Every advantage has its disadvantage : does the right to vote introduce politicians acting selfish to stay in office?

Every advantage has its disadvantage:

Does the right to vote introduce politicians acting selfish to

stay in office?

Abstract: In this paper the existence of a political budget cycle in The Netherlands from 1971 to 2014 is being researched. OLS regression for time-series is used. The dependent variables are the budget deficit as percentage of GDP and the payment on social securities as percentage of GDP. Controlling for other factors, no evidence was found for an increase of the budget deficit or the social security payments during election time. Furthermore, no evidence for a decrease of both dependent variables after elections was found.

Bart van den Berg 10420274 Bachelor thesis 01/2016 Supervisor: Oana Furtuna

Statement of Originality This document is written by Student Bart van den Berg who declares to take full responsibility for the contents of this document. I declare that the text and the work presented in this document is original and that no sources other than those mentioned in the text and its references have been used in creating it. The Faculty of Economics and Business is responsible solely for the supervision of completion of the work, not for the contents.

Introduction

Democracy is widely appreciated in the western world (EBRD, 2011). The traditional democracy, as we know it from the Greek comes from the words demos and kratos,. The word demos can be translated to ‘the people’ and the word kratos means ‘power’ in the ancient Greek language. Democracy thus means: ‘power to the people’ (Ober, 2007). Since the first Greek democracy, the world has changed. The Greek democracy was what is called a direct democracy, in which ordinary citizens can pass their vote on every law. Having the right to vote on every law by every citizen is unthinkable nowadays, even if only the amount of time and money it would consume is thought of. That is why indirect democracy is prevailing now. An indirect democracy gives the citizens of a country the right to choose candidates for office for a given period. In this period the election official takes decisions on behalf of his or her voters (Matsusaka, 2005). With the introduction of indirect democracy there comes a problem. Given that an elected official wants to stay in office, why would he not do anything in his power to get to stay in office? Doing anything within his power means pleasing the voters just before elections so that he get favored above his competitors. By doing this, incumbent’s decisions are driven by self-interest and not by the public interest as it is supposed to. Taking measures by politicians in their own interest of getting re-elected is known as the political budget cycle. According to Rogoff (1987) the political budget cycle means that an incumbent does not want to conduct unpopular measures right before elections. The reason behind this is that it may decrease the probability of getting re-elected. This means that right before elections, government spending will be increased and taxes will be cut. Hence, the budget deficit will increase. Because of information asymmetries between politicians and voters, voters may reward politicians for the cut taxes and increased government spending for which the impact is visible for citizens. After the elections are being held, the government has to conduct austerity measures in order to remain financially stable. The theory of the political budget cycle thus states that: the sooner the next elections are, the higher the budget deficit for that year will be. Of course, the existence of a political budget cycle is in no one but the incumbent’s and his political party’s interest. Therefore, the existence of the political budget cycle will mean that by giving a selected candidate the mandate to represent a citizen’s interest (in other words: the introduction of an indirect democracy) a flaw of democracy is being introduced. In this bachelor thesis, the existence for a political budget cycle in The Netherlands for the timeframe of 1971-2014 is being examined. The research question is: Is there evidence for a political budget cycle in the Netherlands for the period 1971-2014?

Since some voters may not like the idea of the government running high budget deficits theory suggests that the existence of pork-barrel spending may be easier to find. Pork-barrel spending follows the same reasoning as the political budget cycle, only now politicians are afraid want to get ‘caught’. In this case politicians do not want to increase the budget deficit, because voters may not like this (voters know as well as politicians that one day the debt has to be repaid). Instead of simply spending more money, the incumbent will shift the composition of spending to areas that are highly visible to the public and yield results in the short term rather than the long term, such as infrastructure spending and social transfers. The reason why The Netherlands are being examined is that this kind of research has not been done yet in The Netherlands, so every conclusion would contribute to the knowledge on the subject. Furthermore, because of the coalition forming and the proportional elections it may be harder and less effective to use the political budget cycle (Persson & Tabellini, 2001) (for more information as to why, see the literature review). Because it is harder to achieve and less effective, the existence of a political budget cycle would be easier to generalize to political systems where implementing a political budget cycle would yield more result. The timeframe 1971-2014 is chosen, because data that before 1971 is hard to find and political culture before 1971 was very different from the political culture today, as The Netherlands was very pillared back then (Andeweg & Irwin, 2009). The research question is going to be investigated with a time-series regression with the budget deficit and spending on social securities as dependent variables in different regressions. The explanatory variable is time to elections. There are a set of variables added to control for changes in the budget deficit unrelated to the expected time to the next elections. These control variables are GDP, unemployment rate, inflation, economic growth, the i/a ratio, EMU-debt, the interest rate and the position of the governing parties on the political left-to-right scale. There will be multiple deviations from the baseline to investigate which specifications fit the data best. In the upcoming section the literature will be reviewed. Firstly, a comparing research for the US will be discussed. Secondly, research with the same aim as this thesis for a panel of data will be disputed and thirdly, the effectiveness of having a political budget cycle in The Netherlands will be discussed. Lastly, it will be argued on what ground my research is different from earlier research and how this research contributes to knowledge accumulation on the topic. After the literature review the relevance of each used variable will be discussed and motivated. Hereafter, the econometric model will be introduced; its assumptions and limitations will be discussed. Then the hypothesis and used tests will be explained. Subsequently the origin of all the data will be discussed and all the transformations made will be advocated. This is also the part where some summarizing statistics will be presented.

