Is Changing the Minimum Legal Drinking Age an Effective Policy Tool?

In year 1991 regional governments in Spain started a period of implementation of a law that rose the Minimum Legal Drinking Age from 16 to 18 years old. This process was fully completed in year 2015. To evaluate the effects of this change on consumption of legal drugs and its related morbidity outcomes, we construct a regional panel dataset on alcohol consumption and hospital entry registers and compare variation in several measures of prevalence between the treatment group (16-18 years old individuals) and the control group (20-22 years old individuals). Our findings show important differences by gender. Firstly, our main result regarding overall drinking prevalence show reductions ranging from -11.57% for the subsample including both genders to -14.31% for the subsample of males. Secondly, effects on males are driven mainly by reductions in beer with alcohol consumption (-8.98%). Thirdly, effects on wine and/or cava drinking prevalence range from -12.62% for the subsample including both genders to -9.65% for the subsample of females. No effects regarding overall smoking prevalence are found. Fourthly, we do not find evidence that these reductions in alcohol consumption are translated into hospitalizations related to alcohol overdose. To our knowledge, this is the first paper providing evidence on gender-based differences to policies aimed at reducing alcohol consumption. Our results have important policy implications for countries currently considering changes in the Minimum Legal Drinking Age.


Introduction
Undesired and fatal consequences of the abuse of alcohol consumption have been studied from multiple perspectives, ranging from direct effects on individuals (Carpenter, 2004a;Mann, Smart, & Govoni, 2003;Rosenberg, Ventura, Maurer, Heuser, & Freedman, 1996;Wagenaar & Toomey, 2002) to negative externalities exerted on the society as a whole (Carpenter, 2005(Carpenter, , 2007Markowitz, 2000Markowitz, , 2005. According to the latest figures provided by the Report on Survey on Drugs Use in Secondary Schools in Spain (Observatorio Español de las Drogas y las Adicciones (OEDT). Ministerio de Sanidad y Servicios Sociales e Igualdad, 2016), corresponding to survey years 2014/2015, the average age at first use of alcohol considering weekly consumption, has remained almost invariable since year 1996 at around 15 years old. Moreover, around 48%, 61%, and 74% of youngsters, aged 14, 15, and 16 respectively, declared to have consumed alcohol during the last 30 days in years 2014/2015. There is a growing body of evidence pointing at the limitation of access to alcohol consumption as an effective policy tool for preventing unhealthy habits and fatal consequences (Carpenter, 2004b;Carpenter & Dobkin, 2011;Dee, 1999;Deza, 2015;Yörük & Yörük, 2011. In an effort to reduce the prevalence of alcohol consumption and its undesired outcomes, regional authorities in Spain decided to restrict the access of teenagers to alcohol by increasing the Minimum Legal Drinking Age (hereafter, MLDA) from 16 to 18 years old. Figure 1 shows a chronological description of the implementation of the new MLDA in Spain.  Table C1, in Appendix C provides detailed regional information.
Our empirical study takes advantage of this quasi-natural experiment using a differences in differences (hereafter, DiD) method, with the aim of evaluating and quantifying the prospective effects of changing the MLDA on the consumption of legal drugs (i.e. alcoholic drinks and cigarettes) and their related morbidity outcomes.

Differences-in-Differences
We compare variation in prevalence measures between the treatment group (16-18 years old individuals) and the control group (20-22 years old individuals) before and after policy implementation.
The key identifying assumption in our DiD setting is that the variables reflecting the answers of individuals within the treatment group would have followed parallel trends to those variables reflecting the answers of individuals in the control group, if the MLDA had not changed. Figures 1 -7, provided in Appendix B, show graphical evidence to assess the validity of this assumption.

