Fiscal Federalism, Decentralization and Economic Growth: A Meta-Analysis by Lars P. Feld (University of Heidelberg, ZEW, CESifo, CREMA, SIAW-HSG), Thushyanthan Baskaran (University of Heidelberg) and Jan Schnellenbach (University of Heidelberg) Abstract: The distribution of competencies between the different levels of a federal system may have remarkable effects on economic growth, because mainly the regions of a country contribute to national economic development. Thus, a government’s economic policy is reasonably shaped along regional lines. The theoretical discussion in economics focuses however on the efficiency aspects of a decentralized provision and financing of public services; rarely the argument is raised that decentralization or federalism increases growth through a higher ability of the political system to innovate and to carry out reforms. After a discussion of the theoretical arguments on federalism and growth, we address the empirical question in this paper how important the assignment of decision making powers and the design of fiscal federalism are for economic development. Finally, on the basis of existing theoretical and empirical studies on economic growth and federalism, open questions and possible ways of answering them are presented. The authors would like to thank the German Science Foundation (DFG) for funding this project (DFG-SPP 1142). Corresponding Author:
Prof. Dr. Lars P. Feld University of Heidelberg Alfred-Weber-Institut Grabengasse 14 D-69117 Heidelberg Germany
[email protected]
2 1. Federalism and Growth: Introduction What impact can fiscal federalism have on economic growth? If we take the perspective of the standard, Solow-Swan growth model (Solow 1956, Swan 1956), this must appear as a peculiar question. With a given production technology, income per capita in the steady state is fully determined by the savings rate, the growth rate of the population and the rate of depreciation of the stock of capital. Prima facie, it is difficult to see through which mechanisms fiscal federalism, as a principle of organizing government, could have an impact on these parameters that are all exogenous in the Solow-Swan world. As we will see later, there might in fact be a channel through which federalism brings forth a different propensity to save than a unitary state if the saving decision is endogenized in a model with overlapping generations. If federalist frameworks were indeed associated with different propensities to save, different growth rates on the convergence paths towards the respective steady states would be expected for otherwise identical model economies even in the Solow-Swan model. But, this is not the first thought that comes to mind when suspecting that fiscal federalism affects growth. As Oates’ (1999) survey shows, it is rather the efficiency (or inefficiency) properties of fiscal competition, fiscal equalization systems, or intergovernmental fiscal relations in a polity in general that determine the relation between fiscal federalism and economic performance. These properties have been focused by a large number of theoretical and empirical studies during the last decades. And based on these microeconomic analyses, Oates (1993) argues that fiscal federalism affects economic development. Those familiar with modern endogenous growth theory may wonder which transmission mechanisms between these variables supposedly exist. Putting the view of neoclassical growth theory at the beginning of our analysis highlights that it is anything but easy to argue that fiscal federalism has an impact on economic growth at all. In addition, it is even more challenging to come up with convincing evidence that it actually does. Existing empirical studies may leave the impression that such a relationship cannot be empirically established. In this paper, we thus shed some light on both issues by, first, providing a survey of the theoretical literature on fiscal federalism, decentralization and economic growth. We particularly focus on the transmission channels by which federalism affects growth. Second, the empirical evidence on this relation is studied. Because the results are numerous and it is difficult to obtain a clear-cut conclusion by simply providing a survey of empirical results, we conduct a meta-analysis which targets the main outcome of that literature and identifies the factors of the single studies that
3 determine the studies’ outcomes. Finally, as meta-analytical techniques are still infrequently used in economics, we also aim at contributing to the establishment of these methods. The paper is organized as follows: In Section 2, the theoretical studies on the impact of fiscal federalism on economic growth are reviewed. A brief survey of the empirical studies follows in Section 3. They are quantitatively reviewed in Section 4 and are subject to the meta-analysis in Section 5. In Section 6, the results of meta-regressions are reported in order to give an impression which characteristics of the single studies influence the results. This also helps to highlight in which direction future analyses in this field could be undertaken. Section 7 provides conclusions. 2.
How Could Fiscal Federalism Affect Economic Growth? The Theoretical Arguments
Returning to the Solow-Swan framework for the moment, it is well known that there only technological progress can explain persistent positive growth rates (differentials) of per capita incomes. With only labor-augmenting technological progress guaranteeing the existence of a steady state (e.g. Barro and Sala-i-Martin 2001, p. 54), the steady state is then characterized by a constant income per effective unit of labor, but income per physical unit of labor growing at the rate of technological progress, which again is exogenous. Even with its focus on the pure mechanics of economic growth, and with its underlying approach of holding the important parameters exogenous, the Solow-Swan model thus already offers important first suggestions where to look at if we are interested in finding channels through which federalism might affect economic growth, and technological progress appears to be a very promising candidate. This is certainly the case if we interpret technological progress in a rather broad sense, including not only the entrepreneurial spirits that lead the private sector to search for and to implement superior production technologies, but also the potential for progress in the public sector. The latter ranges from the discovery of institutions that solve political control problems more efficiently and thus reduce the scope for wasteful activities in the public sector, to actual technologies for the production of public goods. There are already a number of contributions that link public sector efficiency gains to federalism, such as the investigation of yardstick competition by Besley and Case (1995). Although most of these contributions do not explicitly refer to the issue of economic growth, they are important in offering micro-foundations of public sector innovation. Therefore, we will return to them in detail later. But before having a closer look at endogenized technological change, we will in the following section review some pioneering attempts of incorporating federalism in models that endogenize decisions on saving and investment.
4 2.1 Federalism, Capital Accumulation and Endogenous Saving Decisions The approach of incorporating fiscal federalism in growth models with endogenous saving decisions goes back to Brueckner (1999), who uses a Diamond-OLG growth model that is extended to include the provision of a publicly provided private good, financed through uniform lumpsum taxes. Heterogeneity of the population is captured by distinguishing different demands for the publicly provided good expressed by young and old individuals. A unitary government is assumed to supply some compromise level of the good, which lies in between the utilitymaximizing levels of the two respective age cohorts. This assumption may be thought of as reflecting the fact that in democracies, the enacted policies often represent some bargaining solution that lies between the ideal policies of different constituencies. In a federal system, on the other hand, Brueckner assumes a perfectly working Tiebout (1956) sorting mechanism, which implies that in equilibrium, two jurisdictions are characterized by age-homogeneous populations and each receives its utility-maximizing supply of the publicly provided good. The change of equilibrium policies after the introduction of a federal system also has an impact on the level of equilibrium savings. This is due to the fact that, under the assumed condition of weak complementarity, an increase of the quantity of the publicly provided good would unambiguously increase the marginal utility of the private good consumed by both young and old individuals. If now we have a situation where the desired quantity of the publicly provided good is greater for the young than for the old individuals, then this implies that with the switch to federalism, the marginal utility of private consumption increases for the young, and decreases for the old – compared to the benchmark of a unitary state, with an intermediate quantity of the publicly provided good. The individuals readjust their consumption plans by saving less, in order to consume more of the private good in their young and less in their old age. The converse argument would hold if the old preferred a larger quantity of the publicly provided good than the young. In this case, higher savings would be the result of a switch to federalism. Brueckner then goes on to solve the Diamond-OLG model for steady state solutions, and shows that as long as saving depends positively on the current capital intensity, the steady state capital intensity of a federal system compared to a unitary system does indeed depend on the initial assumptions on the relative demands of the young and the old for the publicly provided good. If the young demand a higher quantity than the old, federalism reduces steady state capital intensity, and if the young demand a lower quantity, then the steady state capital intensity increases after a federalist system is implemented. Wages and output per unit of labor move into the same direction. Summing up, while the growth rate of total output in steady state is once again identical for
5 otherwise identical unitary and federal economies, the model predicts different growth rates on the convergence paths towards the respective steady states, and different steady state levels of income per capita, conditional on an economy being governed by a centralized or a decentralized political regime. A model that assumes a perfectly working Tiebout sorting mechanism obviously presents an extreme case. Nevertheless, it is useful to remember that the main justification for fiscal federalism in the literature has always been diversity of preferences for public goods (e.g. Oates 1972) and the ability of local governments to supply specific levels and types of public goods that are in line with the demands of their respective populations. If this mechanism is at work in the real world, then the argument presented by Brueckner (1999) is likely to also have some empirical relevance – even if sorting in the real world is not perfect. Nevertheless, the ambiguity of the sign of the growth effect – its dependence on the demands for public goods of different groups in the econ1
omy – must appear as somewhat unsatisfactory.
Brueckner (2006) tackles this problem by using a different theoretical framework. Again, the basic approach is an OLG model as in Brueckner (1999), but following the lead of Yakita (2003), an endogenous growth model is used, which Brueckner extends to include a publicly provided good. Individuals can invest into human capital while being young in order to increase their productivity and earnings when being old. Therefore, they earn higher incomes in their older ages, and correspondingly always have a lower demand for the public good when they are young. Assuming again a perfect Tiebout sorting mechanism and the assumption that a unitary government would supply some intermediate level of the public good, the same intuition as in Brueckner (1999) applies. The decisive difference is that now it can be shown that the young always demand a lower demand for public goods than the old and the steady state capital intensity in a federalist system is higher. Another interesting issue, however, is the question of financing locally supplied public goods. Brueckner deliberately leaves these issues out of consideration by assuming a uniform lump-sum tax, but Zou (1994, 1996) has a closer look at the problems, albeit in a very different theoretical framework. In Zou (1994), the focus is on the effects of federal grants-in aid in the form of matching and non-matching grants on the investment decisions of local governments. Growth effects are not explicitly taken into consideration here, although the accumulation of public capi-
1
Similarly, Kellermann (2007) shows that decentralization of public investment decisions under the presence of head taxes may lead to less or more economic growth.
