Entrepreneur: Start & Grow Your Business

Internet purchases, cross-border shopping, and sales taxes.


by Ballard, Charles L.^Lee, Jaimin
National Tax Journal • Dec, 2007 •

INTRODUCTION

Consumers can purchase goods in any of several ways. They can buy in their own jurisdiction, paying sales taxes if the jurisdiction levies a sales tax. Alternatively, consumers can travel across a border and make their purchases in a different jurisdiction, paying the sales taxes that apply in that jurisdiction. A third possibility is that they can buy over the Internet. (1) Currently, state and local governments cannot require an online firm to collect sales taxes or use taxes, unless the firm has a physical presence (nexus) in the taxing jurisdiction. (See Fox and Murray (1997) for a discussion of the relevant Supreme Court rulings.) In 2003, several large retailers announced that they would begin to collect state sales tax on their online sales. Also, several states are attempting to negotiate an agreement for mutual enforcement of these taxes. Nevertheless, the data analyzed in this paper are for 1997 and 2001, which are years in which it is believed that evasion of sales taxes on Internet purchases was very substantial. (2)

In this paper, we combine the analysis of the effects of taxes on Internet shopping with the analysis of cross-border shopping. We analyze Internet shopping behavior, using a data set for the United States that includes the general retail sales-tax rates in the consumer's home county and in adjacent counties. (3) Our results are consistent with the interpretation that Internet shopping, cross-border shopping, and home-county shopping are substitutes.

The literature on cross-border shopping includes both theoretical and empirical contributions. (4) For example, in a theoretical paper, Ohsawa (1999) uses a Hotelling-type model of spatial competition to examine the effects of the size and spatial arrangement of countries on tax rates and government revenues, in a Nash-equilibrium setting. There is also a substantial empirical literature. Gordon and Nielsen (1997) estimate that 3.9 billion Danish Kroner of value added escaped taxation by Denmark in 1992 because of cross-border shopping. FitzGerald (1992) finds substantial evidence of tax-induced cross-border shopping between the Republic of Ireland and the United Kingdom. Ferris (2000) analyzes border crossings from Canada to the United States from 1972 to 1997, and finds that taxes play a significant role, as do exchange rates and other variables. Garrett and Marsh (2002) use 1998 county-level data from Kansas to investigate cross-border lottery shopping between Kansas and its neighboring states. Their results suggest that the Kansas lottery gains revenue from Oklahomans who cross into Kansas, but suffers a net loss of revenue because of Kansans who cross into Missouri and Nebraska.

There is also a recent literature on the effects of taxation on Internet sales. Goolsbee (2000) tests the relationship between sales-tax rates and Internet purchases, using data for 1997 from Forrester Research (a market research company). Goolsbee knows the metropolitan area in which a consumer lives, but not the county of residence. Thus, Goolsbee assumes that the tax rate is uniform throughout each metropolitan area. (5)

Alm and Melnik (2005) also investigate the effects of sales-tax rates on Internet purchases. They use data from the Current Population Survey for 2001. They define the tax rate at the state level, using the lowest sales-tax rate available to the consumer in his/her state of residence.

In this paper, we also use the Current Population Survey. However, we exploit the fact that the county of residence is known for some of the respondents in the sample. Thus, we define the sales-tax rate at the county level. Performing the analysis at the county level plays an important role in our study of cross-border shopping because the extent of cross-border shopping is likely to diminish with distance. (6) Virtually all residents of the United States are fairly close to a county boundary, but many are much farther from the nearest state line. (7)

Goolsbee (2000) and Alto and Melnik (2005) address only the tax-evasion aspect of Internet shopping (i.e., they do not deal with cross-border shopping). Goolsbee finds that Internet purchases are highly sensitive to sales taxes, with tax-price elasticities in the range of two to four. Alm and Melnik also find a significant effect of sales taxes on Internet purchases, although their estimated tax-price elasticities are much smaller, around 0.5.

We also find that Internet purchases are significantly more likely for consumers who face higher sales-tax rates, all else equal. Our estimated tax-price elasticities are closer to those of Alm and Melnik than to those of Goolsbee. Even so, we consider the quantitative magnitude of the effects to be fairly substantial.

We also find that consumers whose home county is adjacent to a county with a lower sales-tax rate are significantly less likely to use the Internet for shopping, all else equal. This does not provide any direct evidence regarding cross-border shopping, but it is consistent with the interpretation that consumers are engaging in cross-border shopping.

