How far to the border? The extent and impact of cross - border casual cigarette smuggling.

(23) One complication with using state or MSA fixed effects is multicollinearity with prices. I run auxiliary regression of home state price on a year trend and state fixed effects and find an [R.sup.2] of 0.82. The associated variance inflation factor (VIF = [R.sup.2]/1 - [R.sup.2]) is 4.42. A VIF less than ten is typically considered an acceptable amount of multicollinearity, so the fixed effects are not soaking up all of the price variation in my regressions.

(24) Using taxes to instrument for prices is also beneficial because the price variation due to cigarette tax changes more likely identifies the demand curve. Much of the evidence on cigarette taxes suggests these taxes are either fully or more than fully passed on to consumers (Chaloupka and Warner, 2000). Using the price data described in the previous section, l regress real state price on real state taxes with state fixed effects and a year trend for 1992-2002. I estimate a coefficient of 1.28 on the tax variable with a standard error of 0.003. Due to this evidence, I will assume throughout that supply is inelastic and that the parameters estimated in the demand function are not confounding supply and demand. This assumption is prevalent in the literature.

(23) The evidence on how states set cigarette excise taxes, while sparse, supports this assumption. The cross-state variation in excise taxes is driven largely by differences in attitudes towards smoking as well as by economic factors that may lead states to increase excise taxes as a way to raise revenue (ACIR, 1985). (26) Results and conclusions are qualitatively similar when I use the individual--level data clustered at the state-MSA level, Results from such regressions are available from the author upon request.

(27) The results and conclusions are unchanged when I use year fixed effects or survey date fixed effects instead of a linear year trend.

(28) Including log distance as a regressor, equation [6] can be interpreted as a specific form of a more general log-linear second order demand function approximation. The second order approximation includes the ln([P.sub.h]), ln([P.sub.h]) - ln([P.sub.b]) and ln(D) terms as well as all squared terms and cross-products. While there are some quantitative differences, the elasticity estimates from the full second order log linear approximation are qualitatively similar to the ones presented and are available upon request. Thus, while the demand model presented in the fourth section is useful in providing an interpretation of the regression coefficients, my results are robust to a more general demand function approximation that embodies fewer assumptions.

(29) One potential bias in identifying the parameter on the log distance, log price difference variable is the existence of Internet smuggling. Goolsbee, Lovenheim, and Slemrod (2007) find evidence using CPS Internet data and taxed state sales of substantial Internet smuggling, which would bias my estimates because one would expect as distance to a lower-price locality increases, the likelihood of smuggling over the Internet would also increase, ceteris paribus. Excluding Internet smuggling might cause an overstatement of the estimated impact of distance on demand. To check whether this is the case, I interact average MSA Internet connectivity calculated from the CPS as described in Goolsbee et al. (2007) with the price difference, log distance interaction term. If the exclusion of the Internet is a source of bias, the coefficient on the triple interaction term should be positive and significant. The point estimates are negative, small and not significant, however, and the other coefficients are quite similar to those in Table 6. Results are available from the author upon request.

(30) Log distance is likely to be correlated with (ln([P.sub.h]) - ln([P.sub.h]))*ln(D). Thus, although the coefficient on In(D) is not statistically differentiable from zero, its exclusion from the regression may affect the coefficients on other variables. I estimate the demand model both including and excluding log distance and find no difference in results.

(31) Chaloupka and Warner (2000) report these studies are consistent in estimating elasticities in a neighborhood of -0.4.

(32) Interestingly, when 1 set [rho] = 1 within 20 miles of the border and [rho] = 0 outside of 20 miles of the border, I find elasticities that are strictly between my full price elasticities and the "naive" elasticities in columns i and ii. The same result occurs when I set [rho] = 0.5 within 25 miles of the border and [rho] = 0 outside of 25 miles. Such methodologies replicate the strategies of Lewit et al. (1981), Lewit and Coate (1982), and Chaloupka (1991), and the results are evidence that exogenously setting [rho] in this manner only partially accounts for smuggling behavior.

(33) If I do not rescale the negative values to zero in equation [10], 1 estimate between 7 and 23 percent of consumers purchase cigarettes in lower-price localities. Thus, my results and conclusions are not sensitive to rescaling.

(34) A central reason for the difference between my estimates and those in Stehr (2005) is due to downward bias in his estimates. He identifies casual smuggling off of the average tax difference between the home state and all border states that have a higher tax than the home state. The main reason for the downward bias is when a state raises its tax level, this average difference will increase by less than the tax increase and can decrease due to the fact the tax increase can change the pool of higher-price states. The first states to drop out will be the lowest price "export" states. My estimates imply a one-cent increase in the home state tax causes a 0.24-cent drop in the average "export" state tax. This effect severely weakens the relationship between ln(consumption) - ln(sales) and the tax difference. Further, utilizing tax differences rather than price differences introduces measurement error as more than ten percent of tax differences have a different sign than the respective price difference. One can expect this measurement error to further obfuscate the smuggling regression in Stehr (2005).

(35) This calculation is based on an average cigarette shelf life of eight months (Wong, Ashcraft, and Miller, 1991 ). They report the shelf life of "normal cigarettes."

(36) Home state price elasticity and percentage smuggling estimates by state-MSA are presented in Appendix Table C-3 in Lovenheim (2007).

(37) For each MSA, I multiply the smuggling percentage by the number of cigarettes smoked. Summing this number within states gives the total number of consumed cigarettes purchased in another jurisdiction. I then attribute these purchases to the closest lower-price state for each MSA to find the sales increases due to smuggling in each state. The denominator in each calculation is the total consumed cigarettes in each state. TABLE 1 STATES THAT TAX CIGARETTE SALES TO NON-TRIBAL MEMBERS ON NATIVE AMERICAN RESERVATIONS State Statute/Case Name Arizona A.R.S. 42-3302 Kansas State v. Oyler Michigan MCLS 205.30c /Individual Tribal Compacts Minnesota Minn. Statute 297F.07/ Individual Tribal Compacts Montana Moe v. Confederated Salish and Kootenai Nebraska Nebraska Department of Revenue (1996) Nevada NRS 370.280 Oklahoma Okl. St. 349 Oregon ORS 323.401 South Dakota Individual Tribal Compacts Washington Washington v. Confederated Colville Tribes Wisconsin Wis. Stat. 139.323 /Individual Tribal Compacts State Year Arizona 1997 Kansas 1990 Michigan 1947 Minnesota 1997/Pre-1992 Montana 1976 Nebraska Pre-1992 Nevada 1947 Oklahoma Pre-1992 Oregon 1979 South Dakota Pre-1992 Washington 1980 Wisconsin 1984 Source: ACIR (1985) updated using LexisNexis searches for state cigarette taxation laws. TABLE 2 TAX CHANGES, PRICE DIFFERENTIALS, AND DISTANCE BY STATE

Average