Do political parties matter? Evidence from German municipalities

1 Introduction

Do political parties matter? That is, do they offer different policy platforms to voters and do they, once elected, implement different policies? Seminal models by Hotelling (1929) and Downs (1957) reject that notion and suggest that parties move to the center of the policy space and capture the median voter. Later work highlights that the median voter theorem relies on a set of non-trivial assumptions and shows that policy divergence may emerge under alternative assumptions (e. g. Wittman (1983) Calvert (1985), Cox and McCubbins (1986), Glaeser et al. (2005)). If politicians, for example, do not only care about being elected – as assumed in Downsian models – but also about implemented policies, parties cannot offer credible platforms that deviate from their representatives’ underlying policy preferences (Alesina (1988)).^[1]

In this study, we empirically test for partisan effects on policy outcomes. The paper adds to the flourishing empirical literature on the topic 1) by assessing partisan effects on the local level (where existing work is still limited and result patterns tend to vary across studies, see below); 2) by relying on a rich dataset on local public spending that allows testing for partisan effects in detailed spending areas; 3) by proposing a new empirical strategy to identify partisan effects on the local level. Our findings point to strong partisan effects: Communities with a left-wing council majority tend to spend more on ‘people-oriented’ public goods and less on infrastructure than communities with a right-wing dominated council.

Our testing ground are German municipalities. In the main analysis, we assess whether the partisanship of local council majorities impacts on the level and composition of local public spending by German municipalities. For that purpose, we exploit detailed data on municipality spending across different spending categories between 1994 and 2006, including e. g. spending for child and youth care, culture and infrastructure. The data is linked to information on the party composition of local councils, which are the local legislative bodies that decide on local spending policies and taxes. To avoid that we confound partisan effects with differences between councils with absolute majorities and coalition majorities (e. g. reflecting common pool problems in council coalitions, see e. g. Persson et al. (2007), Meriläinen (2019)), we focus in the analysis on jurisdictions with a change in the party with the absolute majority in the local council.

Our empirical analysis is a fixed effect approach that compares changes in local spending policies when council majorities change from the main left-wing party SPD to the main right-wing party CDU/CSU or vice versa. To address potential selection into treatment, we rely on an instrumental variable (IV) strategy that exploits variation in party support at the federal level for empirical identification and draws on data from monthly opinion polls on intended voting behavior in federal elections (‘If next Sunday were to be a federal election day for the German national parliament, which party would you vote for?’). There are different potential shifters of intended voting behavior at the federal level, including changes in the perceived competence and popularity of party representatives at the national level (e. g., members of the national parliament or the national government) or perceived changes in their political positions. Exogenous events may also alter federal policy preferences; the ‘Fridays for Future’-protests, e. g., were associated with an increase in the support for the green party in Germany. Less than perfectly informed voters may rely on signals on party performance and party positions at the federal level – which are often more salient due to better media coverage – to proxy for expected party performance and positions at the local level; changes in federal party support may, therefore, trigger changes in local party support. Our empirical identification strategy exploits that in some jurisdictions these federal trends in party support imply that parties gain the majority of council seats, in others not. Common trends, like shifts in the general policy preferences of the electorate (that affect both, policy decisions at the federal and the local level irrespective of the party affiliation of decision makers) or common socio-economic shocks – are absorbed by (state-)year fixed effects in the estimation approach.

We, moreover, couple the instrumental variable strategy with an entropy matching approach (e. g. Hainmueller (2012)) to acknowledge that identification partly relies on differences in the initial vote shares for parties in council elections at the outset of our sample period. Specifically, we match localities with comparable spending patterns in the pre-election period and, on top of that, augment the empirical approach by a rich set of municipal-level control variables to ensure unbiased results.^[2]

Our findings point to significant partisan effects. Specifically, a council seat majority of the main left-wing party SPD is associated with more spending for people-oriented public goods and less spending for infrastructure public goods relative to councils dominated by the main conservative party CDU/CSU. Quantitatively, the share of spending assigned to ‘people goods’ increases by 3.6 percentage points and the spending assigned to infrastructure goods drops by 7.2 percentage points respectively, which is quantitatively significant. Our results, in turn, reject partisan effects on tax setting behavior. In addition, we present a sensitivity analysis, where we use a RD design to assess whether the partisanship of municipalities’ mayors impacts local spending choices. Institutionally, mayors can influence policy outcomes by drafting and proposing legislation (which then has to be enacted by a simple majority of votes in the local council). Our RD design assesses whether spending policies systematically differ between municipalities where SPD mayoral candidates marginally won or marginally lost an election. The findings are broadly consistent with our baseline results: SPD mayors are found to spend more on ‘people-oriented’ goods and less on infrastructure.

