Journal of Public Economics 74 (1999) 97–139
On the determinants of fiscal centralization: Theory and
evidence
Ugo Panizza*
Office of the Chief Economist, Inter-American Development Bank, Stop W-0436,
1300 New York Avenue NW, Washington DC 20577, USA
Received 1 April 1998; received in revised form 1 September 1998; accepted 1 December 1998
Abstract
This paper presents a simple model that unifies most of the results of the literature on
fiscal federalism. The model describes an economy characterized by two levels of
government, one public good, and a private good. The predictions of the model are tested
by using a new set of measures of fiscal centralization. The main findings are that country
size, income per capita, ethnic fractionalization, and level of democracy are negatively
correlated with the degree of fiscal centralization. The model is tested using OLS, Tobit,
and semi-parametric estimators. The paper also shows that the variables included in the
regression are helpful in the prediction of changes in the level of centralization.  1999
Elsevier Science S.A. All rights reserved.
Keywords: Political economy; Median voter; Fiscal federalism; Decentralization
JEL classification: D720; H110; H710; H770
1. Introduction
Across countries we observe very different institutional arrangements and very
different levels of fiscal centralization. The recent years have witnessed strong
movements toward decentralization and secession (Argentina, Colombia, Ethiopia,
the republics of the former Soviet Union, Yugoslavia, India, Czechoslovakia,
*Tel.: 11-202-623-1427; fax: 11-202-623-2481.
E-mail address: [email protected] (U. Panizza)
0047-2727 / 99 / $ – see front matter  1999 Elsevier Science S.A. All rights reserved.
PII: S0047-2727( 99 )00020-1
98
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Canada, Belgium, Italy and Spain). At the same time, a divided country has
reunified (Germany), large free-trade zones have been created, and there have been
important steps towards the economic and monetary union of some European
countries. It is therefore natural that decentralization has been at the center of the
political and economic debate in many countries.
This paper attempts to identify empirical regularities explaining cross-country
differences in the level of fiscal centralization. To set the stage for the empirical
work, Section 2 presents a stylized model describing an economy with two levels
of government, one public good, and one private good. The central government
obtains utility from the size of its budget and the equilibrium level of centralization is determined by a sequential game where the central government moves first.
The key findings of the theoretical model are that the level of fiscal centralization is inversely correlated with: (i) country size; (ii) income per capita; (iii) tastes
differentiation; and (iv) level of democracy. The last point is particularly
interesting because it may account for the wave of decentralization that has
followed the end of the Cold War.
The existing literature on fiscal centralization can be divided into three main
branches. The first branch studies the optimal division of powers between the
central and local governments (Musgrave, 1959; Oates, 1972). One of the main
results of this branch of the literature is the Decentralization Theorem (Oates,
1972) that identifies the conditions under which it is more efficient for local
governments to provide the Pareto-efficient levels of output for their respective
jurisdictions than for the central government to provide an uniform level of output
across all jurisdictions. One of the corollaries of the decentralization theorem is
that the benefits of decentralization are positively correlated with the variance in
demands for publicly provided goods. This is also one of the key results of the
model presented in Section 2.
The second branch of the literature concentrates on the role of organization
costs (Breton and Scott, 1978). A decentralized system can reduce mobility and
signaling costs, but it is likely to increase administrative and coordination costs.
The optimal level of decentralization is the one that minimizes the sum of these
costs.
The third branch of the literature emphasizes the benefits of competition among
jurisdictions. Tiebout (1956) studies how, in a system with many jurisdictions, the
agents can ‘vote with their feet’ and locate in the jurisdiction that has policies that
are closer to their preferences. While Tiebout concentrates on horizontal competition, Breton (1996) studies the benefits of vertical competition. According to this
notion, different levels of government, in an effort to increase their ‘market share,’
provide the citizens with the optimal type and quantity of public goods. Brennan
and Buchanan (1980) claim that horizontal and vertical competition among
different levels of government can be very important in containing the size of their
budgets.
In addition to the fiscal federalism literature, the scope of this paper extends to
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
99
several issues examined in the recent strand of political economy literature that
studies the optimal number and size of nations (Alesina and Spolaore, 1997) and
the optimal amount of public goods in countries with heterogeneous preferences
(Alesina et al., 1996a,b). The model presented in this paper constitutes a step in the
direction indicated by Alesina and Spolaore (1997) who suggest that it would be
interesting to extend their analysis to a framework with multiple levels of
government. The presence of two levels of government is the key difference
between the model presented in this paper and the one studied by Alesina and
Spolaore (1997); Alesina et al. (1996a).
The second part of the paper builds a new data set of measures of fiscal
centralization and uses it to test the predictions of the model. The theory is tested
using both standard techniques (OLS and Tobit) and a semi-parametric estimator
that does not require normally distributed residuals. To the best of my knowledge,
Oates (1972); Wallis and Oates (1988) are the only two attempts to use a large set
of cross-section data to study the existence of empirical regularities explaining
differences in centralization across countries. Oates (1972) finds that only size and
income per capita are significantly correlated with fiscal centralization. Wallis and
Oates (1988) study the fiscal structure of the American states. They build a panel
of state-level measures of fiscal centralization and find a correlation between a few
socio-economic variables (the percentage of urban population, the percentage of
whites, and the percentage of farmers) and fiscal centralization. When Wallis and
Oates test the model using a simple cross section of American states they find that
only size is robustly associated with fiscal centralization.
In the light of the existing literature, the results of this paper seem particularly
interesting. In fact, the empirical analysis identifies a correlation between fiscal
centralization and two variables measuring the degree of democracy and the
degree of ethnic fractionalization. These results are consistent with the model of
Section 2 and contradict the idea that it is impossible to find a unique set of
variables to explain cross-country differences in fiscal centralization (Oates, 1972).
2. A model of fiscal centralization
The model presented in this section extends the framework developed by
Alesina and Spolaore (1997); Alesina et al. (1996a), to an economy with two
levels of government. The model studies a linear country with area S, population
N, and divided into J jurisdictions. S, N and J are assumed to be exogenous.1 To
simplify the analysis, J is assumed to be an odd number.
Since it is not possible to capture in a single model the richness of the vast
literature on fiscal centralization, the focus of the model presented in this section is
1
Alesina and Spolaore (1997) present a model where the size and number of countries is
endogenous but the size of the public good is exogenous.
100
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
simplification and unification. Having a unified framework will be useful to
organize the ideas and guide the empirical analysis presented in Section 4.
The government produces one public good: G (for government). The general
(local plus central) government budget constraint is given by T 5 G, where T
stands for tax receipts. This can be written in per capita levels as: t 5 g. Per capita
consumption of the private good is defined by c. The prices of all goods are
normalized to 1. All individuals have the same income y, on which they pay a
lump sum tax t, and similar preferences, but they differ in their tastes for the type
of public good.2 Education is an example of publicly provided good on which
preferences are often polarized: some citizens may prefer religious as opposed to
secular schools or may favor the use of a specific language. It is assumed that the
individuals are uniformly distributed over the territory and that the individuals are
stratified and sorted according to their preferences for the public good. The latter
implies that there is a one-to-one correspondence between the physical and the
‘ideological’ distance from the center of the country. The theoretical basis for this
assumption is provided by the literature pioneered by Tiebout (1956). According
to this literature, individuals with similar preferences tend to locate in the same
community and stratification is an equilibrium condition. The model of this paper
assumes a situation where Tiebout-style stratification has already taken place and,
since individuals are already sorted according to their preferences, there is no role
for mobility.
Although there is no general consensus supporting the Tiebout hypothesis, many
empirical studies have found a strong support for neighborhood stratification in
American cities (Hamilton et al., 1975; Borjas, 1995). The idea that individuals are
sorted according to their preferences is also central in Alesina and Spolaore’s
work. These authors claim that: ‘‘If there were no relationship between location
and preferences, then there would be no presumption that a country would be
geographically connected’’ (Alesina and Spolaore, 1997, p. 1030).
The preferences of the ith citizen are described by a distance-sensitive utility
function (Jurion, 1983):
Ui 5 g i12 a (u l im 1(12u )l ij ) c ib
(1)
where g is the per capita amount of government expenditure, l im represents
individual i’s distance from the center of her country, l ij the distance from the
center of the jurisdiction, and u the level of centralization (i.e. the share of the
public good that is provided by the central government). The difference in tastes
across individuals is captured by a [[0,1]: a 50 indicates a country with a very
homogeneous population, while a 51 characterizes a country with a population
2
The assumption of homogeneous income allows us to abstract from all the issues linked to income
redistribution. Wildasin (1991) discusses in great detail the cost and benefits of local versus central
redistribution.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
101
with diversified tastes. Hence, a (u l im 1(12u )l ij ) is the distance between individual i’s preferred type of government and the actual type of government
provided in equilibrium. Assuming that S [(0,1) guarantees that the exponents in
Eq. (1) are non-negative for all individuals.
It is easy to show that the above utility function generates the prediction that the
closer an individual’s preferences are to the actual government the higher is the
ratio between public and private goods that the individual will demand. This
intuitive result justifies the use of a utility function like the one in Eq. (1). Not all
the utility functions would generate this prediction. A simpler Cobb-Douglas
utility function of the kind: Ui 5[gi (12 a (u l im 1(12u )l ij ))] g c i b would generate
the non-realistic prediction that all individuals demand the same mix of public and
private goods, no matter what their preferences for the type of public good are. All
the results obtained in this paper can be reproduced using the additive utility
function studied by Alesina et al. (1996a).
By maximizing the utility function of Eq. (1) under the constraint y5c1g, it is
possible to derive the following demand functions for the public and private
goods:
di y
gi 5 ]]
di 1 b
(2)
by
c i 5 ]]
di 1 b
(3)
with di 512 a (u l im 1(12u )l ij ). Note that ≠gi / ≠di .0, ≠c i / ≠di ,0. As already
pointed out, each citizen will have different preferences for the quantity of g and c.
The actual level of g provided in equilibrium will then be determined by voting.
Note that the idea behind Eq. (1) is that utility is decreasing in the distance from
the type of public good provided in equilibrium (i.e. that Udi . 0). Since Udi 5
U idi ln( g), this condition will be satisfied only if g.1. It is then necessary to
assume that g.1. Since di .0, this assumption only requires an appropriate
definition of the units in which y is measured.
To derive the equilibrium level of centralization I assume that the central
government is the first mover and decides the level of centralization. This
assumption may seem at odds with democratic voting over the type and amount of
public good. Its theoretical background relates to the large political science
literature that shows how the agenda setter can, by controlling the order of voting,
manipulate the final outcome of an election.3 The agenda setter has always some
power and, in the model presented in this paper, the level of democracy measures
how much of this power the government is willing to use to achieve its own goals.
3
One of the most important results in this literature is McKelvey (1976) Chaos Theorem. The latter
shows that multidimensional voting is almost always characterized by a situation with no Condorcet
winner.
102
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
After observing u, the citizens vote on the amount of the public good, and then
on the type of the public good. The assumption of sequential decision making
(similar to the one used by Alesina et al., 1996a) reflects the budget process
adopted in many countries (Alesina and Perotti, 1996).
In a democratic system the type of public good provided is the one preferred by
the median voter. Two types of median voters will be considered: the ‘national’
median voter, indicated by med, and the median voter of jurisdiction j, indicated
by mej. Given the assumption on the spatial distribution of individuals, the median
voters will be located at the center of the country and at the center of their
jurisdictions, respectively.4 See Fig. 1.
