The implications of a government shift on labor market outcomes

(1)

The implications of a government shift on labor market outcomes

Measuring the causal effect of government control on unemployment rates in Swedish municipalities

Niclas Östgren

Master thesis I Economics, 15 credits

Spring term 2021

(2)

Abstract

The purpose of this paper is to research if there is a causal effect on labor market outcomes in Swedish municipalities, depending on if there is a left-wing or right-wing rule. By grouping Swedish political parties into two coalitions, left-wing and right-wing, a sharp regression discontinuity design can be applied, where a seat majority in the municipal council shifts discontinuously at 50% of the voter share. Based on the results, there is no estimations that is statistically significant that suggest such a causal effect exists. One should be cautious when drawing conclusions from this paper though, since there is evidence that the estimates could be noisy, and are therefore not certain.

Keywords: Partisan effect, regression discontinuity design, labor market outcomes, Swedish municipalities

(3)

1. Introduction

The purpose of this paper is to research if there is a causal party effect on labor market outcomes in Swedish municipalities, depending on if there is a left-wing or right-wing rule.

To estimate the party effect, this paper employs a regression discontinuity design, henceforth addressed as RDD. When using a RDD to estimate a party effect, there is some rudimental issues that needs to be addressed.

First of, the outcome of municipal elections is endogenously determined. Of course, the winners of a municipal election are not assigned at random, but rather relies on the

preferences of voters, campaign funds, the state of the economy, previous election outcomes, and numerous other characteristics that cannot realistically be measured. If this endogeneity is not controlled for, estimates of a party effect will surely be biased (Lee et al., 2004).

However, by exploiting a natural experiment that occurs in the Swedish election system, the retrieval of election outcomes that were exogenously determined is possible. This is due to the fact that party control in municipal councils shift discontinuously at 50% of the voter share.

The fundamental assumption being made is that municipals where the left-coalition barely wins the election, is directly analogous to municipalities where the right-wing coalition barely won; party control being the only difference. The party effect can therefore be estimated with a RDD, by studying elections where the incumbent party barely wins, i.e., receives votes just above the threshold (Pettersson-Lidbom, 2008). This wis elaborated in section 2 and 4.

This paper will mostly reference the study carried out by Pettersson-Lidbom (2008), and use the same methodology, to research the established question. The data that is used in this paper ranges from 2008 to 2014, as opposed to Pettersson-Lidbom whose data ranges from 1974 to 1994 by, which is the major diversifying quality this paper will poses. This study by

Pettersson-Lidbom will briefly be discussed in section 2.

The results show that there is no significant causal effect on labor market outcomes, regardless of what coalition govern the municipalities. If this a consequence of similarities between left-wing and right-wing labor market policies, limitations in the study, or a

combination of both, is discussed in section 7. Perhaps the most crucial limitation in the study is that there is suspicion that the assignment variable, that determines if a municipality has a left-wing or right-wing rule, is subject to manipulation at the threshold. As mentioned, this

(5)

2

violates the important assumption when using RDD, namely that the party control should be determined exogenously near the threshold. Drawing conclusions based on the results in this paper should therefore be done with caution, and is therefore more suited to serve as a basis for further research and refinement.

The paper is structured in the following way. Section 2 reviews and discusses previous research that utilizes RDD to estimate partisan effects. Section 3 provide necessary

information about the Swedish governing model and political climate, and how parties are grouped into coalitions. Section 4 explains the empirical framework and chosen method.

Section 5 describes the chosen data and how it was processed. Section 6 presents the results.

In section 7, the results are discussed and evaluated. Finally, in section 8, the paper offers some concluding remarks.

(6)

3

2. Literature review and previous research

The seminal paper whose methodology pioneered RDD in econometrics is Lee, Moretti, and Butler’s (2004) study on how voters influenced policies in the US. By examining elections on the US House of representatives, Lee et al. notes how the voting pattern of an individual in the US is endogenous, as political leanings of voters, quality of candidates, campaign funds, incumbent advantages, and other unmeasurable characteristics can determine election

outcomes. By focusing the analysis on electoral races where the incumbent party only wins or loses with a very small margin, the exogenous variation that determine what party holds the seat is isolated. In other words, districts where either a republican or democrat barely won the election are comparable to districts where the other party won. Lee et al. findings imply that voters merely elect policies, e.g., politicians elected rarely compromise their policies.

Lee, Moretti, and Butler’s method has since been used in numerous studies. Fereira and Gyorko (2009) uses a RDD to research whether the mayor of a city in the US have an effect on public spending, crime rates, and city government, finding no effect. Beland (2015) estimates the impact on labor market outcomes in US states, depending on if the governor is either a democrat or republican, finding that states with a democratic governor increases the annual wage and hours worked for black citizens relative to white. A study carried out by Leigh (2007) arrives at the conclusion that on a state level, the party affiliation of the governor only has a small effect on policy outcomes.

