Commodity Tax Competition Among Governments:
Evidence from the United States
∗
Hendrik Vrijburg
†Thesis Groningen University
August 2006
Abstract
This thesis investigates commodity tax competition between US states and the spatial characteristics influencing this strategic interaction. Average Effective Tax Rates (AETRs) are employed instead of the nominal rates used until now in commod-ity tax competition research between US states. We find overwhelming evidence for strategic interaction, however, the impact of spatial characteristics remains unclear.
∗The author would like to thank Jenny Ligthart and Jan Jacobs for their support and advice during
the project.
†Department of Economics, Groningen University, The Netherlands, Phone: 0512538051, E-mail:
1
Introduction
After the District of Columbia implemented a 6 percent gasoline tax in August of 1980. The tax raised gasoline prices by an average of 8 cents a gallon, causing many residents to cross the border into Maryland, where prices were lower. In the first 26 days of the tax’s existence, gasoline sales fell 27 percent in the District, resulting in lower revenues than had been expected. The tax was repealed soon thereafter, replaced with a lower one-cent excise tax, which had a minimal impact on differentiating gasoline prices between the District and Maryland. (Rork, 2003)
Communication technology improved dramatically during the last decade. Consumers can easily observe prices in neighboring states by visiting the Internet website of a par-ticular firm. In addition buying and selling from the Internet, e-commerce, is increasingly becoming common practice. This accompanied by the more customary cross-border shop-ping1 practice, makes the commodity tax base increasingly vulnerable against changes in sales and excise taxes in neighboring states. Politicians fear this vulnerability of the sales tax base.2 Fleenor (1998) provides abundant evidence for cross-border shopping with
re-spect to cigarette taxes,3 Fleenor argues that strategic interaction will result in reduced
sales tax revenue, harming the provision of public goods.
Although various papers have empirically studied governments’ strategic tax setting behavior,4 economists have not reached agreement on the consequences of tax competi-tion. Besides the adverse effects on the provision of public goods, tax competition may potentially have some positive effects. First, public choice arguments state that govern-ments have no incentive, besides moral, to work efficient. Governgovern-ments therefore have a natural tendency to become too large from an efficiency perspective. Competitive pressure may reduce tax revenues and, consequently, depresses the size of the government.5 Second,
1The practice that a citizen chooses to shop in a neighboring jurisdiction. 2See Rork, 2003, page 775 for more US state level evidence.
3For example, cross-border shopping and smuggling accounted for 28%, 33%, 28% and 0% in 1997 for
Massachusetts, New York, Michigan and Kentucky, respectively.
tax competition may yield a positive distributional effect in favor of the poor, this because sales taxes are commonly seen as regressive.6
Furthermore, it is not all clear that tax competition will lead to a race to the bottom. Jurisdictions are heterogeneous entities, which might give rise to non-symmetric responses in tax setting. An important source of heterogeneity between jurisdictions seems spatial characteristics, that is, size, location with respect to neighbors, and geographical character-istics. Geographical characteristics includes border length and border curvature. Following the classical article of Kanbur and Keen (1993) and the recent work by Ohsawa (2003), spatial characteristics are identified as important theoretical determinants. However, until recently, spatial characteristics, apart from jurisdictional size, have not been used in the empirical tax competition literature.
The purpose of the thesis is twofold. First, it analyzes whether evidence can be found on commodity tax competition at the state level in the United States. Second it explores the effect of a state’s spatial characteristics—that is, its size, position with respect to its neighbors and relative border length—on the degree of tax competition. Tax reaction functions of state governments’ competing with other state governments are estimated. To this end, a panel data analysis is conducted for the period 1977–1995, covering 48 states. The focus will be on simple panel data techniques that are accessible via standard software. Advanced topics, such as spatial correlation in the disturbances, are beyond the scope of this thesis and are therefore left for further research. Nevertheless, Chapter 4 will briefly discuss the implications of spatial correlation. The main result is that reaction functions based on average effective tax rates (AETRs) are found to slope upwards in the United States, supporting tax competition theory.
The thesis contributes to the literature in the following ways. First, it poses and empirically tests various new hypotheses on the effect of jurisdiction’s spatial characteristics
6A tax is regressive if the propensity to consume out of wealth decreases with wealth, resulting in
on tax competition. Second, it employs state-level retail sales and excise tax data for the United States, which have not been used so far in this line of work.7 To this end, sales and excise tax data are incorporated into a single measure, that is, the (average) effective consumption tax rate. The benefit of this refinement is that our dependent variable—that is, the average effective consumption tax rate—contains more variability than the different underlying statutory (or nominal) tax rates.
The remainder of the thesis is organized as follows. Chapter 2 discusses the tax com-petition literature. Chapter 3 describes the theory and hypotheses. Chapter 4 discusses the research methodology underlying the empirical analysis. Chapter 5 provides a detailed description of the data used. Chapter 6 reports the empirical results. Finally, chapter 7 concludes.
2
The Literature
Our analysis builds on the fiscal competition literature in which the strategic interaction among governments with respect to fiscal policy is analyzed.8 Following Breuckner (2003),
the literature distinguishes spill-over models and resource-flow models. Both types of models generate reaction functions for each participating jurisdiction. In spillover models, the choice of an instrument (say, z) by the government of jurisdiction i directly influences jurisdiction j and therefore the choice of variable z by jurisdiction j. An example of this type of spill-over is environmental policy. If one jurisdiction decides to reduce CO2 emissions by imposing a high tax on such emissions, the neighboring jurisdictions will also benefit from the reduction in pollution. The government of the first jurisdiction might not internalize the total effects of the policy on the environment of both countries.
In resource-flow models, the choice of a particular variable by jurisdiction i is influenced by the choice of that variable by jurisdiction j because the underlying target is mobile with respect to locating in jurisdiction i or jurisdiction j. An example of such interaction is commodity tax competition, where the tax base, that is private consumption, is mobile because consumers can shop across the border.
Different categories of tax competition exist. First, capital tax competition, which focuses upon the effect of capital taxes on the location of firms or the location of savings deposits. Relocation of firms affects corporate tax revenues of governments. Key theoretical contributions are those by (1999) and Wilson and Janeba (2005).9 Second, articles focusing on tax burden shifting from mobile to immobile tax bases. Mobile tax bases can adjust to circumvent the tax, immobile bases are unable to do so which makes immobile bases superior from an efficiency perspective. This because imposing a tax on an immobile base induces less distortions compared with a tax levied on a mobile base. See for example Rork (2003) and Winner (2005). Finally, the cross-border shopping literature, which focuses upon commodity tax competition, which is the topic of this thesis. In the following, the relevant articles for this thesis will be discussed.
2.1
Theoretical Contributions
In their classic article, Kanbur and Keen (1993) provide the basic theoretical background for our econometric model. They model one-dimensional commodity tax competition between two jurisdictions on a Hotelling line.10 Important to note is that they assume the spatial
size of both jurisdictions to be equal, the jurisdictions differ in population and therefore in density.11 From this model they prove the following. First, the reaction function of
9Besley, Griffith, and Klemm (2001), Devereux, Lockwood, and Redoano (2002), Altshuler and
God-speed (2002) have taken an empirical approach.
10See the classical article by Hotelling (1929).
11The asymmetry in country size is key to the cross-border shopping result. If countries are of equal
each jurisdiction is an upward sloping function of the tax rate of the other jurisdiction. Second, the tax rate is increasing in jurisdiction size, that is, in a Nash equilibrium, the larger jurisdiction will have the highest tax rate.12 Third, ceteris paribus, an increase in the size of the neighbor, resulting in an upward pressure on the tax rate of the neighbor, leads to an increase in the tax rate of the home jurisdiction. Fourth, asymmetries between jurisdictions and the impact of this on strategic interaction are unambiguously harmful to both. In equilibrium both charge a lower tax rate than without strategic interaction. Increasing transportation costs, resulting in more autarky, and reducing the asymmetries between jurisdictions are strictly Pareto improving for both jurisdictions.