The results are the next parts of this paper, where multiple tests will be performed and the results will be interpret. It will also be compared to existing theory and discussed if the results match the existing theory or deviates from existing theory and some explanations as to how this might be possible. This thesis will end with the conclusions, where a summary of the results is given and the main advantages and limitations of the used research method is given. Also, a suggestion for further research is proposed.

Literature review

The research done on the topic prior to my research can roughly be divided into three categories. The first is similar to this thesis, a country specific time-series regression testing for evidence of the political budget cycle, either by measuring budget deficits or examining the composition of government spending prior to elections. Such research has been conducted by Eslava (2005) and Drazen & Eslava (2005) for Colombian municipalities. For Portuguese municipalities, Veiga & Veiga (2007b) have done the same. On the federal level the existence of the political budget cycle is tested by, among others, Laney & Willett (1983) for the United States in the period 1960-1976. The second part of the relevant research focuses on testing the political budget cycle in a panel of countries with different characteristics. Their aim is to see whether the existence of a political budget cycle depends on the nature of the political system. Such research has been done by Persson & Tabellini (2001 & 2003), Brender & Drazen (2005) and Shi & Svensson (2006). The last category important to this research is about the profit of the usage of a polital budget cycle. The main question in these articles is: does it pay off to use a political budget cycle? In other words, is an incumbent more likely to get reelected when he makes use of a political budget cycle? Some of the most important contributions in this field are made by (Veiga & Veiga, 2007a) testing in Portugal, (Drazen & Eslava, 2005) testing for Colombia and (Brender, 2003) testing for Israel. (Brender & Drazen, 2008) and (Brender & Drazen, 2007) use a panel of countries to find evidence for increased reelection prospective by imposing a political budget cycle. A last research of importance to this thesis tested whether a stronger economy boosts the probability of a party of getting reelected in the Netherlands (Ueyama, 2013).

Time Series within a Country:

Eslava (2005) found that in the Colombian municipalities, the budget deficit did not increase prior to elections, she did however find evidence of a shift in the composition of the budget. Furthermore, on the national level a same political budget cycle is visible. Both on the local and the national level government investment rose prior to elections. Government investment includes investments in infrastructure, energy, education, housing and healthcare. The share of income dedicated to infrastructure rose on average by no less than 40%. The consequence is that current expenses decrease prior to elections. Infrastructure projects are mostly financed on the municipal level, so it makes sense that these expeditures are increased at election times. Drazen and Eslava (2005) found in a panel of data of all 1100 Colombian municipalities spanning from 1987 to 2000 essentially the same as Eslava (2005), with investment spending rising and current spending - such as personnel expenditure and social transfers - decreasing. In Colombia, the same as in The Netherlands, a party is elected office instead of a politician direct, so this may indicate that despite the fact that a party instead of a person is elected, there still can be a political budget cycle. Veiga and Veiga (2005b) make use of the expenditure data of all 278 mainland Portuguese municipalities covering from 1979 to 2000. They find evidence for both an increase in the budget deficit in an election year and a shift in the composition of the spending of the municipality. Once again, the investment expenditures increase significantly, because these can be timed very well so that the investments are visible to all the voters just before the elections as a signal of how well the incumbent is doing. Laney and Willett (1983) use national U.S. data from 1960-1976 in search of evidence of the political budget cycle. They do not test the composition of the spending, but find overwhelming evidence that the budget deficit of the government is higher in election years, which supports the theory of the political budget cycle. It may be hard to generalize these results to The Netherlands because of the different timeframe, the predetermined and hard to shift date of elections and the fact that there are only two parties in the United States.

A panel of countries

Most of the existing empirical research has focused on a panel of countries because this makes it easier to determine what nationwide factors, such as the way politicians are chosen, either by proportional representation or winner-takes-it-all and factors like a country being either developed or developing contribute to the political budget deficit. Persson and Tabellini (2001) use a panel of 61 democracies from 1960 to 1998. Their research is very broad, aiming to explain differences in the composition of governments around

the world. One of their goals is to understand political budget cycles depending on the form of government. They find that in parliamentary regimes with proportional elections, such as The Netherlands, social security and welfare spending prior to elections is about 0.2% of GDP higher than it is on average. In presidential regimes the results are mixed, depending on the estimation methods. One explanation of this result is that politicians in proportional election countries seek support nationwide, hence the increase in social transfers, whereas politicians in district-elected countries seek support only in parts of the country, for example swing states. This could be a reason that pork-barrel spending is more widespread in countries where elections are district based. Persson and Tabellini (2003) use the same dataset and timeframe as they did in 2001. Only, they are now solely looking for a political budget cycle. They find that without conditioning on the political system, taxes are cut and painful fiscal adjustments, such as spending cuts and tax raises, are postponed until after the elections. Controlling for political system, they found that the tax-cuts take place irrespectively of the political system but the delay of the fiscal adjustements only takes place in presidential regimes. Only in majoritarian electoral rules they find pre-election spending cuts, whereas they find that in proportional electoral rules welfare spending is increased before elections. This is in line with their results from 2001. Brender and Drazen (2005) use a dataset of 106 countries for the period 1960-2001. Their main question is how the political budget cycle and the degree of democracy are related. They find that the political budget cycle is a widespread phenomenon. It is the strongest in new democracies. Once having controlled for the new democracies in the dataset, the evidence of the political budget cycle is not significant anymore for developed democracies such as the Netherlands. They only tested for the budget deficit and not for the composition of the spending. The results found by Shi and Svensson (2006) are for the most part in line with those of Brender and Drazen (2005). They find that on average the budget deficit of any government in an election year increases by about 1% of GDP. The results are stronger for developing than developed countries. The data is taken from between 1975 and 1995 for 85 countries. The earlier empirical results on cross-country data suggest that it may be hard to find a political budget cycle in the sense that the budget deficit increases in election years. There may be more succes in investigating the composition of the government spending, where theory suggests that the spending on social transfers will be higher in election years. This too is a way of manipulating voters.