In a more related context of Swedish municipalities, Pettersson-Lidbom (2008) estimates the causal effect of party control on fiscal and economic policies. Studying data between 1975- 1994, Pettersson-Lidbom employs a similar method to Lee and Leigh; namely studying municipalities near the 50% threshold of the voter share. Pettersson-Lidbom voices his

concern with the difficulty of measuring a partisan effect when the Swedish model, unlike the US one, employ an electoral system with proportional representation. Nevertheless, because Swedish politics generally have been characterized by a general division into two coalitions of parties, Pettersson-Lidbom argues that the Swedish election system can be treated the same way as a two-party system would. Results suggest that left-wing governments in Swedish municipalities have a higher spending and tax-rates than right-wing governments.

Additionally, left-wing government generally have 7% lower unemployment rate.

(7)

4

Continuing on research on countries with proportional representation; Folke (2014) finds that party representation has a significant effect on immigration policy and environmental policy in Swedish municipalities. Freier and Odendahl (2015) research party effect in close elections in municipalities in Germany, finding that political power does not matter for policy

outcomes. Fiva et al. (2018) study the party effect on policy outcomes in municipalities in Norway, finding that a larger left-wing party representation leads to higher property taxation, childcare spending, and lower elderly care spending.

Beland and Unel (2017) continued the studies on partisan effects, and added to the literature by focusing on a smaller subsection of the economy; namely the labor market for immigrants.

Beland and Unel researched if there is a causal impact on labor market outcomes for immigrants in the US change depending on if there is a democratic or republican governor.

Findings suggested a strong statistically significant relationship between lower unemployment level and higher wages for immigrants in states with a democratic governor. Furthermore, it showed that this relationship differs across immigrants depending on their skill level, and if the labor is condoned in the public or private sector.

However, the validity of using a RDD, to measure election outcomes have also been met with skepticism and been criticized. Eggers, Fowler, Hainmueller, Hall, and Snyder, jr. (2014), mentions that a common critique of the method pioneered by Lee et al., (2004), do not entirely solve for the endogeneity occurring when winners of an election is determined, since incumbent parties tended to be more likely to win the subsequent election. To test the validity of the use of RDD to estimate electoral effects, Eggers et al. researched over 40 000 close elections in the US, finding support that close elections in the US are as good as randomized.

(8)

5

3. Sweden’s municipalities

Sweden has three levels of government. In order of size, these are: the central government, county councils, and municipal councils. Sweden implements a decentralized governing model, meaning that the highest level of government, the central government, delegate power downwards to legislative bodies; in this case counties and municipalities. This gives Swedish municipalities quite a significant amount of autonomy, and is therefore a large part of the economy. Examples of areas that Swedish’s municipalities are responsible for are schooling, elderly care, and social welfare services. Each municipality have their own elections every fourth year, and seats in the municipal council are then distributed proportionally to the political parties. The eight main political parties that are represented in municipal legislature are also the main ones represented on the national level. The political parties are the following (in order of most votes as of national election 2018): The Social Democrats (S), the Moderate party (M), the Swedish Democrats (SD), the Center Party (C), the Left Party (V), the

Christian Democrats (KD), the Liberal Party (L), and the Green Party (MP). On a municipal level, there is also commonplace that smaller local parties receive seats in the council. These parties are addressed as Undefined (U).

Figure 1: Average voter share (%) per party for election years 2006, 2010, 2014, and 2018, on a municipal level

As mentioned by Pettersson-Lidbom (2008), the Swedish electoral system can be treated as a two-party system because of the two coalitions that have formed.A coalition receiving 50%

or more of the voter share would therefore imply a majority of the seats in the municipal council. The determination of what political parties constitutes as a particular coalition, in this

0 5 10 15 20 25 30 35

M C L KD MP S V SD U

(9)

6

paper, is partly based on the definition used by the Swedish Association of Local Authorities and Regions (SKR).¹ Note that MP, according to SKR, is classified as being both left-wing and right-wing. A study carried out by Svaleryd & Vlachos (2009) argues that it is more appropriate to classify MP strictly into the left-wing coalition. ² Since this is a more accurate representation of contemporary Swedish politics, this paper classifies MP solely as a part of the left-wing coalition along with S and V.

An issue that needs to be addressed is that parties are more likely to collaborate with each other to obtain a majority of seats, regardless of political alignment, on a local level than they would on a national level. For example, the 2014 election in the municipality Vimmerby resulted with a rule composed of S and M. As a consequence, there are several rules that cannot be classified as neither left-wing or right-wing. These are called mixed rules.

Table 1, Party control on a municipal level, from the election 2006 to the election 2014

Election year

Number of left- wing majorities

Number of right-wing majorities

Number of mixed rules

Number of other rules

2006 89 135 66 0

2010 108 109 73 0

2014 99 63 127 1

Sum 296 307 266 1

Frequency 34% 35% 30% 0

Note: because of the definition on what constitutes left-wing and right-wing rule, the above table will differ from official published statistics.