Ohsawa (1999) extends the basic model into a one-dimensional multi-jurisdiction set-ting. His model assesses the influence of the spatial configuration of jurisdictions on com-modity tax competition and yields the following results. First, location and size of juris-dictions create differences in the ”market power” of countries, which results in different equilibrium tax rates and revenues. Second, small jurisdictions set lower tax rates than larger jurisdictions. Third, equilibrium tax rates go down from the peripheral jurisdictions to the center. In the conclusion Ohsawa stresses that the spatial configuration of countries plays a critical role in the commodity tax game. The Kanbur and Keen (1993) results for a two-jurisdiction setting can be re-derived in a multi-jurisdiction setting, which proves it’s robustness with respect to the number of jurisdictions.
Nielsen (2001) presents a simple model with two jurisdictions, which produces the same results as Kanbur and Keen (1993). In addition, Nielsen extends his model to incorporate both transportation costs13 for the goods sold and the impact of border control (including
fines for illegal cross-border shopping) on the equilibrium tax rates and revenues. It is
12However, when the tax rate of the large jurisdiction becomes extremely small, the best response of the
small jurisdiction is to set a higher tax rate. See Kanbur and Keen (1993, p.881).
found that both transportation costs for goods and border inspection tend to raise the non-cooperative tax rates and tax revenues.
Ohsawa and Koshizuka (2003) investigate commodity tax competition between two jurisdictions in a more realistic two-dimensional setting. Ohsawa and Koshizuka (2003) prove that the results obtained by Kanbur and Keen (1993) and Ohsawa (1999) also hold in a two-dimensional setting. It is shown that a small jurisdiction sets a lower tax rate than a larger jurisdiction and, per capita revenue for a small jurisdiction is larger than for the larger jurisdiction. These results hold for any costumer distribution and any jurisdic-tional shape. As a third result, the paper proves that the more curved is a jurisdiction’s border, the lower is the tax rate, the tax difference between countries and the number of cross-border shoppers. This last result is an interesting extension of the one-dimensional models.14
2.2
Empirical Contributions
Evers, de Mooij and Vollebergh (2004) estimate fiscal reaction functions for European governments competing for revenues on diesel excises. They find evidence for significant tax competition in diesel excises. A 10% higher tax rate in neighboring countries induces a country to raise its own rate by around 2 to 3%. They prove that this impact is robust with respect to different specifications. In addition they find evidence that the magnitude of the tax rate matters for the degree of tax competition. Countries with a higher tax rate tend to compete more aggressive than countries with a lower tax rate. They don’t find evidence that country size impacts on the degree of tax competition. Rork (2003) investigates strategic interaction at the US state level with respect to different tax rates, namely, excise taxes, general sales taxes and corporate income taxes. The tax rates are
14An interesting hypothesis following from this proposition is that fiscal competition between the small
calculated by dividing the respective tax revenue by gross state product. Rork reports a significant positive slope for cigarette taxation, gasoline taxation, and corporate income taxation. Most notably, Rork observes a negative dependency between a personal income tax and general sales tax change of neighboring states. This would suggest that following a reduction of the tax rate of the neighbor, the home state rises its tax rate. There seems no intuitive explanation for such behavior.
Devereux, Lockwood and Redoano (2004) focus on horizontal and vertical indirect (i.e., excise) tax competition. Their theoretical framework states that when individual demand for the good is relatively price inelastic, and incentives for inter-state arbitrage (cross-border shopping or smuggling) are strong, the tax set in any state is likely to be strongly positively dependent on taxes set in neighboring states, but unresponsive to the various federal taxes. The contrary holds for goods with price-elastic demand. Devereux et al. use a panel data set of 48 US states over the period 1977–1997.15 They find that products with
inelastic demand, for example cigarettes, are not influenced by changes in federal taxes but do respond positively to neighboring states tax rate changes. Furthermore, products with a higher elasticity of demand, for example gasoline, are relatively more influenced by federal taxes and less by neighboring state tax changes.
Egger et al. (2005a), the key article underlying this thesis, is the only one that carries out a cross-country analysis. This article uses panel data for 22 OECD countries and 23 years. The key contribution of this article is the econometric refinement used. Egger et al. use a General Method of Moments (GMM) estimator to estimate the tax reaction functions and deal with spatial correlation.16 Besides, Egger et al. use an AETR as
dependent variable, as defined by Mendoza et al. (1994) instead of nominal rates as used in the literature so far. Their main results are the following. First, strong support for
15They collect state level and federal level unit taxes on cigarettes and gasoline from the World Tax
Database, this data-set is also used by Egger et al. (2005b), and is part of our data set.
the hypothesis that tax reaction functions are upward sloping. Second, relatively smaller jurisdictions tend to set lower tax rates. Finally, a negative impact of the neighboring state’s size on the tax rate of a particular jurisdiction, which confirms Kanbur and Keen’s (1993) theoretical results.
Egger et al. (2005b) investigate excise tax competition among US states and apply new panel data econometrics to the data.17 They estimate tax reaction functions based
on nominal excise tax rates for gasoline, cigarettes, beer, and wine, and find significant evidence for a positively sloped reaction function for all four excise tax rates.
Besides evidence from the United States, some studies have focused on Europe. Besley, Griffith and Klemm (2001) estimate tax reaction functions for OECD countries. They investigate labor taxes, corporate income taxes, property taxes, Value-Added Taxes (VAT) sales taxes, and excise duties, and find significant evidence for strategic interaction. In addition, they state that the reaction is larger for more mobile factors of production. Winner (2005) uses AETRs to estimate the impact of capital mobility and jurisdiction size on the taxation of factor incomes. Winner uses a data set of 23 OECD countries between 1965 and 2000. Most interestingly, like Egger et al. (2005a), Winner uses AETR as defined by Mendoza et al. (1994), as we use in this thesis. Winner observes a significantly negative relationship between capital tax burden and capital mobility, and a significantly positive impact of capital mobility on the labor tax burden. In addition, he finds significant effects of jurisdiction size on capital and labor tax burdens, and evidence for a shift of tax burden from capital to labor.
From the above survey it becomes clear that the commodity tax competition literature at the state level in the United States so far focuses on nominal tax rates. Therefore, our approach to investigate competition in effective tax rates at the level of United States
17Egger et al. (2005b) employ the spatial GMM estimation methods proposed by Kelejian and Prucha
states seems an important extension of the existing literature. The following chapter will discuss the theoretical model.
3
Hypotheses
This chapter will discuss the theoretical arguments underlying the econometric model out-lined in Chapter 4. Following the models of Kanbur and Keen (1993) and Ohsawa and Koshizuka (2003), we come up with the following hypotheses regarding tax competition at the state level in the US and the impact of spatial characteristics on the degree of competition. Each hypothesis will be discussed briefly.
Hypothesis 1 (Egger et al., 2005a) A jurisdiction’s domestic consumption tax rate is positively related to that of its neighbors.
This hypothesis follows readily from differentiated Bertrand competition.18 Commodities
in jurisdictions can be regarded as imperfect substitutes for each other; strategic interaction in policy setting results in upward sloping reaction functions. Both Kanbur and Keen (1993) and Ohsawa (1999) prove this proposition.
Hypothesis 2 (Egger et al., 2005a) Smaller jurisdictions tend to set lower equilibrium tax rates and get more than their size-proportional share in tax revenues.
The tax base of a jurisdiction can be separated into a mobile part (fraction of consumers willing to shop abroad) and an immobile part (fraction of consumers not willing to shop abroad).19 Large jurisdictions have a relative small immobile fraction, this results in a
relative large interest in the immobile fraction leading to less aggressive attitude towards neighboring states tax changes, because reacting is relative costly. This leads to upward pressure on the tax rate of a larger jurisdiction.
