflect ordinal data on consumption of wine, beer, spirits and water, scaled from 1 (no consumption) to 6 (more than one liter per day for wine, beer and water, and more than two shots per day for spirits). 7 Alcohol consumption is higher among males than among females, and among young adults than among adolescents. Water consumption is the same for males and females, as well as for adolescents and young adults. Binge is a dummy for having exceeded with alcohol at least once in the last year (having consumed more than 6 glasses of any alcoholic drinks at a single time). 8 Around 15% of the whole youth sample is involved in binge drinking (about 30% of males and 10% of females; about 10% of adolescents, aged 14-19, and 20% of young adults, aged 20-30). Our regressors of interest are measures of social relations/interactions and of consumption by other household members. Among the former variables there is, first of all, the frequency of friends meeting, which may have an important effect on drinking together. We capture it with a dummy variable for meeting friends more often than once a week (Friends). 9 This variable takes value of one for about half of the whole youth sample (about 65% of males and 40% of females; about 70% of adolescents, aged 14-19, and 40% of young adults, aged 20-30). Second, we consider two dummies, Shows and Dance, capturing going to any kind of shows (cinema, theatre, concerts or sports events) and to dance at least three times a 6 As mentioned above, more than 80% of Italian men aged live with their parents (Manacorda and Moretti, 2006). 7 Intermediate consumption levels are ranked as follows. For wine, beer and water, 2= only seasonally ; 3= only rarely ; 4= 1-2 glasses a day ; 5= between 1/2 lt. and 1 lt. per day. For spirits, 2= exceptionally ; 3= seldom (less than a few shots per week) ; 4= a few shots per week ; 5= 1-2 shots per day. Spirits include super-alcoholic and bitter liquors (Italian amari). 8 Data on Binge are not available in This reduces the sample to 15,994 individuals. 9 If both loneliness and a very rich social life may stimulate alcohol consumption, we might expect a non monotonic, U-shaped relationship between alcohol consumption and the frequency of friends meeting. In a previous version we allowed for this possibility by including a wider set of dummies for different frequencies, but we found no evidence of non monotonicity in the data. Moreover, we did not find any significant difference between those who meet friends on a monthly basis and those who do it less often. 7

8 year. Third, we have two dummies, capturing active participation, through either time or money, to any kind of political parties, unions or professional associations (Olson assoc.), and to any kind of voluntary, cultural or recreational associations (Putnam assoc.). Fourth, we consider a dummy for religious participation (Church), capturing going to church at least once a month. Consumption by other household members may be one of the fundamental drivers of individual alcohol consumption. We go beyond the simple consideration of the effects of average consumption by other group members, which is traditionally studied in the literature on peer effects, and try to be as specific as possible. In particular, we tackle the possibility that drink-specific consumption by different household members has different effects by considering three (drink-specific) dummies, called Father consumer, Mother consumer and Siblings consumer, taking the value of one if the father, the mother or at least a sibling is a usual drinker (or a binge drinker). 10 Usual drinkers are defined as those who drink at least 1-2 glasses per day in the case of wine, beer and water, and at least a few shots per week in the case of spirits. Binge drinkers are those who binge drank at least once in the last year. 3.2 Control variables Besides to social life and to other household members consumption habits, differences in alcohol consumption may obviously be due to a number of individual or contextual variables. To account for this, we include in all our regressions several controls, capturing personal information (age, age squared, sex, health status), education (primary, secondary or college), employment status (employed, unemployed or inactive), a number of individual habits (smoking, practicing sports, reading books and newspapers, talking of politics), individual satisfaction (with respect to last year comparison, economic situation, health, family, friends and leisure), a set of neighbourhood characteristics (crime and urban blight) and regional and survey year fixed effects. To make the coefficients of the dummies for other household members consumption meaningfully interpretable, we also always control for three sets of family characteristics. The first includes general household characteristics, such as the number of household members (Family components), the number of each individual s siblings living in the household (Siblings) and whether the family has moved in the last year (Recent 10 For instance, we investigate whether individual consumption of wine is affected by the presence in the household of a father, a mother or at least a sibling who are usual wine drinkers (and analogously for beer, spirits, water and binge drinking). 8

