Monday, September 25, 2006
Alcohol and Income
Specifically, Peters and Stringham base their conclusions on a regression with the following form:
Ln(income) = a + B*Drink + C*Bar + X’d + e
where Drink = 1 if you drink alcohol, Bar = 1 if you visited a bar in the past month, and X is a vector of basic controls (age, race, gender, education, …).
Students of econometrics will recall that B and C provide unbiased estimates of the relationship between alcohol consumption and income under a variety of assumptions – the key one for our purposes being that the Drink and Bar variables are uncorrelated with the error term. It doesn’t take a lot of imagination to think that this assumption may not hold in this particular case.
First, it is possible (likely even) that an individual’s income affects their propensity to drink or visit bars (rather than strictly vice versa). In this case the relationship is endogenous and the coefficient estimates are unreliable.
Second, setting aside the endogeneity, the variables in the X vector do not adequately eliminate the correlation between the variables of interest and the error term. There are a number of obvious characteristics that determine both a person’s propensity to drink and their ability to earn that are not included in the regression. The most obvious of these is the person’s social capital itself.
I certainly believe that individuals with high social capital earn more money, but I also believe that individuals with more social capital are more likely to drink and to drink at bars. Failure to include sufficient controls for individuals’ social capital implies (setting aside bias introduced by endogenity) the coefficient estimates are largely reflecting the relationship between social capital and income, not the relationship between drinking and income.
To show that I am not just making this up, I turned to the National Longitudinal Survey of Adolescent Health and asked – among those who had not consumed alcohol during the past 12 months when they were first questioned (and were 12-18 years old), were the abstainers with more social capital in the first wave more likely to have started drinking by wave 2 or 3? If we observe that more socially adept adolescents are more likely to switch from abstainers to drinkers, than we can feel more confident that Peters and Stringham’s drinking variables are just proxies for unobserved social skills.
The lazy blogger (as opposed to rigorous academic) regressions I ran show that, among those who were not drinking at the time of the first wave, more socially adept adolescents were more likely to take up drinking by the second or third waves. Specifically, I regressed a binary variable for consumption of alcohol in waves 2 and 3 (which were18 months and 6 years after the first sampling) on three different social capital metrics (ln(size of social network), ln(number of school activities participating in), and whether or not the person had dated in the last 18 months) from the first wave and basic controls for age, gender, race, parent education and income. The relationship between the social capital metrics and the probability that respondents was drinking in the subsequent waves is presented in the table below.
The relationship between social capital and subsequent alcohol consumption is positive and significant nearly every case. To get a rough sense of the magnitudes, doubling the size of a students social network in wave 1 increases the probability that they are drinking by wave 2 by 4.5 percentage points or 14.5 percent (at the mean). Alternatively, doubling the number of high school memberships increases the propensity to drink in the second wave by 4.7 percentage points and having dated increases it by 9.7 percentage points.
While the relationship between drinking and income is at first striking, like the relationship between drinking and health, Peters and Stringham’s results appear to be the result of poor empirical technique. So it is probably best to not use these results to justify policy recommendations (as the Reason Foundation was so eager to do).
Have you seen Chatterji and DiSimone, "High School Alcohol Abuse and Young Adult Labor Market Outcomes," NBER Working Paper No. 12529? Do you think the same argument you make about the here applies there?
THe Chatterji and DiSimone paper essentially makes the same argument that I am making. They aren't arguing that adolescent binge drinking causes higher earnings, but rather that variable serves as a proxy for social skills. So the coefficient they find on binge drinking does not, necessarily, reflect a causal relationship between adolescent drinking and earnings, rather it is merely picking up the effect of unobserved, but valuable, social skills that are correlated with adolescent binge drinking.
Subscribe to Post Comments [Atom]
Subscribe to Posts [Atom]