- Email: [email protected]
Alcohol advertising bans and alcohol abuse: Reply
Journal of Health Economics 12 (1993) 229-234. North-Holland
Alcohol advertising bans and alcohol abuse: Reply Henry Saffer* National Bureau of Economic Research, New York, NY, USA Final version received April 1993
1. Introduction Saffer (1991) found that alcohol advertising increased alcohol consumption and abuse. This is an interesting finding since several older studies of alcohol advertising find no effect. These prior studies were important efforts and represented the state of econometric knowledge at the time they were published. Recent research using econometric models of advertising has provided more detailed standards for this type of work. Berndt (1991) reviews severai of these econometric issues. Four of the more important issues in econometric studies of alcohol advertising are: (1) testing non-linear specifications, (2) simultaneity between consumption and advertising, (3) temporal aggregation bias, and (4) the correct measurement of advertising. The older studies of alcohol advertising ignored some or all of these issues. The issue of measurement is particularly important in distinguishing SafFer $991) from prior studies. Most prior alcohoi advertising studies used expenditure data rather than advertising bans as the measure of advertising. However, expenditure data vary over a limited range. The relationship between alcohol use and alcohol advertising is subject to diminishing returns. In the range that is employed by expenditure studies these returns may be quite small or zero. l This results in an observed marginal efYect of Correspondence to: ID. Henry Saffer, National Bureau of Economic Research, 269 Mercer St., 8th Floor. New York,,, NY IOW3, USA. *I would like to thank Frank Chaloupka, Michael Grossman and Theodore Joyce for helpful commenason an earlier version of this reply. ‘An increase in advertising expenditure by a firm may increase its market share but may ulso lead to a corresponding increase in advertising by rivals. Market shares initial B~cvcls but at a higher level of advertising for all firms. This level of nae by rivals crnd can result in higher than in the expenditures in the n
230
H.
So$er,
Alchohnl
advertisirtg ham
artd alcohol abuse:
Reply
advertising that is also quite small or zero. This does not mean that advertising has no effect on consumption. Advertising must be observed over a wi&r range in order to correctly estimate its effects. A data set which inch::: 9:s plac es -+zhich allow advertising and p&es which ban advertising is thy. :‘ pirical equivalent to measuring advertising over a wide range. Over this range the relationship between aggregate consumption and advertising can be observed. A similar study of cigarette advertising bans In 22 OECD countries conducted by Laugesen and IvIeads ( 1991j also finds that advertising bans reduce consumption. 2. Cultural differences The issue of cultural differences or sentiment must be addressed when using an international data set. It is likely that there are unobserved variations in sentiment towards alcohol across countries. This sentiment may be exogenous or endogenous to consumption and could affect both alcohol consumption and the enactment of alcohol advertising bans. In the case of t;xogeno~s sentiment some type of proxy variable is needed. The variable I used is dichotomous and equal to one for those countries where more than 75 percent of the alcohol consumed is in the form of beer and wine. Nine of the 17 countries in the data set fall in this category.2 An interesting extension of my 1991 study might be to find some alternative or additional measures of sentiment such as religion. Young does not provide any new sentiment proxies. Another method I used to account for exogenous sentiment are fixed effects models. These fixed effects models exclude the advertising ban variables due to collinearity problems. The country dummies are grouped together by advertising policy and a mean coefficient and t-value computed. These mean coefhcients show increases in all three measures of alcohol abuse as advertising increases. Young (1993) ignores the collinearity problem and simply uses both the advertising variables and the country dummies in the same regressions. Since there were only three countries that changed policies during the sample period there exists a linear transformation of the advertising dumr- ‘I variables which is almost identical to a linear transformation of the country aummy variables. Young’s advertising ban coer%cients in these regressions are unstable and the t-v,Jues are meaningless due to the collinearity he has created. 21 also experimented with a per capita production oG pure alcohol as an alcohol attitude measure. The results from these regressions show that alcohol advertising bans reduce alcohol abuse and were similar to the results reported in gaffer I,1991). While the regression results are reasonable, a regression with log per capital ccnsumption as dependent variable and per capital production as an independent variable is approximately the sanle as using per capita exports as the dependent variable. This also is a problem in the liver cirrhosis and highway fatality regressions.