The effects of the usage of a budget cycle

Veiga and Veiga (2007a) and Drazen and Eslava (2005) yield essentially the same results. Changing the composition of spending in favor of highly visible spending and investment expenditures increases the probability of a mayor to get reelected. This is however on the municipal level, so it’s hard to generalize to the federal level. This increased probability comes from assymetric information. Voters do not know that the increase in economic activity and the increase in investment spending is only temporary, and so voters think the incumbent is doing a great job and therefore reward him with their vote (ibid). The results of Brender (2003) are mixed and change per election. Brender and Drazen (2008) argue that in developed economies, politicians running a high budget deficit are punished and are less likely to get reelected. The same is found by Brender and Drazen (2007) with the addition that it doesn’t matter in what year a deficit is run. Right before elections or just after the elections, an incumbent always gets punished the next elections for running budget deficits. The last research focuses solely on The Netherlands. It states that a decrease in the unemployment rate and an increase in the GDP both contribute to raised support of the biggest party in the coalition in the Netherlands. This means that if an incumbant in The Netherlands can manage to somehow boost economic growth in the shortterm without altering the budget deficit, his probability of reelection increases (Ueyama, 2013).

Contribution to the topic

This research is different in the sense that it uses data at the federal level and within one country. Only the research by Laney and Willett (1983) has been done this way. Since that research takes place in the United States, which has a vastly different system of electing politicians the studies are hardly alike. Because theory suggests that the political budget cycle is the hardest to encounter in developed democracies such as the Netherlands, finding it would bring overwhelming evidence to the theory that politicians are self-interested in the sense that their sole purpose is to get reelected. The Netherlands thus is the least likelihood case for finding a political budget deficit, which means its findings would be easier to generalize to other countries.

Methodology

Variables and motivation

The chosen time frame is 1971-2014. There are two important reasons why the time-series begins at 1971. The first is practical, as it was hard to find data from before 1971. The second has to do with the fact that Dutch society was very pillared from the end of World War II to around 1970. Pillarization means that every member of a group in society only lives within this group. In The Netherlands sport clubs, hospitals and schools all were segregated and related to a certain pillar. The same applies for political parties, where the socialist pillar voted for the socialist party and Christians voted for the confessional parties (Andeweg & Irwin, 2009). The consequence is that politicians focused on getting their people to vote – the people in the same pillar as the political party – rather than getting people to shift their vote from one party to another (ibid). This of course has some implications for the theory of the political budget cycle. Since postponing unpopular measures and administering popular measures before elections has the sole purpose of winning votes from other parties, these actions only make sense if there are votes to win. In a pillared society there are hardly any votes to win from other parties, as people tend to vote within their pillar no matter what. Assuming politicians are rational, we wouldn’t expect to find an effect on the budget deficit caused by elections. This is why the time-series only begins after the depillarization kicked in in the 60’s. By 1971, most of Dutch society was depillarized, so people actually did consider changing their vote from election to election. That is why 1971 is a perfect year to start the time series, since a time series of 1971 to 2014 leaves 44 years. An N of 44 (if no leads and lags are used) is not ideal, but sufficient to do a regression. Since the political budget cycle would probably be hard to find before the 70’s, this is the best way to do a time-series regression in the Netherlands on the political budget cycle. As of the variables, in this section the choice of variables will be motivated. Starting with the dependent variables, then the explanatory variables and finishing with the control variables. Please note that there is no regression containing all dependent and independent variables, as most of the election variables are differently operationalized but are used for the same purpose. The purpose of the research is to prove whether or not politicians in the Netherlands apply policy just to gain votes – testing the self-interestedness of politicians. Theory from, among others, Rogoff (1987) suggests that politicians let the budget deficit increase before election time. Postponing unpopular measures such as tax raises and budget cuts and antedating tax cuts and budget raises causes this increase. A way to measure both these effects is to measure the budget deficit as part of GDP. The budget deficit is calculated as the difference between government spending and government income divided by the GDP. Following the

