Escolar Documentos
Profissional Documentos
Cultura Documentos
Abstract
Direct democracy is spreading across the world, but little is known about
its effects on policy. I provide evidence from a unique scenario. In Spain,
national law determines that municipalities follow either direct or representative
democracy, depending on their population. Regression discontinuity estimates
indicate that direct democracy leads to a smaller government, reducing public
spending and revenues by 4%. These findings can be explained by a model in
which direct democracy allows voters to enforce lower special-interest spending.
Consistent with the model, the reduction in spending is entirely driven by
lower current expenditures, while capital (infrastructure) expenditures are not
affected.
A previous version of this paper covered a more limited sample period (20022011). I am
thankful to Manuel Bag
ues for sharing the budget data for the entire period in which the 100inhabitant threshold separated the two systems (19872011).
the data, and Samuel Suskind for outstanding research assistance. I also thank Jorge Alvarez,
Marco Battaglini, Angus Deaton, Bo Honore, Federico Huneeus, Matas Iaryczower, John Klopfer,
Ilyana Kuziemko, John Londregan, Fernanda M
arquez-Padilla, Eduardo Morales, Thomas Romer,
Tom Vogl, and seminar participants in Princeton University for their thoughtful comments and
suggestions. Financial support from the Research Program in Political Economy is gratefully acknowledged. I am solely responsible for any remaining errors.
Introduction
Besley and Case (2003) argue that there is a widespread belief that agency problems lead to
a government that is too large under representative democracy.
2
Direct democracy is a form of democracy in which people decide policy initiatives directly,
as opposed to representative democracy, in which people vote for representatives who then decide policy initiatives. The main forms of direct democracy are participatory or town meetings,
referendums, and initiatives.
3
From 2001 to 2015, 1,645 state-wide ballot measures have been voted in the United States.
Source: www.ncsl.org. Also in the US, direct democracy has been one of the battlegrounds of the
Occupy Wall Street movement.
sorting is concentrated are dropped. The results from these tests provide assurance
of the validity of the empirical strategy. Finally, a pure regression discontinuity
design with no fixed effects yields results that are remarkably similar to the ones
from the main specification with fixed effects, which provides further credibility to
the estimates.
The main finding is that direct democracy lowers expenditures and revenues by
4%. Budget deficits are therefore equal under both systems. I explain the results
with a model, based on Besley and Coate (2008), that I develop to formalize the
common argument that direct democracy curbs special interest spending and aligns
policy with the preferences of the median voter. In the model, a majority of the
population (citizens) prefer low special-interest spending, but they care more for
some other dimension of policy. There is, however, a minority (special interests)
that want high spending and whose vote is determined by special-interest spending.
Under representative democracy, the two issues get bundled, as individuals have only
one vote (to elect a representative who will implement policy). The consequence is
that, even though the majority prefer low spending, parties propose high spending in
equilibrium to attract the votes of the special interests. As a result, special-interest
spending is high in representative democracy (independent of which candidate wins
the election). In contrast, in direct democracy, the two dimensions are unbundled, as
individuals vote separately on issues. In this case, the preferences of the median voter
prevail along the two dimensions, and this results in low special-interest spending.
I present five additional sets of results to shed light on the mechanisms. First,
direct democracy reduces current spending but does not affect capital (infrastructure) spending. This is consistent with the prediction that direct democracy should
reduce special-interest spending.5 Second, the effects are not driven by changes in
the party that is in office. This too is consistent with the model, which predicts
that the two main parties should converge towards high spending in representative
democracy. Third, expenditures and revenues are equally volatile in direct and representative democracy. This suggests that the costs of participation in meetings do
not drive the resultsif that were the case, policy should be more volatile in direct
democracy (Osborne, Rosenthal, and Turner (2000)). Fourth, against an hypothesis of gradual learning, the effects do not change over time: they appear in the
5
As previous literature has noted, current expenditures can be used to capture the extent of
special-interest policies (Besley, Persson, and Sturm (2010)). There is even more reason to believe
that this is the case in the context of small Spanish municipalities (see Section 6).
first term a municipality works under direct democracy, and remain similar as municipalities spend more time under direct democracy. Fifth, direct democracy does
not affect subsequent elections behavior, suggesting that direct democracy does not
dramatically change individuals political views (e.g., being exposed to participatory
meetings does not lead to more voter turnout in subsequent elections).
Even though these results are consistent with the mechanism highlighted by the
model, I consider other potential explanations: elite capture in direct democracy,
differential costs of participation in meetings and elections, and direct democracy
directly affecting what policy individuals prefer. I discuss them in light of the
empirical evidence and conclude that, although it is possible that some of these
other mechanisms are also at work, the weight of the evidence indicates that they
do not drive the results.
This paper contributes to the growing literature on the impact of political institutions on public policy (see, for example, Besley and Case (2003) and Persson
and Tabellini (2005)). More specifically, it expands our knowledge about the policy
implications of participatory institutions. As mentioned, these institutions are increasingly used in villages across the world, and have been subject of recent work,
both theoretical and empirical. At the theory level, several papers have modeled
decision-making in meetings (Aragon`es and Sanchez-Pages (2009), Osborne, Rosenthal, and Turner (2000), Turner and Weninger (2005)). At the empirical level, the
literature has worked on estimating the impact of participatory institutions on several outcomes: for example, Besley, Pande, and Rao (2005) find that village meetings
in South India improved the targeting of welfare programs for the poor; Hinnerich
and Pettersson-Lidbom (2014) find that town meetings reduced welfare spending
in the context of Sweden in the 1920s and 30s; and Olken (2010), in a randomized
controlled trial, finds that the use of plebiscites improved voter satisfaction with
public policy in Indonesian villages. The present paper adds to that literature by
providing the first quasi-experimental evidence on how town meetings affect total
public spending and taxation decisions.
This paper also contributes to the literature, pioneered by De Haan and Sturm
(1997) and Roubini and Sachs (1989), that analyzes how public budgets are shaped
by political institutions, and, in particular, by direct-democracy institutions. Evidence from the US and Switzerland suggests that initiative and referendums reduce
public spending (see Matsusaka (1995) for the US, and Feld and Matsusaka (2003)
and Funk and Gathmann (2011) for Switzerland). Berry (2014), however, has re5
cently questioned the causal interpretation of the findings for the US, and Asatryan,
Baskaran, Grigoriadis, and Heinemann (2013) have found that referendums at the
local level increased expenditures in Germany. This paper contributes to that literature by providing evidence from an unexplored setting, in which, unlike what
happens in the aforementioned countries, the use of direct democracy is determined
by population size, therefore reducing endogeneity concerns.
Finally, this paper expands the work that uses regression discontinuity designs
to estimate the effects of institutional rules on policy (for example, Bordignon, Nannicini, and Tabellini (2013), Ferraz and Finan (2009), Fujiwara (2015), Gagliarducci,
Nannicini, and Naticchioni (2011), Gagliarducci and Nannicini (2013), Hinnerich
and Pettersson-Lidbom (2014), and Pettersson-Lidbom (2012)).
The rest of the paper is organized as follows. Section 2 presents the theoretical
framework. Section 3 provides background on the Spanish government systems.
The empirical strategy and the data are outlined in Sections 4 and 5. Section 6
presents the results. The robustness of the results is the focus of Section 7. Section
8 explores other potential mechanisms. Section 9 concludes the paper.
Theoretical Framework
Social scientists have long been interested in how closely policy reflects the preferences of the median voter under representative systems and in how switching
to systems of direct democracy might change policy. One strand of the literature
argues that, in representative democracy, an agency problem exists between voters and their elected representatives due to free-rider problems in monitoring and
disciplining officeholders (Matsusaka (2005)). This gives elected officials leeway to
pursue costly policies that are not in the interests of their constituents, resulting in
overspending. Frey (1994) argues that a model that pictures politicians as forming
a coalition against taxpayers and voters seems to be an apt illustration of representative democracy. In this view, the elected representatives are a well-defined group
that jointly reaps rents or cartelizes against the interests of citizens. Besley and Case
(2003), in summarizing this literature, conclude that there is a widespread belief
that agency problems lead to a government that is too large under representative
democracy. The use of direct-democracy institutions is a way to address this issue.
By allowing people to vote directly on policy, direct democracy gives them a way
to circumvent representative institutions that may have been captured by elites or
6
other special interests and aligns policy with the preferences of the median voter
(Matsusaka (2005)).
Another strand of the literature emphasizes that, under representative democracy, citizens have only one vote to cast for candidates who have responsibility for
choosing a bundle of issues (Besley and Coate (2008)). This may lead to policies that
are incongruent with the position favored by the median voter in some of the dimensions. By unbundling issues, direct democracy allows the median voters preferences
to prevail along different dimensions.
I formalize these ideas in a simple model, adapted from Besley and Coate (2008).
In the model, policy is two-dimensional. There is a main ideological issue (henceforth, ideology) and a secondary issue (special-interest spending). Although the
majority of individuals prefer low (possibly zero) special-interest spending, there
exists a minority that prefer high spending, perhaps because they will benefit directly from it. The majority care more about ideology but the minority care more
about spending. Under representative democracy, policy is implemented by a representative elected in an election between two candidates, proposed by two political
parties. Assuming the minority is sufficiently large, in the unique equilibrium, both
parties propose high special-interest spending, against the wishes of the majority. In
contrast, under direct democracy, individuals vote separately on the ideological issue
and on special-interest spending, and the position favored by the median voter in
each dimension is chosen. Direct democracy, therefore, results in low special-interest
spending, unlike what happens under representative democracy.
The following example illustrates how the mechanism operates. In the context
of small Spanish municipalities, we can think of special-interest spending as expenditures on local festivities.6 A majority of the population prefer low spending on
festivities, but it is not a very important issue for them. There is, however, a group
of people who are very keen on the fiestas and want high spending on them. Under
representative democracy, political parties need to cater to fiesta-lovers to attract
their vote. Although spending on festivities is not a highly relevant issue in regard
to how the majority vote, it is crucial for the fiesta-lovers. As a result, spending
6
I use festivities merely as an illustration. However, the anecdotal evidence that I gathered by
interviewing mayors and administrative staff actually points to fiestas, which play a large role in
social life in Spain, as one of the most likely sources of expenditures driven by special groups. This
is also consistent with the empirical findings on the types of expenditures. Specifically, spending
on local festivities is accounted for in Chapter 2 of the expenditures budget, which is precisely the
chapter that I find that direct democracy reduces.
Model
There is a continuum of individuals who have to decide on two dimensions of policy.
The first one is a primary policy dimension, which I will call ideological stance,
or, simply, ideology, i. Ideology can be left wing, denoted by i, or right wing,
i, so i {i, i}. We can think of ideology as a salient policy issue, such as the
progressiveness of policy.7 The second issue is targeted or special-interest spending,
denoted by s. Spending also can take two values, s {s, s}, where s denotes low
(maybe zero) spending and s denotes high spending.
Individuals preferences differ over the two issues. With respect to ideology, individuals are divided into leftists and rightists, v {l, r}. A fraction l of individuals
are leftists and prefer i. The rest, r = 1 l , are rightists and prefer i. Let i (v)
denote the optimal ideological policy from the perspective of an individual of type
v (hence, i (l) = i and i (r) = i). Without loss of generality, I assume that leftists
are the majority, thus l > .5. With respect to special-interest spending, individuals
are divided (independently of their preferences in regard to ideology) into citizens
and special interests, w {c, x}. A fraction c of individuals are citizens, c, and
prefer low spending, s. The remaining individuals, x = 1 c , are special interests and prefer high spending, s. The special interests are those who benefit from
special-interest spending, so they do not necessarily constitute an economic elite
they could be, for example, public employees who are favored by the government.
Special interests are a minority, thus x < r . Let s (w) denote the optimal ideological policy from the perspective of an individual of type w (hence, s (c) = s and
s (x) = s).
7
For example, municipalities can use tax deductions to benefit poor individuals. Even if it
may seem surprising that a main ideological issue determines the vote at the local level, the
correlation between votes to the main right (left)-wing party in local and national elections is .63
(.57), suggesting that the determinants of voting are closely related at the two levels.
(1)
where uv,w denotes the utility of type (v, w). Citizens are more concerned about
ideology, so that blc > lc and brc > rc . The special interests, in contrast, care
more about special-interest spending, as they benefit directly from it: blx < lx and
brx < rx .
The insights and conclusions are the same with both approaches.
individual in the party, which is a special interest. The same happens if Party A
switches to a rightist special interestin this case, the probability of winning goes
down to 1/2, as both parties propose the exact same types of candidates. Finally, a
radical switch to a rightist citizen increases the probability of winning (to l ) but
at the cost of sacrificing the preferred policy in both dimensions, so the utility in
the case of winning is zero. The reasoning for Party B is analogous. Finally, note
that both parties running with a citizen of their preferred ideological type cannot be
an equilibrium, as both parties would want to deviate to a special interest of their
preferred ideological type.
Case 2: For each P, sP = s.
In this case, citizens are a majority in political parties, as they are in the entire
population. In this case as well, an equilibrium exists in which both parties run
with a special-interest candidate of their preferred ideological position, under certain
conditions that are indicated in the following proposition.
Proposition 2. (i) If, for each P , sP = s, and, for each v,
c
bvc
> ,
vc
x
then (vA = l, wA = x) and (vB = r, wB = x) is an equilibrium.
(ii) If, additionally,
blc
x + c l
>
lc
x + c l l
or
brc
x + c r
>
,
rc
x + c r r
11
special interests, given the preferences of citizens). Assumption (ii) rules out that
both parties choosing a citizen is an equilibrium.
Case 3: sP = s for some P, sP0 = s for P 6= P0.
This case is a combination of the other two: Citizens are the majority in one party
but not in the other.
Proposition 3. If sl = s, sr = s, and blc /lc > c /x , then (vA = l, wA = x) and
(vB = r, wB = x) is the unique equilibrium. If sl = s, sr = s, and brc /rc > c /x ,
then (vA = l, wA = x) and (vB = r, wB = x) is the unique equilibrium.