An interesting aspect to note from the table is that the amount of mixed rule in municipalities has significantly increased over the election years. This is most likely explained by that the support for Swedish Democrats (SD), both on a municipal and national level, have increased drastically over the election years.³

1 To constitute as a left-wing rule, S and/or V has to be included. To be a right-wing rule, M and/or C, KD, L has to be included. In a mixed rule, one or more of the right-wing parties, in conjunction with one or both of the left- wing parties, has to be included. Local parties, and MP can be included in all rules. Rules where SD is sole majority, rules where SD and/or other parties are included, and rules where one and/or more parties not represented on a national level, are classified as “other rule”.

2 According to Svaleryd & Vlachos (2009), voters of MP generally self-align as left-wing, MP is four times more likely to support a left-wing coalition rather than a right-wing on a municipal level, and has been part of the left- wing coalition that supported the S lead government 1998 and 2002.

3 In the Swedish national election 2006, SD only received 2,93% of the votes. On the national election 2018, they received 17,53% of the votes, becoming Sweden’s third largest party.

(10)

7

4. Methodology

To estimate the party effect on labor market outcomes, this paper employs a “sharp” RDD.

RDD is a form of quasi-experiment where the treatment status depends in whole, or in part on, whether a specific continuous or assignment variable exceeds an arbitrary threshold. In a

“sharp” RDD, the treatment status entirely depends on if the specific variable exceeds the chosen threshold (Stock & Watson, p. 546. 2015). In the context of this paper, if the left-wing coalition receives 50% of the voter share or more, control shifts discontinuously at 50% of the voter share from the right-wing coalition.

Formally, the treatment status is a deterministic function written as:

𝑇_𝑖 = 𝑇(𝑙𝑒𝑓𝑡 𝑣𝑜𝑡𝑒 𝑠ℎ𝑎𝑟𝑒_𝑖) = 1[𝑙𝑒𝑓𝑡 𝑣𝑜𝑡𝑒 𝑠ℎ𝑎𝑟𝑒_𝑖 ≥ 50] (1) Left vote share is the assignment variable that determine what coalition have a majority of seats in the municipal council, and the treatment threshold is 50 % of the voter share.

“Treatment status” in this case is if the left-wing coalition defined won the election. If left vote share exceeds the threshold, T will assume a value of 1. If it does not exceed the

threshold, T will assume a value of 0. Very important to note is that that a treatment status of 0 under these assumptions does not imply a right-wing rule, since it also could be a mixed rule or other rule. Pettersson-Lidbom (2008) addresses this problem by including a separate dummy variable to control for rules that are neither left-wing or right-wing. However, Pettersson-Lidbom also notes that an equally viable approach is to simply omit these observations when conducting the analysis. For simplicity, this paper will choose the latter approach and omit mixed rules and other rules. Treatment can therefore be binarily defined as either left-wing or right-wing.

As discussed, winners of an election are also endogenously determined (Lee et al. 2004). This endogeneity could then be solved by mainly studying close elections near the threshold, as the variation in winners can be seen as quasi-randomly selected. These observations should then therefore have similar characteristics, on average, except the election outcome. However, there is an alternative method that can be used to estimate the treatment effect, which is a control function approach. Instead of only looking at data close to the threshold, the whole data set is used, and then regressing a labor market outcome of interest 𝑌_𝑖 on some polynomial in the treatment-determining covariate, the left-wing vote share.

(11)

8

If the polynomials correctly specify the regression function, the estimates of the treatment effect should yield unbiased estimates, even though the data far away from the threshold is used. This is because the left-wing voter share is the sole systematic determinant of the treatment effect 𝑇_𝑖 , and the control function, if correctly specified, will capture any correlation between 𝑇_𝑖 and the error term (Pettersson-Lidbom, 2008).⁴

These two methods are known as choosing a non-parametric, or parametric estimation approach. Although one method may be preferred over the other to eliminate bias, according to Lee & Lemieux (2010), the two methods essentially compute the same statistic, but with different standard errors. Result from using both of these methods will be presented in this paper, which is beneficial for two reasons. The first reason is that one can see which method is more efficient, i.e., yield lower standard errors. The second reason is that it can serve as a specification check. If the control function is correctly specified, the estimated party effect should not differ between the two methods (Pettersson-Lidbom, 2008).