6 tal that results from local investment decisions obviously must have some relevance for growth. In a framework that involves a government which maximizes present and discounted future utility of a representative citizen, and that also abstracts from private capital accumulation, Zou shows that different schemes of grants-in-aid have very different effects on public sector investment decisions. In particular, only a permanent increase of non-matching grants and a temporary increase of either matching- or non-matching grants are shown to reliably stimulate local public capital accumulation, while other grant schemes involve different ambiguities. Zou (1996) extends the analysis and presents a fully-equipped local growth model with both private and public capital accumulation and consumption, and endogenous decisions on the local level of taxation. Again, however, a government maximizing an infinitely lived, representative citizen’s lifetime utility is at the center of the model – there are no overlapping generations. The somewhat surprising result is that the sign of the effects of different schemes of financing public goods crucially depends on the utility function that is used in the analysis. With a standard utility function that includes only private and public consumption, there are no differential effects from the composition of financing instruments on long-run economic variables. If, on the other hand, an Arrow-Kurz utility function is used, which also includes the public capital stock, then a federal grant that sponsors local public consumption crowds out investments into the local capital stock. Using a local consumption tax to finance the local budget would be associated with faster accumulation of both private and public capital. As we see in the discussion of these models, both the choice between a unitary and a federal government, as well as the choice between different instruments of financing a local budget are characterized by rather ambiguous growth effects. A harmful use of inefficient instruments to finance local budgets can obscure a positive effect on growth that the more fundamental decision to move from a unitary to a federal regime might in fact have. Thus, we have a strong argument here to control for different modes of financing the local budgets in empirical studies relating decentralization to economic growth. 2.2. Agglomeration, Heterogeneity in the Productive Sector and Economic Growth The concept of Tiebout competition between local jurisdictions (Tiebout 1956) has at its centre a notion of efficiency in the consumption of public goods. Sorting of individuals into jurisdictions along the lines of their heterogeneous preferences for governmental activity is presented as a welfare-maximizing process. While this is certainly a laudable property of fiscal federalism, a link from here to economic growth has so far only been established in the literature via the detour of
7 individual saving decisions, as we have seen in the above subsection. It would be much more straightforward to establish such a link if heterogeneity would also be established among individuals and businesses as producers of private goods, which is quite likely an empirically relevant case. In fact, there is a large body of literature in the realm of the New Economic Geography (see Krugman 1991 and Baldwin et al. 2003 for surveys) that lays the theoretical foundations for explaining the spatial allocation of productive activities and that accounts for agglomeration effects. Among the most important reasons for agglomeration effects are economies of scale, which are traded off against transportation costs. Regional immobility of resources is another important feature that has an influence on the spatial allocation of productive facilities in such models, which often predict core-periphery patterns of productive activity (Ottaviano and Thisse 2004). Another important mechanism that drives agglomeration is the existence of different types of knowledge spillovers, which lead businesses that use similar immaterial resources to allocate in close proximity to each other, with Silicon Valley being probably the most prominent example (for a survey, see Döring and Schnellenbach 2006). We can expect these forces of agglomeration to have growth effects by themselves. Countries that attract centers of productive activity benefit from a relatively faster accumulation of physical capital (Baldwin and Martin 2004). Furthermore, agglomeration forces may be at least partially influenced by regional policies, given the fact that the regional conditions that drive agglomeration are not only the result of pure chance (e.g. the presence of natural resources), but more importantly are found in the presence of specific human capital (Camagni 1995) whose accumulation can be influenced by educational and other policies. In this sense, regional policies will have an impact on the speed of capital accumulation, and also on the quality of capital that is accumulated. In an important paper which links a standard notion of fiscal competition with agglomeration effects, Justman et al. (2002) show that regional politicians have an incentive to differentiate the supplies of public infrastructure in different regions, in order to alleviate the pressures of fiscal competition. As a result, these regions will also attract different types of private capital. Justman et al. thus offer a politico-economic rationale for the emergence of regional heterogeneity in productive activities that does not depend on exogenous differences between jurisdictions. In other important contributions linking fiscal competition to agglomeration, Baldwin and Krugman (2004) analyze core-periphery models that lead to full agglomeration of the mobile factor in one region, and show that the resulting existence of an agglomeration rent for the mobile factor implies that further economic integration (e.g. through declining trade costs) will be followed by rising instead of declining tax rates on the income of the
8 mobile factor. Borck and Pfüger (2006) generalize this result to a model economy with only partial agglomeration. From a growth-oriented perspective, the presence of strong agglomeration effects implies that peripheral regions that wish to develop have little other policy alternatives than to attract businesses with a suitable fiscal policy (Brakman et al. 2002). Given that agglomeration effects are relevant empirical phenomena also in unitary states this means that political devolution is likely to enhance the opportunities of peripheral regions to initiate catch-up processes. If this is not only a zero-sum transfer of capital from one region to the other, but instead leads to additional capital accumulation, then it would be plausible to assume that federal states are characterized by higher steady-state per capita incomes, and by higher growth rates on the way to the steady state. In this context, it is important to note that on theoretical grounds, transfers from a rich to a poor region will have temporary effects at best in a new economic geography model (Brakman et al. 2006). The reason is that in this class of models a stable equilibrium also serves as an attractor, i.e. the model economy reverts to the initial equilibrium, even if a different one is theoretically conceivable, as long as the departure from the status quo ante is not sufficiently large. This appears to be in line with empirical results from the United Kingdom that transfers have only very small effects on firm location and that these effects decrease when the number of similar firms already present in a region decreases (Devereux et al. 2007). Put differently: when only very few similar firms are already present in a region, then a subsidy needs to be very large in order to attract additional firms to this region. Thus, the perspective for growth in peripheral regions is likely to be brighter when this region can apply its own fiscal policy tailored by itself, compared to a situation in which the same region receives grants from its neighbors or from central government. Becker and Rauscher (2007), although not working within a New Economic Geography framework, but within a standard Barro (1990) growth model, investigate the effects of fiscal competition with quadratic installation and de-installation costs for physical capital, i.e. with costs of relocating the physical capital stock between regions. Furthermore, they assume benevolent governments. While their main contribution is to find optimal capital tax rates, and showing that the impact of installation costs on the optimal rate of taxing capital is ambiguous, they also show that equilibrium may not exist for high installation costs of capital. In this case, the tax rate chosen by welfare maximizing governments can become so high that capital owners want to reduce the capital stock at a rate that conflicts with the smooth consumption paths desired by consumers. In this sense, a certain intensity of tax competition (reflected in sufficiently low installation costs of capital) may be a prerequisite for the existence of a balanced growth path.
9 This leads us to a more general point. If regions are, for whatever reason, heterogeneous – and there are certainly also a number of reasons for heterogeneity that do not rely on agglomeration externalities – then it is quite likely that a regional government is endowed with a greater ability to enact a growth-promoting policy for its region vis-à-vis a central, unitary government. This point has already been informally made by Oates (1993, 1999). Davoodi and Zou (1998) incorporate the idea into a frugal endogenous growth model which they use as a workhorse for their empirical analysis. Similar to Barro (1990), private and public capital enter a Cobb-Douglas production function, but Davoodi and Zou split the public capital stock into three fractions: federal, state and local capital. This specification guarantees that there exist unique and strictly positive growthmaximizing quantities of public goods for each level of government; it is never optimal to centralize (or to decentralize) public good provision completely. However, it is important to realize that this result is driven entirely by the specification of the production function, so that it can at best serve as a reduced-form sketch of much more complex, underlying processes. Xie et al. (1999) formulate a technically more complex model that, however, in essence relies on the same specification of the production function, and yields similar results. But what exactly is happening behind the curtain of these reduced-form characterizations of economic growth in decentralized political systems? Within the framework of benevolent governments, one could argue that even a unitary government must be able to supply regionally differentiated quantities of public goods, as long as these have a limited spatial reach. Even a unitary, benevolent government would thus be able to supply an efficient quantity and type of public good to each region (Besley and Coate 2003). The unique benefit of decentralization must therefore be found elsewhere. An obvious choice would be informational advantages of decentralized governments; even if a unitary government is benevolent, it may find it difficult to obtain reliable information regarding the efficient types and quantities of public goods in regions far away from the centre (Hayek 1948). We will return to information-related arguments at length in the next subsection. Besley and Coate use a different argument and explain the benefits of decentralization with a political economy argument: With a unitary government, pork-barreling and cost-sharing on the central level, voters in the regions have incentives to strategically delegate public good lovers as representatives to the central level – having a unitary government evokes a fiscal commons problem and leads to an inefficient oversupply of public goods. Although an oversupply of public goods can have negative effects on economic growth, a more important causal influence of decentralization on growth, especially within the context of regional heterogeneity in the productive sector, may have to do with the political management of struc-
10 tural change. Recently, Aghion and Howitt (2006) have argued that in particular for countries whose productive sectors are close to the world technology frontier, a high firm turnover yields higher growth rates. This argument relies on a Schumpeterian endogenous growth model, and is backed by empirical evidence. In a recent book, Caballero (2007) has also argued for the importance of creative destruction for economic growth, and hinted at economic policies such as labor market regulations that may slow down processes of restructuring. But how could federalism have an impact on structural change? This question has, to our knowledge, not been subject to research so far. Nevertheless, there may be some relatively straightforward mechanisms at work. It is a well-known stylized fact that industrial policies in developed countries tend to be directed at preserving incumbent industries. For risk-averse politicians, this is a thoroughly rational choice: In times of a structural crisis, the demise of an unsubsidized incumbent industry may be more or less certain, while the emergence of new industries is uncertain as far as their sizes and types are concerned; preserving the incumbent may thus easily turn out to be the dominant strategy. But if the Besley-Coate overprovision argument holds, then we can also expect the public funds that are directed towards structural preservation to be higher under a unitary than under a federal regime. The problem may be aggravated with sufficiently heterogeneous regions in a unitary state. In this case, a single or a few well-organized regions suffering from asymmetrical shocks could engage in Olson-type interest group politics on the central level and effectively lobby for structural preservation through a transfer of tax revenue into this region. Furthermore, fiscally autonomous regions have a larger set of instruments with which they can react to a crisis of incumbent industries. For example, tax incentives can be used to attract new businesses, compensating for the demise of the old. With a unitary government that normally acts under (e.g. constitutional) provisions to levy uniform nation-wide tax rates, this would obviously not be possible. Thus, there is some reason to believe that local and regional fiscal autonomy would not only enable the supply of efficient, region-specific public goods (from a static perspective) but would also be associated with swifter structural change, and thus higher growth rates. 2.3. Political Competition and Economic Growth Closely related to the positive effects of fiscal federalism that we have discussed above – and that principally stem from the heterogeneity of regions – are effects that have their roots in the competition between sub-central governments. This is, in a sense, the other side of the Tiebout mechanism: Individuals do not only sort into regions according to their preferences, but they can also avoid unfavorable combinations of taxes and public good supplies by relocating themselves
11 or at least their mobile capital. They become endowed with additional instruments to discipline self-interested representatives. However, the use of these instruments may also yield effects that deplete economic efficiency. Oates (1972) has already pointed out that efficiency requires fiscal equivalence, i.e. that those who decide on the provision of sub-central public goods are to be identical with those who consume and finance them. There is a wealth of contributions to the literature that analyze how externalities between jurisdictions distort efficiency, and how efficient compensating transfers might correct for these problems. Discussing these horizontal and vertical externalities in detail here would clearly be beyond the scope of this paper (see Wellisch 2000 for a survey of the issues). Few papers have dealt with the externality issue explicitly within the framework of a growth model. An example is Devereux and Mansoorian (1992), who analyze an endogenous growth model of the Barro (1990) type with two countries whose decisions on tax levels produce fiscal externalities. Whether independently chosen tax levels are too high or too low depends on preferences in this model. In any case, coordination of fiscal policies improves welfare – but not necessarily growth, because the decentralized equilibrium may be characterized by too low public consumption and a too high level of public investment. It is therefore not possible to derive clear-cut predictions regarding growth effects of uncoordinated, decentralized policy from this model. In a different two-country endogenous growth model with imperfectly mobile capital, Lejour and Verbon (1997) show that uncoordinated source taxes on capital returns may actually imply too much redistribution. The reason is a growth externality: If one country levies a tax, it reduces investment at home, but by depleting the equilibrium return to capital in the entire economic union, also in the other region. Contrary to conventional wisdom, efficient coordination would then lead to lower public consumption and lower tax rates compared to an uncoordinated equilibrium. Given the lengthy discussion in the literature of different externalities emerging from decentralized policy, it is useful to recall that there is also a lengthy discussion on how to efficiently offset these externalities. Recently, Bucovetsky and Smart (2006) have shown that pragmatically correcting for horizontal fiscal externalities which lead to an inefficiently low tax effort by local jurisdictions may in fact require only a rather simple fiscal equalization grant. Generally, it appears that possible negative effects on efficiency that are associated with fiscal externalities in competitive federalism can be dealt with by imposing an appropriate institutional framework for jurisdictional competition. Even for the issue of economies of scale in the production of public goods, which may not always be exploited by relatively small jurisdictions, one can argue that voluntary contracts between such polities will eventually solve efficiency problems (Blankart 1996).
12 On the other hand, there are a number of unambiguous benefits resulting from fiscal decentralization. A classical argument is that Leviathan tendencies in government are reduced by the pressures of fiscal competition (Brennan and Buchanan 1980). While this may on first sight appear as a pure distributive issue between taxpayers and representatives, a reduction of rent-extraction in government essentially means that the costs per unit of public good are reduced from the perspective of the taxpayers. This can result in a smaller distortionary tax burden, or in a larger quantity of productive public goods being supplied, which would also accelerate growth (Aschauer 1989). Having fiscal competition as a mechanism to restrict self-interested representatives may be particularly useful if formal constitutional rules are difficult to enforce (Schnellenbach 2004). In this case, federalism can even be thought of as a “market-preserving” institutional framework (Weingast 1995). Given that mobile factors have the option to leave, it becomes difficult for governments to severely intervene into private property rights, and large-scale interventions into the market process become more unlikely.
2
A related and more formal point has been made by Edwards (2005), who models a time-inconsistency problem. In his neoclassical growth model, human capital investment drives growth. But a unitary government cannot commit to low tax rates in the future, and accordingly, the unitary state is characterized by high taxes, low human capital and low growth. Local governments, however, are exposed to the threat of out-migration of fiscally expropriated factors. The exit option helps to solve the time-inconsistency problem, and the decentralized equilibrium is characterized by low taxes, high investment and high growth rates. Madiès and Ventelou (2004) analyze a Leviathan setting, and set up a model which allows them to analyze the trade-off between introducing vertical fiscal externalities, which result from different levels of government sharing the same tax base, and having a more targeted supply of public goods (education) on the local level. They show that under reasonable assumptions, the public good effect over-compensates the inefficiencies introduced by the vertical externalities, which yields positive overall growth effects. An issue that is still very much contested is that of the influence of fiscal federalism and fiscal competition on corruption. The question is of interest here, because corruption probably exerts a negative effect on the efficiency of government. If it increases the costs of providing productive public services, then it is also likely to be associated with negative growth effects. Cai and Treis-
2
Rauscher (2005) assumes a Leviathan government analyzing tax competition in an endogenous growth model with a productive public input. A taming of Leviathan governments by more intense inter-jurisdictional competition induces higher economic growth if the Leviathan’s elasticity of intertemporal substitution, which is also the elasticity of substitution between rents and political support, is not substantially larger than one, i.e., if current and future utility or rents and political support are bad substitutes.