reasoning above, the budget deficit should increase during election times. For the purpose of only comparing political leaders to themself, in another regression the change in budget deficit from year t to year t+1 is used, with every new prime-minister beginning at change 0. This way it’s easier to exclude preferences changing per premier regarding budget deficits. The second dependent variable is social security spending as a percentage of GDP. The hypothesis of altering the budget deficit has been tested and rejected in most developed countries. Persson & Tabellini (2001) did find however, that countries in parliamentary regimes with proportional elections tend to increase spending on social security and welfare prior to elections. The Netherlands is such a country. The reasoning behind this according to Persson & Tabellini is that a lot of people benefit from increasing the social security, while none are harmed because the taxes aren’t raised until at least after the elections. That is why a second regression is run with social security spending as a percentage of GDP as the dependent variable. The existence of a Dutch political budget cycle may be clearer in social spending. The dependent variables all refer somehow to the presence of an election year. There is an election year dummy, which equals one of there is an election held in the last six months of that year or an election is held in the first six months of the next year. The underlying assumption for doing that is that it is hard to change the economy over night, so that a lag of six months is assumed. A dummy without assuming the six months lag is also used in different regressions. Also, expected years to election is used, which equals 3 if the expected years to the election are three. This variable is introduced since a government can be unexpectedly disbanded. The result is that there will be elections without the administration having had the possibility to make usage of a political budget cycle. Because no linearity can be assumed, also dummies for each possible expected number of years to elections are used. Finally, the control variables will be introduced. The control variables all will be used in both the regression with the budget deficit and the spending on social security. The first control variable used is the unemployment rate. As the unemployment rate rises, the government will have less income due to a fall in received income taxes. On the other hand, the government will have to increase its spending on unemployment benefits since more people are unemployed. If unemployment rises, then the budget deficit rises, it is therefore important to control for the unemployment rate. Furthermore, every article researching the existence of the political budget cycle mentioned before in this thesis, made use of the unemployment rate as a control variable. The second control variable is economic growth per capita as percentage of the GDP. As the GDP per capita grows, inhabitants of a country have more money. As a result, the gains of income tax increase. Since the people of a country have more money, they also tend to spend more money. This results in an increase in value-added tax income. Furthermore, economic

growth can result in higher profits for companies, which in turn would increase the corporate tax income received. In short, economic growth ensures more tax income and in therefore a lower budget deficit. Of course, the effect also works the other way around. There comes however a problem of simultaneous causality. The budget deficit – the same applies to social security spending – also influences the growth of GDP per capita. If the government decides to spend more money and collect less, then spending nationwide increases and in turn GDP grows. To avoid this problem, in this thesis it is assumed that economic growth today, will only start to alter the budget deficit next year. This is a reasonable assumption because money increase in the pocket of consumers does not flow to the government right away. This will take time, a year for instance. On the other hand, an increase in the budget deficit will not stimulate the economy right away, this also will take some time. This way, the problem of simultaneous causality is being avoided. Also the problem of GDP being used to calculate both the budget deficit and the economic growth is being circumvented because the years of GDP used in both variables being compared are different. The third control variable is inflation, which is the percentage change of the weighted average price from year t-1 to year t. This variable also is widely used when researching political budget cycles. It is important because inflation tends to make people spend today. After all, the product is cheaper today than tomorrow. More spending yields more tax income. Hence, inflation decreases the budget deficit. In times of deflation people wait with consuming, – the product will be cheaper tomorrow – this will increase the budget deficit. The fourth control variable is the first and only non-economical. This variable represents the position of a government on the left to right scale of politics. This is an important control variable because left and right wing parties have different political ideas, including different ideas about how low or high a budget deficit in a given year should be (Tavares, 2004). Tavares (2004) also found that left parties and right parties have different preferences in both government spending and tax policy. This makes the position of the ruling parties on the left-to-right scale an important variable to control for. The fifth control variable is an interaction variable, interacting the EMU debt and the annual interest rate the Dutch government has to pay. The EMU debt is the total government debt of The Netherlands – determined by the economic and monetary union – divided by the GDP. The annual interest rate is the interest rate of the EMU debt. These variables are used interacting, because when multiplied it precisely represents the percentage of GDP paid on interest. It therefore is important in explaining the budget deficit as a percentage of GDP. When the interest rate rises or the EMU debt rises, then the budget deficit as a percentage of GDP also rises.

The econometric model

In this research, a multiple regression model is used, with the budget deficit being the dependent variable in the main regression and the election dummy being the independent variable. For every regression being conducted, StataSE v.13 is used. Because there are a lot of other factors influencing the budget deficit, a number of variables are used to control for fluctuations unaccounted for by the election dummy. Not only do I make use of multiple variables, it’s also important that the variables are time-dependent. Not only elections this year can influence the budget deficit this year, also the knowledge that elections are next year or the knowledge that elections were last year can influence the budget deficit at year t. This is called the dynamic causal effect of the election dummy on the budget deficit. The period of time between observation t and t-1 is one year. Because of autocorrelation among variables such as the budget deficit and spending on social securities as percentage of the GDP at time t, lag terms are introduced. The ultimate model is an autoregressive distributed lag model. For determining how many lag terms there have to be introduced for variables such as the GDP growth, budget deficit and the unemployment rate, the Akaike Information Criterion (AIC) is used. As far as the stationarity of the budget deficits is concerned, there is need to test for stationarity. Stationarity means the effect of upcoming elections on the budget deficit does not depend on the time (Stock & Watson, 2015b). In other words, there are no trends over time. The associated test for stationarity is the Dickey-Fuller test. If a dependent variable is proven to be nonstationary, the variable needs to be transformed so that there are no trends. Transforming can take place by using differences between years or by using the natural logarithm. The model used is for the budget deficit as dependent variable as follows: % !" !"# ! = !0 + !1 ∗ !"#$%&'( !"##$ ! + !2 ∗ !"#$%&'( !"##$ ! + 1 + !3 ∗ !"#$%&'! !"##$ ! − 1 + !4 ∗ % !" !"# ! − 1 + !5 ∗ % !" !"# ! − 2 + !6 ∗ %∆!"# !"# !"#$%" ! − 1 + !7 ∗ %∆!"# !"# !"#$%" ! − 1 + !8 ∗ !"#$%&'($#") !"#$ ! + !9 ∗ !"#$ ! + !10 ∗ !"#$%&' ! ∗ !"#$%$&#%'#! ! + !(!)