The focus of most of this literature is on the effect of tax rates. Clearly, however, the definition of the tax base also has the potential to affect shopping behavior. There is wide variation among the states in the definition of which goods and services are subject to the retail sales tax. We include a variable that is designed to capture this effect. We find that Internet shopping is significantly more likely for consumers who reside in states in which the base of the sales tax is broader, all else equal. Thus, our results suggest that consumers may use the Internet to avoid sales taxation, because of a high sales-tax rate in their jurisdiction, and/or because the sales tax in their jurisdiction applies to a wide range of purchases.

DATA AND VARIABLES

We use data from the Current Population Survey (CPS) for 1997 and 2001. (8) Our key dependent variable, SHOP INTERNET, measures whether a consumer engages in online shopping. SHOP INTERNET is a binary variable. (9)

The CPS data include some individuals who have access to the Internet, but also some who do not. If we were to restrict the sample to include only those who have access to the Internet, there is the possibility of sample-selection bias. We deal with the issue of sample selection by estimating a system of two probit equations, including a selection equation for Internet access, as well as the equation for Internet shopping. The variable that measures whether the consumer has access to the Internet, ACCESS INTERNET, is also a binary variable.

One key explanatory variable is HOME TAX PRICE, which is the tax price associated with the sales tax of the consumer's home county. HOME TAX PRICE = (1 + [t.sub.h]), where [t.sub.h], is the sales-tax rate in the home county (measured as a proportion). The states can be divided into four categories, with respect to their general retail sales taxes. First, there are no general retail sales taxes in Delaware, Montana, New Hampshire, and Oregon, either at the state level or at the county level. (10) Second, 13 states impose a uniform sales-tax rate throughout the state, (11) and the District of Columbia has a uniform sales-tax rate within its borders. Third, eight states have variation in sales-tax rates across counties, but no variation within counties. (12)

For the remaining 25 states, there are variations in sales-tax rates, both across counties and within counties (i.e., cities within the same county can have different sales-tax rates). However, although we know the sales-tax rates in each city, we only have data on the county of residence. Thus, we calculate a weighted average sales-tax rate for each county, using the populations of cities within the county as weights. All consumers in a given county are then assumed to face the same sales-tax rate.

Our data on sales-tax rates were collected by a variety of methods, including telephone and mail contacts with state and county revenue officials, as well as visits to the websites maintained by some state and county revenue offices. In many counties, the sales-tax rate was the same in 1997 as it had been in 2001. In jurisdictions where the tax rates did change during this period, the changes were usually small. Thus, most of the variation in HOME TAX PRICE in our pooled cross-section data set is across counties, rather than across years.

We also include a variable that measures the relationship between the sales-tax rate in the consumer's home county and the sales-tax rates in the immediately adjoining counties. This variable, TAXRATIO, is (1 + t.sub.h])/(1 + [t.sub.f]), where [t.sub.h] is the sales-tax rate of the home county and [t.sub.f] is the minimum of the sales-tax rates among all of the adjacent counties.

This specification for TAXRATIO does not necessarily capture all aspects of cross-border shopping. For instance, a consumer may engage in cross-border shopping by going two or more counties away, rather than to an adjoining county. Also, we treat everyone in a given county the same, regardless of whether he or she lives close to the county line or miles away. Nevertheless, we argue that this variable provides valuable information regarding the incentives for cross-border shopping.

HOME TAX PRICE and TAXRATIO are calculated on the basis of sales-tax rates. However, the effects of the sales tax in a given jurisdiction will depend on the tax base, as well as on the tax rate. (If an item is excluded from the sales-tax base, its effective sales-tax rate is zero, regardless of the statutory sales-tax rate.) There is considerable variation among the states in the base of the sales tax. (Although counties have some latitude to set the sales-tax rate in 33 of the 46 states with a retail sales tax, the tax base is set at the state level.) (13) There may also be cross-state differences in the intensity of sales-tax enforcement. As a result of these differences, there is variation among the states in sales-tax revenue, even when we hold constant the level of the sales-tax rates. In an attempt to control for these differences, we have created a variable called TAXBASE. In creating TAXBASE, we begin by calculating the sales tax as a percentage of personal income in each state. (14) We then divide by the weighted average of the sales-tax rates in the state. For TAXBASE, as for HOME TAX PRICE, there is considerable variation by region, but relatively little variation across years.