13 man (2004) even go so far as to predict a “state corroding federalism” to be the result of fiscal competition. According to their model, local governments may decide to shield business located under their jurisdiction from central-level law enforcement, which effectively erodes the rule of law. In their case study for China, they argue that local governments also shielded local businesses from investigations related to corruption. This extends an argument made by Bardhan and Mookherjee (2000) as well as Blanchard and Shleifer (2001), stating that local governments may be more easily captured by special interest groups. Martinez-Vazquez and McNab (2003) survey the literature on this issue more thoroughly. These arguments can, however, be contested on the grounds that fiscal competition will generally reduce the scope for redistributive special interest politics, since mobile tax bases can leave when they perceive to be exploited. Furthermore, already Riker (1964) stated that simple institutions such as a strong national party system may countervail unfavorable tendencies towards local capture. Again, it is thus possible to compensate for the problems of federalism by choosing an appropriate institutional framework. Finally, there is another politico-economic avenue via which federalism may have an impact on growth, namely that of fostering political innovation. Seeing the relevance of this requires a step back from the neoclassical standard model, where we got used to have a clear benchmark for an optimal policy and the question is only under which conditions we can expect the optimal policy to be executed. In a world with model uncertainty, in which neither the policymaker nor a consulting economist knows the true, underlying model of the economy, the task of the policymaker is to learn from mistakes and to gradually improve his politics in a long-term process of trial-anderror. Oates (1990, 1999) has hinted at the fact that federalism may be useful in this respect, by speaking of “laboratory federalism” – a system where many, parallel small-scale experiments can be undertaken at the sub-central level. Besley and Case (1995) as well as Salmon (1987) have argued for the relevance of yardstick competition, a mechanism where voters use political results in neighboring jurisdictions to assess the ability of their own representative. Blankart (2007) argues in favor of federalism as a source of political (and other) creativity, and Inman and Rubinfeld (1997) make the point that the US welfare reform of 1996 was a case of laboratory federalism. Again, the argument is not uncontested. Rose-Ackerman (1980) argued that information resulting from political experiments is a pure public good. Given that such experiments are risky, and that failure may lead to a denial of reelection, sufficiently risk-averse representatives will abstain from conducting experiments themselves and instead attempt to free-ride. Kotsogiannis and Schwager (2006) argue that self-interested representatives can even use policy innovations to increase their scope for extracting rents from office, because voters are uncertain about what could have been
14 achieved with a different policy. As far as free-riding is concerned, Strumpf (2002) shows that the argument depends on the degree of heterogeneity between regions. As soon as regions become sufficiently heterogeneous, the learning externality loses relevance, and each region has to experiment to learn about policies being implemented under its specific conditions. Finally, Schnellenbach (2007) shows that for rationally ignorant voters factor migration is necessary to disturb a given equilibrium distribution of policy-related beliefs in the electorate – changing factor prices as a result of migration induce collective learning processes that would not occur otherwise. Only very few papers attempt to incorporate political innovation into growth models. An example is Rauscher (2007), who analyzes an endogenous growth model with Leviathan governments. In this model, tax competition leads to a reduced frequency of political innovation and slower growth for reasonable parameter values. In a sense, this is a standard tax competition result carried into a dynamic model: Standard tax competition models predict an undersupply of public goods due to the increased opportunity cost of public funds, and in the model discussed here the public good is innovation undertaken by the government. The politico-economic arguments that hint into the opposite direction are neglected in this framework. Maybe even more importantly, in order to keep the model tractable, Rauscher (2007) assumes that the private capital stock is constant and that only public capital accumulation drives growth, which is a very unrealistic assumption. Rauscher (2006) allows also for private sector capital accumulation, and reaches the opposite result: Increased mobility of tax bases now yields an increased frequency of political innovation and also higher growth rates. 3.
The Results of Previous Empirical Work
The survey of the theoretical analyses on the impact of fiscal federalism on economic growth helps to identify several transmission channels for such a relation. First, a decentralization of public good provision and its financing allows for a consideration of heterogeneous demand for public goods, and may thus positively affect economic growth (Heterogeneity Channel). Second, tax competition between regions restricts Leviathan governments in the exploitation of mobile tax bases and keeps government interventions at a low scale such that private initiative could fully display its usefulness for economic development (Market-Preservation Channel). Third, given the presence of agglomeration economies and knowledge spillovers, regional fiscal policies do not have much leverage at all. But in particular tax competition is providing for the means to successfully attract businesses and adapt to structural change (Structural Change Channel). Fourth, policy innovation that is induced by fiscal competition may play a role for economic growth in particular
15 in a situation of dynamic structural change when creativity and willingness for experimentation are necessary (Political Innovation Channel). The existing empirical evidence does not test these transmission channels explicitly, but addresses the question from a different perspective. On the one hand, federalism or decentralization may be favorable for economic development and structural change in a country. Then, a top-down perspective is adopted, and the question becomes which role the lower level governments perform for the economic development of a country. On the other hand, it appears important to know which type of internal arrangement of a country favors regional development. Thus, a bottom-up perspective is assumed by focusing on lower level jurisdictions, regions and agglomerations. Following such a distinction in cross country and single country studies, it is however possible to highlight on which transmission channels the empirical studies mainly focus. 3.1 Cross-Country Studies The majority of the cross country studies interprets fiscal federalism as decentralized organization of government activities and measures decentralization by the fraction of sub-federal spending from total government spending. This approach is problematic as theoretical analyses presume autonomy of sub-federal decision-making on provision and financing of public goods and spending decentralization might simply indicate the extent of administrative federalism with states, provinces or cantons providing public services according to federal mandates and financed by the federal government (Treisman 2002, Rodden 2004, Stegarescu 2005). As long as fiscal transfers from other jurisdictions are not controlled for, the estimates for spending decentralization may be biased. Using spending decentralization as a measure for fiscal federalism under these circumstances would then mainly allow for testing the Tiebout channel on economic growth. Given the measurement problems, it is unsurprising that the authors of the cross-country studies on the impact of federalism on economic development end up with controversial results (see Table 1). Davoodi and Zou (1998), for instance, find a weakly significant negative correlation between the degree of fiscal federalism and the average growth rate of GDP per capita for a sample of 46 countries and the period from 1970 to 1989. This effect is not significant for the sub-sample of developed countries. The negative influence for developing countries is however robust, though only weakly significant as well. According to these estimates, an additional decentralization of functions by 10 percent reduces the growth of real GDP per capita in developing countries by 0.7 – 0.8 percentage points. Woller and Philipps (1998) do not report a robust relation between economic growth and decentralization, using a sample with a lower number of de-
16 veloping countries and a shorter period. Also, they analyze, in addition to the average growth rates for five years, the annual growth rates in a panel. Both studies use fixed-effects models. In contrast to Davoodi and Zou, Woller and Philipps consider a common time trend by including a trend variable. Iimi (2005) uses the most recent data for 51 countries – average growth between 1997 and 2001 – and applies an instrumental variable technique. Spending decentralization turns out to be highly significant such that a 10 percent higher decentralization of spending increases growth of real GDP per capita by 0.6 percentage points.
17 Table 1: Study
Empirical studies on the influence of fiscal decentralization or federalism on economic growth in cross-country studies Countries
Period
Method
Main results
Fixed Effects Model, Time Dummies, Unbalanced Panel
10% higher decentralization of spending reduces growth of real GDP per capita in developing countries by 0.7-0.8%-points (10% significance level)
Woller and Philipps (1998)
Fixed Effects 23 Developing Coun- 1974-1991 Model, OLS tries three and five year averages and annual data
No robust significant effect of the decentralization of spending or revenue on growth of real GDP per capita
Yilmaz (2000)
17 Unitary States,
Decentralization of expenditures at the local level increases growth of real GDP per capita in unitary states more than in federal countries. Decentralization at the regional level is not significant
Ebel and Yilmaz (2002)
6 Transition Countries
1997-1999
Eller
22 OECD Countries
Fixed Effects, Decentralization is positively related to 1972-1996, economic growth annual and four Time Dummies year averages
Davoodi and 46 Developing and 1970-1989 Zou (1998) Developed Countries five and ten year averages
1971-1990
13 Federal Countries, annual data Newly Industrialized Countries and Developed Countries
(2004)
Fixed Effects Models, Time Dummies, GLS
Bivariate OLS Decentralization is in general positively related to economic growth
Enikolopov and Zhuravskaya (2003)
21 Developed and 70 Cross-section of OLS, 2SLS Developing and the averages Transition Countries 1975-2000
10% higher decentralization of revenue reduces growth of real GDP per capita in developing countries by 0.14%-points (5% significance level)
Thießen (2003)
21 Developed Countries
Cross-section of OLS the averages of 1973-1998
Decentralization of spending by 10% increases growth of real GDP per capita by 0.15%-points (5% significance level), quadratic term is significantly negative
Thießen (2003a)
26 Developed Countries
Panel data 1981-1995
Decentralization of spending by 10% increases growth of real GDP per capita by 0.12%-points (5% significance level).
Iimi (2005)
51 Developing and Cross-section of OLS, IV Developed Countries the average of 1997 to 2001
Feld, Baskaran and Dede (2004)
19 OECD countries
Panel data 1973-1998
Bodman and Ford (2006)
18 OECD Countries
Cross-section of OLS 1996 and Panel data 1981-1998
GLS
10% higher decentralization of spending increases growth of real GDP per capita by 0.6%-points (1% significance level)
Fixed Effects, No robust effect of spending or revenue decentralization, but a significantly negaTime Dumtive effect of stronger participation in mies revenue sharing arrangements No significant effect of revenue or spending decentralization on economic growth
Source: Own compilation.
Additionally considering institutional aspects, Enikolopov and Zhuravskaya (2003) present evidence for average economic growth of the past 25 years in a cross-section of 91 countries that the effects of fiscal decentralization largely depend on the structure of the party system as well as
18 on the degree of „subordination“ of sub-national levels. According to these results, the age of the most important political parties is favorable to the positive effects of decentralization on economic growth particularly in developing and transition countries. In countries with a – in this respect – weaker party system a 10 percent higher decentralization of revenue decreases the growth of real GDP per capita in developing countries by 0.14 percentage points. These results challenge those by Martinez-Vazquez and McNab (2002) according to which the decentralization of revenue significantly reduces growth of real GDP per capita of developed countries, but not of developing and transition countries. Yilmaz (2000) analyses the different effects of fiscal decentralization in 17 unitary and 13 federal states for the period 1971-1990 with annual data. Decentralization of expenditures to the local level increases the growth of real GDP per capita in unitary states more than in federal countries. However, decentralization to the regional level in federal countries is not significant. Thießen (2003) analyses, similar to Enikolopov and Zhuravskaya (2003), the average growth rates of real GDP per capita for a cross-section of 21 developed countries in the period 1973-1998 and in a companion study (Thießen 2003a) for a panel of 26 countries and the period 1981 to 1995. According to his estimates a 10% stronger decentralization of expenditures increases the growth of real GDP per capita by 0.12-0.15%-points in high-income countries. But the relation between federalism and economic growth might be non-linear as a quadratic term of expenditure decentralization is significantly negative. His results are corroborated by Eller (2004). Feld, Baskaran and Dede (2004) and Feld (2007) use new data provided by Stegarescu (2005) that measure the importance of sub-federal tax autonomy by the extent to which sub-central governments can actually set tax rates or bases. The three decentralization variables, the fraction of subcentral spending from total spending, the fraction of sub-central revenue from total revenue stemming from taxes autonomously decided by state and local governments and the fraction of sub-central revenue from total revenue stemming from revenue sharing systems, do not have a robust effect on real GDP growth per capita in a panel of 19 OECD countries between 1973 and 1998. While expenditure decentralization is significantly negative in the pooled regressions, it is not robust to the inclusion of fixed country effects. Employing a two stage method in which the fixed effects from a first stage regression are explained by fiscal decentralization measures on a second stage does not yield robust results. These disappointing results are basically supported by Bodman and Ford (2006) who do however not use modern panel data methods. Although further analyses are necessary to test the robustness of these results, the impact of fiscal decentralization on economic growth, in particular with respect to the Tiebout channel, remains open.