At this point it is necessary to investigate who will assume the role of ‘central
government.’ On principle, anybody who promises to supply the type of public
goods preferred by the median voter could play the role of central government, but
only one individual can credibly commit to provide such type of public good: the
median voter herself. Any other individual will, once elected, have an incentive to
deviate and implement a policy closer to her own preferences. It is therefore
natural to assume that the central government will have preferences for the public
good identical to those of the national median voter. Besides sharing the
preferences of the national median voter, the central government derives additional
utility from staying in power. Following Brennan and Buchanan (1980) view of
the central government as a budget-maximizing Leviathan I will assume that the
utility that the government obtains from staying in power is a function of the
budget it controls.
The central government chooses u in order to maximize the following utility
function:
Vgov 5 f U med 1 (1 2 f )u g
(4)
Where f [(0,1) indicates the level of democracy: f 50 indicates a dictatorship,
while f 51 indicates a perfectly democratic country. The utility function of Eq.
(4) has a straightforward interpretation: dictators only care about their budget
while democratic governments identify their preferences with the preferences of
Fig. 1. A country with five jurisdictions.
4
The assumption that j is an odd number guarantees that the national median voter is also the
median voter of the central jurisdiction.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
103
the median voter. Given the discretional power of the agenda setter, the central
government will always be able to extract some rent. In the utility function of Eq.
(4) the level of democracy measures how much of this rent the central government
is willing to extract.5
Notice that the probability of being re-elected does not appear in the government’s utility function. This is because, given the structure of preferences, the
median voter is the only individual for which the Condorcet winning policy is
incentive compatible. Therefore she will be re-elected with probability 1.
For the central government the distances from both the center of the jurisdiction
and the center of the country are 0. Eq. (2) can then be rewritten as: V gov 5 f gc b 1
(12 f )u g.
The government maximizes Eq. (4) by solving the model backward. The last
decision (and therefore the first to analyze) is on the type of the public good. The
type of public good chosen in equilibrium is the one preferred by the median voter
(and therefore the central government).
The next step is to determine the amount of the public good to be provided in
equilibrium. By applying the median voter theorem and using the demand function
of Eq. (2) it is possible to derive the following:
Proposition 1. The amount of public good provided in equilibrium is given by:
my
g* 5 ]]
m 1b
F
(5)
G
S
S
with m 5 1 2 a u ] 1 (1 2 u ) ] .
4
4J
Proof. See Appendix A.
In a model with only one level of government, Alesina et al. (1996a) show that
the optimal quantity of public good depends on the ‘median distance from the
median.’ In the framework presented here, it is possible to interpret (u (S / 4)1
(12u )S / 4J) as a weighted average of the median distance from the national
median and the jurisdiction median and (12 m ) as the ‘ideological’ distance from
the median. The latter will be close to 0 if a country is very small or if its
preferences are homogeneous, and close to 1 for large countries with heterogeneous preferences.
By substituting Eq. (5) into the central government’s utility function Eq. (4), it
is easy to show that the central government will choose the value of u that
maximizes the indirect utility function:
5
This could also be interpreted as: ‘Given the institutional structure, how much rent is the central
government capable of extracting.’ This would involve modeling the institutional structure and goes
beyond the scope of this paper.
104
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
my
Vgov 5 f ]]
m 1b
S
by
DS]]
m 1 bD
b
my
1 (1 2 f )u ]]
m 1b
(6)
From the maximization of Eq. (6) it is possible to derive the following:
Proposition 2. The amount of fiscal centralization u is decreasing in: ( i) the level
of taste differentiation a; ( ii) democracy f; ( iii) income per capita y; and ( iv)
country size S.
Proof. See Appendix B.
The proof of Proposition 2 is simple but long and tedious. Its interpretation
however is straightforward. Since (i) the central government obtains direct utility
from the size of its budget, and (ii) dgov 51. m, the amount of public good
demanded by the government is greater than g*. In other words, voters force the
central government to decentralize by choosing an amount of publicly provided
good lower than the amount preferred by the central government. To increase the
amount of public good produced in equilibrium the central government is then
required to reduce the level of centralization. Eq. (14) in Appendix B shows that
the central government chooses the value of u that sets the marginal cost of
centralization (due to the fact that gu ,0) equal to the marginal benefit of
centralization (due to the fact that, for a given level of g, an increase in u increases
the government budget). Since the ideological distance from the center (12 m )
depends on the size of the country and on the level of heterogeneity of preferences,
it is easy to show that ga ,0 and gS ,0. Therefore, an increase in the level of
tastes differentiation or in country size will decrease the amount of public good
and the marginal benefits from centralization. A decrease in g by increasing
Ug gov /Uc gov will also increases the costs of centralization. Both effects will induce
the central government to reduce the level of u.
One of the results of Proposition 2 is in line with Alesina and Spolaore (1997)
finding that democratization should be positively correlated with the equilibrium
number of countries and it proves the claim that their analysis can be applied to
the division of a country into jurisdictions. A corollary to Proposition 2 is that a
perfectly democratic government will set u 50 (see the proof in Appendix B).6
Proposition 2 does not provide any result for the effect on centralization of the
number of jurisdictions J. A change in the number of jurisdictions will have two
opposite effects. On the one hand, an increase (decrease) in the number of
jurisdictions will affect m and cause an increase (decrease) of g and therefore of
the benefits (from the central government’s point of view) of centralization. On the
6
This result allows to compare the model of this paper with the normative analysis. In fact, it is
possible to show that, given the preferences illustrated in Eq. (1), a perfectly democratic government
will behave like a Benthamite social planner (this is proved in Appendix B).
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
105
other hand, since gu J ,0, an increase in J will increase the centralization-elasticity
of the demand for g and therefore the cost of centralization. The final effect on the
equilibrium level of centralization will depend on which of these factors dominates
the other.
3. Taking the model to the data
Proposition 2 generates four predictions. The result that ≠u / ≠a ,0 suggests that
countries with polarized preferences for the type of public good should be more
decentralized than countries with homogeneous preferences. In countries with
homogeneous tastes the central government can induce the citizens to vote for a
high level of expenditure in publicly provided good even when the latter are
provided in a highly centralized fashion.7 Hence, we should find a negative
correlation between the level of centralization and heterogeneity in the
demand of public goods.
Economic theory indicates that the key factors in determining demand are tastes
and income. Since the model assumes constant income, I will concentrate on the
role of taste heterogeneity. The problem in testing the possible presence of a
negative correlation between centralization and taste heterogeneity is how to
measure the latter. In Section 3.1 I provide some justifications for the use of ethnic
fractionalization as a proxy for taste heterogeneity.
The result that ≠u / ≠f is based on the idea that a democratic government will
not try to exploit its agenda setter power. Hence, we should find a negative
correlation between the level of democracy and the degree of centralization.
This results, besides being in line with Alesina and Spolaore (1997) who find that
a world of dictatorships would lead to larger countries, is also consistent with Ades
and Glaeser’s (1995) finding that dictatorships tend to have very large capital
cities.
The central government’s utility function suggests that perfect democracies
should set u 50 and very repressive dictatorships should set u 51. The majority of
countries included in the data set used in this paper fall between these two
extremes. Most of the real-world governments are neither perfect democracies
(because they are run by self-interested politicians with some agenda-setting
power) nor perfect dictatorships (even dictators need to rely on the support of the
group of people who put them in power). It should also be noted that, since some
public goods cannot be efficiently produced by the local governments (these are
goods with large spillover; defense is an example of such a good), even perfect
democracies will have levels of centralization greater than 0.
7
U gov
g
Remember that, in equilibrium, ]
. 1 and therefore the Central Government would like to
U gov
c
increase public good expenditure above its equilibrium level.
106
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
The result that ≠u / ≠y,0 is based on the idea that decentralization is a normal
good. Hence, we should find a negative correlation between centralization and
income per capita. The existing literature does not agree on the relationship
between income per capita and fiscal centralization. Some authors claim that
decentralization is a luxury and therefore should be positively correlated with
income per capita (Wheare, 1964). In support of this idea Oates (1972) finds that
the level of fiscal centralization is inversely correlated with a country’s level of
development. Other authors have observed that this relationship tends to disappear
when only rich countries are examined. Wallis and Oates (1988), for instance,
show that in developed countries centralization is positively correlated with
income per capita. A possible explanation for this finding is that richer countries
tend to have more generous redistributive policies, usually implemented by the
central government. Casual evidence indicates that some rich countries are
extremely decentralized (Switzerland, United States, Canada are examples) while
others are highly centralized (France for instance). The regression analysis of
Section 4 indicates a strong correlation between income per capita and decentralization.
The result that ≠u / ≠S ,0 depends on the fact that, other things equal, the bigger
the country, the smaller m (leading to a large ideological distance from the center
(12 m )), and hence the quantity of g provided in equilibrium. The central
gov
government will then be forced to decentralize in order to reduce the U gov
g /U c
ratio. Hence, we should find a negative correlation between centralization and
country size. While casual evidence shows that there are some countries that
contradict this prediction (Switzerland is more decentralized than France for
example), the regression analysis provides strong support for a negative correlation
between size and centralization.
The empirical analysis uses Area as the relevant measure of size. Support for
using this variable can also be found in the previous fiscal federalism literature.
The latter suggests that the benefits of decentralization can be offset by market
failures. The most important of these market failures relates to the presence of
spillovers across jurisdictions. These externalities are likely to be inversely
correlated with the size of the jurisdictions. So, to the extent that larger countries
have larger jurisdictions, land area (Area) will be an appropriate proxy for size.8
Note that the utility function of the central government (Eq. (4)) depends on the
per capita level of expenditure in g. By changing Eq. (4) into Vgov 5 f U med 1(12
f )u G (i.e. considering total government expenditure rather than per capita
expenditure), it would be possible to show that centralization is increasing in
population size (≠u / ≠N .0).9
8
Switzerland is a notable exception. The small size of its jurisdictions is compensated by
characteristics of the territory that limit the mobility of its citizens, and therefore the spillovers among
jurisdictions.
9
Tables 13 and 14 show a positive (but not statistically significant) correlation between population
and centralization.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
107
Summarizing, the simple framework presented in this section predicts a negative
correlation between fiscal centralization and: (i) taste heterogeneity; (ii) democracy; (iii) income per capita; and (iv) country size.
Note that the predictions of Proposition 2 derive from the reduced form of the
model of Section 2. Estimating the structural form of the model would require an
equation for the behavior of the agents and one equation for the behavior of the
central government. This could be done by estimating the following simultaneous
model:
gi 5 a1 1 a2 y i 1 a3 Fract i 1 a4 Area i 1 a5ui 1 u i
(7)
ui 5 b1 1 b2 Demi 1 b3 y i 1 b4 gi 1 vi
(8)
Eq. (7) estimates the quantity of public good voted by the economic agents. The
equation is identified by the exclusion of the democracy variable (Democracy does
not appear in the demand functions of Eqs. (2) and (3)). Eq. (8) estimates the level
of centralization chosen by the central government. The equation is identified by
the exclusion of Area and Fract, variables that do not enter in the direct utility
function of the central government. The theoretical model predicts a positive sign
for a2 and b4 and a negative sign for all the other coefficients.
The above model is fully identified and could be estimated using two-stage least
squares. Unfortunately, TSLS produce consistent but inefficient estimates and,
with approximately 55 observations, this could be a serious problem.10 The
empirical part of the paper will then concentrate on estimating a reduced form
equation of the kind u 5f( y, Area, Fract, Dem).
One last problem concerns the technique used in estimating the model. By
definition, the level of fiscal centralization cannot be higher than 100%. Therefore,
I am dealing with a censored dependent variable and, assuming well-behaved
residuals, the appropriate estimation technique is the Tobit model. It should be
pointed out that Eq. (4) shows that the central government will set u 5100 only if
f 50. Hence, only very repressive dictatorships should be fully centralized. The
data seem to support this implication. If we exclude very small countries where the
scope for decentralization may be extremely limited (Malta and Singapore, for
instance), we observe that full centralization is found almost only in extremely
autocratic regimes like Myanmar and Zaire (Egypt is an exception).
