Let Hijt be the happiness of individual i who lives in state j at time t, and Tjt be the real level of cigarette taxes in state j at time t. A simple regression that relates happiness to cigarette taxes in the state would be:
where Pj are state fixed effects and n are year fixed effects, respectively. These fixed effects completely control for any fixed differences between states and between years, which means that only within-state variation in cigarette taxes is used in the estimation.
Though it deals with many of the obvious endogeneity problems inherent in using state policy, this approach may still have problems. For example, if states are changing cigarette taxes at different points in their state business cycle, the estimated ”effect” may instead reflect the effect of these economic conditions. Another potential omitted factor from this model is the state spending (or reduced other taxes) that is financed by cigarette taxation. If we find that higher cigarette taxes lead to a general rise in happiness that could simply reflect the fact that these revenues are used in a welfare-enhancing way. Finally, we have the fact that only about a third of our sample smokes on average, so an impact for smokers could be masked in the full sample. buy zyrtec online
To address this problem, we exploit the fact that cigarette taxes should only affect the happiness of those who are smokers (and former smokers). We can therefore compare the effect of taxes on this group to taxes on those who do not smoke. We cannot do so by using direct data on smoking behavior, for three reasons: smoking decisions are endogenous to tax rates; the happiness effect in our model should operate through both current and former smokers; and smoking data are only available for a subset of years in both surveys. We therefore compare the impact of excise taxation on predicted smokers.
Specifically, we first estimate a regression that relates smoking behavior to the observable predictors of smoking we see in the GSS data. Most of the variables are available in both countries, but some are available only in one or the other; we used the broadest set of covariates possible to generate the best possible prediction of smoking behavior. Our predictors are: age category and gender interactions; household income quartile dummies; personal income quartile dummies (Canada only); education categories (high school dropout, high school graduate, some college, and college graduate); education of the respondents mother and father (by the same categories; U.S. only); race (white, black, and other; U.S. only); marital status (married, divorced/separated, widowed, never married); dummies for number of children (U.S.) or household size (Canada);
dummies for full time work, part time work, unemployed, out of labor force, and whether ever worked (U.S. only); religious attendance (8 categorical values in U.S. that rise monotonically with attendance; three dummies for weekly, monthly, or annual attendance in Canada); born in Canada; live in house or apartment (Canada only); own your house (Canada only); language spoken at home (Canada only); and the state/year or province/year unemployment rate. We estimate such an equation for each year that has smoking information, and use that to form a predicted probability of smoking (PREDSMOKijt).