Our paper contributes to a flourishing empirical literature on partisan effects. The more recent literature predominantly uses regression discontinuity designs for empirical identification (e. g. Lee et al. (2004), Ferreira and Gyourko (2009), Albouy (2013) and Curto-Grau et al. (2018)).^[3] Our paper departs from this literature by using a different identification strategy, which arguably identifies partisan effects less locally than RD designs.^[4]

Most of the existing work, moreover, tests for partisan effects on the national or state level (see e. g. Besley and Case (2003), Lee et al. (2004), Potrafke (2011a) and Potrafke (2011b)). Compared to this well-developed literature, analyses on the local level are still relatively limited – despite the fact that insights from higher federal tiers may not carry over to the local level, where interjurisdictional mobility is high and Tiebout sorting into homogenous local units or competition for mobile tax bases may limit the scope for partisan politics (e. g. Ferreira and Gyourko (2009)).^[5] Existing work for the local level, moreover, largely focuses on the identification of partisan effects for policy outcomes other than government spending (see e. g. Sole-Olle and Viladecans-Marsal (2013), Folke (2014), Freier and Odendahl (2015)). Among those papers that consider public spending, most use total spending or broad spending categories (Pettersson-Lidbom (2008) and Ferreira and Gyourko (2009)) so that fine grained compositional spending changes are not observed. Our paper, in turn, is similar to Fiva et al. (2018) in that we can draw on detailed information on the level and composition of government expenditures. Our results show that the use of fine-grained spending data is essential: while we, similar to Ferreira and Gyourko (2009) and in contrast to Pettersson-Lidbom (2008), do not observe partisan effects in overall government spending, our findings point to partisan effects in detailed spending sub-categories, namely spending for social services and streets.

Most existing studies for the local level, moreover, either assess effects of mayoral partisanship on policy outcomes in the context of a majoritarian voting system (e. g. Ferreira and Gyourko (2009) and Gerber and Hopkins (2011)) or focus on the effect of political power of parties in the local council in the context of a proportional representation system (e. g. Folke (2014), Freier and Odendahl (2015), Fiva et al. (2018)). Our main analysis determines the effect of changing seat majorities for the main left-wing and right-wing party in German local councils, in turn. Note, moreover, that estimates for partisan effects at the local level display considerable heterogeneity. Ferreira and Gyourko (2009), for example, reject quantitatively significant partisan effects at the local level, while other work – including this paper – does find significant effects. This may reflect differences in the aggregation level of employed data (see above), but also differences in institutional contexts and testing grounds.^[6]

The remainder of the article is as follows. Section 2 describes the institutional background and the data. The methodology is outlined in Section 3 and Section 4 presents the results. Finally, Section 5 concludes.

2 Institutional background and data

Our empirical analysis assesses the impact of partisanship of local council majorities on overall local spending and the composition of municipal spending using West German localities as a testing ground.^[7] In the following, we describe the institutional background and the data set used for the empirical analysis.

3 Methodology

We draw on an instrumental variable strategy to assess the impact of partisanship of local council majorities on municipal spending choices. The second stage model reads

where yit depicts variables for the size and structure of local spending as described in the previous section. Our main explanatory variable Pit is an indicator that is one if the main left-wing party SPD holds the majority (more than 50 % of the seat shares) in the local council. The model includes municipality fixed effects λi and state-year fixed effects κst, hence absorbing time constant heterogeneity across localities and common shocks to municipality spending in given states over time. In our sample period, we observe 59 jurisdictions with a switch in the dominant party in the local council (44 from SPD to CDU and 15 from CDU to SPD). While this number of treated observations is small, the strong sample requirements ensure that the jurisdictions included in the sample are comparable.