Hypothesis 3 (Egger et al., 2005a) The domestic consumption tax rate is increasing in the jurisdiction size of its foreign competitor.
Intuitively, larger neighbors have a relative higher tax rate; the positive slope implies an increase of the tax rate. In contrast, Nielsen (2001) argues, that small jurisdictions should compete aggressively against large neighbors to capture a mobile base relatively large compared to it’s domestic base. Because the large state has a high tax rate, a low rate of the small jurisdiction attracts a large number of cross-border shoppers. The first three hypotheses are formulated by Egger et al. (2005), who are inspired by Ohsawa’s (1999) work that focuses on one-dimensional cross border shopping. The following two hypotheses are put forward by Ohsawa (1999), who employ a multi-jurisdiction setting, and Ohsawa and Koshizuka (2003), who focus on two-dimensional cross-border shopping.
Hypothesis 4 (Ohsawa, 1999) For equally sized states, the consumption tax rates in peripheral states are significantly higher than those in states situated in the center of the United States.
Part of the border of peripheral jurisdictions is not exposed to competitive pressure; this exerts an upward pressure on the tax rate. For example, Florida has a large unexposed border on the Atlantic Ocean and the Mexican golf, which must have an impact on the expected amount of cross-border shopping.
Hypothesis 5 (Ohsawa and Koshizuka, 2003) The consumption tax rate of a partic-ular state decreases if its border becomes more curved.
Anecdotal evidence at the policy level20suggests that competition between two isolated neighboring jurisdictions is more aggressive than competition between larger clusters of jurisdictions.21 Common sense suggests competitive pressure to rise with the number of jurisdictions. Two opposing forces are at work namely, the aggressive force, the incentive of each jurisdiction to undercut its neighbors to attract a large sales tax base. And, the cooperative force, the incentive of each jurisdiction to stick to the current equilibrium to prevent a destructive fierce competition with its neighbors with which it has to cooperate on other policy issues. When the number of jurisdictions increases, the former becomes more dominant, this because it becomes more difficult to coordinate.
Several hypothesis can be stated based on the above reasoning. In this thesis we examine whether states with more neighbors have a lower tax rate. Intuitively, when there are more neighbors, there is a larger probability of a fiscal change of one of the neighbors in each year. So we expect such a state to change its fiscal policy more often. Second, it is expected that larger clusters of jurisdictions have a lower equilibrium tax rate because competitive pressure is higher. This obviously cannot be tested on US state data, that is on a large cluster of 48 states. This because we cannot vary the number of neighbouring jurisdictions. This brings us to the sixth and final hypothesis.
Hypothesis 6 Jurisdictions with a larger number of neighbors will feature more competi-tive pressure, which leads to a lower equilibrium tax rate.
The following chapter discusses the methodology used to test the above hypotheses.
20A case in point are the two jurisdictions (that is, the Federation and Republica Srpska) in Bosnia and
Herzegovina, which competed aggressively over sales and excise taxes.
21We were not able to come up with a satisfactory theoretical explanation for this phenomenon. It might
4
Empirical Methodology
This chapter will describe the research methodology used in answering the research ques-tions stated in the above chapter. Section 4.1 discusses the choice of an AETR, Section 4.2 describes the econometric model, while Section 4.3 briefly touches upon the issue of spatial correlation.
4.1
Average Effective Tax Rate (AETR)
The research methodology of this thesis follows Egger et al. (2005a) and Winner (2005) in that we are investigating AETRs. This AETR is constructed following the methodology of Mendoza et al. (1994). Mendoza derives the following expression for the AETR on consumption:
τit =
GT R + ET R
C + G − GW − GT R − ET R (1) where GTR is revenue from a general sales taxes on goods and services, ETR is revenue from excise taxes, C is private final consumption, G is government final consumption and GW is wages paid by the government. We will assume that government final consumption (G) equals the compensation of employees (GW ).
Several compelling reasons for the choice of the AETR can be given. First, theory states that the consumption decision by consumers is based upon the total consumption tax rate to be paid for goods. For example, suppose a consumer purchases one unit of a good subject to sales and excise taxes. The sales tax is paid on an excise-tax inclusive base. In formal terms,
where p denotes the tax exclusive sales price, τs is the nominal sales tax rate and τe is a
specific excise tax rate.22 The consumer thus pays tax on tax, the last term of equation (2), which is not measured by just taking nominal rates as in Egger et al. (2005b) and Besley, Griffith and Klemm (2001). Second, effective tax rates include all relevant components of a tax law, including exemptions, deductions, and the degree of enforcement of the tax law. Therefore, it allows us to compare states with very distinct tax structures. For example, Montana does not have a sales tax but does generate a significant amount of consumption tax revenue, reflecting excises on certain commodities. Third, effective tax rates show more variation than nominal tax rates. Table 1 lists some statistics of the the dataset for the United States used by Egger et al. (2005b). Column 1 reports the total number of changes made in the respective category by the states contained in the dataset for the period reported in column 4. Column 2 reports the minimum number of changes for a single state, column 3 reports the maximum number of changes for a single state. Column 5 reports the number of states included in the dataset for each category. Finally, column 6 reports the average frequency with which a particular state changes the respective tax. It becomes clear that on average, nominal tax rate changes occur infrequently. Only gasoline and cigarette excise taxes are changed more or less regularly. Table 2 summarizes the variability in each tax rate. Column 1 reports the mean of the normalized standard deviations for each tax rate category,23 column 2 reports the standard
deviation of this variability measure, to indicate how it varies between states for each tax category. As emphasized above, our dependent variable consists of sales and excise tax rates. Therefore, the AETR used in this thesis incorporates all the changes listed, in one effective consumption tax measure. The tax rate measure therefore exhibits much more
22Excise tax rates are often specific, but could also be ad valorem or be a combination of specific and
ad valorem rates.
23The standard deviation of the particular tax rate of state i is divided by the mean of the tax rate of
variation, as becomes clear from the last row of column 1 in Table 2. A dependent variable with more variability is more informative than a dependent variable with less variation.
Table 1: Nominal taxes rate changes
Total Min Max Years States Average (1) (2) (3) (4) (5) (6) Salesa 220 0 14 46-02 47 12 years Excises Gasoline 463 3 26 50-02 52 6 years Beer 206 1 10 33-02 52 17 years Cigarette 378 1 14 50-02 52 7 years Distilled Spirits 199 2 13 33-02 34 8 years Wine 108 1 8 50-02 34 16 years AETR 867 17 17 77-94 51 continuous
Notes: a 5 states do not have a sales tax.
Table 2: Standard deviations
Category Mean Norm Deviation Standard deviation
Sales 0.4510 0.2204 Wine 0.3784 0.3021 Gasoline 0.4968 0.1309 Beer 0.5870 0.6888 Distilled spirits 0.4141 0.1618 Cigarettes 0.4968 0.1309 AETR 1.8897 0.2717
cross reactions. Note that we do not agree with Rork (2003), who expects no responsiveness of the sales tax rate. Moreover, Rork states that it seems unlikely that residents would consistently do all their shopping in another state to take advantage of a difference in (broader) sales tax rates. Hence, a general sales tax may exhibit less responsiveness to changes in neighbor’s tax rates than an excise tax on a specific good. First of all, why would a consumer move to a neighboring state to only buy some gasoline or a pack of cigarettes when also grocery store products are cheaper due to a lower general sales tax? We think consumers decide the shopping location based on their evaluation of the price of a basket of products. Therefore, responsiveness to a general consumption tax might well be expected. We do agree with Rork that the responsiveness to excise taxes might well be higher because changes in excise taxes occur more often due to a lower revenue base that is at stake. In addition, the argument made by Devereux et al. (2004) seems relevant here. That is price inelastic products are more likely to exhibit fierce tax competition as opposed to price elastic products.