19 Table 2:
Empirical studies on the influence of fiscal decentralization or federalism on economic growth in China, Russia, Ukraine and India
Study
Countries
Period
Zhang and Zou (1998)
28 Chinese Provinces 1987-1993 Fxed Effect Models Annual Data without Time Dummies
Decentralization of expenditure to the provinces reduces growth of real GDP per capita
Lin and Liu (2000)
28 Chinese Provinces 1970-1993 Fixed Effect Models, Annual Data Time Dummies
Revenue decentralization by 10% increases growth of real GDP per capita by 2.7%-points (5% significance level)
Jin, Qian and Weingast (2005)
29 Chinese Provinces 1982-1992 Fixed Effect Models, Annual Data Time Dummies
Expenditure decentralization by 10% increases growth of real GDP per capita by 1.6%-points (10% significance level)
Qiao, Martinez Vazquez and Yu (2002)
28 Chinese Provinces 1985-1998
2SLS with Pooled Data
Expenditure decentralization increases growth of nominal GDP per capita significantly (5% significance level)
Feltenstein and Iwata (2005)
Central Level in China
VAR with Timeseries data
Fiscal decentralization has adverse implications for macroeconomic stability but tends to increase growth
Jin and Zou (2005)
30 Chinese Provinces 1979-1999
Fixed Effects with Corrected Standard Errors
Divergence between local expenditures and revenue (i.e. centralization) increases growth
Zhang and Zou (2001)
29 Chinese Provinces 1987-1993, annual data
OLS, Fixed Effects
Decentralization reduces economic growth
Zhang and Zou (2001)
16 Indian States
1970-1994
OLS
Decentralization increases economic growth
Desai, Freinkman and Goldberg (2003)
80 Russian Regions
1996-1999
OLS with panel corrected standard errors, 3SLS
Decentralization has a positive but non-linear effect on growth
Naumets (2003)
24 Ukrainian Oblasts and Autonomous Republic of Crimea
1998-2000
Fixed-Effects and Random Effects Models
Not robust negative impact of own revenue decentralization on growth of real gross value added
1952-1996
Method
Main results
Source: Own compilation.
3.2 Single Country Studies The empirical results concerning the impact of decentralization on economic growth for individual countries are no less ambiguous. Analyses have been conducted for China, the Ukraine, India, Russia, the U.S., Spain, Switzerland and Germany (see Tables 2 and 3). Zhang and Zou (1998, 2001) report a significantly negative effect of expenditure decentralization on economic growth in 28 (29) Chinese provinces, using annual data between 1987 and 1993. Jin, Qian and Weingast (1999), however, find a weekly significant positive effect of expenditure decentralization on the economic growth of almost the same sample of Chinese provinces over time. The most important difference between the studies – aside the small differences in the explanatory variables used
20 – consists in the fact that Zhang and Zou do not use time dummies. Consequently, the common positive and negative economic shocks in China are inadequately included as compared to Jin et al. Qiao, Martinez-Vazquez and Xu (2002) report similarly positive growth results for expenditure decentralization even without any fixed effects. Lin and Liu (2000) corroborate the result of a positive impact of decentralization on economic growth in Chinese provinces for the period 1970 to 1993 also for the revenue side. Moreover, a higher responsibility of public budgets at the provincial level is associated with increased economic growth. These authors, too, use time dummies in addition to cross-section fixed effects. Jin and Zou (2005) present evidence that a higher divergence between local expenditure and revenue increases growth. The relevance for the estimates of using time dummies points to the strong economic dynamics in China. The sometimes enormously high Chinese growth rates cannot be exclusively covered by structural variables such that dummy variables for the individual years are necessary for specifying the model. The fact that Zhang and Zou neglect them must be interpreted as a mis-specification of the model. Thus, for China, decentralization of government activity has rather a positive impact on economic growth. This assessment is corroborated by the time series analysis by Feltenstein and Iwata (2005). Desai, Freinkman and Goldberg (2003) are in line with the Jin, Qian and Weingast (2005) approach of focusing on revenue retention in transition countries. For 80 Russian regions between 1996 and 1999, they report a significant positive effect of revenue retention rates on growth of gross regional products. In addition, a conditional effect of natural resources and budgetary transfers on the relation between retention rates and cumulative output growth occurs. The effect of retention rates on growth switches from positive to negative when transfers cover more than 45 percent of total revenues, while this effect declines in magnitude, but remains positive when regions become more resource abundant. Similarly, Zhang and Zou (2001) report a positive decentralization effect for India. These results provide some support for the Market Preservation Thesis. It does however not generally hold for transition countries. Naumets (2003) finds a negative, though not robust impact of the share of own revenue from consolidated regional revenue on growth of real gross value added in a panel of 24 Ukrainian regions from 1998 to 2000. Much the same holds for individual developed countries. Exploring American economic development between 1790 and 1840, Wallis (1999) argues that fiscal federalism was an important institutional precondition that fostered economic growth of the U.S. In a time-series analysis for the whole of the USA from 1951 to 1992, Xie, Zou and Davoodi (1999) claim that the U.S. find themselves in a decentralization equilibrium because differences in decentralization at the state level or at the local level do not have statistically significant effects on real GDP growth. Akai and
21 Sakata (2002), however, offer evidence to the contrary for U.S. states. Taking into account additional explanatory variables and various indicators for the degree of fiscal federalism, they underline the positive influence on economic growth. If expenditure decentralization increases by 10 percent, then the growth of GDP per capita increases by 1.6-3.2 percentage points. However, decentralization on the revenue side and indicators for fiscal autonomy of sub-national levels, measured by the share of own revenue in total revenue, do not have any significant impact. Stansel (2005) develops a different approach by testing the impact of local fragmentation on growth of local real per capita money income. This idea is related to the fragmentation hypothesis by Brennan and Buchanan (1980) who argue that a higher fragmentation of a polity into different jurisdictions increases the intensity of inter-jurisdictional competition and thus restricts Leviathan governments. Hence, Stansel (2005) indirectly supports the Market Preservation Hypothesis. Table 3:
Empirical studies on the influence of fiscal decentralization or federalism on economic growth in the U.S., Germany, Spain and Switzerland
Study
Countries
Period
Method
Main results
Xie, Zou and Davoodi (1999)
Central Level in 1951-1992 the USA
Akai and Sakata (2002)
50 US States
1992-1996, Cross- OLS and Fixed EfSection of Aver- fects Model, Time Dummies age Growth Rates, Panel with Annual Data
Expenditure decentralization by 10% increases growth of GDP per capita by 1.6-3.2%-points (robust 10% significance levels)
Stansel (2005)
314 US Metropolitan Areas
1960-1990
Robust OLS
Higher fragmentation is associated with significantly higher growth in (log) real per capita money income
Berthold, Drews and Thode (2001)
16 Laender
1991-1998
Fixed Effect Model without Time Dummies
Higher horizontal and vertical grants significantly reduce growth of nominal GDP per capita
Behnisch, Büttner and Stegarescu (2002)
Central Level in 1950-1990 Germany
Time Series Analysis
Increase of federal share of expenditure in total expenditure has positive effect on German productivity growth
Time Series Analysis, No significant impact of expenditure OLS decentralization on growth of real GDP per capita
Gil-Serrate and 17 Spanish Lopez-Laborda Autonomous Communities (2006)
1984-1995
Fixed and Random Effects, Time trend
Revenue control decentralization has a positive effect on decentralization
26 Swiss Cantons
1980-1998
OLS, 2SLS
Tax autonomy and tax competition are not harmful for economic growth
Feld, Kirchgässner, and Schaltegger (2004, 2005)
Source: Own compilation.
Three studies have also been conducted for Germany. Berthold, Drews and Thode (2001) analyze the effects of horizontal fiscal equalization between states and of supplementary federal grants on regional economic development of the 16 Laender in a panel analysis with annual data from 1991 to 1998. According to their estimates higher grants in horizontal and vertical fiscal
22 relations reduce the growth of nominal GDP per capita of the Laender significantly. However, these econometric results suffer from severe endogeneity problems as slowly growing Laender may receive higher grants. Berthold and Fricke (2007) thus update their study for the more recent years until 2003 and employ an instrumental variable technique. As instruments they use the GDP level, the unemployment and employment rates, the fraction of people receiving social assistance and further variables. Unfortunately, they do not report tests on the validity of the instruments or on over-identification. Their selected list of instrumental variables casts some doubts on the validity of the instruments however as they appear to be correlated with the dependent variable and would thus not satisfy the orthogonality condition. If the instruments were valid, this evidence would partly support the Structural Change Thesis as higher grants would apparently provide incentives to adopt structural change more slowly. Behnisch, Büttner and Stegarescu (2002) indeed contradict these results as they report a positive effect of increasing federal activities – measured by the share of expenditure at the federal level – on German growth of productivity in a time series analysis from 1950 to 1990. In a study for Switzerland, Feld, Kirchgässner and Schaltegger (2004, 2005) analyze the impact of tax competition, fragmentation and grants on economic performance more explicitly. Controlling for expenditure decentralization in a panel of 26 cantons between 1980 and 1998, a higher intensity of tax competition exerts a significantly positive impact on cantonal labor productivity. The stronger a canton finds itself in tax competition, the higher cantonal economic performance. Fragmentation of a canton in its communities does not have any robust effect on labor productivity. The estimation results for vertical matching grants suffer from endogeneity problems and may thus be biased. The effects of tax competition and fragmentation are however not affected by the inclusion of grants. Gil-Serrate and Lopez-Laborda (2006) report evidence pointing in the same direction for Spain. Despite some differences between the studies, the results for the U.S., Spain and Switzerland lend some support for a positive effect of competitive federalism on economic development. While the U.S. studies unequivocally support the Tiebout Thesis and provide only indirect support for the Market Preservation Thesis, the Swiss results are rather in line with the latter. A rigorous analysis of German cooperative federalism is still terribly needed. 4.
Quantitative Literature Review
In the next two sections, this literature on decentralization and economic growth is more thoroughly discussed in a quantitative review. We start with some descriptive statistics on the studies that have been hitherto conducted. The database consists of 26 studies, listed in Tables 1 to 3,
23 which provide point estimates on the effect of a measure of decentralization on a measure of economic growth. 4.1. Descriptive Statistics on Studies The studies in the data set are heterogeneous in many dimensions. We coded the characteristics of an estimated model as dummy and continuous variables. There is of course some subjectivity in classifying idiosyncratic characteristics into distinct groups. For instance, we differentiate countries into two groups only, developed and developing, and thus classifying Transition countries as developing countries. Even though this particular classification might be criticized, inherent subjectivity is unavoidable in both meta-analyses and traditional literature reviews. The major advantage of a meta-analysis compared to traditional reviews is that the subjective judgments of the reviewer can be made explicit. Figure 1: Subject of Studies
Source: Own calculation.
First, both single country and cross-country studies have been conducted. Within each of the two groups, studies can be further subdivided according to whether they consider developed or developing countries. Within the subgroup of single country studies, a further differentiation according to the particular country considered is possible. The majority of studies are single country studies; they number 15 while 11 cross-country studies have been conducted. According to Figure 1 an equal number (i.e. eleven) of studies consider either developed or developing countries, only four consider both simultaneously.
24 Table 4 provides cross-tabulations between single- and cross-country studies and studies on developed and developing countries. Out of the eleven cross-country studies, two are exclusively on countries from the developing world, five are exclusively on countries from the developed world, and four consider countries from the developing and the developed world simultaneously. Out of the fifteen single-country studies, nine cover developing and six developed countries. Table 4:
Cross-Tabulations of Study Type (Single- or Cross-Country) and Study Subject (Developed, Developing or Both Types of Countries) Cross-country
Developing
No
Yes
Total
No
6
9
15
Yes
9
2
11
Total
15
11
26
Developed
No
Yes
Total
No
9
6
15
Yes
6
5
11
Total
15
11
26
Both
No
Yes
Total
No
15
7
22
Yes
0
4
4
Total
15
11
26
Source: Own calculation.