Proof. See Appendix A.
Following the logic of the first two cases, now the requirement that special interests are sufficiently large is required only for the party in which citizens are a
majority.
12
interests may feel more intensely about the issue. Hence, a measure of welfarefor
example, utilitariancould rise or fall.
Institutional Background
In this section, I provide some institutional background and describe the two government systems that are used by Spanish municipalities. Spain is a very decentralized country. It is politically divided into 17 regions, 50 provinces, and more
than 8,000 municipalities, which are administered by local governments. Local governments spend 13% of the overall spending of the country (Gil-Ruiz Gil-Esparza
and Iglesias Quintana (2007)) and have substantial autonomy. National law (Local
Government Regulatory Law, Ley de Bases del Regimen Local ) requires them to
provide a variety of services, including public lighting, waste collection, street cleaning, road paving, household water supply, sewerage, access to villages, and food and
beverage control.9 In addition, they usually provide other services useful to the municipality, such as organizing local festivities or providing tourist information. Local
governments can decide how much to spend on the required goods and services and
whether to provide additional goods and services. There is considerable variation
in expenditures across municipalities: The 9010 ratio is 4.29.10
Local governments obtain revenues both from their own sources, which account
for an average of 52% of total revenues for municipalities close to the threshold, and
from transfers from upper-level governments (48%). Municipalities can set the tax
rateswithin certain limits imposed by national lawand the prices and user fees
of the goods and services that they provide.11 As happens with expenditures, there
9
Larger municipalities are required to provide some additional services. For example, municipalities with more than 5,000 inhabitants must provide a public park and a public library. This
paper, however, focuses on municipalities near the 100-inhabitant threshold, so rules for larger
municipalities are not relevant for the analysis.
10
I calculate this number by taking the average value of expenditures by municipality over time
for those municipalities that were close to the threshold in at least one term. I define being close
to the threshold as lying within the optimal bandwidth in the regression for public expenditures
(54 inhabitants), calculated following Imbens and Kalyanaraman (2012)s procedure. This way, I
do not use the cross-time variation, which would overstate the variation in expenditures.
11
According to national law, municipalities can impose five different taxes: property tax (impuesto de bienes inmuebles), a tax on economic activities (impuesto de actividades econ
omicas), a
motor vehicle tax (impuesto sobre los vehculos de tracci
on mec
anica), a tax on improvements to
real property (impuesto de instalaciones, construcciones y obras), and a tax on the increased value
of urban land (impuesto sobre el incremento de valor de los terrenos de naturaleza urbana). The
first three are required taxes that national law forces municipalities to levy while the last two are
optional taxes. For both required and optional taxes, municipalities are free to set the tax rates
13
14
15
Municipal) imposed the use of direct democracy to all municipalities with fewer than
500 inhabitants. However, this provision was never enforced. During the Second
Spanish Republic, another attempt was made to extend the use of direct democracy
to all municipalities with fewer than 500 inhabitants (1935 Law), but the attempt
never materialized, as the regime lasted only one more year before the onset of the
Spanish Civil War in 1936. The situation did not change during Francos regime.
A 1955 law (Ley de Regimen Local) required the following of the direct-democracy
system only for those municipalities in which that was the traditional form of government. It was not until after the restoration of democracy after Francos death
in 1975 that the rules changed substantially. In 1978, national law (Ley de Elecciones Locales) required all municipalities whose population in the election-year was
smaller than 25 inhabitants to follow the direct democracy-system. In 1985, a reform
extended the requirement to all municipalities with fewer than 100 inhabitants.19
This system was in force until 2011. Since then, no municipality has been required
to follow direct democracy, although municipalities can adopt the direct-democracy
system voluntarily by following the procedure described in the national law. In sum,
a population threshold at 100 inhabitants separated municipalities under direct and
representative democracy from 1987 to 2011. In this paper, I study public finances
in that period.
Data
Data for the public budgets and population size of municipalities are from the Spanish Ministry of Finance (Ministerio de Hacienda) and are publicly available for
82.7% of annual budgets.20 During the sample period, local elections were held
every four years from 1987 to 2011, so the data set covers six terms. Because local
elections are in May or June, the (possible) change in government system happens
19
Municipalities with 100 or more inhabitants could follow a demanding procedure to adopt
the direct-democracy system. Specifically, a majority of the citizens of the municipality had to
sign a petition, and two-thirds of the members of the council and the regional government had to
approve. To the best of my knowledge, no municipality ever used this procedure. This implies that
the regression discontinuity component of the estimation is sharp, as the probability of treatment
jumps from 0 to 1 at the threshold.
20
Municipalities were required to report their budgets to the Ministry of Finance, which makes
them public. However, data from some budgets are missing because either they did not comply
with their obligations or because their data have not been digitized. A placebo test shows that
direct democracy does not have an effect on a dummy variable that indicates whether or not the
observation is available (see Appendix C).
16
halfway through the year. As spending and revenues are determined by the budget
approved at the end of the previous year, I assign election-year observations to the
previous term.21 Figure 2 provides a graphical representation of this information.22
Table 2 shows the summary statistics. Panels A shows the statistics for all
the municipalities and Panel B, for municipalities close to the threshold, to allow
the comparison between the municipalities used for identification with the rest of
municipalities in the country.23
The first three variables are the main outcomes: real expenditures per capita
(Expenditures), real revenues per capita (Revenues), and budget deficit percentage
as a fraction of revenues (Deficit).24 The average municipality close to the threshold
spends more than the average municipality in the country (e1592 per capita per year
vs. e1095). This is because expenditures per capita decrease with population, probably due to economies of scale. Revenues include transfers to the municipalities from
upper-level governments and the municipalities own income. The deficit is defined
as the difference between expenditures and revenues, as a fraction of revenues, and
can be financed by issuing debt or by reducing cash balances. The average deficit is
5.5% for all municipalities and 3.5% for those close to the threshold. Figure 3 shows
the histograms for expenditures and revenues.
The following seven variables provide details on the composition of revenues and
expenditures. Revenues are the share of transfers from upper-level governments over
21
In Spain, the fiscal year coincides with the calendar year, and budgets for year y are approved
in November or December of year y 1.
22
If expenditures and revenues are sticky, expenditures and revenues for municipality m at year y
are partially determined by the government system that municipality m followed in years before y.
When a municipality m switches systems, this implies that some of the expenditures and revenues
for year y may be determined by a government system that does not correspond to the government
system municipality m is following at that year. In general, this will make the estimated coefficients
closer to zero than in a scenario with no changes in government system. These considerations,
however, are not likely to play a significant role in practice. First, most expenditures and revenues
in small municipalities are decided on a year-to-year basis. For example, it is rare that they
incur in multi-year capital expenditures, unlike what happens in larger municipalities. It is true,
nonetheless, that other expenditures, such as personnel, could be more sticky. Second, if sticky
policies were playing a large role in the estimation, we should expect to see that the effects of direct
democracy on policy increase (in absolute value) over time, as municipalities spend more time in a
given system. However, I find that the effects of direct democracy on policy are already produced
in the first year that a municipality switches into direct democracy and do not vary significantly
over time (see Appendix C).
23
I define close to the threshold as lying within the optimal bandwidth in the main regression
for public expenditures (54 inhabitants), calculated following Imbens and Kalyanaraman (2012)s
procedure.
24
To obtain the variables in real terms, I divide the nominal variables by the GDP deflator. Data
for the deflator are from the Spanish National Institute for Statistics.
17
total revenues (Share Transfers) and the share of the municipalities own income
that is derived from taxes (Share Taxes), fees (Share Fees), and rents (Share Rents).
Transfers account for an average of 48% of their own revenues for municipalities close
to the threshold. With respect to their own revenues, municipalities can impose five
different taxes, the most important of which is a property tax.25 Taxes account
for 40% of own revenues for municipalities close to the threshold. Fees (tasas,
precios p
ublicos and contribuciones especiales) account for 28% of own revenues.
Typical fees include those for garbage collection and a water supply. Rents and other
property income, including the disposal of assets, account for the remaining 32%.
From the expenditures side, the main variable is the share of current expenditures
(Share Current Exp), which accounts for an average of 52% of total expenditures.
Capital expenditures (mainly infrastructure investment) account for the remaining
48%. Within current expenditures, the main categories are personnel spending
(Share Personnel ) and expenditures on non-durable goods and services (Share Goods
and Services). They account for an average of 16% and 35%, respectively.26
I use the next variables to study whether the results are driven by different
parties being in office in direct and representative democracy. RW Mayor and
LW Mayor are dummy variables that indicate whether the mayor is from the main
right-wing party (Popular Party, PP) or the main left-wing party (Socialist Party,
PSOE). The next two variables are used in a different regression-discontinuity design to estimate the effect of political parties on policy at the local level. With the
sample restricted to those elections in which the two most-voted parties were the PP
and the PSOE, they are the percent-point difference between the PP and the PSOE
RW Difference, and a dummy variables that indicates whether the PP was the most
voted (so that a positive (negative) value of RW Difference implies RW Winner = 1
(RW Winner = 0)). The summary statistics from these variables suggest that municipalities close to the threshold are more right-wing than the average; for example,
the mayor is from the PP in 53% of observations close to the threshold, compared
with an average of 36%.
The next six variables in Table 2 are from national Congress electoral results; I
use them to test for covariate balance around the threshold. They are provided by
the National Statistics Institute (Instituto Nacional de Estadstica). The variables
considered are the shares of votes for the three main parties in Spainthe right-wing
25
26
See footnote 11 for a more detailed description of the taxes they can introduce.
The remaining current expenditures are interest payments, which are negligible (0.7%).
18
PP (Votes Right), the left-wing PSOE (Votes Left), and the far-left-wing United
Left (Votes Far Left), the difference in the share of votes for the two most-voted
parties (Votes Difference), the percentage of votes for the most-voted party (Votes
Winner ), and voter turnout (Turnout).27 These summary statistics from national
elections also show that municipalities close to the threshold are more right-wing:
The PP obtains an average of 53.8% of the vote in them, while it obtains 41.9% in
the average municipality in the country.
The final six variables are demographic variables, which I employ to study covariate smoothness around the threshold. They are the average age in the municipality
(Mean Age); the share of young (Young), middle-aged (Middle-Aged ), and old (Old )
individuals; the share of immigrants (Immigrants); and the share of EU immigrants
over total immigrants (EU Immigrants).28 The average municipality close to the
threshold is older than the average municipality in the country; for example, the
share of old people is 38.5% versus 27.7% in the whole country. It also has fewer
immigrants: 2.5% versus 4.1%.
Empirical Strategy
19
Conversations with mayors and other local government officials make me think that this is
probably the most relevant reason.
32
Of course, if the mayor is a citizen instead of a special interest, he or she will prefer direct
democracy. This raises the theoretical possibility of two-way sorting. Although it is not possible
to directly test for the existence of two-way sorting, I believe that sorting into direct democracy
is negligible. First, as mentioned, mayors and local government officials indicate that the main
reason for sorting is the first, i.e., that representative democracy is easier to operate. Second, it is
hard to conceive local government officials trying to persuade people to go to register in another
municipality. Also, note that the empirical approach and the robustness checks to assess the validity
of the strategy, in particular the donut regressions, do not depend on the sorting being one-way.
20
Outcomemyt
= m + y + DirDemmt
+f (Populationmt 100)
+ log(Transfersmyt ) + umyt ,
(2)
21
For example, suppose that the whole sorting process is driven by local government
officials who are very good at persuading individuals to register. A violation of the
identification assumption would not be that local government officials characteristics are correlated with other factors that affect policy but, rather, that the precise
period in which the sorting is achieved coincides with a change in those other factors.
Although the identification assumption is, of course, not directly testable, I
perform five sets of tests to assess the validity of the empirical approach, following
previous research that has dealt with similar situations (Pettersson-Lidbom (2012),
Barreca, Guldi, Lindo, and Waddell (2011)). First, I check that municipalities at
each side of the threshold are similar in other time-varying variables that may have
an effect on the outcomes of interest, conditional on the fixed effects. Second, I
consider donut regressions, which address specifically the issue of sorting. In this
test, observations very close to the threshold, where self-selection is concentrated, are
dropped from the analysis. Third, I test whether the effect is similar in municipalities
that switch into direct democracy and in those that switch out of direct democracy.
Finding that the effect is similar would reinforce the credibility of the estimates, as
it would be difficult to explain that correlation from some omitted factor. Fourth,
I estimate a dynamic model to test for pretrends in the outcomes of interest. Fifth,
I conduct placebo tests at other (artificial) population thresholds. The results from
all these tests provide assurance of the validity of the empirical strategy.
I also will show the results for Equation (2) without the fixed effects. Although,
in principle, the sorting could be more problematic without the fixed effects, most
of the robustness checks indicate that it is also a valid approach. An exception is
that covariate balance around the threshold is not as good without the fixed effects
as it is with them (see Table 8). Specifically, without the fixed effects, there is an
imbalance in the transfers received from upper-level governments. The control for
transfers, however, removes that imbalance. The two different empirical approaches
produce remarkably similar estimates, lending further credibility to the estimates.