The regression model, based on Pettersson-Lidbom (2008), estimated in the parametric approach using some polynomial in left-wing voter share:

𝑌_𝑖𝑡 = ц_𝑖 + 𝜆_𝑖 + β𝑇_𝑖𝑡+ 𝑓(𝑙𝑒𝑓𝑡 𝑣𝑜𝑡𝑒 𝑠ℎ𝑎𝑟𝑒)ф + 𝑋_𝑖𝑡+ 𝑢_𝑖𝑡 (2) Where 𝑌_𝑖𝑡 isthe labor market outcome of interest for municipality 𝑖 in time period 𝑡. ц_𝑖 are fixed effects; controlling for characteristics that may vary across municipalities, but not over time. 𝜆_𝑖are time fixed effects; controlling for characteristics that may vary over time, but not over municipalities. The parameter of importance is β, the party effect, which will measure the average difference on labor market outcomes, depending on if there is a left-wing or right- wing rule. 𝑇_𝑖𝑡is the treatment indicator taking the value of one for left-wing rules and zero for right-wing rules. The control function 𝑓(left vote share) is a polynomial of left vote share, and ф is the coefficient of vote share. 𝑋_𝑖𝑡 is a set of covariates, and 𝑢_𝑖𝑡 is the error term. Since it is assumed that treatment status in close elections is as good as randomized, β should be

independent of any covariates 𝑋_𝑖𝑡. It may therefore seem unnecessary to include covariates;

however, one may reduce sampling variation, and also use them to test the validity of the RDD (Lee & Lemieux, 2010). This will be elaborated in section 6. As a result of the reason stated above, only the coefficient for β is of interest and has causal interpretation.

4 Most of the previous studies listed have used a parametric approach (see e.g., Beland, 2005; Fereira & Gyorko, 2009; Folke 2004; Pettersson-Lidbom, 2008).

(12)

9

When using the parametric approach, it is reasonable to assume that the regression function is nonlinear; the purpose of including polynomials in the control function is to correctly specify it. To determine what degree of polynomial one should use is a trade-off between flexibility and precision when estimating coefficients. Adding a high-degree polynomial gives the regression function more flexibility to match the shape of the regression function. However, a high-degree polynomial also means increasing the amount of regressor, which will reduce the precision of the treatment effect (Stock & Watson, p. 313. 2015). Pettersson-Lidbom (2008) uses a first-degree polynomial up to fourth in the left-wing vote share, which this paper adopts.⁵

When using the non-parametric approach, in line with methods previously discussed, only elections that are close to the threshold will be used. This is known as choosing a bandwidth, and there is intuitive pros and cons depending on the size of the bandwidth. Choosing a bandwidth closer to the threshold implies more randomized election outcome, but more omitted observations. A larger bandwidth would of course have the opposite effect. There is no formal way to determine the optimal bandwidth, but nonetheless, it is recommended to include different ones to test how sensitive they are to the polynomials (Lee & Lemieux, 2010). Hence, election outcomes within ±2 and ±4 percentage points from the threshold will be studied in this paper.⁶ The regression model will take on the same form as equation (2), but without the control function and covariates.

5 Important to note is that it has since been proven by Gelman & Zelizer (2015) that controlling for higher-order polynomials in regression discontinuity analysis could lead to noisy causal inference.

6 Choosing ±2 of the voter share is in line with Pettersson-Lidbom. Choosing ±4 of the voter share is done in the purpose of gaining a larger sample of observations.

(13)

10

5. Data

This paper uses panel data composing of all 290 Swedish municipalities, sampled from 2006 to 2018. The data is mainly retrieved from the Swedish Central Bureau of statistics (SCB).

From year 2006 to 2014, there have been four municipal elections. Because elections are held late in the year in September, the election period is lagged by one year, as to more accurately represent results dependent on the ruling coalition. An election period will therefore range from 2007-2010, 2011-2014, and 2015-2018. The outcome variables and covariates are measured as the average value in each corresponding time period. Since some elections were omitted, the panel is unbalanced.

The data set is composed of labor market outcome variables, covariates, and the assignment variable. The assignment variable is the left vote share. The labor market outcomes that will be studied is unemployment rate, the public employment rate, and the private employment rate. Covariates will include population size, proportion of young inhabitants, proportion of elderly inhabitants, and income per capita.⁷. For summary statistics, see table 3-8 in the Appendix.

5.1 Outcome variables

The variable unemployment rate is the unemployment rate in municipalities measured in percent. Unemployment rate on a municipal level is not available on SCB, and the data has instead been retrieved from “The Swedish public employment service”.⁸ Important to note is that the definition of unemployment vary between these two entities, which can cause biased estimations. This will be elaborated in the data limitations later in the section. The variable public employment rate is the share of workers in a municipality that are employed by municipal owned firms and organizations, as well as municipal administrations. The variable private employment rate is the share of workers in a municipality that are employed by private businesses. Both of these variables are measured as nighttime population. Nighttime

population refers to individuals that inhabits the municipality and works, regardless of where they are employed.

7 Note that these are the same covariates used by Pettersson-Lidbom (2008), but with a few exceptions. Because of limited available data, the covariate for proportion of young inhabitants is measured in the age group 0-19, instead of 0-

8 Note that in the references section, the source for this variable is listed as SCB. This is because SCB sources the Swedish Public employment service.