4.2
Econometric Model
Following the hypotheses discussed in chapter 3, the tax reaction function to be estimated is of the following form:
τt = β0 + β1W τt+ β2St+ β3W St+ β4LOC + β5BC + β6N
+X
l
γlXt+ t, (3)
where the βs and γs denote the parameters to be estimated, tis an i.i.d. error term. W τt,
St, WSt, LOC , BC and N are explained elow and Xt is a vector of control variables (also
which makes a logistic transformation necessary to obtain consistent estimates. Therefore, the dependent variable is defined as τit = ln(1−ttit
it). Note that the same transformation is
applied to the tax rate variable on the right-hand side given by τt.
Two different weighting matrices W are used in the regressions. The weighting matrix reflects the importance of each state in influencing tax setting behavior of all other states. The first matrix is constructed using the adjacency of various states, in this case a typical off-diagonal element of W is given by wjk = b1jk/
PK
k=1 1
bjk,j 6= k, while wjj = 0; bjk is a
border dummy which equals 1 when two states have a common border and zero otherwise. The second matrix is constructed using the squared distance between two states. The typical off-diagonal element of W is given by wjk = d12
jk /PK k=1 1 d2 jk , j 6= k and wjj = 0.
Here djk reflects the distance between two fixed points in differing states. We use the
largest city in each state as the fixed point.24
Several explanatory variables are included. To test Hypothesis 1 we include the weighted average effective tax rate of neighboring states (W τit). This weighted average is constructed
by multiplying the vector of AETRs (τit) with one of the weighting matrices (W ).
Fol-lowing Hypothesis 1, a positive slope is expected, which should be smaller than one to ensure the existence of a Nash equilibrium.25 Therefore, we expect 0 ≤ β1 ≤ 1. To test
Hypothesis 2 we include jurisdiction size (St). We experiment with different measures of
state size namely, population, surface, labor force and the logarithm of these variables.26
Following Hypothesis 2, a larger size implies a larger tax rate, therefore β2 ≥ 0. Hypothesis
3 is tested by including the weighted neighboring state size (W St). Following Hypothesis
24Note that for a given state j, the value of state h is smaller than the value for state g when d
jh≤ djg.
States located further away have a smaller impact upon the tax setting behavior of state j. Squaring the distances introduces a non-linearity increasing the weight of states located close to a particular state relative to states located further away. Note that all possible reference points to measure distance are equally disputable. Table A.3 reports the website used for calculating the distances.
25See Egger et al. (2005a), who correctly make this argument.
26Egger et al. (2005a) uses the logarithm of population. The impact of the level of population will
3, we expect β3 ≥ 0. A geographical location dummy (LOC) is included to test Hypothesis
4. Which provides information about the location of a particular state. Several alternative specifications are tested. First, a dummy variable indicating whether a particular state is located in the periphery, that is, it has a part of the border unexposed to domestic US competitive pressure,27 leading to an expectation of β
4 ≥ 0. Second, we introduce a
dummy variable indicating whether a state belongs to the East or the West of the United States.28 Border curvature (BC, defined as border length over state size) is employed to
test Hypothesis 5. More competitive pressure results in a lower tax rate, therefore the expectation is β5 ≤ 0. The number of neighboring jurisdictions (N ) is included to test
Hypothesis 6, which we expect to have a negative impact β6 ≤ 0.
In addition several control variables (X) are included to control for influences not cap-tured by Ohsawa’s model. The control variables can be divided into three broad categories: fiscal, political, and business cycle related variables. Each will be discussed in turn.
Fiscal Variables This category attempts to measure the effect of different fiscal policies. The first measure included is per capita public expenditure. As public expenditure rises, the state needs more revenue and this provides an incentive to raise commodity tax rates. Second, the lagged tax structure, defined as the ratio of indirect tax revenue over total tax revenue capturing the effect of the relative importance of indirect taxation; a higher ratio implies a higher tax rate.
Political Variables This category attempts to measure the effect of the political envi-ronment on the tax structure. Government political orientation is included. To explore
whether Republican states are less aggressive in tax setting than Democratic states which are expected to favor a larger size of the public sector.29
Business Cycle Variables Per capita income by state is included. This tends to mea-sure the impact of changes in the business cycle on the tax setting behavior. It might well be expected that in an economic-downturn state governments change their fiscal policy in a coherent manner. This, however, is not the strategic behavior of our interest. Second, the unemployment rate is used to capture automatic stabilizers. A higher unemployment rate leads to more social security expenditures and therefore has an upward pressure on the tax rate.
Following Breuckner (2003) the estimation of Equation (3) involves the following prob-lems:(i) endogeneity of the explanatory variable (W τt), (ii) correlation between the state
characteristic variables and the error term, and (iii) spatial error dependence. Endogene-ity of the dependent variable follows from the weighted dependent variable (W τit) being
an explanatory variable. We are investigating Nash equilibriums, in which each state is choosing the best reply given the behavior of all other states, this implies theoretically a simultaneous move game. The AETRs of different states are jointly determined and depend upon the random component of all other states. This correlation between the AETR of state j and the error term of all other states i 6= j means that OLS estimates are inconsistent. Second, correlation between the state characteristics and the error term might result from an explanatory variable not included in the model but correlated with the included explanatory variables. Spatial correlation will be discussed in Section 4.3.
29See Reed (2006) for recent evidence of the significant influence of political orientation in the United
We use the Arellano-Bond (1991) dynamic panel data estimator (DPD).30 This is a generalized method of moments (GMM) estimator, which corrects for higher-order auto-correlation by including lags of the dependent and explanatory variables. The model is first differenced to exclude the state fixed effects.31 The use of the DPD solves the endogeneity
problem by instrumenting the weighted neighboring states tax rates. As an instrument the nominal sales tax rate is chosen. It is important to recognize that the GMM method, as used in this thesis, is robust against the distribution of the dependent variable. The control variables will be lagged one period to exclude any endogeneity problems.
4.3
Spatial Correlation
Spatial error dependence is a well-known complication in the spatial economic literature. Spatial error dependence follows from the clustering pattern of high and low values of a variable in space.32 A spatial distribution displays spatial correlation when the probability
of a specific value occurring in a specific location depends on the value of neighboring locations. Consequently, Egger et al. (2005) and Florax and Nijkamp (2003)assume that the error term of 3 is given by it ≡ (I − ρW )µ where µ is an independently and identically
distributed vector of error terms and ρ measures spatial correlation. Solving the problem requires a spatial transformation which is, econometrically, beyond the scope of this thesis. Therefore we retain the assumption of it ∼ N (0, σ2I).
4.4
Specification Tests
This section briefly discusses three tests to address the following issues: endogeneity, spa-tial correlation, and instrument validity. First, we have to test whether endogeneity is
30See Baltagi (2001) for an excellent discussion of this estimator. This estimator is readily available in
Eviews 5.1.
31By first differencing any time-invariant variables drop out.
