Electoral System, Fiscal Rule and Form of Government: Case Studies in Political Economics

Christian Friedrich Pfeil

Birthplace: Schlema

September 2016

Doctoral Thesis submitted to the Faculty of Economics and Behavioural Sciences at the University of Freiburg (Germany) Erstgutachter: Prof. Dr. Lars P. Feld Zweitgutachter: Prof. Dr. Bernd Raﬀelh¨uschen Dekan: Prof. Dr. Alexander Renkl

Datum des Promotionsbeschlusses: 4. Mai 2017 Acknowledgment

Writing a doctoral thesis is impossible without support from friends and colleagues. First of all, my wife Sabine consented to move from East to Southwest Germany and, for a long time, endured the absent-mindedness of her husband. Thank you very much indeed. Likewise, I give thanks to our three children as I often was not sufficiently patient with them. Many thanks to Lars P. Feld for his willingness to supervise this thesis, for the great opportunity to get in touch with the research area ‘Political Economics’ and for an atmo- sphere of academic freedom at his chair and institute. I wish to thank Bernd Raffelhüschen for his readiness to act as second assessor. Many thanks to Toke S. Aidt for his generous hospitality and hence the chance to visit the university of Cambridge (UK). Similarly, I give thanks to Ulrike und Lutz Pfaff who hosted me in 2008 and 2009. Many thanks to my classmate Benedikt Fritz for many insightful discussions and a lot of joint train rides. And I give thanks to Maximilian Grasl for sharing my interest in the unique government structure of Belgium. I wish to thank my colleagues at the University of Heidelberg, the University of Freiburg and the Walter Eucken Institute Thushyanthan Baskaran, Heiko Burret, Christian Con- rad, Annabelle Doerr, Ekkehard Köhler,Sarah Necker, Daniel Nientiedt, Christoph Sajons, Jan Schnellenbach, Johannes Voget and Julia Wolfinger for a considerable number of very interesting conversations. I wish to thank my relatives in Geyer, Halle, Leipzig, Lippstadt and Thum as well as my friends in Eppelheim and Heidelberg (at the United Methodist Church) who encouraged me in various ways.

Eppelheim, June 2017 Christian F. Pfeil Contents

List of Figures III

List of Tables V

1 Prelude 1 1.1 Do Institutions Matter? ...... 1 1.2 Methodological Reflections ...... 3 1.2.1 The Rubin Causal Model ...... 4 1.2.2 The Matching Estimator ...... 6 1.2.3 The Synthetic Control Estimator ...... 8 1.2.4 The Difference-in-Difference Estimator ...... 13 1.2.5 Discussion ...... 14

2 Do Electoral System Changes Impact on Government Spending? 18 2.1 Introduction ...... 18 2.2 Data ...... 21 2.3 The Alteration from the SNTV Rule to the MMM Electoral System in Japan 25 2.3.1 Institutional Setting ...... 25 2.3.2 The Effect on Social Spending ...... 26 2.3.3 The Effect on Old-Age Spending ...... 27 2.4 The Alteration from Plurality Rule to the MMP Electoral System in New Zealand ...... 28 2.4.1 Institutional Setting ...... 28 2.4.2 The Effect on Overall Spending ...... 30 2.4.3 The Effect on Social Spending ...... 31 2.5 The Alteration from ListPR to the MMM Electoral System in Italy in 1994 32 2.5.1 Institutional Setting ...... 32 2.5.2 The Effect on Overall Spending ...... 33 2.5.3 The Effect on Social Spending ...... 34 2.6 The Alteration from the MMM Electoral System to ListPR in Italy in 2006 36 2.6.1 Institutional Setting ...... 36 2.6.2 The Effect on Overall Spending ...... 37 2.6.3 The Effect on Social Spending ...... 38 2.7 Conclusions ...... 38

I 2.8 Tables and Figures ...... 41 2.9 Appendix: Variables ...... 54

3 Does the Swiss Debt Brake Induce Sound Federal Finances? 56 3.1 Introduction ...... 56 3.2 The Design of the Swiss Balanced Budget Rule ...... 59 3.3 Data ...... 62 3.4 Empirical Analysis ...... 64 3.4.1 The Eﬀect on the Budget Balance ...... 64 3.4.2 The Eﬀect on the Government Debt Ratio ...... 67 3.4.3 On the Introduction of Fiscal Rules among the Comparison Units . 69 3.4.4 The Debt Brake in the Medium Run ...... 70 3.5 Conclusions ...... 71 3.6 Tables and Figures ...... 73 3.7 Appendix: Variables ...... 78

4 Do Federalism Reforms in Belgium Cause Economic Growth? 79 4.1 Introduction ...... 79 4.2 Federalism Reforms in Belgium ...... 84 4.2.1 The First State Reform (1970) ...... 84 4.2.2 The Second State Reform (1980) ...... 85 4.2.3 The Third State Reform (1988/89) ...... 86 4.2.4 The Fourth State Reform (1993) ...... 88 4.2.5 The Fifth State Reform (2001) ...... 89 4.2.6 The Sixth State Reform (2011-13) ...... 91 4.2.7 Hypotheses ...... 91 4.3 Data ...... 92 4.4 Empirical Analysis ...... 94 4.4.1 The Growth Effect of the 1989 Reform ...... 94 4.4.2 The Growth Effect of the 1993 Reform ...... 96 4.4.3 The Growth Effect of the 2001 Reform ...... 98 4.5 Conclusions ...... 100 4.6 Tables and Figures ...... 102 4.7 Appendix I: Variables ...... 110 4.8 Appendix II: Comparison Units ...... 111

Bibliography 112

II List of Figures

1.1 Stylised Representation of the Synthetic Control Method ...... 9

2.1 Social Expenditure in Japan ...... 41 2.2 Social Expenditure in Japan, Post-Pre-Ratios ...... 42 2.3 Share of Elderly People in Japan & OECD Countries ...... 42 2.4 Social Expenditure in Chile, Japan, Portugal & Pool Countries ...... 42 2.5 Old-Age Expenditure in Japan ...... 43 2.6 Old-Age Expenditure in Japan, Post-Pre-Ratios ...... 43 2.7 Old-Age Expenditure in Japan (w/o DEU), Post-Pre-Ratios ...... 44 2.8 Old-Age Expenditure in Japan (w/o CHL, DEU), Post-Pre-Ratios . . . . . 44 2.9 Total Outlays in Canada, Ireland, New Zealand, USA & Pool Countries . . 44 2.10 Total Outlays in New Zealand ...... 45 2.11 Total Outlays in New Zealand, Post-Pre-Ratios ...... 45 2.12 Social Expenditure in New Zealand ...... 46 2.13 Social Expenditure in New Zealand, Post-Pre-Ratios ...... 47 2.14 Social Expenditure in New Zealand (w/o DNK, ESP, FIN, SWE), Post- Pre-Ratios ...... 47 2.15 Social Expenditure in New Zealand (w/o DNK, ESP, FIN, IRL, POL, SWE), Post-Pre-Ratios ...... 47 2.16 Total Outlays in Italy, 1994 ...... 48 2.17 Total Outlays in Italy, 1994, Post-Pre-Ratios ...... 48 2.18 Social Expenditure in Italy, 1994 ...... 49 2.19 Social Expenditure in Italy, 1994, Post-Pre-Ratios ...... 49 2.20 Social Expenditure in Italy, 1994 (w/o NLD), Post-Pre-Ratios ...... 50 2.21 Social Expenditure in Italy, 1994 (w/o NLD, NOR), Post-Pre-Ratios . . . 50 2.22 Central Government Expenditure in Italy, 2006 ...... 51 2.23 Central Government Expenditure in Italy, 2006, Post-Pre-Ratios ...... 51 2.24 Central Government Expenditure in Italy, 2006 (w/o SVK), Post-Pre-Ratios 51 2.25 Central Government Expenditure in Italy, 2006 (w/o BEL, SVK), Post- Pre-Ratios ...... 52 2.26 Social Expenditure in Italy, 2006 ...... 52 2.27 Social Expenditure in Italy, 2006, Post-Pre-Ratio ...... 53

3.1 Cyclically Adjusted Budget Balance ...... 73

III 3.2 Cyclically Adjusted Budget Balance, Post-Pre-Ratios ...... 74 3.3 Cyclically Adjusted Budget Balance (w/o CZE, DEU), Post-Pre-Ratios . . 74 3.4 Cyclically Adjusted Budget Balance (w/o CZE, DEU, NZL), Post-Pre-Ratios 74 3.5 Cyclically Adjusted Budget Balance (w/o AUT, CZE, DEU, ITA, JPN, NZL, SWE), Post-Pre-Ratios ...... 75 3.6 Cyclically Adjusted Budget Balance (w/o AUT, CZE, DEU, FIN, ITA, JPN, NOR, NZL, SWE), Post-Pre-Ratios ...... 75 3.7 Federal Government Debt Ratio ...... 76 3.8 General Government Debt Ratio ...... 77 3.9 Cyclically Adjusted Budget Balance, extended post-period ...... 77

4.1 Real GDP Per Capita in Flanders, 1989 ...... 102 4.2 Real GDP Per Capita in Flanders, 1989, Gaps ...... 102 4.3 Real GDP Per Capita in Belgium, 1993 ...... 103 4.4 Real GDP Per Capita in Belgium, 1993, Gaps ...... 103 4.5 Real GDP Per Capita in Flanders, 1993 ...... 104 4.6 Real GDP Per Capita in Flanders, 1993, Gaps ...... 104 4.7 Real GDP Per Capita in Wallonia, 1993 ...... 105 4.8 Real GDP Per Capita in Wallonia, 1993, Gaps ...... 105 4.9 Real GDP Per Capita in Belgium, 2001 ...... 106 4.10 Real GDP Per Capita in Belgium, 2001, Gaps ...... 107 4.11 Real GDP Per Capita in Flanders, 2001 ...... 107 4.12 Real GDP Per Capita in Flanders, 2001, Gaps ...... 108 4.13 Real GDP Per Capita in Wallonia, 2001 ...... 108 4.14 Real GDP Per Capita in Wallonia, 2001, Gaps ...... 109

IV List of Tables

2.1 Social Expenditure in Japan, w-Weights ...... 41 2.2 Social Expenditure in Japan, Predictor Balance ...... 41 2.3 Old-Age Expenditure in Japan, w-Weights ...... 43 2.4 Old-Age Expenditure in Japan, Predictor Balance ...... 43 2.5 Total Outlays in New Zealand, w-Weights ...... 45 2.6 Total Outlays in New Zealand, Predictor Balance ...... 45 2.7 Social Expenditure in New Zealand, w-Weights ...... 46 2.8 Social Expenditure in New Zealand, Predictor Balance ...... 46 2.9 Total Outlays in Italy, 1994, w-Weights ...... 48 2.10 Total Outlays in Italy, 1994, Predictor Balance ...... 48 2.11 Social Expenditure in Italy, 1994, w-Weights ...... 49 2.12 Social Expenditure in Italy, 1994, Predictor Balance ...... 49 2.13 Central Government Expenditure in Italy, 2006, w-Weights ...... 50 2.14 Central Government Expenditure in Italy, 2006, Predictor Balance . . . . . 50 2.15 Social Expenditure in Italy, 2006, w-Weights ...... 52 2.16 Social Expenditure in Italy, 2006, Predictor Balance ...... 52

3.1 Cyclically Adjusted Budget Balance, w-Weights ...... 73 3.2 Cyclically Adjusted Budget Balance, Predictor Balance ...... 73 3.3 Federal Government Debt Ratio, w-Weights ...... 75 3.4 Federal Government Debt Ratio, Predictor Balance ...... 76 3.5 General Government Debt Ratio, w-Weights ...... 76 3.6 General Government Debt Ratio, Predictor Balance ...... 76

4.1 Real GDP Per Capita in Flanders, 1989, Predictor Balance ...... 102 4.2 Real GDP Per Capita in Belgium, 1993, w-Weights ...... 103 4.3 Real GDP Per Capita in Belgium, 1993, Predictor Balance ...... 103 4.4 Real GDP Per Capita in Flanders, 1993, Predictor Balance ...... 104 4.5 Real GDP Per Capita in Wallonia 1993, Predictor Balance ...... 105 4.6 Real GDP Per Capita in Belgium 2001, w-Weights ...... 106 4.7 Real GDP Per Capita in Belgium 2001, Predictor Balance ...... 106 4.8 Real GDP per Capita in Flanders 2001, Predictor Balance ...... 107 4.9 Real GDP Per Capita in Wallonia 2001, Predictor Balance ...... 108

V 1 Prelude

1.1 Do Institutions Matter?

Political instutions are rules of the game. Rules that guide the behaviour of politicians and officials. Hence, political instutions have a bearing on the workings of the political system and thus the society. Electoral rules, fiscal rules and the form of government are examples of political institutions. The electoral system determines how votes of the electorate are transferred into seats in parliament (Taagepera and Shugart, 1989). As candidates or parties aim at maximising the number of seats in parliament they react to such rules. Fiscal rules directly affect the spending behaviour of fiscal policy makers as they, e.g., require the balance of the budget or stipulate a ceiling of expenditures. Finally, a decentralised government structure may afford politicians the opportunity to tailor regional economic policy measures to regional needs (Oates, 1993, 1999). Since political institutions impact on the operation of the polity in a very general way they are often enshrined in the constitution. As political institutions may have far-reaching consequences scholars as well as practitioners, e.g. legislators, would like to learn more about how they work and what the effects are. That is, they want to know whether political institutions are effective in guid- ing the behaviour of politicians. Consequently, section 3 uncovers the effectiveness of the Swiss debt brake that was de facto introduced in 2003. This fiscal rule stipulates that expenditures of the next fiscal year must follow the predicted revenues of that fiscal year and applies to the national level. Thus it requires the balance of the budget. We find that the introduction of this fiscal rule improved the cyclically adjusted budget balance by about 3.6 percentage points on average in a post-intervention period covering five years. While fiscal rules usually are implemented to affect an economic variable political institutions in the form of the electoral rule or the form of government usually are implemented to affect political variables. That is, a society barely would change the electoral rule in order to raise social spending. Likewise, a society barely would establish subnational territorial authorities in order to cause economic growth. However, political institutions may impact on economic variables like government spending or economic growth. Scholars and practitioners thus also would like to know such (side) effects. With respect to the electoral system Persson and Tabellini (2003, 2004) find that governments that run under a majoritarian electoral rule exhibit a level of overall spending that is about five percentage points of GDP lower compared to governments that run

1 under a proportional electoral rule. Regarding the composition of government spending they find that governments that run under a majoritarian rule exhibit social spending as percentage of overall spending that is two or three percentage points of GDP lower than governments that run under a proportional electoral rule. In section 2 I capture several changes of the electoral system. Where Japan and New Zealand switched from a majoritarian rule to a mixed-member electoral system in 1996, Italy switched from a proportional rule to a mixed-member electoral system in 1994. In 2006 Italy switched back to a proportional rule. I find an effect on the overall level of spending in the range of 2.13 to 3.36 percentage points. However, the treatment effect is either poorly statistically significant or insignificant. I find a clear significant effect on social spending in New Zealand (2.08 percentage points) but not in Japan (0.45 percentage points) and Italy (0.42 percentage points and 1.53 percentage points). In section 4 we uncover the growth effects of federalism reforms in Belgium at the national level as well as for Flanders and Wallonia (regional level). While the third state reform (1988/89) further increased the spending power of regions and the communities, the fourth state reform (1993) raised the financial means available to the communities. In 2001 (fifth state reform) this funds were further increased and taxing authority as well as spending power was transferred to the regions. We find a small positive growth effect of the 1989 reform in Flanders. Regarding the 1993 reform we find small negative growth effects at the national level, in Flanders and in Wallonia. Finally, we do not find a growth effect of the 2001 reform at the national level but a small negative growth effect of the 2001 reform in Flanders as well as in Wallonia. All effects we find are not exceptional and thus statistically insignificant. That is, we cannot find clear significant growth effects of federalism reforms in Belgium. In that respect, our results correspond to studies provided by Davoodi and Zou (1998), Thornton (2007) and Xie et al. (1999) who cannot find a significant effect of fiscal decentralisation on economic growth. As the effects we find are rather small in size and clearly insignificant our results are also in line with Asatryan and Feld (2014) who conclude that there is no effect on economic growth at all. The core task of this dissertation can be described as estimating the economic effects that accrue from a change of political institutions. That is, the analyses of this thesis explicitly capture the change of institutions instead of comparing different statuses of political institutions. This is basically a good approach to capture causal effects. However, the number of such changes usually is very small. Bormann and Golder (2013) report roughly 50 electoral system changes in a worldwide sample covering 65 years. With respect to the Swiss debt brake we are interested in the effectiveness of one single institution. In order to nevertheless capture the change of political institutions this thesis employs the Synthetic Control Method that allows for the execution of quantitative case studies (see section 1.2.3). To sum up, I can say that the very rationally designed Swiss debt brake is effective in

2 balancing the budget. The shift from a majoritarian electoral rule to an (almost) proportional electoral system in New Zealand significantly impacts on social spending. Further electoral system changes which do not embrace an alteration from one pure electoral rule (e.g. the majoritarian rule) to the other (e.g. the proportional rule) do not significantly impact on social spending. Likewise, the 2001 federalism reform that is very much based on political compromise does not cause a growth effect. Although the results of this dissertation originate from case studies which are strongly associated with a high level of internal validity and a low level of external validity, I cautiously conclude that political institutions are effective if outstandingly designed. The work presented in section 2 is authored by myself and has been pre-released as Freiburg Discussion Paper on Constitutional Economics No. 16/06. The study presented in section 3 is coauthored with Lars P. Feld. I contributed to this study via managing the data, executing the empirical analysis (including the utilisation of the Hodrick-Prescott filter) and drafting the words. It has been pre-released as CESifo Working Paper No. 6044. The work presented in section 4 is coauthored with Benedikt Fritz and Maximilian Grasl. I contributed to this study via collecting information on reform contents, executing the empirical analysis and providing all words except for text passages related to the historical and political background of federalism in Belgium. Due to the different authorships the personal pronoun changes in the course of this thesis.

1.2 Methodological Reﬂections

As stated above the body of this thesis uncovers economic effects caused by alterations of political institutions. The intuitive approach to precisely uncover such effects would be to compare the economic outcome that emerges under the treatment, that is under the modified political institution, with the economic outcome that emerges in the absence of the alteration. However, this comparison cannot be made for one single unit.1 This is because such a unit either receives the treatment or it does not receive the treatment at one point in time. However, it is possible to compare the unit’s outcome that emerged in the time before it received the treatment with the outcome of the same unit that emerges after the treatment’s implementation. Alternatively, one can compare the outcome of a unit that receives the treatment with the outcome of another unit that does not receive this treatment.2 In the latter case these two outcomes, however, can only be compared directly if the two units are equivalent in all respects except for the treatment. This can be best achieved via the execution of an experiment. Unfortunately, it is not possible to

1Within the scope of this thesis ‘unit’ refers to countries or regions within countries. 2This assumes, of course, that the intervention is permanent existent after it was once introduced (at least for the time under scrutiny). In fact, this assumption holds for all cases covered by the body of this work.

3 conduct an experiment whenever researchers want to investigate the eﬀects of a policy intervention. Or put another way, data that captures an intervention, i.e. observational data, often does not fulﬁl the assumption of random assignment. Consequently, the comparison of treated with untreated units should be based on covariates to assure that units are as similar as possible to each other.

1.2.1 The Rubin Causal Model

The exposition of the former paragraph comprises the line of thought of the Rubin Causal Model (Rubin, 1974). More formally, the starting point is an individual i that in one case receives a treatment and in another case it does not receive the treatment. If it is exposed 1 to the treatment the outcome Yi can be observed. If the individual i is not exposed 0 to the treatment the outcome Yi can be observed. The superscript denotes the binary treatment. The causal treatment eﬀect is then simply the diﬀerence between these two outcomes 1 0 τi = Yi − Yi . (1.1)

1 0 As mentioned, we cannot observe both outcomes Yi and Yi in time t. Thus, the outcome 0 0 Yi becomes the counterfactual outcome Yj that indicates what would have happened to individual i if it were not under the treatment in time t. This outcome usually comes from another individual j that is not exposed to the treatment. The idea of a counterfactual outcome is the basic concept of the Rubin Causal Model. This concept can easily be extended towards a group of treated units (n > 1). The average treatment eﬀect over all treated individuals can then be depicted by

N 1 X τ = Y 1 − Y 0 , (1.2) att N i j i=1 where N is the total number of treated individuals. This effect is called the ’average treatment effect on the treated’ (ATT) and describes the average causal effect of the treatment on a finite number of individuals that were indeed affected by the intervention (Rubin, 1974).3 As the counterfactual outcome stems from another (untreated) unit the question arises how the untreated unit is assigned to the treated unit. Rubin (1977) thus considers an infinite number of units P . From this set P a subset I of units is randomly chosen to receive the treatment where the remaining units J provide the counterfactual outcome. He shows that the obtained treatment effect is unbiased, if an untreated unit j with a value α of a covariate x is assigned to a treated unit i with the same value α of the

3The average treatment effect (ATE), in contrast, gives the expected causal effect of the treatment on a unit randomly drawn from the population. This effect can be derived only from an experiment (Gangl and DiPrete, 2004).

4 covariate x. Finally, he shows that the bias is smaller if the assignment is based on more than one covariate. Since data often is observational in an empirical execution, the creation of the counterfactual outcome cannot be done without assumptions. These basic assumptions can be illustrated by equation

1 0 0 0 Yi − Yj = τi + Yi − Yj + (τi − τj) (1.3)

1 0 were Yi −Yj depicts the empirically obtained treatment effect (Gangl and DiPrete, 2004). The equation’s first term gives the ‘real’ treatment effect. However, this effect is clear only when any treatment to any unit results in a unique outcome. If that holds, the stable unit-treatment value assumption (SUTVA) is satisfied. It is violated, in contrast, if there are spillover effects or other interdependencies between the treated unit as well as the untreated unit.4 That is, the treatment of unit i must not affect the outcome 0 of a comparison unit j. Otherwise, the outcome Yj is biased and cannot be used as counterfactual outcome. The second term captures that treated and untreated unit might be different from each other in principal and thus irrespective of a potential treatment. To minimise this bias treated and untreated unit are matched based on similarity regarding some further covariates as suggested by Rubin (1977). If resemblance is very close the unit homogeneity assumption (UHA) is satisfied. Beyond that, unit i might respond to the treatment unlike unit j would do. The reason might be that unit i expects a certain gain from participating and thus actively tries to receive the treatment. Such self-selective behaviour is covered by the last term (τi − τj) of the equation and should be brought down to zero in order to get an unbiased result. If that is true the conditional independence assumption (CIA) holds.5 This assumption states that conditional on a vector of observable characteristics X, the assignment to the treatment does not impact on the outcome of both the treated as well as the untreated unit. Put another way, the 1 0 outcomes Yi and Yj must be orthogonal to the binary treatment status D, conditional on the covariates in X. More formal,

1 0 Yi ,Yj ⊥⊥ D | X (1.4) must hold. For the assumption to hold, all variables that aﬀect both participation and outcomes need to be covered by the vector X. In the following sections I consider estimators that capture the causal eﬀects of policy interventions (treatments). I discuss the Matching Estimator, the Synthetic Control

4Spillover eﬀects can occur, if, e.g., the treated unit consumes ressources that are not available to the untreated unit anymore. 5This assumption also goes under the names ‘assumption of unconfoundedness’, ‘selection on observables’ and ‘(strong) ignorability’. The latter can be traced back to Rosenbaum and Rubin (1983).

5 Method (SCM) and the Difference-in-Difference (DiD) Estimator. The Synthetic Control Method is the estimator used throughout this thesis. However, I restrain from dealing with the Regression-Discontinuity-Design (RDD) although it is also a reliable method when it comes to the disclosure of causal treatment effects. This is because its basic concept is different. RDD is also based on a comparison between treated and untreated units. However, it wilfully creates an experimental setting and thus compares units that were randomly assigned to a treatment with other units that randomly missed the same treatment (Lee and Lemieux, 2010). In terms of an election, for example, RDD would compare a politician who barely won a plurality rule election with another politician who just lost a plurality rule election. To this, RDD searches for and uses a threshold that separates between treated and untreated units. In terms of the plurality electoral rule the threshold is given by 50 percent of the votes plus one vote.6 The estimators ME, SCM and DiD instead aim at finding and using comparison units that are as similar to the treated units as possible (perfect counterpart). Finally, they use comparison units in a non-randomised setting. That is, they use observational data. In fact, Rubin (1974) shows that randomisation is not necessary if a comparison unit is perfectly identical to the treated unit. As this perfect comparison unit differs from the treated unit in nothing except for the treatment, it can be assumed to respond to the treatment just as the treated unit would do. Consequently, the counterfactual outcome indeed indicates what would have happened to the treated unit if it would have missed the treatment.

1.2.2 The Matching Estimator

This estimator comes very close to the concept of the Rubin Causal Model as it directly compares treated units with untreated units. To this, it searches for an untreated unit that is best suited to a treated observation regarding some observable characteristics X. Eventually, it compares the outcome of the treated unit with the outcome of the assigned untreated unit and an average treatment effect is estimated over all pairs. In its original constitution matching is pairwise and cross-sectional (Todd, 2008). Basically, there is a trade-off concerning the number of covariates included in the empirical model. On the one hand it is useful to employ a large battery of covariates. Firstly, because all observable characteristics that affect both treatment status and outcome need to be included in order to fulfil the CI assumption. Secondly, covariates might to some extent capture unobserved characteristics. Thirdly, the more covariates are included the smaller might be the violation of the unit homogeneity assumption (UHA). On the other hand, employing many covariates is problematic as the dimensionality of the matching process increases with every included covariate (curse of dimensionality). This especially holds if the covariates are continous variables (Todd, 2008).

6In this case the treatment entails holding an office. The winning politician accidentally is appointed to an office where the defeated politican accidentally missed the office.

6 However, Rosenbaum and Rubin (1983) show that if the outcome of the treated unit 1 0 Yi and the outcome of the untreated unit Yj are independent of the participation D conditional on a vector of covariates X, that is, if the CI assumption (see equation 1.4) holds and the estimator is unbiased, then the estimates are also unbiased if the outcomes are independent of the treatment status D conditional on the probability of participating in the treatment. That is, 1 0 Yi ,Yj ⊥⊥ D | P (X) (1.5) also holds. The probability of participating in the treatment P (X) = P rob(D = 1 | X) is determined by means of a vector of observed covariates and is ﬁnally depicted by the one- dimensional propensity score. Thus, the propensity score matching estimator is based on a two-step procedure. For the one thing the propensity score is estimated and for another thing matching is conducted on the basis of the propensity score obtained at the ﬁrst stage. Matches can be bad if the procedure is conducted pairwise without replacement. Under this setting the estimator assigns the closest untreated unit to each treated unit. This untreated unit is not assigned to another treated unit once it is matched (single nearest- neighbour matching). To see this assume that an untreated unit ja is the best match for two treated units ia and ib and assume that it is assigned to unit ia. Consequently, another untreated unit jb is matched with ib resulting in a match (ib, jb) being worse than the match (ib, ja). Thus the quality of pairwise-matching can be improved by conducting it with replacment such that an untreated unit can be assigned to several treated units. In order to keep a certain quality level, the divergence of the untreated unit’s propensity scores from the treated unit’s propensity score can be limited. Hence, an untreated unit is used for matching only if its propensity score Pj is within the tolerance (caliper matching). That is,

kPi − Pjk < (1.6) must hold. If a match cannot be found for a treated unit it is excluded from the analysis. In order to decrease the variance of the estimator the treated unit can be matched with the (weighted) average of a number of untreated units (oversampling). Usually, the ﬁve or ten nearest neighbours are used. The ATT depicted in equation 1.2 thus turns into

N " J # 1 X X τ = Y 1 − w Y 0 (1.7) att N i j j i=1 j=1

PJ were J is the set of untreated units and j=1 wj = 1 holds. The variance decreases because more information is used to construct the counterfactual outcome. However, biases increases because, on average, matches are poorer then. Similarly, (nearly) all units of the control group can be used for the assignment (kernel matching). The counterfactual

7 outcome is the kernel-weighted average of the comparison units. The kernel function determines the weight of each untreated unit in dependence of its distance from the treated unit in terms of the propensity score (Caliendo and Kopeinig, 2008; Heckman et al., 1997; Smith and Todd, 2005). In order to capture time-invariant unobserved characteristics, Heckman et al. (1997) introduced the conditional Difference-in-Difference (DiD) Matching Estimator. This estimator also matches every treated unit with an untreated unit. However, it uses the change in the outcome variable ∆Y = Yt − Yt−1 with Yt being the outcome in the period after the intervention and Yt−1 being the outcome in the period before the intervention. The first difference is represented by ∆Y where the second difference is captured by the comparison between treated unit and untreated unit. In that, it is similar to the standard Difference-in-Difference (DiD) Estimator that is described in section 1.2.4. However, it is nevertheless a non-parametric matching estimator which does not imposes a (linear) functional form on the empirical model. Concerning the bias stemming from different matching estimators Heckman et al. (1997) show that DiD-Matching is in general less biased compared to other (cross-sectional) matching estimators.7 However, it cannot eliminate bias completely. They show that DiD- Matching indeed reduces bias if treated unit and comparison unit are located in the same (economic) setting (e.g. one local labour market). They take this as evidence that DiD- Matching reduces bias stemming from time-invariant unobserved characteristics. Finally, they show that the bias of the Matching Estimator is in principal sensitive to the variables included in the popensity score estimation and that the DiD-Matching Estimator is less prone to insufficient information in the estimation of the propensity score compared to other (cross-sectional) matching estimators.

1.2.3 The Synthetic Control Estimator

The Synthetic Control Method (SCM) combines the analysis of case studies with quantitative procedures regarding the approximation of the treated unit’s trajectory in the pre-intervention period as well as the statistical signiﬁcance of the intervention. Since it allows for the analysis of only one or a few interventions, it is very well-suited to dis- cover the eﬀects of political institutions on policy or economic outcomes. The following description of the method follows Abadie et al. (2010, 2015). In principle the method depicts the trajectory of an outcome variable for one single unit of interest (treated unit) in the pre-intervention period as well as the post-intervention period. The method then aims at approximating the path of the outcome variable of the treated unit in the pre-intervention period as well as possible with that path of a control group. This control group is called ’synthetic’ since it does not consist of only one

7This result is conﬁrmed by Smith and Todd (2005).

8 Y 6

Treated Unit

Treatment

Eﬀect Synthetic Control

- pre-intervention period post-intervention period t

Figure 1.1: Stylised Representation of the Synthetic Control Method comparison unit (e.g. one country) but of (the weighted average of) a positive number of units. Eventually the trajectory of the outcome variable of the synthetic control group is depicted for both periods. So the path of the control group in the post-period shows how the outcome variable of the treated unit would have evolved without the intervention. The difference between both trajectories in the post-period thus provides the Synthetic Control Estimator. A stylised representation of the method is given in figure 1.1 where the solid line represents the outcome variable of the treated unit and the dashed line represents the outcome variable of the synthetic control group, respectively. The effect of the intervention on the outcome variable is then represented by the grey triangle. More formally the method is based on a sample of i + j units. While i depicts the treated unit, j = 1, ..., J is a positive number of potential comparison units. All these units together constitute the donor pool. Units that are not in the pool obviously cannot be part of the control group and ultimately the researcher decides on which units are in the donor pool. Time runs from t = 1, ..., T where Tpre is a positive number of pre- intervention time periods. Tpost, in contrast, entails a positive number of post-intervention 8 time periods. Thus it holds that T = Tpre + Tpost. The approximation of the outcome variable of the treated unit is not obtained by simply using the weighted average of the outcome variables of the synthetic control group but is based on some k = 1, ..., K (economic) characteristics that are predictor variables of 9 the outcome variable. Xi is the (K × 1) vector of these predictor variables for the treated unit in the pre-intervention period. The predictors are averaged over this period.