The assumptions of the econometric model

The assumptions underlying the autoregressive distributed lag model (ADLM) are similar to those in OLS, besides some exceptions. The four most important assumptions as discussed in Stock & Watson (2015b) concerning this research will be explained briefly. 1. X is exogenous, which means the expected conditional mean of the error term is 0 for every given X in the present and past. When the expected conditional mean of the error term is also 0 for future values of X, then X is strictly exogenous. This is the case with the election dummy. While having elections a year from now (t+1) may obviously alter the budget deficit at time t, introducing a lead for the election dummy makes sure that this effect is accounted for in the model and not in the error term. 2. The second least square assumptions does not hold for time-series regression. x1i, x2i, . . . ,xki are not independently and identically drawn since t-1 influences t. A different assumption consisting of two parts comes in place: a. Data are drawn from a stationary distribution. This means that the data today has the same distribution as the data in the past. This holds for the election dummy, since all the other factors differing in time are accounted for. b. We assume that data are independently distributed when the time between them is large. This also holds, the budget deficit of time t might be influenced by the budget deficit at time t-1, but is in no way influenced by t-20. This assumption is called weak dependence 3. Large outliers are unlikely. This holds, since the budget deficit is always within the 10% of GDP range. There has never been a budget deficit of say, 40%. 4. No perfect multicollinearity. This also holds, because no variable can explain another variable for 100%. The dummy variable trap is avoided by not adding all the expected time to elections dummies. It is unclear from theory whether or not the standard errors are homoscedastic. Homoscedasticity means that the standard deviations of the error term are constant among different values of x. Of course this is important to the model, because when the standard errors are heteroscedastic, robust standard errors have to be used. The way to find out about heteroscedasticity is the Breusch-Pagan test. If the error term is heteroscedastic, it might even be the case that the error term is auto-correlated with itself. This means that the error term at time t is in part determined by the error term at time t-1. This is the case when there are omitted auto-correlated variables left in the error term. This means that the error term itself is auto-correlated. Because there is not certainty that there are no auto-correlated omitted variables in the error-term, we have to test

for autocorrelation in the error-term. The proper test to use is the Breusch-Godfrey test. If the error term is proven to be auto-correlated, then the Newey-West model has to be used.

Hypotheses and statistical tests

The hypotheses being tested are whether or not the remaining time to elections influences the budget deficit or social security payments. Because time-series regression is used and lags and leads for the independent variable researched are used in the model, a simple t-test on the beta of the election dummy at time t is a good place to start, but it also is necessary to do an F-test on joint hypotheses. To test if the budget deficit or social security payments are influenced by the time to elections, the proper thing to do is to test if all coefficients of the election variables, lags and leads included, are equal to 0. If this is indeed the case, then time to elections doesn’t influence the dependent variable. If this is not the case, then the H1 hypothesis is true, which means that the time to elections does influence the budget deficit. Hence, the existence of a political budget cycle in the Netherlands would be proven. The right test to conduct is the Granger Causality test. This test answers not only whether or not the coefficient of the explanatory variable at time t is significantly different from zero, but also its lags and leads. This is the test used to find evidence for a political budget cycle in The Netherlands.

The dataset:

The dataset being used consists of data compiled from various data sources. I will give a description of the source and the transformations I made to each variable. I again begin with the dependent variables, then the explaining independent variables and I will end with the control variables. The first dependent variable is the budget deficit of The Netherlands from year to year. The data is obtained from Eurostat. For calculating the budget deficit, the EMU rules are being used. As Eurostat states it: ‘’ Public deficit/surplus is defined in the Maastricht Treaty as general government net borrowing/lending according to the European System of Accounts. The general government sector comprises central government, state government, local government, and social security funds.’’ (Eurostat, 2015)

Variable No. of observations Mean Standard deviation Minimum Maximum Budget deficit as % of GDP 44 2.94 2.41 -1.9 8.6 Table 1 The second dependent variable is the of spending on social security by the Dutch government as a percentage of the GDP of The Netherlands. This variable is compiled by data found at OECD. The data is compiled by: (Social security benefits paid by general government, value)/ (Gross domestic product, value, market prices)*100.