The vector Xit comprises socio-economic control variables for our sample jurisdictions. These include (linear and squared) (ln) localities’ population (in 1000), (ln) employed population, age composition of the jurisdiction as determined by the population shares aged under 20 and above 65 as well as the (ln) unemployment rate and (ln) municipalities’ debt, all lagged by one period. We also control for (ln) higher governments grants, overall and for each of the analyzed subcategories. We report robust standard errors that account for clustering at the municipality and election term level.

The empirical strategy hence resembles a difference-in-differences-design: We compare adjustments in the spending pattern of localities with changing council majorities in the course of local elections to localities where majorities remained constant. The key identifying assumption is that there is a common trend in the size and structure of local public spending of treatment and control communities and spending patterns would have developed similarly in the absence of the treatment. We cannot rule out that this assumption is violated: There may be confounding factors that correlate with treatment assignment and directly impact the dependent variable. If municipalities, for example, face adverse economic shocks or threats of future adverse economic shocks (e. g. important local employers considering relocation), the electorate might push for adjustments in public spending (e. g. increases in social spending and adjustments in investment spending) and, simultaneously, give their vote to the conservative CDU/CSU at increased rates, which are commonly perceived to have more economic competence among their ranks (see e. g. Infratest-dimap (2019)). While actual changes in economic conditions can be absorbed by observed control variables (e. g. the local unemployment rate), signals of future shocks (e. g. considerations about plant relocation of large local employers) may act as a confounder and are genuinely unobserved.

We hedge against this threat by augmenting the fixed effect strategy with an instrumental variable approach. The IV strategy exploits variation in party votes at the local level induced by changing party support at the federal level, drawing on data from monthly opinion polls on intended voting behavior prepared by the ‘Forschungsgruppe Wahlen’ – a scientific society – for one of the two main public television channels in Germany. The poll is based on a representative sample of the German population eligible to vote. Specifically, we use poll results for the following question: “If next Sunday were to be a federal election day for the German national parliament, which party would you vote for?”. There are different potential shifters of intended voting behavior at the federal level, including changes in the perceived competence and popularity of party representatives at the national level (e. g., members of the national parliament or the national government) or perceived changes in their political positions. Exogenous events may also alter federal policy preferences; the ‘Fridays for Future’-protests, e. g., were associated with an increase in the support for the green party in Germany. Less than perfectly informed voters may rely on signals on party performance and party positions at the federal level – which are often more salient due to better media coverage – to proxy for expected party performance and positions at the local level; changes in federal party support may, therefore, trigger changes in local party support. Our empirical identification strategy exploits that in some jurisdictions these federal trends in party support imply that parties gain the majority of council seats, in others not. Common trends in policy outcomes, for example related to shifts in the general policy preferences of the electorate are absorbed by (state-)year fixed effects.

Technically, the instrument is constructed as follows: For municipalities with a SPD majority in the local council in the first election period, we predict the seat share of the main contending party – the CDU – in the following election seatcduˆt+1 as the product of the actual seat share for CDU in the prior election seatcdut and the percentage-change in voting intention for the CDU in federal elections between t and t+1, drawn from the opinion poll by ‘Forschungsgruppe Wahlen’ Δcdupollt: seatcduˆt+1=seatcdut·(1+Δcdupollt). If the initial CDU seat share in a jurisdiction is, for example, 45 % and the intended vote share for the CDU in the federal election increased by 20 % between t and t+1, then we predict a CDU seat share of 54 % (=45%·1.2) in t+1. The instrument logs that the CDU wins the absolute majority if the predicted seat share seatcduˆt+1 is above 50 %: wincduˆt+1=I(seatcduˆt+1>50%). We proceed analogously for municipalities with a CDU majority in the local council at t by predicting the change in seat share for the main contending party, the SPD: seatspdˆt+1=seatspdt·(1+Δspdpollt) and the emergence of a new SPD council majority winspdˆt+1=I(seatspdˆt+1>50%).