Likewise, the (K × J) matrix XJ covers the same K (averaged) predictor variables for

8The year of the intervention itself neither belongs to the pre-intervention period nor to the post- intervention period. 9Lags of the outcome variable are often used as predictors, too.

9 the same time period but for all J potential comparison units in the donor pool. Since the predictor variables differ in their ability to predict the outcome variable, a weight vk is assigned to every predictor variable k based on a regression. In a next step the synthetic control group is obtained by compiling the (J × 1) vector 0 of weights W = (w1, ..., wJ ) . The control group thus is the weighted average of the comparison units. These weights are required to be non-negative and sum up to one so that 0 ≤ wj ≤ 1 and w1 + ... + wJ = 1 holds. Usually some units in the donor pool are not chosen to be part of the synthetic control group, which means that the respective comparison units have a zero w-weight. The W -vector is then chosen to approximate the matrix XJ with the vector Xi as well as possible. That is, it minimises the difference vkXi − XJ W k. The ultimate aim of the optimisation is to minimise the root mean square prediction error (RMSPE)10 of the outcome variable in the pre-intervention period.11 The vector W ∗ is then used to depict the outcome variable of the synthetic control 0 group. Let Yi = (Yi,Tpre+1, ..., Yi,T ) be a (Tpost × 1) vector that contains the values of the outcome variable for the treated unit in the post-intervention period. And for the units in the donor pool the (Tpost × J) matrix YJ contains the values of the outcome variable for all countries in the pool. The estimation of the intervention effect is now given by the difference between the Tpost-values of the outcome variable of the synthetic control group and those of the treated unit. So for any point in time t > Tpre the synthetic control estimator (SCE) can be written as

J X ∗ SCEt = Yi,t − wj Yj,t. (1.8) j=1

In the subsequent analyses of the chapters 2, 3 and 4 the treatment effect of the overall post-intervention period Tpost is calculated as the average treatment effect, i.e. 1 PT (SCE ). Tpost Tpre+1 t One then wants to know whether the observed treatment effect occured by chance. However, the selection of cases is not completely random because the scholar has some leeway in selecting the units for the pool. Furthermore the sample size is rather small and the method does not provide a coefficient of a variable for which a Gaussian distribution in the population can be assumed. Therefore results stemming from the SCM are usually proved by placebo studies. The subsequent analyses rely on cross-unit placebo tests. The intervention is then assigned to and the analysis is conducted for every single country in the donor pool.12 The year of intervention is that of the treated unit. ‘Placebo’ then

1 2 2 10 1 PTpre PJ ∗ RMSPE= Yi,t − w Yj,t Tpre t=1 j=1 j 11The sequence is: 1) find initial v-weights via regression, 2) find optimal w-weights and calculate pre- intervention RMSPE, 3) find alternative v-weights via iteration method, calculate w-weights and pre-intervention RMSPE, 4) repeat step three until pre-intervention RMSPE is minimised. 12An alternative would be in-time placebos. The intervention then is assigned to the treated unit but at

10 means that an intervention is imposed to a country although it obviously did not face one.13 After receiving the results one can then calculate the post-intervention RMSPE as well as the ratio of the post-intervention RMSPE over the pre-intervention RMSPE for every country. The intervention effect can be labeled significant if the treated unit exhibits a post-pre-ratio that is large relative to the ratios of the donor pool countries. The determination of the statistical significance refers to the concept of the p-value as the probability to find a country in the donor pool with a post-pre-ratio of the same size or even larger than that for the treated unit is calculated. This probability can then be compared with a certain significance level (e.g. α = five percent) to evaluate the statistical significance of the result obtained in the analysis.14 In chapters 2 and 3 we rely on this numerical procedure and calculate post-pre-ratios. In chapter 4, however, we rely on a visual procedure to evaluate the statistical relevance of the result. After the execution of the analysis and the placebo studies we calculate the PJ ∗ differences Yi,t − j=1 wj Yj,t for both the treated unit and the comparison units. The subtrahend represents the outcome of synthetic control group and time runs from t = 1, ..., T and thus covers both the pre-intervention period Tpre as well as the post-intervention period Tpost. By means of a chart we can compare the quality of the approximation of the treated unit relative to that of the untreated units. And we can visually evaluate the size of the treatment effect of the treated unit relative to the treatment effects of the comparison units. If the treatment effect of the treated unit is among the largest treatment effects we would qualify the result as economically meaningful. Eventually one can seriously rely on the results only if all distorting influences are excluded. So first of all one shall care for observable as well as unobservable characteristics. The former are basically taken into consideration via the inclusion of predictor variables. With respect to the latter (but also the observable characteristics) Abadie et al. (2010) argue that their influence is smaller the longer the pre-intervention time period is. If it is possible to trace the outcome variable of the treated unit for a long time it can be assumed that this is because the units are similar to each other concerning the aspects that impact on the outcome variable. This holds for both observable and unobservable characteristics. In fact, the length of the pre-intervention period is restricted by data availability. Additionally, the samples are restricted to OECD member countries (chapter 2 and 3) and EU-27 countries plus Norway (chapter 4), respectively. Countries with substantial economic or political structural breaks different from OECD countries thus are excluded.

a point in time where there was no intervention. 13The placebo test for a country j in the pool is conducted without the treated unit. So if we, e.g., run the placebo test for Australia in chapter 2, Japan is not part of the donor pool. 14This signiﬁcance test is rather strong. If there are twenty countries in the pool and the treated unit comes up with the largest post-pre-ratio the utmost signiﬁcance level is 0.05. If, however, the treated unit comes up with the second largest ratio this level jumps up to 0.1.

11 Similarly, the post-intervention period is restricted to about six years to protect the trajectory of the synthetic control group against influences unrelated to the intervention. One particular characteristic that hardly can be measured but is particularly discussed when it comes to Switzerland (see chapter 3) is fiscal conservatism. The reasoning here is that the introduction of the fiscal rule does not cause an improvement of the budget balance by itself because the Swiss citizens and with them the Swiss politicians inherently prefer low spending and slight borrowing. Bohn and Inman (1996) deal with fiscal conservatism in the U.S. states such that they declare the southern states to be fiscally conservative. Furthermore, they include the average percentage of voters who consider themselves to be conservative and declare the Republicans to be more conservative than the Democrats. But the effect of the states’, voters’ or legislators’ (political) conservatism does not turn the effect of the deficit rule into insignificance nor does it reduce the size of the effect considerably. To test for fiscal conservatism Dafflon and Pujol (2001) create an index based on referenda dealing with fiscal affairs regarding the central level submitted to voters in the Swiss cantons. So voters are classified to be conservative if they show a high acceptance rate in favour of tax rate increases, spending cuts or measures that aim at reducing the deficit. They find that the demand for public borrowing is lower in a canton the more conservative the residents are. Consequently, Krogstrup and Wälti(2008) are exclusively testing for voter preferences as omitted variable. In their analysis covering the Swiss cantons the impact of a fiscal rule on the budget balance remains significant even when it is controlled for the fiscal preferences of the voters. Similarly, Funk and Gathmann (2013b) augment their estimation of the effect of direct democracy on government spending by the inclusion of some measures that capture voter preferences. They find that this relationship gets considerably weaker in terms of size and statistical significance once voter preferences are incorporated. Hence it can be questioned what the separate impact of fiscal rules is. Or put differently, we risk to get a treatment effect that is too large if we do not capture fiscal conservatism. More importantly, we would not be able to quantify the real size of the treatment effect as we cannot separate the treatment effect from fiscal conservatism in the post-intervention period. Again, we follow Abadie et al. (2010) in these arguments. If citizens and politicians in Switzerland behave fiscally conservative this should affect the outcome variable in the pre-intervention period. Since this method is based on matching it would assign a positive w-weight to countries that are also fiscally conservative or are at least able to approximate the conservative spending behaviour in Switzerland. This means that the fiscal conservatism of Switzerland is captured by the approximation of the pre-period and the treatment effect can be attributed to the introduction of the fiscal rule.15 Beyond that, Funk and Gathmann (2013b) clearly show that fiscal conservatism

15If, in contrast, Swiss citizens are ﬁscally conservative but the politicians do not act correspondingly this constitutes a classical principal agent problem with too generous spending in the pre-intervention

12 is rather stable over time. In this respect the inclusion of a respective measure would not help to considerably improve the pre-intervention RMSPE as the SC Estimator emulates the variation of the outcome variable over time. Angrist and Krueger (1999) argue that omitting crucial variables would lead to the misspecification of the empirical model. This request is less strong, however, under the Matching Estimator as the covariates mainly contribute to the fulfilment of the homogeneity assumption (Gangl and DiPrete, 2004). Thus we are convinced that our results are unbiased even if fiscal conservatism is not explicitly controlled for. Next, one shall consider reverse causality. Applied to the subsequent analyses this would mean that a certain shape of the treatment effect has led to the intervention. As the treatment effect occurs undoubtedly after the policy intervention one can assume that reverse causality is not a serious obstacle in this context. Finally, it is necessary that there are no spillover effects. This means that there must not be an effect of the intervention on the outcome variable of the treated unit or the donor pool units during the pre-intervention period. The same holds for the outcome variable of any donor pool unit in the post-intervention period. An effect of the intervention on the treated unit’s outcome in the pre-intervention period could come up as an anticipation. However, anticipation effects are captured by the approximation similar to unobservable characteristics or fiscal conservatism in Switzerland. Beyond that, it is not apparent in which way the alteration of the electoral system in Japan, New Zealand or Italy (see chapter 2) or the introduction of the Swiss balanced budget rule (see chapter 3) should impact on the spending behaviour of another OECD member country. Likewise, it is not obvious in which way the implementation of a federalism reform in Flanders or Wallonia (see chapter 4) should impact on economic growth in Aland in Finland, in the Mediterranian region in France or in Baden-Wuerttemberg in Germany. Eventually, Belgium may be too small to have a growth impact on another OECD country.

1.2.4 The Diﬀerence-in-Diﬀerence Estimator

As the name suggests this estimator uses a difference in the outcome twofoldly. It starts from the existence of a group of treated subjects and calculates the average difference in the outcome before and after the intervention (first differentiation). However, the obtained difference might be biased due to influences that are in no relation to the intervention. Thus, another group of subjects is used that does not receive the treatment. For this group of untreated units one can also calculate the average difference in the same outcome variable before and after the (imposed) intervention. As this group does not receive the treatment, the average difference must be attributed to environmental influ-

period. This period does not capture fiscal conservatism because it does not exist on the politicians’ side. The fiscal rule comes up with a separate effect because it forces the politicians to stick to citizens’ preferences.

13 ences that simultaneously impact on both subject groups. This difference is the average counterfactual difference as it shows what would have happened to a treated unit if it did not receive the treatment. Finally, this latter difference gets subtracted from the first difference (second differentiation) and the result is the average treatment effect on the treated group of subjects that seriously can be attributed to the intervention. In terms of the second differentiation the DiD Estimator is related to the Rubin Causal Model. The calculation of the treatment effect can be illustrated by the formula

p p np np τatt = (¯yat − y¯bt) − (¯yat − y¯bt ) (1.9) wherey ¯ is the average outcome of the respective group. The superscripts p and np stand for ’participant’ and ’non-participant’. The subscript bt and at stand for the time period before the implementation of the treatment as well as after its implementation. The treatment eﬀect can easily be implemented in a linear regression model that uses panel data. Such an empirical model can be depicted by

yit = β0 + β1T + β2P + β3TP + it (1.10)

with β0 being the intercept and T being a dummy that takes the value one for outcomes that occur after the policy intervention (zero otherwise). The dummy P takes the value one for outcomes of participants (zero otherwise) and TP is the categorial interaction of T and P . Consequently, the dummy TP takes the value one for outcomes of participants in the post-intervention period. That is, the treatment eﬀect τatt is captured by the coeﬃcient β3. Finally, it is the normally distributed error term. In order to include a vector Xit that comprises further control variables the model can easily be extented to

yit = ψi + ωt + β1TP + β2Xit + it (1.11)

were the dummies T and P are replaced by the unit fixed effects ψi and the time fixed effects ωt. The former interaction term TP now is a dummy that takes the value one for outcomes of participants in the post-intervention period. In this setting the treatment effect τatt is captured by the coefficient β1 (Angrist and Krueger, 1999; Ashenfelter and Card, 1985; Baskaran, 2009).

1.2.5 Discussion

From the previous sections it becomes clear that the three estimators are similar in some respects. Firstly, all three estimators use a counterfactual outcome in order to show what would have happened to a treated unit without the introduction of a policy measure. To this, all estimators need to ﬁnd comparison units that are as similar to the treated unit as possible. That is, the unit homogeneity assumption must hold.

14 Secondly, the Matching Estimator as well as the Synthetic Control Method assign comparison units to the treated unit on the basis of covariates. The SCM additionally selects comparison units regarding their capability to trace the outcome of the treated unit in the pre-intervention period. Likewise, the parallelism of the evolution of the outcome variable needs to be checked under the DiD Estimator, as it uses panel data (common trend assumption). This parallelism is required to hold for the pre-intervention period as well as the post-intervention period. Thirdly, all three estimators use data of the pre-intervention period to some extent. Ei- ther to conduct a pre-post-comparison (DiD Estimator and DiD-Matching Estimator) or to approximate the trajectory of the treated unit’s outcome (Synthetic Control Method). Beyond that, the Matching Estimator (when based on oversampling) and the Synthetic Control Method are similar to each other such that both use a weighted average of comparison units’ outcome to construct the counterfactual outcome. Fourthly, the estimators cannot grip the bias caused by a superimposing treatment. A superimposing treatment is a treatment B that coincides with the treatment A under scrutiny. Eventually, it is the researcher’s task to carefully examine the existence of such superimpositions. Finally, all three estimators build on the stable unit-treatment value assumption (SUTVA) such that none of them can cope with spillover effects. Again, it is the scholar’s task to examine the existence of such effects. The discussion of differences starts with the separation between parametric and non- parametric estimators. Apparently, the DiD Estimator belongs to the former group where the Matching Estimator and the Synthetic Control Method belong to the latter. That is, the former assumes a linear relation between the dependent variable and explaining variables while the Matching Estimator and the Synthetic Control Method do not. This is all the more relevant as the DiD Estimator aims at answering the causal question also for observations beyond the sample, i.e. the population (Angrist and Krueger, 1999; Gangl and DiPrete, 2004). Due to the missing functional form, the non-parametric estimators (DiD-)Matching and SCM are able to create a counterfactual outcome for a treated unit only if there indeed exists a contrastable untreated unit. If such a unit cannot be found in the sample, matching is not possible. That is, the condition of common support is limiting (Rubin, 1977; Todd, 2008). As both the treated unit and the untreated unit need to exhibit the same value of a covariate x, finding a contrastable unit is the more complicated the more covariates are used. In this respect the propensity score is a helpful device as it reduces the dimensionality of the assignment mechanism considerably. However, matching is still impossible, if for a treated unit i with a propensity score si an untreated unit j with the same propensity score sj = si cannot be found. Of course, unit i can be contrasted with a unit j that exhibits the value sj next to si. However, this match can be bad if the distance ksi − sjk is large. Thus it is useful to define a region of common support. Observations

15 outside this region are excluded from the analysis (Caliendo and Kopeinig, 2008). In that, defining a region of common support is in line with the unit homogeneity assumption. Heckman et al. (1997) show that bias stemming from going beyond the common support can be reduced by restricting the region of common support. Basically it holds for the DiD Estimator, the Matching Estimator and the Synthetic Control Method that all observable covariates that are both in relation to the treatment D and to the outcome Y need to be included in the analysis.16 Though, this requirement is valued to be most important for the parametric DiD Estimator. Otherwise, the estimated treatment effect suffers from an omitted variable bias. This is, because the covariates

Xit have a causal interpretation in the linear regression model. Due to this causality, the covariates need to be strictly exogenous under the DiD Estimator. That is, Cov(Xit, it) = 0 must hold. This requirement is less strong under the Matching Estimator and the Synthetic Control Method. The Rubin Causal Model directly attributes a difference in the outcome Y to the treatment D without a causal interpretation of some covariates. Here the covariates mainly contribute to the fulfillment of the unit homogeneity assumption. Of course, covariates used in an empirical model that is based on the Matching Estimator must not be endogenous such that they are a consequence of the treatment (Angrist and Krueger, 1999; Gangl and DiPrete, 2004). What is left, if all observable characteristics are captured, is the potential bias stemming from unobservable characteristics. In this, the three estimators are equal as they all depend on the CI assumption. However, the DiD Estimator is able to capture time-invariant unobservable effects. For this it is not necessary to (precisely) know the unobservable effects (Angrist and Krueger, 1999). This also holds for the Matching Estimator if it comes in as DiD-Matching. The cross-unit Matching Estimator can cope with unobserved characteristics in terms of self-selection as long as the average treatment effect on the treated (ATT) is measured. Heckman et al. (1997) show that the weaker CI asummption

0 Yj ⊥⊥ D | X (1.12) suffices then. As the ATT captures the effect of the treatment ’only’ for the treated units in the sample it is without problems if the ATT covers unobserved characteristics provided that they are inseparable related to the treated units in the sample. Self-selection does not cause a problem then. The average treatment effect on the population (ATE), in contrast, also calculates the impact of the treatment on the untreated units in the sample. To this, the treated units are used as counterparts for the untreated units. To get an unbiased estimation, thus, self-selection of the treated units must be ruled out. That is, conditional independence (see equation 1.4) must be assumed. However, as data often is observational the focus is basically on the ATT. The SC Estimator, finally, cannot incorporate dummy

16This follows from the CI assumption.

16 variables to control for unobserved characteristics. Moreover, Abadie et al. (2010) argue, that using time differences would not cancel out unobserved effects. However, they state that tracing the outcome of the treated unit over a long pre-intervention period helps to control for unobserved factors. The reasoning is, that approximating the treated unit’s outcome with data of untreated units over a long time period is possible only if these units are akin with respect to observed as well as unobserved factors. This also includes time-variant unobserved characteristics. Thus, it is appropriate to build the analysis on an extensive pre-intervention period. The DiD Estimator, the Matching Estimator and DiD-Matching, in contrast, cannot cope with time-variant unobservable characteristics. Besides incorporating dummy variables to capture time-invariant unobservable effects, the DiD Estimator can capture self-selection via the inclusion of further covariates if there exists some knowledge of it. So if, for example, individuals self-select into a job training programme because they were unemployed ahead of the programme, including a variable that records unemployment might help. Similarly, variables that capture self-selection can be used to estimate the propensity score under the matching estimator (Heckman et al., 1997). Finally, the setting of the study might suggest that selection into the programme is beyond the control of the participants. In this case the researcher can deny self-selection by means of argumentation. The three estimators considered also differ with respect to the ease of implementation. In that respect I value the SC Estimator to be the easiest one. The implementation of the estimator itself is smooth and it works well also if data availability is limited. The latter aspect matters when it is about capturing a treatment effect at the macro level. If the analysis is set at the country level it is sometimes difficult to get comprehensive national level information for a large number of variables. The DiD Estimator is more data-demanding but also very easy to implement. The implementation of the Matching Estimator is valued to be most sophisticated as there are many aspects to consider. For a comprehensive review of the implementation issues regarding the Matching Estimator, see Caliendo and Kopeinig (2008). In the following chapters the economic effects of several political institutions are examined. Throughout the chapters the Synthetic Control Method is used. This estimator is prefered since it captures unobservable characteristics even when time-variant. Moreover, it is easy to implement and it is less data-demanding. This is relevant as national level institutions are captured. Most important, however, is the fact that it is able to deal with a small number of observations (treated units). That is, in chapter 2 the alteration of the electoral system is examined. The number of electoral system changes is rather small in a worldwide sample. And it is even more relevant with respect to chapter 3 where the economic effect of one single balanced budget rule, the Swiss debt brake, is considered. Chapter 4 uncovers the growth effect of the alteration of the state structure in one single country.

17 2 Do Electoral System Changes Impact on Government Spending?

2.1 Introduction

“At present, the UK uses the ‘ﬁrst past the post’ system to elect MPs to the House of Commons. Should the ‘alternative vote’ system be used instead?”

This question was the centrepiece of a referendum held in Great Britain in May 2011. Through this the British were given the opportunity to vote for a change of their electoral system. Eventually the referendum was unsuccessful, but there was a lively debate about the political consequences of the suggested system change. The political as well as economic consequences of electoral laws are also discussed among scholars. This is relevant as the electoral system is a crucial element of a representative democracy in which the citizens delegate their decision-making power to elected representatives. With respect to political outcomes Lijphart (1990) finds that the plurality rule and the majoritarian electoral rule cause a clearly more disproportional allocation of seats in parliament than the largest remainder (LR) Hare formula or the d’Hondt formula which are both rules of proportional representation.1 Additionally, he finds that the seat allocation is more proportional the larger the district magnitude is. More comprehensive theoretical as well as empirical analyses of the political consequences of electoral rules are provided, e.g., by Lijphart (1994) or Taagepera and Shugart (1989). Economists are instead interested in the effects on economic outcomes like the overall level of government spending or the relative scope of welfare spending. A considerable effort in discovering such effects was undertaken by Torsten Persson and Guido Tabellini. As a result of the utilisation of a probabilistic voting model in which candidates focus on marginal districts they state that majoritarian electoral rules come up with more geographically targeted spending compared to proportional electoral rules (Persson and Tabellini, 1999). Because of the strong link between a candidate and the local electorate established by the electoral formula itself, deputies support spending programmes that directly benefit their constituency. Under proportional representation, in contrast, the candidate-party link is stronger and candidates support spending that appeals to social rather than geographical groups.

1An allocation of seats in parliament is proportional, if the seat share of each party as percentage of all seats in parliament coincides with the vote share of each party as percentage of all votes.

18 With respect to the overall level of spending, the effect of electoral rules is indirect. As Duverger’s law states, majoritarian electoral rules tend towards two-party competition at the local level. This might eventually result in a two-party political system at the national level. The relationship is to some extent confirmed by Lijphart (1990). For 20 western democracies he finds the average effective number of elective parties at the national level to be 1.12 higher under proportional rule than under majority rule. The number of parties in parliament and in government thus tend to be smaller under majoritarian rule. Accordingly, party systems that are based on the proportional electoral rule are associated with coalition governments where party systems based on the majoritarian electoral rules are associated with single-party governments. From a theoretical model Persson et al. (2007) thus conclude that government spending is higher under proportional rule as coalition governments induce electoral competition among parties. Every party in the coalition tries to implement spending programmes that benefit their clientele. Empirical studies undertaken by Persson and Tabellini (2003, 2004) that capture the reduced-form relationship support this idea and indicate that governments that run under majoritarian electoral rules exhibit a level of overall spending that is about five percentage points of GDP lower than governments that run under proportional electoral rules. Regarding the composition of government spending they find that governments that run under majoritarian electoral rules exhibit a share of social spending as percentage of overall spending that is two or three percentage points of GDP lower than governments that run under proportional electoral rules. The results are statistically significant. Pers- son and Tabellini (2002) confirm the result on overall spending by using non-parametric matching estimators instead of the ordinary least square (OLS) estimator. Persson (2002) confirms the result regarding the effect on social spending but cannot confirm the effect on overall spending. By incorporating more countries in the sample and using more recent data, Blume et al. (2009) confirm the results provided by Persson and Tabellini (2003) regarding the effect on overall spending. Milesi-Ferretti et al. (2002) apply a model of strategic delegation that leads to the same theoretical hypotheses. For OECD countries they find that proportionality raises overall spending as well as transfer spending. In line with the idea that the effect on overall spending is indirect, Persson et al. (2007) provide empirical evidence that coalition governments are more common under proportional rule and that coalition governments exhibit a higher level of government spending compared to single-party governments. They conclude that the effect of the electoral rule on the level of government spending is driven by the type of government. Similarly, some studies find (weak) evidence that government expenditure and transfers are the higher the more parties are part of the coalition in office (Kontopoulos and Perotti, 1999; Perotti and Kontopoulos, 2002; Volkerink and de Haan, 2001). With respect to social spending, Funk and Gathmann (2013a) as well as Iversen and Soskice (2006) find that left-wing parties prevail under proportional representation and

19 on grounds of their preference for redistribution, proportional representation is associated with a higher level of welfare spending. For local governments in Sweden, Pettersson- Lidbom (2008) uses a regression-discontinuity design and finds that overall spending is at a higer level under left-wing party control. Acemoglu (2005) points to methodological aspects and argues that the OLS as well as the matching estimates provided by Persson and Tabellini (2003) cannot definitely rule out the influence of unobservable variables. Thus a reduced-form OLS estimation may fail to capture the causal effect if the relation between the explaining variable of interest and the explained variable is indirect and persistent. In order to capture the causal effect of electoral systems on government spending it thus would be better to focus on the change of the electoral rule and compare spending before and after the alteration for the same observations. At best, nothing but the electoral system changes and the omitted variables bias should be minimised. Consequently, Funk and Gathmann (2013a) apply the difference-in-difference estimator to uncover the effect of electoral system changes on government spending in Switzerland at the cantonal level. They find that the adoption of proportional representation increases welfare spending and decreases spending on roads. They cannot find an effect on the size of government. Bordignon and Monticini (2012) use bootstrap estimation to test the effect of the change of the electoral rule from proportional representation to a mixed-member majoritarian system in Italy in 1993/94 at the national level. They find that the number of parties in government increased significantly which is counter to expectations derived from theory. The number of parties in parliament increased insignificantly. While Bordignon and Monticini (2012) focus on political outcomes, Funk and Gath- mann (2013a) capture the subnational level. But there is no study that captures the alteration of electoral systems at the national level. This is reasonable, as the number of electoral system changes at the national level is rather small. Bormann and Golder (2013) in their dataset report roughly 50 electoral system changes in a worldwide sample covering 65 years. Most of the East European countries introduced a new electoral system together with democratisation in the early 1990s. However, the previous electoral systems did not meet democratic standards. The cases of, e.g., the Czech Republic, Esto- nia, Hungary, Latvia, Poland, the Slovak Republic or Slovenia can thus not be exploited. Beyond that, the application of the difference-in-difference estimator requires the changes of all observations (cases) among the treated units to move in the same direction, e.g., from a majoritarian rule towards a (more) proportional system. Japan and New Zealand both switched this way in 1996.2 While Japan switched from the Single Non-Transferable Vote (SNTV) to the Mixed-Member Majoritarian (MMM) System, New Zealand switched from Plurality Rule to the Mixed-Member Proportional (MMP) System. Italy, however,

2All dates that indicate an electoral system change refer to the time of the ﬁrst election under the new rule and might diﬀer from the date of the adoption the electoral law.

20 switched from an Open List Proportional Rule to the Mixed-Member Majoritarian Rule (with vote linkage), and thus in the opposite direction, in 1994. In 2006, Italy switched (back) to (closed) List Proportional Rule that is accompanied by a majority bonus for the leading party. Albania switched from the Mixed-Member Proportional Rule to the Mixed-Member Majoritarian Rule in 1996 and switched back in 2001. Thus these electoral system changes cannot be investigated together in a difference-in-difference setting. Fac- toring in all these constraints reduces the number of exploitable electoral system changes considerably. In order to investigate the economic effects of electoral system changes on the overall level of government spending as well as on social spending at the national level I apply the Synthetic Control Estimator. This method allows for the analysis of only one or a few treated units. The alterations under scrutiny are the aforementioned changes of the electoral rule in Japan (1996), New Zealand (1996) and Italy (1994, 2006). The chapter proceeds as follows: section 2.2 gives information concerning the data I use. Sections 2.3, 2.4, 2.5 and 2.6 present the empirical analyses. Each section starts with a short description of the institutional setting in the respective country before and after the system change. Section 2.7 concludes.

2.2 Data

To capture the effect on the overall level of government spending in New Zealand and Italy (1994), I use total outlays (disbursements) of general government as percentage of GDP. This includes current as well as capital disbursements. Current outlays consist of current consumption, transfer payments, subsidies and property income paid. This also includes interest payments. The use of total outlays instead of central government expenditure is mainly for data availability reasons. For some countries in the donor pool, the latter is available from 1995 onwards. In case of the utilisation the pre-intervention period would thus entail only one observation for these analyses. To investigate the effect on the overall level of spending in Italy in 2006, I use total central government expenditure as percentage of GDP. This variables includes intermediate consumption, compensation of employees, subsidies, interest payments, taxes, social benefits and social transfers in kind, current transfers and capital transfers (payable), adjustments for the net equity of households in pension funds reserves as well as gross capital formation and net acquisition of non-financial non-produced assets. From a theoretical perspective it is appropriate to use central government data rather than general government expenditure or total outlays, as the focus is on the election of the national assembly. While Persson and Tabellini (2004) use central government spending, Milesi-Ferretti et al. (2002) use data from the general government level. Where Blume et al. (2009) report significant results based on central government data but insignificant

21 results when using general government data, Perotti and Kontopoulos (2002) cannot find a substantial difference in the significance of their results when using general instead of central government data. Concerning the overall level of spending in Japan, total outlays increased steadily in the 1990-95 period while the potential comparison units did not experience such an increase. Hence, the approximation in the pre-intervention period is not possible to a satisfactory extent and the pre-intervention RMSPE is fairly large. Thus I refrain from investigating the impact of the electoral system change on the level of overall spending in Japan. To uncover the effect on the composition of government spending, the dependent variable is social expenditure of the public sector as percentage of GDP. In its all-encompassing definition it covers old-age pensions, survivor pensions, incapacity-related benefits, health spending, family benefits, active labour market policy, unemployment benefits, housing and others. This variable covers cash benefits as well as in-kind transfers and refers to the general government level. In the case of Japan I also use old-age expenditures of the public sector as percentage of GDP. This variable covers public old-age pensions, early-retirement pensions, home-help for the elderly and residential services for the elderly (Adema et al., 2011). In selecting variables that are good predictors of government spending I follow the literature. In reference to Wagner’s Law I include the natural logarithm of GDP per capita. The law supposes government spending to evolve in line with the (economic) development of the society. The sum of imports of goods and services plus exports of goods and services as percentage of GDP is included to capture the extent of trade activity. This is in line with Rodrik (1998) who finds a positive relationship between the economy’s exposure to international trade and government spending. Government spending is implemented to offset the losers of globalisation. Furthermore, several measures of population characteristics are included. The percentage of people aged 65 and older in the total population is considered since elderly people are more in need of health spending. The percentage of people aged 14 and under in the total population is taken into account as young people receive child benefits. Likewise, the natural logarithm of total population is considered. To capture further economic conditions I incorporate the rate of unemployment, unemployment of people aged 15 to 24 (as percentage of total labour force, aged 15 to 24) as well as the density of trade unions (share of wage and salary earners who are trade union members in the total number of employees). In recognition of the fact that countries may differ in some sector-specific characteristics, I use patent applications of residents (per million of population) and high-technology exports (as percentage of manufactured exports). Finally, I take account of employment in the three main economic sectors: agriculture, industry and services (as percentage of total employment). These predictor variables are always accompanied by the lagged outcome variable of certain years which often is the first and the last year of the pre-intervention period. Among these predictors I search for

22 a combination that best supports the approximation of the treated unit’s real data. As mentioned already the sample is restricted to OECD member countries. However, Greece, Mexico and South Korea are discarded from the sample as their electoral rule either was altered several times within a relatively short time period or it features pecu- liarities that makes it difficult to deduce clear hypotheses. Greece amended its electoral rule in 1989, 1993 and 2007. The rule used before 1989 was called reinforced proportional representation. This was basically a proportional rule, since voters casted a vote for a closed list of a party or a coalition of parties at the district level. However, it contained majoritarian elements. Seats were allocated in 56 lower electoral districts pursuant to the Droop/Hagenbach-Bischoff method.3 Residual seats were assigned at a second tier (9 major electoral districts) pursuant to the Hare quota. If there were lower electoral district seats unallocated after the second distribution, they were finally assigned to parties at a third tier that comprised one nationwide constituency. At that tier seats were solely allocated to parties or coalitions that participated in the second tier distribution. Eventually, as the Hare quota at the third tier refered to the vote share a party or coalition obtained nationwide, lower electoral district seats were solely allocated to the party or coalition with the largest vote share at the national level. In the two elections of 1989 Greece used a pure proportional system combined with preferential voting at the district level. Under this rule voters casted a vote for a party list and could indicate a preference for one candidate of that party by marking her name with a cross.4 The third tier was abolished and the number of major districts increased to 13. In 1993, Greece mainly restored the reinforced proportional rule that was used before 1989, but preserved preferential voting. A party threshold of three percent of the national votes was introduced. Beyond that, the third tier was reintroduced, but seats were assigned to parties only. Thus coalitions were excluded from the allocation at the third tier. Moreover, parties were required to obtain the most votes at the national level as well as in the respective major district in order to qualify for the seat allocation at the third tier. In 2007, the second and the third tier were abolished. At the one remaining (first) tier, seats were allocated proportionately to the parties or coalitions of parties that polled three percent of the votes at the national level (party threshold). However, the party or coalition with the largest vote share nationwide obtained a bonus of 40 seats. Preferential voting was preserved (Dimitras, 1994; Lamprinakou, n.d.). Mexico applied a mixed-member proportional electoral system from 1988 to 1993. Un- der this rule voters casted one vote for a candidate in 300 single-member districts and casted a vote for a party list in five multi-member districts (200 seats). The winning party was awarded a majority bonus.5 However, the precise design of that bonus de-

3Most districts were multi-member districts. In single-member districts the plurality rule was applied. 4In some districts voters could mark the name of more than one candidate. 5The party that wins most of the single-member district elections is declared winning party.