Graph 1: the relation between the social security spending and expected years to elections Variable No. of observations Mean Standard deviation Minimum Maximum Social security payment as % GDP 44 14.07 3.10 9.58 19.46 Table 2 10 12 14 16 18 20 Ex p ed itu re o n so cia l se cu rit ie s a s % o f G D P 0 1 2 3 4

The explaining dependent variable is ultimately the time to elections, but it’s been operationalized in a number of ways. Both the dates of elections and the dates elections are called are obtained from parlement.com and double-checked making use of the Freedom house dataset. All the different forms of election date variables will now be explained. Please see appendix a) for the exact election years • Election dummy: a variable that equals 1 if an election is held that year and equals 0 if no election is held that year. • Election dummy, correction for broken years: a variable that equals 1 if in election is held in the last six months of that year or in the first six months of the next year and otherwise equals 0. • Expected years to election: a variable that represents the timespan to the next elections. It counts down from 4 years to 0, but once a cabinet is dissolved it immediately equals 0, because elections are then sooner than expected before. • Expected years to election dummies: works the same as expected years to election, only this time multiple dummies are used because linearity in the latter variable cannot be assumed. The data representing the unemployment rate in the Netherlands is obtained from the CBS (Statistics Netherlands). It is the percentage of people willing to work at least 12 hours per week who are available and actively seeking work divided by the total labor force (CBS, 2015). Economic growth per capita is measured as the percentage change in GDP per capita from year t-1 to t. the data is compiled from data found on CBS via parlement.com. Economic growth is the percentage difference in the value of domestic produced goods and services, correcting for inflation according to CBS (2015). Also the growth in population was found on CBS. The growth in population is the percentage difference in the number of inhabitants in The Netherlands from t-1 to t. The economic growth per capita is in turn found by subtracting population growth from economic growth. The inflation rate is obtained from CBS and is defined as the weighted average increase in prices in The Netherlands per year. Variable RILE is the position of the Dutch government on the left-to-right scale, where a high number indicates that the administration’s ideals lie to the political right. The Freedomhouse database is used, where the elections programs are examined every election again to determine what the position of a political party on the left-to-right scale is. The position of the ruling coalition is then calculated by the weighted average of the seats acquired by the

parties, see appendix a) for the composition of the government throughout the years and the rile scores associated. The EMU debt is acquired from CBS, the annual interest rate for the Dutch government is obtained making use of the Datastream database. The interest rate for bonds with a maturity of 10 years is used. Since all the Dutch bonds interest rates tend to fluctuate in the same way this rate can be used for the debt as a whole.

Methodology of testing and Results

Introduction and sequence of testing

This part consists of essentially three parts. The first part comprises of testing for stationarity with the general Dickey-Fuller test and transforming variables so that all the variables of interest have stationarity. The second part consists of other test being used to determine the kind what kind of regression is suitable. These test are the Breusch-Godfrey test for autocorrelation among the error-term and the Breusch-Pagan test for heteroscedasticity of the error term. After these tests the correct regressions are conducted, beginning with the budget deficit as percentage of GDP as dependent variable. The Akaike Information Criterion (AIC) is used to see what regression best fits the data. Lastly, a Granger causality F-test is used to test for joint hypotheses.The results of the regressions will be discussed, with interpretations of all coefficients proven to be significant. Because theory suggests that the best chances to find a political budget cycle in the Netherlands are to take a look at the spending on social security as a percentage of GDP, the second dependent variable is the payment to social security as a percentage of GDP. With this dependent variable, again AIC and the Granger causality test will be used. Also, the interpretation of the results will be discussed again.

-2 0 2 4 6 8 Bu dg et D ef ici t a s % o f G D P 1970 1980 1990 2000 2010 Year

Stationarity of variables

The general Dickey-Fuller test is used on all variables. This test is used to determine whether or not the variables contain a unit root, meaning the variables are non-stationary. As for the first dependent variable, the budget deficit as a percentage of GDP, graph 1 represents the budget deficit as a function of time. Graph 2 shows that there is hardly any trend visible in the long run, this is mostly due to all the little peaks. The null hypothesis of the test is that the variable has a unit root process; the alternative hypothesis is that the variable was generated by a stationary process. the p-value displayed in appendix b) shows the variable is stationary with about 94% certainty. There is therefore no need to transform the variable any further. Graph 2: the relationship between the budget deficit on the y-axis and years on the x-axis

Also, stationarity of the payment on social securities as a percentage of GDP is tested. As mentioned before, the null hypothesis is nonstationarity. The Dickey Fuller test is afresh used. Before testing, we can obtain a clue concerning stationarity of the dependent variable with the use of graph 3. This time, there does seem to be a trend, because there are far less ‘random’ deviations. 10 12 14 16 18 20 So cia l Se cu rit y e xp en se s as % o f G D P 1970 1980 1990 2000 2010 Year Graph 3: the relationship between the social security expenses on the y-axis and time on the x-axis -1 -. 5 0 .5 1 1. 5 d ss gd p 1970 1980 1990 2000 2010 Year Graph 4: the relation between the difference of the social security expenses on the y-axis and time on the x-axis

Testing for unit roots in appendix b) doesn’t give enough evidence to reject the null hypothesis, we can therefore not be sure the variable contains no unit root. This leads to some problems, as specified in the methodology. We therefore make some transformations to make sure the variable is stationary. The difference from one year to the next makes sure that the variable is stationary, as showed in appendix a). The dependent variable now is the difference in social security spending. Looking at graph 4 above, the difference of the dependent variable indeed seems to contain less of a trend. As for the independent variables, some contain a unit root. This could pose a problem. On the other hand, making all the variables stationary could lead to tossing away useful information (Vliegenthart, 2014). Since the dependent variables are stationary, there is no need to fear for spurious regression. Testing for every independent variable, it is found that the variables unemployment rate, inflation, debt to GDP ratio, the interest the government has to pay on its debt and the interest payment of the government as percentage of the GDP do have a unit root, graph 5 illustrates this. The exact p-values for the Dickey-Fuller test associated can be found in appendix b). 0 5 10 15 1970 1980 1990 2000 2010 Year

Inflation Unemployment rate

interestgov100 ipgdp

Graph 5: the development of the inflation, unemployment rate, the percentage of GDP paid on interest and the average interest rate of the government from 1971-2014

To avoid misspecification of the model, all nonstationary independent variables are transformed into stationarity. Using the logarithm for the unemployment rate and inflation make the variables stationary. For the other variables suffering from nonstationarity, the difference from one year to the next is used. All the lagged variables are treated the same as the original variables.