9 mover). The second includes information on the father s education, employment status, job position (five job variables), health, social life and consumption habits; and the third contains analogous information for the mother. 11 The list and the definition of all variables, together with summary statistics, are presented in Table Model specification As discussed in section 2, the empirical investigation of peer group effects on individual behaviour poses serious identification challenges. In our analysis we cannot rely on a quasi natural experimental setting or on a credible instrumental approach. We thus exploit the richness of our dataset along several dimensions. First, our baseline specification includes social relations variables (Friends, Shows, Dance, Olson assoc., Putnam assoc. and Church), consumption by other household members (Father consumer, Mother consumer and Siblings consumer), together with demographics, health, habits, education, labour market status and satisfaction. Second, many of the existing studies on social interactions and risky behaviours use as a measure for peer behaviour the perceived (by the respondent) rather than the actual peer behaviour (Kawaguchi, 2004; Krauth, 2005, 2007). The use of perceived peer behaviour is problematic since the respondent tends to project his own behaviour on the behaviour of his peers. This may lead to overestimate peer influence. Norton et al. (2003) show that regression estimators are inconsitent when perceived group behaviour is used instead of its correct measure. Our dataset allows us to overcome this issue, since peer behaviour (father, mother and siblings) is directly reported by the peer itself. Third, we try to control for contextual effects, such as variables related to neighbourhood (crime and urban blight), family characteristics (number of family component and of siblings) and for region and year fixed effects. Fourth, we do not limit ourselves to the inclusion as regressors of peers average behaviour, but we separately include father and mother s alcohol consumption and we extensively control for father and mother s characteristics such as education, employment status, job position, age, heath status and social life. By controlling for potential determinants of parents drinking, we clean the interpretation of our coefficients of interest and reduce the risk that our results are driven by unobserved heterogeneity. Fifth, we consider different types of alcohol consumption (wine, beer, spirits and binge drinking) and we use information on the degree of 11 Individual education and employment status, together with parents education and job position, proxy for differences in individual and family income, which are not in the data. 9

10 consumption, while the majority of existing papers only study participation. 12 As a robustness check, we perform regressions for water consumption in order to exclude the possibility of spurious effects. Finally, we split the sample by gender and by age group (14-19 and 20-30), to investigate possible differences between males and females and possible changes from adolescents to young adults. Summing up, for each drinking behaviour we estimate the following equation: y i = α + βs i + γh i + δx i + ɛf i + ζc i + ηp i + ϑz i + u i, (1) where Greek letters denote parameter vectors, y i is individual i s outcome, S i captures i s social interactions outside the household, H i measures whether the considered outcome (e.g., wine drinking in the wine consumption regression) is usual among other household members, X i and F i are vectors of exogenous individual and family characteristics, respectively, C i measures contextual variables, P i captures parental characteristics (separately for the father and the mother), Z i is a vector of additional individual controls and u i is a white noise disturbance. More specifically, S i and H i are the following vectors of dummies: S i =(Friends, Shows, Dance, Olson assoc., Putnam assoc., Church); H i =(Father consumer, Mother consumer, Siblings consumer). X i and F i contain the following information: X i =(Age, Age squared, Sex); 13 F i =(Family components, Siblings number, Recent mover). Contextual variables are C i =(Crime, Urban blight, Region dummy, Year dummy); P i contains, separately for i s father and mother, the following information: P i = (P if, P im ), where P ij =(Age, Health status, Cigarettes, High school, Labour market status, Job position, Friends, Shows, Dance, Church), for j = f, m. 14 Finally, Z i =(Health status, Habits, Satisfaction, Education, Labour market status) We present both ordered probit (probit for binge drinking) and OLS regressions. While ordered probit is better suited for categorial data, OLS are useful to have a synthetic indication of the magnitude of the effect of our variables of interest. Since we have six categories for each consumption variable (two for Binge), presenting all marginal effects from the ordered probit regression, for each category and drink, would be extremely cumbersome. While OLS are clearly problematic with ordinal variables (because they implicitly assume a constant distance between each category and the next), precisely the richness of having six categories helps reducing this concerns. 13 The variable Sex is obviously omitted when running separate regressions for males and females. 14 Variables not in italics are the following sets of regressors: Labour market status=(employed, Unemployed); Job position=(white collar, Executive, Office worker, Blue collar, Other occupation). 15 Habits=(Cigarettes, Sport, Newspapers, Books, Talking of politics); Satisfaction=(Prev. year comparison, Economic satisfaction, Health satisfaction, Family satisfaction, Friends satisfaction); Education=(Primary education, Secondary education, College education); Labour market status is defined as above. 10