3.1. Data
What follows discusses the variables used in the empirical analysis. A
description of the data sources can be found in Appendix C.
10
The parameters obtained by estimating Eqs. (7) and (8) with TSLS have all the expected signs but
they are not statistically significant. The estimation results are available upon request.
108
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
3.1.1. Fiscal centralization
To test the predictions of the model studied in Section 2 it is necessary to build
a data set of measures of fiscal centralization. Identifying such measures is not an
easy task. The main issue is finding a method to quantify the activity of local
governments that results from independent decision making. Oates (1972)
discusses the conceptual problems involved in the choice of the right measure of
fiscal centralization. These problems can be summarized as follows: (i) Different
levels of local governments should be weighted in different ways. For instance,
large regional governments should be considered more centralized than smaller,
city-wide, governments. (ii) Sometimes the local governments collect revenues or
make expenditure but have no autonomy in deciding the tax amount to be
collected or the type of expenditure to be made. This problem is also related to the
identification of the relevant definition of jurisdiction. (iii) The role of intergovernmental grants.
The available data do not allow to address the problems listed above. They
divide between the central government and local governments as a group.
Information on the appropriate decision units and on the use of the intergovernmental grants is not available. It is therefore impossible to apply a
weighting scheme to different levels of local governments or identifying the
number of relevant jurisdictions.
Following Pryor (1968), I define centralization ratios as the percentage of
revenues (or expenditure) of the central government out of the total revenues (or
expenditure) of the public sector. Two measures of fiscal centralization (Total
Revenues and Total Expenditure) for 1975, 1980 and 1985 are built using data
from the IMF (1975, 1980 and 1985) Government Finance Statistics Yearbook.11
The centralization ratios are illustrated in Table 8 of Appendix C. While the data
for 1980 and 1985 are very similar (the correlation between the 1980 and 1985
values is 0.95 for total revenues and 0.98 for expenditure) the centralization ratios
for 1975 are less correlated with the 1980 and 1985 values (the correlation is
between 0.6 and 0.75). This may be an indication of either a change in the level of
decentralization or a change in the methods of data collection.
Oates (1972) tested for the determinants of fiscal centralization using the data
reported in the World Tables (The World Bank, 1969). These data, following the
United Nations accounting conventions, include social security programs in total
public sector revenues and expenditures, but do not include social security
programs in the revenues or expenditure of the central government. Given the fact
that in some countries social security programs represent a big share of the
government budget (sometimes more than 30%), not including social security
11
Until 1988 the IMF provided information on central governments expenditure and revenues and on
general governments (defined as central plus local governments) expenditures and revenues. The
centralization ratios are computed by dividing the former by the latter. A detailed description of the
method used to build the data is provided in Appendix C.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
109
programs makes a big difference. Therefore, the centralization ratios reported by
the World Tables are much lower than the centralization ratios used in this paper.
Since the central government plays a very important role in deciding the size and
the scope of social security programs, the measure of fiscal centralization used in
this paper should be better suited for a study of the determinants of fiscal
centralization than the centralization ratios of the World Tables.12
For most measures Yugoslavia is the most decentralized country. Among the
industrialized countries Switzerland, Canada, and the United States are the most
decentralized. The values for Yugoslavia are suspiciously low. The shares of the
federal government in total revenues and expenditures are always below 30%.
Although Yugoslavia was a well known case of extreme federalism the above
values may be the outcome of an accounting procedure that did not reflect the
actual division of powers between the central and local governments.
3.1.2. Heterogeneity in the preferences for public goods
Since tastes are not directly observable, it is necessary to find a proxy for this
variable. It is not unlikely that different ethnic groups may diverge in their tastes
for publicly provided goods (education is an important example). Therefore,
differences in tastes may be proxied by a measure of ethnic fractionalization
(Fract). The possibility of a relationship between ethnic fractionalization and
tastes was first explored by Oates (1972) who used a dummy variable to
differentiate homogeneous from heterogeneous countries.13
One problem with Oates’ analysis is that a dummy variable dividing the world
between homogeneous and heterogeneous countries is a very coarse measure of
tastes differentiation. A continuous measure of heterogeneity would be more
instructive.
To address this problem, I measure ethnic fractionalization using the data
collected by the Department of Geodesy and Cartography of the State Geological
Committee of the Soviet Union, originally published in the Atlas Narodov Mira
(1964) and then reported by Taylor and Hudson (1972). This variable ranges from
0 (Korea) to 0.9 (Zaire) and measures the probability that two randomly selected
individuals will belong to different ethno-linguistic groups.14 The same measure
has been used in cross-country studies of growth by Canning and Fay (1993);
Mauro (1995); Castilla (1996); Easterly and Levine (1997). Easterly and Levine
study the relationship between this measure of ethnic diversity and similar indices
proposed by various authors and find that ethnic fractionalization is a robust
12
Ideally, one should use data for the activities that are decentralizable and therefore exclude defense
and social security.
13
He used three measures of differentiation: linguistic, racial, and religious. These dummies assumed
a value of 1 for countries that Oates deemed to be homogeneous and 0 for countries deemed to be
heterogeneous.
14
The value for Korea was not included in the original data set of the Atlas Narodov Mira but
reported (using a different source) by Taylor and Hudson (1972).
110
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
predictor of potential ethnic conflicts. Alesina et al. (1996a) use ethnic fractionalization to capture conflicts among groups in the decision of the quantity and type
of public goods. To support this idea, they quote a vast sociological literature that
finds that preferences and conflicts over public policies are more strongly
correlated with ethnic as opposed to income differences.
Most African countries are highly ethnically fractionalized and nine out of the
ten most fractionalized countries are in Africa (the tenth one is India). Zaire,
Cameroon, South Africa, Nigeria, Central African Republic, Kenya and Zambia
have measures of ethnic fractionalization greater than 0.8. Among the industrialized countries, Canada has the highest degree of ethnic fractionalization (0.75),
followed by Belgium (0.55), Switzerland, and the USA (0.5). Yugoslavia is the
European country with the highest degree of fractionalization (0.75).
Of course, ethnic fractionalization is appropriate to test the model of Section 2
only if one assumes that the different ethnic groups are spatially separated. Support
for this assumption comes from the empirical literature aimed at testing Tiebout
(1956) model. It is also often observed that different ethnic groups are located in
different regions of a country and, in some cases, these regions have large
autonomy from the central government.15
3.1.3. Democracy
To test for the fourth hypothesis I use the data on political rights assembled by
Gastil (1990).16 Gastil’s classification assigns the value 1 to perfect democracies
and 7 to countries where most citizens have no political rights. Following Barro
(1996), I transform Gastil’s 1–7 ranking into a 0–1 ranking, where 0 corresponds
to Gastil’s 7 and 1 corresponds to Gastil’s 1. An alternative source of measures of
democracy is provided by the Polity II data set (Gurr et al., 1989). The results
obtained using the Polity II indices of democracy are essentially identical to the
results obtained using Gastil’s definition.
4. Estimations of the determinants of fiscal centralization
This section discusses the methods used to test for the empirical predictions
described in Section 3.1. The results of the various tests are presented using the
centralization ratios for both government revenues and expenditures. As a first
15
A possible interpretation is that the central government ‘bribes’ ethnically diverse regions by
giving them more autonomy. For instance, in a highly centralized country like Italy there are five
special regions that enjoy large fiscal autonomy and transfers from the central government. Two of
these regions are islands (Sardinia and Sicily). The other three are border regions characterized by large
ethnic minorities: French in Valle d’Aosta, German in Trentino Alto Adige, and Slavic in Friuli Venezia
Giulia. Similar examples can be found in Canada, Belgium, India, Spain, Switzerland and the UK.
16
Gastil uses the following definition: ‘‘Political rights are rights to participate meaningfully in the
political process. In a democracy this means the right of all adults to vote and compete for public
office.’’
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
111
step, the model is estimated using ordinary least squares. Then, more appropriate
Tobit and semi-parametric estimation methods are used.
Previous work found evidence for the fact that size and income per capita have
some weight in explaining fiscal centralization, but most attempts of finding any
correlation between fiscal centralization and socio-political variables have not been
successful. I start the analysis by estimating two regressions where measures of
ethnic fractionalization and democracy are added, one at a time, to a basic
specification that includes income per capita and area (these variables are always
included given the existing evidence of their correlation with centralization).
Formally, I estimate linear forms of the following two equations: CENTR5f(Size,
y, Fract), and CENTR5f(Size, y, Dem). The rationale for the above specifications
is to compare the effects on centralization of ethnic fractionalization and
democracy. Next, I test the following equation:
CENTR i 5 a0 1 a1 Size i 1 a2 y i 1 a3 Fract i 1 a4 Dem i 1 u i
(9)
Instead of levels, the estimations use the logs of Area and GDP per capita. The log
specification improves the fit of the regression and indicates the presence of a
non-linear relationship between these two variables and fiscal centralization.
Oates (1972) uses ordinary least squares to estimate an equation similar to
CENTR5f(Area, y, Fract). Since the centralization ratios cannot be higher than
100, the independent variables are censored from above and OLS is not the
appropriate technique to estimate Eq. (9). It is well known that, with censored
data, least squares estimates are biased. Greene (1993) finds that the bias is
proportional to the percentage of limit observations. In the data set used in this
paper the non-limit observations range between 84 and 90% of the total. Therefore
the OLS estimates should have a bias that ranges between 10 and 20%.
Despite its fundamental flaw, I run OLS regressions to compare my results with
Oates’ work. The results, reported in Appendix D, confirm Oates’ finding that size
and income per capita are negatively correlated with fiscal centralization. I also
find that, as predicted by the model of Section 2, democracy and ethnic
fractionalization are negatively correlated with fiscal centralization. These preliminary results, although not very strong (because the coefficients are not always
statistically significant), are encouraging. Oates (1972) found that only size and
income had a role in explaining centralization and that the proxies for demand
differentiation always had the wrong signs.17
4.1. Tobit estimations
This section uses the Tobit model to test the predictions of Section 3.1. The
Tobit model is the standard technique used to estimate equations with censored
17
In his regressions ethnic fractionalization was never statistically significant and always positively
correlated with centralization.
112
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
dependent variables. The assumption of normally distributed residual is crucial for
the consistency of the Tobit estimates.
Table 1 presents the results for the Tobit estimates when the dependent variable
is total revenues. For all regressions, Area and y are negatively correlated with
Table 1
Tobit estimates: Revenues centralization ratios
1975
Area
y
Fract.
22.761
(23.913)***
27.925
(25.206)***
28.779
(21.519)
Dem.
Const.
No. obs.
x2
Area
y
Fract.
189.271
(12.097)***
60
34.52
1980
22.623
(23.039)***
29.75
(24.681)***
215.347
(21.968)*
Dem.
Const.
No. obs.
x2
Area
y
Fract.
206.03
(9.508)***
56
25.55
1985
23.202
(23.865)***
210.961
(25.638)***
218.438
(22.724)***
Dem.
Const.
No. obs.
x2
225.481
(10.69)***
56
37
23.052
(24.892)***
24.528
(22.238)**
210.592
(22.375)**
166.674
(10.384)***
59
37.92
23.049
(23.700)***
27.861
(23.279)***
1.258
0.199
189.264
(8.463)***
58
21.35
23.85
(25.129)***
27.736
(23.098)***
26.876
(21.008)
204.559
(9.17)***
60
39.2
23.045
(24.333)***
24.621
(22.133)**
26.405
(21.081)
212.545
(22.299)**
172.579
(10.000)***
57
37.57
22.639
(23.048)***
29.306
(23.638)***
215.522
(21.983)*
21.919
(20.299)
203.75
(8.873)***
56
25.64
23.253
(23.949)***
28.937
(23.488)***
218.435
(22.732)***
28.247
(21.192)
214.361
(9.417)***
56
38.43
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
113
revenues centralization and have high and statistically significant coefficients. The
coefficients attached to ethnic fractionalization are, as predicted by the theory,
always negative but they are not statistically significant in the 1975 regression.