32Florax and Nijkamp (2003) give an excellent overview of the econometric problems surrounding spatial
pronounced in our benchmark specification, that is, whether the use of GMM is necessary. We apply the Hausman test. The underlying idea of the Hausman test is to compare the performance of the least squares estimator to an instrumental variables estimator. Several forms of the test, we will employ a simple alternative.33 The null-hypothesis to be tested
is that there is no correlation between the variable suspected of endogeneity (weighted neighboring tax rates) and the residuals against the alternative that there is correlation. We assume that among the set of explanatory variables are both an endogenous variable (weighted neighboring tax rates) and strictly exogenous variables (all other variables), which will also be incorporated in the instrument matrix. Assume that z1 is the
instru-ment for the weighted neighboring tax rates (W τt) taking x2 exogenous, so that the full
instrument matrix becomes Z = [z1 x2]. First regress,
W τt = Zπ + ν, (4)
where π is a parameter and νit ∼ N (0, σ2I). Second estimate Equation (3) by OLS and
retrieve the residuals from this regression µ. Under the assumption that W τt and the
exogenous variables are uncorrelated with ν in Equation (4), it follows that under the null-hypothesis of no correlation between W τt and µ, ρ = 0 in
µ = νρ + ω, (5)
because,
µ = (W τt− Zπ)ρ + ω, (6)
where ωit ∼ N (0, σ2I). However, when ρ is significantly different from zero, we must
conclude that W τt has a significant impact on µ and endogeneity is proved. This
esis can easily be tested by including the error term from Equation (4) in the regression Equation (3) and estimate this equation with OLS. When ρ turns out to be significantly different from zero, the assumption of an exogenous x1 is violated and GMM is required.
Second, spatial correlation will be discussed. Above we have made the assumption that the disturbances it∼ N (0, σ2I).34 It is informative to investigate whether our observations
are spatially correlated. When there is no evidence for the spatial correlation of our dependent variable, we need not to be worried about spatial correlation. To test for spatial correlation in the dependent variable Moran’s I test statistic and Geary’s c are commonly used. This thesis will use Moran’s I.35 High positive values of this test statistic point at spatial clustering of similar values over space; negative values indicate the joint occurrence of high and low values in nearby locations. Both violate the assumption of complete randomness. The test statistic is given by:
I = n S0 Pn i=1 Pn j=1wij(yi− y)(yj − y) Pn i=1(yi− y)2 , (7)
where n refers to the number of states, 48 in our case, y refers to the dependent variable, S0 is a normalized weighting factor defined as
P
i
P
jwij, and wij refers to the weight of
state j given state i, which follows from the weighting matrix discussed above.
Finally, the proposed instruments used in the GMM estimator must be valid, meaning that they are independent of unmeasured underlying variables and the error term. Often there is some degree of uncertainty about the validity of the proposed instruments. When the number of instruments is greater than the number of included endogenous variables, the
34As explained above, spatial correlation results from cross-sectional dependence. Following Anselin et al.
(forthcomming), the structure of the correlation or covariance between observations at different locations is derived from a specific ordering, determined by the relative position (distance, spatial arrangement) of the observations in geographic space.
35Following Florax and Nijkamp (2003), both Moran’s I and Geary’s c are special cases of the general
validity of the instruments can be tested via an over-identifying restrictions test. For this a Sargan test is used. Which tests the null hypothesis that the over-identifying restrictions are valid, the Sargan statistic is χ(n−k) distributed.36
5
Data Description
This chapter describes the data used to test the econometric specification discussed in Chapter 4. Section 5.2 discusses the states and time-period under consideration. The correlations are shortly addressed in Section 5.1. Section 5.3 discusses the dependent variable. Sections 5.4 and 5.5 analyze the explanatory variables and control variables, respectively. The data sources are listed in Table A.3 in the appendix.
5.1
Correlations
The correlations between the non-transformed variables are reported in Table A.1 and the correlations between the first differenced variables are reported in Table A.2. Only 11 cases with a correlation ≥ 0.75 are reported. The AETR is negatively correlated with the weighted average tax rates of neighbors (-0.29 and -0.25 for distance and adjacency weights respectively), and strongly positive correlated with the lagged tax structure and the nominal sales tax rate (0.75 and 0.80 respectively). Furthermore, the correlation be-tween population and the labor force is almost unity (0.998) which does not come as a surprise. As discussed in Chapter 4, the DPD estimator first differences the data, there-fore the relevant correlations for identifying multi-colliniarity are the correlations between the first differences of each variable. Not surprisingly, many of the correlations are much lower. The first difference of the AETR is relatively strongly positively correlated with the first difference of lagged tax structure (0.13), tax structure (0.49), sales rate (0.21),
and the weighted average neighboring tax rate (both weighted usign the distance matrix (0.29) and the adjacency matrix (0.31)). Again, population is highly correlated with the labor force (0.81). Changes in Gross State Product (GSP) per capita are almost perfectly correlated with changes in per capita public expenditure, reflecting the effect of the cycle on government spending. No potential multi-colliniarity problems can be identified.
5.2
Jurisdictions
Two states; Alaska and Hawaii and the District of Columbia (DC), have been left out of the data set. The two states are excluded for the compelling reason that they do not have US neighbors. Several reasons resulted in the exclusion of DC. First, DC has a very high tax rate but is extremely small in size.37 This is a violation of the theory, which predicts
small states to aggressively undercut their neighbors to attract a relative large flow of cross-border shoppers compared to their own size. Second, DC is a working district, people spend their money in the surrounding states, where the shopping malls are located. Because of this exceptional status, DC is regarded as an outlier and is excluded which leaves us with 48 states.
5.3
Dependent variable
The AETR is calculated using Equation (1) above. Consumption data is estimated from non-durable retail sales data using the Survey of Buying Power and an estimate for durables consumption.38 The latter is estimated using the ratio of US total retail sales data and US
total consumption data. Data on sales tax revenue and excise tax revenue at the US state level is used for the denominator.
37See Table A.5.
Figure 1: AETR and Nominal sales tax rate, 1977–1995 2.5 3 3.5 4 4.5 5 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 Year T a x R a te
AETR Nominal sales tax rate
Figure 1 shows the average of the 48 AETR’s in each year of the sample. The figure reveals that, apart from a decline from 1977 till 1981, the AETR has been remarkably stable.39 This might be regarded as evidence against a “race to the bottom”, which would
predict a constant decline.40 Also depicted in figure 1 is the Nominal sales tax rate. As can be inferred from figure 1 the nominal sales tax rate follows a different pattern but is in magnitude comparable to the AETR. Let us proceed by examining the differences in AETRs between the states. The bottom row of Table 3 reports the F-statistic resulting from the equality of means test, testing the null hypothesis of a single mean for all cross section units. This hypothesis is rejected. In the top part, the six states with the highest and in the bottom part the six states with the lowest AETR are depicted. Column 2 reports the mean AETR and column 3 the corresponding standard deviations.
The overall average effective tax rate is 4 per cent. The most remarkable feature is without doubt that all states with on average the lowest tax rates are those without a sales
Table 3: Means comparison
State Mean AETR St. Dev. Ranking Nominal p/c rev.
(1) (2) (3) (4) (5) (6) (7) Highest Level Washington 7.63 0.50 1 4 5.93 1 Nevada 6.38 0.77 2 8 5.33 2 Wisconsin 6.07 0.36 3 18 4.95 6 Mississippi 5.95 0.34 4 5 5.79 12 New Mexico 5.84 0.37 5 25 4.29 4 Connecticut 5.74 0.45 6 1 7.13 3 Lowest Level Virginia 2.89 0.34 43 40 3.24 43 Colorado 2.80 0.35 44 44 3.05 44 New Hampshire 1.67 0.48 45 48 0 45 Montana 1.57 0.25 46 48 - 46 Delaware 1.50 0.21 47 48 - 47 Oregon 0.86 0.13 48 48 - 48 Mean 4.02 0.41 - - 4.18
-Description df Value Probability Anova F-statistic (47,864) 227.9530 0.0000
Note: The complete Table is included in Appendix A.5.
It remains to be seen whether it is a wise policy to adopt a low AETR. Those states successful in the strategic interaction, would have the largest per capita revenue. Column 7 reports the ranking of the states in per capita sales tax revenue. The states with the highest sales tax rate are also amongst the states with the highest per capita sales tax revenue, violating the theory that undercutting is strategically beneficial. This survey suggests that states should–from revenue point of view– certainly impose a sales tax because this generally generates more revenues than no sales tax at all.41 The four states without a sales
tax may have opted for this choice because of historical reasons, not for strategic reasons. Governments in these states might suffer from a suboptimal fiscal structure imposed by history reflecting path dependency.