(14)

11 5.2 Covariates

The variable population represents the number of inhabitants in a municipality. The variable income per capita is the average income earned per inhabitant in a municipality aged 16+, expressed in SEK. The variable proportion of young is the share of the population in a municipality that is aged between 0-19 years. The variable proportion of elderly is the share of population in a municipality that is aged 65 years or older. The variable annual income earnings measure the average annual income earned by services and labor, in a municipality.

As mentioned before, the inclusion of the covariates is intended to reduce sampling variation, and serve as a specification check.

5.3 Assignment variable

The assignment variable is based on data from municipal elections in 2006, 2010, and 2014, retrieved from the Swedish Election Authority. The voter share for S, V, and MP, is pooled into a single result, representing left-wing voter share, municipal rule can therefore now be binarily determined. In total, there have been 870 elections in that timeframe. After omitting elections that yielded mixed rules and other rules, the amount of election decreases to 604.

381 of those elections yielded a right-wing government, and 223 a left-wing government.

Within the range of [48, 52] of the left-wing voter share, there is 62 observations, where 29 elections yielded are right-wing rule and 33 a left-wing rule. Within the range of [46, 54] of the left-wing vote share, there is 126 observations, where 63 elections yielded a right-wing rule, and 63 yielded a left-wing rule.

5.4 Data limitations

The variable that describes unemployment rate in municipalities is based on data retrieved from the Swedish Public Employment Service, an authority whose purpose is to assist jobseekers to find employment. SCB defines unemployment as people in the labor force, but without employment, and collect their data by large randomized sample surveys. In

comparison, the statistics for unemployment rate is solely based on the individuals registered as unemployed in the database of the Swedish public unemployment service (The Swedish Public Employment Service, 2020). In other words, individuals can be unemployed, but if they are not using the service, they are not accounted for in the unemployment statistic.

Naturally, the definition of unemployment by SCB is the superior version to collect data, since it is not biased in its collection of data. However, since this type of data for

unemployment rate in municipalities is not kept by SCB, this paper has to make do with the data from the Swedish Public Employment Service. With the established requirements for what constitutes a left-wing or right-wing rule, there is also large number of elections that

(15)

12

yielded a mixed rule, and was therefore omitted from the research. When narrowing the eligible observations by imposing a bandwidth, a significant amount of observations are also lost. A larger sample with more election periods would therefore be ideal, since it most likely also would lead to a larger sample size close to the threshold.

(16)

13

6. Results

As previously mentioned, two implementations of sharp RDD are utilized. The first

implementation is a parametric approach, using all available data and a control function. A first-order polynomial up to fourth-order polynomial in the control function is used, as to see which one will yield the best fit, in line with Pettersson-Lidbom (2008). For a specification check, an additional regression including the selected covariates, with a fourth-order

polynomial in the control function, is done. As previously established, the inclusion of these control variables, if the control function is specified correctly, should not yield a different estimation of the treatment, but only reduce the standard errors.

The second implementation is the parametric approach, imposing a bandwidth when

estimating the data, and therefore only study observations close to the established threshold using linear regression. Like previously discussed, the sample close to the threshold should be exogenously determined, and the inclusion of covariates is therefore unnecessary. If the control function is correctly specified, the difference in the estimated coefficients between the two methods should be very small. The result from these estimations is presented in table 2 bellow.

Table 2: estimated causal effect of left-wing and right-wing rule

Outcomes 1 2 3 4 5 6 7

Log Unemployment rate (%)

0.002 (0.028)

0.000 (0.028)

-0.012 (0.031)

-0.014 (0.030)

-0.012 (0.054)

-0.004 (0.033)

-0.015 (0.029) Log Public employment

rate

0.003 (0.010)

0.005 (0.010)

-0.006 (0.011)

-0.005 (0.011)

0.015 (0.018)

-0.004 (0.012)

-0.005 (0.009) Log Private employment

rate

0.002 (0.005)

0.001 (0.005)

0.003 (0.005)

-0.006 (0.006)

0.001 (0.006)

0.003 (0.004)

Sample Full

604

Full 604

±2 62

±4 126

Full 604 Left vote share

polynomial First Second Third Fourth None None Fourth

Control variables No No No No No No Yes

Note: Robust standard errors are expressed in parentheses. ***p<0.01, **p<0.05, *p<0.1

(17)

14

Column 1-4 uses the entire dataset, the control function with a first-degree polynomial up to fourth, and no additional covariates. Column 5-6 uses data close to the threshold, no

polynomial of the control function, and no additional covariate. Column 7 is a specification check, as discussed above. All labor market outcomes are in logarithmic form, making the interpretation of the coefficient in percent. Each entry is a separate regression, and uses robust standard errors to control for heteroskedacticity. Municipal specific and time fixed effects are also included in all regression, but not reported. Observing the results, none of the party effect estimates are significant on even a 10% level.

