In search of yardstick competition: a spatial analysis of Italian municipality property tax setting

Journal of Urban Economics 54 (2003) 199–217 www.elsevier.com/locate/jue

In search of yardstick competition: a spatial analysis of Italian municipality property tax setting Massimo Bordignon,a Floriana Cerniglia,b,∗ and Federico Revelli c a Faculty of Economics, Catholic University, Largo Gemelli 1, 20123 Milano, Italy b Faculty of Political Sciences, Catholic University, Largo Gemelli 1, 20123 Milano, Italy c Department of Economics, University of Torino, Via Po 53, 10124 Torino, Italy

Received 18 September 2002; revised 18 March 2003

Abstract This paper uses Italian local government data to test for fiscal interaction arising from yardstick competition. To discriminate yardstick competition from competing theories of strategic interaction, we account for the incentives and constraints generated by the electoral system, in particular for the presence of term limits and the size of the majorities supporting the mayors. Estimation of a local property tax setting equation uncovers positive spatial auto-correlation in local tax rates of jurisdictions where the mayors run for re-election in uncertain contests, while interaction is absent where either mayors face a term limit or are backed by large majorities.  2003 Elsevier Inc. All rights reserved. JEL classification: C51; D72; H71 Keywords: Local property tax; Yardstick competition; Spatial auto-correlation

1. Introduction One of the most interesting recent developments in empirical local public finance is the growing evidence of strategic interaction in tax and public spending determination.1 * Corresponding author.

E-mail addresses: [email protected] (M. Bordignon), [email protected] (F. Cerniglia), [email protected] (F. Revelli). 1 Ladd [26], Case [21], Case et al. [22], Kelejian and Robinson [25], Murdoch et al. [27], Besley and Case [5], Shroder [34], Bivand and Szymanski [10,11], Buettner [20], Heyndels and Vuchelen [24], Brueckner [17], Brueckner and Saavedra [19], Figlio et al. [23], Brett and Pinkse [16], Saavedra [32], Revelli [29–31], and Sole0094-1190/$ – see front matter  2003 Elsevier Inc. All rights reserved. doi:10.1016/S0094-1190(03)00062-7

200

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

Unfortunately, such empirical evidence might be consistent with different competing theories of local government behavior. For example, the fact that local tax rates tend to be correlated across neighboring jurisdictions has been variably interpreted as arising from public expenditure spill-overs, tax competition or yardstick competition.2 Worse than that, these theories may have very different, indeed contrasting, normative implications, so that the observer is left without solid grounds to judge this evidence. As usual in empirical analyses, the above indeterminacy arises from one (or both) of the following reasons. Either alternative theories may be observationally equivalent, or the available data set may not be rich enough to allow discrimination among their different predictions. Consequently, solving these problems requires one either to carefully re-examine the implications of the theories to be tested, or to build a better data set. In this paper, we follow both strategies, looking for evidence of yardstick competition across local governments in Italy. We focus on this theory for a number of reasons. First, as forcefully argued by Salmon [33], yardstick competition theory may provide new and powerful insights on the structure of government in democratic countries. If it were really the case that citizens made comparative performance evaluation across governments in order to understand the quality or the competence of their politicians, a new approach to key issues of fiscal federalism would follow. For example, if this evaluation is beneficial to citizens, one would then want to organize, say, the allocation of functions and resources to local governments so as to maximize this behavior, possibly in contrast with the suggestions of traditional theory (as summarised, for example, by Oates [28] or Wildasin [38]). A second motivation is that the interest in these issues is not purely academic. Indeed, several European countries—where one would expect other forms of competition across governments, such as the celebrated Tiebout [37] ‘voting by feet’ to be less relevant than, say, in the United States—are actually involved in deeply reforming their local government structure.3 Knowing if yardstick competition is a true phenomenon may then help them in designing a better institutional framework. In order to investigate the presence of yardstick competition, one has to find specific empirical implications of the theory. As we argue in Section 2, the key insight of yardstick competition theory is that tax setting behavior and the features of the electoral system should be considered at once in the empirical analysis. Consequently, we test the theory on a sample of Lombardian municipalities, by estimating a property tax setting equation that fully exploits both the institutional structure of the Italian system of local government and the techniques developed within the spatial econometrics literature (Anselin [1], Anselin and Florax [3], Anselin et al. [2]). As argued below, careful spatial econometric testing,

Olle [35], all found evidence of interdependence in local government fiscal choices, using a variety of data sets and estimation techniques. 2 See Wilson [39] and Brueckner [18] for recent reviews. 3 In Italy, for example, the idea that decentralization and the ensuing competition across local governments may enhance efficiency in the production of local services has been determinant in shaping the 1997 tax reform and the subsequent constitutional reform (October 2001).

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

201

modelling and estimation are required in order to be able to identify the true process generating the observed spatial pattern in the data. Our results definitely suggest that some form of yardstick competition is going on. First, it turns out that local property tax rates are positively spatially auto-correlated only for those jurisdictions where the mayor can run for re-election and the electoral outcome is uncertain, while correlation is absent for those jurisdictions where either the mayor cannot run for re-election because he faces a binding term limit, or where the mayor is backed by a large majority. Moreover, it turns out that, coherently with yardstick competition theory, only the component of the tax rate that is not explained by observable characteristics—the unpredicted component of the tax rate—shows positive auto-correlation, a result that can hardly be explained by other theories. The rest of the paper is organized as follows. Section 2 briefly discusses yardstick competition theory and identifies its key predictions. Section 3 presents the data set, arguing for its validity as a natural laboratory for testing the theory. Section 4 turns to the empirical specification of a tax setting equation, Section 5 presents and discusses some preliminary estimation results and Section 6 tackles the issue of testing the theory. Finally, Section 7 concludes.