The variation is thus allocated to two instruments: wincduˆt+1 and winspdˆt+1. By definition, the former (latter) instrument does not change for municipalities with a CDU (SPD) majority in the election at t. Allocating the variation to two rather than one instrument, allows us to test whether effects on policy outcomes are symmetric when the SPD wins and loses the local council majority from/to the CDU, as is expected under partisan behavior. If SPD council majorities, for example, were associated with higher social spending than CDU majorities, we would expect to see an increase (a drop) in social spending when the SPD wins (loses) power. If adjustments, in turn, reflected changes related to simply winning an election, policy responses would be aligned. In the context of our example, social spending would then increase/decrease irrespective of the partisanship of the new majority.

In terms of intuition, using winspdˆt+1 as an instrument implies that we compare changes in the size and composition of public spending between localities with unchanged council majorities and municipalities where tailwind from federal politics (e. g. related to variation in the popularity of party representatives in the national parliament) predicts that the SPD wins the majority of local council seats. Analogously, using wincduˆt+1 as an instrumental variable implies that we compare changes in the size and composition of public spending between localities with unchanged council majorities and municipalities where tailwind from the federal level implies that the CDU wins the majority of seats in the local council. The key underlying assumption is that shifts in partisan support at the federal level are irrelevant for the municipal budget outcomes if it were not for their influence on the shifts in partisan majorities at the local level.

We consider this assumption to hold. First note that the difference-in-differences nature of the setup implies that we absorb common changes in underlying policy preferences (that may affect both, policy decisions at the federal and the local level irrespective of the party affiliation of decision makers) as they impact policy outcomes in the treatment group and control group alike. We, in turn, exploit that federal trends in party support imply that parties gain the majority of council seats in some municipalities, but not in others (depending on the initial seat shares in the prior election, see below). Second, the fact that we use shocks to voting intentions rather than actual changes in electoral outcomes at the federal level hedges us against concerns that shifts in federal party support may reallocate resources from politically unaligned to politically aligned municipalities, thereby affecting local budgetary outcomes. To further hedge against this concern, we, moreover, control for federal grants received by localities from higher government tiers in the estimation strategy.

A third concern arises because treatment assignment may correlate with a party’s seat share at t: To see this, consider again the example sketched above: If, in the construction of wincduˆt+1, the share of intended votes for the CDU in national parliamentary elections increases by 20 % between t and t+1, the CDU is predicted to gain the majority of council seats if the seat share at t is high (above 41.7 %), but not if it is low (below 41.7 %).^[14]

The IV strategy is therefore only valid if the initial seat share for SPD or CDU does not affect changes in overall spending and spending composition.^[15] As we cannot rule out such a correlation, we combine our difference-in-differences-IV strategy with a matching approach. The underlying idea is to draw on observed changes in spending patterns in the pre-election period and to re-weight treatment group and control group such that pre-election (and thus plausibly common) trends in public spending are similar in the two groups. While our main concern is that differences in initial seat shares (an observed variable) may lead to different underlying spending trends, matching on pre-trends would also account for other – observed and unobserved – reasons why underlying trends in the dependent variable may differ between the treatment and control group.^[16] Note in this context that, while the number of treated municipalities is rather small, the matching analysis draws on a large pool of control jurisdictions with unchanged council majorities.

The most often used matching approach in the literature is propensity score matching (where the propensity score represents the likelihood of being treated). It can be used to match observations, e. g. to find the closest control unit for every treatment observation, or to weight observations to create balance between control and treatment units (see Imbens2004: for a review). One particular assumption of the propensity score approach – in addition to conditional independence – is that the distribution behind the mean of the matching variables is the same, as otherwise observed differences between treatment and control group remain unaccounted and bias the estimated average treatment effect.