I use nonparametric local linear regressions to estimate Equation (2), as suggested by Hahn and der Klaauw (2001), Porter (2003) and Gelman and Imbens
(2014). A key ingredient to this approach is the bandwidth. A larger bandwidth
increases the efficiency of the estimation at the cost of more bias. I choose a baseline bandwidth according to the procedure suggested by Imbens and Kalyanaraman
(2012) and provide the results at different fractions of that bandwidth.36 I use a
36
The results are very similar if I use the optimal bandwidth suggested by Calonico, Cattaneo,
22
rectangular kernel, as recommended by Imbens and Lemieux (2008) and Lee and
Lemieux (2010). This is equivalent to estimating standard linear regressions over
the interval of the selected bandwidth on both sides of the cutoff point. I cluster standard errors at the municipality level (Bertrand, Duflo, and Mullainathan
(2004)). Because the running variable is discrete, I also cluster standard errors at
the running variable, as recommended by Lee and Card (2008). Therefore, I follow
a multi-clustering approach (Cameron, Gelbach, and Miller (2011)).37
Results
log(Transfersmyt ) + umyt using the observations within the optimal bandwidth for Outcomemyt ,
where Populationj ,mt is a dummy that indicates whether municipality m has population size j at
term t. Then, in the y-axis, I plot the estimated coefficients j . I normalize the coefficients so that
the average bin immediately to the right of the threshold takes the value of zero.
39
Because the outcome variable is the log of expenditures, the exact percent effect on expenditures
is 100 (exp(0.0377) 1) = 3.699 (Halvorsen and Palmquist (1980)).
23
go together with a change in revenues or whether, on the contrary, they are created
by different deficits. According to the preferred specification (Column 1), direct
democracy reduces revenues by 3.2%, and the effect is statistically significant at
the 5% level. The rest of the columns show that the results are robust to different
bandwidths and to the removal of the fixed effects. Finally, the results for deficits
(Panel C) are close to zero and not significant under any specification. These findings
imply that direct democracy reduces expenditures and revenues by a similar amount,
without affecting budget deficits.
Besley, Persson, and Sturm (2010) use current expenditures to measure the extent of specialinterest policies in the United States south.
41
One issue in regard to the estimation for capital expenditures is that 8% of the observations
24
tures is brought about by current expenditures and is consistent with the prediction
of the model that direct democracy curbs special-interest spending.
In Appendix B, I show results for two finer divisions of expenditures and revenues. First, I show that the reduction in current expenditures is driven by spending
on non-durable goods and services, while personnel expenditures are not affected.
This rules out the possibility that the difference in spending is created by wages
paid to city councilors, who are not present in direct democracy.42 Second, I show
that direct democracy lowers revenues in the three main categories: taxes, fees,
and rents. This suggests that the reduction in revenues is across-the-board and not
driven by a specific item, which could be used to benefit some particular group or
class of individuals.
25
the left-wing PSOE were the two most-voted parties. I perform a fuzzy regression
discontinuity design, in which the running variable is the difference in the percentage
of votes for the PP and the PSOE (RW Difference) and the treatment variable is
having a PP mayor (RW Mayor ).43 Thus, when the running variable is positive
(negative), the most-voted party (RW Winner ) is the PP (PSOE). More formally,
I estimate:
(3)
where the coefficient of interest is . As with the main results, I will show results
with and without municipality and year fixed effects. I estimate Equation (3) by
local linear regression, and I cluster standard errors at the municipality level. Tests
for covariate balance and manipulation of the running variable RW Difference, shown
in Appendix C, provide assurance that this is a valid empirical approach.
The first stage is strong. As shown in Panel A of Table 5, being the most-voted
party increases the probability of obtaining the mayor by 55 percentage points.
Panels B to E in Table 5 show the estimated coefficients of interest
, and Figure A1
in the Appendix provides a graphical representation of the first-stage and reducedform results. There is no evidence of differences in policy between PP and PSOE
mayors, as the effects are close to zero and mostly insignificant.44 These results
suggest that the increased spending in representative democracy comes from both
parties, as predicted by the model.
One caveat to this approach is that, to obtain precise estimates, I need to use all municipalities
under representative democracy, not only those close to the threshold.
44
These results replicate the findings by Arenas and Bag
ues (2015) for Spain and are in line with
the findings of Ferreira and Gyourko (2009) for United States cities.
45
Eggers, Freier, Grembi, and Nannicini (2015) follows a similar procedure.
26
(3) letting the outcome variable be the absolute value of the residuals obtained in
the previous step. The coefficient for DirDem from this second stage captures the
effect of direct democracy on the variability of the outcome within municipalities
over time. If the fixed effects are dropped from both stages, then it will capture the
effect on both the within- and across- municipality variability.
The results, shown in Table 6, indicate that there is no effect of direct democracy
on the variability of expenditures or revenues. Panel A shows the results for expenditures, and Panel B for revenues. No effect is significant, and the point estimates
are close to zero: in the main specification, direct democracy leads to expenditures
that deviate from the predicted values just two percentage points less than in representative democracy. The results are also insignificant if, instead of the absolute
value of the residuals, I consider the squared residuals, the log of the absolute value
of the residuals, or the log of the squared residuals. Finally, Panel C shows the
results for deficits. There is evidence that direct democracy leads to less variability
in deficits, but only so in the specifications with the fixed effects. Without them,
there is no significant difference between direct and representative democracy.
4
X
j=2
+ log(Transfersmyt ) + umyt ,
(4)
where Nj,mt is a dummy variable that indicates how many terms municipality m
at year y has been under direct democracy. For a municipality in its first term,
N1,mt = 1 and Nj,mt = 0 for j 6= 1 and, analogously, for a municipality in its second
or third terms. For a municipality in its fourth or longer term, N4,mt = 1 and
46
Additionally, this test addresses the question of whether sticky policies may yield lower estimates (in absolute value); see footnote 22.
27
Nj,mt = 0 for all j 6= 4.47 Therefore, captures the effect of direct democracy in
the first term under direct democracy, and the j indicate how the effect varies in
subsequent terms. An analysis based on years instead of terms spent under direct
democracy yields similar results but are more imprecisely estimated.48
The results are shown in Table 7. For expenditures, the coefficient on DirDemmt
under the optimal bandwidth is -3.4% and significant at the 5% level, indicating
that direct democracy reduces expenditures in the first term. For N2,mt DirDemmt ,
N3,mt DirDemmt and N4,mt DirDemmt , the coefficients are not significant, indicating
that the effect does not vary over time. A similar pattern appears in the estimation
for revenues. For deficits, neither the coefficient on DirDemmt nor the interaction
terms are significant.
28
Robustness
In this section, I present five robustness checks to assess the validity of the empirical
strategy: placebo tests to study covariate balance at the threshold, donut regressions to address specifically the issue of sorting, separate analysis for municipalities
that switch into and out of direct democracy, dynamic specifications to test for pretrends in the outcomes of interest, and placebo tests at other population thresholds.
Appendix C shows additional robustness checks to study whether the results change
when top-coding outliers and whether the number of missing observation changes
at the threshold. Appendix C also includes robustness checks for Equation (3).
7.1
Note that, due to the fixed effects, we need outcomes that vary over time.
Spain is a parliamentary democracy, so there are no direct elections for the executive branch.
Citizens elect the Congress, which, in turn, elects the Prime Minister.
50
29
of votes for the two most-voted parties and the share of votes for the most-voted
party (Votes Differencem,t1 , and Votes Winnerm,t1 ), which give a measure of how
politically polarized the municipality is, and voter turnout (Turnoutm,t1 ), which
has been shown to correlate with variables that also may affect peoples preferences
toward economic policy, such as age, education, or income (Blais and Achen (2010),
Matsusaka and Palda (1999)).51 Finally, I study covariate smoothness in demographic variables: average age in the municipality (Mean Age); the share of young
(Young), middle-aged (Middle-Aged ), and old (Old ) individuals; the share of immigrants (Immigrants); and the share of EU immigrants over total immigrants (EU
Immigrants). These variables will capture how economic conditions are evolving, as
immigrants are more likely to locate in booming municipalities.
The results from these tests are shown Table 8 and represented graphically in
Figure 7. With the preferred specification, which includes fixed effects, everything
is smooth at the threshold. Only the coefficient for Votes Left is significant, and
significance is only at the 10% level and disappears when the bandwidth is cut.
Without fixed effects, there is an imbalance in transfers, which I control for in
the estimation, as explained in Section 5. In addition, the imbalance in Votes Left
becomes larger, but the results are also robust to including it as a control in Equation
(2).
7.2
Donut Regressions
This section considers donut regressions in the spirit of Barreca, Guldi, Lindo, and
Waddell (2011). These regressions exclude from the estimation observations that
are very close to the threshold, where most of the sorting is concentrated. Given
the nature of the sorting process, it is likely that most of the self-selection of municipalities into representative democracy is concentrated in a small window around
the threshold. It does not seem plausible that the population size of municipalities
that are self-selecting is much larger than what is strictly necessary to be above
the threshold or that municipalities with a population size far below the threshold
attempt to cross it.52 Thus, finding a similar effect when municipalities very close
51
The rest of the parties with representation in Congress during the entire period are regional
parties that do not run in the whole country.
52
For example, a municipality that would have naturally had 97 inhabitants may well end
up having 105 inhabitants if enough individuals are willing to register so that their municipality
is above the threshold. However, it is unlikely that this municipality would end up having 120
inhabitants. Likewise, it is unlikely that a municipality with a population far below the threshold
30
to the threshold are excluded would reinforce the credibility of the estimates.
For every outcome variable, I consider four regressions. The first one is the
benchmark regression.
around the threshold (that is, municipalities with 97 to 103 inhabitants), the third
excludes those within an interval of 5%, and the fourth excludes those within an
interval of 10%. Table 9 shows the results. For public expenditures, the coefficient
for expenditures is -3.7% in the baseline specification, and -2.4%, -3.8%, and -4.9%
in the three donut regressions, excluding 3, 5, and 10% of observations around the
threshold, respectively. The results follow a similar pattern in the regressions for
revenues and remain insignificant for deficits.
7.3
31
7.4
If, conditional on the fixed effects, the next periods government system is as good
as random for those municipalities sufficiently close to the threshold, then it should
not help predict this periods outcome variables.53 In this section, I study the timing
of the effect of the government system on the outcomes of interest by estimating the
following model:
Outcomemyt
= m + y + 1 DirDemmt + 2 DirDemm,t+1
+f1 (Populationmt 100)
+f2 (Populationm,t+1 100)
+ log(Transfersmyt ) + umyt .
(5)
The coefficient 2 serves as a robustness check: 2 6= 0 would suggest the existence of an endogeneity problem, which would raise concerns about the validity
of the identification strategy. In particular, 2 < 0 (2 > 0) would mean that
municipalities that switch out of representative into direct democracy are already
experiencing a decrease (increase) in the outcomes before the actual change in government system occurs. 1 captures the contemporaneous effect of the government
system on the outcome variables. Thus, if the model is correctly specified, the estimates for 1 should be similar to the baseline results. Alternatively, one can think of
Equation (5) as testing whether treatment and control municipalities were following
different trends before the realization of this periods variables.54
The results for this test are shown in Table 11. The estimates for 2 are close
to zero and not statistically significant across specifications, and the coefficients for
1 remain similar to the baseline results.
An alternative way to test for an effect on lagged variables, which allows a clear
graphical representation, is to repeat the exercise of Figure 5, but for lagged instead
of contemporaneous outcomes. Specifically, I lag the outcome variables four years,
so that they correspond to the outcomes in the previous term at the same year of
the electoral cycle. These tests, shown in Figures 8 and 9, show that there is no
53
According to Lee and Lemieux (2010), finding a discontinuity in ymt but not in ym,t1 would
be a strong piece of evidence supporting the validity of the regression discontinuity design.
54
It is necessary to flexibly control for population in both periods (i.e. to include the functions
f1 and f2 ) to capture the effect of the government system and not of changes in population.
32
7.5
In this section I conduct placebo tests by estimating the effect of crossing population
thresholds that are irrelevant (e.g., the effect of having more than 115 inhabitants).
Specifically, I create placebo treatments at all other population sizes from 30 to 250
inhabitants, by defining dummies that indicate if the population of a municipalityyear is above or below a given population size.56 I then run equation (2) with
every placebo treatment (so I run 220 regressions per outcome variable). If the
effect of direct democracy has an actual causal effect, then the estimate of direct
democracy on policy, based on the 100-inhabitant threshold, should be an outlier in
the distribution of placebo coefficients.
I show the results for these tests in Figure 10. I show the empirical cumulative
distribution function of point estimates and t-statistics for the 220 regressions considered for each variable. I also show the implied p-values, which are the share of
placebo regressions in which I obtain a point estimate (or t-statistic) that is larger in
absolute value than the one for the true threshold. For expenditures, the implied pvalue is 3.2% for point estimates, and 0% for t-statistics, meaning that the t-statistic
at the 100-inhabitant threshold is the largest of the 220 placebo regressions. For
revenues, the p-values are 5% and 2.3%, respectively.
These tests specifically address the concern that the bandwidths used in the main
specifications are inadequate and lead to artificial significant effects.57 However, if
that were the case, then we should see similarly large effects at other thresholds.
55
33
Mechanisms
Although the evidence presented above is consistent with a theory in which direct
democracy curtails representative-democracy overspending, other explanations may,
of course, be suggested. Here I examine three other potential mechanisms and
discuss their plausibility, given the political and social circumstances in Spain and
the available data.
(a) Direct democracy may be more prone to elite capture, as was recently proposed by Hinnerich and Pettersson-Lidbom (2014). An elite who prefers low spending may exert more influence under direct democracy than in a representative democracy setting. Hinnerich and Pettersson-Lidbom (2014) provide three arguments that
may show that this was a likely situation in the case studied (Sweden at the beginning of the 20th century). First, the lack of political parties in direct democracy
made it more difficult for the citizens to solve their collective action problems (e.g.,
Acemoglu and Robinson (2008)). Second, the chairman of the town meeting, often a member of the elite, had substantial power to set the agenda. Third, many
decisions at meetings were taken by an open vote, and there was the potential for
intimidation by the elite.