However, they are marginally significant in the regression for 1980 and highly
statistically significant in 1985. Conversely, democracy has an important role in
explaining fiscal centralization in 1975 but loses its explanatory power in 1980 and
1985.
Table 2 presents the results for the Tobit estimates when the dependent variable
is total expenditure. Now, Fract has high t-statistics for both 1980 and 1985. As
for revenues, democracy plays a key role in explaining centralization in 1975.
Qualitatively, the results of the Tobit estimations are not very different from the
results of the least squares estimates. As expected, the absolute values of the
coefficients are higher and, as predicted by Greene (1993), the differences range
from 10 to 25%.
To summarize, the Tobit estimations support the idea that size, income, and
ethnic fractionalization are negatively correlated with the degree of fiscal centralization. The coefficients attached to democracy are always negative (supporting the
prediction that democratic countries tend to have more decentralized fiscal
systems), but statistically significant only for 1975.
Tables 13 and 14 in Appendix E show that these results are robust to the
inclusion of population or total GDP as a measure of size. When more than one
measure of size is included in the regression, only Area shows a robust correlation
with fiscal centralization.
The specifications used in Tables 1 and 2 are very parsimonious. To test the
robustness of the results, I augmented the regressions with some variables that are
likely to be correlated with fiscal centralization. One key issue is that most of these
variables are likely to be endogenous with respect to the measure of fiscal
centralization. A variable that might play an important role in determining the
degree of fiscal centralization is defense expenditure. Augmenting Eq. (9) with
defense expenditure does not modify the basic results of Tables 1 and 2. The
coefficients attached to defense expenditure have the expected (positive) sign but
they do not enter in the regression in a statistically significant way.
Four regional dummies (Africa, Latin America, Asia and OECD countries) were
included in the regression to test for possible regional effects. The introduction of
these dummies does not change the results of Tables 1 and 2. Since different
regions may have different slopes, Eq. (9) was also augmented with four
interaction dummies. When sub-Saharan Africa is used, some of the coefficients
attached to the interaction dummies become marginally significant, but the basic
results are similar to the ones of the regressions that do not include the interaction
dummies (see Table 10 in Appendix E). The coefficients attached to ethnic
fractionalization increase and become statistically significant in all the regressions
(including 1975) when interaction dummies for Asia are included in the model
(Table 11). Asian countries seem to be characterized by a positive relationship
114
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Table 2
Tobit estimates: Expenditure centralization ratios
1975
Area
y
Fract.
22.422
(23.572)***
27.208
(24.593)***
22.129
(20.365)
Dem.
Const.
No. obs.
x2
Area
y
Fract.
176.650
(11.179)***
58
26.91
1980
22.761
(23.215)***
29.474
(24.940)***
218.351
(22.561)***
Dem.
Const.
No. obs.
x2
Area
y
Fract.
208.162
(10.143)***
56
29.41
1985
23.076
(23.658)***
210.446
(24.881)***
215.052
(22.127)**
Dem.
Const.
No. obs.
x2
218.608
(9.681)***
54
29.93
22.700
(24.492)***
24.241
(22.108)**
212.117
(22.416)***
159.098
(10.099)***
57
32.04
23.457
(24.147)***
26.873
(22.937)***
21.68
(20.268)
189.795
(8.61)***
58
23.81
23.648
(24.812)***
29.086
(23.403)***
21.06
(20.158)
210.177
(8.872)***
59
34.29
22.917
(24.206)***
23.337
(21.520)
20.197
(20.328)
213.934
(22.448)***
158.259
(9.214)***
55
30.39
22.789
(23.241)***
28.687
(23.582)***
218.536
(22.580)***
23.328
(20.529)
204.015
(9.308)***
56
29.69
23.088
(23.667)***
29.929
(23.504)***
214.953
(22.108)**
21.928
(20.279)
215.558
(8.6)***
54
30.01
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
between fractionalization and centralization and by a negative relationship between
democracy and centralization. In interpreting these results one must be careful
because all the coefficients have very large standard errors and are unstable over
time.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
115
5. Outliers and robustness
This section identifies some important outliers and study their role in determining the results discussed in Section 4. The next section will use a semi-parametric
estimator and bootstrapping techniques to investigate in a more systematic way the
role of outliers and the consequences of the violation of the hypothesis of normally
distributed residual.
There are two important outliers in the data: Yugoslavia and Zaire. Both
countries have high levels of ethnic fragmentation and low levels of democracy,
but Yugoslavia has the most decentralized fiscal structure and Zaire one of the
most centralized. To analyze the role of these two countries, I dropped Zaire and
Yugoslavia from my data set. Table 3 illustrates the results of this experiment
when revenues centralization for 1985 is used as a dependent variable to estimate
Eq. (9). Column 1 shows that, when only Yugoslavia is excluded from the
regression, the coefficient attached to ethnic fractionalization loses most of its
explanatory power and the coefficients and the t-statistics attached to Dem increase
noticeably. It is interesting to note that the results are similar to the 1975 estimates
(that do not include Yugoslavia). Column 2 shows that the opposite happens when
only Zaire is excluded from the regression. The most interesting case is the one
illustrated in Column 3 where both Yugoslavia and Zaire are dropped from the
regression. In this case, both the coefficients attached to Fract and Dem are
statistically significant, but the effect of democracy dominates the effect of ethnic
fractionalization. The coefficients of Column 3 are very similar to the coefficients
of the semi-parametric estimation that will be presented in the next section.
Table 3
Tobit estimations excluding Yugoslavia and Zaire (dependent variable: 1985 revenues centralization
ratio)
Area
y
Fract.
Dem.
Const.
No. obs.
x2
1
(Yugoslavia excluded)
2
(Zaire excluded)
3
(Yugoslavia and Zaire excluded)
23.35
(25.54)***
25.63
(22.88)***
29.65
(21.89)*
215.86
(22.97)***
190.40
(11.25)***
55
53.38
23.28
(24.01)***
28.32
(23.23)***
219.54
(22.90)***
28.34
(21.21)
209.66
(9.21)***
55
37.88
23.37
(25.49)***
25.23
(22.66)***
210.49
(22.06)**
215.85
(22.99)***
187.26
(11.06)***
54
52.48
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
116
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Economists and statisticians do not agree on the way one should treat outliers.
Some think that outliers distort the estimations and they should be excluded from
the model. Others claim that outliers provide interesting information. Paradoxically, Yugoslavia and Zaire seem to be good real-world examples of the situation
described in the model of Section 2. In both countries, we observe two opposing
forces: low levels of democracy that tend to make the country more centralized
and highly ethnically fractionalized populations that push for more decentralization. The outcome probably depends on the cost of repression (a very repressive
central government may face international sanctions). It is reasonable to think that
the latter will depend on the ‘visibility’ of a country. So, high levels of repression
may have been very costly for a country like Yugoslavia (located at the center of
Europe), while the costs of repression may have been lower in Zaire.
The recent history indicates that both democracy and ethnic fractionalization
played an important role in Yugoslavia. This country was kept unified and fairly
stable by a powerful and charismatic leader: Marshall Josip Tito. Tito’s death was
soon followed by the break-up of Yugoslavia.
As pointed out in Section 3.1, the values for fiscal centralization in Yugoslavia
are suspiciously low and might not reflect the true division of powers between
central and local governments. In this case, Yugoslavia would not be a true outlier,
but only an observation not accurately measured. Since Yugoslavia has such an
important weight in determining the role of ethnic fractionalization, it is interesting
to do a sensitivity analysis and explore the effects of different values of its
centralization ratio. To this purpose, I re-estimated the equation for revenue
centralization in 1985 using different values for Yugoslavia. When revenues
centralization for Yugoslavia is between 0.5 and 0.85,18 both ethnic fractionalization and democracy are statistically significant.
While Area, GDP per capita, and ethnic fractionalization are very likely to be
exogenous with respect to fiscal centralization, reverse causation from centralization to democracy may be a problem. The promotion and preservation of
democracy was one of the main reasons why the American founding fathers chose
a highly decentralized system. De Tocqueville in La Democratie en Amerique
praises the highly decentralized American system and claims that such a system is
the best protection for democracy. Although the events of the recent years (Spain,
Czechoslovakia, Poland and the Soviet Union) seem to indicate a clear causation
from democracy to decentralization, it is interesting to explore the reverse
causation issue by using an instrument for the democracy variable. The problem is,
as usual, the identification of a good instrument. An attempt was made by using
the lagged value of democracy. This variable is highly correlated with democracy
and therefore has all the statistical characteristics to be a good instrument.19 The
18
These seem to be reasonable values given the fact that only two countries have centralization
ratios lower than 0.5 (Yugoslavia and Canada).
19
However, even if the use of this variable rules out direct reverse causation, there may still exist a
systematic relationship between democracy and centralization.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
117
instrumental variables estimates (available upon request) are almost identical to the
results of the Tobit estimations with no instruments.
5.1. Semi-parametric estimations
Similarly to OLS, the Tobit model yields consistent parameter estimates if the
residuals are homoskedastic and normally distributed. While ordinary least squares
are very robust, and a violation of these hypotheses leads to inefficient but still
consistent estimates, this is not the case for the Tobit model. Maximum likelihood
methods can only be used to estimate censored regression models if the
distribution of the residuals is known. Greene (1993) discusses the consequences
of a mis-specification of the distribution of the residuals and shows that the
violation of the normality and homoskedasticity hypotheses leads to inconsistent
estimates of the parameters.
In recent years many semi-parametric estimators that do not require any
assumption on the distribution of the residuals have been proposed. The symmetrically trimmed least squares estimator (STLS) and the censored least absolute
deviation estimator (CLAD) proposed by Powell (1984), (1986) do not require
i.i.d. errors but they do not perform well in small samples. The pairwise difference
estimator (PD) proposed by Honore´ and Powell (1994) requires i.i.d. residuals but
performs relatively well in small samples. Monte Carlo simulations have shown
that, with a sample size of 50 observations, the PD estimator is often close to the
best estimator.
Since the data set used in this paper contains less than 70 observations, I
decided to use Honore´ and Powell’s PD to estimate Eq. (9). (Appendix F contains
a description of the Honore´ and Powell pairwise difference estimator.) Table 4
presents the results obtained by using this estimator. The number in parentheses
indicate the probability of a sign switch. These pseudo p-values were obtained by
bootstrapping the model 1000 times.
The semi-parametric estimations of the coefficients of Area and y are similar to
the ones obtained in the OLS and Tobit estimations. The semi-parametric
estimator instead produces very different results for the coefficients attached to
ethnic fractionalization and democracy. The coefficients attached to Fract are
always lower than the Tobit estimates and the coefficients attached to Dem always
higher. Ethnic fractionalization is statistically significant at the 5% level when the
dependent variable is revenues in 1985 and statistically significant at the 10% level
when the dependent variable is revenues in 1980. In the expenditure regressions,
the coefficient attached to ethnic fractionalization is statistically significant (with a
p-value of 0.051) only in 1980. Democracy enters in the regression with a
statistically significant coefficient for both expenditure and revenues in 1975 and
for revenues in 1985.
The results of the semi-parametric estimations are much closer (in the
magnitude of both the parameters and their p-values) to the results of Column 3 of
Table 3 than to the Tobit results for the whole sample. In the Tobit estimations for
118
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Table 4
Semi-parametric estimations
Revenues centralization ratios
1975
Area
y
Fract.
Dem.
Const.
No. obs.
Area
y
Fract.
Dem.
Const.
No. obs.