LOGINCOME is the log of the resident's household income. We expect a positive coefficient for LOGINCOME, since use of the Internet for purchases is almost certain to be a normal activity. (15)

FEMALE, WHITE, and MARRIED are binary variables for the individual's demographic characteristics. HIGHGRAD, COLLGRAD, and PROGRAD are a set of dummy variables for the individual's education level, indicating whether his or her highest level of educational attainment is a high-school diploma (HIGHGRAD), a Bachelor's degree (COLLGRAD), or an M.A., Ph.D., or professional degree (PROGRAD). (The omitted category is the group who have less than a high-school education.) NUMCOMP is the number of computers in the household. AGE15, AGE20, AGE30, AGE50, and AGE60 are a set of age-group dummy variables. For example, AGE15 is equal to one if the consumer's age is from 15 to 19, AGE20 equals one for consumers aged from 20 to 29, and AGE60 equals one for those whose age is greater than or equal to 60. (The omitted category is the group of consumers aged 40-49.) D2001 is a dummy variable, equal to one when an observation is taken in 2001, and zero when it is taken in 1997.

We include the percentage change in the Consumer Price Index (which we call CPI-CHANGE) as an explanatory variable. (16) The inflation rate in the local area may play a role in driving consumer awareness of the desirability of out-of-area shopping methods, such as mail-order shopping and Internet shopping. Therefore, we expect inflation rates to have a positive effect on Internet shopping.

Table 1 presents summary statistics for several of the most important variables in our data set. Table 1 includes separate columns for those without Internet access, those with Internet access who do shop online, and those with Internet access who do not shop online, as well as for the entire sample. Not surprisingly, the sample is predominantly white. More than half of the individuals in our sample are married.

For the sample as a whole, the average sales-tax rate at the county level is about 6.29 percent. The means reported in Table 1 suggest that those with Internet access tend to be somewhat younger, more affluent, and more highly educated than those without Internet access. Those with Internet access are also more likely to be unmarried, white, and male, although the differences among the sub-groups tend to be relatively modest. Of course, caution must be used when interpreting simple univariate means. In particular, the simple means in Table 1 indicate that sales-tax rates are actually lower for Internet shoppers than for Internet users who do not shop, on average. This simple univariate relationship would appear to contradict our hypotheses. However, multivariate analysis suggests that Internet shopping is more likely for those who face higher sales-tax rates, all else equal.

The mean value of TAXRATIO is about 1.0066. This indicates that the average individual in our sample lives adjacent to a county in which the sales-tax rate is about two-thirds of one percentage point lower than in his or her home county.

The mean value of TAXBASE is about 0.41. This means that, on average, each percentage point of sales-tax rate generates sales-tax revenue equal to about four-tenths of one percent of personal income. One important reason why the mean value of TAXBASE is less than one is that, in most states, the sales tax applies to relatively few services.

EMPIRICAL SPECIFICATION

Counties may differ in a variety of ways, even when they are in the same state, and even when they have the same sales-tax rate. (For example, some counties have better Internet infrastructure and/or better transportation networks than others.) Therefore, we control for county effects in all of our econometric specifications.

Our dependent variable, SHOP INTERNET, is equal to one if the consumer shops over the Internet, and zero otherwise. It is possible to use SHOP INTERNET as the dependent variable in a standard probit model, using the sample of Internet users. However, this procedure will not necessarily generate consistent estimates. The coefficients may be subject to sample-selection bias, because a consumer can only shop online if he or she has access to the Internet in the first place.

Thus, a superior estimation strategy is to use maximum-likelihood techniques to estimate simultaneously a probit equation for Internet access and a probit equation for Internet shopping. (For discussion of models of this type, see Alm and Skidmore (1999) and Van De Ven and Van Praag (1981).) The Internet-access equation strengthens the interpretation of the Internet-shopping equation, but the coefficients from the Internet-access equation are also of interest in their own right. (17) The dependent variable in the selection equation for Internet access, ACCESS INTERNET, is equal to one if the consumer has Internet access, and zero otherwise.

Thus, our model consists of equations [1] and [2]:

[1] [SHOP INTERNET.sub.i]

= 1 if [SHOP INTERNET.sup.*.sub.i]

= [x.sub.i][beta] + [u.sub.i] > 0;

= 0 otherwise;

[2] ACCESS INTERNET

= 1 if [ACCESS INTERNET.sub.i]

= [z.sub.i] [gamma]+ [e.sub.i] > 0;

= 0 otherwise.

In equation [1], [SHOP INTERNET.sup.*] is a latent variable, x is a vector of explanatory variables, [beta] is a vector of coefficients, and u is an error term. In equation [2], [ACCESS INTERNET.sup.*] is a latent variable, z is a vector of explanatory variables, [gamma] is a vector of coefficients, and e is an error term.