23 pended on the vote share of that party. If the vote share of the leading party was lower than 51 percent it was awarded 251 seats, which guaranteed an absolute majority. In the case of a vote share between 51 percent and 70 percent, seats of the winning party were assigned proportionately to the vote share. Finally, the seat share of the leading party was restricted to 70 percent (350 seats) even if its vote share exceeded this threshold. In 1991, the terms of the majority bonus were altered. If the winning party received less than 35 percent of the national vote, seat allocation followed a proportional rule. If the winning party received 35 percent of the national vote, it was awarded 251 seats (absolute majority). If the party received a vote share between 36 percent and 60 percent it was awarded 251 seats that assured the absolute majority plus two additional seats for each percentage point above 35 percent. If it received more than 60 percent but no more than 70 percent of the votes it received a number of seats proportional to votes. If it received more than 70 percent of the national vote, the number of seats was limited to 350 (70 percent of the seats in parliament). In 1994, the MMP electoral system was turned into a MMM electoral system and the majority bonus was abolished. However, two limitations of the overall number of seats a party could win (in both tiers) and one exemption from the MMM rule were introduced. A party that gained less than 60 percent of the national vote could not win more than 300 seats (60 percent of the seats in parliament). If a party received a vote share between 60 percent and 63 percent, seat allocation followed a proportional rule.6 Finally, no party could win more than 63 percent of the seats in parliament (315 seats). Since 1997 there are two limitations of the number of seats in place. No party can receive more than 300 seats (60 percent of the seats in parliament) and the share of seats both from the nominal as well as the list tier cannot be higher than eight percentage points above the national vote share (Horcasitas and Weldon, 2003). South Korea used a mixed-member electoral system with linkage between the tiers from 1962 to 1972 and from 1980 to 1996.7 Under this rule voters casted one vote for a candidate in single-member districts. These votes were also used to calculate the seat allocation at the list tier (single ballot). In these two time periods the electoral rule was accompanied by a majority bonus. From 1962 to 1972 the party that received 50 percent of the national vote was awarded two-thirds of the list tier seats. If no party received this absolute majority of votes the party with the simple majority of the vote share received 50 percent of the list tier seats (vote linkage). In 1980 the majority bonus was altered. The party that won most of the district elections received two-thirds of the list tier seats (seat linkage). In 1987 the majority bonus was weakened. The party that won most seats at the nominal tier received a number of seats in parliament that ensured an overall majority. In 1996, the linkage between the tiers was abolished. However, the

6In that small range the rule switched to mixed-member proportional representation but only for the winning party. 7Between 1972 and 1980 South Korea used the SNTV rule to elect its representatives of the national assembly.

24 single ballot was maintained. That is, the seat allocation at the list tier was determined on the basis of each party’s vote share at the nominal tier. In 2003, the dual ballot was introduced. Voters cast one vote for a candidate and one vote for a party list. That is, South Korea uses a standard mixed-member majoritarian electoral system since then (Hicken and Kasuya, 2003; Reilly, 2007). Eventually, the donor pool contains Australia, Austria, Belgium, Canada, Chile, the Czech Republic, Denmark, Estonia, Finland, France, Germany, Hungary, Iceland, Ireland, Israel, Luxembourg, the Netherlands, Norway, Poland, Portugal, the Slovak Republic, Slovenia, Spain, Sweden, Switzerland, Turkey, the United Kingdom and the USA.8 Further countries need to be removed from the pool for data availability reasons. For details see the respective section of the paper.

2.3 The Alteration from the SNTV Rule to the MMM Electoral System in Japan

2.3.1 Institutional Setting

For many years Japan used the Single Non-Transferable Vote to elect the representatives of the national assembly. Under this rule voters casted their vote for one candidate in multi- member districts.9 Hence, candidates faced clear incentives to take a stand for spending that directly favoured their constituency. Moreover, the non-transferability incentivised parties to place more than one candidate in each district.10 This caused intra-party competition and thus further increased the candidates’ appeal to serve their constituency in terms of targeted spending. Due to electoral competition between candidates at the district level campaigning was very expensive in Japan and was considered responsible for the vast extent of corruption in Japan (Reed and Thies, 2003a,b). Japan experienced a long lasting public debate about the causes of the political grievan- ces in which the electoral system was at least partially held responsible. Besides the public claim for reform further aspects have helped the electoral reform to finally push through. On the one hand two party factions (Sakigake and Shinseito) separated from the long time leading Liberal Democratic Party (LDP) and formed independent parties. On the other hand junior politicians within the LDP supported the reform movement as they realised the drawbacks of the SNTV rule, e.g., in terms of the pecuniary campaign efforts (Sakamoto, 1999). Eventually, Japan introduced a mixed-member majoritarian electoral system in 1994. The first election under the new rule took place in October 1996. Under this rule 300 seats

8Treated units are in no case part of the donor pool. 9In Japan the district magnitude amounted to four on average. 10Under non-transferability received votes which are not necessary for a candidate to qualify for a mandate in the district cannot be transferred to another candidate of the same party.

25 are assigned in single-member districts and 200 seats are assigned via list proportional representation in eleven districts. Hence, this system is called two-tier electoral system. While voters cast a vote for one candidate at the nominal tier, they cast a second vote for a party at the list tier. Since there is no linkage between the tiers, the overall seat allocation is not necessarily proportional to the overall vote share a party receives (Reed and Thies, 2003b). The alteration from the SNTV rule to the MMM electoral system brought along a decrease of the district magnitude down to one at the district level as well as the addition of the proportional rule via the introduction of the list tier. Thus this change can be expected to have relaxed the candidate-constituency link and I expect social as well as old-age spending to increase in the wake of this transformation. Unfortunately it is not possible to disentangle the eﬀect of the decrease in the district magnitude from that of the list tier introduction. However, both eﬀects can be expected to lead to an increase in social spending as well as old-age spending.

2.3.2 The Eﬀect on Social Spending

To uncover the eﬀect of the electoral system change in Japan on social spending I use social expenditure as percentage of GDP as dependent variable. To predict social spending I use the natural logarithm of GDP per capita, the share of elderly people in the total population and the density of trade unions. These variables are accompanied by lagged social spending of the years 1990, 1993 and 1995. The pre-intervention period covers the years 1990 to 1995 and the post-intervention period covers the years 1997 to 2002. Due to missing data in the pre-intervention and the post-intervention period, Estonia, Hungary, Israel, the Slovak Republic, Slovenia and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 22 countries. The w-weights are given in table 2.1. Chile and Portugal are used to approximate the development of social spending in Japan. Both countries are rather equally represented. The predictor balance is given in table 2.2. This balance indicates the quality of the approximation in the pre-intervention period for every single predictor variable. The column ‘Treated’ reveals the value of the respective predictor variable for Japan that is averaged over the pre-intervention period. The entire column thus provides the vector

Xi explained in section 1.2.3. Likewise, the column ‘Synthetic’ provides the mean value of the same predictor variable for Synthetic Japan and is related to the matrix XJ . The predictor balance shows that the approximation works very well with respect to lagged social spending. It works well regarding the predictors income, the share of elderly people and the density of trade unions. This is reﬂected by the pre-intervention RMSPE which amounts to 0.295. Figure 2.1 provides the graphical result. In the post-intervention period the trajectory of Japan exceeds that of Synthetic Japan. This indicates that the electoral system change

26 indeed caused an upward shift in social spending. However, the average treatment effect in the post-intervention period amounts to 0.452 percentage points and is thus rather small.11 This might be due to the fact that voters still cast a vote for a candidate. Additionally, the nominal tier and the list tier are not linked under the MMM electoral system. The result of the subsequently conducted placebo tests can be seen in figure 2.2. Japan is located among the countries with a small post-pre-ratio. The probability of finding a country in the donor pool with a post-pre-ratio the size of Japan or even larger is 17/23=0.739. Thus I conclude that the effect of the electoral system change on social spending in Japan is statistically insignificant.

2.3.3 The Eﬀect on Old-Age Spending

The share of elderly people in the total population of Japan increased steadily in the last centuries. It reached 18 percent in 2002, the last year of the post-intervention period. This is illustrated in figure 2.3 where the label ‘OECD Countries’ covers all OECD member countries as of 2010 (except for Japan). Consequently, the social group of the elderly is an important clientele when it comes to electoral competition between the parties. Old-age spending can thus be assumed to be a politically important subcategory of social spending. Beyond that, Synthetic Japan in the previous section consists of Chile and Portugal only. This is reasonable as social spending increased in Chile, Japan and Portugal in the course of the 1990s. In Japan it started with a rate of 11.1 percent in 1990 and ended up with 17.5 percent in 2002. Portugal experienced a very similar increase in social spending. It started with a rate of 12.5 percent and ended up with 20.6%. In Chile it started with a rate of 9.3 percent in 1994 and ended up with 12.7 percent in 1999. Most of the remaining countries in the donor pool exhibit a relatively constant development or even a decrease in social spending.12 This is illustrated in figure 2.4 where the label ‘Pool Countries’ refers to all countries used for the placebo tests of the preceding analysis except for Chile and Portugal. For the donor pool countries it reveals the sidewards movement of the average values of social spending in the range between 20 percent and 23 percent. In order to appreciate the role of old-age spending with respect to the composition of social spending and in order to obtain a Synthetic Italy that consists of a larger number of comparison units I use old-age expenditure as percentage of GDP as dependent variable to uncover the effect of the electoral system change in Japan on old-age spending. To predict old-age spending I use the share of the elderly in the total population, the unemployment rate and the fraction of people employed in the service sector. These

11The distance is larger for the years from 2003 onwards, which is mainly driven by Chile. However, an extended post-intervention period involves the danger of capturing inﬂuences that are unrelated to the intervention. 12While social spending neither increased nor decreased, e.g., in Australia, Austria, Belgium, Germany, Iceland, Luxembourg, Spain, United Kingdom and the USA, it decreased, e.g., in Canada, Finland, Ireland, the Netherlands, Poland and Sweden.

27 variables are accompanied by lagged old-age spending of the years 1990, 1992 and 1995. The pre-intervention period as well as the post-intervention cover the same years as in the previous analysis. The donor pool consists of the same countries used in the prior analysis. The w-weights are given in table 2.3. While Portugal still contributes to Synthetic Japan with a w-weight very similiar in size, Chile does not contribute to Synthetic Japan anymore. Australia, Iceland and Germany are selected for Synthetic Japan instead. Where Australia is represented with a w-weight half the size of Portugal, Iceland and Germany contribute to the synthetic control group with a distinctly smaller w-weight then Aus- tralia. The predictor balance is given in table 2.4. Except for the unemployment rate the approximation of old-age spending in Japan works very well. The pre-intervention RMSPE amounts to 0.055. Figure 2.5 exhibits the graphical result. Old-age spending of Japan exceeds that of Synthetic Japan in every year of the post-period. The effect of the electoral system change on old-age spending is larger than the effect on (overall) social spending. The average treatment effect amounts to 0.937 percentage points. The result of the placebo tests can be see in figure 2.6. Japan stands out with the largest post-pre-ratio among all countries. The p-value equivalent is 1/23=0.044. To test the robustness of this result I eliminate the country with the smallest w-weight (Germany) from the donor pool and rerun the analysis.13 The graph (not shown) looks very similar to the baseline estimation and the pre-intervention RMSPE amounts to 0.103. Running placebo tests again leads to figure 2.7. Again, Japan exhibits the largest post- pre-ratio. The p-value equivalent is 1/22=0.045. Once more I drop the country with the smallest w-weight (Chile).14 The graph (not shown) again looks very similar to the baseline estimation and the pre-intervention RM- SPE amounts to 0.099. Figure 2.8 shows that Japan exhibits the second largest post-pre- ratio. The p-value equivalent amounts to 2/21=0.095. The effect of the electoral system change on old-age spending thus is statistically significant at the ten percent level.

2.4 The Alteration from Plurality Rule to the MMP Electoral System in New Zealand

2.4.1 Institutional Setting

Similar to Japan, New Zealand used its plurality voting system for many years. Under this rule voters casted their vote for one candidate in single-member districts. Furthermore, there was no direct election of the president and no second chamber in place. Thus the

13This leads to the following w-weights: Australia 0.479, Austria 0.072, Chile 0.004, Iceland 0.019 and Portugal 0.425. 14This leads to the following w-weights: Australia 0.453, Austria 0.073, Iceland 0.044 and Portugal 0.431.

28 electoral rule gave rather unrestricted power to one of the two major parties, Labour and National, and to their governments. The governments’ electoral accountability to the voters was thus rather limited. Minority parties in New Zealand were barely represented in parliament as they were not geographically concentrated and as the electoral rule was based on the concept called ‘the-winner-takes-all’. Due to these facts people became increasingly dissatisfied with the political as well as the electoral system. In 1985, the Labour Party installed a Royal Commission to prepare proposals for alternative electoral rules. However, Labour was not in favour of a system change and supported the reform process mainly because of the public claim. In 1986, the Royal Commission proposed the Mixed-Member Proportional System as the best alternative. At that time National began to support the reform process but similarly to Labour, only for strategic reasons. The parties suggested a tentative as well as a binding referendum in the hope that the electorate would lose interest in the topic in consequence of a long lasting reform process. Against the parties’ expectations, however, citizens voted for an electoral system change towards an MMP system in 1993. The first election under the new rule took place in October 1996 (Denemark, 2003). Under the new rule candidates are still running in single-member districts. However, voters cast a second vote for a closed party list. Thus the MMP electoral system also comprises two tiers. The national assembly consists of 120 seats of which 65 seats (54 percent) are filled via the electoral competition at the nominal tier (Massicotte and Blais, 1999). Contrary to the Mixed-Member Majoritarian electoral system the final number of parliamentary seats a party gets is determined by the votes the party polls at the list tier.15 If the number of seats a party wins at the list tier is larger than the number of seats the party obtains at the nominal tier, remaining seats are filled with party list candidates. If, in contrast, the number of seats a party wins at the list tier is smaller than the number of seats the party receives at the nominal tier, these excess seats are granted nonetheless (overhang seats). Furthermore, parties either need to pass a vote share threshold of five percent or need to win at least one district seat to be represented in parliament (Shugart and Wattenberg, 2003). The alteration from plurality rule to the MMP electoral system put a proportional election next to the already existing majoritarian election. Additionally, the overall seat allocation in parliament is determined by the list tier result. Thus this change can be expected to have increased the number of parties in New Zealand. Furthermore, this change can be expected to have strengthened the candidates’ alignment towards their parties. Consequently, I expect the overall level of spending as well as social spending to increase in the wake of this transformation.

15This conjunction is called ‘seat linkage’.

29 2.4.2 The Eﬀect on Overall Spending

To uncover the effect of the electoral system change in New Zealand on the overall level of spending I use general government total outlays as percentage of GDP as dependent variable. To predict overall spending I use the natural logarithm of the total population, the share of the population aged 14 and under in the total population as well as the fraction of people employed in the industry sector. These predictors are accompanied by lagged overall spending of the years 1993 and 1995. Just as the development of total outlays in Japan (see section 2.2) the development of total outlays in New Zealand stands out among the group of OECD countries. As can be seen in figure 2.9 total outlays from 1990 until 1996 decreased continuously in New Zealand. Total outlays also decreased in Canada and the USA but not before 1992. In Ireland it did not decrease before 1993. The average of the remaining donor pool countries moved sidewards in that time. So approximating total outlays in New Zealand is hardly possible in the early years of the 1990s. Starting the pre-intervention period in 1990, 1991 or 1992 results in fairly large values of the pre-intervention RMSPE. Thus the pre-intervention period covers the years 1993 to 1995 and the post-intervention period covers the years 1997 to 1999. Due to missing data in the pre-intervention and the post- intervention period, Chile, Estonia, Hungary, Israel, Poland, the Slovak Republic, Slovenia and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 20 countries. The w-weights as well as the predictor balance are given in table 2.5 and table 2.6, respectively. Synthetic New Zealand consists of Canada, Ireland and the USA. All countries are rather equally represented. The approximation in the pre-intervention period works very well which is reflected by the pre-intervention RMSPE which amounts to 0.512. Figure 2.10 provides the graphical result. In the post-intervention period total outlays of New Zealand clearly exceed total outlays of Synthetic New Zealand. The average treatment effect amounts to 3.36 percentage points. This result receives support from Barker et al. (2003) who state that the effective number of parliamentary parties is higher under the MMP electoral system than under the plurality rule in New Zealand.16 The result of the subsequently conducted placebo tests can be seen in figure 2.11. New Zealand exhibits the fourth largest post-pre-ratio. The probability of finding a country in the donor pool with a post-pre-ratio the size of New Zealand or even larger is 4/21=0.19. Rerunning the analysis with an extended post-intervention period that covers the years up to 2002 results in a post-pre-ratio that amounts to 6/21=0.286.17 Thus I conclude that the effect of the electoral system change on the overall level of spending in New Zealand

16Nishikawa and Herron (2004) in a sample of 53 countries and a timespan covering the years 1990 to 2001, however, can ﬁnd only vague evidence that the number of legislative parties under mixed-member proportional representation is higher compared to plurality. 17However, an extended post-intervention period involves the danger of capturing inﬂuences that are unrelated to the intervention.

30 is statistically insigniﬁcant.

2.4.3 The Eﬀect on Social Spending

To explore the effect of the electoral system change in New Zealand on social spending I use social expenditure as percentage of GDP as dependent variable. To predict social spending I use the unemployment rate and the density of trade unions. These predictors are accompanied by lagged social spending of the years 1993 and 1995. Similar to the overall level of spending the development of social expenditures in New Zealand cannot be approximated adequately if the pre-intervention period starts in 1990, 1991 or 1992. Thus the pre-intervention period covers the years 1993 to 1995 and the post-intervention period covers the years 1997 to 1999. Due to missing data in the per-intervention period and the post-intervention period Estland, Hungary, Israel, the Slovak Republic, Slovenia and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 22 countries. The w-weights are given in table 2.7. All countries of the pool are used to approximate the trajectory of social spending in New Zealand. However, Canada stands out with a relatively large w-weight of 0.55. The predictor balance is given in table 2.8 and indicates that the approximation works very precise. This is reflected by the pre-intervention RMSPE which amounts to 0.019. Figure 2.12 exhibits the graphical result. In the post-intervention period Synthetic New Zealand exhibits a downward trend of social spending. New Zeland also exhibits a downward trend from 1998 to 1999 but an abrupt increase in social spending right after the introduction of the mixed-member proportional system in 1997 and 1998. The average treatment effect amounts to 2.082 percentage points. The result of the placebo tests can be seen in figure 2.13. New Zealand stands out with the largest post-pre-ratio. The p-value equivalent is 1/23=0.044. To test the robustness of the result I eliminate the countries with the smallest w-weights (Denmark, Finland, Spain and Sweden) from the donor pool and rerun the analysis.18 The graph (not shown) looks very similar to the baseline estimation and the pre-intervention RMSPE amounts to 0.013. Running placebo tests again leads to figure 2.14. Again, New Zealand exhibits the largest post-pre-ratio and the p-value equivalent amounts to 1/19=0.053. Once more I drop the countries with the smallest w-weights (Ireland and Poland).19

18This leads to the following w-weights: Australia 0.003, Austria 0.002, Belgium 0.002, Canada 0.521, Chile 0.004, Czech Republic 0.011, France 0.002, Germany 0.002, Iceland 0.027, Ireland 0.001, Luxem- bourg 0.004, the Netherlands 0.121, Norway 0.003, Poland 0.001, Portugal 0.004, Switzerland 0.084, United Kingdom 0.002 and the USA 0.207. 19This leads to the following w-weights: Australia 0.003, Austria 0.003, Belgium 0.003, Canada 0.604, Chile 0.004, Czech Republic 0.007, France 0.002, Germany 0.003, Luxembourg 0.004, the Netherlands 0.051, Norway 0.004, Portugal 0.004, Switzerland 0.215, United Kingdom 0.003 and the USA 0.089.

31 The graph (not shown) again looks very similar to the baseline estimation and the pre- intervention RMSPE amounts to 0.022. Figure 2.15 shows that New Zealand once more exhibits the largest post-pre-ratio. The p-value equivalent is 1/17=0.059. The eﬀect of the electoral system change on social spending in New Zealand thus is statistically signiﬁcant at the six percent level and robust.

2.5 The Alteration from ListPR to the MMM Electoral System in Italy in 1994

2.5.1 Institutional Setting

Until 1993 italian voters were used to elect their representatives via an open list proportional electoral rule, this tradition having lasted more than 30 years. Under this rule voters casted their vote for a party list in 31 electoral constituencies and then indicated the prefered order of party candidates within that list.20 That is, the electoral rule in Italy was proportional in general but provided a deviation from the pure proportional concept in terms of preferential voting in that time. Similiar to the electoral rule used in Greece in 1989 (see section 2.2), candidates acted in a trade-off between fostering the party line and obtaining a personal reputation.21 Consequently, this ordering of candidates added intra-party competition between opponents to the electoral competition between parties. Katz and Bardi (1980) state that this caused the formation of party factions and with this a rather fragmented party system. Similar to Japan a noticeable level of corruption was observed and the electoral system was at least partially held responsible for these occurrences. Although citizens increasingly demanded for a reform, parties in the parliament were not able to find an agreement because interests were too contradictory. In the end, the pressure to reform came from outside the governing parties, namely from Mario Segni, a former member of the Christian Democratic Party. Similar to New Zealand a Committee for Electoral Reform was established to prepare proposals for a referendum. Since the initiation of a referendum in no way depends on the approval of the government in Italy, this was a good chance to impose reform progress on the leading political actors. As several referenda were successful the parliament was forced to legislate towards a new electoral rule taking into consideration the proposals submitted by the committee (Donovan, 1995). Eventually, Italy introduced a mixed-member majoritarian system in 1993. The first election under the new rule took place in march 1994. As it is typical for mixed-member electoral systems voters cast two votes. That is, 155 out of 630 seats in parliament are assigned via proportional representation where 475 candidates (75 percent) are elected in

20The district magnitude in Italy ranges from four to 53. 21A ranking of electoral rules regarding their incentive to foster a ‘personal vote’ is provided by Carey and Shugart (1995).

32 single-member districts at the nominal tier. However, there is a linkage of votes between the two tiers as the calculation of the vote share of party i at the list tier does not rely on the total number of votes the party received at that tier. The procedure is represented by the equation lt lt fl Veff = Vtot − V + 1 (2.1)

lt where Vtot is the total number of votes party i polled at the list tier, however gathered at the constituency level. If the candidate ci of party i carried her election in the single- member district, this total number of votes is diminished by the number of votes the candidate ci needed in order to win her nominal tier election. That is the number of votes V fl the first loser polled in the nominal election plus one vote.22 The total number of votes of party i at the district level adjusted by the winning margin of the candidate ci lt gives the effective number of votes Veff of party i at the district level. The total number of effective votes of all districts is then used to calculate the vote share of party i at the list tier via the Hare quota. (Katz, 1996, 2003; Massicotte and Blais, 1999; Nunez, n.d.). The alteration from the open list proportional representation to the mixed-member majoritarian electoral system put a direct election of candidates next to the already existing proportional election. Thus this change can be expected to have led to a decrease in the number of parties (which was explicitly an objective of the reform). This is in line with Reed (2001) who in comparing the 1994 election with the 1996 election finds that the number of candidates in the nominal tier decreased down to two (Duverger’s law). Furthermore, this change can be expected to have increased the candidates’ alignment towards their local electorate. Consequently, I expect the overall level of spending as well as social spending to decrease in the wake of this transformation.

2.5.2 The Eﬀect on Overall Spending

To uncover the eﬀect of the 1994 electoral system change in Italy on the overall level of spending I use general government total outlays as percentage of GDP as dependent variable. To predict total outlays I use the share of the population aged 14 and under in the total population as well as the fraction of people employed in the service sector. These variables are accompanied by lagged outlays of the years 1989, 1991 and 1993. The pre-intervention period covers the years 1989 to 1993 and the post-intervention period covers the years 1995 to 1999. Due to missing data in the pre-intervention and the post- intervention period, Chile, the Czech Republic, Estonia, Hungary, Israel, Luxembourg, Poland, the Slovak Republic, Slovenia, Switzerland and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 17 countries. The w-weights are given in table 2.9. Synthetic Italy consists of Austria, Finland,

22To keep the description simple the case in which a candidate is aﬃliated with more than one list is disregarded.

33 Norway and Sweden. Austria stands out with a w-weight of 55 percent and Norway contributes to Synthetic Italy with a w-weight of almost 38 percent. The predictor balance is given in table 2.10. It indicates that the approximation in the pre-intervention period works very well. This is reflected by the pre-intervention RMSPE which amounts to 0.293. Figure 2.16 provides the graphical result. In the post-intervention period the trajectory of Synthetic Italy exceeds that of Italy in every year. This indicates that the 1994 electoral system change in Italy reduced the overall level of spending. However, the average treatment effect amounts to 2.13 percentage points and is thus rather small. The result of the subsequently conducted placebo tests can be seen in figure 2.17. Italy exhibits the fourth largest post-pre-ratio. The probability of finding a country in the donor pool with a post-pre-ratio the size of Italy or even larger is 4/18=0.222. Thus I conclude that the effect of the 1994 electoral system change on the overall level of spending in Italy is statistically insignificant. This indicates that the two-candidate competition at the district level did not push through to the national level and is very much in line with Bordignon and Monticini (2012) who find that the number of parties in the coalition governments as well as in the parliament increased in the wake of this electoral system change. Similarly, Ferrara and Herron (2005) argue that small parties form a coalition with larger parties in mixed- member electoral systems if voters cast two votes (dual ballot) and if the tiers operate separately. Hence the number of parties does not decrease due to a shift of the electoral system from a proportional rule towards a mixed-member one.

2.5.3 The Eﬀect on Social Spending

To uncover the eﬀect of the 1994 electoral system change in Italy on social spending I use social expenditure as percentage of GDP as dependent variable. To predict social spending I use the share of elderly people in the total population as well as the unemployment rate. These variables are accompanied by lagged social spending of the years 1990 and 1993. Looking at social expenditure data of Italy in the 1980s reveals that social spending increased slightly from 20.5 percent in 1984 to 21.2 percent in 1989, fell to 19.9 percent in the following year and increased again until 1993 (20.9 percent). As this unique development cannot be approximated adequately by the potential comparison units, the pre-intervention period covers the years 1990 to 1993. Due to missing data in the pre-intervention and the post-intervention period, Estonia, Hungary, Israel, the Slo- vak Republic and Slovenia had to be removed from the sample described in section 2.2. The donor pool thus contains 23 countries. The w-weights are given in table 2.11. Synthetic Italy consists of Luxembourg, the Netherlands and Norway. Luxembourg stands out with a w-weight of almost 79 percent and Norway contributes to Synthetic Italy with a w-weight of 21 percent. The predictor balance is given in table 2.12. Except for the unemployment rate the approximation of

34 social spending in Italy works very well. The pre-intervention RMSPE amounts to 0.09. Figure 2.18 exhibits the graphical result. In the first year of the post-intervention period Italy exhibits a clear decrease of social spending compared with Synthetic Italy. In 1996, however, social spending rockets and it further increases in the subsequent years. Obviously, such a development can hardly be attributed to a change of the electoral system. In fact, Italy passed a major pension reform in 1995 that replaced the public defined-benefit pension scheme by a defined-contribution pension plan. While pension payments were based on benefits with a weak link to payments of contributions under the former rule, pension payments depend much more heavily on contributions since then. Basically, pension expenditures can be assumed to decrease in the wake of the alternation from a defined-benefit to a defined-contribution pension scheme. Franco and Sartor (2006), however, argue that the opposite can occur in the transition from the former to the latter. This is because a remarkable increase of the contribution rate would be necessary in order to keep the level of pension benefits. In a framework where contributions are detracted from wages this would mean a substantial increase in labour costs. Thus the pension reform probably caused an increase of old-age spending in the short run since the public sector beared that cost. To uncover the effect of the 1994 electoral system change on social spending in Italy, I consequently restrict the calculation of the average treatment effect to the year 1995. It amounts to 1.533 percentage points. The result of the placebo tests can be seen in figure 2.19. Italy exhibits the second largest post-pre-ratio. The p-value equivalent is 2/24=0.083. To test the robustness of this result I eliminate the country with the smallest w-weight (the Netherlands) from the donor pool and rerun the analysis.23 The graph (not shown) looks very similar to the baseline estimation and the pre-intervention RMSPE amounts to 0.09. Running plabeo tests again leads to figure 2.20. Italy stands out with the largest post-pre-ratio and the p-value equivalent is 1/23=0.044. Once more I drop the country with the smallest w-weight (Norway) and rerun the analysis.24 Except for the year 1992 in which the approximation is a little less precise the graph (not shown) looks very similar to the baseline estimation and the pre-intervention RMSPE amounts to 0.14. Figure 2.21 shows that Italy exhibits the second largest post- pre-ratio. The p-value equivalent is 2/22=0.091. The effect of the 1994 electoral system change on social spending in Italy is thus statistically significant at the ten percent level.

23This leads to the following w-weights: Luxembourg 0.786 and Norway 0.214. 24This leads to the following w-weights: Belgium 0.124 and Luxembourg 0.876.