1980
1985
23.33
22.82
23.53
(0.000)
(0.002)
(0.000)
24.18
26.68
26.27
(0.026)
(0.002)
(0.002)
24.47
28.45
210.19
(0.275)
(0.095)
(0.044)
212.64
25.46
212.46
(0.025)
(0.19)
(0.030)
170.8
182.38
193.42
(0.000)
(0.000)
(0.000)
57
56
56
Expenditure centralization ratios
1975
1980
22.99
(0.002)
23.98
(0.082)
0.05
(0.492)
213.54
(0.007)
163.96
(20.000)
55
23.04
(0.002)
26.03
(0.001)
210.45
(0.051)
26.78
(0.189)
182.49
(0.000)
56
1985
23.25
(0.002)
26.92
(0.001)
25.71
(0.200)
26.05
(0.240)
190.17
(0.000)
54
Pseudo p-values in parentheses.
the whole sample, Fract has a coefficient of 218.43 (and a t-statistic of 2.73),
which is 80% higher (in absolute value) than the coefficient obtained with the PD
estimator (210.19). At the same time, the coefficient of the Tobit estimation that
excludes Yugoslavia and Zaire is almost identical to the coefficient of the PD
estimation for the whole sample. This is an indication that the Honore´ and Powell
method and the bootstrapping downweigh outliers and implicitly correct for some
of the problems in the data. To conclude, the PD estimator reduces the explanatory
power of ethnic fractionalization and increases the explanatory power of the
democracy variable, but still finds a negative correlation between these two
variables and the level of fiscal centralization.
After exploring the existence of a statistically significant relationship between
centralization and a set of socio-economic variables, it is worth asking if this
relationship is also economically significant. The results of Table 4 indicate that,
for the 1985 sample, a 1 standard deviation increase in ethnic fractionalization is
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
119
associated to a 2.8% decrease in revenues centralization. At the same time, a 1
standard deviation increase of democracy in 1985 is associated to a 4.3% decrease
in centralization. The effects on centralization associated to changes in income per
capita and area size are also very large (ranging from 0.5 to 6%). This allows to
conclude that the relationship between centralization and the set of socio-economic
variables used in this paper is not only statistically significant but also economically significant.
6. The role of history
History is the main reason why many scholars claim that it is not possible to use
economic theory to explain cross-country differences in the division of powers
among different levels of government. Many political scientists have pointed out
that, in each country, intergovernmental fiscal relations are the outcome of a
bargaining process that is generally unpredictable (Oates, 1972). In some cases,
this bargaining process generated a structure of intergovernmental fiscal relations
that, although optimal at the time the process took place, may not reflect the
current preferences of the citizens. In other cases, the level of fiscal centralization
may have been imposed by a dictator (or monarch) or by international pressure
without any consideration for the citizens’ preferences. Since the process of
adjusting to the optimal fiscal structure requires time, many countries may still be
far away from their optimal level of centralization.
Although the model presented in Section 2 is static, it is interesting to study if
some of the countries included in the sample are out of the equilibrium and slowly
adjusting towards it. An ideal way to control for the role of history would be to
include in the regression fiscal centralization measured at the time a given country
achieved its independence, but this variable is not available.20 An alternative
method is to augment the regression with the lagged value of fiscal centralization.
20
Attempts to find proxies for the distance from the optimal fiscal structure were not successful. To
this purpose I explored variables like social unrest, presence of independentist movement, or number of
years since a given country has started its adjustment process. All these variables present conceptual
problems. The first set of variables (social unrest and presence of independentist movement) is likely to
be the effect and not the cause of a sub-optimal fiscal structure. The latter variable has the advantage of
being exogenous but it is not easily measurable. The problem with this variable is to identify a starting
point. While this is not very difficult for countries like the UK or the US that have had a very stable
political system and the same form of government for more than two centuries, other countries pose
very serious problems. How does one deal with cases like France, where there have been several
changes in the constitution in the last 40 years? Or Spain? Should one start counting after the death of
Franco? What about Italy where there is a large consensus for a more decentralized system but nothing
has changed so far? Exogenous shocks, like the end of the Cold War, seriously reshaped the political
equilibrium in Europe and modified the status quo. Unfortunately, the available data stop in 1985 and it
is impossible to control for these events.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
120
To this purpose, I included fiscal centralization in 1975 among the explanatory
variables for fiscal centralization in 1985.21
The results (reported in Table 5) confirm the fact that history is very important.
The lagged value of fiscal centralization absorbs most of the variance of the
regression (but its coefficient is significantly lower than 1) and reduces the
explanatory power of the other variables. Although the explanatory power of Area
and y decrease noticeably, the correlation between Fract and fiscal centralization
is robust to the inclusion of the lagged dependent variable. It is not surprising that
fiscal centralization is highly correlated with its lagged value. Most of the
explanatory variables are either almost constant (like ethnic fractionalization and
area) or highly correlated (like GDP and democracy) over time. The fact that, even
when the lagged value is included, the other explanatory variables still have some
power in explaining fiscal centralization suggests the idea that, while some
countries have reached their equilibrium level of fiscal centralization, others are
still adjusting towards it.
The very recent history shows that for some countries even the overall borders
were not optimal from the point of view of their citizens. It is not surprising that
the borders of these countries had not been fixed by a national movement but
imposed by outside forces (this is the case for the separation of West and East
Germany) or were the output of a highly non-democratic process (Yugoslavia and
Czechoslovakia, for instance). The borders of other countries, especially in Africa,
Table 5
Tobit estimations (dependent variables: 1985 revenues and expenditure centralization ratios)
Area
y
Fract.
Dem.
Centr. 1975
Const.
No. obs.
x2
Revenues centralization
Expenditure centralization
20.82
(22.19)**
21.32
(21.20)
25.93
(22.08)**
23.50
(21.25)
0.87
(16.48)***
37.54
(2.80)***
53
125.37
20.89
(22.00)*
23.27
(22.19)**
27.91
(22.26)**
21.64
(20.44)
0.80
(12.19)***
60.95
(3.66)***
51
93.90
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
21
When the 1975 values were missing, the 1980 values were used as proxy.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
121
were determined by their former colonial authorities. Even countries with a strong
and long democratic history can be far away from the equilibrium level of
centralization. The regression therefore pools together countries that may be very
close to their optimal fiscal structure with countries that are far away from the
optimum and are slowly adjusting to it.
If we accept the idea that some countries are out of the equilibrium, the
residuals of the regression should be correlated with changes in fiscal centralization over time. Countries where the actual level of centralization is higher than the
predicted value (i.e., u i .0) should be moving towards a more decentralized
system and vice versa. Table 6 illustrates some anecdotal evidence in support of
this idea: Belgium, France, Italy, Spain, and the United Kingdom are all countries
where, in 1985, u i .0; since then, all these countries have moved or are moving
towards a more decentralized system. The very large negative residual for
Yugoslavia seems to indicate that this country was too decentralized in 1985; this
is contradicted by the events that followed Tito’s death. The situation is more
complex for Canada; the strong movement for the independence of Quebec seems
to indicate that the country is too centralized (and therefore contrasts with the
negative residual). Another possible interpretation is that, in order to prevent
secession, the central government is ‘bribing’ Quebec with ‘too much’ autonomy.
This idea finds some support in the results of the referendum in which the
residents of Quebec voted against secession from the rest of Canada.
Although Table 6 provides some support for the idea that the residuals are
negatively correlated with future decentralization, relying on anecdotal evidence to
try to identify which countries are moving towards a more or less centralized
system may sound suspicious. A more systematic method to investigate the
problem is to look at the correlation between the residuals of the 1975 regression
and the annual rate of change in fiscal centralization for the 1975–1985 period
(defined as Dc5(Centr1985 /Centr1975)1 / 10 21).22 The results of this experiment
Table 6
Revenues centralization (1985) (actual and predicted values for selected countries)
Country
Actual
Predicted
u
Belgium
France
Italy
Spain
United Kingdom
Canada
Yugoslavia
92.71
89.38
92.76
85.15
86.52
47.37
27.13
78.99
74.22
82.83
76.20
76.49
53.58
81.09
13.71
15.16
10.73
8.95
10.03
26.21
254.41
22
The 1975–1980 rate of change was used when data for 1985 were not available.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
122
Table 7
Correlation between the residuals for the 1975 regression and rate of change in fiscal centralization in
the following decade
Correlation
p value
No. obs.
Revenues centralization
Expenditure centralization
20.40
0.0045
48
20.51
0.0003
45
strongly support the idea that the 1975 residuals are negatively correlated with the
subsequent changes in fiscal centralization. Table 7 provides evidence for the fact
that countries that are out of the equilibrium are slowly moving towards it and that
Eq. (9) is a good specification to analyze the empirical determinants of fiscal
centralization.
Some evidence that countries are adjusting toward an equilibrium can also be
found in the fact that the 1985 equation fits the data much better than the 1975
equation. When an OLS regression is performed on the same set of countries R¯ 2
increases from 0.47 to 0.56. At the same time the sum of squared errors of the
Tobit regression decreases from 4670 to 3842.
One should be careful in interpreting the results of this section. In particular,
they do not provide any additional support for the model of Section 2. The latter is
set in a purely static framework, while the issues discussed in this section are
dynamic. In fact, this section shows that there are changes in the level of
centralization that cannot be explained by the set of regressors used in this paper,
and calls for further research on the transitional dynamics of the decentralization
process.
7. Conclusions
In the real world, democratization has often been followed by decentralization.
This happened in Spain, where the death of Francisco Franco and the return to a
democratic system was soon followed by a massive process of decentralization.
Other examples are Poland, Czechoslovakia, Russia and the Ukraine. The end of
the Cold War also favored the rise of secessionist (or pro-decentralization) political
movements in countries that, although democratic, used to have an extremely rigid
political situation. In Italy, for instance, the end of the ‘Cold War equilibrium’ was
soon followed by the rise of a separatist political party. The model and empirical
analysis presented in this paper try to shed some light on these recent events.
Many authors think that it is not possible to find a single set of variables
explaining the cross-country differences in the degree of fiscal centralization.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
123
Oates (1972) attempt seems to support this view. This paper is more optimist on
the possibility of finding empirical regularities explaining decentralization. I find
that country size, income per capita, ethnic fractionalization, and the level of
democracy are negatively correlated with fiscal centralization.
The main problem with the results described above is their sensitivity to outliers
and to the use of different samples. While Area and GDP per capita are robustly
correlated with centralization in almost all specifications, ethnic fractionalization
and democracy have unstable coefficients, highly dependent on the sample used.
This is due to the fact that outliers play a very important role in determining these
coefficients. Yugoslavia is a particularly troublesome outlier. Its very low level of
fiscal centralization plays an enormous weight in determining the coefficients (and
t-statistics) attached to democracy and ethnic fractionalization. When Yugoslavia is
dropped from the sample, democracy becomes strongly correlated with decentralization. This is also true when more realistic values are used for fiscal centralization
in Yugoslavia.
To conclude, this paper shows that there are some regularities that can be used
to explain the cross-country variations in the degree of fiscal centralization. In
particular, if one believes that outliers are important and that they should not be
eliminated from the analysis, then it is possible to conclude that land area, GDP
per capita, and ethnic fractionalization are associated with fiscal decentralization.
If, however, one believes that outliers should be eliminated, the analysis shows a
strong correlation between fiscal centralization and each of land area, GDP per
capita, and Democracy.
This paper indicates that there are open avenues for further research both on the
empirical and theoretical sides. On the empirical side, it would be interesting to
identify a homogeneous definition of jurisdiction and augment Eq. (9) with a
variable measuring the number of jurisdictions. On the theoretical side there are
many possible extensions. The first relates to extending the results to a case of a
circular country and relaxing the assumption of perfect sorting of the citizens. It
would be interesting to show whether the results of this paper are robust to a case
where preferences are highly, but not perfectly, correlated with location.