11 4 Results: drinking together Let us first focus on the effect of our variables of interest. 4.1 Effect of social life and of consumption by other household members on alcohol consumption In our first specification, we regress Wine, Beer, Spirits and Binge on social life outside the family and on (drink-specific) consumption by other household members, controlling for individual, environmental and parental characteristics. 16 Table 2 shows our first clear evidence of the importance of drinking together for youths. The first four columns present an ordered probit specification, whereas the last four columns present an OLS specification. All forms of alcohol consumption (wine, beer, spirits and even binge drinking) are positively and significantly correlated with the frequency of friends meeting, with going out to shows and to dance, and with participation in any kind of associations (only the coefficient of Shows is not significant for Binge and barely significant for Wine). We thus conclude that alcohol consumption and social life are complements. The only form of social interaction that is negatively and significantly related to alcohol consumption is going to church, possibly because of religion s emphasis on the virtue of moderation. Quantitatively, meeting friends in leisure time more often than once a week (relative to not doing so) is associated to an increase in alcohol consumption of around 6.5% for wine and 10% for beer and spirits, and to an increase in the probability of binge drinking of around 2.5%. 17 Our data allow us to single out the potentially different effects of living with a father, a mother or at least a sibling who is a usual (drink-specific) consumer (relative to living with a father, a mother or all siblings, respectively, who are not usual drinkers). We find a positive and significant correlation between all forms of individual alcohol drinking (again, wine, beer, spirits and binge drinking) and an analogous usual 16 Sign and significance of the effects of both social life and consumption by other household members are robust to several changes in the set of control, as discussed in section 4.3. Even coefficient magnitude (including magnitude comparison between different household members) is not very sensitive to inclusion or exclusion of different controls. 17 Going to shows at least three times a year has a comparable effect for beer and spirits, but not for wine and binging. Going to dance at least three times a year has an effect about twice as large as that of Friends. Active participation in associational life has an effect which is comparable to that of Dance for wine and intermediate between those of Friends and Dance for beer, spirits and binge drinking. Going to church at least once a month has an effect that is comparable in size to that of other associational activities, but reversed in sign. 11

12 behaviour by other household members. Yet, mother s consumption is more influential than father s consumption for wine (and to a smaller degree also for spirits and binge drinking), whereas for beer it is the other way around. This result might reflect the disproportionate frequency of wine drinking during meals in Italy and a higher frequency of meals with the mother than with the father. In all cases, individual drinking (or binge drinking) is more strongly associated to having a sibling who is a usual drinker (or binge drinker) than having either parent displaying analogous behaviour. 18 To investigate whether the effects under analysis are different by gender, we split the sample between males and females. 19 Tables 3 and 4 show that the effects of friends meeting and social interaction outside the household are broadly similar for both genders, with a slight increase in the association of Friends with Spirits and Binge for males relative to females. More interestingly, among males, the father s example is always more important than the mother s one (but for Wine) and the opposite holds (for all drinking behaviours) among females. This points to an important asymmetry in parental influence, with marked gender-specific role models. Both for males and for females, having a sibling who is a usual drinker (or binge drinker) is more strongly associated to individual drinking than having either parent who is a usual drinker (or binge drinker). 20 We then split the sample between adolescents (aged 14 to 19) and young adults (aged 20 to 30), keeping together males and females. 21 (ordered probit yield similar patterns and are not reported). Table 5 presents OLS results Friends meeting and social interaction outside the household have broadly similar effects among adolescents and among young adults, with a stronger complementarity between the frequency of friends meeting and alcohol consumption for the former than for the latter. Conversely, while family influence is the same in the two groups in terms of sign and significance, parental influence is stronger for young adults than for adolescents. While at first sight 18 Having a parent who is a usual drinker (or binge drinker) is associated to a rise in the analogous behaviour between 7.6% and 26.7%. The corresponding figure for siblings ranges between 18.4% and 50.7%. 19 We perform a Chow test to assess whether the regression model is different for males and females and we reject the null hypothesis of coefficients equality for the two groups. In particular, the Chow test assumes the following values: (p<0.0001), (p<0.0001), (p<0.0001) and (p<0.0001) respectively for wine, beer, spirits and binge drinking. 20 The only exception is that for females spirits consumption the mother s example is more important than that of siblings. 21 We perform a Chow test to assess whether the regression model is different for adolescents and young adults and we reject the null hypothesis of coefficients equality for the two groups. In particular, the Chow test assumes the following values: (p<0.0001), (p<0.0001), (p<0.0001) and (p<0.0001) respectively for wine, beer, spirits and binge drinking. 12