We specify equation [1] using the explanatory variables already described, along with a full set of interactions among the variables for race, sex, and marital status. Thus,

[3] [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

We specify the selection equation for Internet access with the same variables that were shown in equation [3]. Thus,

[4] [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

It is possible for the Internet-shopping equation to include all of the variables in [4]. However, if the selection equation and the equation of primary interest include the same set of explanatory variables, identification is achieved exclusively through the non-linearity of the probit. Thus, it is preferable for at least one of the explanatory variables in the sample-selection equation to be absent from the equation of primary interest. (For discussion, see Wooldridge (2002) and Van Ham and Buchel (2004).) We will report on five specifications of the maximum-likelihood model, in which different combinations of variables are excluded from the Internet-shopping equation. For purposes of comparison, one of these specifications includes the same set of explanatory variables in both the selection equation and the equation of interest.

We assume that the error terms in the two equations are distributed according to a bivariate standard normal distribution. Following the notation of Alm and Skidmore (1999), we denote the correlation between u and e as [[rho].sub.ue]. If [[rho].sub.ue] = 0, probit estimates based on the selected sample of those with Internet access would be consistent. However, in the more general case in which [[rho].sub.ue] is nonzero, probit estimates based on the selected sample will be biased, since

[5] E([SHOP INTERNET.sup.*.sub.i] | [X.sub.i], [Z.sub.i], & [ACCESS INTERNET.sub.i] = 1) = [X.sub.i][beta] + E(u | [X.sub.i], [Z.sub.i], & [ACCESS INTERNET.sub.i] = 1) = [X.sub.i][beta] + [[rho].sub.ue](e | [Z.sub.i] & [ACCESS INTERNET.sub.i] = 1).

The likelihood function for this model is

[6] [MATHEMATICAL EXPRESSION NOT REPRODUCIBLE IN ASCII]

where [N.sub.1] refers to those who do not have access to the Internet, [N.sub.2] refers to those who have access to the Internet but do not shop over the Internet, [N.sub.3] refers to those who have access to the Internet and do shop over the Internet, [[PHI].sub.1] is the univariate standard normal density function, and [[PHI].sub.2] is the bivariate standard normal density function.

Results for Internet Access

Our results are shown in Table 2. (18) The first part of Table 2 shows the estimates for the Internet-shopping equation; the second part shows the estimates for the Internet-access equation. We begin our discussion of the results from the system of two equations by looking briefly at the results for Internet access.

For most variables, the coefficients and significance levels in the Internet-access equation are quite similar across the five specifications in Table 2. Thus, readers can get a good sense of the results by scanning down any of the five columns. We have a slight preference for Specification (2), because the interaction terms excluded from the Internet-shopping equation in that specification have insignificant effects in all of our other specifications. (19)

The results in the second part of Table 2 suggest that consumers in a high-tax county are more likely to have Internet access, and that those who live adjacent to a low-tax county are less likely to have Internet access. However, these effects are not statistically significant. (20) Table 2 also indicates that higher-income consumers are significantly more likely to have Internet access than those with lower incomes. Those with at least a high-school education are significantly more likely to have Internet access than those without a high-school education. Not surprisingly, the coefficient on the variable NUMCOMP (number of computers in the household) indicates that households with more computers are significantly more likely to have Internet access. The results indicate that Internet access (as opposed to Internet shopping) decreases monotonically with age.

The estimates of [[rho].sub.ue] (rho) are significantly different from zero in all five of our specifications. (21) Thus, the error term in the Internet-access equation is correlated with the error term in the Internet-shopping equation. This means that the simple probit model based on the selected sample of Internet users is indeed subject to sample--election bias, so that it is appropriate to use the model in which Internet access and Internet shopping are estimated as a system. (22)

Results for Internet Shopping

The results for the Internet-shopping equation are shown in the first part of Table 2. Below, we will consider the coefficients for the tax variables, HOME TAX PRICE, TAXRATIO, and TAXBASE. First, however, we consider some of the results for the non-tax variables.

(i) The coefficient estimate of LOGINCOME is positive and statistically significant. Thus, an increase in income is associated with an increase in the probability that a consumer would engage in online shopping, all else equal.

(ii) The dummy variables for educational attainment (HIGHGRAD, COLLGRAD, and PROGRAD) have positive signs, and most are highly statistically significant. These results suggest that those with a high-school diploma or a Bachelor's degree are more likely to use the Internet for shopping than those with less than a high-school education. Those with a graduate or professional degree are also more likely to shop via the Internet than those with less than a high-school education, although the effect is smaller than the effect of a high-school diploma or a Bachelor's degree.

(iii) The coefficients for the age-related dummy variables (AGE15, AGE20, etc.) indicate that the probability of Internet purchases has an inverse-U-shaped pattern by age, and that this pattern is statistically significant. The probability of Internet purchases rises until consumers are in their thirties, and then declines. Our conjecture is that teenagers are less likely to engage in online shopping because they are less likely to have access to credit, and that the elderly may be less familiar and less comfortable with online shopping, even after controlling for other variables.