As I noted in footnote 12 of my 1991 study, endogeneity problem. Alcohol sentiment or cultural attitudes could, in part. of consumption. The passage of advertising bans could also, function of alcohol sentiment. This suggests that consumption ing bans could be endogenous. Saffer and Grossman
is a potential be a function in part, be a and advertis-
3These problems also occur in the estimation of price effects and have resulted in a ge..eral shift away from estimating separate beer, wine and spirits demand curves. The estimation of all own-effects and cross-effects parameters, in both the case of price and advertising, may be too demanding for the available data.
232
H. Sager, Akhohol adoerrisirig bans and
alcohol abuse: Reply
Young does not provide any explanation of why his results are not symmetric. The third reason for not specifying a set of dissaggregated demand curves is that each demand curve should include an own price variable.’ Young uses only a composite alcohol price variable. There is no way of interpreting the estimated effect of this type of variable in separate regressions of beer, wine and spirits demand. Since the sign of the price variable is uncertain it is hard to judge the validity of the equation specification. A fourth reason for not disaggregating total consumption is that this disaggregation exacerbates the cultural difference problem. There is more variation across countries in beer, wine and spirits consumption due to cultural differences than there is in total consumption. Young includes my sentiment variable which is equal to 1 for countries where more than 75 percent of the alcohol consumed is in the form of beer and wine. In regressions using beer, wine and spirits as dependent variables, this sentiment variable has an almost mechanical relationship with the dependent variable creating biased estimates. Young’s expansion of total consumption into beer, wine and spirits consumption is redundant with total consumption, is diRcult to interpret, makes estimation mere difficult and has no potential of adding any new insight. Several hypotheses about how a ban on one alcoholic beverage affects consumption of another alcoholic beverage are possible. However, Young’s conclusion that advertising bans increase beer, wine and spirits consumption is rather surprising. 4.
Serial correlation
Young states that his Durbin-Watson statistics ‘do not take account of the panel nature of the data’. This is very puzzling since Young states that in testing for higher order serial correlation he is able to account for the panel nature of the data. Apparently, Young computed the Durbin-Watson statistic as if the data were a single time series rather than 17 time series. This cannot be done since the estimation of rho would use last observation of country A followed by the first observation of country B. The value of rho should be computed within each country and then could be averaged over countries. This average value of rho could be used for an overall DurbinWatson statistic. The serial correlation correction Young uses in his regressions use the restrictive assumption that rho is the same for all countries in the data set. It 4Young notes that the expenditure data could LC measured with error which would bias the price variable. However, to the extent that the error in the expenditure data is correlated with the error in the consumption data these errors will cancel and tF : resulting regression will be unbiased.
is unclear how this value of rho was computed. In any case, Young’s rho is close to one for e;?ch equation. This results in transforming the data into approximately first differences. Since only three countries had any change in advertising policies during the sample period, the transformed advertising data approximates two vectors of zeros. The advertising ban coefficients and their t-values resulting from this transformation are meaningless.
5. Mean values of alcohol use by advertising policy
Young’s table 2 is a table of means of the alcohol ;huse measure by advertising ban category and time. This table is mainly descriptive since other factors affecting alcohol abuse are not held constant. Young extends the table to include t-values and f-tests. However, Young chooses a null hypothesis of no difference between the means which implies a two tail test. The null hypothesis should be that one group has a higher mean than the other, which implies a one tail test. Also, Young’s choice of five percent significance level is arbitrary. Significance at 10 percent or even 15 percent is aiso llrzaningful information. The beer, wine and spirits means are not good indicators of alcohol related public health and should be ignored. Using a tvalue of 1.3 to define significance, 6 of the 12 differences are significant. However, even if the b-values resulted in no significant differences it simply points out that other factors affecting demand must be controlled.