13 surprising, this result may be due to the fact that alcohol consumption habits develop slowly over time, so that it takes time for adolescents to follow the path of their parents, and correspondingly the correlation between their own alcohol consumption and the fact of having parents who are usual drinkers raises over time. In one word, we observe convergence towards parental drinking models. Interestingly, this result also extends to siblings, so that we observe the progressive formation of family drinking habits. Our last sample split is both by age and by gender. Table 6 shows results for male and female adolescents. While the general pattern is broadly similar for these two groups, a few differences emerge. First, among male adolescents Friends is only significant for Beer and Binge, but not for Wine and Spirits, suggesting that boys concentrate on beer and binging when drinking together with friends. On the other hand, among female adolescents Friends is significant for all forms of alcohol drinking. Second, the mother s behaviour is not significant for her adolescent sons spirits and binge drinking, whereas it is always significant for her daughters drinking (but for binging). Table 7 shows results for young adults. The fact that the frequency of friends meeting is less significant for alcohol drinking among young adults than among adolescents is confirmed both for males and for females, especially as far as Wine and Binge are concerned. Yet the other variables for social interaction outside the household maintain sign and significance. We would thus simply conclude that wine consumption and binge drinking among young adults are not so much related to the frequency of friends meeting as are beer and spirits consumption, although all forms of alcohol drinking remain broadly complementary to social life. The other distinctive feature among young adults is that the mother s behaviour is not significant for Beer among either males or females. One important concern is related to adolescents participation decision in alcohol consumption. 22 For each drink, we look at whether such decision is affected by our regressors of interest (social interaction outside the household and usual alcohol consumption by other household members). Moreover, we also inquire whether, conditional on participation, adolescents drinking intensity is affected by these variables. Results, presented in tables 8 and 9, confirm that participation in alcohol consumption among adolescents is positively (but for church attendance) and significantly related to social interaction outside the household. For instance, the probability that adolescents make use of alcohol is higher for those who meet friends more often than once a week, 22 We thank a referee for suggesting this analysis. 13

14 with an effect that is more significant for beer and wine than for spirits. It is also confirmed that adolescents participation in alcohol drinking is significantly related to the having parents (and siblings) who are usual drinkers, with the interesting asymmetry that the father s example is more important for beer and spirits, and the mother s example for wine (and for teenage girls participation in spirits drinking). Intensity of alcohol consumption conditional on participation is positively and significantly related to meeting friends for females, but not for males. More broadly, social interaction outside the household is more significantly related to adolescents choice whether or not to participate in alcohol consumption than to their drinking intensity conditional on participation. On the contrary, usual consumption by other household members is highly significant both for adolescents participation and for their drinking intensity conditional on participation. Once again relevant asymmetries by gender and drink emerge in the effect of father and mother s usual consumption on children s drinking intensity conditional on participation. 4.2 Robustness check: water consumption Despite our large set of controls, one might still doubt whether our results are not spurious. For instance, one possibility is that they hold for drinking in general, and not just for alcohol in particular. To control for this possibility, Table 10 shows OLS regressions, with the same specification as for alcohol consumption, but this time for water consumption. The table is organised in three groups of three columns. The first group refers to the full sample, the second only to adolescents and the third only to young adults. In each group, the first column refers to males and females together, the second only to males and the third only to females. The first remarkable result is that in none of these specifications Friends is significant for water consumption. This stands in sharp contrast to our results on alcohol consumption, showing that the complementarity between friends and drinks truly reflects something specific to drinking alcohol together, rather than a general thirst effect (possibly induced by activities carried out with friends). 23 Interestingly, while church attendance is always negatively and significantly related to alcohol drinking, it is positively (but not significantly) related to water consumption. So people who attend church do not drink less in general (quite to the contrary), but they specifically drink less alcohol. Again, this reinforces the idea that religious emphasis on moderation specifically reduces alcohol consumption. Finally, 23 Indeed, going out to shows and to dance and participating in associational activities tend to be positively related to water consumption, although they are only sometimes significant. 14