Observe that standard errors generally are higher, when using a local linear regression in a neighborhood of the threshold, as in column 5, compared to when using a control function as in column 1-4. Implications of this is that the control function approach is the more efficient of the two methods, if only very small. However, when the sample include ±4% of the 50%

threshold in column 7, the standard errors are strictly smaller, albeit very slight, than the standard errors in column 1-4.

Regarding the specification check in column 7, when including covariates when regressing with a fourth-degree polynomial in the control function, the change in estimates compared to column 4 is only negligible. Implications of that result is that the winner of municipal

elections are randomly assigned, which as discussed. is a key assumption in RDD. The estimated effect of treatment differs slightly when comparing the two different methods with each other. Namely, column 5-6 compared to column 1-4. As previously mentioned, this could suggest that the control function is not specified correctly, since column 1-4 and 5-6 should yield the same estimates when compared to each other, and only differ in standard errors. However, Petterson-Lidbom (2008) argues that the control function can still be

specified correctly, given that the difference in estimations is very small. It can also be argued that column 6 uses data too far away from the threshold to yield the same estimates as column 1-4.

(18)

15

7. Discussion

According to the results, no significant party effect on labor market outcomes in Swedish municipalities can be supported.

A possible explanation for the nonsignificant estimates could be the median voter theorem.

Fiva et al. (2018) discuss the importance of political convergence when measuring partisan effects. This theorem states that when there is a conflict of interest when deciding policies in a dimension, the median party will be decisive when choosing the policy. Hence, party policies tend to convergence, to appease the preferences of the median voter (Fiva et al, 2018). In the context of this paper, if the median voter theorem is applicable to the labor market the party effect for labor market outcomes can potentially be diminished.

The next question one should ask themselves is why the results in this paper differ in estimated treatment effect and level of significance from the one presented by Pettersson- Lidbom (2008). Pettersson-Lidbom findings suggest that municipals with a left-wing majority are associated with a lower unemployment rate. On average, -7% in unemployment rate, statistically significant on a 5% level when using a third and fourth-degree polynomial in left voter share. Pettersson-Lidbom also include a variable for government employees per capita, which is analogous to the variable public employment in this paper. Pettersson-Lidbom estimates the party effect to being an average of 3.0%-3.7% higher for municipalities with a left-wing majority. Regarding the variable for private employees, Pettersson-Lidbom does not include it in his study, and a comparison of results can therefore not be made.⁹ Though that a party effect does not exist for private employment rate is perhaps intuitive, since private employment is outside municipalities area of responsibility.

This large difference in results can partially be explained by two different reasons. Firstly, the assignment variable may perhaps not be determined randomly. As discussed previously, a key assumption of RDD is that the assignment variable is free of manipulation and treatment close to the threshold is therefore random. In this case, winners of a municipal election. In a pure two-party system, most notable one being the US, parties cannot form a coalition to win an

9 However, the results for unemployment rate and public employment rate in this paper is similar to the results in a study by Lakomaa and Korpi (2014), who did not find evidence that party control have an effect on neither of the variables. Lakomma and Korpi used the same dataset as Pettersson-Lidbom, albeit with a different definition of what party compositions constitutes as left-wing and right-wing.

(19)

16

election. However, in the Swedish context with proportional representation, this is a norm. A possible explanation for this misspecification of the assignment variable could therefore be left-wing parties can be more accommodative to form rules with other left-wing parties, if it would result in gaining a majority of in the municipal council. For example, the outcome of the municipal election 2014 in Lidköping yielded a left-wing rule composed of S, V, and MP.

S received 40,4% of the votes, V received 6,4% of the votes, and MP received 5,1% of the votes. Any combination of these three parties, that would not include all parties, would not be able to receive a share of votes above the 50% threshold. Hence, all three parties collaborated to obtain majority. This kind of manipulation, to receive a voter share above the threshold, is not accounted for in this paper, this highlights a weakness with the paper’s a method of treating a multiparty system as bipartisan. Looking at figure 2 and 3 in the Appendix, one can observe a slight “jump” in observations just at the right-hand side of the threshold, supporting the suspicion that the assignment variable may be manipulated near the threshold.

Secondly, the data used in this paper is more recent than the one used by Pettersson-Lidbom, and is processed differently. Most notably, this paper group the party MP exclusively as belonging to left-wing, while Pettersson-Lidbom does not include it. It is therefore a

possibility that categorizing MP exclusively as a left-wing party leads to different estimations.

This paper also process data in election periods differently. For example, for the election period 2007-2010, the observations in 2007, 2008, 2009, and 2010 is computed as a single mean for the election period. The same approach is of course done for election period 2011 and 2015. Pettersson-Lidbom uses a different approach, and uses each yearly value in the election periods, effectively quadrupling the amount of observations. The variable for unemployment rate is also based on data retrieved from the Swedish public employment service which, as discussed in section 5, is not an accurate representation of the true

unemployment rate. This data may not necessarily be the same used by Pettersson-Lidbom.