Although Hinnerich and Pettersson-Lidbom (2014) present compelling evidence
in favor of this mechanism in the case that they study, there are several reasons to
doubt that it is relevant in the setting explored in this paper. First, their argument
is based on a context with a conflict between the landed local elite (e.g., farmers)
and citizens (e.g., the agricultural workers) at the time when Sweden was still a
poor, mostly agrarian society. Thus, the landed elite would like to keep the old
labor-repressive economic system, while the citizens would like to have an economic
system based on wage labor. This type of situation is not at all present in contemporary Spain. Second, it is not obvious how the elite could capture the meetings in
the Spanish setting. The first two reasons that Hinnerich and Pettersson-Lidbom
(2014) provided for Sweden do not apply: There are political parties under direct
democracy, and the meetings are mostly deliberative, and, therefore, the agendasetting power of the mayor is limited. Finally, although it is the case that many
decisions are taken by an open vote and the possibility of intimidation cannot be
completely ruled out, conversations with mayors and other local officials make me
34
think that intimidation is not an issue in Spanish town meetings. Third, it is not
obvious why the elite would want a smaller local government. Gobernado Rebaque
(2003) shows that fiscal policy is only very slightly progressive at the local level;
even though some expenditures are progressive, this is almost totally offset by the
regresiveness of local taxes.58 Alternatively, if we think of the elite that would want
to capture the meetings as the special interests of the model, then the elite would
prefer a larger government, which is not consistent with the results.
(b) The differences in policy may be driven by costs of participation. An alternative model for direct democracy is developed by Osborne, Rosenthal, and Turner
(2000). In this model, individuals decide whether to participate, at a cost, in a
meeting whose outcome is a compromise among the participants favorite positions.
In deciding whether to participate, each person compares the cost of participation
with the impact of his or her presence on the compromise. One prediction of the
model is that the outcome will be extremely volatile: in the presence of even a
small amount of randomness, the equilibrium outcome may vary dramatically.59
Therefore, it is reasonable to predict that, if direct democracy behaves according
to Osborne, Rosenthal, and Turner (2000), policy will be more variable than in a
representative-democracy setting.60 The evidence, however, indicates that expenditures and revenues are not more volatile in direct than in representative democracy,
as was discussed in Section 6.
Alternatively, it could be the case that the median attendant is different from the
median voter at the representative-democracy elections, but that the same group
of people consistently attend the meetings, unlike what is predicted by Osborne,
Rosenthal, and Turner (2000). If individuals with a high cost of voting who do not
participate under direct democracy but do participate under representative democracy prefer a larger government, then we can expect a larger government under
representative democracy.61
58
A caveat is that Gobernado Rebaque (2003) studies only large municipalities. However, it is
likely that the overall system is even more regressive in smaller towns; while the regresiveness of
the taxes (mainly the property tax) applies to both, the progressive expenditures singled out by
Gobernado Rebaque (2003) (for example, public transportation) have a very limited role in small
municipalities.
59
One source of such randomness is the players behavior in a mixed-strategy equilibrium. Another source of randomness is exogenous, and arises in a model in which each player is prevented
from attending the meeting with some positive probability.
60
Of course, this depends on the mechanics of the representative-democracy system, but none of
the standard models would predict such big variability.
61
Thus, the median voter theorem may hold in both representative and direct democracy, and we
35
Although it is not possible to test this alternative explanation directly, there are
two reasons why it is a less plausible explanation for the findings. First, that individuals can delegate their vote reduces drastically the cost of participation in meetings.
In fact, once vote delegation is taken into account, the average of represented voters
is 40%. Second, as special interests care more intensely for spending, they should
have a lower net cost of attending the meetings and, therefore, be over-represented
in them. Therefore, this mechanism works against finding that direct democracy
reduces special-interest spending.62
(c) Direct democracy may affect the policy that individuals prefer. For example,
the deliberative nature of the meetings may allow citizens to learn by aggregating information, or it may make the drawbacks of some public expenditures more
salient.63
Although it cannot be completely ruled out that this mechanism is at play,
two tests that were presented in Section 6 suggest that it is not the driving force
of the results. First, if the effects of direct democracy on policy were driven by
gradual learning, they should increase over time as municipalities spend more time
under direct democracy. However, this is not the casethe evidence indicates that
the effects do not change significantly over time.64 Second, if being under direct
democracy changed what policies are favored by individuals, that may bring about
a change in how individuals behave in elections. However, the evidence presented in
Section 6 indicates that political behavior is not affected by being exposed to direct
may still observe different policies, as the set of voters and, therefore, the median voter (of those
who participate) are different in both systems.
62
Although it is difficult to predict how pure costs of participation correlate with spending preferences, it is reasonable to expect that special interests have lower net costs of participation. In a
Downsian framework, an individual decides whether to participate (attend the meeting or delegate
his her vote). If he or she participates, his or her utility is given by U = P B + D, where P is the
probability that he or she is pivotal, B is the benefit that he or she obtains if his or her preferred
policy is implemented compared to the alternative, and D is the consumption-value of the act of
voting, which includes the pure costs of participation (e.g., spending time in the meeting, or writing
the vote delegation). If he or she does not participate, his or her utility is normalized to zero. The
argument above states that, while it is not obvious how D correlates with spending preferences,
special interests should have higher B. Thus, it seems reasonable to assume a positive correlation
between participation and a preference for high special-interest spending.
63
Thus, we may see different policies even if the median voter theorem holds in both representative
and direct democracy and even if there are no previous differences in policy preferences between the
citizens who participate in direct and representative democracy. Under this scenario, differences in
policy arise because direct democracy brings about a shift in the median voters optimal point.
64
Data are yearly, so I cannot rule out the possibility that town meetings affect peoples preferences in less than a year. For example, attending merely one meeting might make individuals change
their minds about policy, thus driving the results. However, this is not a testable proposition.
36
democracy.
Conclusion
This paper has provided empirical evidence on the effects of direct democracy on
policy. Using a regression discontinuity design in the context of Spanish local governments, I have shown that direct democracy results in a smaller government.
Compared with a standard representative-democracy system, direct democracy reduces public spending and revenues by 4%. The reduction in public spending is
driven entirely by lowered current expenditures, which are more easily targetable,
while capital (infrastructure) expenditures are not affected.
This paper expands our knowledge of the consequences of adopting participatory
institutions. It also adds to our understanding of the political determinants of public
finances. An interesting avenue for future work is to study how other outcomesfor
example, corruptionare affected by the use of participatory meetings.
37
References
Acemoglu, D. and J. A. Robinson (2008). Persistence of power, elites, and institutions. The American Economic Review 98 (1), 267293.
Aragon`es, E. and S. S
anchez-Pages (2009). A theory of participatory democracy
based on the real case of Porto Alegre. European Economic Review 53 (1),
5672.
Arenas,
A. and M. Bag
ues (2015). Realmente importa el color del
Blais, A. and C. H. Achen (2010). Taking civic duty seriously: Political theory
and voter turnout. Unpublished manuscript, Princeton University.
Bordignon, M., T. Nannicini, and G. Tabellini (2013). Moderating political extremism: Single round vs runoff elections under plurality rule. Unpublished
manuscript, Bocconi University.
Calonico, S., M. D. Cattaneo, and R. Titiunik (2014). Robust nonparametric
confidence intervals for regression-discontinuity designs. Econometrica 82 (6),
22952326.
Cameron, A. C., J. B. Gelbach, and D. L. Miller (2011). Robust inference with
multiway clustering. Journal of Business & Economic Statistics 29 (2).
Campa, P. (2011). Gender quotas, female politicians and public expenditures:
Quasi-experimental evidence. Working Paper no. 157, Econpubblica, Universita Bocconi .
Casas-Arce, P. and A. Saiz (1995). Women and power: Unwilling, ineffective, or
held back? Journal of Political Economy 123 (3), 641669.
Cellini, S. R., F. Ferreira, and J. Rothstein (2010). The value of school facility
investments: Evidence from a dynamic regression discontinuity design. The
Quarterly Journal of Economics 125 (1), 215261.
Corbi, R., E. Papaioannou, and P. Surico (2014). Federal transfer multipliers:
Quasi-experimental evidence from Brazil. NBER Working Paper No. 20751 .
Curto-Grau, M., A. Sole-Olle, and P. Sorribas-Navarro (2015). Does electoral
competition curb party favoritism? Unpublished manuscript, University of
Barcelona.
De Haan, J. and J.-E. Sturm (1997). Political and economic determinants of oecd
budget deficits and government expenditures: A reinvestigation. European
Journal of Political Economy 13 (4), 739750.
Deshpande, M. (2014). Does welfare inhibit success? The long-term effects of removing low-income youth from disability insurance. Unpublished manuscript,
MIT .
Eggers, A., R. Freier, V. Grembi, and T. Nannicini (2015). Regression discontinuity designs based on population thresholds: Pitfalls and solutions. Discussion
Papers of DIW Berlin, 1503 .
39
Pettersson-Lidbom, P. (2012). Does the size of the legislature affect the size of
government? Evidence from two natural experiments. Journal of Public Economics 96 (3), 269278.
Porter, J. (2003). Estimation in the regression discontinuity model. Unpublished
manuscript, Harvard University.
Roubini, N. and J. D. Sachs (1989). Political and economic determinants of budget
deficits in the industrial democracies. European Economic Review 33 (5), 903
933.
Salanova Alcalde, R. (2009). El concejo abierto en Aragon: la ley 9/2009, de 22 de
diciembre, reguladora de los concejos abiertos. Anuario aragones del gobierno
local (1), 193231.
Sanz, C. (2015). The effect of electoral systems on voter turnout: Evidence from
a natural experiment. Political Science Research and Methods, forthcoming.
Turner, M. and Q. Weninger (2005). Meetings with costly participation: An empirical analysis. The Review of Economic Studies 72 (1), 247268.
Vila, J. V. i. (2012). El endeudamiento de los municipios: Una cuestion de
comportamiento poltico? Presupuesto y Gasto P
ublico 66, 199216.
42
Tables
DirDem
RepDem
Total
19871991
668
7382
8050
19911995
710
7356
8066
19951999
795
7291
8086
19992003
878
7220
8098
20032007
913
7195
8108
20072011
949
7161
8110
DirDem RepDem
Total
1st to 2nd
54
13
67
2nd to 3rd
108
24
132
3rd to 4th
105
22
127
4th to 5th
77
43
120
5th to 6th
91
55
146
Total
435
157
592
Term
Note: The overall number of municipalities is not constant as some municipalities have been divided into two (see Appendix D).
43
mean
sd
p1
p50
p99
count
1095.2
1048.1
5.5
45.4
48.4
33.1
17.9
59.8
26.0
32.1
0.4
0.3
0.5
2.9
41.9
37.6
4.3
22.6
53.9
77.6
46.8
15.2
57.1
27.7
4.0
41.3
6070.6
1681.0
1641.5
22.5
28.4
18.1
16.0
20.0
20.9
13.4
13.8
0.5
0.5
0.5
31.5
19.4
15.3
4.6
17.8
11.3
8.4
6.9
6.5
7.2
11.7
6.4
33.4
49249.9
154.9
148.4
-8.9
-14.0
8.5
1.9
0.1
14.2
2.8
7.5
0.0
0.0
0.0
-72.1
3.2
2.5
0.0
0.2
31.2
51.7
34.0
0.0
36.1
7.3
0.0
0.0
30.0
800.9
764.8
0.0
46.5
48.9
32.4
9.7
60.5
25.0
30.3
0.0
0.0
1.0
2.1
43.6
38.1
3.0
18.5
52.6
78.8
46.2
15.6
57.9
26.2
1.6
36.4
721.0
5384.7
5189.0
60.9
88.3
88.4
75.2
81.8
100.0
59.8
73.6
1.0
1.0
1.0
84.3
84.7
71.7
22.4
76.0
85.1
92.5
64.1
28.3
71.1
59.5
30.7
100.0
83855.0
160593
160593
160593
160593
160593
160593
160593
160593
160593
160593
160288
160288
79930
79930
160466
160466
136299
160466
160466
160466
100541
100541
100541
100541
100421
80784
160593
Summary statistics for all municipalities. The unit of observation is a municipality-year. The sample
size is smaller for election variables because some data are missing from the official files. The sample is
even smaller for Votes Far Left because the United Left party did not run in every election. For RW
Winner and RW Difference, the sample is restricted to elections in which the PP and the PSOE were
the two most-voted parties. For the demographic variables, data are available only from 1996. Also note
that, in municipalities with no immigrants, EU Immigrants is missing by construction. See Appendix
D for a precise definition of the variables.
44
mean
sd
p1
p50
p99
count
1591.6
1546.5
3.5
48.1
39.9
28.0
31.5
52.2
16.4
35.1
0.5
0.2
0.6
14.3
53.8
29.7
3.3
33.0
60.5
78.3
53.7
8.5
52.9
38.5
2.5
45.2
97.6
1937.3
1890.7
22.8
32.2
19.4
17.9
24.6
24.2
11.5
17.8
0.5
0.4
0.5
43.0
20.7
15.3
3.9
21.8
13.1
8.7
5.9
5.4
9.2
11.5
4.9
43.5
33.1
178.4
173.8
-1.9
-36.8
5.9
0.4
0.1
9.9
0.6
5.5
0.0
0.0
0.0
-85.7
2.4
1.1
0.0
0.0
33.3
54.0
39.7
0.0
29.6
14.3
0.0
0.0
41.0
1167.9
1143.2
0.0
51.4
38.2
25.7
26.4
48.8
13.6
32.2
1.0
0.0
1.0
17.1
56.5
28.2
2.2
30.4
59.6
79.3
53.8
7.9
53.1
38.4
0.0
40.0
100.0
7413.0
7157.2
59.7
92.5
88.9
75.9
89.1
100.0
55.2
82.8
1.0
1.0
1.0
94.5
91.5
69.8
17.2
86.1
91.6
94.2
67.4
23.3
73.3
67.4
23.4
100.0
172.0
22531
22531
22531
22531
22531
22531
22531
22531
22531
22531
22487
22487
10495
10495
22474
22474
18586
22474
22474
22474
14968
14968
14968
14968
14961
7133
22531
Summary statistics for municipalities within the optimal bandwidth in the main regression for public
expenditures (54 inhabitants), calculated following Imbens and Kalyanaraman (2012)s procedure. The
unit of observation is a municipality-year. The sample size is smaller for election variables because
some data are missing from the official files. The sample is even smaller for Votes Far Left because
the United Left party did not run in every election. For RW Winner and RW Difference, the sample
is restricted to elections in which the PP and the PSOE were the two most-voted parties. For the
demographic variables, data are available only from 1996. Also note that, in municipalities with no
immigrants, EU Immigrants is missing by construction. See Appendix D for a precise definition of the
variables.