35 2.6 The Alteration from the MMM Electoral System to ListPR in Italy in 2006

2.6.1 Institutional Setting

The mixed-member majoritarian electoral system contributed to the consolidation of the party system in the sense that two coalitions competed for political power. However, these coalitions consisted of a fairly large number of parties and were thus rather heterogenous. Due to the existence of the list tier small parties could aﬀord to keep their sovereignty. The party system was thus heavily fragmented under the MMM electoral rule. The debate on returning to a proportional electoral rule thus never stopped and led to the revision of the electoral rule (Bull and Pasquino, 2007). Beyond that, the introduction of a new electoral system that would give a premium of votes to the winning party was just in the interest of the parties of the governing centre- right coalition ‘House of Freedoms’ (Baldini, 2011). This coalition lost ground in the 2005 regional elections and was faced with negative polls in the run-up to the forthcoming general election in 2006. While the Union of Centre (one party of that coalition) was in favour of a return to proportional representation since it hoped to become the third largest party, Forza Italia (another coalition partner) was in favour of a return to proportional representation as it underperformed at the nominal tier of the MMM electoral system due to its geographical dispersion in the country. Finally, the government headed by Silvio Berlusconi altered the electoral system thanks to its parliamentary majority and introduced a proportional electoral rule in 2005. Un- der this rule voters cast their vote for a closed party list in 26 multi-member districts. Preference voting was not reintroduced. The electoral rule is equipped with a majority bonus assigned to the leading party or coalition. Tat is, the party or coalition of parties that polls the plurality of votes is immediately awarded 340 seats of the assembly, which amounts to 55 percent. Additionally, several party thresholds were adopted. Coalitions of parties need to have at least ten percent of the national vote to be represented in parliament. Parties that are part of a coalition need to have at least two percent of the national vote to be represented in parliament. Parties that are running on their own, i.e. without being part of a coalition, need to have at least four percent of the national vote to be represented in parliament (Nunez, n.d.). Mirror-inverted to the previous electoral system change in Italy the alteration from the MMM electoral system to list proportional representation removed the nominal tier and thus the election of candidates. Thus this change can be expected to have strengthened the candidates’ alignment towards their parties. Furthermore, small parties are not forced to participate in coalitions anymore. Consequently, I expect the overall level of spending as well as social spending to increase in the wake of this transformation.

36 2.6.2 The Eﬀect on Overall Spending

To uncover the effect of the 2006 electoral system change in Italy on the overall level of spending I use total central government expenditure as percentage of GDP as dependent variable. To predict overall spending I use the natural logarithm of GDP per capita, the share of elderly people in the total population and the unemployment rate. These predictors are accompanied by lagged overall spending of the years 2001, 2004 and 2005. Total central government spending steadily decreased in Italy from 38.6 percent in 1994 to 26 percent in 2000. It increased in 2001 up to 27.8 percent and remained rather constant until 2005 (26.7 percent). As this development cannot be approximated adequately by the potential comparison units, the pre-intervention period covers the years 2001 to 2005. The post-intervention period covers the years 2007 to 2011. Due to missing data in the pre- intervention and the post-intervention period, Chile, Israel and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 25 countries. The w-weights are given in table 2.13. Synthetic Italy consists of Belgium, the Slovak Republic, Spain and Sweden. Sweden stands out with a w-weight of almost 58 percent and Spain contributes to Synthetic Italy with a w-weight of 30 percent. Belgium and the Slovak Republic contribute to the approximation with distinctly smaller w-weights. The predictor balance is given in table 2.14 and indicates that the approximation works very well. The pre-intervention RMSPE amounts to 0.068. Figure 2.23 provides the graphical result. In the post-intervention period government expenditures of Italy exceed government expenditures of Synthetic Italy. Moreoever, the trajectories run parallel over the entire period. The average treatment effect amounts to 2.2 percentage points. The result of the subsequently conducted placebo tests can be seen in figure 2.23. Italy stands out with the largest post-pre-ratio. The probability of finding a country in the pool with a post-pre-ratio the size of Italy or even larger is 1/26=0.038. To test the robustness of the result I eliminate the country with the smallest w-weight (the Slovak Republic) from the donor pool and rerun the analysis.25 The graph (not shown) looks very similar to the baseline estimation and the pre-intervention RMSPE amounts to 0.07. Running placebo tests again leads to figure 2.24. Again, Italy exhibits the largest post-pre-ratio. The p-value equivalent is 1/25=0.04. Once more I drop the country with the smallest w-weight (Belgium) from the donor pool and rerun the analysis.26 The approximation in the pre-intervention period works somewhat imprecise. This is reflected by the pre-intervention RMSPE which amounts to 0.167. However, the treatment effect looks very similar to the baseline estimation. Figure 2.25 shows that Italy exhibits the second largest post-pre-ratio. The p-value equivalent is 2/24=0.083. The effect of the 2006 electoral system change on central government expenditures in Italy thus is statistically significant at the nine percent level.

25This leads to the following w-weights: Belgium 0.121, Spain 0.303 and Sweden 0.576. 26This leads to the following w-weights: Spain 0.326 and Sweden 0.674.

37 2.6.3 The Eﬀect on Social Spending

To uncover the effect of the 2006 electoral system change in Italy on social spending I use social expenditure as percentage of GDP as dependent variable. To predict social spending I use the sum of imports and exports as percentage of GDP as well as the fraction of people employed in the industry sector. These predictors are accompanied by lagged social spending of the years 2001, 2003 and 2005. The pre-intervention period covers the years 2000 to 2005 and the post-intervention period covers the years 2007 to 2012. Due to missing data in the pre-intervention and the post-intervention period, Switzerland and Turkey had to be removed from the sample described in section 2.2. The donor pool thus contains 26 countries. The w-weights are given in table 2.15. Synthetic Italy consists of France, Portugal, Slovenia and the United Kingdom. France and the United Kingdom contribute to Syn- thetic Italy with a w-weight of 41 percent each. The predictor balance is given in table 2.16. It indicates that the approximation works very well with respect to lagged social spending. It works fairly well regarding the sum of imports and exports and employment in the industry sector. This is reflected by the pre-intervention RMSPE which amounts to 0.104. Figure 2.26 exhibits the graphical result. In the post-intervention period social spending of Italy exceed social spending of Synthetic Italy. Again, the trajectories run parallel over the entire period. After a small reduction of social spending in the year right after the intervention it rises by about 3 percentage points followed by a sideward movement from 2009 to 2012. The average treatment effect amounts to 0.42 percentage points. The result of the placebo tests can be seen in figure 2.27. Italy is located among the countries with a small post-pre-ratio. The p-value equivalent is 15/27=0.556. Thus I conclude that the effect of the 2006 electoral system change on social spending in Italy is statistically insignificant.

2.7 Conclusions

This paper investigates the eﬀects of electoral system changes in Japan, New Zealand and Italy on the overall level of spending as well as on social spending. In 1996 Japan replaced the Single Non-Transferable Vote by the Mixed-Member Majoritarian System. In the same year New Zealand abolished the Plurality Rule and introduced the Mixed-Member Proportional System. Italy switched from an Open List Proportional Representation to the Mixed-Member Majoritarian System with a vote linkage of tiers in 1994 and turned (back) to (closed) List Proportional Representation with a majority bonus for the leading party in 2006. By using the Synthetic Control Method I ﬁnd that the electoral system changes from a majoritarian rule to a proportional rule in New Zealand and Italy (2006) increased the

38 overall level of spending by 2.2 and 3.36 percentage points. Where the former result is statistically significant at the 9 percent level the latter result is statistically insignificant. The change of the electoral system from a proportional rule to a majoritarian rule in Italy (1994) reduced the level of overall spending by 2.13 percentage points. This result is statistically insignificant. Concerning the composition of government spending I find that the electoral system changes from a majoritarian rule to a proportional rule in Japan, New Zealand and Italy (2006) increased social spending by 0.42, 0.45 and 2.08 percentage points.27 Where the treatment effect in New Zealand is statistically significant at the six percent level the effect on social spending in Japan and Italy (2006) is insignificant. The electoral system change in Japan increased old-age spending by 0.934 percentage points. This result is significant at the ten percent level. The change of the electoral system from a proportional rule to a majoritarian rule in Italy (1994) decreased social spending by 1.53 percentage points, the result being statistically significant at the ten percent level. That is, I cannot find a clear significant effect of the electoral system changes under review on the overall level of spending. I can find a clear significant effect on social spending only in New Zealand. This might be due to the fact that New Zealand switched from a pure majoritarian rule to an almost pure proportional rule. Italy and Japan switched from a pure proportional or a pure majoritarian rule to a mixed-member rule. In that, this paper also contributes to the literature dealing with the effects of mixed- member electoral systems. These results are in contrast to the outcomes provided by Persson and Tabellini (2002, 2003, 2004) with respect to the size of the effect and regarding the statistical significance.28 They find a significant difference between the majoritarian rule and the proportional rule that amounts to five percentage points with respect to the overall level of spending and two or three percentage points regarding the effect on social spending. My results, however, correspond to the results provided by Persson (2002) and Funk and Gathmann (2013a) and lend some support to the argumentation of Acemoglu (2005). As the literature shows the effect of electoral systems on social spending might be indirect and thus caused by a simultaneous shift in the ideology of parties and the government. Funk and Gathmann (2013a) as well as Iversen and Soskice (2006) point to the fact that left-wing parties prevail under proportional representation. If that is true the Synthetic Control Method is not able to disentangle adequately the changes within the party system from the electoral system change. Indeed, the ideological composition of government changed in Japan due to the election in 1996 but the percentage of right- wing parties in government increased with a simultaneous increase in social spending and

27Japan switched from the SNTV to the MMM electoral system. However, the MMM electoral system is more proportional than the SNTV. 28Concerning the eﬀect on the overall level of spending these results are also in contrast to the outcomes provided by Blume et al. (2009).

39 old-age spending. The 1994 electoral system change in Italy caused an increase in the proportion of left-wing parties. Nevertheless social spending decreased slightly in 1995. The 2006 electoral system change in Italy caused a shift from right-wing dominance to left-wing dominance which in 2008 switched back to a right-wing dominance. However, social spending increased in 2007 and 2008. In New Zealand there was no change in party ideology in the time under review (Armingeon et al., 2013). So there is no clear indication that the effect on social spending is driven by government ideology. As these results are derived from case studies, the external validity is limited. Never- theless I would conclude that the chance to observe a sizeable effect of an electoral system change, e.g. like in the case of Great Britain (see section 2.1), on overall spending is rather slight. An effect on social spending might be expected if the superseding electoral rule is much more (dis)proportional than the superseded.

40 2.8 Tables and Figures

State w-Weight State w-Weight Chile 0.465 Portugal 0.535

Table 2.1: Social Expenditure in Japan, w-Weights

Variable Treated Synthetic lnGDP 10.217 9.374 POP65 13.134 10.613 Union density 24.535 22.774 Social Expenditure 1990 11.121 11.291 Social Expenditure 1993 12.6 12.71 Social Expenditure 1995 14.077 13.989

Table 2.2: Social Expenditure in Japan, Predictor Balance 18 16 14 12 Social Expenditure (% of GDP) 10 1990 1995 2000 2005

Year

Japan Synthetic Japan

Figure 2.1: Social Expenditure in Japan

41 Poland Chile Sweden Spain Canada Finland Japan Norway Australia Austria Switzerland Iceland Denmark Netherlands France United Kingdom USA Belgium Portugal Luxembourg Czech Republic Germany Ireland

0 5 10 15 20 Post-Period RMSPE / Pre-Period RMSPE

Figure 2.2: Social Expenditure in Japan, Post-Pre-Ratios 25 20 15 10

Share of Elderly People (% Total Population) 1980 1990 2000 2010

Year

Japan OECD Countries (Average)

Figure 2.3: Share of Elderly People in Japan & OECD Countries 25 20 15 Social Expenditure (% of GDP) 10

1990 1995 2000 2005

Year

Chile Japan Portugal Pool Countries (Average)

Figure 2.4: Social Expenditure in Chile, Japan, Portugal & Pool Countries

42 State w-Weight State w-Weight Australia 0.248 Iceland 0.147 Germany 0.053 Portugal 0.552

Table 2.3: Old-Age Expenditure in Japan, w-Weights

Variable Treated Synthetic POP65 13.134 13.176 Unemployment rate 2.5 6.208 Employment, Service 59.083 59.713 Old-Age Expenditure 1990 4.03 4.024 Old-Age Expenditure 1992 4.33 4.33 Old-Age Expenditure 1995 5.24 5.237

Table 2.4: Old-Age Expenditure in Japan, Predictor Balance 8 7 6 5 Old-Age Expenditure (% of GDP) 4 1990 1995 2000 2005

Year

Japan Synthetic Japan

Figure 2.5: Old-Age Expenditure in Japan

Germany Chile Netherlands Sweden Poland Iceland France Switzerland Norway Belgium Portugal Australia United Kingdom Czech Republic Austria Finland Denmark Ireland Canada USA Luxembourg Spain Japan

0 5 10 15 20

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.6: Old-Age Expenditure in Japan, Post-Pre-Ratios

43 Netherlands Chile Sweden Poland France Iceland Portugal Australia Czech Republic Switzerland United Kingdom Finland Belgium Norway Canada Ireland Spain USA Denmark Austria Luxembourg Japan

0 2 4 6 8 10

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.7: Old-Age Expenditure in Japan (w/o DEU), Post-Pre-Ratios

Netherlands Sweden France Poland Iceland Portugal United Kingdom Australia Finland Czech Republic Switzerland Norway Belgium Canada Ireland Spain USA Denmark Austria Japan Luxembourg

0 5 10

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.8: Old-Age Expenditure in Japan (w/o CHL, DEU), Post-Pre-Ratios 55 50 45 40 35 Total Outlays (% of GDP) 30 1990 1995 2000 2005

Year

Canada Ireland New Zealand USA Pool Countries (Average)

Figure 2.9: Total Outlays in Canada, Ireland, New Zealand, USA & Pool Countries

44 State w-Weight State w-Weight Canada 0.297 USA 0.318 Ireland 0.385

Table 2.5: Total Outlays in New Zealand, w-Weights

Variable Treated Synthetic lnPOP 15.102 17.081 POP14 23.037 22.569 Employment, Industry 24.567 25.206 Total Outlays 1993 45.1 44.79 Total Outlays 1995 41.7 42.026

Table 2.6: Total Outlays in New Zealand, Predictor Balance 46 44 42 40 Total Outlays (% of GDP) 38 36 1992 1994 1996 1998 2000

Year

New Zealand Synthetic New Zealand

Figure 2.10: Total Outlays in New Zealand

USA Spain Switzerland Germany Ireland Australia Sweden Czech Republic Portugal United Kingdom Austria Norway Canada Finland Netherlands Denmark Iceland New Zealand Luxembourg France Belgium

0 10 20 30 40

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.11: Total Outlays in New Zealand, Post-Pre-Ratios

45 State w-Weight State w-Weight State w-Weight Australia 0.003 France 0.002 Portugal 0.003 Austria 0.002 Germany 0.002 Spain 0.001 Belgium 0.002 Iceland 0.018 Sweden 0.001 Canada 0.55 Ireland 0.002 Switzerland 0.137 Chile 0.002 Luxembourg 0.003 United Kingdom 0.003 Czech Republic 0.006 the Netherlands 0.091 USA 0.167 Denmark 0.001 Norway 0.002 Finland 0.001 Poland 0.002

Table 2.7: Social Expenditure in New Zealand, w-Weights

Variable Treated Synthetic Unemployment rate 8.233 8.297 Union density 30.154 30.157 Social Expenditure 1993 19.9 19.922 Social Expenditure 1995 18.6 18.62

Table 2.8: Social Expenditure in New Zealand, Predictor Balance 20 19 18 Social Expenditure (% of GDP) 17 1992 1994 1996 1998 2000

Year

New Zealand Synthetic New Zealand

Figure 2.12: Social Expenditure in New Zealand

46 Chile Germany Austria Luxembourg Portugal Sweden Canada USA Denmark France Belgium Australia Czech Republic Iceland Spain Ireland United Kingdom Netherlands Norway Switzerland Poland Finland New Zealand

0 20 40 60

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.13: Social Expenditure in New Zealand, Post-Pre-Ratios

Austria Chile USA Portugal France Luxembourg Canada Germany Iceland Netherlands United Kingdom Czech Republic Norway Ireland Belgium Switzerland Australia Poland New Zealand

0 20 40 60 80 100

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.14: Social Expenditure in New Zealand (w/o DNK, ESP, FIN, SWE), Post-Pre- Ratios

Austria Chile Portugal France Canada Luxembourg Germany Australia USA Netherlands Iceland United Kingdom Czech Republic Belgium Norway Switzerland New Zealand

0 20 40 60 80 100

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.15: Social Expenditure in New Zealand (w/o DNK, ESP, FIN, IRL, POL, SWE), Post-Pre-Ratios

47 State w-Weight State w-Weight Austria 0.552 Norway 0.376 Finland 0.062 Sweden 0.011

Table 2.9: Total Outlays in Italy, 1994, w-Weights

Variable Treated Synthetic POP14 16.113 18.237 Employment, Service 58.78 61.849 Total Outlays 1989 51.5 51.564 Total Outlays 1991 54 53.957 Total Outlays 1993 56.4 56.506

Table 2.10: Total Outlays in Italy, 1994, Predictor Balance 56 54 52 50 Total Outlays (% of GDP) 48 1990 1995 2000

Year

Italy (1994) Synthetic Italy (1994)

Figure 2.16: Total Outlays in Italy, 1994

Australia Sweden Finland Portugal Canada Spain USA Denmark Norway Iceland Netherlands France Belgium Austria Italy United Kingdom Germany Ireland

0 10 20 30 40

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.17: Total Outlays in Italy, 1994, Post-Pre-Ratios

48 State w-Weight State w-Weight Luxembourg 0.786 Norway 0.212 the Netherlands 0.001

Table 2.11: Social Expenditure in Italy, 1994, w-Weights

Variable Treated Synthetic POP65 15.479 14.112 Unemployment rate 9.85 2.659 Social Expenditure 1990 19.9 19.766 Social Expenditure 1993 20.9 20.934

Table 2.12: Social Expenditure in Italy, 1994, Predictor Balance 23 22 21 Social Expenditure (% of GDP) 20

1990 1992 1994 1996 1998

Year

Italy (1994) Synthetic Italy (1994)

Figure 2.18: Social Expenditure in Italy, 1994

Ireland United Kingdom Sweden Poland Switzerland Norway Finland Czech Republic Turkey Canada Belgium Austria USA Netherlands Australia Spain Luxembourg Portugal Chile Denmark Iceland France Italy Germany

0 5 10 15 20

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.19: Social Expenditure in Italy, 1994, Post-Pre-Ratios

49 United Kingdom Sweden Poland Switzerland Norway Finland Ireland Turkey Czech Republic USA Chile Austria Australia Canada Spain Luxembourg Portugal Belgium Denmark France Iceland Germany Italy

0 5 10 15 20

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.20: Social Expenditure in Italy, 1994 (w/o NLD), Post-Pre-Ratios

Sweden Poland United Kingdom Switzerland Finland Ireland Turkey Czech Republic Australia Luxembourg Austria USA Chile Belgium Spain Canada Portugal Iceland Denmark France Italy Germany

0 5 10 15

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.21: Social Expenditure in Italy, 1994 (w/o NLD, NOR), Post-Pre-Ratios

State w-Weight State w-Weight Belgium 0.12 Spain 0.302 Slovak Republic 0.002 Sweden 0.576

Table 2.13: Central Government Expenditure in Italy, 2006, w-Weights

Variable Treated Synthetic lnGDP 10.247 10.292 POP65 19.131 17.095 Unemployment rate 8.66 7.669 CG Expenditure 2001 27.835 27.788 CG Expenditure 2004 26.283 26.288 CG Expenditure 2005 26.651 26.676

Table 2.14: Central Government Expenditure in Italy, 2006, Predictor Balance

50 CnrlGvrmn xedtr nIay 06(/ V) Post-Pre-Ratios SVK), (w/o 2006 Italy, in Expenditure Government Central 2.24: Figure CnrlGvrmn xedtr nIay 06 Post-Pre-Ratios 2006, Italy, in Expenditure Government Central 2.23: Figure CnrlGvrmn xedtr nIay 2006 Italy, in Expenditure Government Central 2.22: Figure Slovak Republic United Kingdom United Kingdom Czech Republic Czech Republic Central Government Expenditure (% of GDP) Luxembourg Luxembourg Netherlands Netherlands Switzerland Switzerland Germany Germany Denmark Denmark Australia Australia Slovenia Slovenia Hungary Hungary Portugal Portugal 25 26 27 28 29 30 Sweden Belgium Sweden Belgium Canada Canada 2000 Norway Norway Estonia Estonia Finland Finland Iceland Iceland Poland France Austria Poland France Austria Ireland Ireland Spain Spain USA USA Italy Italy 0 0 Italy (2006) Post-Period RMSPE/Pre-Period Post-Period RMSPE/Pre-Period 10 10 51 2005 Year 20 20 Synthetic Italy(2006) 30 30 2010 40 40 Austria Spain Norway Switzerland Slovenia France Germany Hungary Czech Republic Luxembourg United Kingdom Estonia Denmark USA Sweden Canada Portugal Poland Australia Finland Iceland Netherlands Italy Ireland

0 10 20 30 40

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.25: Central Government Expenditure in Italy, 2006 (w/o BEL, SVK), Post-Pre- Ratios

State w-Weight State w-Weight France 0.415 Slovenia 0.076 Portugal 0.089 United Kingdom 0.419

Table 2.15: Social Expenditure in Italy, 2006, w-Weights

Variable Treated Synthetic Trade 50.786 59.678 Employment, Industry 31.417 25.93 Social Expenditure 2001 23.4 23.338 Social Expenditure 2003 24.2 24.346 Social Expenditure 2005 24.9 24.732

Table 2.16: Social Expenditure in Italy, 2006, Predictor Balance 28 27 26 25 24 Social Expenditure (% of GDP) 23 2000 2005 2010 2015

Year

Italy (2006) Synthetic Italy (2006)

Figure 2.26: Social Expenditure in Italy, 2006

52 Australia Norway Poland Slovak Repulic Israel United Kingdom Portugal Hungary Netherlands Estonia Ireland Chile Italy Austria Canada Slovenia Luxembourg Czech Republic Spain France Finland Sweden Iceland Belgium Denmark Germany USA

0 5 10 15

Post-Period RMSPE / Pre-Period RMSPE

Figure 2.27: Social Expenditure in Italy, 2006, Post-Pre-Ratio

53 2.9 Appendix: Variables

CG Expenditure Total Central Government Expenditure as Percentage of GDP, Source: OECD.Stat and Own Calculations.

Employment, Agriculture Employment in Agriculture as Percentage of Total Employ- ment, Source: World Development Indicators.

Employment, Industry Employment in Industry as Percentage of Total Employment, Source: World Development Indicators.

Employment, Service Employment in Services as Percentage of Total Employment, Source: World Development Indicators.

High-Tec Exports High-Technology Exports as Percentage of Manufactured Exports, Source: World Development Indicators. lnGDP Natural Logarithm of GDP Per Capita, Constant 2000-US$, constant PPPs, Source: OECD.Stat and Own Calculations. lnPOP Natural Logarithm of Total Population, Source: World Development Indicators and Own Calculations.

Old-Age Expenditure Old-Age Expenditure of the Public Sector as Percentage of GDP, Source: OECD.Stat and Own Calculations.

Patents Patent Applications of Residents per Million People, Source: World Develop- ment Indicators and Own Calculations.

POP14 Population Aged 0-14 as Percentage of Total Population, Source, World Devel- opment Indicators.

POP65 Population aged 65 and Above as Percentage of Total Population, Source, World Development Indicators.

Social Expenditure Social Expenditure of the Public Sector as Percentage of GDP, Source: OECD.Stat.

Total Outlays Total Outlays (Disbursements) of the General Government Level as Per- centage of GDP, Source: OECD Economic Outlook No. 88 database.

Trade Imports of Goods and Services Plus Exports of Goods and Services as Percentage of GDP, Source: World Development Indicators and Own Calculations.

Unemployment Rate Total Unemployment as Percentage of Total Labour Force, Source: World Development Indicators.

54 Union Density Wage and Salary Earners who are Trade Union Members as Percentage of Employees, Source: OECD.Stat.

Youth Unemplyoment Unemployment of People Aged 15-24 as Percentage of Total La- bor Force Aged 15-24, Source: World Development Indicators.

55 3 Does the Swiss Debt Brake Induce Sound Federal Finances?

3.1 Introduction

On December 2, 2001 Swiss citizens voted in favour of the introduction of a fiscal rule at the national level. After the debt to GDP ratio of the Swiss federal government increased from 13 percent in 1991 to 25 percent in 1997, government, parliament and the general public wanted to break this trend and start with a consolidation of the federal budget. This is in line with the deficit bias of political decision-makers in developed economies already emphasized in the survey by Alesina and Perotti (1995). In addition, the institutional logic of the budgetary process at the Swiss federal level induces a deficit bias. Tax rate increases require a change of the Swiss constitution whereas an increase in government expenditure only needs a simple majority of representatives (Geier, 2011). As a result, spending increases were rather financed by budget deficits than tax increases because of political reasons and institutional restraints. Switzerland introduced the federal debt rule to abolish deficit biases and finally arrive at sound federal finances. For the new debt rule in Germany, introduced for similar reasons, the Federal Ministry of Finance recommended a budget that is ’close-to-balance’. Due to the difficulty of precisely defining public investment spending the German debt brake does not include any investment orientation or golden rule. Borrowing for non-cyclical reasons is however allowed for in the size of 0.35 percent of GDP.1 Borrowing for cyclical reasons is symmetric under the Ministry’s proposal since the proposal includes a compensation account which is debited if revenues fall below expenditures. Similar to the Swiss rule the compensation account is armed with an upper limit. So a record of the compensation account larger than the limit must be cut back under the limit (Baumann et al., 2008; BMF, 2008; Kastrop and Snelting, 2008).2 In 2011 the European Union member states agreed on the European Fiscal Compact in response to the European debt crisis. Under this treaty the respective states consented to implement a balanced budget rule at the general government level. Although it is the member states’ responsibility to decide on the details of their fiscal rule the compact

1The German debt brake applies to the federal level and the states. Borrowing for structural reasons is allowed for the federal level only. 2However, the German fiscal rule differs from the Swiss one e.g. in terms of the estimation of the output gap and the exceptions from the rule. For a lengthy description of the enacted version of the German fiscal rule, see Feld (2010).

56 requires some features following the Swiss balanced budget rule and the German debt brake. So it calls for a balanced budget and limits the structural deficit to 0.5 percent of GDP. The national rules must be accompanied by an automatic correction mechanism that becomes effective in the case of non-compliance with the rule. In line with the compact, Austria, Cyprus, Finland, France, Ireland, Italy, Luxembourg, Latvia, Portugal and Spain introduced a balanced budget rule that allows for a small structural deficit in general. In Austria the rule is accompanied by a compensation account of German style. Spain and Portugal implemented an automatic correction mechanism in terms of a financial plan that corrects the deviations. In Latvia the structural deficit of the subsequent years automatically gets reduced in the case of non-compliance (Burret and Schnellenbach, 2013). In Israel the experience with fiscal rules indicates that they were hardly obeyed in the past. For Israel it was important to obtain a fiscal rule that directly impinges on spending trends and that directly punishes systematic deviations from the rule. Debrun et al. (2008) value the compensation account to be helpful in a way that systematic deviations from the fiscal target do not fall into oblivion. Finally, they recommend an expenditure rule for Israel including a feedback mechanism that adjusts the expenditure growth ceiling according to the long-run debt target. Despite this political popularity, it is contested whether fiscal rules actually contribute to fiscal sustainability. On the one hand, a debt reducing effect of fiscal rules is supported by some empirical studies. Bohn and Inman (1996) consider the budget surplus of the general fund of 47 U.S. states covering the 1970-91 time span and find a clear positive and statistically significant effect for strong rules, especially the no-carry-over deficit rule. This result for the U.S. states is confirmed by Alesina and Bayoumi (1996) and Hou and Smith (2010). The former study reports supportive evidence also for a broader measure of the budget surplus. Hou and Smith (2010), in constrast, find that narrow budget measures are better constrained by fiscal rules than broad ones. Additionally they find that fiscal rules are more effective if they contain a technical requirement instead of charging political actors. In focusing on constitutional limits on guaranteed bond indebtedness in the U.S. states for the 1961-90 period, Kiewiet and Szakaly (1996) find that guaranteed debt is lower when the issuance of bonds is prohibited at all. A qualified majority in parliament or revenue-based limitations are, in contrast, not helpful in reducing the debt level. On the other hand, there is also some evidence that calls the effectiveness of formal fiscal rules into question. In their 1980 cross-section analysis covering the U.S. states Abrams and Dougan (1986) do not find a robust significant effect on government spending if borrowing is allowed for. Although using a panel structure of the 1961-89 period and focusing on debt growth to deal with the nonstationarity of debt data, Clingermayer and Wood (1995) cannot find a significant effect of formal debt constraints in the U.S. states. Instead they confirm that limitations of revenue or spending are not helpful in

57 reducing debt growth. In also considering the U.S. states, von Hagen (1991) does not find a significant effect of a debt rule or a strong balanced budget rule on debt per capita or the debt ratio based on a test for equal means. However, he finds that the median-debt- per-capita is lower in the country group with a debt rule or a strong balanced budget rule. Eventually, the study of Hou and Smith (2009) can also be interpreted as questioning the effectiveness of formal fiscal rules. They find that the no-carry-over deficit rule is obeyed as informal rule even if it is not fixed by law. This basically means that politicians and eventually voters behave in a fiscally conservative way. For Switzerland there is evidence for the 1980-98 time period, the 1986-97 time span and the 1980 to 2011 period that strong statutory fiscal constraints significantly reduce deficits and debt at the cantonal level (Burret and Feld, 2014, 2016; Feld and Kirchgaessner, 2001, 2008; Schaltegger, 2002). Regarding the local level fiscal rules reduce deficits. Beyond also providing evidence for the effectiveness of fiscal rules Luechinger and Schaltegger (2013) show that fiscal rules improve the accuracy of budget projections. The authors interpret this result as a decrease in strategic behaviour during the budget process. Using financial market data from 1981 to 2007, Feld et al. (2013) report that cantonal fiscal rules reduce spreads, although the re-establishment of a no-bailout regime regarding local jurisdictions has a quantitatively larger negative effect on cantonal spreads. However, there is no evidence yet on the effects of the Swiss federal debt brake on federal finances. The literature could be summarised such that fiscal rules are the more effective the more restrictive the design of the rule is. That is, a strong balanced budget rule works better than a revenue rule or a debt rule, and a technical rule works better than a political one. Since the Swiss federal fiscal rule meets these requirements it can be assumed that it might work and the premature praise seems justifiable. Admittedly, this does not provide evidence that the rule is indeed effective. Thus, this is the first paper that analyses the effects of the Swiss federal debt rule. As the Swiss debt brake is legally effective only since a dozen years, standard time series methods cannot be used. While the 2003 budget was the first under the debt brake, the government had to apply a transitional period that lasted until 2005. Therefore, we employ the Synthetic Control Method to analyse the effect of the Swiss debt rule’s introduction on the cyclically-adjusted budget balance, the central government debt ratio as well as the general government debt ratio. Aside other results, we find an improvement of the budget balance by about 3.6 percentage points in a post-intervention period covering five years. The chapter proceeds as follows: section 3.2 explains the mechanics of the Swiss fiscal rule, surveys the literature that discusses the features of this rule and deals with the start-up difficulties of the years 2003 to 2005. Section 3.3 reports the data, section 3.4 provides the results and section 3.5 concludes.

