Tax Sensitivity in Electronic Commerce*

I. Introduction

The need for empirical work on the tax sensitivity of online shoppers increases each year as sales from e-commerce continue to expand in the US and state and federal legislation takes form under the Streamlined Sales Tax Project (SSTP). The point of contention is a little-known, and widely evaded, tax called the use tax. This tax, which mimics a jurisdiction's sales tax, allows states to collect tax revenue on remote purchases of tangible goods if they are intended to be used in the consumer's home jurisdiction. The problem is that, due to Supreme Court cases such as National Bellas Hess vs. the Department of Revenue of the State of Illinois (1967) and Quill Corporation vs. North Dakota (1992), companies without a 'substantial presence' in a state cannot be required by that state to collect and remit sales taxes. This leaves the burden of remitting these taxes on the buyer; however, limited knowledge of the tax, along with its ease of evasion, has led to a high level of non-compliance, as documented in a study by the Washington State Department of Revenue (2003).

The SSTP is in the process of addressing the non-compliance issue by both working to simplify each state's tax code and lobbying Congress to change the requirement for compelling a firm to collect taxes within a state from the current requirement of 'physical nexus' tc· the weaker standard of 'economic presence'. This new wording would force businesses that do substantial business within a state to collect and remit sales taxes to that state even if they do not have a physical presence there. If this is successful, previously untaxed e-commerce and mail-order sales will become subject to the full sales tax liability. A movement towards a system in which Internet and mail-order sales are taxed would have two main effects: first, there would be a potentially large increase in previously uncollected sales tax revenues for each state; and second, there would be a shift of marginal shoppers away from online purchases back towards brick-and-mortar stores. In this study, I address the latter effect by estimating online consumers' tax sensitivities at the county level. I find that the current market structure has led to much lower tax sensitivities than have previously been estimated. I further find that only consumers living in high-tax locations and those with multiple years of online experience are significantly sensitive to sales tax rates when deciding whether to shop online. This is an important contribution since all previous studies, done at more aggregate levels, find significant, and sometimes very large, tax sensitivities for all consumers.

The first effect, regarding current and future losses in state sales tax revenue from use tax evasion, has been well documented in studies by Goolsbee and Zittrain (1999), Cline and Neubig (1999), McQuivey and DeMoulin (2000) and the US General Accounting Office (2000). Bruce and Fox (2001), using the concepts from these studies, estimate the annual revenue loss from business to consumer (B2C) commerce alone to be between $1.7 billion and $2.6 billion. Furthermore, each study predicts continued growth in state revenue losses in the future as the number of Internet users increases.

There has been considerably less work done on the second effect, regarding what the potential impact of collecting these taxes would be on Internet shopping. Though literature on the topic of e-commerce tax sensitivity is growing, empirical work is still hampered by a lack of reliable data. The breakthrough papers on the subject are Goolsbee (2000 and 2001), with his latter paper updating and expanding on his initial study. His research is based on the concept that online shopping allows all consumers to act as if they were living on a tax border. This draws from research done by authors such as Fox (1986) and Walsh and Jones (1988), who found that individuals living in border towns looked at tax differentials across borders when deciding where to make their purchases. Goolsbee's assertion is that ecommerce now also allows individuals living away from borders to make purchase decisions based on tax differentials, thanks to the increasing number of products sold over the Internet and the widespread evasion of use taxes. His samples are from 1997 and 1998 surveys,1 which had 110,000 and 85,000 household responses respectively.2 In the first study, he finds that imposing taxes on all Internet sales would reduce the number of online buyers by up to 24 per cent. His later work still finds significant tax sensitivity and separates the effects on the relatively non-sensitive new Internet users from those on the very sensitive experienced users.

Since Goolsbee's papers, three other studies have confirmed his finding that sales taxes play a major role in online shopping. Brynjolfsson and Smith (2001), examining price dispersions of online booksellers through shopbots,3 find that buyers in 1999 were twice as sensitive to changes in taxes and shipping costs as they were to changes in prices, even if the end price is the same in both cases. They also find that shoppers prefer to buy from brandname outlets online and are willing to pay a premium to do so. Contrary to this, Ellison and Ellison (2003), analysing shopbot data on the sale of memory modules from 2000 to 2001, find a more intuitive result - that online shoppers are actually more sensitive to price differences than to tax differences. They are also able to confirm Goolsbee's result by finding that there are indeed more online sales to states with high sales tax rates than to those with lower rates.4 Since this study focuses on a specialised product, usually with a high price and bought through a shopbot, the shoppers are thought to be more sophisticated than the average online shopper. This leads to large tax sensitivities, where a 1 per cent increase in a state's tax could lead to an 8 per cent increase in online sales to that state. The fact that both studies analyse data from Internet shopbots may indicate that they are picking up only the most advanced Internet shoppers, thus those most sensitive to tax rate changes. This means that the general policy implications from their studies are harder to apply to the population as a whole. It is also important to note that both studies focus on price dispersion online and only look at taxation as a secondary issue.