Testing to yield the right model

The error term could be auto-correlated. To test for autocorrelation, the Breusch-Godfrey test is used. If it is auto-correlated the Newey-West regression model should be used. If not, linear regression for time-series can be used. The regression consisting of all variables said to be important by theory are included, so that there is no auto-correlated omitted variable within the error term. This is regression 3 mentioned in appendix c) for the budget deficit as dependent variable. The resulting p-value arising from the Breusch-Godfrey test is 0.4122. As the null hypothesis is that there is no autocorrelation among the error-term, this hypothesis is not rejected. The same applies to the second and the third lag term of the error, those are also not significantly different from 0. We can now assume that, as long as we incorporate enough variables in the model, the error term in the model with the budget deficit as dependent variable is not auto-correlated. The same test is conducted with the social security payment as dependent variable. Because the difference in the social security spending in year t is quite heavily correlated – a correlation of 0.5473 – with the difference in social security spending at t-1, it is important to incorporate this lag term in every regression. Once done, in none of the regressions there is any autocorrelation of the error term left. Regression 3 in appendix d) for example yields a p-value of 0.4890. Now that it is clear all variables are stationary and the error term is not auto-correlated, it is time to conduct another test. It is important to know whether or not the standard errors are heteroscedastic, because using homoscedastic only standard errors when the data is in fact heteroscedastic would result in wrong standard errors and as a consequence invalid t-values. The way to find out is by using the Breusch-Pagan test. The outcome of the test on heteroscedasticity varies per input of variables. In regression 3 and 4 of both appendix c) and d) the standard errors are both homoscedastic. Other regressions conducted with a mix of variables that explained the variance in the budget deficit poorly – and therefore not included in the appendix – did have heteroskedastic standard errors. Since homoscedastic standard errors cannot be used when the standard errors are heterescedastic, but robust standard errors can be

2015a) in this section robust standard errors are used if not indicated otherwise. Please note that the differences in outcomes between using robust or homoscedastic-only standard errors are very small. Now that all the needed tests are done, a linear regression for time-series can be used. The key to finding the right amount of lag terms is minimizing the Akaike Information Criteria. Adjusting the amount of lag terms used does this. To yield the right amount of lag terms per variable, the AIC score per regression is compared. This gives a broad idea of the number of lags terms to be used. Another way of deviating from the baseline is checking how significant variables are and variables proving to be not significant time after time omitting from the regression. Because of the small N, the problem of overfitting could arise if too many control variables are added. That is why the focus is mostly on regression 4,5 and 8 of appendix c) and on regression 3 and 4 in appendix d) when discussing the impact of the explanatory variables. Furthermore, in some regressions instead of the election dummy, the corrected election dummy is used or the ‘expected years to election’ dummy is used. See appendix c) and d) for a selection of the most important regressions used. Since the explanatory variable can be operationalized in a number of ways, in some regressions the election dummy is used, in others the election dummy counting for broken years or the expected time to elections dummy is used. The best way to measure the effect of election from intuition is the election dummy, because all the other variables use January to December as a year. It therefore makes more sense to also use the election dummy from January to December

Results with the budget deficit as a percentage of GDP as dependent

variable

The main interest is the p-value of the betas on the election dummy and the p-value of the joint hypothesis Granger Causality test, but in this part also the coefficient on the control variables will be discussed. Since regression 3, 5 and 8 of appendix c) fit the data best, mainly the findings of these regressions will be discussed. Regression 3 will be discussed mainly because it incorporates all variables that could be relevant given theory. In all of the regressions the election dummy seems not to be relevant. Furthermore, in none of the conducted regression the Granger Causality test is significant. This implicates that in none of the regressions can be proven that at least one of the used dummy variables (lag, time t and lead) influences the budget deficit.

The fact that the election dummy cannot be proven to be different from zero is for the most part in line with theory. Theory already suggested that a political budget cycle would be hard to find in The Netherlands because The Netherlands is both a developed country and makes use of a proportional election system. These characteristics both contributed to not finding the existence of a political budget cycle using panel data in prior research mentioned in the theoretical review. Another factor contributing to making it harder to find a political budget cycle is the relatively short time series. This leads to estimations being less precise and thus lower t-values. Both the corrected election dummy and the expected time to election dummies proved to be not significant. This could be because the corrected election time uses broken years, whereas no other variables use broken years. For the expected time to elections dummy I would say it is quite a theoretical concept, maybe politicians do not worry about expected years to elections and they only start to worry about getting reelected once the elections are issued. Also, it makes sense that when the election dummy is not significant, the same variables operationalized in a different way are not significant. Mostly regression 3 will be discussed, as all of the variables are included there. What clearly matters is the unemployment rate. A higher unemployment rate leads to a higher budget deficit. This is in line with theory, because theory states that with more unemployment, the government has to pay more unemployment benefits and receives less income tax. Hence, the budget deficit increases. The significance for this variable is quite strong in regression 7 and 8 and less strong in others. Also the percentage of GDP spent on interest seems an important variable, but only if one year is used. An explanation for this is that the correlation from one year to the next of the interest spent as a percentage of GDP is 0.82. If both the term at time t and the lag are incorporated the terms explain the same for the most part and are therefore separately not significant. If only one of the two is used however, an increase in the change of the percentage of GDP paid to interest tends to raise the budget deficit given the positive beta. This also makes sense, because if the interest to be paid raises, the amount of interest paid as percentage of GDP increases and thus the budget deficit raises if other government expenses and other income remains constant. The last variable that seems to be important is the lag term of the dependent variable; the budget deficit as percentage of GDP at t-1. This positive beta means that a high budget deficit in year t-1 tends to raise the budget deficit in year t. An explanation here for is somewhat harder, but a possible explanation might be that it reflects the preference of politicians to have high or low budget deficits year after year. Given that the position of the government doesn’t matter, this must mean that the preference for a high or low budget deficit is not correlated to

