Does yardstick competition evolve in time

XXV Conferenza Siep
Pavia, 27 settembre 2013
Does yardstick competition evolve in time?
Evidence from Italian Municipalities
Fabio Padovano
CREM-CNRS et Centre Condorcet, Université de Rennes 1, Rennes, France
Ilaria Petrarca
Department of Economics, University of Verona, Verona Italy
Introduction
• This work tests whether the repetition of strategic interactions
among local governments makes yardstick competition (YC) more
or less intense and selective in time
• The empirical analysis follows the evolution of this environment in
Italy
– The pattern of interaction, as we expected, is not constant over time
– The disciplining force of YC wanes off in time
– YC seems to have led to a selection of better mayors through time
2
Issues in the literature
• The literature developed since the 1990s (Besley and Case, 1995;
Brueckner, 2003; Bordignon et al., 2003)
• It is wide and well known, but still incomplete: strategic
interaction is analyzed as a one-shot game
• Theoretical issue: YC models implicitly rely on the assumption of
an informational spillover expiring after each election
– Implicit prediction of a constant mimicking in time (Bordignon et al.,
2003)
• Empirical issues: the time dimension has been neglected in the
literature
– Cross-sectional data (Fiva and Rattso, 2007; Santolini, 2008;
Bordignon et al., 2003; Allers and Elhorst, 2005; Dubois and al.,
2010)
– or estimation of average interaction on panel data (Besley and Case,
1995; Schaltegger and Küttel, 2002; Solé-Ollé, 2003; Vermeir and
Heyndels, 2006; Bosch and Solé-Ollé, 2007; Dubois et Paty, 2010)
– As the time dimension increases, interactions may vary their
magnitude and/or direction
3
The dynamics of yardstick
competition
• What happens if we remove the assumption of a constant
spillover?
• We should observe an accumulation process of the interjurisdictional informational spillovers
– The circulation of information about mayors’ fiscal performance improves
(Franzese, 2001)
– Voters ‘learn’ how to implement the comparative electoral strategy envisaged in
YC models (Meseguer, 2009)
– Mimicking behavior becomes progressively less political rewarding and the
average quality of the selected mayors improves (Sangupta and Van Long,
2008), a result actually in line with the logic of the original Besley and Case
(1995) model
• The scope for pooling equilibria decreases we should observe YC
becoming less intense – i.e., similarities across jurisdictions in the
policy variables should decrease
• So far, there is no reference literature in public finance
4
Industrial organization literature
•
•
‘Learning by prices’ games as the ‘learning by tax rates’ game of YC
Benabou and Gertner (1992) studied strategic competition in a market
with two sellers of different types trading a homogeneous good to a
customer, assuming the price as a performance indicator that reveals the
true type of the seller
–
–
•
Dynamic extension by Bar-Isaac (2003): customers observe both present
and past prices
–
–
•
Seller = politician, customers = voter, price set by the seller = tax rate set by the politician,
markup = rent diverted, incentive to sell = electoral rewards
They reached the same theoretical results as Besley and Case (1995): existence of a pooling
equilibrium in which the bad seller mimics the good seller by reducing the markup and still can
be considered as the good type by the customers
In time the probability of observing a pooling equilibrium decreases up to the point where only
the good seller survives in the market
By similarity, we expect that when such a learning process applied to YC reduces the
probability of mimicking and only the good incumbent finds it optimal to run for re-election
Limitation: game theoretical results that are hard to translate into testable
empirical predictions
5
Political science literature
•
•
‘Learning from experience’
Meseguer (2006 and 2009)
– Model of policy diffusion in this field analyze the decision of a policy maker (the
agent) under the hypothesis that the decision to adopt a policy depends on the
realization of alternative policies previously implemented in the domestic and in
the neighboring countries
– If several policy makers face a decision and are exposed to the same stock of
information, their beliefs on the performance of the policy converge and they will
select the best performing one among the feasible set of alternatives.
•
Testable predictions
– Bayesian econometrics, dynamic updating of agent’s beliefs
– Empirical test on a sample of south-American countries during the 90s
– The implementation of institutional and economic reforms has been determined
by a learning process consistent with the theory
6
Macroeconomic literature
•
Per capita income convergence: estimation of the β coefficient indicating
the speed of convergence
– The structural equation is re-estimated on time subsamples of the full panel
dataset to observe the variation in the growth rate of the GDP through time.