58 3.2 The Design of the Swiss Balanced Budget Rule

The Swiss fiscal rule basically consists of a ceiling of total central government expenditures, i.e., expenditures in the next fiscal year must follow the predicted revenues for that fiscal year. Additionally, predicted revenues are multiplied by a factor that corrects for the position within the budget cycle. The basic mode of operation can be illustrated by equation (3.1)

Et = k · Rt (3.1) where the expenditure ceiling E for ﬁnancial year t equals the revenue forecast R for the same year t multiplied by the adjustment factor k. This factor is given by

Y T k = t (3.2) Yt

T with Yt being the potential real GDP (trend) and Yt being the predicted real GDP of the year t. In case of an under-utilisation of capacities, the ratio of potential real GDP over predicted real GDP is larger than one and expenditures can exceed predicted revenues. In times of booms, in contrast, the factor k forces the budget to generate a (yearly) surplus which results in a balanced budget over the cycle. Eventually, the calculation of the factor k relies on an adjusted Hodrick-Prescott filter instead of using a production function approach. If total expenditures exceed the ceiling, the (additional) deficit is booked in a compensation account. Deficits in this account must be redeemed in the subsequent fiscal years. However, the terms of the amortisation are not specified. If the deficit of the compensation account gets to large, i.e., it exceeds six percent of effective expenditures of the last fiscal year, the government must reduce it below six percent of expenditures within three years. Beyond that, the calculation of the expenditure ceiling comprises investment spending but disregards windfall revenues (which therefore do not endanger the rules’ stringency) and the social security system. The law also considers the need for extraordinary expenditures. Yet, this is possible only for a finite number of cases which are conclusively stated in the law. Every single case needs the confirmation by the majority of parliament. Since 2010 the rule also covers the extraordinary budget (BV, 2002; FHG, 2006; FHG, 2010; Geier, 2011). As mentioned the Swiss debt rule was given credit from scientists and practitioners alike. It is esteemed that the rule targets the deficit instead of debt as the former is under direct control of policy-makers. Furthermore, a balanced budget target is welcomed because a surplus target would certainly raise the desire to spend these surplus amounts in line with short-run political demands instead of using them to unburden future generations. In addition, a balanced budget target can be better understood by the public than a

59 somewhat arbitrary threshold for the debt ratio (Danninger, 2002; Debrun et al., 2008). The rule’s way of considering the business cycle is also acknowledged. Instead of demanding the budget to balance ’over the cycle’ the rule is based on revenue and GDP forecasts of the next financial year only. Prediction errors which often go along with medium term forecasts of macroeconomic variables are thus minimised (Danninger, 2002). The centrepiece of the rule, the compensation account, is explicitly appreciated by Debrun et al. (2008). Thanks to this error correction mechanism the violation of the fiscal rule does not jeopardise the sustainability of public finances which is the ultimate aim of a fiscal rule. Finally, the rule is commended for being enshrined in the constitution and for the fact that the escape clauses are listed conclusively in the law (Danninger, 2002). However, there is also some criticism. Danninger (2002) points out that the Hodrick- Prescott filter comes up with an end-point bias that is not convincingly solved in the Swiss debt brake. When calculating the adjustment factor k for the fiscal year t the trend-GDP value of that year is always the last value of the trend-GDP time series. The HP-filter formula (not shown) reveals that its first term minimises a deviation of the trend from the real values. Its second term, in contrast, penalises changes in the slope and thus smoothes the trend time series. If a GDP forecast of the year t + 1 shall not be used in order to avoid prediction errors, this second term of the HP-filter formula cannot be calculated for the year t because a trend-GDP value of the year t + 1 does not exist. Thus the trend-GDP value is biased towards the real GDP value in year t. To correct for this the Swiss Federal Finance Administration (SFFA) uses a modified HP-filter since 2004. Under this modification the smoothing factor λ is multiplied by the factor 1.5 for the penultimate trend-GDP value and is multiplied by factor 3 for the last trend-GDP value. Although the modified HP-filter cannot eliminate the end-point bias it clearly shrinks it (Bruchez, 2003a). An alternative would be to use a production function approach which additionally might be better able to capture the country’s economic situation than the HP-filter. However, this approach is based on many parameter assumptions which might be subject to political leverage (Colombier, 2004, 2006). A minor point of criticism related to the filtering procedure is the somewhat arbitrary choice of the smoothing factor λ = 100. Danninger (2002) argues that there is no theoretical reasoning behind it. The SFFA, however, points to the fact that the choice of the smoothing factor is in line with European Union standards (Colombier, 2006). The calculation of the expenditure ceiling at least implicitly assumes the GDP-elasticity of revenues to be one and fixed over time. Danninger (2002) notes that this is not necessarily the case. Colombier (2004) admits that some studies report a GDP-elasticity of revenues larger than one but counters that these studies are not relevant as they include revenue categories legally irrelevant to the Swiss debt brake. He concludes that there is no clear evidence of a GDP-elasticity of revenues different from one. An alternative would

60 be to employ a revenue trend instead of revenue forecasts combined with a GDP trend. However, a reliable revenue trend estimation method would be necessary then. If the revenue trend is based on a HP-filtering procedure the end-point bias would still be in place. Another point of criticism refers to long-lasting recessions. Redeeming deficits from the compensation account might become problematic under such circumstances because it would take place in a situation of under-utilisation of capacities. Fiscal policy acts in a procyclical way then. Similarly, fiscal policy becomes procyclical in a long-lasting recession because the HP-filter adjusts the GDP trend downwards if the effective GDP decreases over a longer time period. This reduces the output gap and forces the fiscal policy to be more restrictive although the recession is still running (Danninger, 2002). Finally, it is argued that investment spending will become too low under the rule. Since it is well known that investment spending positively impacts on economic growth it might be helpful or even necessary to foster investment spending. Moreover, politicians tend to cut investment spending when forced to cut overall spending. Yet, the problem with investment spending is its precise definition. There is a danger that consumption spending is declared as investment spending which would undermine the functioning of the rule. This is precisely the reason why the authorities in Germany departed from the golden rule of public finances. The SFFA argues that a special fund might be put in place in the case of a need of investment spending. However, this again leads to a decrease in transparency (Colombier, 2004, 2006). Before starting the analysis some words are in place regarding the introduction phase of the rule (2003-05). The budget of the year 2003 and the medium-term financial plan 2004-06 were both the first of its kind under the new rule. To calculate the expenditure ceiling, however, the SFFA used the classical HP-filter instead of the modified one at that time. Because of the end-point bias the expenditure ceiling was inappropriately low and the budget ran into a deficit. Moreover, the adjustment factor k does not react very precisely to corrections of the GDP forecast. So at the time of the budget preparation in year t − 1 only a GDP forecast is available. That is the calculation of the expenditure ceiling is based on this GDP forecast and the respective trend GDP value. If then later that year the GDP forecast is being revised the GDP trend changes due to the relatively T large impact of the last trend GDP value on the overall trend. However, the ratio of Yt T over Yt is affected only weakly because both values Yt and Yt have changed. This was precisely the case in 2003 when the adjustment factor k did not appropriately react to the revision of the GDP forecast (Colombier, 2004). Consequently, the federal government introduced a transition period and debt ratio increased in the years 2003 to 2005 in spite of the introduction of the debt brake (see figure 3.7 and the discussion below).

61 3.3 Data

First and foremost we are interested in the effect of the Swiss debt brake on the budget balance. Thus we look at the total federal government budget balance which is total federal government revenues net of total federal government expenditure. These variables are given in national currency and in current prices. While total revenue entails total sales (market output and output for own final use), payments for non-market output, subsidies, property income, total taxes, total social contributions, other transfers and capital transfers, total expenditure mainly covers intermediate consumption, compensation of employees, subsidies, interest payments, taxes, social benefits, social transfers in kind, current transfers, capital transfers, gross capital formation and net acquisition of non-financial non-produced assets (OECD, 2007). The budget balance is expressed in percent of GDP and is adjusted by the output gap to obtain the cyclically adjusted budget balance (see section 3.4.1). Additionally we are interested in the effect on government debt measured by the federal government debt ratio as percentage of GDP. This covers debt issuance of the federal government level (marketable and non-marketable) excluding state and local government debt as well as social security funds. Moreover, we look at general government debt as percentage of GDP. This covers all government units (federal, state, local) as well as nonmarket-nonprofit institutions that are financed by government units. Public corpora- tions are excluded (Abbas et al., 2010). In order to find out whether compliance with the rule is achieved by an increase in revenues or a decline in expenditures we also intend to test the effect of the debt brake on the revenue ratio as well as the expenditure ratio. However, these variables are rather volatile and an approximation in the pre-intervention period is not possible to a satisfying extent. We also subtract social expenditures from government expenditures without any particularly helpful result. In choosing variables that are good predictors of debt development we follow the literature (e.g. Bohn and Inman, 1996). Firstly, we consider the natural logarithm of real GDP (per capita) as Wagner’s law claims that government spending is higher the more developed the economy is. Secondly, we consider the annual growth rate of GDP, the growth of unemployment as well as the rate of unemployment as debt development clearly depends on the business cycle. Moreover, the growth of GDP reflects a country’s capability to repay debt. Thirdly, the natural logarithm of total population is included as it captures the idea that some government spending categories are affected by economies of scale. Finally, we include the percentage of people aged 65 and older and the sum of imports and exports as an indicator for the openness of a country. These two variables describe important determinants of social spending and thus government spending. Whereas the share of the elderly is assumed to be positively correlated to health spending the the-

62 ory behind the trade indicator is that governments of open economies are incentivised to compensate the losers of globalisation (Rodrik, 1998). These predictor variables are always supported by the lagged outcome variable of certain years which often is the ﬁrst and the last year of the pre-intervention period. Among these predictors we search for a combination that best supports the approximation of the real Swiss data. As mentioned already we restrict our sample to OECD member countries. However, data is not available for all of these nations and all variables of interest. Thus we basically restrict our sample to countries for which data is available for the budget balance, the central government debt ratio and the general goverment debt ratio. So our sample contains Australia, Austria, Belgium, Canada, the Czech Republic, Denmark, Finland, France, Germany, Ireland, Italy, Japan, Luxembourg, the Netherlands, New Zealand, Norway, Portugal, the Slovak Republic, Spain, Sweden, the United Kingdom and the USA. Hence our donor pool basically consists of 22 countries. With respect to Denmark (1997), New Zealand (2002) and Sweden (2003) there was one single observations missing in the general government debt ratio time series. However, since this variable works as dependent variable the whole time series cannot be used if one observation is missing. In order to be able to use this data we generated the missing observations by employing a linear interpolation. As government debt generally moves rather smoothly we think that a linear interpolation for one single observation is appropriate. For the United Kingdom central government debt ratio data was missing for the 1993-1997 time span in our data source which is OECD.Stat. In order to be able to include the United Kingdom we use central government debt data from the OECD Fiscal Decentralisation Database instead. This data covers the total liabilities excluding insurance technical reserves and is given as percentage of GDP. For the years 1998 to 2010 for which both sources provide data the average deviation amounts to 2.69 percentage points. This shows that the grafted time series behaves very similar to the original data. For the Czech Republic and the Slovak Republic both central government debt data and general government debt data start in 1992 or 1993. We thus choose the latter year to be the starting point of the respective pre-intervention periods. So the pre-intervention period starts a little later than without this restriction but this is for the beneﬁt of including these countries in the analysis. Likewise, we choose the post-intervention period of the central government debt analysis to end in 2009 (instead of 2010) in order to be able to include Japan in this analysis. For the three countries Israel, South Korea and Poland we have data for two of the dependent variables. We choose Israel to be part of the analysis of the budget balance and the general government debt ratio as data is available for these two series only. South Korea and Poland are included as robustness check for the analysis of the central and general government debt ratio.

63 3.4 Empirical Analysis

3.4.1 The Eﬀect on the Budget Balance

At first glance the variable ’government expenditure’ is the rule’s target figure. However, expenditures are related to predicted revenues of the same fiscal year. This renders the Swiss fiscal rule a balanced budget rule. Thus we first check its impact on the budget balance, and by applying the HP-filter to available GDP data, we calculate the cyclically adjusted budget balance. Unlike the SFFA we use the classical Hodrick-Prescott filter (with a smoothing factor λ = 100). On the one hand this is because the classical filter is easier to implement. On the other hand the SFFA needs this modified filter because the classical one has an end-point bias (see section 3.2). However, we do not face this problem. Whereas trend-GDP data is available for the 1980-2013 or 1990-2013 time span our pre-intervention period starts in 1995 and ends in 2007 in order to estimate the effect of the debt brake as precisely as possible. The Synthetic Control Method is able to capture the short run effect of a policy intervention. A large extension of the period around the intervention thus runs the risk of capturing structural shocks that are not related to the Swiss debt rule’s introduction. Our calculation of the cyclically adjusted budget balance is in line with the standard methodology used, e.g., in the European Union fiscal framework (Mourre et al., 2013) and can be expressed by B Y − Y T CAB = − (3.3) Y Y T B Y −Y T with Y as the nominal budget balance to GDP ratio and Y T as the percentage of the output gap on the trend-GDP. Regarding the minuend of the equation we use GDP data in current prices since revenues and expenditures are also expressed in current prices. Consequently, we also use (trend) GDP data in current prices in the subtrahend of the equation. Concerning the minuend, price changes should not affect the public finance variables very much since they are expressed as percentage of GDP. Similarly, price move- ments should largely be cancelled out in the subtrahend as the nominal business cycle element Y − Y T is also expressed as percentage of nominal GDP. Of course, this assumes that revenues and expenditures are affected by price changes in the same way as the output variable. The output gap usually is multiplied by the factor that captures the cyclical sensitivity of the budget balance. However, since the SFFA in their calculation of the budget balance implicitly assumes the revenue elasiticity to be one (Colombier, 2004) we refrain from using an overall elasticity different from one. To test the effect of the rule’s introduction on the budget balance we start the pre- intervention period in 1995. This is because revenue and expenditure data starts in 1995 for most of the countries in the sample. We rely on the consistent sample as explained in section 3.3 which is enriched by Israel since revenue and expenditure data is available

64 for it. This leaves us with 23 countries in the donor pool. Among the predictor variables named in section 3.3 we use the natural logarithm of total population and the growth rate of unemployment as these variables proved to be helpful in minimising the pre-intervention RMSPE. Additionally, we use the lagged cyclically adjusted budget balance of the years 1995, 1999 and 2000. Since the budget of 2003 was the first under the rule we declare the year 2002 to be the year of intervention. Although the referendum took place in 2001 this is unproblematic because the referendum took place at the end (December) of that year. The post-intervention period consequently starts in 2003 and ends in 2007 as it shall cover no more than five years. Referring to what is recommended in section 1.2.3 one could argue that the pre-intervention period is short. However, the budget balance is rather volatile because it is composed of revenue and expenditure data which are rather volatile, too. As this affects the approximation we value a 7-year-long pre-intervention period to be reasonable. The w-weights are given in table 3.1. All countries used for this analysis are also part of the synthetic control. In fact, this makes the Synthetic Switzerland a real ’synthetic’ control since Switzerland is not replicated by a small number of countries with large w-weights each. Instead, many countries contribute to the Synthetic Switzerland with a w-weight smaller than one percent. And although four countries (Ireland, Israel, the Netherlands and the Slovak Republic) stand out with a relatively large w-weight, none of these weights is larger than 30 percent.3 The predictor balance is given in table 3.2. This balance shows the quality of the approximation in the pre-intervention period indicating it for every single predictor variable. The column ’Treated’ provides for the value of the respective predictor variable for Switzerland that is averaged over the pre-intervention period. The entire column thus provides the vector Xi explained in section 1.2.3. Likewise, the column ’Synthetic’ provides the mean value of the same predictor variable for Synthetic

Switzerland. So this column is related to matrix XJ . As can be seen the average fit between Switzerland and the synthetic control is very precise for all explaining variables. This is also reflected by the pre-intervention RMSPE which amounts to 0.43. Figure 3.1 now exhibits the graphical result. Except for small deviations between the treated unit and the synthetic control in 1996, 1997, 1998 and 2001, the fit is absolutely congruent. For the years after the rule’s introduction we see a clear positive effect. That is, the budget balance would have performed worse without the fiscal rule. The average treatment effect is then calculated as explained in section 1.2.3. It amounts to 3.678 percentage points. The result of the subsequently conducted cross-country placebo test

3Fiscal consolidation took different paths across these countries. Ireland showed an expenditure restraint and increased indirect taxes (OECD, 2003). The Slovak Republic, besides the revenue-neutral tax- rate-cut-cum-base-broadening reform, mainly reduced social security benefits (Moore, 2005). The Netherlands adjusted on the expenditure side via the freeze of wages, job cuts, the reduction in healthcare benefits and cuts in subsidies. To roughly 20 percent this country consolidated its public finances via the abolition of tax exemptions (OECD, 2004).

65 can be seen in figure 3.2. Switzerland clearly stands out with the largest post-pre-ratio among the countries. The probability to find a country in the pool with a post-pre-ratio of the size of Switzerland or even larger is 1/24=0.042. To test the robustness of this result we eliminate the countries with the smallest w- weights from the pool. So we discard the Czech Republic and Germany in a first step. These countries contribute to the Synthetic Switzerland with a w-weight of 0.003 and 0.004, respectively.4 The w-weights as well as the graph (not shown) are very similar to the baseline estimation with the pre-intervention RMSPE amounting to 0.44. Conducting placebo tests again leads to figure 3.3 showing that Switzerland again comes up with the highest post-pre-ratio. The p-value equivalent is 1/22=0.045. We then drop New Zealand as it shows up with a w-weight of 0.006.5 Again, there are no considerable changes compared to the baseline estimation concerning the graph (not shown). The pre-intervention RMSPE accounts for 0.441. Figure 3.4 shows that Switzerland can assert its position in the ranking of the post-pre-ratios. The p-value equivalent increases only slightly to 1/21=0.048. In a third step we drop Austria, Italy, Japan and Sweden which show up with a w- weight of 0.008.6 The pre-intervention RMSPE amounts to 0.462. Once more the graph is very similar to the one of the baseline estimation and thus not shown. And once more Switzerland comes up with the largest post-pre-ratio (see figure 3.5). The p-value equivalent accounts for 1/17=0.059. Finally, we discard Norway and Finland which contribute to the Synthetic Switzerland with a w-weight of 0.01 and 0.011, respectively.7 Still the parameters are very similar. This holds for the pre-intervention RMSPE (0.48), the graph of the cyclically adjusted budget balance (not shown) and the p-value equivalent (1/15=0.067). The latter indicates that our result is robust and significant at the seven percent level. Eventually, figure 3.6 reveals that Switzerland keeps the first position among the post-pre-ratios. Moreover, the distance to the country that comes second is distinct in all the figures of post-pre-ratios. From the baseline estimation as well as from the robustness checks we conclude that

4This leads to the following w-weights: Australia 0.01, Austria 0.007, Belgium 0.01, Canada 0.013, Denmark 0.013, Finland 0.007, France 0.016, Ireland 0.28, Israel 0.143, Italy 0.008, Japan 0.007, Luxembourg 0.01, the Netherlands 0.174, New Zealand 0.006, Norway 0.007, Portugal 0.21, Slovak Republic 0.236, Spain 0.008, Sweden 0.007, the United Kingdom 0.009 and the USA 0.009. 5This leads to the following w-weights: Australia 0.01, Austria 0.008, Belgium 0.011, Canada 0.012, Denmark 0.014, Finland 0.009, France 0.014, Ireland 0.281, Israel 0.144, Italy 0.008, Japan 0.008, Luxembourg 0.011, the Netherlands 0.173, Norway 0.009, Portugal 0.02, Slovak Republic 0.235, Spain 0.009, Sweden 0.008, the United Kingdom 0.009 and the USA 0.009. 6This leads to the following w-weights: Australia 0.014, Belgium 0.014, Canada 0.018, Denmark 0.019, Finland 0.011, France 0.02, Ireland 0.269, Israel 0.142, Luxembourg 0.016, the Netherlands 0.166, Norway 0.01, Portugal 0.029, Slovak Republic 0.232, Spain 0.012, the United Kingdom 0.013 and the USA 0.012. 7This leads to the following w-weights: Australia 0.017, Belgium 0.02, Canada 0.019, Denmark 0.027, France 0.02, Ireland 0.268, Israel 0.137, Luxembourg 0.026, the Netherlands 0.16, Portugal 0.034, Slovak Republic 0.226, Spain 0.016, the United Kingdom 0.016 and the USA 0.013.

66 the Swiss balanced budget rule indeed contributes to an improvement of the (cyclically adjusted) budget balance.

3.4.2 The Eﬀect on the Government Debt Ratio

Beyond analysing the effect on the budget balance we also check the effect of the debt rule on the debt ratio. Admittedly, the rule does not directly aim at reducing the debt ratio. However, a fiscal rule should help to achieve sustainable public finances more or less irrespective of what the target variable precisely is. And the status of financial sustainability can best be captured by the debt ratio. So at first we look at the federal government debt ratio as dependent variable and subsequently we look at the general government debt ratio. The general government level is more appropriate when it comes to the overall fiscal sustainability. However, the Swiss fiscal rule applies to the federal government level with the cantons having their own fiscal rules. Moreover, the comparison between the federal and the general government level can indicate whether there is some dislocation from the former to the latter. Starting with the central government debt ratio we again rely on the sample mentioned in section 3.3. Israel is not added because central government debt data is missing for the years before 1997. So there are 22 countries in the pool. In this analysis we use the percentage of people aged 65+ and the rate of unemployment as well as the lagged outcome variable of the years 1993, 1997, 2001 as predictors. Although data is available for the earlier years we start the pre-intervention period in 1993 in order to be able to include the Czech Republic and the Slovak Republic into the analysis. Due to an intersection of both trajectories in the post-intervention period (2003-05) we extend this period to more than five years. In order to be able to include Japan, however, the post-period ends in 2009 instead of 2010. The post-period thus covers seven years. The w-weights are given in table 3.3. In contrast to the budget balance analysis a considerably smaller number of countries is used to approximate the debt development in Switzerland. Two of these four countries show up with relatively large weight of about 30 percent and almost 60 percent, respectively. So almost 90 percent of the debt trend of Switzerland is represented by France and Luxembourg. The predictor balance (table 3.4) points to a very precise fit in the pre-intervention period. Only the approximation in terms of the unemployment rate is a little less precise. The pre-intervention RMSPE accounts for 0.836. Figure 3.7 now provides the result. As can be seen the approximation is relatively good except for the 1997-99 period. The hump in Switzerland’s data scarcely can be imitated by any synthetic control. In the post-period one can see that debt in Switzerland exceeds that of Synthetic Switzerland for three years after the introduction (2003-05). We ascribe this effect to the rule’s introduction phase explained in section 3.2. In the years to follow, however, the debt ratio clearly decreased below the level of 2001. The debt of the synthetic

67 control, in contrast, increases remarkably from 2008 onwards. To some extent this might be driven by the situation in France which weightily contributes to Synthetic Switzerland and strongly suffered from the debt crisis. Sadly, we cannot run the placebo tests due to the intersection of the two trends in the post-period. And we cannot say what the trend of Synthetic Switzerland would have been without the debt crisis. Thus we hardly can tell anything about the debt progress of Switzerland without the introduction of the fiscal rule. As mentioned earlier, South Korea and Poland are not part of our basic sample. So we now include these two countries. The post-period again ends in 2009. Searching for the smallest pre-intervention RMSPE uncovers that the natural logarithm of real GDP and the share of elderly people are the predictors that best help to minimise the pre-intervention RMSPE. Instead of Japan, the Czech Republic (0.059) and Finland (0.035) now contribute to the synthetic sibling. France (0.401), Luxembourg (0.473), and Sweden (0.033) are still part of the synthetic control whereas France and Luxembourg again explain about 90 percent of the synthetic control’s trajectory. The pre-intervention RMSPE worsens only slightly (0.92). The graph of the post-period looks very much the same (and is thus not shown) which is reasonable since neither South Korea nor Poland contribute to the synthetic control. If we instead simply include South Korea and Poland in the previous analysis and do not search for the best fit the pre-intervention RMSPE amounts to 0.942. The new w-weights are: Austria 0.129, Finland 0.156, Japan 0.066 and Luxembourg 0.65. The post-period trajectories of Switzerland and its synthetic sibling still move very similar to the baseline analysis (not shown). Next we turn to the general government debt ratio as dependent variable. Again we rely on the sample described in section 3.3. Since general government debt data is available we include Israel in the analysis. This leaves us with 23 countries in the pool. As explaining variables we employ the natural logarithm of population, the rate of unemployment and the lagged debt ratio of the years 1993, 1997 and 2001. Whereas the pre-intervention period again runs from 1993 to 2001 we once more extend the post-period up to 2010 due to an intersection of the post-period trajectories. The post-period thus covers eight years. The w-weights are now given in table 3.5. Similar to the federal government debt analysis a limited number of countries contribute to Synthetic Switzerland. France is not part of the synthetic control group anymore and the w-weight of Luxembourg is considerably smaller compared to the central government debt analysis. In contrast, Denmark and the United Kingdom explain almost two-thirds of Switzerland’s debt development. The predictor balance shows that the approximation works very precisely regarding the lagged outcome variable and rather well with respect to the remaining predictors. The pre-intervention RMSPE amounts to 0.712. The graphical result can be seen in figure 3.8. In the first five years of the pre- intervention period the approximation is very accurate. Then Synthetic Switzerland,

68 to some extent, tries to copy the hump of the 1997-99 period. In the post-period general government debt of the Synthetic Switzerland evolves differently compared to the federal government debt analysis. After a slight increase in 2003 and 2004 it declines in 2005, 2006 and 2007. Similar to the central government debt analysis it rises from 2008 onwards. Once again, we cannot conduct the placebo tests due to the intersection of the trajectories. And we hardly can state what the path of Synthetic Switzerland would have been without the financial crisis. Since the debt of the synthetic control starts to decrease in 2005 we cannot deduce what a potential treatment effect from 2008 onwards would look like. Nevertheless we can assume the nonexistence of debt dislocation effects. This is because the trajectories of the federal government debt ratio as well as the general government debt ratio of (real) Switzerland look very similar in the post-intervention period. That is the pressure to put federal public finances in order is not realised by stressing subnational government levels. This makes sence as the cantons have fiscal rules themselves. In a next step, we again include South Korea and Poland as robustness check. Our search for the best fit of the pre-intervention period, however, leads us to the same predictor variables and the same Synthetic Switzerland with a very similar pre-intervention RMSPE (0.711) and a graphical result (not shown) compared to the previous analysis.

3.4.3 On the Introduction of Fiscal Rules among the Comparison Units

As mentioned in section 1.2.3 the outcome variable of any comparison unit must not be affected by shocks in the time period under consideration for the result to be unbiased. This also means that the introduction of a fiscal rule in any country of the synthetic control group would bias the values of their cyclically adjusted budget balance. This is especially relevant for the post-intervention period. The best way to get rid of this problem is to drop all countries from the donor pool that introduced a fiscal rule between 1995 and 2007.8 This applies to 16 out of 23 countries in the donor pool. That is, Australia, Aus- tria, Canada, Czech Republic, Denmark, Finland, France, Israel, Japan, Luxembourg, Norway, Portugal, Slovak Republic, Spain, Sweden and the United Kingdom either introduced (changed) a national fiscal rule (central or general government level) or joined the Maastricht Treaty (IMF, 2009; Budina et al., 2012). A serious analysis would thus not be possible anymore. Hence we remove every single affected country from the pool and check what the trajectory of Synthetic Switzerland in the post-period looks like without this one country. When we drop the countries with a w-weight smaller than 0.02 (see table 3.1) the treatment effect remains unchanged in every single case. The pre-intervention RMSPE is

8As we do not obtain a treatment eﬀect in either of the debt ratio analyses this section deals with the (cyclically adjusted) budget balance analysis only.

69 always very close to that reported in section 3.4.1. Thus we argue that there is no bias stemming from these countries in the post-period. When we look at the four countries with a relatively large w-weight, we find that Ireland and the Netherlands did not introduce a fiscal rule in the respective time period. Dropping Israel leads to a pre-intervention RMSPE of 0.593 which is only slightly poorer than the RMSPE of the baseline estimation. Again, a remarkable treatment effect can be observed (not shown) and, in fact, this one is even larger than the baseline effect. It amounts to 5.817 percentage points. Yet, this makes sense because Israel expanded its Deficit Reduction Law by a provision to limit the growth of central government expenditures in 2004 (IMF, 2009; Budina et al., 2012). The only country that is problematic in this sense is the Slovak Republic. The approximation of the years 1996, 1997, 1998, 2000 and 2001 is rather imprecise if we remove it from the pool. The pre-intervention RMSPE now amounts to 1.247. However, this country did not introduce a fiscal rule at the national level in the respective time period but joined the European Union in 2004. Similar to the case of Israel we argue that joining the Maastricht Treaty might have improved the budget balance rather than worsened it. The gap between the trajectory of Switzerland and Synthetic Switzerland would be larger if the Slovak Republic did not join the European Union. So this case but even more the Israeli case suggests that the treatment effect we see in figure 3.1 is a conservative estimation. Overall, the budget balance analysis benefits from the fact that many comparison units contribute with a small w-weight which creates a ’really’ Synthetic Switzerland.

3.4.4 The Debt Brake in the Medium Run

The Swiss fiscal rule aims at balancing the budget in the medium run. In the light of a positive GDP growth this would lead to a decrease of the debt ratio. However, Schips et al. (2003) show that exactly balancing the budget is not possible over the business cycle. This is because revenues follow a positive trend over time whereas the factor k does not follow a trend. The cause of the positive revenue trend can be seen in the stable relationship between revenues and GDP. In conducting a simulation based on GDP and revenue data of the 1988-2001 period they show that the rule might bring debt back to its initial level in the medium run but that this is not necessarily the case. Thus a decreasing debt ratio cannot be expected in the medium run. Bruchez (2003b) also finds that a structurally balanced budget is not possible in the medium run. But beyond that he theoretically as well as numerically shows that the expected average deficit in the long run is rather small. Thus he concludes that the Swiss fiscal rule is able to stabilise the debt level in the long run. And this, in turn, means that it promotes the decrease of the debt ratio in the long run. This is supported by Geier (2004) who applies the rule to three artificial series of GDP-data and revenue data. That is a sinusoidal business cycle, a random business cycle and a random walk with drift. He finds that the budget is almost always balanced over the long term.

70 The study of Bodmer (2006) does not value the effectiveness of the fiscal rule against a certain numerical target. Neverthelesss he finds that the deficits of the 1989-2002 period would have been considerably smaller if the rule had been in place at that time already. He concludes this from a simulation where he uses provisional GDP data for year t − 1 as well as the revenue estimate and the GDP estimate for year t for every single year of the 1989-2002 time period. Under the assumption that the rule is complied with, he calculates the notional expenditure limit and the notional deficit which is then compared to the true deficit in year t. To contribute to the discussion about the effectiveness of the fiscal rule in the medium run we extend the post-intervention period up to 2013 and rerun the budget balance analysis. Here we need to discard Australia, Japan and New Zealand as revenue and eypenditure data is not available for 2013. However, their individual contribution to the Synthetic Switzerland is very small (see table 3.1). This leaves us with twenty countries in the donor pool. The pre-intervention RMSPE increases only slightly (0.446) and the w-weights of the remaining countries change only marginally, too. The impact of dropping these countries on the change in the development of Synthetic Switzerland can thus be neglected. Figure 3.9 shows that the clear distance between the cyclically adjusted budget balance of both Switzerland and Synthetic Switzerland persists up to 2013. The average treatment effect accounts for 4.706 percentage points. This suggests that deficits would have been larger without the rule and the rule is effective also in the medium run. However, one should be cautious in relying on this medium-term result as the Synthetic Control Method is able to capture the short run effect of a policy intervention. A large extension of the post-period, in contrast, runs the risk of capturing structual shocks of both Switzerland and Synthetic Switzerland that are not related to the rule’s introduction. However, the treatment effect lends support to the idea that the Swiss fiscal rule is effective in improving the budget balance also in the medium run.