45
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Exp.
-0.0377
(0.0154)
21615
1633
YES
Optimala
(2)
Log Exp.
-0.0455
(0.0150)
29575
2038
YES
1.5 x Opt.
(3)
Log Exp.
-0.0306
(0.0185)
11082
1080
YES
.5 x Opt.
(4)
Log Exp.
-0.0580
(0.0192)
21667
1685
NO
Opt.
(5)
Log Exp.
-0.0622
(0.0183)
29627
2090
NO
1.5 x Opt.
(6)
Log Exp.
-0.0792
(0.0194)
11127
1125
NO
.5 x Opt.
(5)
Log Rev.
-0.0630
(0.0182)
29627
2090
NO
1.5 x Opt.
(6)
Log Rev.
-0.0789
(0.0180)
11127
1125
NO
.5 x Opt.
(5)
Deficit
0.0365
(0.378)
44955
2935
NO
1.5 x Opt.
(6)
Deficit
-0.261
(0.503)
24129
1815
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Rev.
-0.0322
(0.0148)
21275
1619
YES
Optimalb
(2)
Log Rev.
-0.0426
(0.0144)
29575
2038
YES
1.5 x Opt.
(3)
Log Rev.
-0.0245
(0.0176)
11082
1080
YES
.5 x Opt.
(4)
Log Rev.
-0.0523
(0.0185)
21325
1669
NO
Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Deficit
-0.445
(0.457)
36656
2418
YES
Optimalc
(2)
Deficit
-0.458
(0.445)
44911
2891
YES
1.5 x Opt.
(3)
Deficit
-0.573
(0.511)
24079
1765
YES
.5 x Opt.
(4)
Deficit
0.116
(0.402)
36712
2474
NO
Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors,
clustered at both municipality and the running variable, are in parentheses. Fixed effects are
municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. a Optimal BW = 54, b Optimal BW = 53. c Optimal
BW = 122. p < 0.10, p < 0.05, p < 0.01.
46
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Curr.
-0.0364
(0.0171)
19846
1550
YES
Optimala
(2)
Log Curr.
-0.0439
(0.0164)
27837
1944
YES
1.5 x Opt.
(3)
Log Curr.
-0.0248
(0.0207)
9770
1011
YES
.5 x Opt.
(4)
Log Curr.
-0.0746
(0.0246)
19896
1600
NO
Opt.
(5)
Log Curr.
-0.0833
(0.0253)
27886
1993
NO
1.5 x Opt.
(6)
Log Curr.
-0.100
(0.0241)
9815
1056
NO
.5 x Opt.
(5)
Log Cap.
0.00746
(0.0355)
30610
2254
NO
1.5 x Opt.
(6)
Log Cap.
0.0271
(0.0448)
12219
1228
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Cap.
0.0138
(0.0412)
23367
1786
YES
Optimalb
(2)
Log Cap.
-0.00595
(0.0406)
30545
2189
YES
1.5 x Opt.
(3)
Log Cap.
0.0275
(0.0497)
12154
1163
YES
.5 x Opt.
(4)
Log Cap.
0.0353
(0.0375)
23431
1850
NO
Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered
at both municipality and the running variable, are in parentheses. Fixed effects are municipality
and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 49, b Optimal BW = 64. p < 0.10, p < 0.05,
p < 0.01.
47
RW Winner
Observations
Municipalities
Fixed effects
Bandwidth
(1)
RW Mayor
0.566
(0.0315)
18772
2845
YES
Optimala
(2)
RW Mayor
0.550
(0.0225)
27717
3471
YES
1.5 x Opt.
(3)
RW Mayor
0.631
(0.0569)
9476
1843
YES
.5 x Opt.
(4)
RW Mayor
0.542
(0.0225)
18873
2946
NO
Opt.
(5)
RW Mayor
0.524
(0.0181)
27806
3560
NO
1.5 x Opt.
(6)
RW Mayor
0.568
(0.0329)
9549
1916
NO
.5 x Opt.
RW Mayor
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Exp.
0.00546
(0.0198)
49206
4496
YES
Optimalb
(2)
Log Exp.
-0.0139
(0.0160)
63115
4968
YES
1.5 x Opt.
(3)
Log Exp.
0.0163
(0.0311)
27515
3459
YES
.5 x Opt.
(4)
Log Exp.
0.0286
(0.0346)
49276
4566
NO
Opt.
(5)
Log Exp.
0.0313
(0.0277)
63176
5029
NO
1.5 x Opt.
(6)
Log Exp.
0.0705
(0.0530)
27604
3548
NO
.5 x Opt.
RW Mayor
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Rev.
0.00110
(0.0254)
33635
3801
YES
Optimalc
(2)
Log Rev.
-0.00278
(0.0200)
46966
4415
YES
1.5 x Opt.
(3)
Log Rev.
-0.00922
(0.0391)
17604
2742
YES
.5 x Opt.
(4)
Log Rev.
0.0562
(0.0466)
33717
3883
NO
Opt.
(5)
Log Rev.
0.0326
(0.0364)
47037
4486
NO
1.5 x Opt.
(6)
Log Rev.
0.0188
(0.0642)
17704
2842
NO
.5 x Opt.
Results from a fuzzy regression discontinuity design in which the running variable is RW Difference (the
difference in the percent of votes between the PP and the PSOE) and the treatment variable is RW
Mayor (having a PP mayor). The sample is restricted to elections where the PP and PSOE were the two
most-voted parties. Each column is a separate local linear regression with a uniform kernel. Standard
errors, clustered at the municipality level, are in parentheses. Fixed effects are municipality and year
fixed effects. The optimal bandwidth (BW) is based on Imbens and Kalyanaraman (2012). a Optimal
BW = 9 percentage points, b Optimal BW = 28 percentage points, c Optimal BW = 17 percentage
points. p < 0.10, p < 0.05, p < 0.01.
48
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
|
uexp,mt |
-0.0205
(0.0132)
6950
826
YES
Optimala
(2)
|
uexp,mt |
-0.0134
(0.0106)
10211
1039
YES
1.5 x Opt.
(3)
|
uexp,mt |
-0.0140
(0.0265)
3389
570
YES
.5 x Opt.
(4)
|
uexp,mt |
-0.0133
(0.0112)
40457
2678
NO
Optimalb
(5)
|
uexp,mt |
-0.00564
(0.0110)
49930
3196
NO
1.5 x Opt.
(6)
|
uexp,mt |
-0.00715
(0.0119)
28402
2017
NO
.5 x Opt.
(4)
|
urev,mt |
-0.0155
(0.0113)
40457
2678
NO
Optimald
(5)
|
urev,mt |
-0.0118
(0.0110)
49930
3196
NO
1.5 x Opt.
(6)
|
urev,mt |
-0.00642
(0.0118)
28402
2017
NO
.5 x Opt.
(4)
|
udef,mt |
0.259
(0.361)
40457
2678
NO
Optimalf
(5)
|
udef,mt |
0.323
(0.339)
49930
3196
NO
1.5 x Opt.
(6)
|
udef,mt |
0.182
(0.446)
28402
2017
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
|
urev,mt |
-0.00866
(0.0104)
8133
907
YES
Optimalc
(2)
|
uexp,mt |
-0.00523
(0.00901)
12213
1136
YES
1.5 x Opt.
(3)
|
urev,mt |
-0.00664
(0.0203)
4103
628
YES
.5 x Opt.
Panel C: Deficit
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
|
udef,mt |
-0.947
(0.377)
17837
1435
YES
Optimale
(2)
|
udef,mt |
-0.641
(0.319)
25726
1845
YES
1.5 x Opt.
(3)
|
udef,mt |
-1.172
(0.604)
9008
965
YES
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. The dependent variable is
the absolute value of the residuals obtained from estimating Equation 3. Standard errors, clustered
at both municipality and the running variable, are in parentheses. Fixed effects are municipality
and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 15, b Optimal BW = 150, c Optimal BW = 13, d Optimal
BW = 150, e Optimal BW = 48, f Optimal BW = 150. p < 0.10, p < 0.05, p < 0.01.
49
50
-0.00960
(0.0711)
11082
1080
YES
.5 x Opt.
-0.0758
(0.0550)
-0.0493
(0.0271)
-0.0383
(0.0359)
29575
2038
YES
1.5 x Opt.
-0.0184
(0.0336)
(3)
Log Exp.
-0.0258
(0.0218)
-0.0200
(0.0189)
(2)
Log Exp.
-0.0393
(0.0156)
-0.00400
(0.0417)
21275
1619
YES
Optimalb
-0.0450
(0.0316)
-0.00513
(0.0232)
(4)
Log Rev.
-0.0287
(0.0162)
-0.0333
(0.0345)
29575
2038
YES
1.5 x Opt.
-0.0563
(0.0264)
-0.0201
(0.0185)
(5)
Log Rev.
-0.0367
(0.0150)
0.0211
(0.0657)
11082
1080
YES
.5 x Opt.
-0.0899
(0.0500)
-0.0141
(0.0328)
(6)
Log Rev.
-0.0211
(0.0210)
-0.387
(0.917)
36656
2418
YES
Optimalc
0.966
(1.045)
0.207
(0.510)
(7)
Deficit
-0.566
(0.490)
-0.520
(0.905)
44911
2891
YES
1.5 x Opt.
0.928
(1.043)
0.211
(0.508)
(8)
Deficit
-0.613
(0.480)
-0.391
(1.145)
24079
1765
YES
.5 x Opt.
1.246
(1.338)
-0.327
(0.620)
(9)
Deficit
-0.644
(0.546)
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered at both municipality and the
running variable, are in parentheses. Fixed effects are municipality and year fixed effects. I calculate the optimal bandwidth following
Imbens and Kalyanaraman (2012)s procedure. a Optimal BW = 54, b Optimal BW = 53, c Optimal BW = 122. p < 0.10,
p < 0.05, p < 0.01.
Observations
Municipalities
Fixed effects
Bandwidth
-0.00786
(0.0437)
21615
1633
YES
Optimala
-0.0393
(0.0327)
DirDem x N3
DirDem x N4
-0.0137
(0.0228)
DirDem x N2
DirDem
(1)
Log Exp.
-0.0342
(0.0168)
Variable
Log Transfers
(1)
Mean
6.25
53.9
29.6
3.3
33
60.5
Turnout (%)
78.4
53.7
Young (%)
8.6
Middle-Aged (%)
52.9
Old (%)
38.5
Immigrants (%)
2.5
EU Immigrants (%)
45.3
(2)
Opt. BW Value
61
[N = 24079]
94
[N = 33493]
76
[N = 29602]
77
[N = 24515]
58
[N = 23834]
58
[N = 23834]
55
[N = 22715]
104
[N = 23364]
79
[N = 20186]
83
[N = 20888]
148
[N = 27291]
150
[N = 27398]
62
[N = 7874]
Fixed effects
(3)
OBW Res.
-0.0384
(0.0436)
0.409
(0.616)
-0.863
(0.491)
0.226
(0.199)
0.933
(0.903)
-0.0174
(0.536)
-0.105
(0.391)
-0.112
(0.184)
-0.00899
(0.228)
0.0693
(0.421)
0.122
(0.424)
-0.252
(0.333)
-0.430
(3.619)
YES
(4)
.5 x OBW
0.0176
(0.0485)
0.424
(0.632)
-0.493
(0.560)
0.149
(0.243)
0.430
(1.062)
0.113
(0.636)
-0.216
(0.480)
-0.230
(0.191)
-0.0790
(0.244)
0.242
(0.423)
-0.139
(0.426)
-0.191
(0.327)
-0.229
(3.806)
YES
(5)
OBW Res.
-0.147
(0.0544)
1.315
(1.368)
-2.034
(0.923)
0.214
(0.226)
1.872
(1.293)
0.156
(0.798)
-0.709
(0.466)
0.0305
(0.388)
0.102
(0.324)
-0.883
(0.655)
1.054
(0.777)
-0.167
(0.226)
-2.164
(3.304)
NO
(6)
.5 x OBW
-0.133
(0.0600)
2.146
(1.513)
-2.897
(0.914)
0.124
(0.271)
2.263
(1.464)
0.602
(0.911)
-0.419
(0.500)
-0.178
(0.388)
0.00998
(0.327)
-0.910
(0.677)
0.373
(0.812)
-0.0977
(0.240)
-3.887
(4.270)
NO
Column 1 shows the mean of the variables for municipalities close to the threshold. Column 2 shows the optimal
bandwidth value and number of observations for a placebo test that estimates the effect of direct democracy on
the corresponding variable. Columns 3 to 7 show the results for the placebo tests: each column is a separate local
linear regression with a uniform kernel. Standard errors, clustered at both municipality and the running variable,
are in parentheses. Fixed effects are municipality and year fixed effects. The optimal bandwidth (BW) is based
on Imbens and Kalyanaraman (2012). p < 0.10, p < 0.05, p < 0.01.
51
52
(2)
Log Rev.
-0.0228
(0.0180)
19948
1618
Opt.
YES
3%
(1)
Log Rev.
-0.0322
(0.0148)
21275
1619
Optimalb
YES
0
(4)
Log Exp.
-0.0496
(0.0288)
17473
1616
Opt.
YES
10%
(5)
Log Exp.
-0.0580
(0.0192)
21667
1685
Opt.
NO
0
(4)
Log Rev.
-0.0408
(0.0276)
17133
1602
Opt.
YES
10%
(5)
Log Rev.
