# Do Parties Matter for Economic Outcomes? A Regression-Discontinuity Approach

### Text-only Preview

Do Parties Matter for Economic Outcomes ?

A Regression-Discontinuity Approach?

Per Pettersson-Lidbom?

July 12, 2007

**Abstract**

A long-standing issue in political economics is to what extent party control makes a

difference in determining fiscal and economics policies. This question is very difficult to

answer empirically since parties are not randomly selected to govern political entities.

This paper uses a regression-discontinuity design, i.e. party control changes

discontinuously at 50 percent of the vote share, which can produce “near” experimental

causal estimates of the effect of party control on economic outcomes. The method is

applied to a large panel data set from Swedish local governments with a number of

attractive features. The results show that there is an economically significant party effect:

left-wing governments spend and tax 2-3 percent more than right-wing governments.

Left-wing governments also have 7 percent lower unemployment rates, which is partly

due to left-wing governments employing 4 percent more workers than right-wing

governments.

Key words: political parties, party control, partisan politics, regression-discontinuity

design, natural experiments, unemployment, government employees, fiscal policy

? An earlier version of this paper has been circulated under the title “Do Parties Matter for Fiscal Policy

Choices? A Regression-Discontinuity Approach”. The idea of using a discontinuity as a source of

identifying information of the party effect originates from a conversation with David Strömberg. The author

gratefully acknowledges helpful comments from the editor Roberto Perotti, Torsten Persson, Jakob

Svensson, Justin Wolfers, two anonymous referees, and seminars participants at MIT, UC Berkeley,

Harvard University, Princeton University, University of Pennsylvania, Göteborg University and Uppsala

University. The views expressed in the paper are mine, as is the responsibility for any mistakes. Financial

support from Jan Wallander’s Research Foundation is gratefully acknowledged.

? Department of Economics, Stockholm University, S-106 91 Stockholm, Sweden; e-mail: [email protected]

2

**1. Introduction**

This paper estimates the

*causal*effect of party control on fiscal and economic policies.1

Estimating the party effect without bias is a very challenging identification problem since

parties are not randomly selected to govern political entities. For example, since voters

select parties to govern, there may be an omitted variable problem due to unmeasured

voter preferences.2 Thus, a correlation between party control and some policy outcome

does not necessarily imply causation. The large empirical literature dealing with partisan

cycles in macroeconomic outcomes (e.g., growth, unemployment and inflation) is also

plagued by similar problems of endogeneity of party control.3 Voters may, for example,

elect conservative governments when recession is anticipated which will lead to a

spurious relationship between party control and economic outcomes. While many studies

claim to find strong empirical support for partisan differences in some macroeconomic

outcomes, Faust and Irons (1999) argue that there is only weak evidence that party control

is of importance when issues of simultaneous causality bias and omitted variable bias are

properly accounted for in a vector autoregression framework.

The causal party effect could be convincingly estimated if parties in government

could be randomized over political entities since randomization ensures that there is no

systematic difference between political entities with governments of various stripes. In

that case, the average difference in economic outcomes between entities with different

party control is an unbiased estimate of the true party effect. However, such an

experiment would not be feasible since it would clash with our notion of democracy.

Thus, we are left with drawing inference from non-experimental data. Nevertheless, we

can still try to approximate the evidence generated by a randomized controlled trial,

namely using a quasi-experimental design.

1 For evidence on the party effect for fiscal policies, see Besley and Case (2003) for a survey of work on

U.S. states, and Blais et al. (1993) for a survey of cross-country studies and U.S. states. See also Imbeau et

al (2001) for a meta-analysis of studies using OECD data. For evidence on macroeconomic outcomes, see

Alesina et al. (1997) and the references cited therein.

2 For work that stresses the endogeneity of other political institutions see, for example, Aghion et al. (2004,

2005).