There are eleven studies (two cross-country and nine single-country analyses) which exclusively focus on developing countries. The majority of these studies (seven) are single-country studies for China. There are also eleven studies that exclusively focus developed countries. Five of them are cross-country studies on OECD countries, three have been conducted for the United States, one for Germany and Switzerland, and three for other countries (see Figure 2). There is also a large variation on the publication status of the studies in the database. According to Figure 3, the majority of studies, thirteen, are journal articles. Six studies are described by their authors as unpublished manuscripts, two as either working or discussion papers, two are masters or PhD. thesis, and one is a book chapter.
25 Figure 2: Countries Considered in Studies
Source: Own calculation. Figure 3: Publication Status of Studies
Source: Own calculation.
4.2. Summary Statistics on Estimated Models In each of the studies, a number of models are estimated. These regressions result in several point estimates of the effect of decentralization on economic growth. Since we will focus on
26 these point estimates in the meta-regressions, we provide separate summary statistics for them. First, according to Figure 4 most of the estimates, about 55%, in our sample derive from journal articles, about 30% are obtained from unpublished manuscripts, while all other publication types together contribute about 20% to the meta-analysis. Figure 4: Publication Status of Models
Source: Own calculation.
Table 5 presents cross-tabulations on the point estimates between study type and study subject. About 50% of estimates derive from cross-country models. The majority of these cross-country models focus exclusively on developed countries. Out of the 50% of point estimates from singlecountry studies, the majority focus on developing countries. Only about 9% of cross-country models use data from both developed and developing countries simultaneously. According to Table 5 and Figure 5, there are however significantly more estimates from studies on developed than on developing economies when taken together.
27 Figure 5: Subject of Estimated Models
Source: Own calculation. Table 5:
Cross-Tabulations of Model Type (Single- or Cross-Country) and Model Subject (Developed, Developing or Both Types of Countries) Cross-country
Developing
No
Yes
Total
No
18.4
41.3
59.7
Yes
30.9
9.3
40.3
Total
49.3
50.7
100
Developed
No
Yes
Total
No
30.9
13.9
44.8
Yes
18.4
36.8
55.2
Total
49.3
50.7
100
Both
No
Yes
Total
No
49.3
46.1
95.5
Yes
0
4.5
4.5
Total
49.3
50.7
100
Source: Own calculation.
The point estimates can be further differentiated by the particular models specified. Figure 6 provides information on whether the dependent variable and the decentralization variable have been specified in level or log form. A large majority of point estimates originate from models specified in the level-level form, other specifications are quite rare, and not a single model was estimated with a specification that facilitates an interpretation of the effect as elasticity.
28 Figure 6: Model Specification
Source: Own calculation. Figure 7: Type of Data
Source: Own calculation.
In Figure 7, we collect information on the type of data used. The majority of estimates use panel data. However, there is also a significant number of estimates deriving from cross-section models. Time-series models, on the other hand, are quite rare.
29 Table 6: Cross Tabulations Regarding the Use of Fixed Cross-Section and Time-Effects (in %) Fixed Effects Time Effects
No
Yes
Total
No
45.9
24.0
69.9
Yes
7.7
22.4
30.1
Total
53.6
46.4
100
Source: Own calculation.
Figure 8 and Tables 6 and 7 provide further information on the specification of estimated models. Almost half of the point estimates derive from models where fixed effects for cross-section units have been included. A smaller number of estimates derive from models where time fixedeffects are included. About 22.4% of all estimated models use both cross-section and time effects. About 17% of models take nonlinearities, notably interaction and quadratic terms, into account. However, none use both simultaneously, i.e. about 10% of models use only quadratic terms, and about 7% exclusively interaction effects. Only a very small number control for potential endogeneity (other than that deriving from unobserved heterogeneity for which it is possible to control by including fixed effects). Table 7: Cross Tabulations of Nonlinear Terms (in %) Interaction Quadratic
No
Yes
Total
No
83.2
6.9
90.1
Yes
9.9
0
9.9
Total
93.1
6.9
100
Source: Own calculation.
Several measures of decentralization have been used in the literature, for instance expenditure and revenue shares, the divergence between central and sub-national government spending or revenue, and the tax autonomy of sub-national governments. Figure 9 provides information on the relative frequency of these measures in our sample. According to this figure, the expenditure share of sub-national governments or closely related measures is used as the decentralization variable in about 35% of models, tax autonomy is used in about 25% of models, and each of the remaining measures, except the weighted average of expenditure and revenue decentralization, is used in about 10% of all estimated models. The distribution of the estimated coefficients is depicted in Figure 10, and summary statistics are provided in Table 8. Since the estimated coefficient is a measure with a dimension, the spread originating simply due to the use of particular units can be substantial. Therefore, it is not surprising that the minimum value of the coefficient is -3623 while the maximum is +5391 (see Table 8).
30 In order to maintain some informative content in the histogram, we have excluded 15 extreme outliers, and included all coefficients with an absolute value of less than 10. Figure 8: Model Specification
Source: Own calculation. Figure 9: Decentralization Measure
Source: Own calculation.
31 According to Figure 10, the estimated coefficients seem to be dispersed around 0, and only a few estimates exhibit large negative values. The majority of observations has an absolute value smaller than 2. In generating the histogram, we had to exclude one estimate from Zhang and Zou (2001) because the coefficient and the t-statistic are mutually inconsistent. Both the t-statistic and the coefficient from this observation are set to missing, but values of variables that do not indicate problems are still used for calculating descriptive statistics. Hence, the number of observations in calculating the mean and standard deviations (including the outliers) of the coefficients in Table 8 is 374 while it is 375 for the remaining variables. Figure 10: Estimated Coefficients
Source: Own calculation.
Figure 11 presents a histogram on the t-statistics. This histogram does not exclude outliers. The number of observations is nevertheless only 346, i.e. less than the full sample of 375. This is due to the fact that most authors (except Iimi 2005) do not provide separate t-statistics for nonlinear terms at their average value. Nor do they provide the variance-covariance matrix of the estimated coefficients, prohibiting us from calculating them on our own. Therefore, we have to exclude these point-estimates, in addition to the estimate from Zhang and Zou (2001). The varying number of observations only affects the descriptive statistics, but not the meta-regressions where observations are immediately discarded if the value for one variable is missing.
32 Table 8: Summary Statistics of Continuous Variables Variable
N
Mean
SD
Min
Max
t-Stat
346
0.158
2.73
-23
22.32
Coefficient
374
15.54
370.06
-3623.8
5391.1
Coefficient without Outliers
359
-0.13
0.83
-8.84
2.43
Control Variables
375
7.42
2.82
1
16
Observations
375
184.73
168.49
14
534
Source: Own calculation.
As the histogram on the estimated coefficients, the t-statistic histogram is centered around zero. However, the spread is larger, and there is a significant number of observations with absolute tvalues larger than 2. Figure 11: Estimated t-Statistics
Source: Own calculation.
In Figure 12, we provide a histogram on the number of control variables other than the constant or country and time fixed effects. On average, a model has about 7 to 8 controls. However, the dataset also contains both estimates from bivariate regressions and from models with a more extensive list of controls. In Figure 13, we depict a histogram on the number of observations indicating that a large number of models have less than 100 observations. However, there are also models with more than 300 observations. On average, a study has about 180 observations
33 Figure 12: Number of control variables
Source: Own calculation. Figure 13: Number of observations
Source: Own calculation.
5.
Meta-Analysis
The aim of a meta-analysis is to explore the true size of an effect and to evaluate its significance. In the context of decentralization and economic growth, the idea is to pool the estimates from several studies and to calculate an average effect. Since in economics most studies provide several
34 estimates of the effect, researchers usually choose one out of these, either the estimate preferred by the original authors (the so called “preferred estimate”), or an estimate according to some rule, for instance the median estimate (see for example Rose and Stanley 2005). We focus on the minimum, maximum and the median estimate of each study. There are a number of methods for summarizing results of multiple independent tests. One widely used method for inquiring whether the accumulated evidence suggests that a significant effect exists, is Fisher’s combined test (Jarell and Stanley 2004). It uses the fact that the p-values of individual tests are normally distributed under the null hypothesis of no effect. Since the p-values are normally distributed, the f-statistic, calculated as the sum of the logarithms of the p values over L estimated models times minus two, has approximately a χ 2 - distribution with 2L degrees of freedom, i.e. L
f = −2∑ ln Pi ~ χ 2 (2L)
(1)
i =1
We calculate this statistic for all three sorts of estimates, the minimum, maximum and median from each of the 26 studies. For the minimum estimates, we obtain an f-value of
f min = 52.03 < 62.83 = χ 0.95 (46) , i.e. the null hypothesis of no significant effect is not rejected 2
at the five-percent level. For the median estimate, we obtain f med = 121.04 > 65.17 = χ 0.95 (48) , 2
and
a
clear
rejection
of
the
null.
For
the
maximum
estimates,
we
obtain
f max = 230.64 < 65.17 = χ 0.95 (48) , and as expected an even clearer rejection of the null than 2
for the median estimates (see also Table 9). Overall, the results seem to suggest that decentraliza3
tion has in general a significant effect on decentralization.
Note that the degrees of freedom vary between the minimum and the two other estimates, and none are equal to the number of studies. This is due to the fact that some studies do not report appropriate t-values for several of their estimates even though they report coefficients (see the discussion in the last section). Without the t-values, we cannot calculate the p-values for these estimates. By coincidence, these estimates are sometimes the median, maximum or minimum coefficient. If this is the case, we exclude them.
3
When calculating the p-values, we take the number of “real” controls and the intercept into account in determining the degrees of freedom. However, we do not conduct a degrees of freedom correction for models with time- or cross-section fixed effects as it does not make any difference asymptotically.
35 Table 9: f-Tests on Significance of Effects f-Statistic
T min
f min = 52.03, p > 0.05
T median
f median = 121.04, p < 0.01
T max
f max = 230.64.10, p < 0.01
Source: Own calculation.
Even though the f-tests suggest a significant relationship, this result should be viewed with skepticism since this test has several drawbacks. First, it assumes independence across studies, and second, that the estimated coefficients are unbiased. We expect some between-study correlation, for example because authors tend to include controls in their models that have proven to be significant in earlier studies. There is also no way in applied econometrics of knowing whether a model is mis-specified. But biased estimates which seemingly suggest that there is a significant effect even though there is none tend to have a disproportionate effect on the likelihood of rejecting the null. Even worse, the test does not consider the direction of the effect, i.e. a number of significant positive and negative estimates would lead to a rejection of the null without giving any indication on whether decentralization actually increases or reduces economic growth (see Jarell and Stanley 2004. for a more thorough discussion of this test). In a second step, we thus calculate the weighted average estimate across all studies. Two different estimators are available, the random effects and the fixed effects estimator. These expressions do not have the usual meaning in the meta-analytic literature. In the meta-analytic context, these estimators are differentiated by the fact that they use different weights in calculating the average estimate. The fixed effects estimator assumes that all studies estimate a common effect and that differences in estimated coefficients are only due to sampling errors. Under this assumption, the pooled effect size is determined according to the following formula:
∑ wT = ∑w L
Tw
i =1 L
i =1
i i
i
,
(2)
36 where Ti, i=1,..L, are the individual estimates and wi weights that are inversely proportional to the precision of the estimate. The weights are defined as wi=1/vi with vi as the variance of the estimated effect i. The random effects estimator assumes that each study estimates a different effect which is randomly drawn from a population of all effects. What is of interest here is the mean of the population. Hence, the variability in estimated coefficients has two components, first the within-study sampling error, and secondly the variance associated with the random effects error. This, in turn, leads to different weights wi in equation (2). The weights for the random effects estimator are given by wi=(1/vi)+z, with z as the estimated variance between studies (see Dominicis et al. 2006 or Sutton et al. 2000, for an extensive discussion). The so called Q-test can be applied in order to explore whether heterogeneity is present among the estimated coefficient. This test is based on the following statistic
∑ − ∑ L
L
Q = ∑ wi (Ti ) 2 i =1
i =1 L
wi Ti
~ χ 2 ( L − 1) ,
(3)
w i =1 i
where all variables are defined as in (2) and wi are the fixed effects weights. Table 10: Q-Statistic and Pooled Coefficients Q-Statistic
Pooled Coefficient, Random Effects,
H 0 : Tw = 0 T min
Qmin = 633.10, p < 0.01
Tw = −0.097, p < 0.01
T median
Qmedian = 92.55, p < 0.01
Tw = 0.001, p = 0.57
T max
Qmax = 596.01.10, p < 0.01
Tw = 0.128, p < 0.01
Source: Own calculation.