-0.0523
(0.0185)
21325
1669
Opt.
NO
0
(3)
Log Rev.
-0.0344
(0.0208)
19115
1615
Opt.
YES
5%
(3)
Log Exp.
-0.0386
(0.0217)
19455
1629
Opt.
YES
5%
(6)
Log Rev.
-0.0517
(0.0228)
19999
1669
Opt.
NO
3%
(6)
Log Exp.
-0.0559
(0.0236)
20341
1685
Opt.
NO
3%
(7)
Log Rev.
-0.0526
(0.0264)
19165
1665
Opt.
NO
5%
(7)
Log Exp.
-0.0582
(0.0268)
19507
1681
Opt.
NO
5%
(8)
Log Rev.
-0.0359
(0.0321)
17183
1652
Opt.
NO
10%
(8)
Log Exp.
-0.0511
(0.0324)
17525
1668
Opt.
NO
10%
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
DirDem
-0.445
-0.0610
-0.224
-0.752
0.116
0.326
0.371
-0.00513
(0.457)
(0.545)
(0.587)
(0.621)
(0.402)
(0.438)
(0.423)
(0.432)
Observations
36656
35329
34494
32510
36712
35386
34552
32570
Municipalities
2418
2417
2414
2402
2474
2474
2472
2462
Bandwidth
Optimalc
Opt.
Opt.
Opt.
Opt.
Opt.
Opt.
Opt.
Fixed effects
YES
YES
YES
YES
NO
NO
NO
NO
Excluded
0
3%
5%
10%
0
3%
5%
10%
Each column is a separate local linear regression with a uniform kernel. Observations within the interval around the
threshold indicated in the Excluded row are dropped; e.g., municipalities with 97 to 103 inhabitants for Column
2. Standard errors, clustered at both municipality and the running variable, are in parentheses. Fixed effects
are municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 54, b Optimal BW = 53. c Optimal BW = 122. p < 0.10, p < 0.05,
p < 0.01.
Observations
Municipalities
Bandwidth
Fixed effects
Excluded
DirDem
Observations
Municipalities
Bandwidth
Fixed effects
Excluded
DirDem
(2)
Log Exp.
-0.0242
(0.0188)
20288
1632
Opt.
YES
3%
(1)
Log Exp.
-0.0377
(0.0154)
21615
1633
Optimala
YES
0
53
(2)
Log Rev.
-0.0355
(0.0152)
28814
2037
1.5 x Opt.
YES
Into
(2)
Deficit
-0.434
(0.471)
43978
2889
1.5 x Opt.
YES
Into
(1)
Log Rev.
-0.0259
(0.0157)
20613
1618
Optimalb
YES
Into
(1)
Deficit
-0.424
(0.485)
35803
2415
Optimalc
YES
Into
(3)
Deficit
-0.692
(0.544)
23391
1762
.5 x Opt.
YES
Into
(3)
Log Rev.
-0.0175
(0.0186)
10526
1069
.5 x Opt.
YES
Into
(3)
Log Exp.
-0.0238
(0.0194)
10526
1069
.5 x Opt.
YES
Into
(4)
Deficit
0.193
(0.429)
35859
2471
Opt.
NO
Into
(6)
Log Exp.
-0.0643
(0.0216)
10572
1115
.5 x Opt.
NO
Into
(5)
Log Rev.
-0.0498
(0.0190)
28865
2088
1.5 x Opt.
NO
Into
(6)
Log Rev.
-0.0637
(0.0203)
10572
1115
.5 x Opt.
NO
Into
(5)
Log Exp.
-0.0486
(0.0193)
28865
2088
1.5 x Opt.
NO
Into
(7)
Log Rev.
-0.0417
(0.0225)
19710
1617
Opt.
YES
Out of
(7)
Log Exp.
-0.0470
(0.0235)
20044
1631
Opt.
YES
Out of
(5)
Deficit
0.104
(0.396)
44021
2932
1.5 x Opt.
NO
Into
(6)
Deficit
-0.236
(0.555)
23440
1811
.5 x Opt.
NO
Into
(7)
Deficit
-0.298
(0.596)
34905
2414
Opt.
YES
Out of
(4)
Log Rev.
-0.0372
(0.0198)
20662
1667
Opt.
NO
Into
(4)
Log Exp.
-0.0430
(0.0206)
20998
1683
Opt.
NO
Into
(8)
Deficit
-0.283
(0.590)
43080
2888
1.5 x Opt.
YES
Out of
(8)
Log Rev.
-0.0564
(0.0218)
27916
2036
1.5 x Opt.
YES
Out of
(8)
Log Exp.
-0.0600
(0.0227)
27916
2036
1.5 x Opt.
YES
Out of
(9)
Deficit
-0.527
(0.651)
22488
1761
.5 x Opt.
YES
Out of
(9)
Log Rev.
-0.0303
(0.0309)
9603
1065
.5 x Opt.
YES
Out of
(9)
Log Exp.
-0.0399
(0.0327)
9603
1065
.5 x Opt.
YES
Out of
(10)
Deficit
0.453
(0.427)
34962
2471
Opt.
NO
Out of
(10)
Log Rev.
-0.0292
(0.0276)
19761
1668
Opt.
NO
Out of
(10)
Log Exp.
-0.0333
(0.0278)
20097
1684
Opt.
NO
Out of
(11)
Deficit
0.368
(0.404)
43124
2932
1.5 x Opt.
NO
Out of
(11)
Log Rev.
-0.0526
(0.0237)
27968
2088
1.5 x Opt.
NO
Out of
(11)
Log Exp.
-0.0497
(0.0239)
27968
2088
1.5 x Opt.
NO
Out of
(12)
Deficit
0.00980
(0.554)
22539
1812
.5 x Opt.
NO
Out of
(12)
Log Rev.
-0.0597
(0.0341)
9655
1117
.5 x Opt.
NO
Out of
(12)
Log Exp.
-0.0554
(0.0361)
9655
1117
.5 x Opt.
NO
Out of
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered at both municipality and the running variable, are in parentheses.
Columns with switches Into (Out of) exclude from the sample the municipality-years corresponding to terms in which a municipality switched from direct
(representative) to representative (direct) democracy. FE are municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. a Optimal BW = 54, b Optimal BW = 53. c Optimal BW = 122. p < 0.10, p < 0.05, p < 0.01.
Obs.
Mun.
BW
FE
Switches
DirDem
Obs.
Mun.
BW
FE
Switches
DirDem
Obs.
Mun.
BW
FE
Switches
DirDem
(2)
Log Exp.
-0.0385
(0.0158)
28814
2037
1.5 x Opt.
YES
Into
(1)
Log Exp.
-0.0318
(0.0165)
20947
1632
Optimala
YES
Into
DirDem
DirDem(t+1)
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Exp.
-0.0515
(0.0152)
(2)
Log Exp.
-0.0635
(0.0149)
(3)
Log Exp.
-0.0387
(0.0188)
(4)
Log Exp.
-0.0603
(0.0159)
(5)
Log Exp.
-0.0649
(0.0154)
(6)
Log Exp.
-0.0768
(0.0179)
-0.0146
(0.0145)
21615
1633
YES
Optimala
-0.0163
(0.0138)
29575
2038
YES
1.5 x Opt.
0.00761
(0.0175)
11082
1080
YES
.5 x Opt.
-0.0154
(0.0169)
21667
1685
NO
Opt.
-0.0199
(0.0160)
29627
2090
NO
1.5 x Opt.
-0.0264
(0.0218)
11127
1125
NO
.5 x Opt.
DirDem
DirDem(t+1)
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Rev.
-0.0465
(0.0144)
(2)
Log Rev.
-0.0609
(0.0142)
(3)
Log Rev.
-0.0330
(0.0179)
(4)
Log Rev.
-0.0580
(0.0149)
(5)
Log Rev.
-0.0674
(0.0146)
(6)
Log Rev.
-0.0777
(0.0164)
-0.00889
(0.0143)
21275
1619
YES
Optimalb
-0.0151
(0.0133)
29575
2038
YES
1.5 x Opt.
0.00999
(0.0172)
11082
1080
YES
.5 x Opt.
-0.00846
(0.0160)
21325
1669
NO
Opt.
-0.0179
(0.0155)
29627
2090
NO
1.5 x Opt.
-0.0263
(0.0202)
11127
1125
NO
.5 x Opt.
DirDem
DirDem(t+1)
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Deficit
-0.476
(0.485)
(2)
Deficit
-0.443
(0.478)
(3)
Deficit
-0.489
(0.520)
(4)
Deficit
0.322
(0.367)
(5)
Deficit
0.327
(0.361)
(6)
Deficit
0.0464
(0.413)
-0.203
(0.450)
36656
2418
YES
Optimalc
-0.259
(0.435)
44911
2891
YES
1.5 x Opt.
-0.400
(0.481)
24079
1765
YES
.5 x Opt.
-0.285
(0.441)
36712
2474
NO
Opt.
-0.366
(0.414)
44955
2935
NO
1.5 x Opt.
-0.532
(0.502)
24129
1815
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered
at both municipality and the running variable, are in parentheses. Fixed effects are municipality
and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 54, b Optimal BW = 53. c Optimal BW = 122. p < 0.10,
54
Figures
t-1
e4
e1
budget
election
budget
e2
t+1
e3
budget
population
figures
e4
budget
budget
election
55
e1
budget
Observations
1000
2000
Expenditures
2000
4000
6000
8000
Expenditures (euros per capita per year)
Observations
1000
2000
Revenues
2000
4000
6000
Revenues (euros per capita per year)
8000
500
Observations
1000 1500 2000 2500
Observations above the 99th and below the 1st percentiles have been excluded to facilitate
comprehension. Bins are 200-euro wide.
100
200
Population
300
400
56
.1 .05
Log Expenditures
.05 .1 .15 .2
.25
.3
Expenditures
50
100
Population
150
.1 .05
Log Revenues
.05 .1 .15
.2
.25
.3
Revenues
50
100
Population
150
Deficit
1
2
Deficit
50
100
Population
150
200
100+OBW
P
I estimate
j=100OBW
servations within the optimal bandwidth for Outcomemyt , where Populationj ,mt is a dummy that
indicates whether municipality m has population size j at term t. In the y-axis, I plot the estimated
coefficients j , averaged to 4-inhabitant-wide bins. I normalize the coefficients so that the average
bin immediately to the right of the threshold takes the value of zero. The lines are linear fits on j ,
fitted separately for observations above and below the threshold.
57
.1
.05
Log Current
.05
.1
.15
.2
Current Expenditures
50
100
Population
150
.1
Log Capital
.1
.2
.3
.4
Capital Expenditures
50
100
Population
150
100+OBW
P
j=100OBW
within the optimal bandwidth for Outcomemyt , where Populationj ,mt is a dummy that indicates whether municipality m has population size j at term t. In the y-axis, I plot the
estimated coefficients j , averaged to 4-inhabitant-wide bins. I normalize the coefficients so
that the average bin immediately to the right of the threshold takes the value of zero. The
lines are linear fits on j , fitted separately for observations above and below the threshold.
58
Votes Right
Votes Left
6 4 2 0 2 4
100
150
Population
Share Young
6 4 2 0 2 4
100
150
Population
50 100150200250
Population
Covariate
smoothness.
50
100
150
Population
50 100 150
Population
50 100 150
Population
Share Immigrants
Share EU Immigrants
Share Immigrants
10 5 0
5
Share Old
10 5 0 5 10
Share Old
7:
Votes Winner
Share Young
Turnout
4 2 0 2
50
100+OBW
P
Votes Difference
50
Turnout
Figure
50 100 150
Population
Votes Difference
15
105 0 5 1015
50 100 150
Population
Votes Winner
5 0 5 10
100
150
Population
50
Votes Left
50 100150200250
Population
estimate
Share EU Immigrants
10 5 0 5 10
Log Transfers
.1 0 .1 .2 .3 .4
Votes Right
105 0 5 10152025
Transfers
50
Outcomemyt
100
150
Population
m + y +
j=100OBW
optimal bandwidth for Outcomemyt , where Populationj ,mt is a dummy that indicates whether
municipality m has population size j at term t. In the y-axis, I plot the estimated coefficients j ,
averaged to 4-inhabitant-wide bins. I normalize the coefficients so that the average bin immediately
to the right of the threshold takes the value of zero. The lines are linear fits on j , fitted separately
for observations above and below the threshold.
59
.2
.1
.4
Lagged Expenditures
50
100
Population
150
.2
.1
.3
.4
Lagged Revenues
50
100
Population
150
Deficit (y4)
1
0
Lagged Deficit
50
100
150
Population
200
Figure 8: Effect of direct democracy on lagged main outcomes. I estimate Outcomem,y4 ,t1 =
m + y +
100+OBW
P
j=100OBW
vations within the optimal bandwidth for Outcomem,y4 ,t , where Populationj ,mt is a dummy that
indicates whether municipality m has population size j at term t. In the y-axis, I plot the estimated
coefficients j , averaged to 4-inhabitant-wide bins. I normalize the coefficients so that the average
bin immediately to the right of the threshold takes the value of zero. The lines are linear fits on j ,
fitted separately for observations above and below the threshold.
60
.1
.3
50
100
Population
150
.1
.5
50
100
Population
150
100+OBW
P
I estimate
j=100OBW
um,y4,t1 using the observations within the optimal bandwidth for Outcomem,y4 ,t , where
Populationj ,mt is a dummy that indicates whether municipality m has population size j at
term t. In the y-axis, I plot the estimated coefficients j , averaged to 4-inhabitant-wide bins.
I normalize the coefficients so that the average bin immediately to the right of the threshold
takes the value of zero. The lines are linear fits on j , fitted separately for observations
above and below the threshold.