3.5 Conclusions

The Swiss federal debt rule has become a role model for other countries. Firstly, Germany adapted it to its constitutional framework, before the Eurozone countries in the fiscal compact accepted to introduce their own national debt rules. Secondly, other countries, e.g., Israel followed the ideas of the Swiss debt rule in reforming their own legislation. In all these countries, the effectiveness of fiscal rules is politically challenged in order to revise them before they can show any effect. Even in the Swiss case, it is difficult to provide evidence regarding the effectiveness of the debt rule using traditional econometric methods. This paper is the first that provides evidence for the effectiveness of the Swiss balanced budget rule. In contrast to previous studies that try to uncover the impact of this fiscal

71 rule using simple simulation exercises, we rely on the Synthetic Control Method. We find that the rule improved the cyclically adjusted budget balance by about 3.6 percentage points of GDP on average in the five years after its introduction. This treatment effect is robust and persists also in the medium run. Concerning the effect on the debt ratio of both the federal government level and the general government level we cannot provide clear-cut results. For methodological reasons, this is mainly due to the difficulties arising from the 2003-05 transition period that causes an intersection of the debt trajectories in the post-intervention period. Additionally, the debt trend of the synthetic control group may be biased by the European debt crisis from 2008 onwards. Overall, the evidence provided in this paper strongly suggests a causal impact of the Swiss federal debt brake on federal finances. With lower structural deficits, the federal government could return to sound public finances. There is no reasons to believe that such intelligent fiscal rules have a different effect in Germany or in other countries. The op- position against fiscal rules may simply originate from the expectation that they actually restrain governments from being profligate.

72 3.6 Tables and Figures

State w-Weight State w-Weight State w-Weight Australia 0.009 Germany 0.004 Norway 0.006 Austria 0.006 Ireland 0.286 Portugal 0.014 Belgium 0.008 Israel 0.145 Slovak Republic 0.237 Canada 0.013 Italy 0.007 Spain 0.007 Czech Republic 0.003 Japan 0.007 Sweden 0.006 Denmark 0.011 Luxembourg 0.009 United Kingdom 0.007 Finland 0.005 the Netherlands 0.181 USA 0.009 France 0.015 New Zealand 0.005

Table 3.1: Cyclically Adjusted Budget Balance, w-Weights

Variable Treated Synthetic lnPOP 15.779 15.783 Unemp growth -4.33 -4.328 CAB 1995 -1.292 -1.297 CAB 1999 0.408 0.414 CAB 2000 -2.426 -2.422

Table 3.2: Cyclically Adjusted Budget Balance, Predictor Balance 2 0 -2 -4 -6 Cyclically Adj. Budget Balance (% of GDP) -8 1995 2000 2005 2010

Year

Switzerland Synthetic Switzerland

Figure 3.1: Cyclically Adjusted Budget Balance

73 France Israel Czech Republic Luxembourg Japan Canada Australia United Kingdom Italy Sweden Netherlands Slovak Republic Germany Finland USA Austria Belgium Spain Denmark New Zealand Ireland Norway Portugal Switzerland

0 2 4 6 8

Post-Period RMSPE / Pre-Period RMSPE

Figure 3.2: Cyclically Adjusted Budget Balance, Post-Pre-Ratios

France Israel Japan Canada Italy Australia Luxembourg Slovak Republic Sweden Finland United Kingdom Netherlands USA Spain Austria Belgium Denmark New Zealand Ireland Norway Portugal Switzerland

0 2 4 6 8

Post-Period RMSPE / Pre-Period RMSPE

Figure 3.3: Cyclically Adjusted Budget Balance (w/o CZE, DEU), Post-Pre-Ratios

France Israel Japan Canada Italy Finland Sweden Slovak Republic Australia United Kingdom Netherlands Luxembourg USA Spain Austria Belgium Denmark Ireland Norway Portugal Switzerland

0 2 4 6 8

Post-Period RMSPE / Pre-Period RMSPE

Figure 3.4: Cyclically Adjusted Budget Balance (w/o CZE, DEU, NZL), Post-Pre-Ratios

74 France Israel Finland Canada United Kingdom Netherlands Spain Slovak Republic Luxembourg Belgium Denmark Ireland Norway Australia Portugal USA Switzerland

0 2 4 6 8

Post-Period RMSPE / Pre-Period RMSPE

Figure 3.5: Cyclically Adjusted Budget Balance (w/o AUT, CZE, DEU, ITA, JPN, NZL, SWE), Post-Pre-Ratios

France Israel Canada Ireland Luxembourg United Kingdom USA Spain Slovak Republic Netherlands Belgium Denmark Portugal Australia Switzerland

0 2 4 6 8

Post-Period RMSPE / Pre-Period RMSPE

Figure 3.6: Cyclically Adjusted Budget Balance (w/o AUT, CZE, DEU, FIN, ITA, JPN, NOR, NZL, SWE), Post-Pre-Ratios

State w-Weight State w-Weight France 0.308 Luxembourg 0.571 Japan 0.031 Sweden 0.091

Table 3.3: Federal Government Debt Ratio, w-Weights

75 Variable Treated Synthetic POP65 14.946 14.816 Unemp rate 3.451 5.678 CGDR 1993 19.25 19.362 CGDR 1997 25.274 25.293 CGDR 2001 24.822 24.91

Table 3.4: Federal Government Debt Ratio, Predictor Balance 35 30 25 20 Central Government Debt Ratio (% of GDP) 1990 1995 2000 2005 2010

Year

Switzerland Synthetic Switzerland

Figure 3.7: Federal Government Debt Ratio

State w-Weight State w-Weight Denmark 0.328 Luxembourg 0.189 Japan 0.167 United Kingdom 0.316

Table 3.5: General Government Debt Ratio, w-Weights

Variable Treated Synthetic lnPOP 15.774 16.29 Unemp rate 3.451 5.461 GGDR 1993 49.6 49.67 GGDR 1996 60.5 60.433 GGDR 2001 61 61

Table 3.6: General Government Debt Ratio, Predictor Balance

76 Ccial dutdBde aac,etne post-period extended Balance, Budget Adjusted Cyclically 3.9: Figure GnrlGvrmn etRatio Debt Government General 3.8: Figure

Cyclically Adj. Budget Balance (% of GDP) General Government Debt Ratio (% of GDP)

-10 -5 0 5 50 60 70 80 2000 1990 1995 Switzerland Switzerland 2005 77 2000 Year Year Synthetic Switzerland Synthetic Switzerland 2010 2005 2015 2010 3.7 Appendix: Variables

CAB Cyclically Adjusted Budget Balance as Percentage of GDP, Source: OECD.Stat and Own Calculations.

CGDR Total Central Government Debt Ratio as Percentage of GDP, Stocks: Outstand- ing Amounts, Source: OECD.Stat.; for the United Kingdom we use Total Liabilities Excluding Insurance Technical Reserves as Percentage of GDP, Source: OECD Fis- cal Decentralisation Database.

GDP Growth Annual Growth Rate of GDP, Source: OECD.Stat.

GGDR General Government Debt Ratio as Percentage of GDP, Source: Historical Public Debt Database. lnGDP Natural Logarithm of GDP Per Head, US$, Constant Prices, Constant PPP’s, Reference Year 2005, Source: OECD.Stat and Own Calculations. lnPOP Natural Logarithm of Total Population, Source: World Development Indicators and Own Calculations.

Pop65 Population Aged 65 and Above as Percentage of Total Population, Source: World Development Indicators.

Trade Imports of Goods and Services Plus Exports of Goods and Services as Percentage of GDP, source: World Development Indicators and Own Calculations.

Unemp Growth Annual Growth Rate of Unemployment, Source: OECD.Stat - Economic Outlook No. 88 database and Own Calculations.

Unemp Rate Total Unemployment as Percentage of Total Labour Force, source: World Development Indicators.

78 4 Do Federalism Reforms in Belgium Cause Economic Growth?

4.1 Introduction

Belgium is an artiﬁcial creation made by actors that do not have much in common.

Although possibly controversial these words are a good description of the conditions present at the time of the state formation in 1830. Indeed, Dutch-speaking Flemings and French-speaking Walloons joined forces before 1830 in order to, e.g., dissociate from a French occupation or the protestant (northern) Netherlands.1 Regarding the internal (economic) affairs, however, differences were conspicuous. While the northern provinces were successfully engaged in maritime trade the Walloon region became a precursor of the beginning industrial revolution. Thus merchants from the north urged King William I. to facilitate free trade and remove economic borders, while the bourgeois from the south hoped for some protection against foreign competitors in order to develop their industries (Claus and Baumann, 1980; Hecking, 2003). That is, a very exceptional coalition fought for the independence of the southern provinces from the Netherlands: mainly the French- speaking elites as well as catholic clerics argued for an independent Belgian state. In the succeeding Belgian Revolution the French-speaking elites gained an influential role and so did the French language. Firstly, the revolutionists followed the French model that is based on a centralised state structure and civil law. To codify law they used the French language as it was, unlike the Dutch, a standardised language at that time. Accordingly, all lawyers and civil servants in the entire country needed to be French- speaking. This made it very unlikely for native Dutch-speaking Belgians to fill higher social positions. Secondly, due to the economic success of the coal and steel industry in Wallonia and the census suffrage, the French speaking part of the country was dominating and many Flemish had to cope with poor and unequal living conditions as peasants or miners in Wallonia. The Dutch-speaking Belgians thus felt more and more repressed

1Incited by the French revolution the idea of sovereignty gained so much importance that the ‘United States of Belgium’ where declared in 1790. However, that ambition could not be put into eﬀect, as –ﬁrstly– the Habsburgian army quickly defeated the insurgents and –secondly– the French army occupied the territory in 1792 and annexed it to the French republic (Clauss and Baumann, 1980). At the Congress of Vienna later in 1814/1815 the Belgian territories were combined with the Dutch provinces to form the United Kingdom of the Netherlands under the rule of the absolutistic calvinistic King William I of Orange-Nassau (Hecking, 2003).

79 regarding their language and culture. This conflict is also known as the language dispute.2 The language dispute could not be resolved but even aggravated and economic differences persisted over decades. In order to improve or even enable the coexistence of Flem- ings and Walloons within one single polity Belgium transformed its institutional setup seriously and turned it from a unitary state into a federal system. Given the circumstances suggested above a constitutional reform in Belgium had to meet three important challenges. Firstly, it need to cope with the language dispute. Views could not be more diverse in that respect. From the Flemish point of view equal rights for both languages need to be conferred. That is, the Flemish culture must be recognised as an independent and equal part of the Belgian state. Solid and immutable territorial boundaries for rights of language use are considered a prerequisite for this culture to exist. Adjustments of language boundaries would denote a sustainable and therefore ultimate discrimination of the endangered culture. In that respect the matter of language reflects a corporate feeling and embodies a territorial component. The Francophone point of view, however, is diametrical different from the Flemish perspective. As the decision on language use is a very individual one, everyone should be able to decide freely which language to use and political as well as administrative boundaries should reflect these preferences. Consequently, territorial boundaries should follow the distribution of individual language preferences and should thus be adjusted on a regular basis. Secondly, the economic differences had to be mastered. As mentioned, Wallonia became a precursor of the industrial revolution. The coal and steel industry grew steadily. Flanders, in contrast, lagged behind in terms of its economic development and remained mainly agricultural. The distinction between the rich and developed south versus the poor and rural north reversed after World War II. The coal and steel industry declined while the Flemish harbours began flourishing. Consequently, the developing service sector in Flanders became the new economic powerhouse of Belgium and since 1966 Flanders accounted for the larger part of the GDP with this share growing steadily (Hecking, 2003). Wallonia more and more became dependent on intra-Belgian solidarity, e.g., in terms of financial means. Finally, a Belgian state reform need to tackle the problem of how to safeguard individual rights when they are incongruent and overlapping with territorial demands. Most notably the municipalities of the region Brussels-Capital need to meet this challenge. Regionally situated in Flanders the majority of the citizens is French-speaking. As the French-speaking population spreads to the suburbs of Brussels they are hit by the strict language regulations of surrounding Flanders. Symptomatic of the problems of incongru-

2A small German-speaking minority in the East-Belgian area of Eupen and St. Vieth became Belgian later in 1920 as a result of World War I. As this minority covers about 70.000 citizens it is clear that it hardly inﬂuences the dispute of the main actors (Hecking, 2003).

80 ent regional and cultural demarcations was the long-lasting dispute over the constituency of Brussels-Halle-Vilvoorde (BHV). Until 2012 this constituency comprised the bilingual region Brussels-Capital as well as the arrondissement Halle-Villevoorde which is part of the region Flanders. Where the region Brussels-Capital formally is a bilingual region with a de facto majority of francophone residents the region Flanders is purely Dutch-speaking. Actually, six institutional reforms were executed in 1970, 1980, 1988/89, 1993, 2001 and 2011-13. In 1970 a Dutch-speaking, a francophone and a germanophone community and three (economic) regions (Flanders, Wallonia and Brussels-Capital) were established and decision-making power in the realm of cultural policy was delegated to the communities. In 1980 decision-making power in the realm of economic policy was assigned to the regions Flanders and Wallonia and further competencies were delegated to the communities. The third state reform (1988/89) initially assigned discretionary competencies to the region Brussels-Capital and further increased the spending power of Flanders, Wallonia and the communities. Beyond that, it redesigned the sources of revenue of both regions and communities. In 1993 the financial means available to the communities were raised. In 2001 this funds were further increased and taxing authority as well as spending power was transfered to the regions. The sixth state reform (2011-13) implemented the partition of the electoral and judicial district Brussels-Halle-Vilvoorde (BHV). Financial means of the region Brussels-Capital were increased and the decision-making power of regions and communities was widened. Finally, fiscal autonomy of the regions was expanded. In summary, it can be said that Belgium made some very innovate - and admittedly complex - institutional arrangements. The reforms responded to the language dispute via the introduction of three cultural communities. The economic differences where addressed by the introduction of the regions. Difficulties stemming from the overlap of individual rights were answered by granting the Brussels-Capital region a special status. Ultimately, devolution in Belgium resulted in a unique two-tier federal government structure. The aim of this paper is to contribute to the literature that uncovers the economic effects of political institutions. To this, we exploit the federalism reforms of 1989, 1993 and 2001 and investigate the growth effects at the national as well as the regional level. Justifications of why we focus on these three reforms can be found in section 4.2.7. In using observational data related to rearrangements of the state structure in Belgium we appreciate that these federal reforms are based on political compromise. The question thus is whether we can find an effect on the economy if the agreements deviate from what one would declare an ideal federalism reform. We value the fact that the reforms were enacted for political reasons an advantage as this minimises the self-selection problem. That is, we can preclude that a positive impact of federalism reform in Belgium is caused by the objective to promote growth. To see what can be expected from the reforms in Belgium we review the theoretical as

81 well as the empirical literature in the next paragraphs.3 A seminal argument regarding decentralisation was provided by Tiebout (1956). In his model an individual chooses were to live within a country. It observes different locations with different bundles of taxes and public good levels. Eventually, it chooses the location with the bundle that best fits its preferences. That is, a decentralised form of government can provide a variety of public good levels and can thus more efficiently meet the preferences of the consumers (Oates, 1972). An increase in the number of options thus is expected to be welfare enhancing. With respect to the supply side of the economy Oates (1993, 1999) argues that measures of economic policy conducted at the regional or local level can better be tailored to pecularities of the regional or local economy compared to similar measures conducted at the central level. Likewise, measures of economic policy implemented at the regional level might better support necessary adjustments to the economic structure via the selection of policy instruments that correspond to regional needs (Baskaran et al., 2014). Again, this is expected to positively affect economic growth. Beyond that, decentralisation can lead to better (economic) policy outcomes. Oates (1972, 1999) states that subnational entities can act as fields for experimentation in cases where the optimal level (design) of a public good (service) is unidentified. Due to competition between the entities the good practice of one subnational jurisdiction spreads over to other subnational jurisdictions or even the federal level. This mechanism of spillover is more precisely worked out by Besley and Case (1995) who call it ‘yardstick competition’. The professional competence of a politician is pit against the (economic) outcomes of her neighbours’ policies. In general this can be expected to have a positive effect on economic growth. Another idea states that decentralisation and the associated competition can have a bearing on the behaviour of politicians such that it reduces rent-seeking. It starts from the assumption that government maximises its net surplus. If to this end a subnational government executes immoderate taxation, the citizens respond with out-migration. Out- migration, however, is more difficult to implement if there exists only one (central) government in a polity. Decentralisation can thus indirectly help to curb a Leviathan government and its taxing power (Brennan and Buchanan, 1980). Similarly, a federal government structure can help to prevent the government from restricting market-based economic activity. If there are subnational jurisdictions and if economic policy competencies are settled at the regional level, enterprises can move to another region if government aims at confiscating the wealth of enterprises. In this sense federalism helps to preserve the market economy (Weingast, 1995). Both aspects can be expected to positively affect growth. Decentralisation could also have a negative impact on growth. Cai and Treisman, 2004 stress that regional governments may shield enterprises from tax collection or (environ-

3This review is guided by Baskaran et al. (2014) who more comprehensively discuss the literature.

82 mental protection) policies of the federal government in order to attract mobile capital. Eventually, the federal government cannot push through its market regulation policy and the federal state structure contributes to the corrosion of the state. Beyond that, a federal state structure might be associated with a higher level of corruption. Martinez-Vazquez and McNab (2003) point to the larger number of politicians who can be bribed. However, local politicians are closer to their electorate and can thus better be monitored by the citizens. Both corrosion and corruption suggest a negative impact on growth. Besides the theoretical literature there exists a vast number of empirical studies that investigate the effect of decentralisation on economic growth. Most of these studies use the share of subnational spending (or revenue) as percentage of total government spending (or revenue) as dependent variable. Davoodi and Zou (1998) cannot find a significant relation between fiscal decentralisation and the growth rate in developed countries by using cross- country data.4 They employ a panel of 46 countries from 1970 to 1989. The same holds for the study of Xie et al. (1999) that is built on time series data on federal, state and local spending in the USA from 1948 to 1994. Akai and Sakati (2002), in contrast, find a positive impact of fiscal decentralisation on the growth rate at the local level.5 They ascribe this result to the utilisation of US state level data rather than cross-country data which is not subject to cultural differences between observations. However, when they use ‘own’ revenues of local governments as percentage of total local revenues the effect on growth turns insignificant. In line with this Thornton (2007) accentuates that the share of subnational revenues as percentage of overall revenues might be a poor measure of fiscal decentralisation. That is, modifications of tax sharing arrangements affect subnational revenues but do not necessarily imply a real increase or decrease in taxing authority. By using ‘own’ revenues of subnational governments as percentage of total revenues he finds an insignificant effect of fiscal decentralisation on the growth rate. Consequently, Baskaran and Feld (2013) use two revenue measures in a cross-country setting. The first measure captures the share of subnational revenues as percentage of general government revenues where the second measure uses revenues that emerge from taxation that is controlled by the subnational authorities. However, they find an insignificant negative effect when using the first measure and a significant negative effect when using the second measure. The endogeneity problem is explicitly addressed by Iimi (2005) who uses lagged data of the independent variables to instrument the explained variable. In a cross-country setting he finds the relation between fiscal decentralisation and growth to be positive and significant.6 Alternatively, the source of measurement errors might be the simultaneous execution of political or administrative decentralisation. When taking this into account

4Fiscal decentralisation is captured by the share of subnational spending as percentage of total spending. 5Fiscal decentralisation is captured by the share of local revenues (spending) as percentage of local plus state revenues (spending). 6Fiscal decentralisation is captured by the share of local government expenditures as percentage of total government expenditures.

83 Rodriguez-Pose and Ezcurra (2011) find the relation between fiscal decentralisation and the growth rate to be significantly negative in a panel study of 21 OECD countries in the 1990-2005 period.7 Thießen (2003) addresses the nonlinearity of the relation between fiscal decentralisation and economic growth. A nonlinear relation implies that fiscal decentralisation promotes growth when it is at low levels but it dampens growth when it is at high levels. He finds clear indications of a nonlinear relation when using the share of subnational government expenditures as percentage of overall government expenditures. Eventually, there is no consensus about the set of explaining variables an empirical growth model must contain in order to correctly estimate the effect of fiscal federalism on growth. Asatryan and Feld (2014) employ a bayesian model averaging approach on a panel of 23 OECD countries over the 1975-2000 period and find that there is no relation between fiscal federalism and economic growth at all. Neither positive nor negative.8 Moreover, their quantitative literature review indicates that the results of previous empirical studies depend on the precise decentralisation measure used as well as the inclusion or exclusion of fixed effects. Thus they argue that an unambiguous conclusion regarding the effect of decentralisation on growth cannot be drawn as long as it is contested which measure should be used ideally and whether an empirical model without fixed effects must be regarded as mis-specified. Regarding federalisation in Belgium there is to our best knowledge one study that explicitly captures the effect of the (entire) institutional reform on economic growth. Pagano (2013) compares GDP growth rates of Belgium with GDP growth rates of the EU-15, the Euro Zone, the Netherlands, Luxembourg, France and Germany. Additionally, he employs Pearson’s correlation coefficient and runs some basic regressions but cannot find a significant difference between these growth rates. The chapter proceeds as follows: in section 4.2 we describe the contents of all federalism reforms carried out in Belgium and present the hypotheses we deduce from a confrontation of the reform contents with the literature. Section 4.3 reports the data we use, section 4.4 presents the results of our empirical analysis and section 4.5 concludes.

4.2 Federalism Reforms in Belgium

4.2.1 The First State Reform (1970)

The reform of the year 1970 (law of 24th December 1970) amended the constitution in stating that Belgium will consist of three so-called ‘communities’ and three regions

7Fiscal decentralisation is captured by the share of subnational revenues (expenditures) as percentage of total government revenues (expenditures). 8This conclusion can also be drawn from the qualitative literature review carried out by Baskaran et al. (2014).

84 from now on. Each entity will have its own parliament that will take decisions in its area of competence. Article 59bis of the constitution established a Dutch-speaking, a francophone and a germanophone community and delegated considerable responsibilities for language and cultural affairs.9 In the bilingual zone of Brussels these competencies were executed in parallel by the Dutch-speaking and the francophone community. The capital Brussels thus very well illustrates the main idea of the communities: shaping policy that affect individuals shall be possible independent of territorial boarders. French- or Dutch-speaking citizens of Brussels shall live close to each other in the same area and still be members of different ‘communities’. Article 107quater stipulated the establishment of the regions Flanders, Wallonia and Brussels-Capital. However, it was not realised, as the course of the borders was too much contested. The foundation of both communities and regions marked the starting point for the unique Belgian two-tier model of sub-national autonomy (Clauss and Baumann, 1980; Hecking, 2003). While the formation of communities was meant to limit the conflict of different languages and cultures, it carried unforeseen additional consequences. The party system is an important element of consensual politics in Belgium. Until then a compromise among the three major parties Christian-Democrats, Liberals and Socialists was found, no matter how difficult it was to obtain. As a consequence of the 1970 state reform, however, pressure emerged for parties to split along these new boundaries and all parties split into a Flemish, Walloon – and some – even into a German branch (Clauss and Baumann, 1980). The Belgian party system was falling to pieces at that time and became the most fragmented one in Europe. Thus it can be stated that this constitutional reform ignited more centrifugal forces than it eased tensions. However, the reform also introduced a new legislative procedure that limits the possibility for one group to dominate the others. So-called ‘special majority laws’ were introduced. In order to pass such a law two-thirds of all members of the House of Representatives plus the plurality of the members of each language group must agree. Hence, not even in case of a demographic prevalence of one language group this group can dominate the others (Deschouwer, 2012).

4.2.2 The Second State Reform (1980)

In the context of the second reform (laws of 8th and 9th August in 1980) two autonomous regions – Flanders and Wallonia – were established and economic decision-making power was assigned to them. Among these were issues of territorial relevance, like regional planning, social housing, local environmental protection, economic policy, public transport, road infrastructure and the surveillance of the local administration. As the area of the germanophone community is too small to conduct regional policy in an eﬃcient way, these tasks are accomplished by the neighbouring region Wallonia. The regions were funded by

9The competence for school education was not devolved.

85 transfers from the central government that were assigned by a formula based on criteria like population, territorial size and revenues from personal income tax. These transfers represented more than 90 percent of the revenues of Flanders and Wallonia. The status of the communities was strengthened by an amendment of article 59bis of the constitution. That is, the communities were allocated additional competencies in the realms of education, health, culture and services for the disabled. Hence, they are not only in charge of cultural issues, but also of individual-related issues since then (Hecking, 2003).10 Communities and regions also gained taxing authority. However, without actual importance to the communities as there is no territory on which they could enforce tax collection. And without actual importance for the regions as they were not allowed to impose a tax on a base that is already exploited by the central government (Gerard, 2001; Verdonck and Deschouwer, 2003). As the negotiating parties could not reach an agreement on the status of Brussels and its peripheral municipalities the capital did not become an autonomous region like Flanders or Wallonia. Respective decisions on regional policy were furthermore accomplished by the central government. In order to demarcate the city of Brussels a provisionary border between Brussels and the region of Flanders was set up (Hecking, 2003). Shortly after the second constitutional reform the Flemish side merged the institutions of the Dutch-speaking community and the Flemish region into one Flemish parliament and one Flemish government. That is, one parliament and one government execute both individual as well as territorial policies since then.11 On the Walloon side, though, there still lasts a distinction between these policies with divided responsibilities being anchored in diﬀerent institutions (Hecking, 2003).

4.2.3 The Third State Reform (1988/89)

Under the third reform (law of 12th January of 1989) the region Brussels-Capital was established as third region with equal competencies under two conditions: the region was kept in sharp and durable regional boundaries and the Dutch-speaking minority received particular safeguards and participation pari passu in the administration of the region. Additionally, further responsibilities were delegated to the communities and regions (laws of 8th August of 1988). That is, communities were given far-reaching exclusive competencies in all matters of education – now also including school and university education – as well as media policy. The regions received far-reaching exclusive competencies in the ﬁelds of regional economic policy, housing issues, environmental policy and public services (Gerard, 2001; Hecking, 2003). Finally, the sources of revenue of both communities and

10The social security system, however, is under the control of the central government. 11However, the Dutch-speaking community and the Flemish region exist further on from a legal point of view (Alen and Ergec, 1994).

86 regions were redesigned in the special financing act of 16th January of 1989. This law settles that communities receive tranfers from the federal government flowing from four different sources. In terms of the value added tax (VAT) transfer the communities yearly receive a fixed amount of financial means which grows in dependence of the consumer price index and is additionally adapted to the change in the (country-wide) number of inhabitants under the age of 18 (students).12 The resulting amount of funds is finally split between the communities following the share in the number of students of the year 1989.13 Actually, the relation to the value-added tax is purely notional. There is no reference to VAT revenue. The personal income tax (PIT) transfer, the second source of funds, is very similar to the VAT transfer. It is also based on a yearly fixed amount of financial means and it also grows in dependence of the consumer price index.14 However, the PIT transfer is divided between the communities based on their contribution to the overall (national) personal income tax revenue (derivation principle) instead of the students population. In the capital Brussels the commission of the francophone community (COCOF) receives 80 percent of the revenue that accrues to Brussels where the commission of the Dutch- speaking community (VGC) receives 20 percent of the revenue that accrues to the capital. The radio-TV fee is collected by the federal government and then refunded to the communities in accordance to its origin. Again, revenues in the capital Brussels are allocated between the communities COCOF and VGC following the 80/20 rule. Finally, the communities receive a yearly funding from the federal government in return for teaching foreign students at the universities. This amount also grows in line with the consumer price index (Moniteur Belge, 1989).15 Moreover, communitities are allowed to levy own-source taxes. But for the citizens of Brussels it holds that they cannot be obliged to tie themselves to one of the two communities. Thus citizens could avoid to be taxed by turning to the other community. This is considered to be the reason why communities do not levy taxes. For all sources of revenue it holds that the utilisation is unrestricted. Communitites are totally free in where to allocate the funds (Gerard, 2001; Van der Stichele and Verdonck, 2001). The special financing act of 16th January 1989 also redesigned the sources of revenue of the regions flowing from four different sources. The first source, the personal income tax (PIT) transfer, is a yearly fixed amount of financial means that grows in line with the consumer price index.16 The transfer is allocated between the regions according to their contribution to the national personal income tax revenue (derivation principle).

12The VAT transfer originally amounted to 296 385 million BEF (7 347 million Euro). 13The francophone community receives 42.45% of the total amount where the Dutch-speaking community receives 57.55% of the total amount. 14The PIT transfer originally amounted to 85 186 million BEF (2 111 million Euro). 15The francophone community received 1 200 million BEF (29 million Euro) where the Dutch-speaking community received 300 million BEF (7 million Euro) in 1989. 16The PIT transfer originally amounted to 234 691 million BEF (5 817 million Euro).

87 Beyond that, for regions with PIT revenue below the national average (per capita) an equalisation mechanism (the national solidarity measure) was installed to close the gap.17 This amount changes in line with the consumer price index. The third source of revenue are regional taxes. They are collected by the central government and regions have taxing authority to a variable extent. In terms of the tax on gambling and betting, the tax on coin-operated amusement devices and the tax on the opening of drinking establishments regions have full taxing authority. Regions can set the tax rate but cannot redesign the tax base in case of the real estate tax and the inheritance tax. In terms of the registration fee on the transfer for payment of real property both the tax rate and the tax base are set by the central government.18 Beyond that, regions receive income from charging, e.g., the hunting and fishing fee or the forestry operation fee (own source but non-tax revenue). Finally, a region receives a specific purpose grant if it creates a job for a former unemployed person. The central government awards this grant in recognition of unemployment benefits that it does not need to pay anymore. In addition to these four sources of revenue regions are allowed to raise a piggyback tax on top of the personal income tax. This surcharge is designed as share of the amount of personal income tax paid by a tax payer. From 1994 onwards regions are furthermore allowed to grant a discount on the personal income tax. The total amount of all granted discounts, however, must be smaller than the amount of the PIT transfer the region receives from the federal government (Moniteur Belge, 1989). Again, there are no restrictions regarding the utilisation of these transfers. Thus, regions are totally free in where to allocate the funds (Gerard, 2001; Van der Stichele and Verdonck, 2001). What can be concluded from these arrangements is that all sub-national entities heavily depend on fiscal transfers granted by the federal government. This especially holds for the communities. The regions, by and large, are allowed to collect piggyback taxes and tax reductions. Van der Stichele and Verdonck (2001), however, state that they never made use of this means.

4.2.4 The Fourth State Reform (1993)

Between 1989 and 1993 it turned out that the francophone community was incapable of ﬁnancing its educational services. On the one hand wages in the education sector consti- tuted a large part of the community’s budget. On the other hand the VAT transfer (the largest source of revenue) was not tied to (real) GNP growth which led to the underfunding of the community. To solve this problem governing parties reached an agreement (Saint-

17An aﬄicted region anually receives 468 BEF (11 Euro) per resident and per percentage point diﬀerence between the PIT revenue of the respective region and the national average. 18This fee is only partly refunded to the regions. However, the regions are allowed to collect piggyback taxes or grant tax reductions on this fee.

88 Michel) and passed a special financing act in 1993. However, the basic configuration of the financing scheme introduced in 1989 was maintained. As part of the financing act the PIT transfer of the communities increased once-off by 4 500 million BEF (111 million Euro) in 1993. Additionally, it was minorly tied to (real) GNP growth in 1994. However, this link increased gradually. It started at ten percent in 1994 and ran up to 97.5 percent in 1999 (ZDDU,¨ 2008).19 The radio-TV fee is now collected at the community level and revenue is entirely payed back to the communities.20 Changing the tax base and the tax rate requires the approval of the communities but still remains the competence of the central government. The allocation of revenue in the capital Brussels still follows the 80/20 rule. In response to the underfunding the francophone community transfered some of its spending responsibilities to the Walloon Region and the francophone community of Brus- sels, COCOF (the Saint-Quentin agreement). In return the Walloon Region and the COCOF received a financial transfer from the francophone community. However, this transfer was insufficient to cover the expenditures that accrued to the Walloon Region and the COCOF (Gerard, 2001; Van der Stichele and Verdonck, 2001). Besides the increase in financial means the 1993 reform amended article one of the constitution such that it now begins with the statement that ‘Belgium is a federal state composed of Communities and Regions’. Additionally, the former bilingual province of Brabant was divided into the provinces Flemish Brabant and Walloon Brabant. The reform also introduced direct elections of the regional councils and the councils of the communities, so that they are no longer merely committees of the federal House of Rep- resentatives. Actually, the councils received full decision-making power in the realm of international affairs. This allocation of responsibilities also makes the Belgian federalism unique as nowhere else subnational levels have such far-reaching competencies with respect to international trade (Hecking, 2003).