Considering Table 10, the Q-test indicates for all three variants of the coefficients, the median, the minimum, and the maximum the presence of considerable heterogeneity across studies, i.e. the random effects estimator is appropriate. Hence, we only display the random effects estimates in the following, and provide further details on our results in Figures 14, 15 and 16, in particular on the weights each study received.
37 Figure 14: Pooled Estimate on Tmin
Source: Own calculation.
Figure 15: Pooled Estimate on Tmed
Source: Own calculation.
38 Figure 16: Pooled Estimate on Tmax
Source: Own calculation.
The random effects pooled estimate calculated with the minimum coefficients is -0.097 and significantly different from zero, the median estimate is 0.001 and not significantly different from 0, and the maximum estimate is 0.128 and significantly different from 0. As expected, the pooled coefficient from the minimum estimates is lower than both the estimate from the median and maximum estimate. However, the fact that the pooled estimate changes its sign between the minimum and the maximum estimate makes it difficult to provide a consistent assertion on the effect of decentralization on economic growth. However, it seems sensible to consider the minimum and maximum estimates as outliers and to focus the median estimate. Based on the median estimate, decentralization does not have a significant effect on economic growth. 6.
Meta-Regressions
6.1. The Meta-Regression Model In this section, the extent to which idiosyncratic characteristics of an empirical model determine the magnitude of the estimated effect is studied, i.e. we explore whether the choice of a particular measure of decentralization, a particular subject of study, or a particular specification changes estimated effect. The most widely used technique for explaining heterogeneous
39 findings across models is meta-regressions. The meta-regression approach to decentralization and economic growth is based on the premise that individual studies estimate models whose specification roughly resembles the following equation Yi = X i β + γDEC i + ν i
(i=1,…, N) ,
(4)
where Y is a measure of economic growth, X a vector of control variables including a constant,
DEC a measure of decentralization, N the number of observations, and ν an error that either conforms to the assumptions of the classical regression model or can be transformed appropriately. By fitting the model to the data by some technique, usually OLS, an estimate of the true ∧
effect of decentralization on economic growth, γ , is obtained. This framework suggests that a meta-regression model can explain diverse findings in the original literature. We propose a meta-regression of the form ∧
γ j = a + Z jα + ε j
(j=1,…,L) ,
(5)
∧
where γ is the estimated coefficient and Z j a vector of variables which describe the characteristics of the estimated model, and ε j the meta-regression error term. The variables used in the meta-regression analysis are listed in Table 11. Most of them are dummy variables, for example indicating whether fixed effects have been included, whether the study uses only single-country or cross-country data, and if nonlinearities are taken into account. Some variables are continuous, notably the estimated coefficient, the t-statistic, the standard deviation, the number of controls and the sample size. In contrast to the last section, where we chose one particular estimate per study, we use the information from all estimated models in the meta-regressions. First, including all observations provides a larger sample leading to several desirable characteristics as more powerful tests and more accurate estimates. And second, with a quasi-panel dataset, we can control for unobserved study-level heterogeneity. The potential number of point estimates is therefore 375. Since the estimated coefficients derive from models that directly or by transformation conform to the assumptions of the classical linear model, the meta-regression errors should be normally distributed. There might be, however, a problem with heteroscedasticity. Since the individual models use different sample sizes, different units and different specifications, the variance of the esti-
40 mates will be related to the characteristics of a study. Indeed, the Breusch-Pagan test rejects the 4 null of homoscedasticity at conventional significance levels ( χ 2 (1) = 1525.37, p < 0.01 ). The
White-test arrives at the same conclusion ( χ 2 (104) = 337.68, p < 0.01 ). Both tests were calculated using the full set of coefficients in Table 13. Table 11: Dependent Variables in the Meta-Regressions Variable
Definition
Variable
Definition
JART
Journal Article
EXPDEC
Expenditure Decentralization
WORKP
Working Paper
REVDEC
Revenue Decentralization
DISSP
Discussion Paper
TAXAUT
Tax Autonomy
UNPMAN
Unpublished Manuscript
REVEXPDEC
Rev. + Exp. Decentralization
INBOOK
Article in Book
OTHERM
Other Measure
THESIS
Dissertation and Masters Thesis
CONTEND
Control for Endogeneity
PUB
JART+INBOOK
UNPUB
WORKP+DISSP+UNPMAN
CCOUNT
Cross-Country Study
CSEC
Cross-Section Dataset
DEVELOPING
Developing Country
PANEL
Panel Dataset
DEVELOPED
Developed Country
FIXEDEFF
Fixed Effects
TIMEEFF
Time Fixed Effects
QUAD
Quadratic Term
INTERAC
Interaction Term
CHINA
Estimate with data from China
USA
Estimate with data from USA
GER
Estimate with German Data
SWISS
Estimate with Data from Switzerland
OTHERC
Estimates with Data from countries except USA, SWISS, GER and CHINA
Dummy Variables
Continuous Variables YEAR
Year of Study
COEF
Estimated Coefficient
TSTAT
t-Statistic
NCON
Number of Control Variables excluding constants and fixed effects
OBS
Number of Observations
4
This particular test was derived by regressing the squared residuals on the fitted values. We did also experiment with regressing the residuals on linear combinations of the explanatory variables. The null was also rejected in this variation of the test. The test is based on pooled OLS estimates.
41 Note: Variables in the meta-regressions and their definitions; not all variables were used in the regressions because several are mutually exclusive and their inclusion would lead to perfect collinearity (for instance the country dummies). Consult table 13 for the set of coefficients used in the regressions.
Another problem with using the plain coefficients is that different studies apply different measures of decentralization and economic growth. They also use different units. The estimated coefficients are not dimensionless and therefore not directly comparable across studies without some standardization according to the precision of the study. Therefore, as in the last section, we use the variability of the estimate as a scaling parameter, i.e. instead of the plain coefficient in model (3) the t-statistic is used as the dependent variable, i.e. ∧
tj =
γj sd
= κ + Z jα + η j
(j=1,…L)
(6)
The t-statistic is a dimensionless measure that can be compared across different models.5 A RESET test based on the pooled OLS does not indicate that the full model in Table 13 has omitted variables ( F (3,317 ) = 1,26, p = 0.29 ). Hence, we record that the functional form of (6) is probably correct and no omitted variables are present. However, based on pooled OLS estimators, both the Breusch-Pagan ( χ 2 (1) = 12.04, p < 0.01 ) and the White-test ( χ 2 (104) = 149.69, p < 0.01 ) indicate that model (4) exhibits considerable heteroscedasticity. This presumption is confirmed by a LR-test for panel level heteroscedasticity based on a GLS estimate which rejects the null of panel-level homoscedasticity ( χ 2 (24) = 505.70, p < 0.01 ). Hence, we record that any estimator of the coefficients in model (4) should control for heteroscedasticity. The Breusch-Pagan LM test does not reject the null of a zero variance in individual effects ( χ 2 (1) = 1,00, p = 0.32 ). Hence, the random effects estimator is not preferred against the pooled OLS estimator. Since the Wooldridge-test (2002) for autocorrelation in the idiosyncratic errors in panel data models suggests that the null of no autocorrelation cannot be rejected ( F (1,22) = 1.956, p = 0.18 ), we conclude that a pooled GLS estimator which takes account of
5
It is usually recommended to divide both the left- and the right-hand size with the standard error in order to handle heteroscedasticity. However, the division of both sides by the standard error does not produce homosecdastic errors even though it seemingly reduces the degree of heteroscedasticity. The Breusch-Pagan 2 test still rejects the null ( χ (1) = 3.67, p < 0.1 ), and so does the White-test 2 ( χ (97) = 329.33, p < 0.01 ). Since dividing both sides by the standard error has several drawbacks (for example, the dummy variables are transformed to continuous variables) but seemingly no positive consequences with regard to the heteroscedasticity, we divide only the LHS with the standard error.
42 the heteroscedasticity is probably the most efficient estimator. Alternatively, the pooled OLS estimator with a robust variance matrix can be applied. Next, we need to explore whether there is an endogeneity problem due to unobserved heterogeneity. To this end, we compare the consistent fixed-effects estimates to the more efficient GLS estimates using the Hausman approach. The traditional Hausman test cannot be applied because the difference variance matrix deriving from our sample is negative semi-definite (presumably because the assumptions underlying the traditional Hausman test are not valid for our data, in particular because the errors are heteroscedastic). Instead, we use the robust artificial regression approach proposed in Wooldridge (2002) which is equivalent to the Hausman test. The test is obtained by estimating an artificial regression of the following form ∩
∩
∩
∩
t j = ω j + X j α + P j β +η j , ∩
∩
∩
(7) ∩
where t j , ω j , and X j are generally the quasi-demeaned variables and P j the set of time-varying fully demeaned data. Since the random effects estimator is not the most efficient estimator, there ∩
is no need to quasi-demean the variables in X j , instead we use the plain values. The test rejects the null of no correlation between individual error component and independent variables if the set of coefficients in the vector β is statistically significant. We find that the test does reject the null ( F (17,328) = 13.28, p < 0.01 ), i.e. that a fixed effects specification is appropriate. 6.2. Results We present our results in Table 12. Since we have coded most of the study characteristics as dummy variables, we had to exclude several variables that are mutually exclusive (for example a study can only be either single- or cross country so that the inclusion of both variables would lead to perfect collinearity with the intercept). The excluded variables provide the base model, and the estimated coefficients should be interpreted as deviations. In order to save degrees of freedom, we merge journal articles and in book publications into the variable PUB and all other publications except dissertations into UNPUB. We feel that this is feasible because there seems to be no meaningful difference between working and discussion papers and the like. We provide in Table 12 pooled OLS estimates with robust White-corrected standard errors, GLS estimates, random effects estimates, and finally fixed effects estimates with robust standard errors. Based on the tests in the last subsection, we will mainly discuss the GLS and the fixed effects results. Note that
43 the robust pooled OLS and random effects estimates are identical for the full set of coefficients. The standard errors are different, however.