However, Pettersson-Lidbom does not elaborate on how his variable for unemployment rate was computed, and it is therefore uncertain whether the different results are a cause of different definitions of unemployment rate. A similar argument can be made for Pettersson- Lidbom’s variable for government employees per capita, since it is not certain how the data for that variable was retrieved either.

(20)

17

8. Conclusion

The purpose of this paper was to study if there was a causal party effect on labor market outcomes in Swedish municipalities, depending if a left-wing coalition or right-wing coalition had a majority of seats in the municipal council.

Based on the results, there is no estimations that is statistically significant which suggest that such a causal effect exists, implying that the color of the coalition that wins a majority in a municipal election do not affect the labor market outcomes chosen in any significant way.

This could imply that mean that left-wing and right-wing coalitions in Swedish are similar when determining labor market policies, similarly to the phenomenon explained by the median voter theorem. Though this may of course not necessarily be the sole explanation for the different estimates, when compared to Pettersson-Lidbom (2008), as discussed in section 7. The conclusion of this paper is therefore that there exists no party effect on labor market outcomes in Swedish municipalities, but it cannot be determined with certainty.

A suggestion when conducting further research in the field, or to refine the study in this paper, is to adopt a more careful approach when assessing manipulation of the assignment variable.

Since it is such a fundamental assumption, this cannot be stressed enough. Another suggestion is to process the data in election periods both as a single mean, as well as separate

observations, then do the established regressions on both datasets. The dataset which yields more significant estimates should then be the more efficient method.

(21)

18

9. References

9.1 Written sources

Beland, L. P. (2015). Political parties and labor-market outcomes: Evidence from us states. American Economic Journal: Applied Economics, 7(4), 198-220.

Beland, L. P., & Unel, B. (2018). The impact of party affiliation of US governors on immigrants’ labor market outcomes. Journal of Population Economics, 31(2), 627-670.

Eggers, A. C., Fowler, A., Hainmueller, J., Hall, A. B., & Snyder Jr, J. M. (2015). On the validity of the regression discontinuity design for estimating electoral effects: New evidence from over 40,000 close races. American Journal of Political Science, 59(1), 259-274.

Ferreira, F., & Gyourko, J. (2009). Do political parties matter? Evidence from US cities. The Quarterly journal of economics, 124(1), 399-422.

Fiva, J. H., Folke, O., & Sørensen, R. J. (2018). The power of parties: evidence from close municipal elections in Norway. The Scandinavian Journal of Economics, 120(1), 3-30.

Folke, O. (2014). Shades of brown and green: party effects in proportional election systems. Journal of the European Economic Association, 12(5), 1361-1395.

Freier, Ronny, and Christian Odendahl. "Do parties matter? Estimating the effect of political power in multi-party systems." European Economic Review 80 (2015): 310-328

Gelman, A., & Zelizer, A. (2015). Evidence on the deleterious impact of sustained use of polynomial regression on causal inference. Research & Politics, 2(1), 2053168015569830.

Lakomaa, E., & Korpi, M. Bloc-party Politics and Economic Outcomes. What Are the Effects of Local Parties?

Lee, D., Butler, M., & Moretti, E. (2004). Do voters affect or elect policies? Quarterly Journal of Economics, 119 (3), 807–859.

(22)

19

Lee, D., & Lemieux, T. (2010). Regression Discontinuity designs in economics. Journal of Economic Literature, 48 (2), 281–355.

Leigh, A. (2008). Estimating the impact of gubernatorial partisanship on policy settings and economic outcomes: A regression discontinuity approach. European Journal of Political Economy, 24(1), 256-268.

Pettersson-Lidbom, P. (2008). Do parties matter for economic outcomes? A regression- discontinuity approach. Journal of the European Economic Association, 6(5), 1037-1056.

Svaleryd, H., & Vlachos, J. (2009). Political rents in a non-corrupt democracy. Journal of Public Economics, 93(3-4), 355-372.

(23)

20 9.2 Data sources

Statistics Sweden. (2021). Andel öppet arbetslösa i befolkningen, procent efter region, kön, utbildningsnivå, bakgrundsvariabel och år.

https://www.statistikdatabasen.scb.se/pxweb/sv/ssd/START__AA__AA0003__AA0003B/Int Gr1KomKonUtb/table/tableViewLayout16/# (Accessed on April 5)

Statistics Sweden. (2021). Folkmängden efter region, civilstånd, ålder och kön. År 1968–

2020.

https://www.statistikdatabasen.scb.se/pxweb/sv/ssd/START__BE__BE0101__BE0101A/Bef olkningNy/?rxid=f5790986-ce9d-4cdc-9658-fbc9c700b7ef# (Accessed on April 5)

Statistics Sweden. (2021). Registerbaserad arbetsmarknadsstatistik (RAMS).

https://www.statistikdatabasen.scb.se/pxweb/sv/ssd/START__AM__AM0207__AM0207J/Na ttSNI07FoR/ (Accessed on April 12)