(iv) The coefficient on the dummy variable for the year 2001, D2001, is large, positive, and highly significant. This is not surprising, since Internet usage increased very substantially between 1997 and 2001. Under the specification reported here, in which the year enters only as a dummy variable, we constrain the coefficient on HOME TAX PRICE to be the same in both years. However, it is possible that the behavioral response to taxes may have changed between 1997 and 2001. We tested this by including an interaction term between the tax-rate variable and the year. We found a significant effect in the equation for Internet access. However, the effect in the equation for Internet shopping was extremely small. (The z value on the coefficient for the interaction term was -0.01.) The results from the regression with the interaction term are available upon request.

Our assessment of these results is that most of the coefficients for the non-tax variables in the Internet-shopping equations (shown in the top portion of Table 2) can be interpreted in reasonable ways. Many of the coefficients are highly significant, and the magnitudes are economically meaningful.

The results in Table 2 also lend support to our hypotheses regarding the influence of sales taxes on Internet shopping. In each of the specifications in Table 2, the coefficient for the tax price in the local county (HOME TAX PRICE) has the expected positive sign, and it is statistically significant at the five-percent level. These results indicate that, all else equal, a resident of a county with a higher sales-tax rate is substantially more likely to use the Internet for shopping than a resident of a county with a lower sales-tax rate. These results are consistent with the notion that Internet shopping is used, in part, as a mechanism for sales-tax evasion.

In each of the specifications in Table 2, the coefficient for TAXRATIO in the Inter net-shopping equation has the expected negative sign, and is also significant at the five-percent level. The negative coefficient on TAXRATIO indicates that a consumer whose county is adjacent to a lower-tax county is less likely to use the Internet for shopping than he or she would otherwise be, all else equal. This result is consistent with an interpretation that involves cross-border shopping: All else equal, the tax benefits from Internet shopping are reduced if a lower-tax county is nearby. Because the sales-tax burden can be reduced, simply by driving across the county line to shop, those who live near a lower-tax county have less of an incentive to shop via the Internet, all else equal.

Recall that TAXBASE is defined as the sales-tax revenue as a proportion of personal income in a state, normalized by the weighted-average sales-tax rate in the state. In each of the specifications in Table 2, the coefficient for TAXBASE has the expected positive sign, and it is significant at the one-percent level. This suggests that, even after we control for the tax rates themselves, the relative amount of sales-tax revenue collected has an effect on Internet-shopping behavior. Some states may have a high value of TAXBASE because the sales tax applies to more goods and services, while others may have a high value of TAXBASE because of more stringent sales-tax enforcement. In either case, a high value of TAXBASE means that (holding constant the tax rates), the sales tax reaches further. Thus, shoppers have a stronger incentive to shop via the Internet. The results for TAXBASE point in the same direction as the results for HOME TAX PRICE: Sales taxes encourage Internet shopping, either because the sales-tax rate itself is high or because the sales tax is widely applied.

Our interpretation of the results for HOME TAX PRICE and TAXRATIO is that shopping in the home county, shopping in an adjacent county, and shopping on the Internet are all substitutes. Goolsbee (2000) and Alm and Melnik (2005) estimate the effect of the tax price on Internet shopping; they also find that a consumer who faces a high tax price would be more likely to engage in online shopping. Thus, broadly speaking, our results are consistent with those of Goolsbee and Alm and Melnik.

When we recover the marginal effects associated with the probit coefficients, we can calculate the elasticity of Internet shopping with respect to HOME TAX PRICE. Note that an increase in the tax price in the home county will also lead to an increase in the ratio of the tax price in the home county to the lowest of the tax rates in adjacent counties. Thus, a complete calculation of the elasticity of Internet shopping with respect to HOME TAX PRICE would also include the indirect effect through a change in TAXRATIO. A one-percent increase in HOME TAX PRICE would increase TAXRATIO by slightly more than 0.01. This, in turn, would decrease the probability of using the Internet for shopping, all else equal. For our preferred equation (Specification (2)), if we combine the direct effect through HOME TAX PRICE with the indirect effect through TAXRATIO, the resulting tax-price elasticity is about 0.198. (23)

CONCLUSION

A previous literature has provided estimates of the effect of sales-tax rates on Internet shopping, and another literature has considered the effect of taxes on cross-border shopping. Each of these is important to state and local governments, which have experienced a decline in their ability to raise revenues through sales taxes. We integrate these two avenues of research, by analyzing empirically the determinants of Internet shopping in the United States, using data from the Current Population Survey for 1997 and 2001.