– Barro and Sala-i-Martin (1991) and Blanchard (1991) verify the convergence
pattern predicted by the neoclassical theory of exogenous growth by applying
this methodology on a dataset of US states over various periods from 1840 to
1988
•
Recently this literature introduced models of growth convergence that
account for the geographic interaction of GDP growth among
observations spatial econometrics regressions, widely used also in the
YC literature
– Arbia et al. (2005), applied this methodology on a dataset of Italian provinces
during the period 1951-2000
7
The Italian dataset - 1
•
All the 8100 Italian Municipalities,1995-2004 (Padovano, 2007)
•
In 1993 a series of reforms created the suitable environment for YC
•
In 1995 the ICI tax rate has been distinguished between residential (HICI)
and business (BICI) property
HICI best suited for YC
•
– Our dataset follows the evolution of YC in Italy from its beginning to its end
– Introduction of the direct election of the mayor since 1995
– Introduction of the property tax rate (‘Imposta Comunale sugli Immobili’, ICI) to
finance municipal expenditures
– Main source of local tax revenue (about 55% of local tax revenue)
– Highly autonomous local tax revenue
• Level ‘b’ in OECD scale (from ‘a’ to ‘e’)
• Tax rate is set by the mayor in the range [0.004,0.007]
– The least mobile tax base in the Municipality
•
• More than 80% Italians own the house in which they live in
The development of YC came to a halt, or at last witnessed a structural
change, when in 2008 HICI was abolished by the central government and
replaced with transfers
8
The Italian dataset - 2
2. The longitudinal dimension of the dataset is exploited to relax the
assumption that all variables are on their long-run equilibrium steady state
3. The presence of strategic interactions among Italian municipalities
consistent with the YC hypothesis has already been verified
•
– Padovano and Petrarca (2011, forthcoming EJPE) tested the consistency
between the responsibility hypothesis and strategic interactions on the same
panel dataset
The literature on YC in Italy usually focused on business ICI tax rate
(ICIB) and failed to verify strategic interactions driven by electoral
concerns (Bordignon et al., 2003; Depalo and Messina, 2011)
9
Methodology
•
•
We estimate the spatial correlation in local ICIH tax setting among the
municipalities on incremental time datasets
The baseline time interval 1996-1998 has been incremented by one year
at a time until 2004
– Since the spatial GMM estimator applied requires at least 3 years to build the
necessary instruments, the shorter subsample is 1996-1998
– In total this procedure yields seven regressions
•
We use a yearly specification of the breaks because
–
–
•
a priori there are not evident breaks in the YC theory to control for and
we are interested in exploiting the whole information provided by the dataset
The spatial estimation follows the linear regression GMM panel data model
of Kapoor et al. (2007) with time and space fixed effects
10
The empirical specification
• Each observation i=1,…,N is observed for t=1,…,T periods.
Data are generated according to the following process:
ICIHit = β’Zit + uit
–
–
–
–
–
ICIHit: Nx1 vector of observations on the dep. variable in period t
Zit: NxK matrix of observations on exogenous regressors in period t
β’: Kx1 vector of regression parameters
uit:Nx1 vector of disturbance terms
The disturbances are assumed to be both correlated over time and across spatial
units, as well as heteroskedastic; they follow a Cliff and Ord first order spatial
autoregressive process:
uit = ρWi uit + εt
– 0<ρ<1: spatial autoregressive coefficient
– Wi : NxN spatial weights matrix (contiguity)
– ε: Nx1 vector of innovations following a one-way error component model
grouped by time periods: εit,N = μi,N + νit,N
11
The vector Z
Variable
BICI lag
Grants
Area
Pop
Depratio
Touristic
GDP
Right wing
Definition
ICI business tax rate lagged one period
Transfers from the central government
Surface area
Population
Dependency ratio
Touristic dummy
GDP per capita
Partisanship of executive dummies
Expected sign
+
NA
NA
NA
NA
+
-
Left wing
Partisanship of executive dummies
+
Center wing
Partisanship of executive dummies
NA
Local list
Partisanship of executive dummies
Electoral year dummy
Term limit dummy
Elec_year * Term limit
Union dummy
Domestic Stability Pact dummy
Number of neighbors
Province capital dummy
Coast dummy
NA
Elec_year
Term limit
Elec_tl
Union
DSP
N_neighbors
Provcap
Coast
+
+
NA
+
+
12
Benchmark estimation
Dep. var
•
Ln HICI
Spatial lag
0.157
Spatial error
0.034
BICI lag
0.236
***
0
***
Grants per capita
Area
-0.005
Population
-0.001
***
•
Depratio
0
Touristic
0.006
GDP per capita
0.005
Left wing
0.001
Right wing
-0.007
Local list
0.001
Electoral_year
-0.006
***
Term limit
-0.004
***
Electoral_term limit
0.003
*
Union
0.003
*
DSP
-0.021
***
N_neighbors
0.005
Observations
60255
Hausman test χ2
p-value
•
The Table shows the estimation of the
tax setting during the period, 1995-2004
Presence of a significant spatial pattern,
both in the form of spatial lag (+0.16)
The covariates are significant and show
the expected sign
–
***
•
the electoral variables suggest that, consistently
with the yardstick competition hypothesis, term
limited incumbents facing re-election set higher
tax rates than non term limited ones
Does the spatial coefficient, and hence
YC, remains stable through time?