More recently, Alm and Melnik (2005) use the 2001 Current Population Survey to estimate tax sensitivity. They find a positive and significant general tax sensitivity, but only at a quarter of the marginal level found by Goolsbee. This study is the most similar to mine as we both use the 2001 Current Population Survey to analyse the relationship between sales tax rates and the propensity to shop online. My study differs from theirs, however, in several significant ways. First, they assign everyone in a state the same sales tax rate equal to the state rate plus the lowest local rate charged in the state, while I use a detailed tax data-set to attempt to match each person with their actual local sales tax rate. I then account for the possibility that consumers may shop outside their home town in an effort to avoid high sales taxes by alternatively assigning each person a tax rate equal to the lowest effective rate of any city within their county of residence. second, I analyse the impact that having multiple years of Internet experience has on an individual's tax sensitivity. This is important because it indicates that the current drop in tax sensitivity may simply reflect a large influx of new users who are not sensitive to changes in the tax rate. Third, I update the general model to include issues such as broadband use and online security which have not been fully considered in the past. Last, I outline the trend away from online-only firms, which have been the source of tax-free sales on the Internet, towards multi-channel firms that offer consumers more services and convenience but are required to collect sales taxes on all online sales.

My results indicate that under models similar to both Goolsbee (2000) and Alm and Melnik (2005), the general tax sensitivity of online shoppers in 2001 is extremely low and not significantly different from zero. Further analysis based on a splined tax rate indicates that only those facing the highest tax rates are still sensitive to rate changes. I also find that users who have been online for many years are significantly more sensitive to changes in the tax rate than are new adopters. Finally, including broadband and security variables in the model does not improve the significance of the tax coefficient even though both new variables are highly significant.

The outline of the paper is as follows. Section II introduces the data that are used in this study. Section III describes the model and econometric techniques I utilise to measure tax sensitivity, while Sections IV and V present the results from the multivariate probit regressions. Section VI outlines a possible explanation for my results and Section VII concludes.

II. Data

The primary data used in this paper are from the 'Computer and Internet Use' supplement in the 2001 Current Population Survey (CPS), which is jointly sponsored by the US Census Bureau and the Bureau of Labor Statistics. The supplement on computer and Internet usage has been given periodically since 1994. Each year it is compiled, the questions get more detailed and in-depth. The sample is drawn from 72,199 households, with a reference person answering for each person in the household. A complete description of the survey's design and methodology is available in US Census Bureau (2002).

III. Model

In order to isolate the effect of tax rates on an individual's decision to shop online, one must first understand the factors that motivate this decision. The probability that an individual will shop online should increase as the cost savings from shopping online increase:

(8) Cost savings from shopping online = tPQ - FC - Shipping + Z

where t is the tax rate, P is the price of the good,8 Q is quantity purchased and FC is fixed costs such as buying a computer, getting Internet access and acquiring the skills needed to use the Internet. The first term can be interpreted as the tax savings from shopping online. The last term picks up other non-pecuniary influences on a person's decision to shop online and could be thought of as the benefits minus the costs of online shopping. Benefits would include things such as convenience, increased selection, avoiding hassle from salespersons, less time needed to make a purchase and less driving. Costs would include security and confidentiality concerns, shipping time, inability to feel or try on a product, difficulty returning goods9 and no immediate customer service for questions or concerns about a product. Equation (8) is used to develop a regression specification and to help predict the signs on the variable coefficients.

The dependent variable in the model is whether an individual has 'used the Internet to purchase products or services' in 2001. The responses are conditional on having access to the Internet from home, work or school. The dependent variable is set to 1 if the person has purchased a product online and to 0 otherwise. Similarly to Aim and Melnik (2005), I use the selection equation described in Van de Ven and Van Pragg (1981) to correct for any bias that may result from the conditional nature of the dependent variable. This process estimates a selection equation and the conditional probit equation of interest simultaneously to allow for correlation in the error terms. The dependent variable in the selection equation equals 1 if the person has online access and equals 0 otherwise. The independent variables for this equation are similar to those used for online shopping in Table 1, but to aid in selection it also includes variables asking whether respondents have online access at work, whether they have online access at school and whether they own their own business. These variables are included due to their direct impact on online access and to better identify the selection equation.