6. Summary
Young reexamines the work first presented in Saffer (1991) on the relationship between alcohol advertising bans and alcohol use. While there are several interesting possibilities for extending Saffer (Igal), Young fails to provide any new insights. Most of his paper is devoted to recreating my data and results ? He tries to extend my work with three new specifications. Each of these new specifications is flawed. His errors include an iaappropriate application of fixed effects models, inappropriate dependent variables, and incorrect serial correlation computations. The results of these errors are a series of inconsistent advertising ban coeff’icients.He concludes that these inconsistent results are evidence that advertising bans have no effect of alcohol abuse. It might be better to at ‘least provide a series of consistent regression models before coming to any conclusions.
‘Young states in footnote 6 that he had private correspondence with Michele Hainut of the OECD regarding the quality of the Portuguese data. 1 wrote to her about this statement. She said she had never made the statement that Young attributes to her and had Liticoncerns about any of the OECD data.
234
t4. Sager. fflchohoi adllertising bans and alcohol abuse: Reply
cferences Berndt, E., 1991, The practice of econometrics: classic and contemporary (Addison-Wesley Publishing Co, Reading, MA). Bruun, K. et al., 1975, Alcohol control policies in a public health perspec!ive, The Finnish Foundation for Alcohol Studies 25. Chaloupka, F. and H. Saffer, 1992, Clean air laws and the demand for cigarettes, Contemporary Policy issues IO, 72-83. Duffy, J. and G. Cohen, 1978, Total alcohol consumption and excessivedrinking, British Journal of Addiction 73. Laugesen, M. and C. Meads, 1991, Tobacco advertising restrictions, price, income and tobacco consumption in OECD countries, 1960-1986, British Journal of Addiction 86, 1343-1354. Saffer, H. and M. Grossman, 1987, Drinking age laws and highway mortality rates: cause and effect, Economic Inquiry 25,403-417. Saffer, L-i., 1991, Alcohol advertising bans and alcohol abuse: an international perspective, Journal of Health Economics IO, 65-79. Young, D., 1993, Alcohol advertising bans and alcohol abuse: comment, Journal of Health Economics 12, OOO-OOO.
Alcohol advertising bans and alcohol abuse: Reply Henry Saffer* National Bureau of Economic Research, New York, NY, USA Final version received April 1993
1. Introduction Saffer (1991) found that alcohol advertising increased alcohol consumption and abuse. This is an interesting finding since several older studies of alcohol advertising find no effect. These prior studies were important efforts and represented the state of econometric knowledge at the time they were published. Recent research using econometric models of advertising has provided more detailed standards for this type of work. Berndt (1991) reviews severai of these econometric issues. Four of the more important issues in econometric studies of alcohol advertising are: (1) testing non-linear specifications, (2) simultaneity between consumption and advertising, (3) temporal aggregation bias, and (4) the correct measurement of advertising. The older studies of alcohol advertising ignored some or all of these issues. The issue of measurement is particularly important in distinguishing SafFer $991) from prior studies. Most prior alcohoi advertising studies used expenditure data rather than advertising bans as the measure of advertising. However, expenditure data vary over a limited range. The relationship between alcohol use and alcohol advertising is subject to diminishing returns. In the range that is employed by expenditure studies these returns may be quite small or zero. l This results in an observed marginal efYect of Correspondence to: ID. Henry Saffer, National Bureau of Economic Research, 269 Mercer St., 8th Floor. New York,,, NY IOW3, USA. *I would like to thank Frank Chaloupka, Michael Grossman and Theodore Joyce for helpful commenason an earlier version of this reply. ‘An increase in advertising expenditure by a firm may increase its market share but may ulso lead to a corresponding increase in advertising by rivals. Market shares initial B~cvcls but at a higher level of advertising for all firms. This level of nae by rivals crnd can result in higher than in the expenditures in the n
230
H.