3 I use the word endogeneity as a catchall for problems with selection, omitted variables and simultaneous

causality since all these problems will make the explanatory variable (party control) correlated with the

error term.

3

In this paper, the source of identifying information of the party effect comes from

an institutional feature of the election system, that is, party control changes

discontinuously at 50 percent of the vote share which makes it possible to implement a

*regression discontinuity design*. The general idea of the regression-discontinuity design is

to compare the outcomes for units (e.g., political jurisdictions) whose value of an

underlying targeting variable (e.g., vote share) is “just below” and “just above” a fixed

threshold (e.g., 50 percent of the votes) since they will, on average, have similar

characteristics except for the treatment (e.g., party control). In other words, those units

slightly below the threshold will provide the counterfactual outcome for those units

slightly above, since the treatment status will be “as good as randomly assigned” in a

neighborhood of the treatment threshold. The inference from a regression discontinuity

analysis can therefore be as credible as that from a randomized experiment (e.g., Lee

2003). In particular, the regression discontinuity approach shares the same attractive

feature as a randomized controlled trial, namely that it can actually be tested whether

treatment status is likely to be “as if” randomized.

I employ the regression-discontinuity design on a data set from Swedish local

governments. The use of this data set offers some attractive features in the search for a

causal party effect on economic outcomes. First, it is a large panel data set (288

municipalities over a 21-year period) making it possible to use a regression discontinuity

design since there must be enough data “close” to the treatment threshold for the method

to be useful. Second, Swedish local governments are very homogeneous. In particular,

they operate within a common political framework and face the same institutional setting.

Thus, economic outcomes and political parties are quite comparable across political

entities, which is otherwise a major obstacle in cross-country studies. One potential

weakness with the data set, however, is the multi-party feature of the Swedish political

system. Nevertheless, the Swedish political map has been characterized by a very clear

dividing line between socialist and non-socialist parties leading to a quite stable two-bloc

4

system.4 Hence, to a first approximation we can treat the Swedish electoral system as

bipartisan.5

The results of this paper show that party control has a causal effect on spending,

taxes and unemployment. The party effect is also quite substantial. For example, left-wing

governments spend, as a share of income, about 2-3 percent more and have about 7

percent lower unemployment rates than right-wing governments. Left-wing governments

also employ about 4 percent more workers than right-wing governments. I also present

evidence in support for party control being as “good as randomly assigned” among those

municipalities that are close to the treatment threshold of 50 percent of the vote share,

which provides strong support for a causal interpretation of my results.

This paper is related to, but distinct from, the literature that investigates whether

representatives from different political parties vote differently.6 Specifically, Lee et al.

(2004) make use of a similar regression-discontinuity design in their study of the voting

records of Democratic and Republican congressmen in the U.S. House of Representatives

from 1946 to 1995.7 Although their analysis is interesting, it does not say whether, or to

what extent, parties are of importance for policy outcomes since the mapping between

votes and policy outcomes is not analyzed. For example, many votes in Congress are

supported by large supermajorities. If the difference in voting between Democrats and

Republicans mainly arises in these types of votes, then the effect on policy is nil.

The paper is organized as follows. Section 2 describes the regression discontinuity

design and how it is implemented in this paper. Section 3 describes the data, while

section 4 presents the results. Section 5 discusses the findings and concludes.

4 For an overview of the Swedish political system, see Petersson (1994). For a detailed description of local

governments in Sweden, see Gustafsson (1988).

5 For example, Alesina et al. (1997) also classify Sweden as a bipartisan system (along with the U.S. and

other political systems with a clear left-right division) in their empirical analysis.

6 See, for example, Levitt (1995), Snyder and Groseclose (2000), and McCarty et al. (2001).

7 The first version of this paper was written in May 2001 (Pettersson-Lidbom 2001) while the first version

of the Lee et al. paper is from 2002. My paper is cited in their working paper, Lee et al. (2002), but not in

the published version Lee et al. (2004).