In the general model, the main variable of interest is the tax rate, which is presented as i+t to conform to previous studies. It is expected to have a positive impact on Internet purchasing, as can be determined from equation (8). This implies that an increase in the sales tax rate would lead to a higher probability that local residents will shop online.

The explanatory variables and their summary statistics are shown in Table 1 and include household income and individual education level. Both the household income and education questions ask the respondent to select a range in which household income or education falls. These variables are therefore set up as categorical variables that correspond to given ranges. Given the poor response rate for household income in the CPS data, I use the hot-deck method of imputation to estimate around 13,000 missing values.10 Other explanatory variables include whether the respondent lives in a metropolitan area, age, race, number of children in the household, marital status, gender, experience using the computer and Internet, and regional dummies. Unlike the data used by Goolsbee (2000 and 2001), the CPS does not have a straightforward variable that accounts for computer and Internet experience. I therefore construct these variables by setting them equal to the sum of the number of activities other than shopping online that are done (i) via the computer or (ii) via the Internet. For example, if an individual responded that he only uses the Internet for one activity other than shopping online, the value of the Internet experience variable would be 1, while if he uses it for 10 other activities, the variable would equal 10. Therefore the variables are set up to be larger the more technical experience a person has accumulated.11 Table 1 shows that online buyers on average have higher household incomes and more education and are predominantly white compared with those who do not shop online. Online buyers are unexpectedly older than non-buyers, but this could be due to the large percentage of teens in the sample who do not have access to credit cards.

The expected results are determined by referring to equation (8) and are as follows: household income is expected to have a positive coefficient, since higher incomes should lead to larger quantities purchased (Q). Education is also expected to have a positive coefficient, since those who are educated tend to have more exposure to computers and a better understanding of technology. Age is likely to have a negative coefficient since younger people have more exposure to advanced technology; however, older individuals are more likely to have the means to shop online. Black, Hispanic and Native American should all have negative coefficients, reflecting larger confidentiality concerns, lack of minority-related content online and possible language barriers. A possible gender divide should make the coefficient for female be negative, though this divide is less documented. Marital status is expected to have a positive coefficient, reflecting a household network externality as spouses learn from one another. Computer and Internet experience are both expected to have positive coefficients since more experienced users are more likely to be comfortable shopping online. The dummy for metropolitan area status could have either a positive or a negative sign. A negative sign may reflect possible agglomeration of businesses nearby or increased selection from adjacent suburbs. A positive sign could reflect a more entrenched Internet infrastructure, a large amount of local content online, more users nearby to learn from or a broader acceptance of technology. For a full analysis of whether online activities are a substitute or complement for big cities and the impact of the level of local content, see Sinai and Waldfogel (2004).

IV. Baseline results

The initial results use a similar specification to Goolsbee (2001) to give a baseline regarding how the results from the CPS data-set in the updated year compare with previous findings.12 The baseline results are presented in column 1 of Table 2 and include dummies for three of four regions. Almost all of the results have the expected sign and are statistically significant at the 5 per cent level. These include the positive and significant coefficients on advanced education and higher income, and the negative and significant coefficients on Black and Hispanic, indicating the continued racial digital divide. The notable exception to the expected results is that the coefficient on the tax rate is found to be not statistically different from zero. Thus, an increase in the tax rate does not appear to have a significant effect on an individual's decision to shop online. This result is in stark contrast to what has been found in the previous literature and implies that there is much lower tax sensitivity than has been previously estimated by Goolsbee (2001). It also indicates that tax sensitivity is lower when tax rates are measured at the local level rather than at more aggregate levels as seen in previous studies.

The tax specification in column 1 attempts to match individuals' tax rates as accurately as possible with the prevailing tax rate in their cities of residence. However, some individuals may be willing to shop at locations outside of their city of residence in order to achieve some tax savings. To check for this possibility, I assign each individual a sales tax rate equal to the lowest sales tax rate of any city within the individual's county of residence. This allows me to now compare the decision to shop online with the decision to shop locally at the lowest-taxing location within the county. The results from using this alternative tax variable are presented in column 2 of Table 2. The coefficient on the tax rate is larger than under the baseline regression, and closer to that found by Aim and Melnik (2005), but is still not significantly different from zero.

Taken together, the base results presented in Table 2 indicate that the sensitivity to local sales tax rates documented by Goolsbee just three years earlier has largely disappeared and can now only be seen in the highest-taxrate jurisdictions. Therefore, only in these jurisdictions do the tax savings from shopping online exceed the costs associated with doing so. In the next sections, I will outline checks on the robustness of these results and will discuss the likely cause of my findings.