15 water consumption by other household members is always positively and significantly related to individual water drinking. This suggests that the formation of household drinking habits is something general and not specific to alcohol. Yet the patterns of influence by other household members are different between water and alcohol. In particular, for water the mother is more influential than the father, who is in turn more influential than siblings. This holds for adolescents as well as for young adults, and for females as well as for males. Moreover, the increase in correlation between individual alcohol drinking and analogous parental behaviour, observed when passing from adolescents to young adults, is absent for water consumption. So, beyond the general pattern of household drinking habit formation, both the gender specificity of parental influence and the gradual convergence towards parental behaviour are indeed specific to alcohol and cannot be found for water. 4.3 Robustness check: effect of different sets of controls While we treat the association of individual alcohol drinking with social interaction outside the household and with siblings usual drinking as a pure correlation (we do not have convincing instruments to disentangle the direction of causality), we suggest that association with parents usual drinking might reflect a causal effect. Although plausible, such claim might still be disputed, for at least two reasons: first, because parents habits of alcohol consumption might be influenced by their children s drinking behaviour; and second, because the correlation between children and parents behaviour might reflect some common unobserved characteristics, rather than a causal influence. We do not have good instruments to rule out the first possibility. So for that part, the causal interpretation is admittedly weak and mainly relies on the verbal argument that parents drinking habits are likely to be pre-determined with respect to their children s drinking behaviour. To the extent that this is particularly likely to be true for adolescent children, the fact that the observed correlation is present for that sub-sample may be seen as supportive of the causal interpretation. Yet the richness of our data allows us to do much more to control for the role of unobserved heterogeneity. Indeed, the results presented so far include a long list of controls, at the individual, contextual and parental level, which are likely to absorb much of unobserved heterogeneity. Of course, one can never be sure that there are no relevant omitted variables, but one way of gaining confidence in the presented estimates is to show that results are very robust to the progressive inclusion of different (and long) sets of controls, which are likely to capture relevant dimensions of heterogeneity. 15

16 Tables 11, 12, 13 and 14 present, for each type of alcohol drinking behaviour, eight possible specifications of the model. 24. The first three columns only includes (besides a constant) our regressors of interest, namely social interaction outside the household (S i ) and usual drinking behaviour by other household members (H i ), first separately and then together. These specifications are obviously the most exposed to the omitted variable problem, so the estimated coefficients are likely to reflect various forms of unobserved heterogeneity. Yet it is interesting that most parameter estimates coincide in sign and significance with those of the most complete specification reported in column (8), which reproduces, for the various drinks, the last four columns in Table Column (4) additionally controls for exogenous individual characteristics, namely age (and age squared) and sex (X i ). Even with this parsimonious specification, all coefficients of interest coincide in sign and significance with the complete specification, and are even very close in magnitude. 26 We find that all forms of alcohol consumption significantly increase with age, but at a decreasing rate, and that women drink significantly less alcohol than men. 27 the set of controls. This finding is robust to the different specifications of Column (5) also controls for exogenous family characteristics, namely the number of family components, the number of siblings, and whether the family recently moved (F i ). These additional controls have no effect on sign and significance of the parameters of interest and a very small effect on estimates magnitude. In this specification, youths who live in larger families or have less siblings in the household tend to drink significantly more wine and spirits; they are also significantly more prone to binging, as are recent movers. Yet these family characteristics tend to lose significance as we expand the set of controls. Column (6) adds contextual variables, namely region and year dummies, together with neighbourhood characteristics such as crime, and urban blight (C i ). This reduces the likelihood that the effects of our variables of interest, particularly social interaction outside the household, reflect unobserved heterogeneity along some contextual dimension. Again, changes in magnitude for the parameters of interests are minimal and 24 For the sake of space, we only present OLS estimates for the full sample. 25 The only change in sign is for the coefficient of F riends on W ine, which becomes negative but not significantly different from zero. The only other changes in significance concern the effects of Shows for W ine and Binge and of Motherconsumer for Beer. 26 The only exception is Shows in the W ine regression, which is significant in column (4) but not in column (8). 27 For the sake of space, coefficients on controls are not presented in Tables 11 to 14, but are available from the authors upon request. 16