2. Yardstick competition and fiscal interaction Strategic interaction in fiscal variables across spatially close jurisdictions may occur for a variety of reasons. Yardstick competition theory points towards one such potential link. The main idea is that neighboring jurisdictions’ choices may provide voters with a positive informational externality on the quality of their own government: by observing fiscal choices in neighboring jurisdictions, citizens should be able to evaluate more precisely if the choices of their own governments are appropriate, or if they include some waste— political rents—which citizens dislike. Hence, voters should condition their electoral decision on the observation of the relative setting of the fiscal variables in their own and in neighboring jurisdictions.4 However, this simple story has several caveats. First, citizens must be able to observe fiscal choices in other jurisdictions beside their own. Second, the economic and institutional conditions in these jurisdictions must be sufficiently close to make a comparison meaningful. Third, it has to be true that electoral behavior is the main tool citizens have at their disposal to punish bad incumbents. For example, if mobility costs are sufficiently low, instead of voting them down, citizens could escape incompetent governments by simply emigrating or moving away taxable assets. Fourth, one needs to consider the strategic reaction of politicians to voters’ behavior. Politicians, as rational agents, are aware of the kind of calculations citizens make. Hence, they might change their 4 We do not provide here a formalisation of a yardstick competition model (see the working paper version of

this work for some theoretical analysis [14]). Starting from the seminal work of Besley and Case [5], however, a number of theoretical papers have recently appeared that look at the welfare and behavioral implications of yardstick competition—see Belleflamme and Hindricks [4], Besley and Smart [7,8], Bordignon and Minelli [15], Bodenstein and Ursprung [12], Wrede [40].

202

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

fiscal choices to influence voters’ behavior, and one has to establish how, in order to come up with testable hypotheses linking fiscal variables to electoral results. Indeed, the crucial point about the theory is not about tax setting as such, but in tax setting as linked to the incentives and constraints generated by the electoral system. On these grounds, the following are predictions of the theory. First, yardstick competition theory predicts a quite different behavior between incumbent politicians who run for re-election and incumbents who do not.5 Since the only reason to mimic other government policies is to be re-elected, governments facing a term limit should not be too worried about competing with their neighbors. Second, and more generally, the theory suggests that only incumbent governments that face uncertain electoral outcomes should interact strategically with their neighbors. If an incumbent is pretty confident of re-election in spite of her tax setting behavior, we should again not expect to find her fiscal choices to be affected by her neighbors’. Third, yardstick competition theory is based on the idea that neighboring jurisdictions’ policies are relevant in that they contain information on general economic conditions and on shocks that are common to nearby governments. Reasonably, though, neighbors’ tax rates also reflect neighbors’ specific characteristics and preferences, and informed voters should take those differences into account when evaluating their own government (Besley and Case [5]). Consequently, comparative performance evaluation conditional on neighbors’ circumstances—such as level and quality of public good provision—should yield correlation in the components of tax rates that cannot be explained by observable characteristics. Finally, a fourth implication of the theory is in terms of the popularity of the incumbent. As argued by Besley and Case [5], the probability of re-election should be negatively correlated with the difference between the tax rate selected by the incumbent and those of his neighbors. Unfortunately, the considerable political changes that occurred in Italy around the mid 1990s make the estimation of a popularity function on Italian municipality data extremely troublesome.6 Consequently, in the rest of the paper we focus on the estimation of a tax rate determination equation which allows for an effect from neighbors’ policies.

3. The data set: property taxation and politics in a sample of Italian cities In order to explain how we capture the key predictions of the theory in our empirical analysis, it is useful to begin by presenting the institutional scenario. We build a 5 See Case [21] and Besley and Case [5,6] for an analysis of the impact of term limits on US state governors’ policies. 6 Most of the mayors that were elected in the early 1990s either did not run again at the next election in the second half of the 1990s—making it impossible to estimate the impact of their tax choices on their popularity—or, when they did run again, they were supported in the following elections by coalitions that were radically different in terms of number and composition of supporting parties, and often included parties that did not even exist five years earlier. See Section 3.

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

203

comprehensive data set on local property taxation in a sample of municipalities in Lombardia, a large region in the north of Italy. The local property tax rate, named ICI (Imposta Comunale sugli Immobili), was introduced in 1993 together with a massive reform of the municipal electoral system, in an attempt to increase the administrative power and the accountability of city governments.7 Since then, the decision about the local property tax rate has become the crucial financial decision an Italian mayor must make each year. On average, property tax revenues alone cover more than 50% of total tax revenues of local governments, and more than 25% of total local government spending. There is also no question about the visibility of this decision; as a large majority of Italians are home-owners, and therefore have to pay this tax. In particular, ICI applies to both domestic and business properties. In both instances, the property tax base is defined by national procedures and regulations and it is therefore uniformly determined across local jurisdictions. Municipalities are free to choose different tax rates on domestic and business dwellings, in the 0.4 to 0.7 percentage points on the property tax base range, although they usually tend to set a uniform tax rate on domestic and business properties. However, domestic property taxation may be accompanied by a lump sum deduction, so that the statutory domestic tax rate does not coincide with the effective domestic tax rate. Unfortunately, we do not have enough information on the tax bases of municipalities to be able to recover the effective domestic ICI tax rate imposed in each municipality. Moreover, even if we could do so, it is doubtful that it would be very useful for our aims, as it is hard to believe that the average citizen himself had the time or the information needed to compute these effective tax rates and to use them for comparative evaluation across local governments. Consequently, although it might be in principle preferable to employ the domestic tax rate for our analysis (as we would make sure that the taxpayer is also a voter in the same jurisdiction), in what follows we use the business property tax rate instead, for which comparison of tax rate levels across municipalities is both easy (given the small size of Italian municipalities) and meaningful (as the tax base is commonly determined by the Central government). Note, however, that given the economic structure of Italian towns— which is characterized by an overwhelming presence of small business in the economy— this is not a major problem for our analysis, as in most instances payers of the business ICI in a municipality also reside in the same municipality. Note further that, in our sample, business property tax revenues constitute on average more than 70% of total local property tax revenues. As for the sample of municipalities we use here, it comprises 143 adjacent Comuni (municipalities) in the Province of Milan. The characteristics of the Comuni in the sample are described in Table 1, from which it emerges that the municipalities differ considerably in terms of size, population, demographic and socio-economic characteristics, even after excluding 44 jurisdictions in the Milan Province with less than 4000 residents. Beyond ICI 7 For more details on Italian local governments, see the relevant chapter in Ter-Minassian [36]. Bordignon [13] updates some of this evidence.