V. Further results

1. Computer and Internet experience

According to equation (8), individuals with more technical experience should exhibit higher sensitivity to changes in the tax rate. Goolsbee (2001) found that the more years a user has had access to a computer the more sensitive that user becomes to changes in the sales tax rate. To test for this, interaction terms are added to the preceding regressions to determine whether there are any interaction effects between computer or online experience and tax rate sensitivity. The results are presented in column 1 of Table 3 and indicate that neither the interaction between Internet experience and the tax rate nor the interaction between computer experience and the tax rate is significantly different from zero. However, as described by Ai and Norton (2003), the coefficient and t-statistic on the interaction term alone in non-linear models such as probit are not sufficient to determine the sign, magnitude or significance of the interaction effect. Therefore I use the method they describe, which checks the magnitude and statistical significance across all observations, to confirm my results. I find that this check does confirm the counter-intuitive results that indicate a lack of significance for the interaction variables across all observations.

One likely cause of this counter-intuitive result is that the experience variables generated above measure the variety of current online activities rather than past online experience. Since I do not have direct data on how long an individual has been online, I use a proxy that equals the percentage of people in each MSA that were online three years prior to the survey. This gives an estimated probability of whether any one individual in that MSA was online three years earlier. Column 2 of Table 3 indicates that the coefficient for this new online experience variable is positive and highly significant. Furthermore, column 3 of the table shows a potentially significant interaction effect between the new online experience variable and the tax rate. To check for significance, I again employ the Ai and Norton (2003) method; the results are presented in Table 4. It shows that the mean zstatistic for the interaction effect equals 2.67, indicating that the interaction effect is indeed significant across the majority of observations. Therefore, using this alternative measure of online experience, which is more similar to that used by Goolsbee (2001), I am able to achieve the expected interaction result.

The fact that users who have been online for a long period of time are more tax sensitive than new users implies that overall tax sensitivity may increase in the future as more users gain familiarity with the Internet. This is an important finding since it opens up the possibility that the recent drop in general tax sensitivity found by Aim and Melnik (2005) and by me may be transitory.16

2. Confidentiality and broadband

The next specification includes two new variables that have tremendous importance when it comes to online activity. The first of these variables is broadband/high-speed access to the Internet, which allows Internet users to access information and complete transactions more quickly than with dialup. Along with eliminating the need to 'sign on'17 in order to access the Internet, this increased speed streamlines the process of getting information on products and making purchases online. To pick up this growing vehicle of e-commerce, I include a dummy variable that equals 1 if the individual has high-speed access to the Internet and equals 0 otherwise.

The second variable I add deals with confidentiality concerns, which remain a major roadblock for many people when it comes to shopping online. Apprehension over giving credit card numbers or personal information over the Internet has slowed the growth of e-commerce and has become a central concern of companies doing business online. To account for this important aspect of Internet activity regarding confidentiality, I use a question in the CPS that asks, 'Compared to providing personal information over the telephone, how concerned are you about providing personal information over the Internet?'. This question is used to build a dummy variable equal to 1 if the individual is more concerned about giving information over the Internet than they are over the phone, and equal to 0 if they are equally or less concerned. The coefficient on this variable is expected to be negative, since it reflects a substantial cost for Internet shopping.

Table 5 presents the results from including these variables in the regression model. Column 1 shows that both variables' coefficients have the expected signs and are strongly significant. There is not a dramatic effect on the coefficient for tax rate with the inclusion of these variables, but the overall fit of the specification, based on the likelihood-ratio test, is better than in Table 2. Column 2 of Table 5 presents the results using a splined tax rate and exhibits a similar overall impact to that of column 3 in Table 2. This may indicate that while both confidentiality concerns and broadband access are important in a consumer's decision to shop online, they may vary in importance between high- and low-tax-rate areas.

To test whether broadband use or confidentiality concerns have a larger marginal effect in high-tax areas, I include tax rate interaction terms for both variables. The results from doing this are presented in column 3 of Table 5 and indicate a lack of significance for both interaction coefficients. This result is confirmed by using the Ai and Norton (2003) method discussed above. Thus, while both coefficients are of the expected sign and are significant when included individually, neither has a significant interactive effect with the tax rate.

3. Alternative baseline regression

Since my study uses a similar data-set and method to Aim and Melnik (2005), it is important to check whether the difference in our results is caused by the alternative tax rate data we employ. In their study, every individual in a state is assigned an identical tax rate equal to the state rate plus the lowest city tax rate found within the state. My tax variable, however, attempts to assign each individual the actual tax rate that exists in their local jurisdiction. This means that my data allow for variation in the tax rate faced by individuals within the same state.