Analysis
We constructed each regional outcome variable y str as prevalence per treatment status, for each year before and after policy implementation. Our treatment dummy variable d_treatment s takes on value 1 for the treatment group, and value 0 for the control group. Our pre-post policy dummy variable d_policy tr takes on value 1 for the year of implementation and subsequent years, and 0 for all years prior to the year of the legal change. Variable DD str is the interaction between dummy variables d_treatment s and d_policy tr . Our econometric model is the following: Equation 1 includes region fixed-effects (α r ), year fixed-effects (ψ t ), as well as region-specific linear trends (θ rt ), and an error term ( str ). Standard errors were clustered at the regional level and computed using wild-bootstrapping (Bertrand, Duflo, & Mullainathan, 2004). Furthermore, regional size differences are taken into account by using as analytical weights the corresponding population per treatment status, region and year. The coefficients of interest that would quantify the causal effect of this policy reform, provided our parallel trends assumption holds, would be a statistically significant estimate of β 3 . The National Health Survey, (Encuesta Nacional de Salud or ENS), and The Hospital Morbidity Survey (Encuesta de Morbilidad Hospitalaria or EMH) are the two main data sources used in this study. While ENS available waves correspond to years 1991, 1993, 1995, 1997, 2001, 2003, 2004, 2006, and 2007, EMH available waves correspond to each natural year between the 1991-2007 period. In order to use the same available data from both sources, we only used yearly datasets corresponding to ENS available waves. From these foregoing sources, we extracted data for the same thirteen regions that shifted the MLDA between years 1994-2002 (see Figure 2). Data for the four remaining regions that shifted the MLDA in years, 1991, 2011, 2014, 2015, were not included due to a lack of enough pre or post policy survey datasets. Three regional panel datasets were prepared, the first including males and females altogether, the second considering only females, and the third including just males. We only considered individuals aged 16-18 or 20-22. Data regarding regional population were extracted from the Population Statistics Database provided by the National Statistics Institute (Instituto Nacional de Estadística or INE). 2  Table 1 show two statistically significant DD str estimated coefficients of -0.06 and -0.08, both significant at the 5% level, corresponding to causal effects of -11.57% and -14.31% in overall drinking prevalence for the subsample including both genders and the subsample of males, respectively. However, for the case of overall smoking prevalence, Table A1, shows that none of the DD str estimated coefficients in any the three subsamples is statistically significant. Figure 1 in Appendix B provide graphical evidence suggesting that our parallel trends assumption holds.

Drink type Prevalence
Firstly, for the beer with alcohol case, the third column in Table 2, corresponding to the subsample of males, shows a DD str estimated coefficient of -0.07, statistically significant at the 1% level, suggestive of a causal effect of -8.98%. Secondly, for the mixed drinks and/or liquors case, the first column in Table 3 regarding the subsample including both genders shows a DD str estimated coefficient of -0.04, statistically significant at the 10% level, that corresponds to a causal effect of -9.53%, whereas the third column, with regard to the subsample of males, shows a DD str estimated coefficient of -0.08, statistically significant at the 10% level, that implies a causal effects of -16.66%. Thirdly, the first and second columns in Table 4 for the wine and/or cava case, show estimates, at the 5% level, of -0.06 and -0.08 corresponding to an implied effect of -12.62% and -15.16% respectively. Interestingly, these latter effects are identified for the subsample of both genders and the subsample of just females, correspondingly. Figures 3, 4, and 5, in Appendix B, provide graphical evidence supporting the validity of our parallel trends assumption.

Morbidity Outcomes
Tables A2 and A3 in Appendix A show that none of the DD str estimates is statistically significant. Figures 6 and 7 in Appendix B supports our parallel trends assumption.

Discussion
Firstly, our main result regarding overall drinking prevalence show reductions ranging from -11.57% for the subsample including both genders to -14.31% for the subsample of males. Secondly, effects on males are driven mainly by reductions in beer with alcohol consumption (-8.98%) and to a lesser extend to reductions in mixed drinks and/or liquors consumption (-16.66%). Thirdly, effects on wine and/or cava drinking prevalence range from -12.62% for the subsample including both genders to -9.65% for the subsample of females. No effects regarding overall smoking prevalence are found. Fourthly, we do not find evidence that these reductions in alcohol consumption are translated into hospitalizations related to alcohol overdose.
We argue that the mechanism of transmission of this policy is closely related to bench drinking or "botellón" given that the identified effects are observed on popular drink types amongst teenagers. Nonetheless, analysing the degree of effective enforcement in public areas as well as the existing alternative ways youngsters use to have access to alcoholic drinks could help to put these findings in context. These effects can be considered as a lower bound given the usual limitations of surveys of this sort (i.e. underreporting). Finally, there may also be unobserved confounding factors that were not controlled by comparison with the 20-22 cohort.

Conclusions
Our findings provide evidence to argue that shifting the MLDA from 16 to 18 years old caused important reductions in alcohol consumption. To our knowledge we are the first to provide evidence regarding gender-based differences related to policies aimed at reducing alcohol consumption. This results suggest that the inclusion of gender perspectives in the process of policy design can contribute to identify more effective policy levers. Furthermore, a quite interesting exercise would be to assess the findings of this study to those that could be obtained from a more focused set of surveys such as the Survey on Alcohol and other Drugs in Spain (Encuesta sobre alcohol y otras drogas en España, EDADES) 3 . We believe our results have important policy implications for countries currently considering changes in the Minimum Legal Drinking Age. If this reduction had an impact on the prospective consequences of excessive drinking, such as performance on standardized tests, crime rate, or traffic accidents, remains as key topics for future research.

Source of Financial Support
This study was supported by Ramón Areces Foundation. Grant: XV Concurso Nacional. Ciencias Sociales 2016.

Conflict of Interest
The authors have no conflict of interest.             ieb@ub.edu www.ieb.edu