204

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

revenues, the rest of local current public spending is funded by other minor own sources of revenue (charges and fees) and by grants from the central government. The latter are unaffected by the property tax rates chosen by local authorities, and we will therefore consider them as exogenous variables in the ensuing analysis.8 The data set also includes detailed information on a number of variables describing the political complexion of the municipalities in the sample, including the political affiliation of local governments (the coalition of parties supporting a mayor), the share of votes and seats earned by the incumbent at the previous election, and whether the incumbent mayor faces a binding term limit. Political affiliation varies widely in the sample. Looking at the current political situation, about 55% of local authorities are controlled by center–left governments, about 36% from center–conservative governments, and the remaining municipalities are controlled by rightwing or Northern League parties. As for term limits, about 45% of the mayors in the sample (64 out of 143) are in their second period in office and cannot consequently run for re-election.

4. Empirical specification and choice of variables Our empirical strategy to test yardstick competition theory basically consists in exploiting the cross-section information available in our data set to investigate whether the mayors who should not be concerned with election outcomes—either because they cannot run for re-election, or because they can rely on a widespread support—do actually behave differently in their tax setting decisions from the remaining mayors. A potential drawback of this approach has to do with the very cross-sectional nature of the data. However, while the availability of a panel data set would be beneficial in many respects—particularly because of the possibility of controlling for fixed jurisdiction effects—simply exploiting the cross-sectional variability to identify behavioral differences between mayors does not pose dramatic econometric problems in this context. Actually, it is difficult to believe that the fact that a mayor is in his second term of office at a given point in time is correlated with a fixed jurisdiction effect. A more serious objection concerns the potential endogeneity of the mayor status in the tax setting equation. In fact, while yardstick competition theory predicts that mayors in their second term should set higher taxes (Besley and Case [6]), it might be argued that, on average, mayors in their second term of office are ‘better’ than first term mayors. After all, second term mayors have managed to gain re-election after their first term in office, making the set of mayors facing a term limit at a given point in time a selected sample of better than average politicians. As a result, the estimate of the effect of the mayor status on tax setting behavior may be biased, as the mayor status variable—the term limit dummy variable—is (negatively) 8 The data source of all demographic and socio-economic variables is ISTAT (National Institute of Statistics) and it refers to the calendar year 1991 (National Census of Government). Property tax rates have been provided by the Association of Italian City Councils (ANCI); data on grants come from the Budget Office of the Regional Government and refer to the budgetary year 1996.

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

205

correlated with the error term in the tax setting equation: the mayor survived to a second term because he is of good quality, and consequently he sets low tax rates (independently of the fact that he cannot run for re-election). Ideally, in order to tackle this problem one needs a proper instrument for the mayor status variable, in particular a variable that is correlated with the status of the mayor (that is, a variable that affects his re-election chances), but is not correlated with his quality. Understandably, though, no such variable is available. However, the fact that the likelihood of a mayor being in her second term of office is negatively correlated with the error term in the tax setting equation implies that we will end up underestimating the real difference in tax setting behavior between mayors that are constrained by the elections and mayors who are not. Consequently, any difference in tax setting behavior between first-term and second-term mayors that emerges form the estimation of a tax setting equation on our sample of municipalities can safely be interpreted as a lower bound, and definitely cannot be attributed to potential endogeneity of the mayor status variable. In the empirical model, local choices about property tax rates are assumed to depend on a series of local characteristics (including the traditional internal determinants of local tax setting policies: Wildasin [38]), as well as on neighboring jurisdictions’ tax rates. In particular, the local business property tax rate for the year 2000—a (143 × 1) vector t of municipalities in our sample—is related to the matrix of explanatory variables, X, that includes (see Table 1 for summary statistics): (a) structural characteristics of the jurisdiction: area, population, and urbanization rate; (b) socio-demographic characteristics of the resident population: percentage of children, percentage of elderly people, and rate of unemployment; (c) fiscal variables: grants from central government and disposable income per capita; (d) political variables: a political control dummy that is intended to capture systematic ideological differences between right-wing and left-wing governments; an election Table 1 Summary statistics Variable

Obs.

Mean

St. dev.

Min

Max

Business property tax rate (❤) Area Population Urbanization rate Percentage young Percentage old Percentage unemployed Grants per head (million liras) Income per head (million liras) Political dummy: left-wing party Election year dummy Second term mayors (term limit) First term mayors, of which: safe (share of vote > 1/3) unsafe (share of vote < 1/3)

143 143 143 143 143 143 143 143 143 80 83 64 79 50 29

5.905 1150.112 25,269 18.408 21.225 16.585 7.840 0.323 13.694

0.625 1579.958 110,327 15.325 1.981 3.232 1.677 0.095 2.143

4 198 4,005 1.808 15.1 5.9 4.751 0.057 10.086

7 18158 1,321,555 85.180 29 25.1 13.603 0.856 23.967

206

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

year dummy that allows for opportunistic behavior of incumbents in election years;9 a term limit dummy to control for potential differences in tax setting behavior between first term and second term mayors; and a confidence of re-election dummy, based (as we explain in greater detail below) on the share of the votes got by the incumbent in the previous election. Moreover, in order to account for spatial interaction among neighboring jurisdictions, we need to include a measure of the tax rates set by close-by jurisdictions as a determinant of a given jurisdiction’s tax rate: (e) neighboring governments’ property tax rates. Estimation of a tax setting equation that allows for an effect from neighbors’ policies poses a number of econometric problems, mainly deriving from the fact that own and neighbors’ choices are made simultaneously (Case [21]). Moreover, spatial dependence between neighboring governments’ tax rates might simply arise from the presence of common shocks, that one could easily mistake for substantive interaction (Brueckner [17]). Notice further that the issue of disentangling the correlation due to actual mimicking behavior from the one that is simply due to common shocks is crucial in our context, because the very existence of spatially correlated shocks to tax setting policies is itself a necessary condition to make comparative performance evaluation meaningful. We will consequently proceed as follows. First, not allowing for the variables at point (e) above, that is in the absence of any consideration of potential strategic interaction among neighboring governments, the model of local tax determination could be expressed in matrix form in a straightforward conventional way, as model (C): Model (C): t = Xb + e,