5.4
Explanatory Variables
This section tries to identify whether some of the explanatory variables exhibit interesting regularities in comparison with the dependent variable. It might be expected that those variables displaying a strong correlation with each other will also display a strong regu-larity. Investigating each variable in some more detail provides insights into distinctive characteristics of particular states.42 Table 4 lists the states with the six highest and the
six lowest values for each particular variable.43
First, the weighted average of tax rates of neighboring jurisdictions [Hypothesis 1], is calculated by multiplying the vector of AETRs with a weighting matrix. As noted above, two weighting matrices will be used, a contingency matrix and a distance matrix. Both
41This issue highlights an ongoing discussion in the states of Oregon, New Hampshire, Montana and
Delaware about imposing a sales tax. Politicians and economists in these states seem to recognize the revenue potential for imposing a sales tax. However, citizens in these states simply oppose a sales tax.
42Of course, correlations tell only part of the story, Chapter 6, therefore, considers a regression analysis
including all relevant variables.
are constructed by the author, in case of the distance matrix, the distance between the two largest cities in any pair of jurisdictions is used.
Hypothesis 1 shows that a jurisdiction will change its tax rate in the same direction as the change in the tax rate of the neighboring jurisdiction(s). In addition, it might be expected that a jurisdiction with neighbors with a relative high tax rate will have a tendency to choose a relative high tax rate of their own. This is however, not confirmed by Table 4, compare columns (1) and (3). The positive impact of state size [Hypothesis 2] is captured by the population size (in millions of persons) of a particular state. Alternatively, the state’s labor force (in millions of persons) or the spatial size (squared miles) are included as a robustness check, note that those are not included in Table 4. An extension uses the logarithm of the labor force to compare our results with those found by Egger et al. (2005a). Because of the high correlation between labor force and population, we will only report the features of the population data. Surprisingly, only Delaware and Montana are in line with the theory, so that no clear pattern emerges. Spatial size might well affect the strategic interaction in a larger jurisdiction, because citizens must travel on average a larger distance to reach the border. Distance increases transportation costs and therefore affects the likelihood of cross-border shopping negatively. However, when considering spatial size, not reported, there is also no clear pattern.
and a zero otherwise. Table 4 reports the location of the six high-tax and low-tax states. The first letter reported denotes East (E) or West(W), the second periphery [sea bordered (s), Mexico bordered (m), Canada bordered (a)] or in the center (c). Border curvature [Hypothesis 5] is measured by the ratio of border length to area size. Intuitively, this ratio is higher for jurisdictions who have a longer border for a given area size. As can be inferred from Table 4, no clear pattern emerges. The number of neighboring states [Hypothesis 6] is constructed by the author. A cross-section analysis, not reported, reveals no interesting pattern. This is also true for the variables concerning spatial size, which are also not reported.
5.5
Control Variables
Following Section 4.2 we will discus each of the different classes in turn.
Fiscal Variables The first fiscal variable, public expenditure per capita is depicted in Figure 2. The figure shows us that public expenditures per capita are consistently in-creasing, which demands more and more tax revenues and gives an upward pressure on tax rates. The summary statistics from Table 4 reveal that there are no states with pub-lic expenditures significantly different from the mean; only Washington is more that one standard deviation from the mean.
Figure 2: United States: Per capita public expenditure, 1977–1995 0 500 1000 1500 2000 2500 3000 3500 4000 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 Year D o ll a r s
Political Variables The variable political orientation is constructed by the author, the political party in charge in the state is inferred from the political party of the governor of the state. It is expected that Democrats (liberals) set higher tax rates than Republicans (conservatives). The final column of Table 4 lists the dominant political party in a state during the period 1977-1995. Dominant is defined as being in charge 75 per cent of the period. When the dominance is less than 75 per cent, neutral (N) is reported. Remarkably, most high tax states are Democratic and the low tax states are predominantly neutral.
Figure 3: United States: Tax structure, 1977–1995 0.47 0.48 0.49 0.5 0.51 0.52 0.53 0.54 0.55 0.56 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 Year R a ti o
Wisconsin has a significantly higher unemployment rate. Remarkably, Wisconsin also has a high tax rate.
6
Empirical Results
This chapter reports the results from the regression analysis. We make use of a balanced panel containing 48 states for the period 1977–1995. Section 6.1 reports the outcomes of the the Hausman test for the benchmark regression and the Moran I statistic for spatial correlation for the AETR. Section 6.2 reports the results of the benchmark case. Finally, Section 6.3 provides the robustness analysis.
6.1
Spatial Correlation and Endogeneity Tests
are significant of a 10% significance level, indicating the possible presence of an endogeneity problem.44
In addition, Table 5 reports the result of the Moran I test statistic for spatial correlation in the dependent variable, testing the null-hypothesis of spatial independence. The t-value reported indicates that Moran’s I is significantly different from the expected value. This gives strong evidence for the spatial clustering of high and low tax rates. This indicates that our results must be interpreted with caution, note that spatial correlation cannot be solved by using a appropriate weighting matrix.45 In subsequent research, spatial econometric techniques will be applied to improve the reliability of our results.
Table 5: Test statistics
endogeneity t-statistic p-value
Residuals 1.70 0.09
spatial correlation value p-value Moran I test statistic 0.9923
Expected value N −1−1 −0.0213 variance 0.0433 standard deviation 0.208
t-value 4.87 0.00
6.2
Benchmark Results
Table 6 reports the results for the benchmark regressions.46 In these regressions population
is used as a measure for size and the contingency matrix is used for weighting. The robustness section below will discuss variations in these measures.
44Note that both OLS and GMM are consistend estimators when there is no endogeneity problem.
However, OLS is the most efficient estimator. In the presence of endogeneity, OLS is inconsistend and GMM is consistend. Choosing GMM implies that we are sure that our estimates are consistend.