One recently proposed approach to overcome the lack of co-variante balancing in a selection-on-observables framework is entropy-balancing (Hainmueller (2012)). The main advantage of this method is that covariate balancing is not just assumed but enforced in a constrained, nonlinear estimation approach. The approach obtains weights for each targeted moment of the balancing/matching variables for treatment and control group subject to the balancing constraints. The resulting weights can then be used in a weighted regression.^[17]

Therefore, we apply entropy balancing to balance spending patterns of treatment and control jurisdictions in the year before the election (but will show that using propensity score matching yields similar results to the ones presented in our main model). As our analysis separately assesses adjustments in local spending in the wake of changes in council-majorities from SPD to CDU and vice versa (see above) – and underlying spending trends may differ for localities where council majorities switch from SPD to CDU and those where council majorities switch from CDU to SPD-, we create two subsamples. Each of the two subsamples includes two consecutive election periods of jurisdictions. In the first subsample (Panel A), we include two types of jurisdictions: (i) jurisdictions which had a CDU majority in both election periods and (ii) jurisdictions which had a CDU majority in the first election period and a SPD majority in the second election period. In the second subsample (Panel B), we include (i) jurisdictions with a SPD majority in the local council in both election periods and (ii) jurisdictions with a SPD majority in the first election period and a CDU majority in the second election period.^[18]

We use the same set of matching variables for both subsamples. Specifically, we include the two year change in (ln) real expenditures, (ln) financial expenditures and the local business tax rate as well as the voluntary spending shares for general public goods, people public goods, culture and infrastructure public goods in the year prior to the second-period election. Further, we match on state dummies. We target for all variables (except for the state dummies) mean and variance of the distribution. The resulting balancing statistics for Panel A and Panel B are shown in Table A1 and Table A2 in the Appendix. The descriptive statistics as well as the point estimates for the treatment indicator obtained when regressing the matching variables on the treatment variable suggests that the entropy balancing is highly effective as all variables in the re-weighted sample are very similarly distributed for treatment and control group. Please note that we do not match on other jurisdiction characteristics as we control for them in the regressions. In particular, in our baseline approach we do not match on the levels of the matching variables as we include in all estimations municipality fixed effects. We address this, however, in a robustness check. Combining both subsamples gives us our final estimation sample, which includes 1,364 two-election periods (or 1,058 unique jurisdictions) and 12,702 municipality-year observations.

4 Results

Table 4 reports our baseline findings. Row (1) shows the results of our preferred estimation strategy, which is the instrumental variable approach using the entropy-balanced sample. The dependent variables are the log of municipalities’ overall real expenditures, financial expenditures, voluntary expenditures and localities’ spending composition as captured by voluntary spending shares for general expenditures, ‘people-oriented’ expenditures, cultural expenditures and infrastructure spending. The main explanatory variable is a dummy variable indicating that the SPD holds the majority of council seats, instrumented as described in the previous section. The sample accounts for all communities in our data, i. e. those that experienced a switch from SPD to CDU and those that experienced a switch from CDU to SPD (as well as the control municipalities, where majorities did not change).

The F-statistic for the first stage is around 15 and thus suggests that our instruments are relevant. Moreover, the first stage point estimates are as expected (the coefficient estimate for the instrument predicting switches from a CDU to a SPD majority is 0.57 (se: 0.32) and the coefficient estimate for the instrument predicting a switch from a SPD to a CDU majority is −0.63*** (se: 0.13)). See the first stage results reported in Column (1) of Table A5 in the Appendix.

The results suggest no impact of the partisanship of local council majorities on a jurisdiction’s local business tax rate choice, general spending or spending for culture. SPD majorities are, however, observed to devote a higher share of overall voluntary spending to people public goods and less to infrastructure public goods. The effect sizes are with 3.6 and −7.2 percentage points substantial given that the mean spending share of the respective categories is around 18 % and 22 % respectively. Our result, furthermore, suggest that SPD majorities engage in lower voluntary spending (as categorized by us). This could explain why the reduction of infrastructure expenditures is larger than the increase in people-oriented expenditures.