2245.48
0
Notes: Source: Padovano and Petrarca, 2011. Dep.var. ln HICI,
continuous variables in ln. 6695 observations per year.
Robust estimations. Significance levels: *10%, **5%,***1%
13
Evolutionary estimation
1995-1998
1995-1999
1995-2000
1995-2001
1995-2002
1995-2003
1995-2004
Spatial lag
Coef.
0.76
0.605
0.511
0.302
0.262
0.229
0.157
p-value
***
***
***
***
***
***
***
Spatial error
Coef.
-0.71
-0.505
-0.35
-0.068
-0.03
-0.009
0.034
LM test p-value
***
***
***
***
***
***
***
Notes: Spatial panel regression with time and space fixed effects..
Dep.var. ln HICI, continuous variables in log. 6695 observations per year.
Significance levels: *10%, **5%,***1%.
•
•
•
Spatial coefficients are statistically significant
The signs of the coefficients confirm the general pattern of interaction
mainly in the outcome (spatial lag) and not in the residuals (spatial error)
The absolute value of the coefficients decreases as the length of the time
interval increases
14
The time dynamics of the spatial
correlations
1
0.8
0.6
0.4
0.2
0
-0.2
1995-1998 1995-1999 1995-2000 1995-2001 1995-2002 1995-2003 1995-2004
-0.4
-0.6
-0.8
Spatial lag
Spatial error
•Progressive decrease of the spatial coefficients over time
•consistent with the hypothesis of a progressive reduction of spatial interaction caused
by reduced incentives for the incumbent to mimic
•A lower share of the domestic tax rate is determined by the neighbors’ tax rate:
municipalities with high (low) tax rates are still observed near municipalities with high (low)
tax rates, but these similarities become less evident
15
The Regional analysis
• To verify the robustness
and generality of the
results we have
estimated the dynamics
of the spatial coefficients
on the geographic
subsamples of the 15
Italian Ordinary regions
• Umbria dropped because
of too small sample size!
16
The dynamics across the Regions - 1
• All the regions replicate the national pattern
– Safe for Veneto and Lazio
• Among them, there are the regions with the largest density of
municipalities measured as the number of municipalities per
square km(Piemonte, Liguria, and Lombardia) and the highest
average of neighbors (Basilicata, EmiliaRomagna, Molise),
• A large density of municipalities in a region is associated to a
decrease of interaction
– in Piemonte the there is a discontinuity in 2001
– in Lombardia the decrease follows the ‘first order’ electoral years (1999, 2001)
• All in all, the disaggregation broadly indicates, in line with
the theory, that regions in which the potential informational
spillover is larger are associated to a faster decreasing
pattern of yardstick competition in time.
17
Municipal elections
Obs.
% Electoral Obs.
1995 4667
69.7
1996
246
3.7
1997 1243
18.6
1998
535
8.0
1999 4308
64.3
2000
315
4.7
2001 1062
15.9
2002
680
10.2
2003
300
4.5
2004 4054
60.5
Source: Italian Ministry of the Interiors
•More electoral years: 1995,1999 and 2004
•Also 1997 and 2001, but less
18
The ‘more dense’ Regions
Piemonte
Lombardia
Liguria
19
The ‘more linked’ Regions
Emilia Romagna
Basilicata
Molise
20
Some exceptions
Veneto
• Veneto and Lazio: the
higher interaction
• The spatial coefficients
do not converge to
lower absolute values
Lazio
– Lazio takes a median
position with respect the
density of municipalities and
the average number of
neighbors per municipality
– The presence of Rome in
Lazio, that Italy's largest
Municipality and the national
capital, surrounded by much
smaller municipalities, may
explain this rather odd
result
21
Concluding remarks
• Interactions in local tax setting do not remain stable over
time
– The common assumption in the literature of an informational spillover
expiring after each election is not plausible at all need for a
development of the theory
• Convergence of the spatial correlation coefficients towards
the lowest levels of interaction
– At the national level the spatial lag is always positive and the spatial
error is always negative, but their absolute values decrease in the ‘first
order’ electoral years 1999 and 2001 when more than 60% of the
municipalities held elections
• YC is most effective in the regions where the density of
municipalities is higher, that is where the potential to
create intense informational spillovers is greatest
22
By density
1.40
1.20
1.00
PIE
LIG
0.80
LOM
0.60
TOS
0.40
BAS
PUG
0.20
0.00
2004
2003
2002
2001
2000
1999
1998
23
By link
1.40
1.20
1.00
0.80
0.60
0.40
0.20
0.00
2004
2003
MOL
2002
BAS
2001
EMI
2000
LIG
CAL
1999
1998
PUG
24