61
.8
.6
CDF
.4
.2
0
.2
.4
CDF
.6
.8
Expenditures: tstatistics
.04
.02
0
.02
Point Estimates
pvalue = 0.032
.04
0
1
tstatistics
pvalue = 0
.8
.6
CDF
.4
.2
0
.2
.4
CDF
.6
.8
Revenues: tstatistics
.04
.02
0
.02
Point Estimates
pvalue = 0.05
.04
1
0
tstatistics
pvalue = 0.023
Figure 10: Placebo tests for other population thresholds. The graphs show the empirical
cumulative distribution functions for the point estimates (or t-statistics) obtained from running the main regressions for expenditures and revenues at all (fake) population thresholds
between 30 and 250 inhabitants. The vertical lines show the point estimates (or t-statistics)
for the effects of direct democracy, obtained from the (true) 100-inhabitant threshold. The
p-value below each graph shows the share of point estimates (or t-statistics) that are larger
in absolute value than the one for the 100-inhabitant threshold.
62
Supplementary Appendices
Appendix A: Proofs
Proof of Proposition 1
I need to show that if, for each P , sP = s, then (vA = l, wA = x) and (vB = r, wB =
x) is the unique equilibrium.
I first show that (vA = l, wA = x) and (vB = r, wB = x) is an equilibrium.
In the proposed equilibrium, Party A wins the election with probability l , and
its median voter, who is a leftist special interest, gets utility blx +lx if that happens.
If Party B wins, the median voter in Party A gets utility lx . Therefore, the expected
utility of the median voter in Party A, which I will denote by u(A), is u (A) =
l (blx + lx ) + r lx = lx + l blx in the proposed equilibrium.
Party A has three possible deviations: (a) If it deviates to proposing (vA =
l, wA = c), Party A reduces both the probability of winning (to c l ) and the utility
in case of a win (to blx ), so u(A) = c l blx + (1 c l )lx < u (A). (b) If it deviates
to (vA = r, wA = c), Party A increases the probability of winning from l to c but
at the cost of getting zero utility if that happens (it is proposing its least preferred
policy): u(A) = x lx < u (A). (c) If it deviates to (vA = r, wA = x), then both
parties propose the same policy and win with 1/2 probability, and u(A) = lx <
u (A).
Now consider Party B. In the proposed equilibrium, its median voter, who is
a rightist special interest, has expected utility u (B) = r (brx + rx ) + l rx =
rx + r brx . Consider the three possible deviations: (a) If it deviates to proposing
(vB = r, wB = c), u(B) = c r brx + (1 c r )rx < u (B). (b) If it deviates to
(vB = l, wB = c), u(B) = x rx < u (B). (c) If it deviates to (vB = l, wB = x),
u(B) = rx < u (B). This completes the proof that (vA = l, wA = x) and (vB =
r, wB = x) is an equilibrium.
I now show that the equilibrium (vA = l, wA = x) and (vB = r, wB = x) is
unique. To do so, I will show that (vA = l, wA = c) and (vB = r, wB = c) is
not an equilibrium (it is trivial to show that other possible proposals cannot be an
equilibrium). Under (vA = l, wA = c) and (vB = r, wB = c), u (A) = l blx . If A
deviates to (vA = l, wA = x), u(A) = (c l + x )(blx + lx ) > u (A). Thus, A prefers
to deviate, and (vA = l, wA = c) and (vB = r, wB = c) cannot be an equilibrium.
63
Proof of Proposition 2
(i) I need to show that, if for each P , sP = s, and for each v,
bvc
c
> ,
vc
x
then (vA = l, wA = x) and (vB = r, wB = x) is an equilibrium.
Under the proposed equilibrium, the median voter in Party A, who is a leftist
citizen, has an expected utility of u (A) = l blc . Consider his or her three possible
deviations: (a) If it deviates to (vA = l, wA = c), Party A increases the utility in
case of a win (to blx + lc ) but at the cost of reducing the probability of winning
(to c l ). Because
blc
lc
>
c
x ,
u (A).
blc
lc
>
c
x ,
by assumption,
u (A).
For Party
brc
x + c r
>
,
rc
x + c r r
64
deviates, if
blc
x + c l
>
lc
x + c l l
or
brc
x + c r
>
,
rc
x + c r r
Proof of Proposition 3
The proof of Proposition 3 follows directly from the proof of Proposition 1 and the
proof of Proposition 2.
especiales) account for 28% of own revenues. Typical fees include garbage collection
and water supply. Rents (property rents and the disposal of assets) account for
the remaining 32%. The results, shown in Table A6, point to an effect of a similar
extent in the three categories. This suggests that the reduction in revenues is acrossthe-board and not driven by a specific item, which could be used to benefit some
particular group or class of individuals.
Number of Observations
As explained in Section 4, some observations are missing because some municipalities
did not comply with their obligation of providing the data or because their data have
not been digitized. In Table A1, I show the share of available observations by year.65
The final data set contains 160,594 observations out of a possible 194,318 (82.6%),
and this share does not vary much over time, ranging from 77.7% in 1990 to 89.2%
in 2004. As shown in Appendix C, there is no effect of direct democracy on an
observations being missing.
Budget Variables
Log Expenditures is defined as the sum of all chapters of the expenditures budget,
net of the transfers to other tiers of governments of institutions (Chapters 4 and 7):
9
P
),
65
The overall number of municipalities changes across years because, during the sample period,
some municipalities were divided into two. I drop from the analysis the nine newly created municipalities within the maximum bandwidth (150 inhabitants) as well as the municipalities from which
they seceded.
67
where m denotes municipality, y denotes the year, t denotes the term, and Ek is
expenditures on Chapter k of the expenditures budget.
Analogously, Log Revenues, Log Transfers, Log Current, Log Capital, Log Taxes,
Log Rents and Log Fees are defined as:
8
P
Log Revenuesmyt
= log(
Log Transfersmyt
Log Currentmyt
Log Capitalmyt
Log Taxesmyt
Log Feesmyt
Log Rentsmyt
k=1
),
Populationmyt
R4,myt + R7,myt E4,myt E7,myt
),
log(
Populationmyt
E1,myt + E2,myt + E3,myt
log(
),
Populationmyt
E6,myt
),
log(
Populationmyt
R1,myt + R2,myt
log(
),
Populationmyt
R3,myt
log(
),
Populationmyt
R5,myt + R6,myt
log(
),
Populationmyt
Expendituresmyt Revenuesmyt
).
Revenuesmyt
Elections Variables
I use the following variable from national elections:
VotesCastmt
,
ElectoralCensusmt
PPVotesmt
= 100
,
ValidVotesmt
PSOEVotesmt
= 100
,
ValidVotesmt
IUVotesmt
= 100
,
ValidVotesmt
Turnoutmt = 100
VotesRightmt
VotesLeftmt
VotesFarLeftmt
68
where
VotesCastmt = VotesForCandidatesmt + BlankVotesmt + SpoiltVotesmt ,
ValidVotesmt = VotesForCandidatesmt + BlankVotesmt .
That is, Turnout measures the proportion of citizens who cast a vote over the
set of potential voters (the electoral census). There is no voter registration in Spain:
Potential voters are all citizens of Spain and other EU countries as well as countries
under Reciprocity Treaties; older than 18, and not disenfranchised by court order.66
Valid votes include votes for candidates and blank votes but not spoiled votes. I use
this denominator because it is relevant for the allocation of seats, as it is used to
determine whether parties reach the election threshold to get seats (3% in national
elections). Accordingly, it is the one that is normally reported by the media.
Observations for year y are from the Congress elections that take place during the
term t to which year y belongs. Because Congress elections have always alternated
perfectly with local elections, there is only one Congress election per term. For
example, for years 1988-1991, I consider the Congress elections of 1989. As explained
in the text, I lag the variables to use them as placebos.
Demographic Variables
The age distribution data are provided by the National Institute for Statistics in
intervals of five years, up to 85 or more until 2010 and up to 100 or more from
2011. To calculate the average age Mean Age, I use the mid-points of those intervals.
For ages 85100 (for which there is no five-year information until 2011), I calculate
the average age in 2011 (89.01) and use it for the rest of the years. For ages 100
or more, I assume the average is 102.5 years.
I define the share of young people Young as the share of individuals in the first
four intervals (19 years old or younger), Middle-Aged as the share of individuals in
the next nine intervals (ages 2064), and Old as the share of individuals in the final
five intervals (ages 65 or older).
66
69
Appendix Tables
Table A1: Number of Available Observations by Year
Year
Observations
Municipalities
Share Available
1988
6,873
8,059
85.3%
1989
6,792
8,063
84.2%
1990
6,268
8,066
77.7%
1991
6,475
8,074
80.2%
1992
6,655
8,079
82.4%
1993
6,649
8,086
82.2%
1994
6,485
8,086
80.2%
1995
6,405
8,094
79.1%
1996
6,489
8,095
80.1%
1997
6,702
8,097
82.8%
1998
6,707
8,098
82.8%
1999
6,615
8,101
81.7%
2000
6,805
8,104
84.0%
2001
6,574
8,107
81.1%
2002
6,344
8,108
78.2%
2003
7,020
8,108
86.6%
2004
7,236
8,109
89.2%
2005
7,139
8,109
88.0%
2006
7,067
8,110
87.1%
2007
6,555
8,111
80.8%
2008
6,630
8,112
81.7%
2009
6,498
8,112
80.1%
2010
6,487
8,114
79.9%
2011
7,154
8,116
88.1%
Total
160,594
194,318
82.6%
Note: The overall number of municipalities is not constant as some municipalities have been divided into two.
70
Table A2: Effect of Direct Democracy on Public Finances (Results Not Controlling
for Transfers)
(2)
Log Exp.
(3)
Log Exp.
(4)
Log Exp.
(5)
Log Exp.
(6)
Log Exp.
DirDem
-0.0581
(0.0263)
-0.0825
(0.0260)
-0.0232
(0.0312)
-0.147
(0.0394)
-0.161
(0.0395)
-0.156
(0.0438)
Observations
Municipalities
Fixed effects
Bandwidth
22483
1639
YES
Optimala
30792
2042
YES
1.5 x Opt.
11522
1086
YES
.5 x Opt.
22531
1687
NO
Opt.
30842
2092
NO
1.5 x Opt.
11563
1127
NO
.5 x Opt.
(2)
Log Rev.
(3)
Log Rev.
(4)
Log Rev.
(5)
Log Rev.
(6)
Log Rev.
DirDem
-0.0492
(0.0257)
-0.0772
(0.0256)
-0.0122
(0.0306)
-0.138
(0.0394)
-0.162
(0.0394)
-0.152
(0.0430)
Observations
Municipalities
Fixed effects
Bandwidth
22130
1627
YES
Optimalb
30792
2042
YES
1.5 x Opt.
11522
1086
YES
.5 x Opt.
22175
1672
NO
Opt.
30842
2092
NO
1.5 x Opt.
11563
1127
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Deficit
(2)
Deficit
(3)
Deficit
(4)
Deficit
(5)
Deficit
(6)
Deficit
-0.917
(0.870)
-1.054
(0.840)
-1.552
(0.939)
0.228
(0.576)
0.0354
(0.525)
-0.435
(0.692)
38100
2427
YES
Optimalc
46559
2898
YES
1.5 x Opt.
25042
1773
YES
.5 x Opt.
38149
2476
NO
Opt.
46598
2937
NO
1.5 x Opt.
25089
1820
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors,
clustered at both municipality and the running variable, are in parentheses. Fixed effects are
municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. a Optimal BW = 54, b Optimal BW = 53. c Optimal
BW = 122. p < 0.10, p < 0.05, p < 0.01.
71
(2)
RW Mayor
(3)
RW Mayor
(4)
RW Mayor
(5)
RW Mayor
(6)
RW Mayor
DirDem
0.00859
(0.0251)
0.0130
(0.0247)
0.00304
(0.0273)
0.00266
(0.0299)
0.00974
(0.0286)
0.00126
(0.0385)
Observations
Municipalities
Fixed effects
Bandwidth
31333
2144
YES
Optimala
38273
2509
YES
1.5 x Opt.
18213
1464
YES
.5 x Opt.
31380
2191
NO
Opt.
38328
2564
NO
1.5 x Opt.
18261
1512
NO
.5 x Opt.
(2)
LW Mayor
(3)
LW Mayor
(4)
LW Mayor
(5)
LW Mayor
(6)
LW Mayor
DirDem
-0.00684
(0.0217)
-0.0138
(0.0214)
0.0187
(0.0235)
-0.0260
(0.0254)
-0.0262
(0.0243)
-0.00118
(0.0282)
Observations
Municipalities
Fixed effects
Bandwidth
30584
2095
YES
Optimalb
37500
2466
YES
1.5 x Opt.
17359
1415
YES
.5 x Opt.
30629
2140
NO
Opt.
37556
2522
NO
1.5 x Opt.
17406
1462
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered
at both municipality and the running variable, are in parentheses. Fixed effects are municipality
and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 89, b Optimal BW = 85. p < 0.10, p < 0.05,
p < 0.01.
72
73
(3)
Votes Right
2.466
(1.300)
40365
2678
NO
Optimala
(4)
Votes Right
0.880
(1.314)
28334
2017
NO
.5 x Opt.
(5)
Votes Left
-0.783
(0.509)
33576
2256
YES
Optimalb
(6)
Votes Left
-0.281
(0.550)
20502
1580
YES
.5 x Opt.
(1)
Votes FarLeft
0.230
(0.199)
25619
2099
YES
Optimalc
(2)
Votes FarLeft
0.315
(0.222)
14301
1397
YES
.5 x Opt.
(3)
Votes FarLeft
0.181
(0.230)
25668
2148
NO
Optimalc
(4)
Votes FarLeft
0.0527
(0.283)
14358
1454
NO
.5 x Opt.
(5)
Turnout
0.127
(0.408)
19018
1506
YES
Optimald
(6)
Turnout
-0.523
(0.622)
9742
1011
YES
.5 x Opt.
(2)
Votes Right
0.497
(0.608)
28282
1965
YES
.5 x Opt.