I therefore run a regression that is similar to Alm and Melnik's but that incorporates my tax rate variable. Table 6 mimics the explanatory variables used in their study, which include categorical income variables in $10,000 increments, whether the respondent lives in a metropolitan area, the number of individuals living in the household and the respondent's race, education, sex and age.18 The table indicates that under this similar specification, the coefficient on the tax rate variable is still not significantly different from zero. This implies that the different results obtained by the two studies are driven, foremost, by the level of aggregation in tax assignments.

VI. Result interpretations and trend analysis

Similar to my results in Section V.1, Goolsbee (2001) found that new users were less sensitive to tax rates when shopping online than more experienced users. He also concluded that new users become more sensitive to tax rates over time as they become better educated on the tax code and its applications to e-commerce. Therefore, he predicts that during times when many new users are going online, the overall tax sensitivity will move down, and that during slower periods of Internet adoption, tax sensitivities will rise. This conclusion may partially explain why Alm and Melnik (2005) and my current study find weaker tax sensitivities than he found for individuals in 1998. An alternative, or additional, explanation for this erosion in tax sensitivity over time is an industry shift away from online-only firms to more dynamic multi-channel outlets which create a taxable nexus.

Ernst and Young (2001), in their Global Online Retailing report, look at the transformation of e-commerce up to 2001. They note that starting in 2000 there has been a substantial shift away from 'pure play' (online-only) firms towards firms with both a physical and an online component. This is largely due to the technology shake-up that saw many 'pure play' companies go out of business. Since 'pure play' firms, which were so popular in the 1990s, are less likely to have nexus in multiple states, they are able to offer more tax-free sales than their brick-and-mortar rivals. Ernst and Young find not only that the number of online-only firms has decreased significantly, but that the number of brick-and-mortar stores establishing an online presence has grown and will continue to grow. This means that there has not been a dramatic change in the number of goods sold online, but that there has been a drop in the number of sites that can sell their products tax-free. This finding is confirmed by Gallo and McAlister (2003), who find that between 2000 and 2001 there was a noticeable drop in the number of 'pure play' firms among the top 50 e-retailers.19 Further, they find that while in 2000 there were five 'pure play' firms among the top 10 e-retailers, in 2001 that number was down to two. This translates to a lower overall tax sensitivity for online sales.

A similar study done by McKinsey & Company and Salomon Smith Barney (2000)20 found that 'pure play' firms were losing money on most sales because the revenues were not covering the overhead costs of warehousing, maintaining a website and brand advertising. These findings led Joanna Barsh, a director at McKinsey & Company, to state: 'The notion of a "pure play" is turning out to be the wrong play'. The study found that a multi-channel approach that combined an online presence with physical storefronts was more appropriate. The changing online environment documented in these reports is one likely cause of my findings and predicts that future results may show similar or lower tax sensitivity for the majority of online consumers depending on the pace of Internet adoption by new users.

VII. Conclusion

Previous work has found that in the late 1990s, when the number of online-only stores was at its height, a small change in local sales tax rates would lead to a large jump in the percentage of online shoppers in that location. The implication was that if every business were suddenly required to collect sales taxes from online sales, the total level of e-commerce would drop significantly. This study finds that this result no longer holds true, as there has been a decrease in the tax sensitivity of online shoppers. In fact, I find that when analysed at the city or county level, the decision to shop online is no longer significantly affected by changes in local sales tax rates, except in areas where the tax rate is very high and for those with multiple years of online experience. Therefore if policymakers are able to induce firms to collect use taxes, either through legislation or through incentive programmes, the negative effects would be limited to high-tax jurisdictions and to those with high levels of technical experience. Thus online firms would see an overall decrease in sales, but it would be far less than what was previously concluded. 'Pure play' firms would be the most harmed, especially if they are currently only able to stay in business as a result of demand for their tax-free goods.

I find that the decrease in tax sensitivity is likely driven by two primary changes. The first is a shift away from 'pure play' firms towards multi-channel outlets that establish nexus, therefore subjecting more purchases to sales tax collection. The second change is a possible influx of new Internet users who are relatively less tax sensitive than long-time users. Other influences that may be driving my results could include brick-and-mortar stores responding to tax differentials by lowering prices or local stores adding additional in-store shopping features. Regardless of origin, the results indicate that the current lack of general sensitivity to tax rates should mitigate the negative consequences of proposed legislation that would force collection of sales taxes on every online sale. Future research in the area could examine how states and businesses would react if use tax collection were mandatory either in a subset of states or across the US. In each case, given tax sensitivities like the ones found in this study, it is of interest to consider firms' location decisions, along with local governments' decisions on what the new sales tax rates should be.