So$er,
Alchohnl
advertisirtg ham
artd alcohol abuse:
Reply
advertising that is also quite small or zero. This does not mean that advertising has no effect on consumption. Advertising must be observed over a wi&r range in order to correctly estimate its effects. A data set which inch::: 9:s plac es -+zhich allow advertising and p&es which ban advertising is thy. :‘ pirical equivalent to measuring advertising over a wide range. Over this range the relationship between aggregate consumption and advertising can be observed. A similar study of cigarette advertising bans In 22 OECD countries conducted by Laugesen and IvIeads ( 1991j also finds that advertising bans reduce consumption. 2. Cultural differences The issue of cultural differences or sentiment must be addressed when using an international data set. It is likely that there are unobserved variations in sentiment towards alcohol across countries. This sentiment may be exogenous or endogenous to consumption and could affect both alcohol consumption and the enactment of alcohol advertising bans. In the case of t;xogeno~s sentiment some type of proxy variable is needed. The variable I used is dichotomous and equal to one for those countries where more than 75 percent of the alcohol consumed is in the form of beer and wine. Nine of the 17 countries in the data set fall in this category.2 An interesting extension of my 1991 study might be to find some alternative or additional measures of sentiment such as religion. Young does not provide any new sentiment proxies. Another method I used to account for exogenous sentiment are fixed effects models. These fixed effects models exclude the advertising ban variables due to collinearity problems. The country dummies are grouped together by advertising policy and a mean coefficient and t-value computed. These mean coefhcients show increases in all three measures of alcohol abuse as advertising increases. Young (1993) ignores the collinearity problem and simply uses both the advertising variables and the country dummies in the same regressions. Since there were only three countries that changed policies during the sample period there exists a linear transformation of the advertising dumr- ‘I variables which is almost identical to a linear transformation of the country aummy variables. Young’s advertising ban coer%cients in these regressions are unstable and the t-v,Jues are meaningless due to the collinearity he has created. 21 also experimented with a per capita production oG pure alcohol as an alcohol attitude measure. The results from these regressions show that alcohol advertising bans reduce alcohol abuse and were similar to the results reported in gaffer I,1991). While the regression results are reasonable, a regression with log per capital ccnsumption as dependent variable and per capital production as an independent variable is approximately the sanle as using per capita exports as the dependent variable. This also is a problem in the liver cirrhosis and highway fatality regressions.
As I noted in footnote 12 of my 1991 study, endogeneity problem. Alcohol sentiment or cultural attitudes could, in part. of consumption. The passage of advertising bans could also, function of alcohol sentiment. This suggests that consumption ing bans could be endogenous. Saffer and Grossman
is a potential be a function in part, be a and advertis-
3These problems also occur in the estimation of price effects and have resulted in a ge..eral shift away from estimating separate beer, wine and spirits demand curves. The estimation of all own-effects and cross-effects parameters, in both the case of price and advertising, may be too demanding for the available data.
232
H. Sager, Akhohol adoerrisirig bans and
alcohol abuse: Reply
Young does not provide any explanation of why his results are not symmetric. The third reason for not specifying a set of dissaggregated demand curves is that each demand curve should include an own price variable.’ Young uses only a composite alcohol price variable. There is no way of interpreting the estimated effect of this type of variable in separate regressions of beer, wine and spirits demand. Since the sign of the price variable is uncertain it is hard to judge the validity of the equation specification. A fourth reason for not disaggregating total consumption is that this disaggregation exacerbates the cultural difference problem. There is more variation across countries in beer, wine and spirits consumption due to cultural differences than there is in total consumption. Young includes my sentiment variable which is equal to 1 for countries where more than 75 percent of the alcohol consumed is in the form of beer and wine. In regressions using beer, wine and spirits as dependent variables, this sentiment variable has an almost mechanical relationship with the dependent variable creating biased estimates. Young’s expansion of total consumption into beer, wine and spirits consumption is redundant with total consumption, is diRcult to interpret, makes estimation mere difficult and has no potential of adding any new insight. Several hypotheses about how a ban on one alcoholic beverage affects consumption of another alcoholic beverage are possible. However, Young’s conclusion that advertising bans increase beer, wine and spirits consumption is rather surprising. 4.