5

**2. Empirical framework**

In this section, I will discuss the regression-discontinuity method and how it is

implemented in this paper.

In the “sharp” regression-discontinuity design, treatment status is a deterministic

function of some underlying continuous variable, that is,

(1)

*Ti*=

*T*(

*xi*) =1[

*xi*?

*x*],

where 1[.] is an indicator function and

*x*is a continuous variable or an assignment

variable, and

*x*is a treatment threshold separating the units into two mutually exclusive

groups: those units receiving treatment (

*T*=1) and those which do not (

*T*=0). The idea is to

compare the outcomes for units whose value of the underlying targeting variable is “just

below” and “just above” the treatment threshold

*x*, since they will on average have

similar characteristics except for the treatment. In other words, those units slightly below

the threshold will provide the counterfactual outcome for those units slightly above, since

the treatment status will be randomized in a neighborhood of treatment threshold. In our

context, the vote share is the assignment variable that assigns parties to political entities

and where the treatment threshold is at 50 percent of the proportion of votes.

In practice, the regression-discontinuity design can be implemented in a number

of ways.8 The simplest possible approach is to just compare average outcomes in a small

neighborhood on either side of the treatment threshold. This approach could, however,

produce very imprecise measures of the treatment effect, since the regression-

discontinuity method is subject to a large degree of sampling variability and this

procedure would therefore require very large sample sizes. An

*equivalent*, but much more

efficient, method is to use all available data and a control function approach, that is, to

regress the outcome of interest, say

*Yi*, on a low-order polynomial in the treatment-

determining covariate

*xi*, i.e., the control function, and the binary treatment indicator

*Ti*.

This procedure will yield an unbiased estimate of the treatment effect, unless the control

function is misspecified, since

*xi*is the

*only*systematic determinant of

*Ti*and therefore the

8 See Hahn et al (2002) for a non-parametric approach.

6

control function will capture

*any*correlation between

*Ti*and the population error term.

The control function approach is my preferred method since there is only a limited

number of observations close to the threshold in my data set (i.e., there are only 89

municipalities within ±2 percentage points from the 50 percent threshold). Nevertheless, I

will also present results where I only use data close to the threshold, i.e., in the range [48,

52], as a specification check since the estimate from the control function approach and the

discontinuity sample should be the same (except for sampling variability) if the control

function is correctly specified.

In this paper, a panel data set from Swedish local governments will be used to

estimate regression models of the form

(2)

*Yit=*µ

*i +*?

*t +*?

*Tit + f*(

*Left vote share*)?

*+ vit*

where

*Yit*is an economic outcome (e.g., spending per capita, taxes, unemployment, and

government employees per capita) for local government

*i*in time period

*t*, µ

*i*is a locality-

fixed effect, ?

*t*is a time-specific effect,

*Tit*is a treatment indicator taking the value of 1 for

left-wing governments and zero for right-wing governments, and

*f*(

*Left vote share*) is a

control function, i.e., some low-order polynomial in

*Left vote share*. The parameter of

interest is ?

*? the party effect ? which measures the average difference in economic*

outcomes between left- and right-wing governments.9 The main reason for including

fixed municipality and time effects is to enhance efficiency since there is no need to

include additional covariates except for

*f*(.) in (2) to get an unbiased estimate of ?.

However, Hoxby (2000) argues that a “within-unit” regression-discontinuity method is

“more powerful and less subject to bias” than a cross-section discontinuity analysis when

there is only a limited number of observations close to the threshold. Thus, specification

(2) takes into account her concern since it only uses the within-municipality variation to

identify the party effect. A number of other controls (e.g., income, population size,

proportion of people below 15, and proportion of people above 65) will also be added to

7

(2) as a way of checking whether party control is as good as randomly assigned. The

inclusion of these additional covariates should not significantly affect the estimate of the

party effect since party control should be as good as randomly assigned conditional on