The data indicate whether consumers used the Internet for shopping, and we use this binary variable as our dependent variable. Our data set also includes the sales-tax rate in the consumer's local county, a measure of the sales-tax rates in adjacent counties, a measure of the breadth of the sales tax in the consumer's state, and a wide range of economic and demographic variables.

We estimate a system of equations, with a probit selection equation for Internet access, as well as a probit equation for Internet shopping. In all of our estimates, we find that the probability of Internet shopping is higher for those with higher incomes, all else equal. The probability of Internet shopping increases rapidly with age until consumers are in their thirties, and then decreases. Those with less than a high-school education are less likely to use the Internet for shopping than those with more education, all else equal. And, not surprisingly, we find a sharp increase in the propensity to use the Internet for shopping between 1997 and 2001.

The effects of the tax-rate variables on Internet shopping are always of the predicted sign, and they are statistically significant in all cases. In addition, the coefficients for the tax-rate variables are fairly robust across our different model specifications. The sales-tax rate in the consumer's own county has a positive effect on online shopping, all else equal. Our interpretation is that those who live in areas with high sales-tax rates are more likely to use the Internet for shopping, because sales-tax evasion on Internet sales is not difficult, and the benefit from evasion is greater when the sales-tax rate is higher. In addition, we find that Internet purchases are less likely for those who reside in a county that is adjacent to another county with a lower sales-tax rate, all else equal. We interpret this as evidence of cross-border shopping. When we account for the direct marginal effect of the home-county tax price, as well as the indirect effect on the relationship between the home-county tax rate and the lowest tax rate in an adjacent county, the tax-price elasticity is about 0.20. These elasticity estimates are closer to those of Alm and Melnik (2005) than to the higher estimates of Goolsbee (2000). However, this does not suggest to us that the behavioral responses are terribly small. We hope we have contributed to an emerging consensus that the effects are fairly substantial, although there remains controversy in the literature about the precise quantitative magnitude of the effects.

Goolsbee (2000) and Alm and Melnik (2005) focus exclusively on tax rates. However, holding constant the statutory sales-tax rate, a state could have high sales-tax revenues because the sales tax applies to more items, or because the sales tax is enforced more rigorously, or both. In either case, a more extensive sales tax would provide a larger incentive for Internet shopping, all else equal. In this paper, we introduce a new variable that is designed to capture the effect of differences in the sales-tax base. This new variable is sales-tax revenue as a proportion of personal income in a state, normalized by the weighted average of the sales-tax rates in the state. We find that, even after controlling for the tax rates themselves, the relative amount of tax revenue has an important effect on Internet shopping behavior: All else equal, Internet purchases are more likely for those who reside in a state in which sales-tax revenues are a larger fraction of personal income, after normalizing by sales-tax rates. We interpret this as additional evidence that consumers respond to the tax-related incentive to use the Internet for shopping.

Goolsbee (2000) and Alm and Melnik (2005) make very substantial contributions. We believe that our paper advances the profession's understanding beyond these earlier papers in three ways. First, we investigate cross-border shopping and sales-tax evasion through Internet shopping in a single framework. We consider both the effect of the tax rate in the consumer's local jurisdiction, and the effects of tax rates in adjacent jurisdictions.

Second, we define some of the tax-rate variables at the county level. We believe that this allows us to measure the tax-rate variables with greater precision. We also believe that this is crucial for our study of cross-border shopping. It would be very difficult to study cross-border shopping in a meaningful way, unless tax rates can be defined over reasonably small geographic areas.

Third, we introduce a new variable that is designed to capture differences in the sales-tax base. Previous papers in this literature control for tax rates, but not for the tax base. Our results indicate that tax rates and the tax base have similar effects: High sales-tax rates encourage Internet shopping, as do sales taxes that are widely applied.

The dependent variable in our study indicates whether the consumer uses the Internet for shopping. Unfortunately, we do not have information on the dollar value of the Internet shopping that is undertaken by the consumers in our sample. Consequently, any attempt to use our results to estimate the effects of the Internet on sales-tax revenue would have to be fairly speculative.

Our data are taken from a time period when the Internet was still in its infancy. As this new medium of commerce becomes more and more a regular feature of the economic landscape, we anticipate that the determinants of Internet shopping may change, possibly by a substantial amount. Thus, we do not consider our estimates to be the last word on the subject. Instead, we look forward to further research.