Table 6: Benchmark results
1 2 3 4 5
Lagged AETR 0.320 ∗ ∗∗ 0.358 ∗ ∗∗ 0.371 ∗ ∗∗ 0.367 ∗ ∗∗ 0.334 ∗ ∗∗
(0.037) (0.022) (0.021) (0.025) (0.027)
Hypothesis [1]:
Weigh. AETR Neighbors 0.58 ∗ ∗∗ 0.603 ∗ ∗∗ 0.588 ∗ ∗∗ 0.778 ∗ ∗∗ 0.623 ∗ ∗∗
(0.03) (0.026) (0.023) (0.065) 0.026) Interaction terma −49.33 ∗ ∗∗ (10.52) Hypothesis[2]: Population 13.13 6.453 4.08 −152.6 ∗ ∗∗ (15.7) (9.46) (8.42) (30.52) Log(Population) 0.310 ∗ ∗ (0.126) Hypothesis[3]: Population Neighbors −54.5 −53.4 ∗ ∗∗ −52.7 ∗ ∗∗ −60.9 ∗ ∗∗ (38.8) (16.1) (16.8) (20.3) Log(Population Neighbors) −0.5 ∗ ∗ (0.227) Hypothesis[4]: Peripheral 0.005 0.006 (0.012) (0.011) Hypothesis[5]: Border Curvature −0.67 (0.415) Hypothesis[6]: Number of Neighbors 0.002 (0.002) Control variables:
Lagged Tax Structure 0.871 ∗ ∗∗ 0.827 ∗ ∗∗ 0.782 ∗ ∗∗ 0.884 ∗ ∗∗ 0.836 ∗ ∗∗
(0.12) (0.106) (0.079) (0.099) (0.079)
Log(GSP per Capita) 0.068 ∗ ∗ 0.086 ∗ ∗∗ 0.078 ∗ ∗∗ 0.086 ∗ ∗∗ 0.075 ∗ ∗
(0.029) (0.019) (0.013) (0.017) (0.03) Statistics Adjusted R-square 0.48 0.49 0.50 0.48 0.48 Instrument Rank 48 48 48 49 48 J-statistic 41.02 43.71 43.81 43.7 42.97 Sargan 0.75 0.65 0.65 0.69 0.68
Notes: standard deviations in parenthesis below parameter estimates; ***/**/* denote significance at the 1, 5 or 10 per cent level, respectively
In columns (1)–(5), five regression results are reported. Each row reports the coefficient of the explanatory variable mentioned in the left column in the table. The explanatory variables are grouped as either testing a hypothesis or being a control variable. The most remarkable features of each will be discussed briefly. First the benchmark model, column (1), this is the model as outlined in methodological section. The weighted neighboring AETR, being endogenous, is instrumented. Note the reported J-statistic47 and the
instru-ment rank. The J-statistic is the value of the GMM objective function, evaluated at the estimated coefficients. This statistic is commonly used for a test of over-identifying restric-tions. This J-statistic equals the Sargan test statistic. Under the null hypothesis that the over-identifying restrictions are valid, the Sargan statistic is distributed χ(n−k), where n equals the number of instruments and k the number of estimated coefficients. The proba-bilities reported clearly state that the null-hypothesis is not rejected. Model (2) excludes the insignificant spatial characteristics of states to focus upon the impact of size, which has a significant robust positive sign in earlier empirical research, and location.48 Model (3)
mainly focuses on the impact of size. Model (4), includes an interaction term, to capture the notion that larger states might be less aggressive. The last model (5) reports the results from the impact for the growth rate of population and the growth rate of population in neighboring states, resembling the specification of Egger et al. (2005a).49 As expected,
the lagged AETR and the slope of the reaction function, that is the responsiveness of the tax rate of jurisdiction i with respect to a change in tax rate of jurisdiction j, are signifi-cantly positive and robust, with coefficients around 0.33 and 0.63 respectively, which is in accordance with earlier research on both nominal tax rates and AETRs.50 The slope of the
reaction function is smaller than one, as required for the existence of a Nash-equilibrium.
47See for a further explanation Hansen and Singleton (1982) and Newey and West (1987). 48See Egger et al.(2005a), Winner (2005).
49Note that Egger et al. (2005a) do not include the lagged AETR, this is due to a different estimator
used.
The coefficient is slightly smaller than the values found by Egger et al. (2005a). The coefficient can be interpreted as a ‘corrected elasticity’, reflecting the log transformation of the effective tax rate taken on both sides of the equation.51 Also the lagged tax structure and per capita gross state product are consistently significant with estimates of 0.85 and 0.08, respectively. The adjusted R-square is consistently around 0.50.
Table 7 reports the same regression while using the distance matrix instead of the contingency matrix. The result display a large degree of similarity. Only the signifi-cant negative impact of neighboring states’ size disappears. The coefficients of the lagged AETR, the weighted AETR of neighbors and the lagged tax structure increase slightly. Interestingly, geographical location turns out be significant at the 10 % and 5 % level in the second and third model. The adjusted R-square is slightly increased and the null hy-pothesis of no over-identifying restrictions is not rejected. Note that the control variables; unemployment rate, per capita public expenditure, and the political orientation index are not included, those are discussed in the robustness analysis section.
6.3
Robustness Analysis
Table 8 reports the results of the robustness analysis. Model (1) reports the benchmark model using the labor force to measure size instead of population. Model (2) uses spatial size to measure the impact of size. Both models also include the unemployment rate which has a significant positive influence on the tax rate. However, this measure is not robust against changes in the model. The third model uses the logarithm of the labor force to test the robustness of benchmark model (5). Model (4) includes in addition to the logarithm of
51Assuming tit
1−tit = tit; τt = ln(tt) and W τt = W ln(tt). An elasticity is defined as
∂Y ∂X
X
Y. First
differentiating Equation (3) with respect to W τjt gives ∂W τ∂τit
jt = β1. Recognize that ∂τit ∂W τjt = ∂ln(tit) ∂W ln(tjt) = ∂tit ∂tjt tjt
tit which is the elasticity of the corrected AETR of state i with respect to a change in the corrected
AETR of state j. By related arguments it is easy to show that βi =t1
it
∂tit
∂Hi for i = 2, 3, 4, 5, 6, which reflects
a proportional change. Hi resembles the remaining explanatory variables in Equation (3), note that those
Table 7: Distance matrix
1 2 3 4 5
Lagged AETR 0.406 ∗ ∗∗ 0.407 ∗ ∗∗ 0.394 ∗ ∗∗ 0.372 ∗ ∗∗ 0.429 ∗ ∗∗
(0.024) (0.02) (0.018) (0.027) (0.021)
Hypothesis [1]:
Weigh. AETR Neighbors 0.697 ∗ ∗∗ 0.669 ∗ ∗∗ 0.658 ∗ ∗∗ 0.969 ∗ ∗∗ 0.625 ∗ ∗∗
(0.046) (0.038) (0.034) (0.076) (0.037) Interaction terma −0.062 ∗ ∗∗ (0.001) Hypothesis[2]: Population −12.4 −0.004 −0.181 ∗ ∗∗ (0.016) (0.009) (0.046) Log(Population) 0.116 (0.072) Hypothesis[3]:
Population Neighbors −3.3 0.016 6.00E − 06 −0.186∗
(0.055) (0.034) (0.030) (0.104) Log(Population Neighbors) Hypothesis[4]: Peripheral 0.013 0.014∗ 0.015 ∗ ∗ 0.013 (0.010) (0.008) (0.007) (0.009) Hypothesis[5]: Border Curvature 0.066 (0.354) Hypothesis[6]: Number of Neighbors −0.001 (0.003) Control variables:
Lagged Tax Structure 0.723 ∗ ∗∗ 0.736 ∗ ∗∗ 0.805 ∗ ∗∗ 0.728 ∗ ∗∗ 0.702 ∗ ∗∗
(0.132) (0.110) (0.090) (0.086) (0.076)
Log(GSP per Capita) 0.0727 ∗ ∗ 0.64 ∗ ∗ 0.065 ∗ ∗∗ 0.049 ∗ ∗ 0.052 ∗ ∗∗
(0.032) (0.023) (0.009) (0.021) (0.013) Statistics Adjusted R-square 0.52 0.53 0.53 0.53 0.54 Instrument Rank 48 48 48 48 48 J-statistic 40.57 40.69 42.38 39.12 41.8 Sargan 0.77 0.76 0.70 0.82 0.72
labor, the political orientation index and substitutes the growth rate in per capita public expenditure for the growth rate of per capita GSP. Model (5) includes border length to test for Hypothesis 5. Model (6) and (7) examine different measures of location.52
The Sargan over-identifying test is never rejected. Model (1) shows that the labor force does not have a significant impact, which resembles the result for population size. The log of the size of the region turns out not to be robust against a change in the measure of size, see models (3) and (4). None of the location variables are robustly significant. Border length does not perform better than border curvature in the test of Hypothesis 5. Moreover, the lagged dependent variable and the positive slope of the reaction function are remarkably stable over different specifications, as are the control variables lagged tax structure and GSP per capita. A one per cent rise in the share of indirect taxes leads to a rise in the AETR of approximately 0.8 per cent rise in the AETR. Public expenditure per capita is a good substitute for GSP per capita. A one per cent rise in per capita GSP leads to a rise in AETR between [0.049–0.089] per cent. Political orientation turns out not to be significant.53
The interaction term has a robust negative sign.54 The robust finding of a negative sign for the interaction term highlights an interesting extension for further research. Although not yet modeled theoretically, it seems intuitive that larger states are less aggressive (as indicated by the negative sign), although the coefficients are far too large to be intuitive. The adjusted R-square is always around 0.50. The robust upward slope of the reaction function replicates earlier research conducted with nominal tax rates.55 Overall, there is
abundant support for the presence of strategic interaction among state governments, which confirms Hypothesis 1.