Serial correlation
Young states that his Durbin-Watson statistics ‘do not take account of the panel nature of the data’. This is very puzzling since Young states that in testing for higher order serial correlation he is able to account for the panel nature of the data. Apparently, Young computed the Durbin-Watson statistic as if the data were a single time series rather than 17 time series. This cannot be done since the estimation of rho would use last observation of country A followed by the first observation of country B. The value of rho should be computed within each country and then could be averaged over countries. This average value of rho could be used for an overall DurbinWatson statistic. The serial correlation correction Young uses in his regressions use the restrictive assumption that rho is the same for all countries in the data set. It 4Young notes that the expenditure data could LC measured with error which would bias the price variable. However, to the extent that the error in the expenditure data is correlated with the error in the consumption data these errors will cancel and tF : resulting regression will be unbiased.
is unclear how this value of rho was computed. In any case, Young’s rho is close to one for e;?ch equation. This results in transforming the data into approximately first differences. Since only three countries had any change in advertising policies during the sample period, the transformed advertising data approximates two vectors of zeros. The advertising ban coefficients and their t-values resulting from this transformation are meaningless.
5. Mean values of alcohol use by advertising policy
Young’s table 2 is a table of means of the alcohol ;huse measure by advertising ban category and time. This table is mainly descriptive since other factors affecting alcohol abuse are not held constant. Young extends the table to include t-values and f-tests. However, Young chooses a null hypothesis of no difference between the means which implies a two tail test. The null hypothesis should be that one group has a higher mean than the other, which implies a one tail test. Also, Young’s choice of five percent significance level is arbitrary. Significance at 10 percent or even 15 percent is aiso llrzaningful information. The beer, wine and spirits means are not good indicators of alcohol related public health and should be ignored. Using a tvalue of 1.3 to define significance, 6 of the 12 differences are significant. However, even if the b-values resulted in no significant differences it simply points out that other factors affecting demand must be controlled.
6. Summary
Young reexamines the work first presented in Saffer (1991) on the relationship between alcohol advertising bans and alcohol use. While there are several interesting possibilities for extending Saffer (Igal), Young fails to provide any new insights. Most of his paper is devoted to recreating my data and results ? He tries to extend my work with three new specifications. Each of these new specifications is flawed. His errors include an iaappropriate application of fixed effects models, inappropriate dependent variables, and incorrect serial correlation computations. The results of these errors are a series of inconsistent advertising ban coeff’icients.He concludes that these inconsistent results are evidence that advertising bans have no effect of alcohol abuse. It might be better to at ‘least provide a series of consistent regression models before coming to any conclusions.
‘Young states in footnote 6 that he had private correspondence with Michele Hainut of the OECD regarding the quality of the Portuguese data. 1 wrote to her about this statement. She said she had never made the statement that Young attributes to her and had Liticoncerns about any of the OECD data.
234
t4. Sager. fflchohoi adllertising bans and alcohol abuse: Reply
cferences Berndt, E., 1991, The practice of econometrics: classic and contemporary (Addison-Wesley Publishing Co, Reading, MA). Bruun, K. et al., 1975, Alcohol control policies in a public health perspec!ive, The Finnish Foundation for Alcohol Studies 25. Chaloupka, F. and H. Saffer, 1992, Clean air laws and the demand for cigarettes, Contemporary Policy issues IO, 72-83. Duffy, J. and G. Cohen, 1978, Total alcohol consumption and excessivedrinking, British Journal of Addiction 73. Laugesen, M. and C. Meads, 1991, Tobacco advertising restrictions, price, income and tobacco consumption in OECD countries, 1960-1986, British Journal of Addiction 86, 1343-1354. Saffer, H. and M. Grossman, 1987, Drinking age laws and highway mortality rates: cause and effect, Economic Inquiry 25,403-417. Saffer, L-i., 1991, Alcohol advertising bans and alcohol abuse: an international perspective, Journal of Health Economics IO, 65-79. Young, D., 1993, Alcohol advertising bans and alcohol abuse: comment, Journal of Health Economics 12, OOO-OOO.