*f*(.). Here, it is important to

*not*include variables that are themselves affected by the

treatment, such as intermediate outcomes, since these will bias the estimate of the

treatment effect.10 For example, including the lagged economic outcome

*Yit-1*among the

control variables is not advisable in our context of measuring the causal effect of party

control since it is an intermediate outcome,11 and therefore affected by the treatment

itself, i.e., party control. Nevertheless, it is possible to include the economic outcome

from a

*previous*treatment since that guarantees that it is a pretreatment variable, i.e., it

was measured before the

*current*treatment was chosen. Thus, one should only control for

pretreatment characteristics to avoid bias. In practice, however, the covariates are often

recorded at the same time as the outcome, subsequent to treatment. In this case, it must be

assessed on a case-by case basis whether a particular covariate should be used as a control

variable.

A final comment about specification (2) is that it is only the party effect ? that has

a causal interpretation since

*f*(

*.*) is allowed to be correlated with the error term

*vit*. Thus, it

is not valid to interpret the coefficient on vote share ?

*as measuring the causal impact of*

voter preferences on economic outcomes. In other words, in the regression discontinuity

approach, it is totally irrelevant whether the vote share can be considered as a good

measure for voter preferences.

9 The estimated treatment effect from a regression-discontinuity design will typically not be the average

treatment effect but a marginal treatment effect (see, e.g. Hahn et al 2001). This issue will be discussed

below.

10 See Rosenbaum (1984) and Imbens (2004) for a discussion of the choice of covariates.

11 This is related to the term-in-office being longer than one year. The term-in-office in Sweden is three

years.

8

**3. Data**

To test whether party control is of importance for economic outcomes, I will use a panel

data set from Swedish local governments, but before turning to the description of the

data, it might be helpful to briefly digress on the workings of Swedish local governments.

As of 2005, there are 291 local governments (or municipalities) in Sweden which

cover the entire country. Local governments play an important role in the Swedish

economy, both in terms of the allocation of functions among different levels of

government and economic significance. They are, for example, responsible for the

provision of day care, education, care of the elderly and social welfare services. To

quantify their economic importance, note that in the 1980s and 1990s their share of

spending out of GDP was in the range 20 to 25 percent and they employed roughly 20

percent of the total Swedish workforce. Swedish local governments also have the

constitutional right of self-government, no restriction on borrowing and no balanced

budget rules.12 Moreover, only 20 percent of their income come from grants, whereas the

rest mostly comes from a proportional income tax, which each municipality can set freely.

In other words, they have a relatively large degree of fiscal freedom.

To implement the regression-discontinuity method, the mechanics of the Swedish

election system need to be discussed in some detail. The election schedule is fixed and

elections were held every third year on the third Sunday of September during the sample

period.13 During the same period, voter turnout has been very high, close to 90 percent, in

the local elections. The decision-making body in each of the municipalities is an elected

municipal council and the Swedish Elections Act prescribes that in elections to the

municipal council, seats should be proportionally distributed among parties on the basis

of the election results in each constituency, where the distribution is based on the adjusted

odd-number method. As a result, the election system is entirely party based, i.e., a closed-

list system, and has several political parties.14 The multi-party issue raises the question of

12 As from 2000, however, there is a balanced budget rule in place.

13 As from 1994, elections are held every fourth year.

14 Whether the proportional election system is a cause of the multitude of parties or whether the number of

parties is caused by a heterogeneous distribution of voter preferences is still in dispute.