Introducing income heterogeneity would be another interesting extension.
Section 2 shows that the equilibrium level centralization is decreasing in income.
Therefore, a model where individuals are sorted according to their income should
predict that the policies of the central government are closer to the preferences of
the rich. This could be used to explain the fact that upper classes tend to have
higher political participation and influence than poorer classes.
The results of Section 6 seem to indicate that the adjustment toward the optimal
fiscal structure takes time. This cannot be captured by the static model of Section
2. It would then be an interesting exercise to extend the model to a dynamic
framework and try to obtain some information on why some countries reached the
equilibrium faster than others or explain the overshooting that characterized other
countries (for instance Poland).
124
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Acknowledgements
This paper, based on the second chapter of my Ph.D. dissertation at the Johns
Hopkins University, was written while I was at the Department of Economics of
the University of Turin. I would like to thank Nada Choueiri, Ernesto Stein, two
anonymous referees, seminar participants at the JHU, University of Adelaide,
University of South Florida, Inter-American Development Bank, University of
Cagliari, and University of Turin for helpful comments. I am also grateful to
Laurence Ball and Louis Maccini for useful comments and constant support, and
Marisa Apa Domino for suggesting the use of the Honore´ and Powell estimator
and sharing her program with me. The usual caveats apply. The opinions
expressed in this paper are my own and do not necessarily reflect the views of the
Inter-American Development Bank.
Appendix A
Proof of Proposition 1
We want to show that g*5( m y /( m 1 b )) (with m 512 a [u (S / 4)1(12u )(S /
4J)]) is the amount of public good preferred by the median voter. This is
equivalent to showing that half of the population would like to consume an amount
of g larger than g* and the other half of the population would like to consume and
amount of g smaller than g*.
In Eq. (1) it was shown that gi 5(di y) /(di 1 b ), with di 512 a (u l im 1(12u )l ij ).
Note that, since we assume a continuum of individuals, for half of the population
l im ,(S / 4) and for the other half l im .(S / 4). Also, for half of the population
l ij ,(S / 4J) and for the other half l ij .(S / 4J). Therefore, when u 51 or u 50,
Proposition 1 is trivially true.
Since (≠gi / ≠di )5( b y /(di 1 b )2 ).0, individuals with larger d (and therefore
smaller l im and l ij ) prefer a larger amount of public good. It is therefore possible to
divide the population into four groups as shown in Fig. 2. For a fraction F of
Fig. 2. Preferences for the size of the public good.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
125
individuals (all located in the first quadrant), l im ,(S / 4) and l ij ,(S / 4J), therefore
di , m and gi ,g*. For a fraction F of individuals (all located in the fourth
quadrant) l im ,(S / 4) and l ij ,(S / 4J), therefore di . m and gi .g*.23
Next, we show that each individual in the second quadrant can be mapped into
an individual in the third quadrant with opposite preferences. Let us start by
defining pl the individuals in the second quadrant and lp the individuals in the
lp
third quadrant. Define l lp
m and l j the distance from the center of the country and
the center of the jurisdiction of an individual i belonging to the third quadrant,
lp
then there is an individual z in the second quadrant with l pl
m 5(S / 2)2l m and
lp
pl
pl
lp
l j 5(S / 2J)2l j . At this point, we are ready to show that (d z 2 m )5 2(d i 2 m )
i.e., for each individual in the second quadrant there is an individual in the third
quadrant with opposite preferences who will vote in the opposite way.
F
S
S
lp
lp
d lp
1 (1 2 u )]
z 2 m 5 1 2 a [u l m 1 (1 2 u )l j ] 2 1 1 a u ]
4
4J
S
S
5 a u ] 2 l lp
1 (1 2 u ) ] 2 l jpl
m
4
4J
FS
D
S
G
DG
(10)
pl
Using the fact that l lp
j 5(S / 2J)2l j it is possible to rewrite Eq. (10) as:
FS
D
S
S
S
d lp
2 l mlp 1 (1 2 u ) l jpl 2 ]
z 2m 5a u ]
4
4J
DG
(11)
Let us now look at d pl
j 2m
F
S
S
pl
pl
d pl
1 (1 2 u )]
i 2 m 5 1 2 a [u l m 1 (1 2 u )l j ] 2 1 1 a u ]
4
4J
S
S
5 a u ] 2 l pl
1 (1 2 u ) ] 2 l pl
m
j
4
4J
FS
D
S
G
DG
(12)
lp
pl
Using the fact that l pl
m 5S / 22l m it is possible to rewrite d j 2 m as:
FS
D
S
S
S
lp
d pl
1 (1 2 u ) ] 2 l pl
j 2 m 5 a u lm 2 ]
j
4
4J
DG 5 2 (d
lp
z
2 m)
(13)
This proves the fact that for each voter in the second quadrant there is a voter in
the third quadrant with opposite preferences. Therefore, given the distribution of
the voters and the fact that the preferences are single peaked g*, is a unique
Condorcet winner. Q.E.D.
23
It is possible to show that F 5((J21)S / 4J) when ((J21) / 2) is odd and F 5((J 11)S / 4J) when
((J21) / 2) is even. But the only thing that really matters for the proof is that the first and the fourth
quadrants have the same number of individuals.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
126
Appendix B
Proof of Proposition 2
The proof of this proposition requires some simple, but long and tedious,
algebra. I start by establishing some basic results:
a S(J 2 1)
mu 5 2 ]]] , 0; muu 5 0;
4J
F
G
S
S
2 S(J 2 1)
ma 5 2 u ] 1 (1 2 u ) ] , 0; mau 5 mua 5 ]]]] , 0;
4
4J
4J
F
G
1
1
2 a (J 2 1)
mS 5 2 a u ] 1 (1 2 u ) ] , 0; mu S 5 mSu 5 ]]]] , 0;
4
4J
4J
a (1 2 u )S
2 aS
mJ 5 ]]]
. 0; mJu 5 mu J 5 ]]
, 0;
2
4J
4J 2
b ymu
2 2b y( mu )2
≠g
≠c
] 5 2 ] 5 ]]]
]]]]
,
0;
g
5
, 0;
uu
≠u
≠u ( b 1 m )2
( b 1 m )3
b yma
mau ( b 1 m ) 2 2ma mu
≠g
≠c
] 5 2 ] 5 ]]]
, 0; gua 5 gau 5 b y ]]]]]]
, 0;
≠a
≠a ( b 1 m )2
( b 1 m )3
b ymS
mSu ( b 1 m ) 2 2mS mu
≠g
≠c
] 5 2 ] 5 ]]]
, 0; gu S 5 gSu 5 b y ]]]]]]
, 0;
2
≠S
≠S ( b 1 m )
( b 1 m )3
b ymJ
mJu ( b 1 m ) 2 2mJ mu ,
≠g
≠c
] 5 2 ] 5 ]]]
, 0; gu J 5 gJu 5 b y ]]]]]]
50
≠J
≠J ( b 1 m )2
.
( b 1 m )3
Using the fact that (≠g / ≠u )5 2(≠c / ≠u ), it is possible to write the government’s
F.O.C. as:
S
D
≠Vgov
12f
]] 5 f gu U med
2 U cmed 1 ]] u 1 (1 2 f )g 5 0
g
≠u
f
(14)
For the median voter of the central jurisdiction (and therefore the government),
Ug 2Uc 5U med ((1 /g)2( b /c))5(U med ( b 1 m ) /y)((12 m ) /m ),0 (because m ,1).
Therefore the first term of Eq. (14) is negative and the second is positive. By
deriving implicitly Eq. (14), it is possible to study the effect of the exogenous
variables on the level of fiscal centralization.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
S
f S
12f
Vuu 5 f guu U med
2 U cmed 1 ]]u
g
f
127
D
D
12f
med
med
1 gu gu (U med
1 (1 2 f )gu
gg 1 U cc 2 2U cg ) 1 ]]
f
(15)
med
med
Since guu ,0 and (U med
gg 1U cc 22U cg ),0, Vuu ,0. This result, besides being
necessary to compute the effect of the independent variables on the level of
centralization, also shows that the government’s utility function is concave and
therefore the F.O.C. of Eq. (14) is necessary and sufficient for a maximum.
S
D
12f
med
med
Vua 5 f gua U med
2 U med
1 ]]u 1 f gu ( ga (U med
g
c
gg 1 U cc 2 2U cg ))
f
1 (1 2 f )ga , 0
(16)
S
D
12f
med
med
Vu S 5 f gu S U med
2 U med
1 ]]u 1 f gu ( gS (U med
g
c
gg 1 U cc 2 2U cg ))
f
1 (1 2 f )gS , 0
(17)
Vuf 5 gu (U med
2 U cmed ) 2 ( guu 1 g) , 0
g
(18)
Note that Vuf ,0 because Eq. (14) requires that ( guu 1g).0.
S
D
12f
med
med
Vu J 5 f gu J U med
2 U med
1 ]]u 1 f gu ( gJ (U med
g
c
gg 1 U cc 2 2U cg ))
f
,
1 (1 2 f )gJ 50
(19)
.
To investigate the role of a change in income let us start by noting that
gu 5g( bmu /m ( m 1 b )), therefore Eq. (14) can be rewritten as:
F
≠Vgov
]] 5 g (1 2 f )
≠u
S
S
bmu
12f
U med ( b 1 m ) 1
1 f ]]] ]] u 1 ]]]] ] 2 1
f
y
m
m( m 1 b )
DDG
Since g50 is a minimum, it must be true that
F
S
bmu
U med ( b 1 m ) 1 2 m
(1 2 f ) 1 f ]]] u 1 ]]]] ]]
y
m
m( m 1 b )
It is now possible to show that:
S D
b
Vu y 5 fmu (1 2 m ) ]
my
2
gU med , 0
DG
5 0.
50
(20)
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
128
Therefore:
Vuf
Vu y
Vua
Vu S
≠u
≠u
≠u
≠u
]5 2]
, 0; ] 5 2 ] , 0; ] 5 2 ] , 0, ] 5 2 ]
≠a
Vuu
≠S
Vuu
≠f
Vuu
≠y
Vuu
Vu J ,
≠u
, 0, and ] 5 2 ] 50
≠J
Vuu .
A corollary to Proposition 3 is that f 51 will lead the central government to set
u 50. This is easily shown by observing that, when f 51, Eq. (14) becomes
(≠Vgov / ≠u )5gu (U med
2U cmed ),0. Q.E.D.
g
To prove that a Benthamite social planner will set u 50 it is sufficient to show
that the social planner’s utility function is strictly decreasing in u. Let us start by
writing the social planner’s problem:
N
max U
sp
u
N
E
E g c di
0
0
di b
5 Ui di 5
≠U sp
]] 5 ln( g)
≠u
N
E
(21)
N
≠di
gu
Ui ] di 1 ]
≠u
g
0
E U [d 2 m ]di
i
i
(22)
0
N
Now note that the ln ( g)e Ui (≠di / ≠u ) di , 0 (because g.1). Proposition 1 shows
0
that m is the median (and, given the assumption
of uniform distribution of
N
individuals, also the mean) value of di . Therefore e [di 2 m ]di 5 0, but since in0
dividuals for which di . m have levels of utility higher than those of individuals for
N
which di , m, then e Ui [di 2 m ]di . 0, the second term of Eq. (22) will then be nega0
tive and (≠U sp / ≠u ),0. Q.E.D.
Appendix C
Data sources
This appendix describes the source of the data used.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
129
Table 8
Centralization ratios
Revenues
1975
Argentina
Australia
Austria
Barbados
Belgium
Bolivia
Brazil
Canada
Cent. Afr. Rep.
Chile
Colombia
Congo
Costa Rica
Cyprus
Denmark
Dominican Rep.