Table 12: Meta-Regressions
YEAR PUB UNPUB CROSSC DEVELOPING DEVELOPED DIV EXPDEC REVEXPDEC TAXAUT OTHER LOGLEV LEVELLOG CROSS PANEL
Pooled OLS 1
Pooled OLS 2
RE 1
RE 2
GLS 1
GLS 2
FE 1
FE 2
b/se 0.243*** (0.073) 0.525 (0.647) 0.764 (0.665) -0.915* (0.489) -0.262 (0.674) 0.157 (0.681) 0.297 (0.484) 0.840* (0.458) 1.244* (0.641) -0.701 (0.815) 1.249 (1.246) -2.887 (1.875) 0.686 (0.493) 0.988** (0.419) -0.273 (0.614) Pooled OLS 1
b/se 0.273*** (0.081) 1.735 (1.166) 1.992** (0.922) -2.332* (1.301) 0.566 (1.054) 0.385 (1.232) -0.201 (0.902) 0.259 (0.403) 0.709 (0.543) -1.042 (0.867) 1.106 (1.560) -3.321* (1.929) 1.096 (1.185) 0.688 (0.500) -0.530 (0.636) Pooled OLS 2
b/se 0.323*** (0.118) 1.211 (1.077) 1.713 (1.044) -1.169* (0.709) 0.673 (0.887) 0.377 (0.847) -0.546 (1.063) 0.774 (0.601) 0.626 (1.300) -1.034 (0.718) -1.063 (1.012) -1.210 (1.726) 0.634 (1.180) 0.756 (0.815) -0.630 (0.811) RE 1
b/se 0.273*** (0.083) 1.735** (0.831) 1.992** (0.803) -2.332** (0.939) 0.566 (0.880) 0.385 (0.861) -0.201 (0.845) 0.259 (0.623) 0.709 (1.324) -1.042 (0.734) 1.106 (0.885) -3.321** (1.335) 1.096 (0.999) 0.688 (0.567) -0.530 (0.747) RE 2
b/se 0.150*** (0.046) -0.608 (0.378) -0.005 (0.385) -0.682*** (0.256) 0.355 (0.423) 1.306*** (0.386) 0.446 (0.418) 0.780** (0.315) 0.554 (0.460) -1.021*** (0.322) 3.068*** (0.492) -5.374*** (0.905) 0.307 (0.413) 1.830*** (0.258) -0.158 (0.344) GLS 1
b/se 0.233*** (0.045) 1.188*** (0.372) 0.019 (0.412) -1.578*** (0.389) 0.616 (0.458) 1.937*** (0.460) 0.334 (0.372) 0.388 (0.244) 0.552 (0.349) -0.694*** (0.265) 3.576*** (0.462) -6.015*** (0.893) -0.060 (0.490) 1.121*** (0.225) -0.460 (0.310) GLS 2
b/se
b/se
0.539*** (0.142) 2.510*** (0.271) 0.942** (0.440) 0.245 (0.411) -0.496 (0.526) -2.992*** (0.687) -23.143** (9.758)
0.490*** (0.119) 2.450*** (0.237) 0.811* (0.444) 0.175 (0.414) -0.450 (0.533) -2.916*** (0.676) -23.067** (9.806)
0.055 (0.209) 1.566 (1.060) 0.047 (0.699) FE 1
0.075 (0.209) 0.647 (1.082) -0.709 (0.656) FE 2
45 FIXED TIME OBS NUMC
b/se -0.526 (0.360) -0.226 (0.539) 0.004** (0.002) -0.142 (0.101)
QUAD INTERACT CHINA USA SWISS GER CONST N F
-487.134*** (147.213) 346 11.293
b/se 0.157 (0.615) 0.221 (0.590) 0.004** (0.002) -0.134 (0.086) 1.198** (0.567) 3.275*** (1.103) -2.727*** (0.452) -0.572 (2.432) -2.558 (1.579) 0.425 (0.872) -547.193*** (161.611) 346 25.861
χ2 R2
0.108
0.158
b/se -0.172 (0.493) -0.036 (0.606) 0.004** (0.002) -0.145* (0.082)
b/se -0.155 (0.256) -1.298*** (0.274) 0.004*** (0.001) -0.050 (0.038)
-646.104*** (237.406) 346
b/se 0.157 (0.503) 0.221 (0.561) 0.004*** (0.001) -0.134* (0.079) 1.198** (0.587) 3.275 (2.026) -2.727*** (0.683) -0.572 (1.180) -2.558 (1.702) 0.425 (1.507) -547.193*** (166.317) 346
-301.653*** (91.666) 346
b/se 0.348 (0.226) -1.036*** (0.293) 0.004*** (0.001) -0.049 (0.034) 0.470 (0.358) 2.774** (1.335) -2.907*** (0.290) -1.478*** (0.534) -0.678 (0.630) 1.425*** (0.482) -467.100*** (90.155) 346
36.17
89.93
547.89
970.93
b/se -0.321 (0.456) -0.575 (0.612) 0.004*** (0.001) -0.082 (0.056)
b/se 0.554 (0.401) -0.325 (0.548) 0.003** (0.001) -0.009 (0.054) 0.484 (0.474) 0.731*** (0.228) -4.232*** (0.508)
1.924 (1.529) 346 10.328
2.800* (1.529) 346 20.938
0.5057
0.5313
Note: The dependent variable is the t-statistic. We discard 29 Observations, 28 due to the fact that the original papers did not provide separate t-statistics for interaction terms at the sample averages. One paper reports an estimated model with inconsistent coefficients and t-statistics. The original number of Observations is 375. Standard Errors are in Parenthesis. Stars indicate significance at the 10%(*), 5%(**) and 1%(***) level.
First, all estimators which can estimate within-study constant variables suggest that there is a positive and significant time trend in the reported t-statistic, i.e. the coefficient for YEAR varies between 0.150 and 0.323 and is significant. Both published and unpublished papers tend to report higher t-statistics than dissertations and masters thesis which we use as base. This effect is particularly strong for published studies which have a highly significant coefficient in the GLS estimation. With regard to the study-type, we use single-country studies as base. We find that cross-country studies tend to report significantly lower t-statistic than single-country studies. With regard to the model specification, we use level-level models as base. We find that log-level models report significantly lower coefficients than the base model. Second, with regard to variables that vary between studies we find that both the fixed effects and GLS estimates suggest that studies which consider either exclusively developing or developed countries tend to report higher t-statistics than studies which use data on both kinds of countries simultaneously. However, the spread is larger for studies on developed countries, suggesting that decentralization tends to increase growth more in developed than in developing countries. With regard to the effect of the decentralization measures, we use revenue decentralization as the base. Both the fixed effects and the GLS estimator suggest that models that use the divergence between the central and sub-national government revenue as decentralization measure tend to report higher t-statistics than those which use revenue decentralization. Models that use expenditure or a pooled measure of expenditure and revenue decentralization do not report a t-statistic that is significantly different from that of the base. Both the fixed effects and GLS estimators suggest that models which use tax autonomy report significantly lower t-statistics. The results for models that use other measures of decentralization are a little ambiguous. Whilst the GLS estimator suggests a positive effect, the fixed effects estimator suggests a negative effect. Overall, it is thus also not decided whether the use of some measure of tax autonomy or tax competition induces significantly higher or lower t-statistics. With regard to the type of data used, the time-series characteristic serves as base. We find that cross-section models tend to report on average higher t-statistics than the base while panel-data models tend to report lower t-statistics. Models which use cross-section fixed effects tend to report higher t-statistics while models which use time-series fixed effects report lower t-statistics. Trivially, the number of observations is in all models positively and significantly related to the reported t-statistic while the number of controls is negatively related. Models which take nonlinearities into account report larger t-statistics. Finally, single country studies on China report sig-
47 nificantly lower t-statistics than studies on other countries, which we use as base. So do studies on the USA and Switzerland. Studies on Germany tend to report larger t-statistics. 7.
Concluding Remarks
Our review of theoretical and empirical studies and the meta-analysis of the estimation results conducted so far reveals an interesting, though still preliminary outcome. With respect to the theoretical studies, it is possible to derive arguments as to how fiscal federalism, decentralization or fiscal competition might affect economic growth which are related to microeconomic analyses from the theory of fiscal federalism. First, an advantage of decentralized provision and financing of public goods, namely tailoring them to heterogeneous preferences of individuals, shows up in economic growth as resulting different savings rates might affect the transition to a steady state (Tiebout channel). Second, and more importantly, decentralization allows designing regional economic policies to the necessities of a regional economy, and thus increases growth (Structural Change channel). If this is accompanied by sub-federal experimentation, an argument could, third, be made that political innovation serves as another growth enhancing mechanism (Political Innovation channel). Finally, to the extent to which fiscal federalism actually contributes to solve political economy problems, there might as well be another way how economic growth is enhanced (Market-Preservation channel). It should be noted, however, that the theoretical literature on federalism and growth is still in its beginning. Moreover, a growth-enhancing or decreasing effect of federalism (or more precisely fiscal competition) in many cases heavily depends on particular assumptions. It is therefore still open which transmission channel is the most persuasive one if there is an effect on economic growth at all. Empirical studies do also not provide strong support for an impact of federalism, decentralization or fiscal competition on economic growth. Overall, the empirical evidence is rather inconclusive whether there is an effect at all. The studies also suffer from the fact that often the autonomy of sub-federal jurisdictions is not properly measured. Moreover, the current study design does not allow identifying particular transmission channels of the impact of fiscal federalism on economic growth. In our survey on the empirical literature, we report a tendency in the studies that there is a mild positive effect of fiscal federalism when it is properly measured as a type of competitive government. This interpretation is only partly supported by the meta-analysis. While Fisher’s combined test indicates a significant effect of fiscal federalism or decentralization on economic growth, a more differentiated test strategy, including the Q-test, leaves us with no significant impact on economic growth at the median.
48 The meta-regressions shed more light on which study characteristics tend towards support for a positive impact of fiscal federalism on economic growth. Most importantly, such a relationship is rather to be expected in developed, industrialized countries lending support to the suspicion that fiscal federalism or fiscal decentralization is having much different effects in less developed than in developed countries. In addition, single country studies tend to indicate a positive effect of federalism on growth which may be the result of their possibility to consider the specific differences within a country more strongly. Finally, while it remains undecided how a switch from expenditure centralization measures to tax autonomy indicators affects the results in favor of fiscal federalism or not, it may be useful to consider the single institutions of a country’s fiscal federalism in order to identify their effect. Future research will thus need to come up with a clearer identification strategy to detect the transmission channels by which fiscal federalism affects growth. Future empirical strategies will consist in a further differentiation as to what federalism or decentralization actually means. References Aghion, P. and P. Howitt (2006), Appropriate Growth Policy: A Unifying Framework, Journal of the European Economic Association 4, 269 – 314. Akai, N. and M. Sakata (2002), Fiscal Decentralization Contributes to Economic Growth: Evidence from State-Level Cross-Section Data for the United States, Journal of Urban Economics 52, 93 – 108. Arrow, K. and M. Kurz (1970), Public Investment, the Rate of Return, and Optimal Fiscal Policy, John Hopkins University Press, Baltimore. Aschauer, D. (1989), Is Public Infrastructure Productive?, Journal of Monetary Economics 23, 177 – 200. Baldwin, R.E. and P. Krugman (2004), Agglomeration, Integration and Tax Harmonization, European Economic Review 48, 1 – 23. Baldwin, R.E. and P. Martin (2004), Agglomeration and Regional Growth, in: J.V. Henderson and J.F. Thisse (eds.), Handbook of Regional and Urban Economics, Vol. 4, Elsevier, Amsterdam, 2671 – 2711. Baldwin, R., R. Forslid, P. Martin, G. Ottaviano and F. Robert-Nicoud (2003), Economic Geography and Public Policy, Princeton University Press, Princeton. Bardhan, P. and D. Mookherjee (2000), Capture and Governance at Local and National Levels, American Economic Review, Papers and Proceedings 90, 135 – 139. Barro, R.J. (1990), Government Spending in a Simple Model of Endogenous Growth, Journal of Political Economy 98, S103 – S125. Barro, R.J. and X. Sala-i-Martin (2001), Economic Growth, MIT Press, Cambridge (Mass.). Becker, D. and M. Rauscher (2007), Fiscal Competition in Space and Time: An Endogenous-Growth Approach, mimeo, University of Rostock.
49 Behnisch, A., T. Büttner and D. Stegarescu (2002), Public Sector Centralization and Productivity Growth: Reviewing the German Experience, ZEW Discussion Paper No. 02-03, Mannheim. Berthold, N. and H. Fricke (2007), Volkswirtschaftliche Auswirkungen der finanziellen Ausgleichssysteme in Deutschland, Discussion Paper No. 93, University of Würzburg. Berthold, N., S. Drews and E. Thode (2001), Die föderale Ordnung in Deutschland – Motor oder Bremse des wirtschaftlichen Wachstums?, Discussion Paper No. 42, University of Würzburg. Besley, T. and A.C. Case (1995), Incumbent Behavior: Vote-Seeking, Tax-Setting, and Yardstick Competition, American Economic Review 85, 25 – 45. Besley, T. and S. Coate (2003), Centralized versus Decentralized Provision of Local Public Goods: A Political Economy Analysis, Journal of Public Economics 87, 2611 – 2637. Blanchard, O. and A. Shleifer (2001), Federalism with and without Political Centralization: China versus Russia, IMF Staff Papers 48, 171 – 179. Blankart, C.B. (1996), Braucht Europa mehr zentralstaatliche Koordination?, Wirtschaftsdienst 76, 87 – 91. Blankart, C.B. (2007), Föderalismus in Deutschland und in Europa, Nomos, Baden-Baden. Bodman, P. and K. Ford (2006), Fiscal Decentralization and Economic Growth in the OECD, Unpublished Manuscript, University of Queensland, Brisbane. Borck, R. and M. Pflüger (2006), Agglomeration and Tax Competition, European Economic Review 50, 647 – 668. Brakman, S., H. Garretsen and C. Van Marrewijk (2002), Locational Competition and Agglomeration: The Role of Government Spending, CESifo Working Paper No. 775, Munich, September 2002. Brakman, S., H. Garretsen and C. Van Marrewijk (2006), Agglomeration and Aid, CESifo Working Paper No. 1750, Munich, June 2006. Brennan, G. and J.M. Buchanan (1980), The Power to Tax, Cambridge University Press, Cambridge. Brueckner, J.K. (1999), Fiscal Federalism and Capital Accumulation, Journal of Public Economic Theory 1, 205 – 224. Brueckner, J.K. (2006), Fiscal Federalism and Economic Growth, Journal of Public Economics 90, 2107 – 2120. Bucovetsky, S. and M. Smart (2006), The Efficiency Consequences of Local Revenue Equalization: Tax Competition and Tax Distortions, Journal of Public Economic Theory 8, 119 – 144. Caballero, R.J. (2007), Specificity and the Macroeconomics of Restructuring, MIT Press, Cambridge (Mass.). Cai, H. and D. Treisman (2004), State Corroding Federalism, Journal of Public Economics 88, 819 – 843. Camagni, R.P. (1995), The Concept of Innovative Milieu and its Relevance for Public Policies in European Lagging Regions, Papers in Regional Science 74, 317 – 340. Davoodi, H. and H. Zou (1998), Fiscal Decentralization and Economic Growth: A Cross-Country Study, Journal of Urban Economics 43, 244 – 257. Desai, R.M., L.M. Freinkman and I. Goldberg (2003), Fiscal Federalism and Regional Growth: Evidence from the Russian Federation in the 1990s, World Bank Policy Research Working Paper 3138, Washington, D.C.