Acknowledgments

We are grateful to Jeff Biddle, Steven Haider, Yoon-Sang Kim, Larry Martin, Therese McGuire, Paul Menchik, Ed Outslay, Leslie Papke, Jean-Francois Wen, Jeff Wooldridge, and two anonymous referees for helpful comments and suggestions. We are also grateful to Adam Cogswell and Michael Gallagher for excellent research assistance. Any errors are our responsibility.

REFERENCES

Aim, James, and Mikhail I. Melnik. "Sales Taxes and the Decision to Purchase Online." Public Finance Review 33 No. 2 (March, 2005): 184-212.

Alm, James, and Mark Skidmore. "Why Do Tax and Expenditure Limitations Pass in State Elections?" Public Finance Review 27 No. 5 (September, 1999): 481-510.

Asplund, Marcus, Richard Friberg, and Fredrik Wilander. "Demand and Distance: Evidence on Cross-Border Shopping." Stockholm School of Economics. Mimeo, 2005.

Bruce, Donald, John Deskins, and William F. Fox. "Has Internet Access Taxation Affected Internet Use?" Public Finance Review 32 No. 2 (March, 2004): 131-47.

Bruce, Donald, and William Fox. "Cross-Border Shopping." In The Encyclopedia of Taxation and Tax Policy (2nd edition), edited by Joseph J. Cordes, Robert D. Ebel, and Jane G. Gravelle, 72-3. Washington, D.C.: The Urban Institute Press, 2005.

Engel, Charles, and John H. Rogers. "How Wide Is the Border?" American Economic Review 86 No. 5 (December, 1996): 1112-25.

Ferris, J. Stephen. "The Determinants of Cross Border Shopping: Implications for Tax Revenues and Institutional Change." National Tax Journal 53 No. 4, Part 1 (December, 2000): 801-24.

FitzGerald, John D. "The Distortionary Effects of Taxes on Trade in Border Areas: The Republic of Ireland-UK Border." In Tax Harmonization and Financial Liberalization in Europe, edited by Georg Winckler, 39-56. New York: St. Martin's Press, 1992.

Fox, William F., and Matthew N. Murray. "The Sales Tax and Electronic Commerce: So What's New?" National Tax Journal 50 No. 3 (September, 1997): 573-92.

Garrett, Thomas A., and Thomas L. Marsh. "The Revenue Impacts of Cross-Border Lottery Shopping in the Presence of Spatial Autocorrelation." Regional Science and Urban Economics 32 No. 4 (July, 2002): 501-19.

Goolsbee, Austan. "In a World Without Borders: The Impact of Taxes on Internet Commerce." Quarterly Journal of Economics 115 No. 2 (May, 2000): 561-76.

Gordon, Roger H., and Soren Bo Nielsen. "Tax Evasion in an Open Economy: Value-Added vs. Income Taxation." Journal of Public Economics 66 No. 2 (November, 1997): 173-97.

Ohsawa, Yoshiaki. "Cross-Border Shopping and Commodity Tax Competition Among Governments." Regional Science and Urban Economics 29 No. 1 (January, 1999): 33-51.

Van De Ven, Wynand P.M.M., and Bernard M.S. Van Praag. "The Demand for Deductibles in Private Health Insurance--A Probit Model with Sample Selection." Journal of Econometrics 17 No. 2 (November, 1981): 229-52.

Van Ham, Maarten, and Felix Buchel. "Unwilling or Unable? Spatial, Institutional and Socio-Economic Restriction on Females' Labor Market Access." Institute for the Study of Labor Discussion Paper No. 1034. Bonn, Germany: Institute for the Study of Labor, 2004.

Wooldridge, Jeffrey M. Econometric Analysis of Cross Section and Panel Data. Cambridge, MA: The MIT Press, 2002.

(1) It is also possible for consumers to buy from smugglers in order to reduce or eliminate their sales-tax payments. However, the data analyzed in this paper do not shed any light on smuggling.

(2) Typically, "evasion" refers to illegal methods of reducing tax liability, whereas "avoidance" refers to legal methods. When consumers reduce their taxes by Internet shopping, it is not immediately clear whether they should be considered to be engaging in evasion or avoidance: The consumers are violating laws, but the Supreme Court has ruled that those laws cannot be enforced. In any event, in this paper, we will usually refer to "evasion."

(3) Throughout this paper, we use "county" to refer to counties as well as to county equivalents, such as the parishes of Louisiana.

(4) For a concise summary of the literature on cross-border shopping, see Bruce and Fox (2005).

(5) Goolsbee deals with this problem in a variety of ways. For example, in one specification, he restricts his sample to observations from states with uniform sales-tax rates within the state.