(1)

where b is a vector of parameters to be estimated, and e is a vector of i.i.d. error terms with mean zero and variance σ 2 . However, in the presence of a spatial process at the local level, model (C) would be mis-specified, and a local tax determination equation should instead be written either as a ‘spatial lag dependence’ model—model (L)—or as a ‘spatial error dependence’ model—model (E) (Anselin [1]): Model (L): t = Xb + pW t + e,

(2)

Model (E): t = Xb + u,

(3)

u = rW u + e,

where p in Eq. (2) is the parameter measuring the interaction between own (t) and a measure of neighbors’ property tax rates (W t)—the customary slope of a reaction function—while r in Eq. (3) is the spatial auto-correlation coefficient in the error term, and W is a square matrix that assigns neighbors to each locality. A test of yardstick competition theory clearly hinges on the identification of the neighboring jurisdictions, whose fiscal choices should be easily observed by voters. In 9 Mayors in power in 2000 were elected at different dates between 1995 and 1999. Consequently (and given

a four to five years term of office), a proportion of them (about 58% in the sample) faced elections during the year 2000.

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

207

the ensuing analysis, we follow the literature by identifying neighbors with spatially close jurisdictions, on the basis of the argument that information flows more easily across close-by jurisdictions and that adjacent localities are more likely to experience similar economic shocks. In particular, the weights matrix W is based on a simple border-sharing criterion: two municipalities are held to be neighbors if they have a common border.10 While being pretty similar from an empirical point of view, models (L) and (E) derive from substantially different behavioral hypotheses. According to model (L), tax rates are correlated independently of the levels of the Xs. This means that, if p > 0, it is more likely that the own tax rate is high if tax rates in the neighborhood are high, irrespective of the reason why tax rates are high in neighboring jurisdictions. According to model (E), on the other hand, only the components of the tax rates that are not explained by the Xs tend to be correlated. Whether yardstick competition should generate a spatial process as the one described by model (L) or as the one described by model (E) basically depends on the voter’s information set on tax rates and their determinants in own and neighboring jurisdictions (Besley and Case [5]). As long as voters have a reasonable knowledge of the deterministic factors affecting local tax rates in own and neighboring jurisdictions—a reasonable assumption in the Italian context, due to the small municipality size—then yardstick competition should definitely generate the spatial process described by model (E). However, it could be argued that the very model (E) could simply arise from the fact that neighboring jurisdictions really are hit by correlated shocks, without incumbents behaving strategically and pretending they are. Consequently, the crucial test of the theory needs to rest on the discrepancy in behavior between politicians in different political and electoral conditions (first term versus second term, and safe versus unsafe mayors), that is in differences in the r parameter among different groups of politicians: if such correlation turns out to be relevant only for incumbents running for re-election in uncertain contests, then it would be hard to explain it in terms of the sheer existence of non-behavioral common shocks.

5. Estimation results As a first test of the local fiscal interaction hypothesis, we estimate in this section the basic spatial models outlined above. In general, if local governments do not set taxes in isolation but affect each other in making their fiscal choices, we should expect to find that tax rates in neighboring jurisdictions are correlated. First, and in order to convince ourselves that a spatial framework is required to analyze local tax setting, we have preliminary performed a spatial auto-correlation test on the raw level of the local property tax rate as well as on the ordinary least squares (OLS) residuals 10 We checked for the validity of this criterion by building weighting matrices based on alternative criteria (such

as similarity among jurisdictions in per capita income levels, rate of unemployment and urbanization rate, as in Case et al. [22], as well as matrices whose spatial weights take into account the population size of neighbors (as in Brueckner and Saavedra [19]). None of them, though, led to significant improvements upon the basic unweighted border sharing specification (results are available from the authors upon request).

208

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

of a non-spatial tax setting model such as model (C). According to the Moran spatial test (Anselin [1]), both raw tax rates and the OLS residuals from the tax setting equation indeed appear to be positively spatially auto-correlated, and the test is in both instances significant at above the 99% level of confidence.11 As a result, a model of local tax determination where spatial interaction is explicitly allowed for, either model (L) or model (E), should be employed instead of model (C). Unfortunately, the Moran test is unable to establish which of the two spatial models is appropriate to explain local tax setting decisions. A more precise indication of what is the most likely source of spatial dependence is given by the two robust Lagrange multiplier (LM) tests on the OLS residuals of model (C), developed by Anselin et al. [2]. The LM test for spatial error dependence (that is robust to the presence of spatial lag dependence) suggests at the 95% level of confidence that the errors in the tax determination equation are spatially auto-correlated.12 On the other hand, the LM test for spatial lag dependence does not point conclusively towards the presence of substantive dependence in the tax rates. The best way to discriminate among the two competing models, then, is to estimate them by maximum likelihood (ML) methods and compare their respective likelihood values (Brueckner [17]). Table 2 shows the results of estimating a local tax determination equation with the level of the business property tax rate in year 2000 as the dependent variable. Column 1 of the table shows the OLS estimation results of a non-spatial tax rate determination equation, while column 2 shows the benchmark OLS estimation results of a tax determination model where a spatially weighted average of neighboring jurisdictions’ tax rates is used as an explanatory variable. As it does not take into account the simultaneity of own and neighbors’ decisions, OLS yields inconsistent estimates of the parameters (Anselin [1]). Consequently, columns 3 and 4 in Table 2 report the ML estimates of the spatial lag dependence and of the spatial error dependence models, that are obtained using standard non-linear estimation techniques.13 As argued above, in all instances the spatial weights matrix used is based on a standard border-sharing criterion and is row-standardised. The results show that there is no statistically significant evidence of systematic differences between right-wing and left-wing governments in property tax rate setting policies,14 while there is significant evidence of opportunistic behavior: tax rates tend to be systematically lower in election years. The structural and demographic variables appear to have a significant impact on the level of the local tax rate. In particular, while area has a positive and significant effect on the tax rate, population size has a negative and highly significant impact, suggesting that a larger population size allows to exploit economies of scale in public service provision.15 On the other hand, the rate of urbanization has a