9

how to define treatment or party control. However, as previously discussed, the Swedish

political map has been characterized by a very clear dividing line between socialist and

non-socialist parties leading to a quite stable two-bloc system.15 Hence, to a first

approximation we can treat the Swedish electoral system as bipartisan,16 and define the

treatment indicator

*Ti*as 1 for left-wing majorities and zero otherwise. The party effect

should thus more accurately be addressed as a majority coalition effect but, for simplicity,

I retain the former name.17

There is also one caveat with my data that needs to be mentioned: the existence of

several small parties?often one-issue parties?at the local level which are not part of the

two blocs. These parties sometimes hold the balance of power, which creates a problem

in defining party control since these are not easily classified along the left-right

ideological spectrum. I call these kinds of constellations undefined majorities.18 The

problem with undefined majorities is solved by including a separate dummy variable for

the undefined majority, however. The party effect will now be correctly identified as the

average difference in policy outcomes between left-wing and right-wing majorities.19

Table 1 summarizes the number of left-wing, right-wing and undefined

governments in every election period during the sample period 1974-1994. There was a

left-wing majority in 826 cases, and a right-wing majority in 833 cases. Thus, the two

blocs have been in power almost the same number of times.20 Table 1 also shows that

there has been an undefined majority in 312 cases, which corresponds to 15 % of all

15 For a general overview of the Swedish political system, see Petersson (1994). For a detailed description

of local governments in Sweden during the period of investigation, see Gustafsson (1988).

16 For example, Alesina et al. (1997) also classify Sweden as a bipartisan system (along with the U.S. and

other political systems with a clear left-right division) in their empirical analysis.

17 To define the left-wing majorities and the right-wing majorities, I have relied on the standard

classifications of parties along the left-right spectrum as discussed by Petersson (1994). According to this

classification, the left-wing bloc includes the Social Democratic Party and the Leftist Party while the right-

wing bloc includes five parties: the Conservative Party, the Centrist Party, the Liberal Party, the Christian

Democratic Party and the New Democratic Party. The Christian Democratic Party is only included in the

right-wing majority from the year 1988, however, and the New Democratic Party only from the year 1991.

18 This classification is compiled from the distribution of seats in local councils. If either of the blocs

receives more than 50 percent of the seats it is defined accordingly, otherwise it is classified as undefined.

19 Another approach would be to altogether exclude these observations from the analysis. It turns out that it

is of no importance for the results about the party effect presented below which of these two approaches I

use.

20 This might be surprising given the Social Democratic party hegemony at the national level.

10

observations. Table 2 shows the frequency of government changes for the localities. The

number of government changes is very unequally dispersed among the different

municipalities. For example, 122 municipalities (42 percent of the total sample) had no

change of power (69 had left-wing and 45 right-wing governments). It is important to

stress that the 122 municipalities with zero turnovers will not be part of identifying the

party effect, since only the within-municipality variation will be used, as discussed in

section 2.

Turning to the economic outcomes, nine different variables will be used in the

empirical analysis: total expenditures per capita, total expenditures as a share of income,

current expenditures per capita, current expenditures as a share of income, total revenues

per capita, total revenues as a share of income, proportional income tax rate, the

unemployment rate, and the number of local government employees per capita. The

difference between total and current expenditures per capita is mainly that investments

are included in the former. Roughly 85 percent of total spending is classified as current

spending. Total revenues per capita include tax receipts from a proportional income tax

rate, fees and governmental grants. Since total revenues might reflect non-discretionary

local government decisions, using the income tax rate itself is a more discretionary

measure.21 The unemployment rate is only available from 1979 and therefore I will lose 5

years of data, as compared to the other outcomes, when I use this variable as the

economic outcome of interest. Total expenditures, current expenditures, total revenues

and income are expressed in 1991 prices. Total expenditures as a share of income, current

spending as a share of income, total revenues as share of income, the proportional tax

rate, the unemployment rate and government employment per capita are expressed as

percentages.22 Table 3 presents summary statistics for the nine outcome variables. Table 3

also presents summary statistics for a standard set of controls in the local public finance

literature (see e.g., Besley and Case 2003): average income, proportion of people of age 0

to 15, proportion of people older than 65 and population size. I consider these variables as

not affected by the treatment, which is the key requirement for using them as controls as

21 On average, about 55 % of the total revenues come from the income tax.