(6) For example, in a study of Swedes who cross into Denmark to purchase alcohol, Asplund, Friberg, and Wilander (2005) find that distance from the border has a substantial negative effect on cross-border shopping. For further evidence on the importance of distance, see Engel and Rogers (1996).

(7) By defining the sales-tax rate at the county level, we also reduce the extent of measurement error in this variable. In some states and metropolitan areas, there is substantial variation in sales-tax rates. By defining the sales-tax rate at the level of the county, we capture most of this variation. The details of our tax-rate calculations are described below. One shortcoming of our approach is that information on county of residence is not available for a substantial fraction of the individuals in the CPS data set.

(8) The CPS data are available at http://www.census.gov/cps/cpsmain.htm. The CPS data for 1998 and 2000 also contain information on certain Internet activities, but the survey questions in those years are not useful for our purposes. Specifically, the Internet-shopping variable in 1998 and 2000 asks whether the consumer uses the Internet for paying bills or other commercial activities, as well as for shopping.

(9) We would like to know not just whether consumers use the Internet for shopping, but how much they spend when they shop online. Unfortunately, in this data set, a person who spends $10 per year via the Internet is coded the same as one who spends $1,000 per year.

(10) In Alaska, the state government does not have a general sales tax, but local governments impose their own sales taxes.

(11) These are Connecticut, Hawaii, Indiana, Kentucky, Maine, Maryland, Massachusetts, Michigan, Mississippi, Rhode Island, Vermont, Virginia, and West Virginia.

(12) These are Florida, Georgia, Nevada, North Carolina, Pennsylvania, South Carolina, Wisconsin, and Wyoming.

(13) It should be noted that an important part of the differences among the sales-tax bases in different states is due to differential taxation of business-to-business transactions.

(14) We use data for sales--tax revenue from the Census of Governments, at http: // www.census.gov/govs/www/estimate.html, and data for personal income from the Bureau of Economic Analysis, at http://www.bea.gov/bea/regional/spi/.

(15) LOGINCOME may also act as a proxy for the opportunity cost of time. It would be valuable to have data on wage rates in addition to the income data. Unfortunately, wage-rate data are unavailable for a substantial portion of the sample.

(16) The Bureau of Labor Statistics reports the Consumer Price Index (CPI) for 26 metropolitan areas annually (see http://www.bls.gov/cpi/home.htm). However, the data are normalized in such a way that comparisons of price levels cannot be made across the metropolitan areas. Thus, we use the percentage changes in the price level in the 26 metropolitan areas.

(17) See Bruce, Deskins, and Fox (2004) for an empirical study of the effects of tax variables and other influences on Internet access.

(18) The z-statistics reported in Table 2 are robust z-statistics, adjusted for cluster sampling within counties.

(19) It is not clear that any of our variables is an ideal instrument. As can be seen from Table 2, most of the variables have strongly significant effects on both Internet access and Internet shopping. We have excluded a wide variety of different combinations of variables from the Internet-shopping equations. The results are fairly insensitive to these exclusions.

(20) Our estimate of the effect of the sales-tax rate on Internet access is of the same sign as the estimate of Bruce, Deskins, and Fox (2004). By contrast, however, their estimate (based on data at the state level) is strongly statistically significant.

(21) In Specifications (1), (2), (3), and (5), the p-value for rho is 0.0000. The p-value is 0.0677 for Specification (4).

(22) However, for purposes of comparsion, we also ran a simple probit model using the selected sample of consumers with Internet access. (The results are available upon request.) For many of the variables in the Internet-shopping equation, the coefficient estimates and levels of significance for the selected sample are very similar to those reported in Table 2. This suggests that, even though the simple probit model suffers from a selection bias that is statistically significant, the quantitative magnitude of the bias is not necessarily very large. We also estimated the linear probability model. The results, which are available on request, are very similar to the probit results, in terms of the magnitude of the marginal effects, as well as the significance levels.

(23) We also estimated the model using HOME TAX PRICE as the only tax-related explanatory variable. (In other words, in this case, we eliminated TAXRATIO and TAXBASE from both the selection equation and the Internet-shopping equation.) The results from this specification are probably most directly comparable with the results of Goolsbee and Aim and Melnik. In this case, the tax-price elasticity is about 0.399.

Charles L. Ballard

Department of Economics, Michigan State University, East Lansing, Michigan 48824-1038

Jaimin Lee

Korea Transport Institute, Goyang-city, Gyeonggi-do 411-701, Korea TABLE 1 MEANS AND STANDARD DEVIATIONS FOR SELECTED VARIABLES (a)

Those with Internet Access

All Internet Non-Internet

Individuals Shoppers Shoppers No. of observat