Furthermore note that, given that we have two instruments, one for the change from CDU to SPD majorities and one for the change from SPD to CDU majorities, we are able to assess directly whether the observed policy responses are symmetric and hence reflect partisan behavior (i. e. wether we find similar coefficient estimates for α1 in Equation 1 irrespective of whether we use communities where council majorities switch from SPD to CDU or communities where council majorities switch from CDU to SPD for identification); if that was not the case, effects might, e. g., also reflect policy adjustments in the wake of winning an election. The Hansen test is above 0.3 in all models despite the ones using log overall spending and financial spending as dependent variables, indicating that we, in most instances, obtain similar estimates for the effect of SPD council majorities, irrespective of whether identification relies on switches from SPD to CDU (as predicted by our first instrument) or switches from CDU to SPD (as predicted by our second instrument).^[19]

This finding is corroborated when we reestimate the model in a subsample of localities that either switch from a CDU to a SPD majority or have an unchanged CDU council majority, denoted as ‘Panel A’ in Row (2) of Table 4, and a subsample of localities that either switch from a SPD to a CDU majority or have an unchanged SPD council majority, denoted as ‘Panel B’ in Row (3) of Table 4. As the first stage F-statistic is below 10 in these subsample models, we report the results of reduced form estimates, where we regress policy outcomes on the excluded instrument respectively.^[20] The result pattern resembles our baseline findings. Importantly, the point estimate for people-oriented and infrastructure spending have the opposite signs and are statistically significant in both subsamples, suggesting that the identified policy adjustments reflect partisan effects rather than behavioral adjustments to gaining legislative power.

To assess whether endogeneity of some of our control variables may act as a confounder, Row (1) in Table A6 in the Appendix reports the results when excluding our set of control variables. This leaves the result pattern largely unchanged.^[21] The effect size is slightly reduced but the estimates are within the confidence intervals of our baseline results. In Row (2), results are reported when estimating not on the municipality-year level but when collapsing the data on the municipality-election term level. Results are very similar. Moreover, we assess whether estimates change when we not only match on changes in pre-election spending but also on the levels. Since the matching does not converge when matching is in addition on state-indicators, Row (3) in Table A6 shows the results when matching only on changes in pre-election spending patterns but not on state-indicators. The precision of the estimates is reduced and the effect sizes are somewhat larger but again within the confidence intervals of our baseline results.^[22] Row (4) then shows the results when matching on levels as well as changes in pre-election spending. The balancing statistics suggest again a successful strategy (see Table A3 and Table A4). The results are very similar to the ones in Row (3), suggesting that the municipality fixed effects absorb level differences in the baseline model.^[23] In the last row of Table A6, we, moreover, rerun our baseline model using standard propensity score matching rather than entropy matching, which yields qualitatively and quantitatively unchanged results.

Table 5 further refines the analysis and assesses which subcategories within the broader categories of ‘people-oriented’ spending and infrastructure spending are driving the results. Increased ‘people-oriented’ spending under SPD-dominated local councils is suggested to particularly reflect higher social spending (around 50 %) and higher spending for recreational goods (35 %), while lower infrastructure expenditures is suggested to particularly reflect lower spending for streets (90 %) and lower spending for economic promotion (10 %). Note that the Hansen tests again support the notion that the observed differences reflect partisan behavior (as symmetric effects of SPD-majorities are observed when identification relies on changes from SPD to CDU majorities and vice versa). Reduced form estimates in the subsamples of Panel A and B largely confirm these findings (see Rows (2) and (3) of Table 5).

Finally, Table 6 intends to refine the interpretation of our results. Specifically, our main findings suggest that spending shares for infrastructure drop under SPD-dominated local councils, while spending shares for people-oriented public goods increase; in quantitative terms, the decline in infrastructure spending, however, appears to outweigh the increase in people-oriented expenditures. The baseline table already suggests that this pattern is matched by a drop in overall voluntary spending under SPD-dominated councils. On top of that, we cannot exclude that spending in other categories, namely voluntary general spending, may also increase with SPD-dominated local councils – albeit the positive coefficient estimate for the SPD dummy does not gain statistical significance in the latter instance.^[24] The results in Column (1) of Table 6, moreover, suggest that the share of voluntary and mandatory expenditures assigned to the general spending category is higher in municipalities with SPD-dominated local councils. Within this category, adjustments in administration and public safety expenditure explain around one third of this response and adjustments in public company expenditures explain the remaining two thirds (cf. Columns (2) and (3)). Reduced form estimates in the subsamples of Panel A and Panel B (Rows (2) and (3)) confirm these findings. This suggests that some of the general expenditure categories that we assigned to mandatory spending are influenced by local partisan politics.