4.2.5 The Fifth State Reform (2001)

The 2001 state reform once more increased the amount of ﬁnancial means available to the communities. On top of the basic VAT transfer settled in 1989 (part A) communities received 198 million Euro in 2002, 149 million Euro in 2003 and 2004, 372 million Euro in 2005 and 124 million Euro in 2006 (part B). From 2007 to 2011 the communities annually received 25 million Euro (part C). Part B and C grow in line with the consumer price index. Part C is additionally tied to (real) GNP growth. Since 2012 the entire VAT transfer is tied to (real) GNP growth. Where the formula that splits part B and

19The PIT transfer of the regions also started to grow in line with real GNP in 1994. The link increased gradually until 1999 following the same schedule (ZDDU,¨ 2008). 20From some authors it can be concluded that this fee was only incompletely refunded to the communities before 1993. However, no details are provided (Gerard, 2001; Van der Stichele and Verdonck, 2001).

89 C between the communities increasingly follows the derivation principle, part A of the VAT transfer will further on be split between the communities following the share in the number of school-aged children (ZDDU,¨ 2008). Moreover, regions were granted taxing authority. That is, they are fully in charge of the base of the inheritance tax and the real estate tax. Regions also can set the tax rate and the tax base of the registration fee and thus retain the revenues to the full extent. The latter implies a financial loss for the federal government which is compensated by a permanent reduction of the PIT transfer amount. The radio-TV fee was turned into a regional tax and regions are fully in charge of the tax rate and the tax base (ZDDU,¨ 2008). Finally, piggyback tax arrangements (see section 4.2.3) were reformed. That is, regions are allowed to charge a lump sum premium or give a lump sum discount on regional taxes as well as the personal income tax. The granting of taxing authority to the regions under the fifth state reform was expected to perceptibly increase the level of tax competition. In order to limit tax competition premiums and reductions on the personal income tax were limited to 3.25 percent of the tax yield collected in the respective region. As of 2004 the limit accounts for 6.75 percent. Moreover, regions are not allowed to decrease the progressivity of the personal income tax scale but they can increase it (Van der Stichele and Verdonck, 2001; Verdonck and Deschouwer, 2003). Beyond raising financial means of the communities and assigning taxing authority to the regions the reform delegated further spending authorities to the regions in the realms of agriculture, fishery, foreign policy and the surveillance of local policy. However, to a limited extent (Blöchliger and Vammalle, 2012). The arrangement of this federalism reform impressively shows that shaping policy in Belgium is very much based on making concessions in return for enforcing own interests. On the one hand the francophone community once more pointed out that the grants it received were not sufficient to provide a proper level of education services. Thus, the francophone community demanded in increase in financial means.21 On the other hand the relatively prosperous Flanders took a stand for both an expansion of spending authorities and an increase in taxing authority for the regions to be able to develop its regional economy independent of the economic situation in Wallonia. Finally, the negotiating parties made the Lambermont agreement and passed a special act in 2001 (Blöchliger and Vammalle, 2012; Hecking, 2003).

21The reason for this insuﬃcient amount of funds might be seen in the fact that the VAT transfer of the communities was not tied to the GNP growth in 1993. Besides that it might be relevant that the region Wallonia and the francophone community did not merge so far. Consequently, the possibility to pool ﬁnancial means did not exist (Verdonck and Deschouwer, 2003).

90 4.2.6 The Sixth State Reform (2011-13)

In the context of the sixth state reform the electoral and judicial district Brussels-Halle- Volvoorde (BHV) was split. While the electoral district of Brussels now is congruent with the region Brussels-Capital the arrondissment Halle-Vilvoorde now belongs the electoral district Flemish Brabant which is congruent with the province Flemish Brabant. Thus, this was another step towards a clear demarcation between the two language groups. In return, the metropolitan area Brussels was established. However, it is not settled yet what the competencies of this area should be. Beyond that, the communities were given further competencies in the realm of health care and they were initially assigned the responsibility for children’s allowances. That is, the division of the nationwide social security system has been partly enacted (Gerard, 2014). Regions were assigned further authorities with respect to professional training, economic policy, energy politics, agricultural policy, city planning, housing and local administration. With respect to the realisation of revenues tax authority was extended. While premiums and reductions on the personal income tax were limited to 6.75 percent of the tax yield under the ﬁfth state reform (2001) these limitations were abolished in 2011. The region Brussels-Capital was granted additional ﬁnancial means in compensation for its loss in tax revenue due to inbound commuters (Reuchamps, 2013; Gerard, 2014).

4.2.7 Hypotheses

As mentioned previously, we investigate the growth effects of the 1989 reform, the 1993 reform and the 2001 reform. In principle, we would prefer to exploit all state reforms. However, we cannot investigate the 1970 reform as well as the 1980 reform for data availability reasons. That is, data goes back to 1980. Thus the 1970 reform is not covered by data at all. Likewise, we cannot consider the second state reform (1980) as data does not embrace the pre-intervention period. This period is used to approximate the outcome variable of the treated unit. Finally, we cannot utilise the sixth state reform (2011-13) as data ends in 2012. That is, data does not comprise the post-intervention period which means that we cannot estimate a possible treatment effect. In the following we thus hypothesise what the growth effect of the 1989 reform, the 1993 reform and the 2001 reform will be. Basically we expect a positive growth effect of federalism reforms as local governments can better respond to local needs and as regional governments can better tailor economic policy measures to regional challenges (Oates, 1972, 1993, 1999). This can be assumed to hold for the communities as they might better be able to adapt their services to the needs of the Flemings and Walloons. And it should apply to the regions as Flanders and Wallonia differ with respect to their economic structure. Regions thus should be able to choose policy instruments that correspond to the respective economic context.

91 With respect to Wallonia decentralisation can be assumed to be helpful to cope with the transformation of the economic structure. Actually, Wallonia demanded more regional autonomy based on this argument (Verdonck and Deschouwer, 2003). The 1989 state reform delegated far-reaching exclusive decision-making power to the communities and the regions. Where regions were allowed to exclusively decide on regional economic policy affairs communities were granted responsibility in all matters of education. Thus we expect a positive growth effect of the 1989 reform. Beyond that, the revenue sources of both the communities and the regions were redesigned. However, all subnational entities still heavily hinge on transfers granted by the central government. The amounts of transfers can hardly be influenced by this entities. Thus, we do not expect a growth effect of the changes made with respect to the sources of revenue. The 1993 reform increased the amount of financial means available to the communities. The francophone community called for such an increase as it was incapable of financing its educational services. Essentially, the reform complied with this demand. Jennes (2014) thus classifies these transfers an (implicit) bailout while Oates (1999) points out that local decision making probably becomes lavish if revenues intensively depend on transfers from the central government. As further spending authorities have not been delegated we assume that the provision of public services is less efficient after the reform. Consequently, we expect a negative growth effect of the 1993 reform. The reform of 2001 also increased the amount of financial means available to the communities. Beyond that, it considerably granted taxing authority to the regions. Where the former can be expected to be growth-deteriorative the latter can be expected to be growth-enhancing. However, the reform implemented some measures that aim at miti- gating the level of tax competition. Eventually, it is an empirical matter which effect dominates.

4.3 Data

To capture the growth effects of the 1989 reform, 1993 reform and the 2001 reform in Bel- gium, Flanders and Wallonia we use real GDP per capita as dependent variable. Gross domestic product is defined as total gross value added plus taxes minus subsidies on products (both payable or receivable per unit of some good or service produced or transacted). The resulting GDP is then deflated to 2005 constant Euros. We use the level of economic performance as we aim at investigating the once-off economic effects, that is the growth effects, of several federalism reforms.22 In choosing variables that are good predictors of gross domestic product we basically follow the literature. Firstly, we use a measure of capital formation and call this variable

22Besides that, we cannot use the growth rate of GDP as the Synthetic Control Method emulates the outcome variable over several years. As the growth rate evolves rather volatile an approximation of the Belgian GDP with data of some comparison units is hardly possible.

92 ‘capital’. We use gross fixed capital formation which consists of resident producers’ ac- quisitions of fixed assets less disposals plus certain additions to the value of non-produced assets realised by the productive activity of producers or institutional units. This variable is given in 2005 constant Euros. We divide this variable by the gross domestic product in order to obtain an investment ratio. Secondly, we use the number of all persons who are permanently (at least one year) settled in the respective territory (even when temporarily absent). We call this variable ‘population’. In contrast to other empirical studies we do not use the population growth rate. This is because our dependent variable captures the level of economic performance rather than growth rate of GDP. Thirdly, we employ the AP OP dependency ratio. This is calculated from the formula DEP = 1 − POP . The quotient’s numerator AP OP gives the number of both employed and unemployed people (active population) and thus excludes children, students and pensioners. The denominator POP is the population variable just described. Fourthly, we use the unemployment rate. This variable is calculated from the formula UNEMP = 1 − EMP . The variable EMP gives the share of employed persons in the number of active persons (AP OP ) and thus comprises all persons engaged in some productive activity either as employee or self-employed. Combining the employment measure and the population measure in one variable might be problematic as the first counts where people live while the second counts where people work. This might lead to some distortions stemming from commuting. However, we cannot subtract such effects out. Finally, we employ gross value added of certain sectors of the economy. That is, we divide gross value added of a respective sector by gross value added of the total economy. Gross value added is the net result of output valued at basic prices less intermediate consumption valued at purchasers’ prices deflated to 2005 constant Euros. Output entails the products created during the respective time period. Intermediate consumption is the value of goods and services consumed as inputs in the same time period. We consider gross value added in the agricultural sector, the industry sector, the construction sector, the services sector, the financial and business services sector as well as the non-market services sector. The construction sector is part of the industry sector with the difference refering to manufacturing and energy. Likewise, the financial and business services sector is part of the services sector with the latter additionally covering wholesale trade, retail trade, transport, accomodation and food service activities, information and communication. The financial and business services sector entails financial and insurance activities, real estate activities, scientific and technical service activities and administrative service activities. The non-market services sector covers public administration, defence, education, human health and social work activities but also services in terms of arts, entertainment and recreation. These predictor variables are always accompanied by the lagged outcome variable of certain years which often is the first and the last year of the pre-intervention period. Among these predictors we search for a combination that best supports the

93 approximation of the treated unit’s real data. All raw data we use is taken from the European Regional Database provided by Cam- bridge Econometrics. This database entails several variables regarding population, employment and value added following the NUTS-classification.23 The database covers the 1980-2012 time span for the EU-27 countries plus Norway.24 Data for the East European countries, Malta and Cyprus (EU27 minus EU15) starts in 1990. With respect to Ger- many regional data prior to 1991 is provided for the western Länder. Since 1991 regional data is also provided for the five eastern Länderand the capital Berlin. Similarly, national level data for Germany includes these five Länderand Berlin as of 1991. To investigate the growth effect in Belgium we use national level data (NUTS-0 level). To capture the growth effect in Flanders and Wallonia we use regional data of the NUTS-1 classification. For some countries data of the NUTS-1 level are not provided as they are small in terms of their territory. This applies to Denmark, Ireland, Latvia, Lithuania, Luxembourg, Malta and Norway. In these cases we use national level data instead.

4.4 Empirical Analysis

This section provides the results of our analysis. Section 4.4.1 contains the results of the 1989 reform analysis and section 4.4.2 presents the results of the 1993 reform analysis. Section 4.4.3 comprises the results of the 2001 reform analysis. In each section we investigate the eﬀect of federalism at the state level (Belgium) as well as for Flanders and Wallonia separately (regional level). The capital Brussels constitutes a separate region since 1989. However, both communities oﬀer services in Brussels (via the commissions COCOF and VGC). As we cannot disentangle these activities the capital Brussels is excluded from our analysis in principle. We exhibit the composition of Synthetic Belgium but due to the large number of subnational comparison units we do not depict the composition of Synthetic Flanders and Synthetic Wallonia. The set of countries and subnational entities (comparison units) used for the analysis are given in appendix II.

4.4.1 The Growth Eﬀect of the 1989 Reform

As mentioned in section 4.3 we search for a combination of predictor variables and lagged outcomes in order to find an approximation of real GDP per capita that minimses the pre-intervention RMSPE. With respect to Belgium we find approximations that exhibit a very small pre-intervention RMSPE in the range of 43 to 85. These approximations indicate a small negative growth effect of the 1989 reform. Approximations with a pre- intervention RMSPE only slightly larger (in the range of 115 to 138) exhibit no growth effect of the 1989. That is, the GDP trajectory of Synthetic Belgium runs very close to

23Data we use follow the NUTS-classiﬁcation of the year 2006. 24Treated units are in no case part of the donor pool.

94 the GDP trajectory of Belgium in the post-intervention period. Where both France and the Netherlands contribute to the synthetic control group with relatively large w-weights under the former set of approximations France contributes to Synthetic Belgium with very large w-weights under the latter set of approximations.25 That is, we do not find a definite growth effect of the 1989 reform in Belgium among the approximations that exhibit a small pre-intervention RMSPE. Similarly, the search for an approximation that minimises the pre-intervention RMSPE does not lead to a definite treatment effect in the case of Wallonia. We find approximations that exhibit a very small pre-intervention RMSPE in the range of 46 to 56. These approximations suggest a positive growth effect of the 1989 reform. The region Greece- Kentriki contributes to Synthetic Wallonia with a relatively large w-weight in the range of 0.398 to 0.474. If we include the predictor variable ‘gross value added of the services sector’ Greece-Kentriki still contributes to Synthetic Wallonia with a large w-weight in the range of 0.424 to 0.428. However, the w-weight of Ireland increases considerably (to the range of 0.449 to 0.453) and the treatment effect turns negative. The pre-intervention RMSPE is very small also in the latter set of approximations (in the range of 46 to 51). The treatment effect of the 1989 reform in Wallonia thus clearly depends on the inclusion of one single predictor variable. That is, we do not find a definite growth effect of the 1989 reform in Wallonia among the approximations that exhibit a small pre-intervention RMSPE. The search for an approximation of real GDP per capita that minimises the pre- intervention RMSPE results in a distinct treatment effect in the case of Flanders. To predict real GDP we use gross value added of the non-market servives sector as well as gross value added of the industry sector. These predictors are accompanied by lagged real GDP per capita of the years 1980, 1984 and 1988. The pre-intervention period covers the years 1980 to 1988 and the post-intervention period covers the years 1990 to 1992. We limit the post-intervention period to three years in order to avoid the distortion of the estimated treatment effect. That is, extending the post-intervention period beyond 1992 involves the danger of capturing the growth effect of the 1993 reform. The donor pool contains 56 countries. That is, the donor pool contains all subnational units listed in appendix II net of the german Länder. This is because the dataset covers the western german Länderuntil 1990 and additionally covers the eastern german Länderas well as the capital Berlin as of 1991 (see section 4.3). All comparison units contribute to Synthetic Flanders. Where Italy-North West (0.129) and the West Midlands of the United Kingdom (0.274) contribute with relatively large w-weights all remaining comparison units contribute with w-weights in the range of 0.001

25France exhibits a w-weight in the range of 0.287 to 0.554 and the Netherlands exhibit a w-weight in the range of 0.211 to 0.5 under the former set of approximations. France exhibits a w-weight in the range of 0.858 to 0.961 under the latter set of approximations.

95 to 0.019. The predictor balance is given in table 4.1. This balance indicates the quality of the approximation in the pre-intervention period for every single predictor variable. The column ‘Treated’ reveals the value of the respective predictor variable for Belgium that is averaged over the entire pre-intervention period. The entire column thus provides the vector Xi explained in section 1.2.3. Likewise, the column ‘Synthetic’ provides the mean value of the same predictor variable for Synthetic Belgium and is related to the matrix

XJ . The predictor balance indicates that the approximation in the pre-intervention period works well. The pre-intervention RMSPE amounts to 100.048. Figure 4.1 provides the graphical result. In the post-intervention period the trajectory of Flanders exceeds that of Synthetic Flanders in every year. This indicates that the 1989 reform improved the economic performance of Flanders relative to the situation in which Flanders did not experience this reform (Synthetic Flanders). The average treatment effect amounts to 743 Euro per capita and 3.412 percent, respectively. In order to evaluate the randomness of the result we apply the procedure explained in section 1.2.3. The result can be seen in figure 4.2. The bold line displays the differences of outcomes between Flanders and Synthetic Flanders in every year of the pre-intervention period as well as the post-intervention period. The grey lines accordingly depict the differences of outcomes between the comparison units and their synthetic siblings. In the post-intervention period the bold line slightly deviates from the zero line. As many comparison units exhibit an accidental treatment effect larger than the effect in Flanders we value the growth effect of the 1989 reform in Flanders to be economically inconsequential.

4.4.2 The Growth Eﬀect of the 1993 Reform

To uncover the growth effect of the 1993 federalism reform in Belgium we use real GDP per capita as dependent variable. To predict real GDP we use the unemployment rate as well as gross value added of the non-market services sector. These predictors are accompanied by lagged real GDP per capita of the years 1981, 1986 and 1992. The pre-intervention period covers the years 1980 to 1992 and the post-intervention period covers the years 1994 to 2000. The donor pool contains 15 countries. The w-weights are given in table 4.2. Unsurprisingly, the neighbouring countries France and the Netherlands contribute to Synthetic Belgium with the largest weights of 57 percent and 25 percent, respectively. Ireland, Norway and Spain also contribute to Synthetic Belgium but to a smaller extent (eleven percent or less). The predictor balance is given in table 4.3. It indicates that the approximation in the pre-intervention period works very well and this is reflected by the pre-intervention RMSPE which amounts to 69.613. Figure 4.3 provides the graphical result. In the post-intervention period the trajectory of Belgium deceeds that of Synthetic Belgium in every single year. This indicates that the 1993 reform worsened the economic performance of Belgium. However, the average treatment effect in the post-intervention period amounts to 782 Euro per capita or 2.86

96 percent and is thus rather small. In order to evaluate the randomness of the result we apply the visual procedure explained in section 1.2.3. The result can be seen in figure 4.4. The bold line repesenting the outcome differences between Belgium and Synthetic Belgium slightly deviates from the zero line at the end of the post-intervention period. This indicates that the growth effect of the 1993 reform in Belgium is not exceptional and thus insignificant. Next we turn to the growth effect in Flanders. To predict real GDP we use gross value added of the services sector as well as gross value added of the financial services sector. These predictors are accompanied by lagged real GDP per capita of the years 1980, 1986 and 1992. The pre-intervention period as well as the post-intervention period cover the same years as in the previous analysis. The donor pool contains 66 subnational entities. All comparison units contribute to Synthetic Flanders. Where Ireland (0.119) and Portugal-Continental (0.103) contribute with relatively large w-weights all remaining comparison units contribute with w-weights in the range of 0.004 to 0.069. The predictor balance is given in table 4.4. It indicates that the approximation in the pre-intervention period works very well. The pre-intervention RMSPE amounts to 88.581. Figure 4.5 exhibits the graphical result. In the post-intervention period the trajectory of Flanders deceeds that of Synthetic Flanders. This indicates that the 1993 reform worsened the economic performance of Flanders. The average treatment effect amounts to 621 Euro per capita and 2.29 percent, respectively. In order to evaluate the randomness of the result we apply the procedure explained in section 1.2.3. The result can be seen in figure 4.6. Compared to the comparison units the deviation of the bold line from the zero line is small in the post-intervention period. Thus we value the growth effect of the 1993 reform in Flanders to be economically inconsequential. Eventually, we investigate the growth effect of the 1993 federalism reform in Wallonia. To predict real GDP we use the dependency ratio as well as gross value added of the financial services sector. These predictors are accompanied by lagged real GDP per capita of the years 1981, 1985, 1989 and 1992. The pre-intervention period as well as the post- intervention period cover the same years as in the previous analysis. The donor pool contains 66 subnational entities. Except for Portugal-Azores and Portugal-Madeira all comparison units contribute to Synthetic Wallonia. Where Greece-Attiki (0.259), Spain-North West (0.195), Germany- Bremen (0.118), Ireland (0.117) and Portugal-Continental (0.115) contribute to the synthetic control group with relatively large w-weights, remaining comparison units contribute with w-weights in the range 0.001 to 0.027. The predictor balance is given in table 4.5 and indicates that the approximation in the pre-intervention period works very well. However, this is somewhat in contrast to the pre-intervention RMSPE which amounts to 166.642.

97 Figure 4.7 exhibits the graphical result. In the post-intervention period the trajectory of Wallonia deceeds that of Synthetic Wallonia. This indicates that the 1993 reform also worsened the economic performance of Wallonia. The average treatment effect amounts to 861 Euro or 4.22 percent and is thus larger than the treatment effect in Flanders. This is reasonable as Wallonia called for an increase in financial means due to its deficient budgets in the years before the fourth state reform. In order to evaluate the randomness of the result we apply the procedure explained in section 1.2.3. The result can be seen in figure 4.8. Compared to the comparison units the deviation of the bold line from the zero line is small in the post-intervention period. This leads us to the conclusion that the growth effect of the 1993 reform in Wallonia is not meaningful in an economic sense.

4.4.3 The Growth Eﬀect of the 2001 Reform

To uncover the growth effect of the 2001 federalism reform in Belgium we use real GDP per capita as dependent variable. To predict real GDP we use the number of persons residing in Belgium (population) as well as gross value added of the services sector. These predictors are accompanied by lagged real GDP per capita of the years 1994, 1997 and 2000. The pre-intervention period covers the years 1994 to 2000 and the post-intervention period covers the years 2002 to 2008. The donor pool contains 27 countries. The w-weights are given in table 4.6. Except for the East European countries Czech Republic, Latvia, Poland, Romania and Slovakia all countries contribute to Synthetic Belgium. Where Denmark (0.587) and Bulgaria (0.267) contribute to the synthetic control group with relatively large w-weights all remaining countries contribute with w-weights in the range of 0.001 to 0.067. This is also true for the neighbouring countries France and the Netherlands which both contribute to Synthetic Belgium with a w-weight of 0.002. The predictor balance is given in table 4.7 and indicates that the approximation in the pre-intervention period works very well. This is reflected by the pre-intervention RMSPE which amounts to 81.967. Merely the variable ‘population’ works a little less precise with an average deviation of about 1100 persons. Figure 4.9 provides the graphical result. In the post-intervention period the trajectory of Belgium exceeds that of Synthetic Belgium in years 2002 to 2005. In 2006 and 2007, however, the reverse is true and in 2008 Belgium again outperforms Synthetic Belgium. This indicates that there is no growth effect of the 2001 reform in Belgium. This is confirmed by figure 4.10. Except for a small deviation in 2006 the bold line representing the outcome differences between Belgium and Synthetic Belgium strictly follows the zero line. Next we turn to the growth effect in Flanders. To predict real GDP per capita we use the dependency ratio as well as gross value added of the industry sector. These predictors are accompanied by lagged real GDP per capita of the years 1994, 1996 and 2000. The

98 pre-intervention as well as the post-intervention period cover the same years as in the previous analysis. The donor pool contains 95 subnational entities. Except for Berlin, the capital of Germany, all comparison units contribute to Syn- thetic Flanders. Where Hungary-North (0.216), Norway (0.125) and Luxembourg (0.121) contribute to the synthetic control group with relatively large w-weights, remaining comparison units contribute with w-weights in the range of 0.001 to 0.042. The predictor balance is given in table 4.8 and indicates that the approximation in the pre-intervention period works very precise. The pre-intervention RMSPE amounts to 133.307. Figure 4.11 exhibits the graphical result. In the post-intervention period the trajectory of Flanders deceeds that of Synthetic Flanders in every year. This indicates a small negative growth effect of the 2001 reform in Flanders. The average treatment effects amounts to 369 Euro per capita and 1.23 percent, respectively. In order to evaluate the randomness of the result we apply the procedure explained in section 1.2.3. The result can be seen in figure 4.12. In the post-intervention period the bold line representing the outcome differences between Flanders and Synthetic Flanders strictly follows the zero line. As many comparison units exhibit an accidental treatment effect larger than the effect in Flanders we value the growth effect of the 2001 reform in Flanders to be economically inconsequential. Eventually, we investigate the growth effect of the 2001 federalism reform in Wallonia. To predict real GDP we use our measure of capital formation as well as gross value added of the industry sector. These predictors are accompanied by lagged real GDP per capita of the years 1994, 1996 and 2000. The pre-intervention period as well as the post-intervention period cover the same years as in the previous analysis. The donor pool contains 95 subnational entities. All comparison units contribute to Synthetic Wallonia. Where Bulgaria-Central contributes with a w-weight of 0.159 all remaining comparison units contribute with w-weights in the range of 0.004 to 0.075. The predictor balance is given in table 4.9 and indicates that the approximation in the pre-intervention period works very well. The pre-intervention RMSPE amounts to 22.395. Figure 4.13 exhibits the graphical result. In the post-intervention period the trajectory of Wallonia deceeds that of Synthetic Wallonia. This indicates that the 2001 reform worsened the economic performance of Wallonia. The average treatment effect amounts to 291 Euro per capita and 1.33 percent, respectively. In order to evaluate the randomness of the result we apply the procedure explained in section 1.2.3. The result can be seen in figure 4.14. In the post-intervention period the bold line representing the outcome differences between Wallonia and Synthetic Wallonia strictly follows the zero line. As many comparison units exhibit an accidental treatment effect larger than the effect in Wallonia we conclude that the growth effect of the 2001 reform in Wallonia is not meaningful in an economic sense.

99 As can be seen from figures 4.9, 4.11 and 4.13 there is no effect on growth at the national level and a small negative effect both in Flanders and Wallonia. This is remarkable at first glance as the economic performance of subnational entities should coincide with the economic performance of the national level. That is, we would expect a positive growth effect in one of the regions if there is a negative effect in the other region and no effect at the national level. However, negative effects in both regions can occur as Synthetic Flanders and Synthetic Wallonia are compiled independent of each other. As mentioned above Synthetic Flanders is mainly built with data of Hungary-North, Norway and Luxembourg where Synthetic Wallonia draws on data of Bulgaria-Central. Beyond that, many other comparison units contribute to both synthetic control groups with varying w-weights.

4.5 Conclusions

This paper investigates the growth effects of federalism reforms in Belgium. In 1989 communities and regions were assigned competencies in the realms of education and regional economic policy, respectively. The 1993 reform increased the financial means available to the communities. The 2001 reform again increased the financial means available to the communities but it also granted taxing authority to the regions. For all these reforms we investigate the effect of federalism at the state level (Belgium) as well as for Flanders and Wallonia separately (regional level). By using the Synthetic Control Method we find a small positive growth effect of the 1989 reform in Flanders. Regarding the 1993 reform we find small negative growth effects at the national level, in Flanders and in Wallonia. We find, that the negative growth effect is a little bit larger in Wallonia relative to Flanders. This might be seen in the light of the financial distress that occured in Wallonia in the years before the reform. Wallonia demanded these transfers (and obtained an implicit bailout) which did not incentivise the community to improve the efficiency of the provision of its public services (Jennes, 2014). Finally, we do not find a growth effect of the 2001 reform at the national level but a small negative growth effect of the 2001 reform in Flanders as well as in Wallonia. Thus it seems that the delegation of tax authority to the regions could not be converted into positive economic growth. This may be due to the restrictions that were implemented to confine the level of tax competition. All effects we find are not exceptional and thus statistical insignificant. That is, we cannot find a clear significant growth effect of federalism reforms in Belgium. As can be deduced from the description of reform contents devolution in Belgium is very much based on political compromise. This can be illustrated using the 2001 reform. Where the francophone community demanded an increase in financial means Flanders took a stand for an expansion of spending authorities and an increase in taxing authority. Eventually, both demands were satisfied. Although it was not the objective of the reform

100 to cause economic growth one might see the zero effect at the national level as well as the small negative effects at the regional level in the light of this political compromise. Referring to the literature on the relation between fiscal decentralisation and economic growth presented in section 4.1 our results are in contrast to Akai and Sakati (2002) and Iimi (2005) who find a clear positive effect of fiscal decentralisation on economic growth. Baskaran and Feld (2013) and Rodriguez-Pose and Ezcurra (2011) find a (clear) negative effect on growth. However, they correspond to Davoodi and Zou (1998), Thornton (2007) and Xie et al. (1999) who cannot find a significant effect of fiscal decentralisation on economic growth. As the effects we find are rather small in size and clearly insignificant our results are also in line with Asatryan and Feld (2014) who conclude that there is no effect on growth at all. Eventually, our results correspond to Pagano (2013) who compares GDP growth rates of Belgium with GDP growth rates of the EU-15, the Euro Zone, the Netherlands, Luxembourg, France and Germany and cannot find a significant difference between these growth rates.