Egypt
Ethiopia
Finland
France
Gambia
Germany
Greece
Guatemala
Honduras
Hungary
Iceland
India
Indonesia
Iran
Iraq
Ireland
Israel
Italy
Jordan
Kenya
Korea
Kuwait
Lesotho
Luxembourg
Malawi
Malaysia
Malta
Mauritius
Mexico
Expenditure
1980
1985
1975
1980
1985
84.36
76.96
72.37
100.00
92.70
86.28
78.20
47.37
70.46
85.04
73.48
100.00
72.72
79.01
75.66
100.00
91.75
69.08
53.97
77.49
50.55
74.50
78.00
75.25
100.00
95.03
95.82
70.59
52.16
94.25
80.03
96.96
90.01
96.28
83.55
94.89
81.90
95.80
96.46
97.75
97.47
67.76
98.29
69.77
98.49
100.00
97.52
69.35
89.38
96.38
98.53
56.62
98.45
100.00
69.30
98.84
100.00
98.56
72.11
90.71
97.33
61.85
71.03
76.21
78.62
72.81
100.00
60.86
79.29
74.71
99.96
94.68
71.27
53.18
79.18
47.42
100.00
97.53
82.68
98.24
85.00
79.77
96.93
97.43
66.60
98.02
100.00
70.26
89.82
93.29
62.20
97.38
97.71
69.97
90.32
94.60
62.94
63.95
73.07
88.11
93.54
62.12
97.43
95.57
85.18
95.51
94.65
92.00
91.41
85.54
70.80
89.96
63.88
96.00
93.76
79.79
67.34
96.55
98.49
100.00
95.69
87.87
74.83
73.79
97.63
91.70
87.58
74.06
70.36
97.45
95.16
92.22
93.26
94.66
92.76
93.74
94.33
94.23
100.00
100.00
99.82
100.00
88.96
86.48
100.00
91.29
82.77
84.11
99.98
92.23
82.51
100.00
91.73
94.11
90.43
100.00
81.71
94.33
84.67
100.00
81.57
80.35
87.95
81.61
74.26
97.23
98.69
87.73
95.59
91.11
94.35
98.88
89.38
73.69
98.58
97.47
95.54
96.70
86.90
76.08
77.94
98.95
95.72
74.92
96.84
95.52
77.31
95.45
96.35
100.00
100.00
90.14
96.51
85.85
100.00
86.88
94.45
86.51
100.00
99.08
87.52
81.41
130
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Table 8. Continued
Revenues
1975
Myanmar
Netherlands
New Zealand
Norway
Pakistan
Panama
Paraguay
Peru
Poland
Portugal
Romania
Senegal
Singapore
South Africa
Spain
Sri Lanka
Sudan
Sweden
Switzerland
Thailand
Trinidad
Tunisia
United Kingdom
United States
Uruguay
Venezuela
Yemen
Yugoslavia
Zaire
Zimbabwe
No. obs.
Mean
S.D.
100.00
97.14
Expenditure
1980
96.35
96.78
95.63
88.80
82.13
72.75
97.69
95.51
96.48
88.73
90.14
66.95
100.00
95.09
94.31
71.24
52.15
92.87
97.42
84.41
60.43
89.68
99.28
100.00
72.75
57.00
87.31
13.15
1985
1975
100.00
95.90
66.66
96.67
83.11
69.16
83.75
97.20
96.86
97.68
95.57
84.28
89.56
1980
1985
100.00
94.91
89.61
79.40
78.82
98.20
95.14
87.37
100.00
95.79
90.85
90.55
100.00
84.66
90.37
98.26
100.00
86.28
87.37
97.70
74.02
55.46
96.40
74.33
47.94
96.05
85.99
68.18
91.59
89.93
68.63
92.50
100.00
29.62
100.00
81.87
55.00
86.90
14.49
100.00
28.36
82.16
98.09
95.01
86.15
91.26
92.57
91.03
83.43
100.00
100.00
82.20
88.87
96.51
100.00
85.45
85.16
94.54
69.31
53.25
91.61
99.76
73.17
53.28
94.14
94.20
96.48
93.13
69.61
53.81
94.44
96.88
86.53
60.11
90.50
84.44
65.36
92.52
86.86
63.35
90.93
98.55
100.00
29.17
100.00
77.64
56.00
85.49
15.12
100.00
27.14
100.00
82.05
56.00
86.19
14.88
100.00
55.00
86.96
13.21
85.10
54.00
85.84
14.74
1. Data on fiscal centralization. The data on fiscal centralization were computed
as the percentage of central government revenues (expenditure) on the total
(central plus local) government revenues (expenditure). The data on central and
total government (general government, in the IMF terminology) revenues and
expenditures were obtained from the IMF (1975, 1980 and 1985) Government
Finance Statistics Yearbook. When the full set of data was not available for the
years considered in this study (i.e., 1975, 1980 and 1985) but it was available
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
2.
3.
4.
5.
131
for the previous or following year the available data were used. The data for
1975 were not available for Zimbabwe and Sudan (1976 was used instead of
1975). The data for 1980 were not available for the Central African Republic,
Hungary, Pakistan, Venezuela and Egypt. For the first two countries 1981 data
were used and for the last three countries 1979 data were used. Data for 1984
instead of 1985 were used for Bolivia, Colombia, Kenya, Malawi, Switzerland
and Uruguay.
Data on ethnic fractionalization. The data on ethnic fractionalization,
originally published in the Atlas Narodov Mira, were obtained from Taylor and
Hudson (1972).
Data on democracy. The data on democracy are from Gastil (1990); Barro
(1996). The data were available for 1975, 1978 and 1988 only. The 1978 data
were used to measure democracy in 1980 and the average between 1978 and
1988 was used to measure democracy in 1985.
Data on area and population. World Development Indicator (World Bank,
1990).
Data on GDP. The data on GDP are from Summers and Heston (1991).
Appendix D
OLS estimations
Given the fact that the dependent variable is censored at 100, OLS is not the
appropriate technique to estimate the equations described in Section 4. This
appendix presents the results of least squares regressions in order to compare them
with Oates (1972) results.
The OLS estimates are presented in Table 9. When the dependent variable is
total revenue, Area and GDP per capita have negative and statistically significant
coefficients.24 Fract has always a negative coefficient, not statistically significant
for 1975 and statistically significant at 2.5% and 1% level for 1980 and 1985.
Democracy seems to play an important role in explaining fiscal centralization in
1975, but this effect disappears in 1980 and 1985.
The expenditure regressions do not fit the data as well as the revenue
regressions, but the results of the two sets of regressions are similar. As for the
revenue regressions, Democracy has an important weight in explaining centraliza-
24
I tried to introduce the square of GDP per capita to control for non-linearities, but its coefficient
turned out not to be statistically significant.
132
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
Table 9
Least squares estimates
Revenues centralization ratios
1975
Area
y
Fract.
Dem.
Const.
No. obs.
2
R̄
F
Area
y
Fract.
Dem.
Const.
No. obs.
R̄ 2
F
1980
1985
22.51
22.177
(24.082)***
(22.800)***
23.723
28.834
(21.868)*
(23.633)***
25.480
214.982
(20.994)
(22.054)**
211.947
20.087
(22.592)***
(20.014)
157.066
191.66
(9.980)***
(8.964)***
56
56
0.40
0.31
13.32
12.05
Expenditure centralization ratios
22.474
(23.532)***
27.643
(23.326)***
214.613
(22.409)***
24.862
(20.774)
188.645
(9.488)***
56
0.39
19.32
1975
1980
1985
22.498
(23.901)***
23.363
(21.592)
20.850
(20.147)
211.987
(22.183)**
151.744
(9.095)***
55
0.35
10.55
2.183
(23.016)***
27.691
(23.549)***
213.853
(22.195)**
21.179
(20.206)
183.642
(9.531)***
56
0.33
14.46
22.396
(23.267)***
28.93
(23.430)***
2 10.993
(21.682)
20.003
(0.000)
194.645
(8.602)***
54
0.34
15.81
t-statistics in parentheses.
Parameters indicated are statistically significant at: * 5%; ** 2.5%; *** 1%.
tion in 1975, but no importance in 1980 and 1985 (the coefficients attached to
democracy in 1980 and 1985 are higher in the revenues regressions). Fract is
statistically significant at the 2.5% level in the 1980 regression and statistically
significant at the 5% level in the 1985 regression.
Even if fundamentally flawed, the OLS analysis is instructive. Using the OLS
estimate it is possible to compare the results of this paper with Oates (1972) and
conclude that the differences are not only due to different estimation techniques,
but also to the use of different data sets.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
133
Table 10
Tobit estimates: Interaction dummies for sub-Saharan Africa
Revenues
Area
y
Fract.
Dem.
Af *Area
Af *y
Af *Frac
Af *Dem
Const.
No. obs.
x2
Expenditure
1975
1985
1975
1985
23.33
(24.78)***
25.82
(22.62)***
23.25
(20.51)
212.73
(22.34)**
13.12
(1.71)*
223.31
(21.92)*
242.84
(21.04)
63.08
(1.55)
186.15
(10.21)***
57
43.95
23.31
(24.09)***
29.58
(23.38)***
220.57
(22.91)***
27.27
(21.07)
223.02
(21.63)
48.29
(1.50)
121.99
(1.77)*
2494.06
(21.45)
220.55
(8.79)***
56
42.41
23.08
(24.37)***
23.22
(21.39)
2.05
(0.31)
213.38
(22.29)**
6.30
(1.31)
29.27
(21.10)
232.20
(20.91)
25.11
(0.75)
158.69
(8.54)***
55
32.50
23.24
(23.89)***
211.91
(23.87)***
215.41
(22.09)**
20.71
(20.10)
25.35
(20.45)
22.59
(20.11)
44.61
(0.68)
195.93
(0.83)
234.22
(8.46)***
54
33.09
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
Appendix E
Sensitivity analysis
Tables 10–14 show the result of the sensitivity analysis.
Appendix F
Description of the Honore´ and Powell pairwise difference estimator
Maximum likelihood estimators can be used to estimate the censored regression
model
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
134
Table 11
Tobit estimates: Interaction dummies for Asia
Revenues
Area
y
Fract.
Dem.
As*Area
As*y
As*Frac
As*Dem
Const.
No. obs.
x2
Expenditure
1975
1985
1975
1985
22.80
(23.71)***
26.67
(22.49)***
214.02
(22.06)**
28.39
(21.26)
22.94
(21.42)
4.74
(1.42)
33.61
(2.02)*
221.51
(21.44)
185.32
(9.69)***
57
43.07
22.40
(22.98)***
213.77
(24.97)***
222.34
(23.27)***
9.42
(1.20)
21.81
(20.39)
7.99
(1.82)*
20.03
(20.001)
256.98
(22.89)***
232.22
(10.71)***
56
54.41
21.99
(23.03)***
28.88
(23.55)***
211.21
(21.78)*
28.26
(21.35)
27.72
(24.19)***
10.38
(3.43)***
55.10
(3.76)***
217.03
(21.27)
193.36
(11.02)***
55
49.71
22.20
(22.53)***
215.29
(24.40)***
219.76
(22.60)***
15.90
(1.76)*
23.24
(21.44)
7.25
(2.19)**
17.14
(0.79)
234.58
(21.95)*
237.77
(9.03)***
54
40.70
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
y i 5 maxh0, x i9 b 1 ´i j
(23)
only if the distribution of ´i is known. In particular, if ´i |N(0, s 2 ), the Tobit
model produces consistent estimates of b. Maximum likelihood estimators are in
general biased if the distribution of ´i is mis-specified. In the recent years, a large
number of semi-parametric estimators that do not require assumptions on the form
of the distribution of ´i have been proposed. Among these estimators, the pairwise
difference (PD) estimator studied by Honore´ and Powell (1994) seems to perform
particularly well in small samples. This appendix briefly describes the general idea
of PD.