Statistics Sweden. (2021). Sammanräknad förvärvsinkomst för boende i Sverige hela året efter region, kön, ålder och inkomstklass. År 1999 – 2019.

https://www.statistikdatabasen.scb.se/pxweb/sv/ssd/START__HE__HE0110__HE0110A/Sa mForvInk1/?rxid=2578c27d-2719-4271-842d-0d8a7a87c78a# (Accessed on April 12)

Swedish Association of Local Authorities and Regions. (2021). Styre i kommuner efter valet 2018.

https://skr.se/skr/demokratiledningstyrning/valmaktfordelning/valresultatstyren/styreikommun ereftervalet2018.26791.html (Accessed on 22 Mars 2021)

The Swedish Public Employment Service. (2020). Därför är SCB:s och Arbetsförmedlingens siffror så olika.

https://arbetsformedlingen.se/om-oss/press/nyheter/nyhetsarkiv/2020-07-24-darfor-ar-scbs- och-arbetsformedlingens-siffror-sa-olika (Accessed on 24 May 2021)

The Swedish Election Authority. (2020). Valresultat 2018.

https://www.val.se/valresultat/riksdag-region-och-kommun/2018/valresultat.html (Accessed on 23 Mars 2021)

(24)

21

Appendix

Figure 2: Distribution of left-wing voter share in Swedish municipalelections (mixed and other rules omitted)

Figure 3: Distribution of left-wing voter share in close Swedish municipal elections. [48, 52].

Table 3: Summary statistics, all elections

(25)

22 Variables

Number of

observations Mean

Standard

deviation Minimum Maximum Labor market outcome

Unemployment rate (%) 604 12.183 3.197 4.9 21.725

Public employment rate (%) 604 25.662 5.394 7.775 41.8

Private employment rate (%) 604 66.674 5.730 50.35 81.425

Assignment variable

Left vote share 604 45.854 11.809 11.6 82.6

Control variables

Population size 604 35107.94 78333.19 2434.75 942762.5

Proportion of young, 0-19 604 22.920 2.640 16.7 31.3

Proportion of elderly, 65+ 604 22.018 l4.249 11.025 33.75

Income per capita 604 245.932 41.175 183.65 515.025

Note: Income per capita is expressed in 1000s of SEK.

Table 4: Summary statistics, close elections [48, 52]

Variables

Number of

observations Mean

Standard

deviation Minimum Maximum Labor market outcome

Unemployment rate (%) 62 13.932 2.626 8.95 18.525

Public employment rate (%) 62 26.345 4.355 16.525 37.95

Private employment rate (%) 62 65.761 4.902 52.225 76.075

Assignment variable

Left vote share 62 50.000 1.138 48 52

Control variables

Population size 62 53101.42 85615.28 2755.75 503695

Proportion of elderly, 65+ 62 21.589 4.397 12.2 28.75

(26)

23

Table 5: summary statistics, close elections [46, 54]

Variables

Number of

observations Mean

Standard

deviation Minimum Maximum Labor market outcomes

Unemployment rate 126 13.571 2.690 7.75 21.675

Public employment rate 126 26.697 4.370 12.3 39.425

Private employment rate 126 65.352 5.279 50.65 80.725

Assignment variable

Left vote share 126 49.905 2.434 46 54

Control variables

Population size 126 48093 81625.68 2755.75 530219.8

Proportion of elderly, 65+ 126 21.870 3.889 12.2 28.75

Note: Note: Income per capita is expressed in 1000s of SEK.

Table 6: All elections, difference in means Variables

Left-wing means (1)

Right-wing means (2)

Difference in means (1) – (2) Labor market outcome

Unemployment rate (%) 14.337 10.923 3.414

Public employment rate (%) 28.385 24.061 4.324

Private employment rate (%) 64.155 68.148 -3.993

Control variables

Population size 25905.43 40494.19 -14588.76

Proportion of young, 0-19 21.805 23.573 -1.768

Proportion of elderly, 65+ 23.511 21.144 2.367

Income per capita 233.955 252.941 -18.986

(27)

24

Table 7: close elections [48, 52], differences in means Variables

Left-wing means (1)

Unemployment rate (%) 13.745 14.101 -0.356

Control variables

Population size 51497.84 52393.39 -895.55

Proportion of young, 0-19 22.881 22.396 0.485

Proportion of elderly, 65+ 21.209 21.944 -0.735

Income per capita 237.275 236.580 0.695

Table 8: close elections [46, 54], differences in means Variables

Left-wing means (1)

Unemployment rate (%) 13.940 13.201 0.739

Control variables

Population size 39800.31 56385.89 -16585.58

Proportion of young, 0-19 22.556 22.719 -0.163

Proportion of elderly, 65+ 22.191 21.550 0.641

Income per capita 235.688 237.986 -2.298

The implications of a government shift on labor market outcomes