101 4.6 Tables and Figures

Variable Treated Synthetic Share of GVA, Non-Market Services 0.229 0.230 Share of GVA, Industry 0.236 0.237 Real GDP Per Capita 1980 17655.11 17738.77 Real GDP Per Capita 1984 18210.08 18300.42 Real GDP Per Capita 1988 20699.63 20800.54

Table 4.1: Real GDP Per Capita in Flanders, 1989, Predictor Balance 24000 22000 20000 Real GDP Per Capita 18000

16000 1980 1985 1990 1995

Year

Flanders (1989) Synthetic Flanders (1989)

Figure 4.1: Real GDP Per Capita in Flanders, 1989 10000 5000 0 -5000

-10000 1980 1985 1990 1995

Year

Figure 4.2: Real GDP Per Capita in Flanders, 1989, Gaps

102 State w-Weight State w-Weight France 0.569 Norway 0.024 Ireland 0.106 Spain 0.043 the Netherlands 0.258

Table 4.2: Real GDP Per Capita in Belgium, 1993, w-Weights

Variable Treated Synthetic Unemployment Rate 0.095 0.078 Share of GVA, Non-Market Services 0.247 0.25 Real GDP Per Capita 1981 18949.38 18947.03 Real GDP Per Capita 1986 20276.4 20280.48 Real GDP Per Capita 1992 23520.75 23519.94

Table 4.3: Real GDP Per Capita in Belgium, 1993, Predictor Balance 30000 25000 Real GDP Per Capita 20000

1980 1985 1990 1995 2000

Year

Belgium (1993) Synthetic Belgium (1993)

Figure 4.3: Real GDP Per Capita in Belgium, 1993 10000 5000 0

-5000 1980 1985 1990 1995 2000

Year

Figure 4.4: Real GDP Per Capita in Belgium, 1993, Gaps

103 Variable Treated Synthetic Share of GVA, Services 0.232 0.232 Share of GVA, Financial Services 0.205 0.205 Real GDP Per Capita 1980 17655.11 17658.94 Real GDP Per Capita 1986 19104.99 19107.55 Real GDP Per Capita 1992 22733.71 22730.05

Table 4.4: Real GDP Per Capita in Flanders, 1993, Predictor Balance 30000 25000 20000 Real GDP Per Capita

15000 1980 1985 1990 1995 2000

Year

Flanders (1993) Synthetic Flanders (1993)

Figure 4.5: Real GDP Per Capita in Flanders, 1993 10000 5000 0

-5000 1980 1985 1990 1995 2000

Year

Figure 4.6: Real GDP Per Capita in Flanders, 1993, Gaps

104 Variable Treated Synthetic Dependency Ratio 0.601 0.601 Share of GVA, Financial Services 0.219 0.219 Real GDP Per Capita 1981 14907.89 14906.37 Real GDP Per Capita 1985 15007.44 15006.67 Real GDP Per Capita 1989 16453.62 16458.38 Real GDP Per Capita 1992 17474.62 17481.89

Table 4.5: Real GDP Per Capita in Wallonia 1993, Predictor Balance 22000 20000 18000 Real GDP Per Capita 16000

14000 1980 1985 1990 1995 2000

Year

Wallonia (1993) Synthetic Wallonia (1993)

Figure 4.7: Real GDP Per Capita in Wallonia, 1993 10000 5000 0 -5000 1980 1985 1990 1995 2000

Year

Figure 4.8: Real GDP Per Capita in Wallonia, 1993, Gaps

105 State w-Weight State w-Weight State w-Weight Austria 0.001 Greece 0.001 Norway 0.015 Bulgaria 0.267 Hungary 0.002 Portugal 0.001 Cyprus 0.001 Ireland 0.001 Slovenia 0.002 Denmark 0.587 Italy 0.001 Spain 0.001 Estonia 0.001 Lithuania 0.001 Sweden 0.005 Finland 0.002 Luxembourg 0.038 United Kingdom 0.001 France 0.002 Malta 0.001 Germany 0.067 the Netherlands 0.002

Table 4.6: Real GDP Per Capita in Belgium 2001, w-Weights

Variable Treated Synthetic Population 10180 11309.76 Share of GVA, Services 0.228 0.231 Real GDP Per Capita 1994 23883.45 23878.77 Real GDP Per Capita 1997 25597.84 25591.34 Real GDP Per Capita 2000 27864.34 27858.69

Table 4.7: Real GDP Per Capita in Belgium 2001, Predictor Balance 32000 30000 28000 Real GDP Per Capita 26000 24000 1995 2000 2005 2010

Year

Belgium (2001) Synthetic Belgium (2001)

Figure 4.9: Real GDP Per Capita in Belgium, 2001

106 Ra D e aiai lnes20,PeitrBalance Predictor 2001, Flanders in Capita per GDP Real 4.8: Table Ra D e aiai egu,20,Gaps 2001, Belgium, in Capita Per GDP Real 4.10: Figure Ra D e aiai lnes 2001 Flanders, in Capita Per GDP Real 4.11: Figure aibeTetdSynthetic 0.556 27575.55 Treated 0.242 27573.95 24239.19 2000 24240.68 0.557 23365.01 Capita Per 0.242 GDP 1996 Real 23368.95 Capita Per GDP 1994 Real Capita Per GDP Real Industry GVA, of Share Ratio Dependency Variable

-5000 0 5000 10000 15000 Real GDP Per Capita

24000 26000 28000 30000 32000 1995 1995 Flanders (2001) 2000 2000 107 Year Year Synthetic Flanders(2001) 2005 2005 2010 2010 Ra D e aiai alna20,PeitrBalance Predictor 2001, Wallonia in Capita Per GDP Real 4.9: Table Ra D e aiai lnes 01 Gaps 2001, Flanders, in Capita Per GDP Real 4.12: Figure Ra D e aiai alna 2001 Wallonia, in Capita Per GDP Real 4.13: Figure aia .0 0.201 0.202 Synthetic 20053.3 Treated 0.189 20107.01 18107.37 2000 18150.55 17464.55 Capita Per 0.189 GDP 1996 Real 17503.76 Capita Per GDP 1994 Real Capita Per GDP Real Industry GVA, of Share Capital Variable

-5000 0 5000 10000 15000 Real GDP Per Capita

16000 18000 20000 22000 24000 1995 1995 Wallonia (2001) 2000 2000 108 Year Year Synthetic Wallonia(2001) 2005 2005 2010 2010 Ra D e aiai alna 01 Gaps 2001, Wallonia, in Capita Per GDP Real 4.14: Figure

-5000 0 5000 10000 15000 1995 2000 109 Year 2005 2010 4.7 Appendix I: Variables

Active Population Number of Employable (both Employed and Unemployed) Persons, Source: European Regional Database.

Capital Gross Fixed Capital Formation, Divided by Gross Domestic Product, Constant 2005 Euros, Source: European Regional Database and Own Calculations.

Dependency Ratio 1 - (Active Population / Population), Source: European Regional Database and Own Calculations.

Employment Rate Persons Engaged in some Productive Activity (Employee or Self- Employed), Divided by Active Population, Source: European Regional Database and Own Calculations.

Population Number of Persons Permanently Settled in the Respective Region, Source: European Regional Database.

Real GDP Per Capita Total Gross Value Added Plus Taxes Minus Subsidies On Prod- ucts, Constant 2005 Euros, Divided by Number of Persons Permanently Settled in the Respective Territory, Source: European Regional Database and Own Calcula- tions.

Share of GVA, Financial Services Gross Value Added of the Financial and Business Ser- vices Sector, Devided by Total Gross Value Added, Source: European Regional Database and Own Calculations.

Share of GVA, Industry Gross Value Added of the Industry Sector, Devided by Total Gross Value Added, Source: European Regional Database and Own Calculations.

Share of GVA, Non-Market Services Gross Value Added of the Non-Market Services Sector, Devided by Total Gross Value Added, Source: European Regional Database and Own Calculations.

Share of GVA, Services Gross Value Added of the Services Sector, Devided by Total Gross Value Added, Source: European Regional Database and Own Calculations.

Unemployment Rate 1 - Employment Rate, Source: European Regional Database and Own Calculations.

110 4.8 Appendix II: Comparison Units

Units Used to Create Synthetic Belgium Austria, Bulgaria, Cyprus, Czech Republic, Denmark, Estonia, Finland, France, Germany, Greece, Hungary, Ireland, Italy, Latvia, Lithuania, Luxembourg, Malta, the Netherlands, Norway, Poland, Portugal, Romania, Slovakia, Slovenia, Spain, Sweden, United Kingdom.

Due to data availability countries marked by underlining cannot be used in the 1989 analysis as well as the 1993 analysis.

The donor pool of the 2001 analysis contains all countries listed.

Units Used to Create Synthetic Flanders and Synthetic Wallonia Austria: East, South, West / Bulgaria: North, Central / Cyprus / Czech Republic / Denmark / Estonia / Finland: Mainland, Aland / France: Ile de France, Bassin Parisien, North, East, West, South West, Central East, Mediterranian, Outre Mer / Germany: Baden-Wuerttemberg, Bavaria, Berlin, Brandenburg, Bremen, Hamburg, Hesse, Mecklenburg West Pomerania, Lower Saxony, North-Rine Westphalia, Rhineland Palatinate, Saarland, Saxony, Saxony-Anhalt, Schleswig-Holstein, Thurin-gia / Greece: Voreia, Kentriki, Attiki, Kriti / Hungary: Central, Transdanubia, North / Ireland / Italy: North West, North East, Central, South, Islands / Latvia / Lithuania / Lux- embourg / Malta / the Netherlands: North, East, West, South / Norway / Poland: Central, Poludniowy, Wschodni, Polmocno-Zachodni, Poludniowo-Zachodni, Polnocni / Portugal: Continental, Azores, Madeira / Romania: Region One, Region Two, Region Three, Region Four / Slovenia / Slovakia / Spain: North West, North East, Madrid, Central, East, South, Canaries / Sweden: East, South, North / United Kingdom: North East, North West, Yorkshire, East Midlands, West Midlands, East of England, London, South East, South West, Wales, Scotland, Northern Ireland.

Due to data availability subnational units marked by underlining cannot be used in the 1989 analysis as well as the 1993 analysis.

The donor pool of the 2001 analysis contains all subnational units listed.

111 Bibliography

[1] Abadie, A., Diamond, A., Hainmueller, J., 2010, Synthetic Control Methods for Comparative Case Studies: Estimating the Eﬀect of California’s Tobacco Control Program, Journal of the American Statistical Association 105 (490), 493-505.

[2] Abadie, A., Diamond, A., Hainmueller, J., 2015, Comparative Politics and the Synthetic Control Method, American Journal of Political Science 59 (2), 495-510.

[3] Abbas, S. A., Belhocine, N., ElGanainy, A., Horton, M., 2010, A Historical Public Debt Database, International Monetary Fund, Working Paper No. WP/10/245.

[4] Abrams, B. A., Dougan, W. R., 1986, The Eﬀects of Constitutional Restraints on Governmental Spending, Public Choice 49, 101-16.

[5] Acemoglu, D., 2005, Constitutions, Politics, and Economics: A Review Essay on Persson and Tabellini´s The Economic Eﬀects of Constitutions, Journal of Economic Literature 43 (4), 1025-48.

[6] Adema, W., Fron, P., Ladaique, M., 2011, Is the European Welfare State Really More Expensive?: Indicators on Social Spending, 1980-2012; and a Manual to the OECD Social Expenditure Database (SOCX), OECD Social, Employment and Mi- gration Working Papers No.124.

[7] Akai, N., Sakata, M., 2002, Fiscal Decentralization Contributes to Economic Growth: Evidence from State-Level Cross-Section Data for the United States, Jour- nal of Urban Economics 52, 93-108.

[8] Alen, A., Ergec, R., 1994, Das FöderaleBelgien nach der Vierten Staatsreform von 1993, Ministerium der AuswärtigenAngelegenheiten, des Außenhandels und der Entwicklungszusammenarbeit, Brüssel.

[9] Alesina, A., Bayoumi, T., 1996, The Costs and Beneﬁts of Fiscal Rules: Evidence from the U.S. States, National Bureau of Economic Research, Working Paper No. 5614.

[10] Alesina, A., Perotti, R., 1995, The Political Economy of Budget Deﬁcits, Interna- tional Monetary Fund, Staﬀ Papers 42 (1), 1-31.

112 [11] Angrist, J. D., Krueger, A. B., 1999, Empirical Strategies in Labor Economics, in: Ashenfelter, O., Card, D., (eds.), Handbook of Labor Economics, Volume 3A, Amsterdam, 1277-1366.

[12] Armingeon, K., Kn¨opfel,L., Weisstanner, D., Engler, S., Potolidis, P., Gerber, M., 2013, Comparative Political Dataset 1. 1960-2011, Bern.

[13] Asatryan, Z., Feld, L. P., 2014, Revisiting the Link Between Growth and Federalism: A Bayesian Model Averaging Approach, Journal of Comparative Economics 43 (3), 772-781.

[14] Ashenfelter, O. Card, D., 1985, Using Longitudinal Structure of Earnings to Es- timate the Eﬀect of Training Programs, Review of Economics and Statistics 67, 648-60.

[15] Baldini, G., 2011, The Diﬀerent Trajectories of Italian Electoral Reform, West Eu- ropean Politics 34 (3), 644-63.

[16] Barker, F., Boston, J., Levine, S., McLeay, E., Roberts, N. S., 2003, An Initial Assessment of the Consequences of MMP in New Zealand, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 297-322.

[17] Baskaran, T., 2009, Did the Maastricht Treaty Matter for Macroeconomic Perfor- mance? A Diﬀerence-In-Diﬀerence Investigation, Kyklos 62 (3), 331-58.

[18] Baskaran, T., Feld, L. P., 2013, Fiscal Decentralization and Economic Growth in OECD Countries: Is There a Relationship?, Public Finance Review 41 (4), 421-45.

[19] Baskaran, T., Feld, L. P., Schnellenbach, J., 2014, Fiscal Federalism, Decentraliza- tion, and Economic Growth: A Meta-Analysis, Economic Inquiry 54 (3), 1445-1463.

[20] Baumann, E., Doennebrink, E., Kastrop, C., 2008, A Concept for a New Budget Rule for Germany, CESifo Forum 9 (2), 37-45.

[21] Besley, T., Case, A., 1995, Incumbent Behavior: Vote-Seeking, Tax-Setting, an Yardstick Competition, The American Economic Review 85 (1), 25-45.

[22] Bl¨ochliger, H., Vammalle, C., 2012, Reforming Fiscal Federalism and Local Gov- ernment: Beyond the Zero-Sum Game, OECD Fiscal Federalism Studies, OECD Publishing.

[23] Blume, L., M¨uller, J., Voigt, S., Wolf, C., 2009, The Economic Eﬀects of Consti- tutions: Replicating - and Extending - Persson and Tabellini, Public Choice 139, 197-225.

113 [24] Bodmer, F., 2006, The Swiss Debt Brake: How it Works and What Can Go Wrong, Swiss Journal of Economics and Statistics 142 (3), 307-330.

[25] Bohn, H., Inman, R. P., 1996, Balanced-Budget Rules and Public Deﬁcits: Evidence from the U.S. states, Carnegie-Rochester Conferences Series on Public Policy 45, 13-76.

[26] Bordignon, M., Monticini, A., 2012, The Importance of the Electoral Rule: Evidence from Italy, Economics Letters 117, 322-325.

[27] Bormann, N.-C., Golder, M., 2013, Democratic Electoral Systems Around the World, 1946-2011, Electoral Studies 32, 360-369.

[28] Brennan, G., Buchanan, J. M., 1980, The Power To Tax, Cambridge.

[29] Bruchez, P.-A., 2003a, A Modiﬁcation of the HP Filter Aiming at Reducing the End-Point Bias, Swiss Federal Finance Administration, Working Paper No. 2003/03, Bern.

[30] Bruchez, P.-A., 2003b, Will the Swiss Fiscal Rule Lead to Stabilization of the Public Debt?, Swiss Federal Finance Administration, Working Paper No. 2003/04, Bern.

[31] Budina, N., Kinda, T., Schaechter, A., Weber, A., 2012, Fiscal Rules at a Glance: Country Details from a New Dataset, International Monetary Fund, Working Paper No. WP/12/273.

[32] Bull, M., Pasquino, G., 2007, A Long Quest in Vain: Institutional Reforms in Italy, West European Politics 30 (4), 670-91.

[33] Bundesministerium der Finanzen (BMF), 2008, Modell einer neuen Schuldenregel, Kommission von Bundestag und Bundesrat zur Modernisierung der Bund-Laender- Finanzbeziehungen, Kommissionsdrucksache 096.

[34] Bundesverfassung der Schweizerischen Eidgenossenschaft (BV), 2002, http://www.admin.ch/opc/de/classiﬁed-compilation/19995395/index.html

[35] Burret, H. T., Feld, L. P., 2014, Vertical Eﬀects of Fiscal Rules: The Swiss Experi- ence, CESifo Working Paper No. 5043.

[36] Burret, H. T., Feld, L. P., 2016, Eﬀects of Fiscal Rules 85 Years‘ Experience in Switzerland, CESifo Working Paper No. 6063.

[37] Burret, H. T., Schnellenbach, J., 2013, Implementation of the Fiscal Compact in the Euro Area Member States, German Council of Economic Experts, Working Paper 08/2013.

114 [38] Cai, H., Treisman, D., 2004, State Corroding Federalism, Journal of Public Eco- nomics 88, 819-43.

[39] Caliendo, M., Kopeinig, S., 2008, Some Practical Guidance for the Implementation of Propensity Score Matching, Journal of Economic Surveys 22 (1), 31-72.

[40] Carey, J. M., Shugart, M. S., 1995, Incentives to Cultivate a Personal Vote: a Rank Ordering of Electoral Formulas, Electoral Studies 14 (4), 417-439.

[41] Clauss, J. U., Baumann, F., 1980, Die Regionalisierung Belgiens - Arithmetik eines “Sprachenstreits”, in: Gerdes, D. (ed.), Aufstand der Provinz. Regionalismus in Westeuropa, Frankfurt a. M., 84-106.

[42] Clingermayer, J. C., Wood, B. D., 1995, Disentangling Patterns of State Debt Financing, The American Political Science Review 89 (1), 108-120.

[43] Colombier, C., 2004, Eine Neubewertung der Schuldenbremse, Swiss Federal Fi- nance Administration, Working Paper No. 2004/02, Bern.

[44] Colombier, C., 2006, Die Schweizer Schuldenbremse - nachhaltiger und konjunktur- gerechter als der neue Stabilit¨ats-und Wachstumspakt?, Schmollers Jahrbuch 126, 521-533.

[45] Daﬄon, B., Pujol, F., 2001, Fiscal Preferences and Fiscal Performance: Swiss Can- tonal Evidence, International Public Management Review 2 (2), 54-76.

[46] Danninger, S., 2002, A New Rule: The Swiss Debt Brake, International Monetary Fund, Working Paper No. WP/02/18.

[47] Davoodi, H., Zou, H., 1998, Fiscal Decentralization and Economic Growth: A Cross- Country Study, Journal of Urban Economics 43, 244-57.

[48] Debrun, X., Epstein, N., Symansky, S., 2008, A New Fiscal Rule: Should Israel “Go Swiss?”, International Monetary Fund, Working Paper No. WP/08/87.

[49] Denemark, D., 2003, Choosing MMP in New Zealand: Explaining the 1993 Electoral Reform, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 70-95.

[50] Deschouwer, K., 2012, The Politics of Belgium: Governing a Divided Society, Bas- ingstoke.

[51] Dimitras, P. E., 1994, The Greek Parliamentary Election of October 1993, Electoral Studies 13 (3), 235-239.

115 [52] Donovan, M., 1995, The Politics of Electoral Reform in Italy, International Political Science Review 16 (1), 47-64.

[53] Feld, L. P., 2010, Sinnhaftigkeit und Eﬀektivit¨atder deutschen Schuldenbremse, Perspektiven der Wirtschaftspolitik 11 (3), 226-245.

[54] Feld, L. P., Kalb, A., Moessinger, M.-D., Osterloh, S., 2013, Sovereign Bond Market Reactions to Fiscal Rules and No-Bailout Clauses - The Swiss Experience, CESifo Working Paper No. 4195.

[55] Feld, L. P., Kirchgaessner, G., 2001, The Political Economy of Direct Legislation: Direct Democracy and Local Decision-Making, Economic Policy 16 (33), 330-367.

[56] Feld, L. P., Kirchgaessner, G., 2008, On the Eﬀectiveness of Debt Brakes: The Swiss Experience, in: Neck, R., Sturm, J.-E. (eds.), Sustainability of Public Debt, MIT Press, Cambridge/London, 223-255.

[57] Ferrara, F., Herron, E. S., 2005, Going It Alone? Strategic Entry under Mixed Electoral Rules, American Journal of Political Science 49 (1), 16-31.

[58] Finanzhaushaltsgesetz (FHG), 2006, http://www.admin.ch/opc/de/classiﬁed-compilation/20041212/index.html.

[59] Finanzhaushaltsgesetz (FHG), 2010, http://www.admin.ch/opc/de/classiﬁed-compilation/20041212/index.html.

[60] Franco, D., Sartor, N., 2006, NDCs in Italy: Unsatisfactory Present, Uncertain Future, in: Holzmann, R., Palmer, E. (eds.), Pension Reform. Issues and Prospects for Non-Financial Deﬁned Contribution (NDC) Schemes, Washington DC, 467-492.

[61] Funk, P., Gathmann, C., 2013a, How Do Electoral Systems Aﬀect Fiscal Policy? Evidence from Cantonal Parliaments, 1890-2000, Journal of the European Economic Association 11 (5), 1178-1203.

[62] Funk, P., Gathmann, C., 2013b, Voter Preferences, Direct Democracy and Govern- ment Spending, European Journal of Political Economy 32, 300-319.

[63] Gangl, M., DiPrete, T. A., 2004, Kausalanalyse durch Matchingverfahren, DIW Berlin, Discussion Paper No. 401.

[64] Geier, A., 2004, Application of the Swiss Fiscal Rule to Artiﬁcal Data, Swiss Federal Finance Administration, Working Paper No. 2004/05, Bern.

[65] Geier, A., 2011, The Debt Brake - the Swiss Fiscal Rule at the Federal Level, Swiss Federal Finance Administration, Working Paper No. 2011/15, Bern.

116 [66] Gerard, M., 2001, Fiscal Federalism in Belgium, Paper Submitted to the Inter- national Symposium on Fiscal Imbalance, The Commission on Fiscal Imbalance, Quebec City.

[67] Gerard, M., 2014, Economic Aspects of Constitutional Change: The Case of Bel- gium, Oxford Review of Economic Policy 30 (2), 257-276.

[68] Hecking, C., 2003, Das Politische System Belgiens, Opladen.

[69] Heckman, J. J., Ichimura, H., Todd, P. E., 1997, Matching As An Econometric Evaluation Estimator: Evidence from Evaluating a Job Trainng Programme, Review of Economics and Statistics 64, 605-654.

[70] Hicken, A., Kasuya, Y., 2003, A Guide to the Constitutional Structures and Elec- toral Systems of East, South and Southeast Asia, Electoral Studies 22, 121-51.

[71] Horcasitas, J. M., Weldon, J. A., 2003, Reforming Electoral Systems in Mexico, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 209-230.

[72] Hou, Y., Smith, D. L., 2009, Informal Norms as a Bridge between Formal Rules and Outcomes of Government Financial Operations: Evidence from State Balanced Budget Requirements, Journal of Public Administration Research and Theory 20, 655-678.

[73] Hou, Y., Smith, D. L., 2010, Do State Balanced Budget Requirements Matter? Testing Two Explanatory Frameworks, Public Choice 145, 57-79.

[74] Iimi, A., 2005, Decentralization and Economic Growth Revisited: An Empirical Note, Journal of Urban Economics 57, 449-61.

[75] International Monetary Fund (IMF), 2009, Fiscal Rules - Anchoring Expectations for Sustainable Public Finances, Fiscal Aﬀairs Department, Washington DC.

[76] Iversen, T., Soskice, D., 2006, Electoral Institutions and the Politics of Coalitions: Why Some Democracies Redistribute More than Others, The American Political Science Review 100 (2), 165-181.

[77] Jennes, G., 2014, Has Belgian Fiscal Decentralisation Reduzed the Size of Govern- ment and the Budget Deﬁcit?, VIVES Discussion Paper No. 42.

[78] Kastrop, C., Snelting, M., 2008, Das Modell des Bundesﬁnanzministeriums f¨ureine neue Schuldenregel, Wirtschaftsdienst 88 (6), 375-382.

117 [79] Katz, R. S., 1996, Electoral Reform and the Transformation of Party Politics in Italy, Party Politics 2 (1), 31-53.

[80] Katz, R. S., 2003, Reforming the Italian Electoral Law, 1993, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 96-122.

[81] Katz, R. S., Bardi, L., 1980, Preference Voting and Turnover in Italian Parliamen- tary Elections, American Journal of Political Science 24 (1), 97-114.

[82] Kiewiet, D. R., Szakaly, K., 1996, Constitutional Limits on Borrowing: An Analysis of State Bonded Indebtedness, Journal of Law, Economics and Organization 12, 62-97.

[83] Kontopoulos, Y., Perotti, R, 1999, Government Fragmentation and Fiscal Policy Outcomes: Evidence from OECD Countries, in: Poterba, J. M., von Hagen, J. (Eds.), Fiscal Institutions and Fiscal Performance, Chicago, 81-102.

[84] Krogstrup, S., W¨alti,S., 2008, Do Fiscal Rules Cause Budgetary Outcomes?, Public Choice 136 (1), 123-138.

[85] Lamprinakou, C., n.d., Electoral System Change in Europe since 1945: Greece; taken from the ESCE project, http://www.electoralsystemchanges.eu.

[86] Lee, D. S., Lemieux, T., 2010, Regression Discontinuity Designs in Economics, Journal of Economic Literature 48, 281-355.

[87] Lijphart, A., 1990, The Political Consequences of Electoral Laws, 1945-85, American Political Science Review 84 (2), 481-496.

[88] Lijphart, A., 1994, Electoral Systems and Party Systems: A Study of Twenty-Seven Democracies, 1945-1990, Oxford.

[89] Luechinger, S., Schaltegger, C. A., 2013, Fiscal Rules, Budget Deﬁcits and Budget Projections, International Tax and Public Finance 20, 785-807.

[90] Martinez-Vazquez, J., McNab, R. M., 2003, Fiscal Decentralization and Economic Growth, World Development 31 (9), 1597-1616.

[91] Massicotte, L., Blais, A., 1999, Mixed Electoral Systems: A Conceptual and Em- pirical Survey, Electoral Studies 18 (3), 341-366.

[92] Milesi-Ferretti, G. M., Perotti, R., Rostagno, M., 2002, Electoral Systems and Public Spending, The Quarterly Journal of Economics 117 (2), 609-657.

118 [93] Moniteur Belge, 1989, Loi SpécialeRelative au Financement des Communautéset des Régions,vol. 159, pp. 850-91.

[94] Moore, D., 2005, Slovakia’s 2004 Tax and Welfare Reforms, IMF Working Paper WP/05/133.

[95] Mourre, G., Isbasoiu, G.-M., Paternoster, D., Salto, M., 2013, The Cyclically- Adjusted Budget Balance Used in the EU Fiscal Framework: an Update, European Commission, Economic Papers 478.

[96] Nishikawa, M., Herron, E. S., 2004, Mixed Electoral Rules’ Impact on Party Sys- tems, Electoral Studies 23, 753-768.

[97] Nunez, L., n.d., Electoral System Change in Europe since 1945: Italy; taken from the ESCE project, http://www.electoralsystemchanges.eu.

[98] Oates, W. E., 1972, Fiscal Federalism, New York.

[99] Oates, W. E., 1993, Fiscal Decentralization and Economic Development, National Tax Journal 46, 237-43.

[100] Oates, W. E., 1999, An Essay on Fiscal Federalism, Journal of Economic Literature 37, 1120-49.

[101] Organisation for Economic Co-operation and Development (OECD), 2003, Ireland, OECD Economic Surveys 2003/9.

[102] Organisation for Economic Co-operation and Development (OECD), 2004, Nether- lands, OECD Economic Surveys 2004/9.

[103] Organisation for Economic Co-operation and Development (OECD), 2007, National Accounts of OECD countries, General Government Accounts, Volume IV, OECD Statistics.

[104] Pagano, G., 2013, The Current Institutional Setup of Belgium: Bonus or Handicap for the Economy?, Directorate General for Economic and Financial Aﬀairs (DG ECFIN) Workshop: “Medium-Term Challenges for the Belgian Economy”, Brussels.

[105] Perotti, R., Kontopoulos, Y., 2002, Fragmented Fiscal Policy, Journal of Public Economics 86, 191-222.

[106] Persson, T., 2002, Do Political Institutions Shape Economic Policy?, Econometrica 70 (3), 883-905.

119 [107] Persson, T., Roland, G., Tabellini, G., 2007, Electoral Rules and Government Spending in Parliamentary Democracies, Quarterly Journal of Political Science 2 (2), 155-88.

[108] Persson, T., Tabellini, G., 1999, The Size and Scope of Government: Comparative Politics with Rational Politicians, European Economic Review 43, 699-735.

[109] Persson, T., Tabellini, G., 2002, Do Constitutions Cause Large Governments? Quasi-Experimental Evidence, European Economic Review 46 (4-5).

[110] Persson, T., Tabellini, G., 2003, The Economic Eﬀects of Constitutions, Cambridge.

[111] Persson, T., Tabellini, G., 2004, Constitutional Rules and Fiscal Policy Outcomes, The American Economic Review 94 (1), 25-45.

[112] Pettersson-Lidbom, P., 2008, Do Parties Matter for Economic Outcomes? A Regression-Discontinuity Approach, Journal of the European Economic Association 6 (5), 1037-56.

[113] Reed, S. R., 2001, Duverger’s Law is Working in Italy, Comparative Political Studies 34 (3), 312-27.

[114] Reed, S. R., Thies, M. F., 2003a, The Causes of Electoral Reform in Japan, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 152-172.

[115] Reed, S. R., Thies, M. F., 2003b, The Consequences of Electoral Reform in Japan, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed-Member Electoral Systems. The Best of Both Worlds?, Oxford, 380-403.

[116] Reilly, B., 2007, Electoral Systems and Party Systems in East Asia, Journal of East Asian Studies 7, 185-202.

[117] Reuchamps, M., 2013, The Current Challenges on Belgian Federalism and the Sixth Reform of the State, in: Lopez-Basaguren, A., Escajedo San Epifano, L. (eds.), The Ways of Federalism in Western Countries and the Horizons of Territorial Autonomy in Spain, Berlin, 375-392.

[118] Rodriguez-Pose, A., Ezcurra, R., 2011, Is Fiscal Decentralization Harmful for Eco- nomic Growth? Evidence from the OECD Countries, Journal of Economic Geogra- phy 11, 619-43.

[119] Rodrik, D., 1998, Why Do More Open Economies Have Bigger Governments?, Jour- nal of Political Economy 106 (5), 997-1032.

120 [120] Rosenbaum, P. R., Rubin, D. B., 1983, The Central Role of the Propensity Score in Observational Studies for Causal Eﬀects, Biometrika 70 (1), 41-55.

[121] Rubin, D. B., 1974, Estimating Causal Eﬀects of Treatments in Randomized and Nonrandomized Studies, Journal of Educational Psychology 66 (5), 688-701.

[122] Rubin, D. B., 1977, Assignment to Treatment Group on the Basis of a Covariate, Journal of Educational Statistics 2 (1), 1-26.

[123] Sakamoto, T., 1999, Explaining Electoral Reform. Japan versus Italy and New Zealand, Party Politics 5 (4), 419-438.

[124] Schaltegger, C. A., 2002, Budgetregeln und ihre Wirkung auf die ¨oﬀentlichen Haushalte: Empirische Ergebnisse aus den US-Bundesstaaten und den Schweizer Kantonen, Schmollers Jahrbuch 122, 369-413.

[125] Schips, B., Frick, A., Graff, M., Kobel Rohr, R., Lampart, D., Müller,C., 2003, Gutachten zu ausgewählenProblemen der Schuldenbremse, Schlussbericht, Kon- junkturforschungsstelle der ETH Zürich.

[126] Shugart, M. S., Wattenberg, M. P., 2003, Mixed-Member Electoral Systems: A Deﬁnition and Typology, in: Shugart, M. S., Wattenberg, M. P. (eds.), Mixed- Member Electoral Systems. The Best of Both Worlds?, Oxford, 9-24.

[127] Smith, J. A., Todd, P. E., 2005, Does Matching Overcome LaLonde’s Critique of Nonexperimental Estimators?, Journal of Econometrics 125, 305-53.

[128] Taagepera, R., Shugart, M. S., 1989, Seats and Votes: The Eﬀects and Determinants of Electoral Systems, New Haven.

[129] Thießen, U., 2003, Fiscal Decentralisation and Economic Growth in High-Income OECD Countries, Fiscal Studies 24 (3), 237-74.

[130] Thornton, J., 2007, Fiscal Decentralization and Economic Growth Reconsidered, Journal of Urban Economics 61, 64-70.

[131] Tiebout, C. M., 1956, A Pure Theory of Local Expenditures, Journal of Political Economy 64 (5), 416-24.

[132] Todd, P. E., 2008, Matching Estimators, in: Durlauf, S. N., Blume, L. E. (eds.), The New Palgrave Dictionary of Economics, Second Edition, Volume 5, Basingstoke, 445-452.

[133] Van der Stichele, G., Verdonck, M., 2001, The Lambermont Agreement: Why and How?, Paper Submitted to the International Symposium on Fiscal Imbalance, The Commission on Fiscal Imbalance, Quebec City.

121 [134] Verdonck, M., Deschouwer, K., 2003, Patterns and Principles of Fiscal Federalism in Belgium, Regional and Federal Studies 13 (4), 91-110.

[135] Volkerink, B., de Haan, J., 2001, Fragmented Government Eﬀects on Fiscal Policy: New Evidence, Public Choice 109 (3-4), 221-242.

[136] von Hagen, J., 1991, A Note on the Empirical Eﬀectiveness of Formal Fiscal Re- straints, Journal of Public Economics 44, 191-210.

[137] Weingast, B. R., 1995, The Economic Role of Political Institutions: Market- Preserving Federalism and Economic Development, The Journal of Law, Economics, & Organization 11 (1), 1-31.

[138] Xie, D., Zou, H., Davoodi, H., 1999, Fiscal Decentralization and Economic Growth in the United States, Journal of Urban Economics 45, 228-39.

[139] Zentrale Dienststelle f¨urDeutsche Ubersetzungen¨ (ZDDU),¨ 2008, Sondergesetz Bez¨uglich der Finanzierung der Gemeinschaften und Regionen, Konsolidierung, Malmedy.

122