The censoring in Eq. (23) has the effect that y i 2x i9 b 5maxh2x i9 b, ´i j and
y j 2x 9j b 5maxh2x j9 b, ´j j are not identically distributed even if ´i and ´j are i.i.d.
Honore´ and Powell show that it is possible to obtain i.i.d. residuals by censoring
y i 2x 9i b and y j 2x 9j b from below at x 9j b and x 9i b, respectively. In particular, let
e ij ( b );maxh´i , 2x j9 b, 2x i9 b j and e ji ( b );maxh´j , 2x i9 b, 2x j9 b j, then e ij ( b )
and e ji ( b ) have the same conditional distribution. Furthermore, since i and j are
randomly sampled, e ij ( b ) and e ji ( b ) are conditionally independent and therefore
(e ij ( b )2e ji ( b )) is symmetrically distributed around 0 conditional on x i and x j .
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
135
Table 12
Tobit estimates: Sensitivity analysis using different definitions of size (1985)
Revenues (1985)
23.97
(23.33)***
1.34
(0.84)
Area
Pop
Inc
y
Frac
Dem
Const.
No. obs.
x2
1.35
(0.85)
28.74
210.09
(23.42)***
(23.48)***
219.38
219.38
(22.84)***
(22.84)***
28.83
28.83
(21.28)
(21.28)
200.29
200.29
(7.14)***
(7.14)***
56
56
39.13
39.13
Expenditure (1985)
23.81
(23.11)***
1.36
(0.82)
Area
Pop
Inc
y
Frac
Dem
Const.
No. obs.
x2
23.97
(23.33)***
29.73
(23.44)***
215.95
(22.22)**
22.45
(20.35)
201.30
(6.37)***
54
30.67
23.81
(23.11)***
1.36
(0.82)
211.09
(23.51)***
215.95
(22.22)**
22.45
(20.35)
201.30
(6.37)***
54
30.67
2 3.97
(23.33)***
1.31
(0.75)
0.00
(0.04)
28.77
(23.31)***
219.39
(22.84)***
28.87
(21.27)
201.06
(6.06)***
56
39.13
23.83
(23.12)***
1.04
(0.58)
0.00
(0.44)
210.06
(23.43)***
216.14
(22.24)**
22.83
(20.41)
209.39
(5.90)***
54
30.87
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
Area, Pop, Inc and y are in logs, except for the last column, where Inc is in levels.
This allows Honore´ and Powell to use the moment condition of Eq. (24) to define
an estimator for b.
E[ j (e ij ( b ) 2 e ji ( b ))ux i , x j ]
(24)
where j (?) is an odd function. Since it is not always easy to show that a set of
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
136
Table 13
Tobit estimates: Sensitivity analysis using different definitions of size (1975)
Revenues (1975)
23.71
(23.47)***
1.08
(0.41)
Area
Pop
Inc
y
Frac
Dem
Const.
No. obs.
x2
23.71
(23.47)***
1.08
(0.41)
24.43
25.50
(22.04)**
(22.28)**
26.57
26.57
(21.11)
(21.11)***
213.38
213.38
(22.41)**
(22.41)**
162.37
162.37
(7.72)***
(7.72)***
57
57
38.26
38.26
Expenditure (1975)
22.89
(22.66)***
20.50
(20.04)
Area
Pop
22.89
(22.66)***
20.50
(20.04)
23.30
(21.34)
1.99
(0.33)**
213.89
(22.40)**
158.74
(7.40)***
55
30.40
Inc
y
Frac
Dem
Const.
No. obs.
x2
23.35
(21.52)
1.99
(0.33)
213.89
(22.40)**
158.74
(7.40)***
55
30.40
23.67
(23.47)***
1.39
(1.02)
0.00
(0.67)
24.08
(21.84)*
26.09
(21.03)
213.00
(22.35)**
154.30
(6.41)***
57
38.70
22.88
(22.70)***
0.03
(0.02)
0.00
(0.17)
23.25
(21.43)
22.08
(20.34)
213.78
(22.36)**
156.61
(6.26)***
55
30.42
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
Area, Pop, Inc and y are in logs, except for the last column, where Inc is in levels.
moment conditions are satisfied at a given point, Honore´ and Powell define the
estimator as the minimand of the following random function:
SD
n
Sn (b) 5 2
21
O s( y , y , d )
i
i ,j
where d 5(x i 2x j )9b and:
j
(25)
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
137
Table 14
Tobit estimates: Sensitivity analysis using different definitions of size (1975)
Revenues (1975)
23.71
(23.47)***
1.08
(0.41)
Area
Pop
Inc
y
Frac
Dem
Const.
No. obs.
x2
1.08
(0.41)
24.43
25.50
(22.04)**
(22.28)**
26.57
26.57
(21.11)
(21.11)***
213.38
213.38
(22.41)**
(22.41)**
162.37
162.37
(7.72)***
(7.72)***
57
57
38.26
38.26
Expenditure (1975)
22.89
(22.66)***
20.50
(20.04)
Area
Pop
Inc
y
Frac
Dem
Const.
No. obs.
x2
23.71
(23.47)***
23.35
(21.52)
1.99
(0.33)
213.89
(22.40)**
158.74
(7.40)***
55
30.40
22.89
(22.66)***
20.50
(20.04)
23.30
(21.34)
1.99
(0.33)**
213.89
(22.40)**
158.74
(7.40)***
55
30.40
23.67
(23.47)***
1.39
(1.02)
0.00
(0.67)
24.08
(21.84)*
26.09
(21.03)
213.00
(22.35)**
154.30
(6.41)***
57
38.70
22.88
(22.70)***
0.03
(0.02)
0.00
(0.17)
23.25
(21.43)
22.08
(20.34)
213.78
(22.36)**
156.61
(6.26)***
55
30.42
t-statistics in parentheses.
Indicated parameters are statistically significant at: * 5%; ** 2.5%; *** 1%.
Area, Pop, Inc and y are in logs, except for the last column, where Inc is in levels.
J ( yi ) 2 ( y j 1 d )j ( yi )
J ( yi 2 y j 2 d )
s( y i , y j , d ) 5
J ( y j ) 2 (d 2 yi )j (2y j )
5
if d # 2 y j
if 2 y i , d , 2 y i
if y i # d
where J (d): R→R 1 is an even (i.e., J (2d)5 J (d)) loss function that must meet
the following criteria: J (d) is convex; J (d) is strictly increasing for d .0; and
138
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
J (0)50 is differentiable everywhere except at 0. j (d)5 J 9(d) is the right
derivative of J (d) at 0. In my estimations I use J (d)5udu. A possible alternative
is J (d)5d 2 .25 The following steps have been followed in the estimation reported
in Table 4.
1. Since fiscal centralization cannot be higher than 100, the censored regression
model to be estimated can be described as: y i 5minh100, x i9 b 1 ´i j. To
transform this model into one similar to Eq. (23), it is necessary to add 100 to
both sides of the equation and multiply all the variables by 21.26
n
2. Generate 2 pairs by matching all the observations. For each pair of
observations the dependent variables are y i and y j and the independent variables
(x i 2x j ).
3. Estimate b by minimizing Eq. (25).
4. Use bootstrap techniques to build confidence intervals. The pseudo p-values
were generated by bootstrapping the data with 1000 replications.
SD
References
Ades, A., Glaeser, E., 1995. Trade and circuses: Explaining urban giants. Quarterly Journal of
Economics 110, 195–227.
Alesina, A., Baqir, R., Easterly, W., 1996. Public goods and ethnic division, Mimeo, The World Bank.
Alesina, A., Perotti, R., 1996. Budget deficits and budget institutions, Mimeo, International Monetary
Fund.
Alesina, A., Perotti, R., Spolaore, E., 1996. Together or separately? Assessing the costs and benefits of
political and fiscal unions. European Economic Review 39, 751–758.
Alesina, A., Spolaore, E., 1997. On the number and size of nations. Quarterly Journal of Economics
112, 1027–1056.
Barro, R., 1996. Determinants of democracy, Mimeo, Harvard University.
Borjas, G., 1995. Ethnicity, neighborhoods, and human capital externalities, The American Economic
Review 85, 365–390.
Brennan, G., Buchanan, J., 1980. The Power to Tax: Analytical Foundations of a Fiscal Constitution,
Cambridge University Press, Cambridge, UK.
Breton, A., 1996. Competitive Governments, Cambridge University Press, Cambridge, UK.
Breton, A., Scott, A., 1978. The Economic Constitution of Federal States, University of Toronto Press,
Toronto.
Canning, D., Fay, M., 1993. The effect of transportation networks on economic growth, Discussion
Paper, Department Of Economics, Columbia University.
Castilla, M., 1996. Political instability, investment, and economic growth: An empirical investigation,
Mimeo, The Johns Hopkins University.
Easterly, W., Levine, R., 1997. Africa growth tragedy: Policies and ethnic divisions. Quarterly Journal
of Economics 112, 1203–1250.
25
The reason for using udu instead of d 2 is that with the second function the conditions for
consistency and asymptotic normality are stricter.
26
After the estimation, it is necessary to add 100 to the constant.
U. Panizza / Journal of Public Economics 74 (1999) 97 – 139
139
Gastil, R.D., 1990. The comparative survey of freedom: Experience and suggestions. Studies in
Comparative International Development 25, 25–50.
Greene, W., 1993. Econometric Analysis, Prentice-Hall, New York.
Gurr, T., Jaggers, K., Moore, W., 1989. The Polity II Codebook, University of Colorado at Boulder,
Center for Comparative Politics.
Hamilton, B., Mills, E., Puryear, D., 1975. The Tiebout Hypothesis and residential income segregation.
In: Mills, E., Oates, W. (Eds.), Fiscal Zoning and Land Use Control, Heath-Lexington, Lexington,
MA.
´ B., Powell, J., 1994. Pairwise difference estimators of censored and truncated regression
Honore,
models. Journal of Econometrics 64, 241–278.
International Monetary Fund, 1975, 1980 and 1985. Government Finance Statistics Yearbook,
Washington, DC.
Jurion, B., 1983. A theory of public services with distance-sensitive utility. In: Thisse, J., Zoller, H.
(Eds.), Location Analysis of Public Facilities, North Holland, Amsterdam.
Mauro, P., 1995. Corruption and growth. Quarterly Journal of Economics 110, 681–712.
McKelvey, R., 1976. Intransitivities in multidimensional voting models. Journal of Economic Theory
12, 472–482.
Musgrave, R., 1959. The Theory of Public Finance, McGraw-Hill, New York.
Oates, W., 1972. Fiscal Federalism, Harcourt Brace Jovanovich, New York.
Powell, J., 1984. Least absolute deviations for the censored regression model. Journal of Econometrics
25, 303–325.
Powell, J., 1986. Symmetrically trimmed least squares estimations for Tobit models. Econometrica 32,
1435–1460.
Pryor, F., 1968. Public Expenditure in Communist and Capitalist Nations, George Allen and Unwin,
London.
Summers, R., Heston, A., 1991. Penn World Table (Mark 5): An expanded set of international
comparisons. Quarterly Journal of Economics 106, 326–368.
Taylor, C., Hudson M., 1972. World Handbook of Political and Social Indicators, Yale University Press,
New Haven.
Tiebout, C., 1956. A pure theory of local expenditures. Journal of Political Economy 64, 416–424.
Wallis, J., Oates, W., 1988. Decentralization in the public sector: An empirical study of state and local
government. In: Rosen H. (Ed), Fiscal Federalism: Quantitative Studies, University of Chicago
Press, Chicago.
Wheare, K., 1964. Federal Government, Oxford University Press, London.
Wildasin, D., 1991. Income redistribution in a common labor market. The American Economic Review
81, 757–774.
The World Bank, 1969. The World Tables, Washington, DC.
World Bank, 1990. World Development Indicator.