11 The Moran test on the OLS residuals was performed using a row-standardized spatial weights matrix, based on a border-sharing criterion. The test is distributed as a standard normal (Anselin [1]) 12 The LM test is distributed as a χ 2 and takes on the value of 4.80. [1] 13 See Anselin [1] for details. 14 For a work which considers ideological differences and mimicking behavior, see Sole-Olle [35]. 15 We also included a squared population term to allow for a non-linear relationship, but the coefficient turned out not to be significantly different from zero.

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

209

Table 2 Business property tax rate determination equation

Constant Area Population Urbanization rate

ML non-spatial model (C)

OLS spatial lag model (L)

ML spatial lag model (L)

ML spatial error model (E)

(1)

(2)

(3)

(4)

3.974***

3.965***

3.970***

(3.20)

(2.81)

(3.09)

3.966*** (2.87)

0.034*** (3.73)

0.031*** (3.40)

0.032*** (3.56)

0.035*** (3.61)

−0.577*** (4.05)

−0.556*** (3.93)

−0.564*** (4.03)

−0.619*** (4.20)

1.388*** (2.73)

1.313*** (2.60)

1.340*** (2.66)

1.541*** (2.77)

Percentage young

−0.162** (2.57)

−0.155** (2.50)

−0.158*** (2.60)

−0.144** (2.31)

Percentage old

−0.117*** (2.93)

−0.105*** (2.63)

−0.109*** (2.86)

−0.100** (2.37)

Unemployment rate

−0.026 (0.65)

−0.028 (0.70)

−0.027 (0.72)

−0.040 (0.94)

Grants per capita

0.080 (0.11)

0.234 (0.33)

0.176 (0.25)

0.499 (0.64)

Income per capita

−0.021 (0.66)

−0.022 (0.69)

−0.022 (0.72)

−0.019 (0.60)

0.015 (0.13)

0.005 (0.10)

0.008 (0.17)

0.023 (0.28)

Left-wing party dummy Election year dummy

−0.563** (2.34)

Spatial parameter, p

−0.602*** (2.58)

−0.588** (2.52)

0.250* (1.85)

0.158 (1.50) 0.302** (2.43)

Spatial parameter, r Moran I test z(0, 1)

2.70***

2 Robust LM test for spatial error χ[1] 2 Robust LM test for spatial lag χ[1] 2 LR test χ[1]

4.80**

Log likelihood Observations

−0.688*** (2.93)

2.59 2.20 −608.80 143

−607.71 143

143

5.28** −606.16 143

Notes. Dependent variable = ordinary ICI rate (business property tax rate) in year 2000. Absolute t-statistics in parentheses; p = spatial auto-correlation coefficient in dependent variable; r = spatial auto-correlation coefficient in errors; spatial matrix for computing neighbors’ taxes = geographical (border-sharing) and rowstandardized. *** 99% level of statistical significance. ** 95% level of statistical significance. * 90% level of statistical significance.

210

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

positive effect, and higher percentages of unemployed, elderly and young people have a negative effect on tax rates. The fiscal variables do not appear to have any systematic impact on the tax rate, with the coefficients being rather imprecisely estimated. While for the income variable this could be due to large measurement error, theory itself does not univocally predict the effect of lump-sum grants on local tax rates; for instance, the existence of a ‘flypaper effect’ would require a very small (negative) effect of grants on the local tax rate. Finally, as far as the spatial parameters are concerned, the spatial lag model (L) yields an ML estimate of the spatial coefficient p of about 0.16, that is not statistically significant. Compared to the estimate of p in column 2 (OLS), it appears that OLS leads, as expected, to an upward bias in the estimation of the spatial lag parameter p. Moreover, it appears that the increase in likelihood of model (L) with respect to a model that arbitrarily sets p = 0 is not statistically significant. On the other hand, the spatial error model (E) yields an ML estimate of the spatial coefficient r of about 0.3, that is statistically significant at the 95% level, and leads to a substantial improvement with respect to a non-spatial specification that sets r = 0. A likelihood ratio (LR) test of model (E) with respect to the non-spatial model (C) yields a value of 5.3, which allows to reject the restriction r = 0 at above the 97.5% level of confidence.16 The empirical evidence is consequently strongly in favour of model (E), and prompts us to investigate further whether the observed spatial auto-correlation in tax rates can really be attributed to strategic tax setting.

6. Detecting yardstick competition The above empirical evidence largely supports model (E), which is compatible with a yardstick competition framework where voters have some information on the nonstochastic determinants of local taxation. However, model (E) is also compatible with the presence of spatially correlated shocks to tax setting behavior that have no behavioral significance. To rule out the latter interpretation, we need to exploit the difference in the sample between different types of mayors, starting with the role played by the existence of term limits. If the correlation in tax rates were simply due to common shocks, we should expect tax rates to be correlated among neighboring jurisdictions, whether or not the mayor faces a term limit. On the other hand, if yardstick competition generates spatial auto-correlation in taxes, we should find stronger correlation among those jurisdictions where mayors seek re-election. Table 3 shows the estimation results of a tax setting equation where we explicitly allow for different behavior of mayors, depending on whether they can run for re-election or they face a term limit. One way of expressing such tax setting equation in the spatial error dependence specification is as in model (E ): Model (E ): t = Xb + u,

u = r1 ZW u + r2 (I − Z)W u + e,

(4)

16 Twice the difference in the log likelihood between the restricted and unrestricted model is distributed as χ 2 with degrees of freedom equal to the number of restrictions (Bivand [9]).