50 Devereux, M.B. and A. Mansoorian (1992), International Fiscal Policy Coordination and Economic Growth, International Economic Review 33, 249 – 268. Devereux, M.P., R. Griffith and H. Simpson (2007), Firm Location Decisions, Regional Grants and Agglomeration Externalities, Journal of Public Economics 91, 413 – 435. Dominicis, L., de Groot, H.L.F and R.J.G.M. Florax (2006), Growth and Inequality: A Meta-Analysis, Tinbergen Institute Discussion Paper, Amsterdam. Döring, T. and J. Schnellenbach (2006), What Do We Know about Geographical Knowledge Spillovers and Regional Growth?: A Survey of the Literature, Regional Studies 40, 375 – 395. Ebel, R.D. and S. Yilmaz (2002), On the Measurement and Impact of Fiscal Decentralization, Unpublished Manuscript, World Bank, Washington. Edwards, R.A. (2005), The Structure of Authority, Federalism, Commitment and Economic Growth, Economic Theory 25, 629 – 648. Eller, M. (2004), The Determinants of Fiscal Decentralization and its Determinants on Economic Growth: Empirical Evidence from a Panel of OECD Countries, Diploma Thesis. Enikolopov, R. and E. Zhuravskaya (2003), Decentralization and Political Institutions, CEPR Discussion Paper No. 3857, London. Feld, L.P. (2007), Fiscal Federalism and Economic Growth in OECD Countries, forthcoming in: A. Bergh and R. Höijer (eds.), Perspectives on Institutional Competition, Edward Elgar, Cheltenham. Feld, L.P., T. Baskaran and T. Dede (2004), Fiscal Federalism and Economic Growth: Cross-Country Evidence for OECD Countries, Unpublished Manuscript, Philipps-University Marburg. Feld, L.P., G. Kirchgässner and C.A. Schaltegger (2004), Fiscal Federalism and Economic Performance: Evidence from Swiss Cantons, Unpublished Manuscript, Philipps-University Marburg. Feld, L.P., G. Kirchgässner and C.A. Schaltegger (2005), Fiskalischer Föderalismus und wirtschaftliche Entwicklung: Evidenz für die Schweizer Kantone, Jahrbuch für Regionalwissenschaft/Review of Regional Research 25, 3 – 25. Feltenstein, A. and S. Iwata (2005), Decentralization and Macroeconomic Performance in China: Regional Autonomy Has Its Costs, Journal of Development Economics 76, 481 – 501. Gil-Serrate, R. and J. López-Laborda (2006), Revenue Decentralization and Economic Growth in the Spanish Autonomous Communities, Unpublished Manuscript, University of Zaragoza. Hayek, F.A. von (1948), Individualism and Economic Order, University of Chicago Press, Chicago. Iimi, A. (2005), Decentralization and Economic Growth Revisited: An Empirical Note, Journal of Urban Economics 57, 449 – 461. Inman, R.P. and D.L. Rubinfeld (1997), Rethinking Federalism, Journal of Economic Perspectives 11 (4), 43 – 64. Jarell, S.B. and T.D. Stanley (2004), Declining Bias and Gender Wage Discrimination? A Meta-Regression Analysis, The Journal of Human Resources 39, 828 – 838. Jin, H., Y. Qian and B.R. Weingast (2005), Regional Decentralization and Fiscal Incentives: Federalism, Chinese Style, Journal of Public Economics 89, 1719 – 1742.
51 Jin, J. and H. Zou (2005), Fiscal Decentralization, Revenue and Expenditure Assignments, and Growth in China, Journal of Asian Economics 16, 1047 – 1064. Justman, M., J.-F. Thisse and T. van Ypersele (2002), Taking the Bite Out of Fiscal Competition, Journal of Urban Economics 52, 294 – 315. Kellermann, K. (2007), Fiscal Competition and a Potential Growth Effect of Centralization, mimeo, University of Fribourg. Kotsogiannis, C. and R. Schwager (2006), Political Uncertainty and Policy Innovation, Journal of Public Economic Theory 8, 779 – 805. Krugman, P. (1991), Geography and Trade, MIT Press, Cambridge (Mass.). Lejour, A.M. and H.A.A. Verbon (1997), Tax Competition and Redistribution in a Two-Country Endogenous-Growth Model, International Tax and Public Finance 4, 485 – 497. Lin, J.Y. and Z. Liu (2000), Fiscal Decentralization and Economic Growth in China, Economic Development and Cultural Change 49, 1 – 23. Madiès, T. and B. Ventelou (2004), Federalism in an Endogenous Growth Model with Tax Base Sharing and Heterogeneous Education Services, Papers in Regional Science 83, 1 – 18. Martinez-Vazquez, J. and R.M. McNab (2002), Cross-Country Evidence on the Relationship between Fiscal Decentralization, Inflation, and Growth, in: National Tax Association (ed.), Proceedings of the 94th Annual Conference on Taxation 2001, Washington, D.C, 42 – 47. Martinez-Vazquez, J. and R.M. McNab (2003), Fiscal Decentralization and Economic Growth, World Development 31, 1597 – 1616. Naumets, I. (2003), Fiscal Decentralization and Local Public Sector Efficiency, Unpublished Dissertation, National University of Kiew. Oates, W.E. (1972), Fiscal Federalism, Harcourt Brace Jovanovich, New York. Oates, W.E. (1990), Decentralization of the Public Sector: An Overview, in: R.J. Bennett (ed.), Decentralization, Local Governments and Markets, Clarendon Press, Oxford, 43 – 58. Oates, W.E. (1993), Fiscal Decentralization and Economic Development, National Tax Journal 46, 237 – 243. Oates, W.E. (1999), An Essay on Fiscal Federalism, Journal of Economic Literature 37, 1120 – 1149. Ottaviano, G.I.P. and J.-F. Thisse (2004), Agglomeration und Economic Geography, in: J.V. Henderson and J. Thisse (eds.), Handbook of Regional and Urban Economics, Volume IV, North-Holland, Amsterdam, 2564 – 2608. Qiao, B., Martinez-Vazquez, J. and Y. Xu (2002), Growth and Equity Tradeoff in Decentralization Policy: China‘s Experience, Georgia State University, International Studies Program, Working Paper 0216, Georgia. Rauscher, M. (2005), Economic Growth and Tax-Competing Leviathans, International Tax and Public Finance 12, 457 – 474. Rauscher, M. (2006), Interjurisdictional Competition and Innovation in the Public Sector, Unpublished Manuscript, University of Rostock.
52 Rauscher, M. (2007), Tax Competition, Capital Mobility and Innovation in the Public Sector, German Economic Review 8, 28 – 40. Riker, W. (1964), Federalism: Origins, Operation, Significance, Little, Brown & Co, Boston. Rodden, J. (2004), Comparative Federalism and Decentralization: On Meaning and Measurement, Comparative Politics 36, 481 – 500. Rose, K. and T.D. Stanley (2005), A Meta-Analysis of the Effect of Common Currencies on International Trade, Journal of Economic Surveys 19, No. 3. Rose-Ackerman, S. (1980), Risk-Taking and Reelection: Does Federalism Promote Innovation?, Journal of Legal Studies 9, 593 – 616. Salmon, P. (1987), Decentralization as an Incentive Scheme, Oxford Review of Economic Policy 3 (2), 24 – 43. Schnellenbach, J. (2004), The Evolution of a Fiscal Constitution When Individuals Are Theoretically Uncertain, European Journal of Law and Economics 17, 97 – 115. Schnellenbach, J. (2007), Rational Ignorance Is Not Bliss: When Do Lazy Voters Learn from Decentralised Policy?, mimeo., University of Heidelberg. Solow, R.M. (1956), A Contribution to the Theory of Economic Growth, Quarterly Journal of Economics 70, 65 – 94. Stansel, D. (2005), Local Decentralization and Local Economic Growth: A Cross-Sectional Examination of US Metropolitan Areas, Journal of Urban Economics 57, 55 – 72. Stegarescu, D. (2005), Public Sector Decentralization. Measurement Concepts and Recent International Trends, Fiscal Studies 26, 301 – 333. Strumpf, K.S. (2002), Does Government Decentralization Increase Policy Innovation?, Journal of Public Economic Theory 4, 207 – 241. Sutton, A.J., Abrams, K.R., Jones, D.R., Sheldon, T.A., and F. Song (2000), Methods for Meta-Analysis in Medical Research, Wiley, New York. Swan, T.W. (1956), Economic Growth and Capital Accumulation, Economic Record 32, 344 – 361. Thießen, U. (2003), Fiscal Federalism in Western European and Selected Other Countries: Centralization or Decentralization? What Is Better for Economic Growth, Unpublished Manuscript, DIW Berlin. Thießen, U. (2003a), Fiscal Decentralization and Economic Growth in High Income OECD Countries, Fiscal Studies 24, 237 – 274. Tiebout, C.M. (1956), A Pure Theory of Local Expenditures, Journal of Political Economy 64, 416 – 424. Treisman, D. (2002), Defining and Measuring Decentralization: A Global Perspective, Unpublished Manuscript, UCLA. Wallis, J.J. (1999), Early American Federalism and Economic Development, 1790-1840, in: A. Panagariya, P. Portnoy and R. Schwab (eds.), Environmental and Public Economics: Essays in Honor of Wallace E. Oates, Cheltenham, 283 – 309. Weingast, B.R. (1995), The Economic Role of Political Institutions: Market-Preserving Federalism and Economic Development, Journal of Law, Economics and Organisation 11, 1 – 31.
53 Wellisch, D. (2000), Theory of Public Finance in a Federal State, Cambridge University Press, Cambridge. Woller, G.M. and K. Phillips (1998), Fiscal Decentralization and LDC Economic Growth: An Empirical Investigation, Journal of Development Studies, 34, 139 –148. Wooldridge, J. M. (2002), Econometric Analysis of Cross-Section and Panel Data, MIT Press, Cambridge. Xie, D., H. Zou and H. Davoodi (1999), Fiscal Decentralization and Economic Growth in the United States, Journal of Urban Economics 45, 228 – 239. Yakita, A. (2003), Taxation and Growth with Overlapping Generations, Journal of Public Economics 87, 467 – 487. Yilmaz, S. (2000), The Impact of Fiscal Decentralization on Macroeconomic Performance, in: National Tax Association (ed.), Proceedings of the 92nd Annual Conference on Taxation 1999, Washington, D.C, 251 – 260. Zhang, T. and H. Zou (1998), Fiscal Decentralization, Public Spending, and Economic Growth, Journal of Public Economics 67, 221 – 240. Zhang, T. and H. Zou (2001), The Growth Impact of Intersectoral and Intergovernmental Allocation of Public Expenditure: With Applications to China and India, China Economic Review 12, 58-81. Zou, H. (1994), Dynamic Effects of Federal Grants on Local Spending, Journal of Urban Economics 36, 98 – 115. Zou, H. (1996), Taxes, Federal Grants, Local Public Spending, and Growth, Journal of Urban Economics 39, 303 – 317.