17 changes in sign and significance are totally absent. Residing in an area characterised by high crime rates is not significantly related to alcohol consumption (but for a small effect on spirits, which is not robust across specifications). Urban blight significantly raises all forms of alcohol consumption, although its coefficient loses some significance when additional controls are introduced. Column (7) adds a long set of controls for a number of parental characteristics and habits (separately for father and mother), including likely determinants of their own alcohol consumption, such as their age, health, education, employment status, job position, the frequency of friends meeting, going to shows and to dance, attending church and smoking (P i ). This long list of parental controls should clean the interpretation of our variables of interest, and in particular of parents usual alcohol consumption, from much of the effects of unobserved heterogeneity. Once again, introducing these additional controls has a negligible effect on our parameters of interest. A bit surprisingly, the general picture is that most of these family characteristics are not individually significant. Finally, column (8) presents the most complete specification by adding a number of individual characteristics, such as health, education, labour market status, job position, habits and satisfaction (Z i ). While this reduces the magnitude of almost all coefficients of interest, it leaves all signs unaffected and it makes only one variable lose significance, namely Shows in the W ine regression. 28 Health significantly raises consumption of wine and beer, but not of spirits and binge drinking. Notice that, especially among adolescents, health has a low variance. College graduates consume more wine, but otherwise education has no significant effect. Relative to non participants to the labour force, unemployed drink significantly less wine and employed significantly more beer and spirits. Individual habits are also significantly correlated with alcohol consumption. In particular, alcohol consumption is positively and significantly correlated with smoking, practicing sports and talking of politics, but not significantly with reading books and newspapers. To minimise the risk that the correlations we are after are driven by unobserved individual characteristics, the set Z i also includes a number of happiness variables. This allows to capture the effect of social relations net of individual satisfaction for relational life, and the effect of consumption by other household members net of family satisfaction. Moreover, satisfaction for other life domains arguably captures much of the unobserved heterogeneity that might induce individuals to drink more or less. 28 The only parameters that are not reduced in magnitude are that of F atherconsumer for W ine and Beer and that of M otherconsumer for Beer and Spirits, whose increase is anyway negligible. 17

18 Two clear results emerge. First, all forms of alcohol consumption are negatively and significantly related to family satisfaction. Second, binge drinking is positively and significantly related to satisfaction for leisure time, but negatively and significantly related to satisfaction for all other life domains (economic condition, health, family and friends). Yet most satisfaction variables are not significant for other forms of alcohol consumption. Taken together, results in columns (4) to (8) in Tables 11 to 14 show three things: first, the estimate of our parameters of interest is extremely robust to different specifications of the set of controls; second, the most complete specification of column (8), which is the baseline for all the results presented in the paper, is likely to yield a conservative estimate of the effects of social interaction outside the household and of family drinking; third, our results do not appear to be driven by unobserved heterogeneity. As mentioned above, this last claim should be taken with all the caveats, because our strategy of progressively including different sets of controls cannot be considered a final proof that nothing important is missing from the model. 5 Conclusions Based on a large sample of Italian individuals in 2002, 2005, 2006 and 2007, we present evidence that alcohol consumption is a socially enjoyed good and that family influence is important for drinking behaviour. More specifically, after controlling for a great number of covariates, as well as for potential unobserved heterogeneity and possible spurious thirst effects, we find that almost any form of social life (from meeting friends to going to shows and to dance, to social and political involvement) significantly raises consumption of any kind of alcoholic beverage (wine, beer and spirits) and even binge drinking. Our results suggest that individual drinking is positively associated to drinking behaviour by other household members. In particular, we run separate regressions by gender (males, females) and by age (adolescents and young adults) to investigate the possible role of asymmetric social influences in the determination of alcohol use or abuse. We find that youths siblings are more influential than parents and that parents display marked gender-specific influence patterns. While siblings estimated influence need not reflect a causal effect (reverse causation might be at work here), we argue that estimated parental influence is more likely to be causal. We also find that the association of individual alcohol drinking with friends meeting is stronger for adolescents than for young adults, whereas the opposite holds for the association with parental 18