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

211

Table 3 Business property tax rate determination equation: the impact of term limits ML spatial error model (E )a

ML spatial error model (E )b

(1)

(2)

0.036∗∗∗ (3.91)

0.036∗∗∗ (3.91)

−0.645∗∗∗ (4.46)

−0.625∗∗∗ (4.50)

Urbanization rate

1.603∗∗∗ (2.85)

1.530∗∗∗ (2.85)

Percentage young

−0.172∗∗∗ (2.81)

−0.181∗∗∗ (3.00)

Percentage old

−0.112∗∗∗ (2.80)

−0.118∗∗∗ (3.00)

Unemployment rate

−0.061 (1.45)

−0.059 (1.41)

Grants per capita

0.346 (0.44)

0.205 (0.30)

Income per capita

−0.031 (1.01)

−0.031 (1.00)

0.052 (0.52)

0.061 (0.63)

3.977∗∗∗ (3.45)

3.980∗∗∗ (3.63)

Area Population

Left-wing party dummy Term limit Constant Election year dummy

−0.736∗ (1.87)

Spatial parameter, r1

0.160 (0.70)

No term limit Constant Election year dummy

3.979∗∗ (2.38) −0.702∗∗ (2.50)

−0.694∗ (1.78) –

3.982∗∗ (2.46) −0.677∗∗ (2.42)

Spatial parameter, r2

0.474∗∗∗ (2.68)

0.503∗∗∗ (2.83)

Log likelihood

−603.02

−603.15

143

143

Observations Notes. See Table 2. a [r Z + r (I − Z)]W . 1 2 b r (I − Z)W . 2

where I is the (143 × 143) identity matrix, and Z is simply a (143 × 143) matrix of dummies, in the sense that element z(i, j ) of matrix Z equals 1 if i = j and the incumbent in jurisdiction i faces a binding term limit, while it equals 0 otherwise. In other words, if a mayor faces a term limit, his interaction with neighboring jurisdictions’ policies is

212

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

captured by parameter r1 , while spatial interaction is captured by r2 for mayors running for re-election. The theoretical prediction is that r1 = 0 and r2 = 0. Clearly, if r1 = r2 = r, one ends up with model (E). Indeed, the spatial process in the error term u can be expressed as
−1 u = I − r1 ZW − r2 (I − Z)W e. (5) Consequently, the log likelihood function for model (E ) above is   L = −(N/2) ln(π) − (N/2) ln σ 2 − (1/2)σ −2 (t − Xb) H  H (t − Xb) + J,

(6)

where H = [I − r1 ZW − r2 (I − Z)W ], and J = det |H | is the Jacobian for the transformation of the vector e into the vector t (Anselin [1]). Overall, model (E ) allows for heterogeneous behavior of first term versus second term mayors on three accounts. First, it allows the constant term in the regression equation to differ among mayors that can run for re-election and mayors that cannot. Second, it allows for different behavior in election years among the two groups. Finally, it admits different spatial interaction patterns. The results in Table 3 show that the overall difference in property tax rate setting behavior is indeed very small, as the difference in the constant terms is not significant. As argued above, this could be due to the fact that while first term mayors are bound by their election prospects, second term mayors are a selected sample of better than average mayors and consequently do not tend to raise tax rates. Moreover, and unexpectedly, both kinds of mayors tend to set lower tax rates in election years. While it is hard to explain this result in terms of yardstick competition (as lame ducks should not adopt an opportunistic behavior), it is likely that party discipline may play an important role in local politics and could force incumbents in their second term of office not to raise taxes on the eve of an election at their party’s expense. Finally, it turns out that for those municipalities where the mayor steps down because of a binding term limit, there is no evidence of spatial auto-correlation in tax rates. Parameter r1 is estimated not to be significantly different from zero. On the other hand, mayors that can run for re-election appear to be significantly affected by neighbors’ tax rates. The estimate of r2 in Table 3 is a high and statistically significant (99% level of confidence) value of almost 0.5, and an LR test of model (E ) with respect to the restricted model (E) shows that the likelihood of the restricted model is significantly lower (at the 95% level of confidence). Moreover, column 2 of Table 3 shows the results of estimating a model that uses matrix (I − Z)W as a spatial weighting matrix and imposes r1 = 0. It appears that the spatial interaction parameter for incumbents running for re-election again equals a large value of about 0.5, while the restriction on mayors facing a term limit (r1 = 0) does not lead to a significant decrease in likelihood. Consequently, the evidence suggests that the spatial interaction pattern is indeed affected by the status of the mayors, in that only politicians running again for office appear to be affected by their neighbors’ policies. As a result, it is very unlikely that the observed spatial auto-correlation in local property tax rates can be attributed to common shocks. A further test of the yardstick competition hypothesis consists in exploiting the fact that while some incumbent mayors hardly passed the post in the previous election, some

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

213

other mayors got a large majority of the votes. Consequently, it is reasonable to expect that the latter should feel free of implementing their preferred fiscal policies, irrespective of neighbors’. Consequently, we build a ‘strong majority’ dummy that equals one if the mayor got more than 1/3 of the total votes at the first election round, and interact it with neighbors’ taxes. The spatial error dependence model can now be expressed as in model (E ): Model (E ):

t = Xb + u, u = r1 ZW u + r2 (I − Z)F W u + r3 (I − Z)(I − F )W u + e,

(7)