(7)
Turnout
-0.372
(0.476)
19065
1553
NO
Optimald
(7)
Votes Left
-2.019
(0.965)
33629
2309
NO
Optimalb
(8)
Turnout
-0.422
(0.549)
9787
1056
NO
.5 x Opt.
(8)
Votes Left
-2.155
(0.934)
20554
1632
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered at both municipality and the running variable,
are in parentheses. Fixed effects are municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 150, b Optimal BW = 102, c Optimal BW = 87, d Optimal BW = 47. p < 0.10, p < 0.05, p < 0.01.
Observations
Municipalities
Fixed effects
Bandwidth
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
DirDem
(1)
Votes Right
0.774
(0.599)
40306
2619
YES
Optimala
74
(3)
Votes Right
2.551
(1.280)
40398
2678
NO
Optimale
(4)
Votes Right
1.089
(1.286)
28354
2017
NO
.5 x Opt.
(5)
Votes Left
-0.502
(0.515)
35036
2335
YES
Optimalf
(6)
Votes Left
-0.226
(0.533)
22229
1660
YES
.5 x Opt.
(1)
Votes FarLeft
0.182
(0.212)
28705
2260
YES
Optimalg
(2)
Votes FarLeft
0.347
(0.237)
17601
1585
YES
.5 x Opt.
(3)
Votes FarLeft
-0.0900
(0.202)
28765
2320
NO
Optimalg
(4)
Votes FarLeft
-0.136
(0.237)
17656
1640
NO
.5 x Opt.
(5)
Turnout
-0.101
(0.396)
18584
1483
YES
Optimalh
(6)
Turnout
0.329
(0.568)
9401
990
YES
.5 x Opt.
(2)
Votes Right
0.372
(0.539)
28302
1965
YES
.5 x Opt.
(7)
Turnout
-0.455
(0.458)
18630
1529
NO
Optimalh
(7)
Votes Left
-1.705
(0.962)
35093
2392
NO
Optimalf
(8)
Turnout
-0.177
(0.571)
9443
1032
NO
.5 x Opt.
(8)
Votes Left
-1.989
(0.920)
22284
1715
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered at both municipality and the running
variable, are in parentheses. Fixed effects are municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. e Optimal BW = 150, f Optimal BW = 111, g Optimal BW = 104, h Optimal BW = 46. p < 0.10,
Observations
Municipalities
Fixed effects
Bandwidth
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
DirDem
(1)
Votes Right
0.606
(0.530)
40339
2619
YES
Optimale
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Pers.
0.00789
(0.0272)
17293
1415
YES
Optimala
(2)
Log Pers.
-0.00248
(0.0256)
24906
1811
YES
1.5 x Opt.
(3)
Log Pers.
0.0302
(0.0321)
8529
932
YES
.5 x Opt.
(4)
Log Pers.
-0.0557
(0.0451)
17340
1462
NO
Opt.
(5)
Log Pers.
-0.0474
(0.0414)
24957
1862
NO
1.5 x Opt.
(6)
Log Pers.
-0.0449
(0.0539)
8574
977
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Serv.
-0.0428
(0.0197)
21910
1651
YES
Optimalb
(2)
Log Serv.
-0.0499
(0.0190)
30027
2063
YES
1.5 x Opt.
(3)
Log Serv.
-0.0315
(0.0238)
11074
1080
YES
.5 x Opt.
(4)
Log Serv.
-0.0759
(0.0282)
21960
1701
NO
Opt.
(5)
Log Serv.
-0.0850
(0.0270)
30071
2107
NO
1.5 x Opt.
(6)
Log Serv.
-0.131
(0.0258)
11118
1124
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors,
clustered at both municipality and the running variable, are in parentheses. Fixed effects are
municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. a Optimal BW = 43, b Optimal BW = 55. p < 0.10,
p < 0.05, p < 0.01.
75
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Taxes
-0.0576
(0.0208)
32317
2192
YES
Optimala
(2)
Log Taxes
-0.0545
(0.0207)
39286
2554
YES
1.5 x Opt.
(3)
Log Taxes
-0.0440
(0.0218)
19062
1506
YES
.5 x Opt.
(4)
Log Taxes
-0.101
(0.0357)
32367
2242
NO
Opt.
(5)
Log Taxes
-0.0884
(0.0353)
39344
2612
NO
1.5 x Opt.
(6)
Log Taxes
-0.124
(0.0356)
19109
1553
NO
.5 x Opt.
(4)
Log Fees
-0.0537
(0.0487)
24377
1834
NO
Opt.
(5)
Log Fees
-0.118
(0.0465)
32004
2231
NO
1.5 x Opt.
(6)
Log Fees
-0.130
(0.0522)
12645
1211
NO
.5 x Opt.
(4)
Log Fees
-0.0537
(0.0487)
24377
1834
NO
Opt.
(5)
Log Fees
-0.118
(0.0465)
32004
2231
NO
1.5 x Opt.
(6)
Log Fees
-0.130
(0.0522)
12645
1211
NO
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Fees
-0.111
(0.0416)
24325
1782
YES
Optimalb
(2)
Log Fees
-0.129
(0.0406)
31950
2177
YES
1.5 x Opt.
(3)
Log Fees
-0.107
(0.0447)
12597
1163
YES
.5 x Opt.
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Log Fees
-0.111
(0.0416)
24325
1782
YES
Optimalc
(2)
Log Fees
-0.129
(0.0406)
31950
2177
YES
1.5 x Opt.
(3)
Log Fees
-0.107
(0.0447)
12597
1163
YES
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. Standard errors, clustered
at both municipality and the running variable, are in parentheses. Fixed effects are municipality
and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman
(2012)s procedure. a Optimal BW = 94, b Optimal BW = 62, c Optimal BW = 68. p < 0.10,
p < 0.05, p < 0.01.
76
77
(2)
Log Rev.
-0.0347
(0.0142)
21275
1619
YES
Opt.
YES
(1)
Log Rev.
-0.0322
(0.0148)
21275
1619
YES
Optimalb
NO
(4)
Log Exp.
-0.0383
(0.0179)
11082
1080
YES
.5 x Opt.
YES
(5)
Log Exp.
-0.0580
(0.0192)
21667
1685
NO
Opt.
NO
(4)
Log Rev.
-0.0311
(0.0172)
11082
1080
YES
.5 x Opt.
YES
(5)
Log Rev.
-0.0523
(0.0185)
21325
1669
NO
Opt.
NO
(3)
Log Rev.
-0.0431
(0.0138)
29575
2038
YES
1.5 x Opt.
YES
(3)
Log Exp.
-0.0464
(0.0144)
29575
2038
YES
1.5 x Opt.
YES
(6)
Log Rev.
-0.0502
(0.0183)
21325
1669
NO
Opt.
YES
(6)
Log Exp.
-0.0554
(0.0190)
21667
1685
NO
Opt.
YES
(7)
Log Rev.
-0.0602
(0.0179)
29627
2090
NO
1.5 x Opt.
YES
(7)
Log Exp.
-0.0593
(0.0180)
29627
2090
NO
1.5 x Opt.
YES
(8)
Log Rev.
-0.0826
(0.0179)
11127
1125
NO
.5 x Opt.
YES
(8)
Log Exp.
-0.0827
(0.0193)
11127
1125
NO
.5 x Opt.
YES
(1)
(2)
(3)
(4)
(5)
(6)
(7)
(8)
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
Deficit
DirDem
-0.445
-0.299
-0.279
-0.308
0.116
0.117
0.0982
-0.0781
(0.457)
(0.343)
(0.339)
(0.383)
(0.402)
(0.327)
(0.313)
(0.396)
Observations
36656
36656
44911
24079
36712
36712
44955
24129
Municipalities
2418
2418
2891
1765
2474
2474
2935
1815
Fixed effects
YES
YES
YES
YES
NO
NO
NO
NO
Bandwidth
Optimalc
Opt.
1.5 x Opt. .5 x Opt.
Opt.
Opt.
1.5 x Opt.
.5 x Opt.
Top-Coding
NO
YES
YES
YES
NO
YES
YES
YES
Each column is a separate local linear regression with a uniform kernel. The top-coded regressions exclude the observations with a dependent-variable value above the top or below the bottom 1% of the observations within the
bandwidth. Standard errors, clustered at both municipality and the running variable, are in parentheses. Fixed effects
are municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and Kalyanaraman (2012)s
procedure. a Optimal BW = 54, b Optimal BW = 54, c Optimal BW = 105. p < 0.10, p < 0.05, p < 0.01.
Observations
Municipalities
Fixed effects
Bandwidth
Top-Coding
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
Top-Coding
DirDem
(2)
Log Exp.
-0.0397
(0.0148)
21615
1633
YES
Opt.
YES
(1)
Log Exp.
-0.0377
(0.0154)
21615
1633
YES
Optimala
NO
DirDem
Observations
Municipalities
Fixed effects
Bandwidth
(1)
Obs.
0.0126
(0.0171)
39704
2025
YES
Optimalc
(2)
Obs.
0.0164
(0.0166)
49460
2411
YES
1.5 x Opt.
(3)
Obs.
0.0152
(0.0194)
20964
1363
YES
.5 x Opt.
(4)
Obs.
-0.0136
(0.0179)
39704
2025
NO
Opt.
(5)
Obs.
-0.00120
(0.0177)
49460
2411
NO
1.5 x Opt.
(6)
Obs.
-0.0180
(0.0226)
20964
1363
NO
.5 x Opt.
Each column is a separate local linear regression with a uniform kernel. The dependent variable
is a dummy that indicates whether the observation is available or missing. Standard errors,
clustered at both municipality and the running variable, are in parentheses. Fixed effects are
municipality and year fixed effects. I calculate the optimal bandwidth following Imbens and
Kalyanaraman (2012)s procedure. a Optimal BW = 73 inhabitants. p < 0.10, p < 0.05,
p < 0.01.
78
Table A9: Placebo Tests and Covariate Smoothness for Effects of Political Parties
on Policy
(1)
Mean
(2)
Opt. BW Value
(3)
OBW Res.
(4)
.5 x OBW
(5)
OBW Res.
(6)
.5 x OBW
Log Transfers
6.25
53.8
29.7
3.28
23
[N = 41857]
12
[N = 24079]
10
[N = 20661]
9
[N = 15113]
13
[N = 25441]
15
[N = 29365]
11
[N = 21658]
12
[N = 16207]
14
[N = 18157]
14
[N = 18916]
20
[N = 25609]
16
[N = 21267]
15
[N = 16717]
-0.0284
(0.0365)
-0.479
(0.519)
0.671
(0.558)
-0.00246
(0.333)
-0.182
(0.919)
0.0355
(0.466)
-0.549
(0.452)
-0.0659
(0.173)
-0.0235
(0.199)
0.243
(0.250)
-0.0609
(0.217)
-0.125
(0.424)
3.484
(2.669)
YES
-0.0131
(0.0543)
-1.193
(0.860)
1.529
(0.944)
-0.586
(0.532)
-0.420
(1.494)
-0.581
(0.773)
-0.276
(0.748)
0.0549
(0.292)
0.212
(0.302)
-0.0872
(0.366)
-0.0501
(0.337)
-0.704
(0.737)
2.112
(4.496)
YES
0.0456
(0.0504)
-0.450
(0.979)
1.271
(0.938)
-0.885
(0.470)
0.193
(1.040)
0.763
(0.601)
-1.728
(0.754)
0.691
(0.679)
-0.998
(0.593)
-0.396
(0.625)
1.557
(0.870)
0.0441
(0.637)
3.253
(3.259)
NO
0.0822
(0.0717)
-1.791
(1.344)
2.197
(1.327)
-0.987
(0.649)
0.128
(1.422)
1.019
(0.821)
-1.487
(1.002)
1.529
(0.889)
-1.796
(0.785)
-0.794
(0.840)
1.130
(1.182)
0.460
(0.948)
1.357
(4.400)
NO
Variable
33
60.5
Turnout (%)
78.3
53.7
Young (%)
8.55
Middle-Aged (%)
52.9
Old (%)
38.5
Immigrants (%)
2.46
EU Immigrants (%)
45.2
Fixed effects
Column 1 shows the mean of the variables for municipalities close to the threshold. Column 2 shows the optimal
bandwidth value and number of observations for a placebo test that estimates the effect of RW Mayor on the corresponding variable. The test is based on a fuzzy regression discontinuity design in which the running variable is RW
Difference (the difference in the percent of votes between the PP and the PSOE) and the treatment variable is RW
Mayor (having a PP mayor). The sample is restricted to elections where the PP and PSOE were the two most-voted
parties. Columns 3 to 7 show the results for the placebo tests: each column is a separate local linear regression with
a uniform kernel. Standard errors, clustered at both municipality and the running variable, are in parentheses. Fixed
effects are municipality and year fixed effects. The optimal bandwidth (BW) is based on Imbens and Kalyanaraman
(2012). p < 0.10, p < 0.05, p < 0.01.
79
Appendix Figures
Expenditures
.2
RW Mayor
.4
Log Expenditures
6.5
.6
.8
RW Mayor
10
5
0
5
RW Difference (%)
10
30
30
Deficit
Deficit(%)
4
6
Log Revenues
6.5
10
Revenues
15
0
15
RW Difference (%)
20
10
0
10
RW Difference (%)
20
10
10
Capital Expenditures
5.5
6.5
Current Expenditures
5
0
5
RW Difference (%)
10
0
RW Difference (%)
10
15
0
RW Difference (%)
15
Figure A1: The effect of political parties on policy under representative democracy. The x-axis
shows the difference in the percent of votes between the PP and the PSOE in local elections. The
sample is restricted to elections in which those two parties were the two most-voted parties. The
bins are local averages of the dependent variables at 2-percentage-points intervals. The lines are
fits of linear regressions run separately at both sides of the threshold.
80
1000
Observations
500
0
100
50
0
RW Difference (%)
50
100
81