52Not reported, but also not significant are the following location dummies: state borders Mexico, and
state borders Canada.
53Political orientation is also not significant in the regressions not reported in table 8. 54Note however that the sign changes when labor is included instead of population.
55See Besley, Griffith and Klemm (2001) for OECD countries and Devereux, Lockwood and Redoano
Table 8: Robustness Analysis
1 2 3 4 5 6 7
Lagged AETR 0.316*** 0.362*** 0.347*** 0.396*** 0.359*** 0.382*** 0.367*** (0.049) (0.04) (0.028) (0.03) (0.019) (0.025) (0.02) Weigh. AETR Neighbors 0.627*** 0.608*** 0.594*** 0.560*** 0.581*** 0.587*** 0.602***
(0.038) (0.03) (0.028) (0.029) (0.024) (0.027) (0.024) Population 0.001 0.007 (0.01) (0.009) Labor Force 19.2 (32.4) Log(Labor Force) 0.127 0.089 (0.09) (0.091)
Spatial Size 4.00E-07
(4.00E-07)
Population Neighbors -0.043** -0.048**
(0.02) (0.018) Labor Force Neighbors -97.9
(77.3)
Log(Labor Force Neighbors) -0.273* -0.182
(0.143) (0.161)
Spatial Size Neighbors -0.0004
(0.001) Peripheral 0.005 (0.014) East 0.004 (0.01) Border Ocean 0.009 (0.01) Border Length 0.008** (0.004) Number of Neighbors 0.003 (0.003)
Lagged Tax Structure 0.871*** 0.856*** 0.81*** 0.760*** 0.842*** 0.732*** 0.816*** (0.138) (0.128) (0.075) (0.089) (0.069) (0.085) (0.095) Log(GSP per Capita) 0.089* 0.060*** 0.081*** 0.063*** 0.081*** 0.085***
(0.047) (0.016) (0.031) (0.01) (0.024) (0.013)
Log(Public Expenditure per Capita) 0.054*
(0.029) Unemployment Rate 0.003*** 0.004*** (0.001) (0.001) Political Orientation -0.003 (0.011) Statistics Adjusted R-square 0.45 0.46 0.48 0.50 0.49 0.5 0.49 Instrument Rank 48 48 48 47 48 48 48 J-statistic 37.7 40.94 43.94 41.3 44.02 43.69 43.58 Sargan 0.86 0.75 0.64 0.71 0.64 0.65 0.65
The coefficient of population size [Hypothesis 2] is significantly positive in one out of 17 reported regressions, therefore this thesis does not find a positive impact of size on the AETR. Furthermore, less than half of the regressions report a significant impact of neighboring state size, all with a negative sign. Therefore, Hypothesis 3 has a weak empirical foundation. This finding is in line with Egger et al. (2005a, 2005b). Egger et al. (2005a) obtain a significant positive impact of the logrithm of the labor force, whereas Egger et al. (2005b) do not find a significant impact of labor force in all regressions.
Hypotheses 4, 5 and 6 are not confirmed by the data. The significantly positively sign of border length is not robust against any changes in the specification. The same holds for the significance of a peripheral location in model 3 in the distance specification. Hypothesis 5 is potentially too far-fetched, a theoretical impact does not necessarily directly translate into an empirical impact. Especially in the United States, where borders are not extremely curved, the impact may well be negligible. Policymakers may not be concerned with the relative length of their border on cross-border shopping considerations. Hypothesis 6 can not be confirmed in the regression analysis, which states that, ceteris paribus, a larger number of neighboring states do not imply a lower AETR. This can be explained by reasoning that the tax rate of a particular state does not depend on the number of competitors, but that it depends on the weighted average of the tax rates of the neighbors. Another argument is that the strategic interaction cannot be considered as local events between two states. Theory would imply strategic competitive changes of tax rates to spread through the whole United States, ultimately resulting in a new Nash equilibrium.
7
Conclusion
This thesis finds evidence for commodity tax competition between US states, which con-firms earlier research on this topic. For this purpose, we make use of a panel data set of 48 states for the years 1977–1995. The key contribution is the use of effective consump-tion tax rates, which are theoretically superior to the statutory (or nominal) tax rates, to estimate tax reaction functions. In addition, we investigate the impact of size and geo-graphical characteristics on the strategic behavior of states. For this purpose we estimate tax reaction functions employing the Arellano-Bond dynamic panel data estimator.
We find a robust positive slope of the reaction function with respect to tax rates of neighboring states. However, state size and geographical characteristics are not found to be significant, and there is some indication of a negative impact of neighboring state size. The relative constant pattern of effective consumption tax rates seems to contradict the fear of politicians for a race to the bottom of tax rates.
Appendix
Figure A.1: Nominal excise rates, 1977–1995
0 5 10 15 20 25 30 35 1977 1978 1979 1980 1981 1982 1983 1984 1985 1986 1987 1988 1989 1990 1991 1992 1993 1994 1995 Year P e rc e n ta g e
Figure A.2: Map USA
Table A.4: Descriptive statistics
Variable Mean Maximum Minimum Standard deviation
AETR, transformed −3.24 −2.39 −4.95 0.41
Weighted AETR neighboring states (n) −3.26 −2.81 −4.51 0.23
Weighted AETR neighboring states (d) −3.22 −2.69 −4.36 0.22
Population (Millions of persons) 4.98 31.70 0.41 5.19
Labor force (Millions of persons) 2.48 15.43 0.21 2.55
Spatial size (square miles) 61643.00 261796.00 1045.00 46341.00
Population neighbors (n) 4.89 10.48 0.66 2.43 Population neighbors (d) 4.92 17.04 1.57 2.48 Labor force (n) 2.40 5.00 0.30 1.17 Labor force (d) 2.41 8.12 0.71 1.16 Spatial size (n) 56.29 107.69 8.61 26.41 Spatial size (d) 67.18 102.65 45.38 14.92 Border curvature 0.02 0.10 0.01 0.02 Border length 0.90 1.61 0.10 0.41 Number of neighbors 4.45 8.00 1.00 1.60 Political orientation 0.59 2.00 0.00 0.51 Tax structure 0.51 0.87 0.11 0.15 Per capita GDP 14649.00 31045.00 5232.00 5086.00
Per capita Public expenditure 1.87 4.39 0.58 0.80
Table A.5: State characteristics
State Abb. AETR Population Tax structure pcap GDP
mean st. dev. mean st.dev mean st. dev. mean st. dev.
References
Alesina, A. and Perotti, R. (1998). The political economy of adjustments. Brookings Papers on Economic Activity, pages 197–248.
Altshuler, R. and Goodspeed, T. J. (2002). Follow the leader? evidence on european and u.s. tax competition. Working Paper Rutgers University.
Anselin, L Florax, R. and Rey, S., editors (forthcomming). Advances in spatial economet-rics. Berlin: Springer-Verlag.
Baltagi, B., editor (1995). Econometric Analysis of Panel Data. New York: John Wiley & Sons Ltd.
Besley, T. Griffith, R. and Klemm, A. (2001). Fiscal reaction functions. Working Paper London School of Economics.
Brueckner, J. K. (2003). Strategic interaction among governments: An overview of theo-retical studies. International Regional Science Review, 26:175–188.
Devereux, M. P., Lockwood, B., and Redoano, M. (2002). Do countries compete over corporate tax rates? CEPR Discussion Paper, No. 3400.
Egger, P., Pfaffermayr, M., and Winner, H. (2005a). Commodity taxation in a ’linear’ world: A spatial panel data approach. Regional Science and Urban Economics, 35:527– 541.