where, similarly to matrix Z, F is a (143 × 143) matrix of dummies whose element f (i, j ) equals 1 if i = j and the incumbent in jurisdiction i got more than 1/3 of the votes at the first election round, while it equals 0 otherwise. The theoretical prediction is in this case that r1 = r2 = 0, i.e., that incumbents facing a term limit (z(i, i) = 1), as well as incumbents who do not face a term limit but are pretty confident of re-election (z(i, i) = 0, f (i, i) = 1), should not compete with their neighbors, and that r3 = 0—that is, mayors who run for reelection and are not supported by large majorities (z(i, i) = f (i, i) = 0) should interact strategically with their neighbors. The likelihood function is as for model (E ), except for the Jacobian term that takes the form   (8) J = detI − r1 ZW − r2 (I − Z)F W − r3 (I − Z)(I − F )W . The results in the first column of Table 4 show that, as far as the overall tax rate level is concerned, there are no significant differences among mayors according to their status. However, mayors that run for re-election in uncertain contests appear to reduce tax rates to a larger extent in election years. Moreover, a substantially heterogeneous pattern emerges as far as spatial interaction in tax setting policies is concerned. On the one hand, there is no evidence of spatial auto-correlation in tax rates either among mayors facing term limits or among mayors that run for re-election and are backed by large majorities: the estimated spatial coefficient r1 is positive, but not statistically significant, while the estimated spatial coefficient r2 is negative and not significant. On the other hand, there is significant evidence of positive spatial auto-correlation for mayors that run for re-election in uncertain contests, with a large estimate of r3 of about 0.75 (with a t-statistic in excess of 3). Moreover, an LR test of model (E ) with respect to the restricted model (E) shows that the likelihood of the restricted model is significantly lower (at the 95% level of confidence). As a further check of the above result, column 2 in Table 4 reports the estimates of a restricted model that imposes r1 = r2 = 0, and uses the weighting matrix (I − Z)(I − F )W to build neighbors’ tax rates. Such model imposes the restriction that only incumbents that are in their first term of office and are not supported by large majorities tend to react to their neighbors’ policies (parameter r3 ). The estimated coefficients, as well as the log likelihood of this specification, turn out to be almost identical as the results in column 1 of Table 4. To sum up, the results in Tables 3 and 4 do offer support to the thesis that local tax rates are set strategically, with electoral considerations uppermost in politicians’ minds, and that mayor candidates look at their neighbors when setting their own tax rates in order not to

214

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

Table 4 Business property tax rate determination equation: the impact of term limits and confidence of re-election ML spatial error model (E )a

ML spatial error model (E )b

(1)

(2)

0.038∗∗∗ (4.11) −0.663∗∗∗ (4.65)

0.037∗∗∗ (4.24) −0.636∗∗∗ (4.67)

Urbanization rate

1.691∗∗∗ (3.02)

1.559∗∗∗ (3.01)

Percentage young

−0.163∗∗ (2.66) −0.112∗∗∗ (2.79)

−0.177∗∗∗ (2.86) −0.122∗∗∗ (3.15)

Unemployment rate

−0.053 (1.24)

−0.039 (0.98)

Grants per capita

0.486 (0.66) −0.027 (0.88)

0.259 (0.38) −0.027 (0.86)

0.044 (0.47)

0.047 (0.47)

3.973∗∗∗ (3.25)

3.977∗∗∗ (3.43)

Area Population

Percentage old

Income per capita Left-wing party dummy Term limit Constant Election year dummy Spatial parameter, r1 No term limit—confidence of re-election Constant Election year dummy Spatial parameter, r2

−0.741∗ (1.93) 0.222 (1.41)

−0.713∗ (1.83) –

3.975∗∗ (2.53) −0.679∗∗ (2.12)

3.979∗∗ (2.55) −0.646∗∗ (1.98)

−0.014 (0.12)

–

No term limit—no confidence of re-election Constant 3.965∗∗ (2.24)

3.970∗∗ (2.31)

−1.226∗∗ (2.29)

−1.192∗∗ (2.22)

Election year dummy Spatial parameter, r3 Log likelihood Observations

0.753∗∗∗ (3.08) −600.02 143

Notes. See Table 2. a [r Z + r (I − Z)F + r (I − Z)(I − F )]W . 1 2 3 b r (I − Z)(I − F )W . 3

0.802∗∗∗ (3.18) −600.21 143

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

215

lose electoral support. In particular, the results show that positive spatial auto-correlation emerges only with regard to mayors having a re-election concern, while mayors facing a binding term limit, as well as the ones that are confident of re-election, do not appear to be affected by their neighbors’ policies. Consequently, it is unlikely that the observed spatial auto-correlation in tax rates can be driven by spatially auto-correlated shocks. Moreover, the results show that it is not raw tax rates that are positively spatially autocorrelated, but only that part of the tax rate which cannot be explained by local conditions, suggesting that local policymakers believe that the electorate is aware of the determinants of local property tax rates.

7. Concluding remarks Yardstick competition is potentially one of the most interesting theories that have been put forward to explain fiscal interaction among local governments. Knowing if it is something more than just a theoretical construct is therefore important. Finding ways to test the theory, discriminating it from other competing explanations of interaction among neighboring local authorities—such as tax competition or expenditure spill-overs—is not an easy task. Moreover, disentangling empirically the sheer presence of common shocks to local taxation and spending from substantive strategic interaction requires adopting a proper empirical framework, especially if one considers that some form of spatial autocorrelation among neighboring jurisdictions in the unobserved determinants of local policy making is one of the pre-requisites of the theory itself. We have exploited the features of the Italian local government electoral system in an attempt to disclose evidence of yardstick competition behavior, starting from the observation that local business property tax rates tend to exhibit positive spatial autocorrelation. The empirical analysis has been performed on a cross-section of Italian municipality data that also includes political and electoral information. The results of the estimation of a spatial tax setting equation provide a rather satisfactory explanation of local tax determination. It turns out that positive spatial auto-correlation emerges only with regard to mayors having electoral concerns, while mayors facing a binding term limit as well as the ones that are confident of re-election do not appear to be affected by their neighbors’ policies. Consequently, it is unlikely that the observed spatial auto-correlation in tax rates can be driven by spatially auto-correlated shocks. Rather, it indeed appears to be driven by strategic considerations.

Acknowledgments Previous versions of this paper have been discussed in seminars held at the Universities of Rome III, Milan Bicocca, Brescia, Pavia, Munich, Honolulu, Marseille, and at the International Conference organized by MEP (Master in Public Economics) at the Catholic University of Milan. We thank all participants, in particular Rosella Levaggi, Ben Lockwood, Piergiovanna Natale, Fabio Padovano, Jorn Rattso, and Patrizio Tirelli. We are also grateful to Aldo Vaccaro from the Home Ministry for providing data on political

216

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

variables, and to two anonymous referees and the editor, Jan Brueckner, for helpful and constructive comments. All remaining errors are our responsibility.