Finally, we assessed the robustness of our findings to using an RD methodology. To do so, we draw on data provided by Freier and Odendahl (2015) for two German states, Bavaria and North-Rhine-Westpfalia, and use an RDD to determine whether the partisanship of the mayor impacts on local spending patterns.^[25]

Mayors in German municipalities can influence policy outcomes by proposing and drafting legislation (which then needs to be enacted by a majority vote of the local council). In both considered states, the mayor is directly elected with an absolute majority of votes. If no candidate wins the absolute majority, run-off elections take place. Since focusing only on run-off elections could introduce a selection bias, we account for all (final) elections and the victory margin of SPD and CDU mayor candidates.^[26] We run two sets of models: one comparing the spending policies of localities with SPD mayors to that of communities with mayors from all other parties/mayors without party affiliation and one comparing spending policies of communities with CDU mayors to those with mayors from all other parities/mayors without party affiliation. Focusing on races between SPD and CDU candidates alone would strongly reduce our sample size as candidates from other parties and independent candidates often run for office. In the RD strategy, we thus test for partisan effects asking whether SPD mayors and CDU mayors respectively implement systematically different spending policies than mayors from other parties/mayors without party affiliation.

Figure 1

Victory Margin SPD Candidates.

Figure 2

Victory Margin CDU Candidates.

One important requirement for a valid RDD is no manipulation/sorting at the threshold (McCrary (2008)). We assess this assumption by using the approach outlined by Cattaneo et al. (2019) and find no sorting when comparing races marginally won and lost by SPD candidates. When comparing races marginally won and lost by CDU candidates, there is evidence for sorting around the threshold (SPD candidates: p-value: 0.80 and CDU candidates: p-value 0.01, see also Figures 1 and 2).^[27] We thus in the following focus on the races marginally won and lost by SPD candidates.

The base results are reported in Table 7 and compare spending outcomes in jurisdictions with and without SPD mayors. We choose the victory margin using the RD bandwidth window selection framework developed by Cattaneo et al. (2014) and our set of municipality control variables as well as all pre-determined outcome variables. This test recommends a victory margin of 1.9 % based on a minimum p-value for the covariance balance test of 0.1. This leaves us with 36 observations.

Panel 1 of Table 7 reports the base estimate. The result pattern resembles our baseline findings: SPD mayors tend to engage in less infrastructure spending and more spending on people-goods than mayors from other parties/with no party affiliation. The estimates are relatively imprecise, however, and none of the estimated coefficients gains statistical significance. To improve efficiency, Panel 2 reestimates the baseline model controlling for the lagged dependent variable.^[28] In this specification, the coefficient estimate for infrastructure spending turns out marginally significant, suggesting that SPD mayors allocate less resources to infrastructure investments.^[29] Appendix B, moreover, presents numerous robustness checks, including placebo tests, checks for the distribution of observed characteristics left and right of the threshold, robustness tests where we adjust the bandwidth of the RD model and specifications, where we opt for a more flexible modeling of the forcing variable. We, moreover, present results based on the recently proposed bias corrected local polynomial RD estimator (Calonico et al. (2014)) which has been found to be closest to experimental estimates (Hyytinen et al. (2018)). From our perspective, the RD analysis offers two insights: First, the, in part, failed specification tests suggest that RD design is not an ideal approach to identify partisan effects in our empirical context as we cannot rule out sorting at the threshold (see Figure 2). This supports our decision to opt for a different identification approach in our main analysis. Second, we, nevertheless, consider it reassuring that the RD model in the sub-analyses for the SPD mayoral candidates – where specification tests are passed – yields results that are similar to our base analysis.

5 Conclusion

The aim of this paper was to assess the role of partisanship of West German local council majorities on the size and composition of local public spending. We combine a instrumental variable fixed effect regression approach with an entropy balanced matching strategy to empirically identify the effect of interest. The results point to stark partisan effects: SPD council majorities are found to spend more on people-oriented public goods and less on infrastructure public goods. The driving subcategories are spending for recreation and social spending on the one hand and spending for streets and economic promotion on the other.