References [1] L. Anselin, Spatial Econometrics: Methods and Models, Kluwer Academic, Dordrecht, 1988. [2] L. Anselin, K. Bera, R. Florax, M. Yoon, Simple diagnostic tests for spatial dependence, Regional Science and Urban Economics 26 (1996) 77–104. [3] L. Anselin, R. Florax (Eds.), New Directions in Spatial Econometrics, Springer–Verlag, Heidelberg, 1995. [4] P. Belleflamme, J. Hindricks, Yardstick competition and political agency problems, Working paper No. 441, Department of Economics, Queen Mary, University of London, 2001. [5] T. Besley, A. Case, Incumbent behavior: Vote seeking, tax setting and yardstick competition, American Economic Review 85 (1995) 25–45. [6] T. Besley, A. Case, Does electoral accountability affect economic policy choices? Evidence from gubernatorial term limits, Quarterly Journal of Economics 110 (1995) 769–798. [7] T. Besley, M. Smart, Globalization and electoral accountability, Unpublished paper, 2001. [8] T. Besley, M. Smart, Does tax competition raise voter welfare? Unpublished paper, 2001. [9] R. Bivand, Regression modeling with spatial dependence: an application of some class selection and estimation methods, Geographical Analysis 16 (1984) 25–37. [10] R. Bivand, S. Szymanski, Spatial dependence through local yardstick competition: theory and testing, Economics Letters 55 (1997) 257–265. [11] R. Bivand, S. Szymanski, Modeling the spatial impact of the introduction of compulsive competitive tendering, Regional Science and Urban Economics 30 (2000) 203–219. [12] M. Bodenstein, H. Ursprung, Political yardstick competition, economic integration and distributional choice in a federation, CESifo working paper No. 501, 2001. [13] M. Bordignon, Problems of soft budget constraint in intergovernmental relationships: the case of Italy, Working paper R-398, Inter-American Development Bank Research Network, 2000. [14] M. Bordignon, F. Cerniglia, F. Revelli, In search for yardstick competition: Property tax rates and electoral behavior in Italian cities, CESifo working paper No. 644, 2002. [15] M. Bordignon, E. Minelli, Rules transparency and political accountability, Journal of Public Economics 80 (2001) 73–98. [16] C. Brett, J. Pinkse, The determinants of municipal tax rates in British Columbia, Canadian Journal of Economics 33 (2000) 695–714. [17] J. Brueckner, Testing for strategic interaction among local governments: the case of growth controls, Journal of Urban Economics 44 (1998) 438–467. [18] J. Brueckner, Strategic interaction among governments: an overview of empirical studies, International Regional Science Review (2003), forthcoming. [19] J. Brueckner, L. Saavedra, Do local governments engage in strategic property tax competition? National Tax Journal 54 (2001) 203–229. [20] T. Buettner, Local business taxation and competition for capital: the choice of the tax rate, Regional Science and Urban Economics 31 (2001) 215–245. [21] A. Case, Interstate tax competition after TRA86, Journal of Policy Analysis and Management 12 (1993) 136–148. [22] A. Case, J. Hines, H. Rosen, Budget spillovers and fiscal policy interdependence, Journal of Public Economics 52 (1993) 285–307. [23] D. Figlio, V. Kolpin, W. Reid, Do states play welfare games? Journal of Urban Economics 46 (1999) 437– 454. [24] B. Heyndels, J. Vuchelen, Tax mimicking among Belgian municipalities, National Tax Journal 60 (1998) 89–101. [25] H. Kelejian, D. Robinson, A suggested method of estimation for spatial interdependent models with autocorrelated errors, and an application to a county expenditure model, Papers in Regional Science 72 (1993) 297–312.

M. Bordignon et al. / Journal of Urban Economics 54 (2003) 199–217

217

[26] H. Ladd, Mimicking of local tax burdens among neighboring counties, Public Finance Quarterly 20 (1992) 450–467. [27] J. Murdoch, M. Rahmatian, M. Thayer, A spatially autoregressive median voter model of recreation expenditures, Public Finance Quarterly 21 (1993) 334–350. [28] W.E. Oates, Fiscal Federalism, Harcourt Brace Jovanovich, New York, 1972. [29] F. Revelli, Spatial patterns in local taxation: tax mimicking or error mimicking? Applied Economics 33 (2001) 1101–1107. [30] F. Revelli, Testing the tax mimicking versus expenditure spillover hypotheses using English data, Applied Economics 34 (2002) 1723–1731. [31] F. Revelli, Reaction or interaction? Spatial process identification in multi-tiered government structure, Journal of Urban Economics 53 (2003) 29–53. [32] L. Saavedra, A model of welfare competition with evidence from AFDC, Journal of Urban Economics 47 (2001) 248–279. [33] P. Salmon, Decentralization as an incentive scheme, Oxford Review of Economic Policy 3 (1987) 24–43. [34] M. Shroder, Games the states don’t play: welfare benefits and the theory of fiscal federalism, The Review of Economics and Statistics 77 (1995) 183–191. [35] A. Sole-Olle, Tax mimicking and electoral control: An empirical analysis of local tax setting in Spain, Unpublished paper, Universitat de Barcelona, 2001. [36] T. Ter-Minassian (Ed.), Fiscal Federalism in Theory and Practice, International Monetary Fund, Washington, 1997. [37] C. Tiebout, A pure theory of local expenditures, Journal of Political Economy 64 (1956) 416–424. [38] D. Wildasin, Urban Public Finance, Harwood Academic, New York, 1986. [39] J. Wilson, Theories of tax competition, National Tax Journal 53 (1999) 269–304. [40] M. Wrede, Yardstick competition to tame the Leviathan, European Journal of Political Economy 17 (2001) 705–721.