the position of the government from left to right. This may be hard to argue on theoretical grounds however. There are also variables that appeared to matter from theory, but actually don’t. The first thing noticed is that the change in inflation does not explain the budget deficit. A small beta for inflation was expected, given theory, but none is found. Furthermore, the economic growth from a year earlier proves not to be significant. A possible explanation is that economic growth doesn’t tend to differ very much from year to year, so that an N of 42 might not be enough to prove the relation. The same applies as possible explanation as to why the beta of inflation is zero. Likewise, the position of the government does not matter at all for the level of budget deficit. This is highly remarkable, as the political right has a reputation of being far more conservative with the government finances. This regression proved this reputation to be untrue in The Netherlands for the years 1971-2014. An explanation could be that coalition forming leads to a government being in the political middle for most of the time, this makes it hard to distinguish different behavior from the political left and right.

Results with social security payments as dependent variable

The results are afresh discussed, starting with the dependent variable. As was the case with the budget deficit as dependent variable, again we find no evidence for a chance in social security spending with elections being on the horizon. In none of the regressions the Granger Causality test has a p-value below 0.05. So again, there is no proof that at least one of the dependent variables is unequal to zero in the regressions being conducted. Looking at the dependent variables separately we see that only the lagged election dummy is proven to be positive in regressions 4, 5 and 6 of appendix b. This is however very weak evidence given the fact the Granger causality test wasn’t significant. Still, a little explanation will be provided. The most obvious explanation would be that politicians want to keep the promises they made during election time. This in turn could lead to a rise in the change of social security spending the year after elections. Another, less plausible explanation might be that politicians systematically fail their timing when it comes to gaining votes by conducting popular measures. As for the other variables, it seems that especially the lagged change in social securities and unemployment rate influence the change in the spending on social security, as does a change in the percentage spent on interest and a change in the EMU debt to a lesser degree.

The lagged rate in social securities has a positive coefficient, meaning that a rise in the change of the social securities in t-1, leads to a rise in the change of social securities at time t. An explanation for this is that it takes some time to conduct the measures and social security police preferred by an administration, so that a rise in the change in t-1 means that at time t the rise will carry on. The natural logarithm of the unemployment rate and the natural logarithm of the lagged unemployment rate are never used in one regression, because the correlation between the two is 0.9505. This implies almost perfect multicollinearity, which means the two variables explain almost exactly the same. Therefore, in no regression both variables are used at the same time. In most regressions the coefficient of unemployment is significant. This is expected, as a change in unemployment tends to change spending on unemployment benefits and thus a change in the payment on social securities. The direction of the coefficient however is very surprising. A positive relation is expected because a rise in unemployment rate increases spending on unemployment benefits and decreases GDP, both contributing to a higher percentage of GDP being spent on social securities. Despite this expectation, the coefficient is proven to be negative, which means a percentage increase in the unemployment rate decreases the change in payment on social securities. From a theoretical point of view, this makes no sense. An explanation could be that the regressions with a significant coefficient of the unemployment rate suffer from omitted variable bias. If there’s a third variable influencing both unemployment and the spending on social security in the opposite way, then such a beta is possible. An example could be that a higher unemployment rate would lead to regulations to get social benefits become stricter in order to be able to pay every person entitled to social benefits. Also, social benefits could be cut if unemployment increases to make sure the government doesn’t default. If such policy shoots over, an increase in the unemployment rate would yield a decrease in the spending on social securities. Such a mechanism is however highly unlikely, and no proof is found for such a mechanism. The EMU debt being positively correlated makes more sense, because both the EMU debt and the social securities use the GDP to be calculated. If the GDP grows, then holding constant the debt, the EMU debt as a percentage of GDP decreases. The same is true for the social security payments as a percentage of GDP. This correlation could well be caused by only this mechanism, so that this correlation doesn’t necessarily imply causality. The same applies to interest payments as percentage of GDP, because here also the GDP is used to calculate its value. All the other variables do not have a significant impact on the change of social security spending. The economic growth per capita doesn’t change the percentage spent on social securities. Most likely because when welfare social security benefits tend to increase to, so that

the social security as percentage of GDP doesn’t change if both the social security payments and the GDP increase with the same percentage. Inflation is in no regression significant. This does not come as a big shock as only a small effect was expected and an N of 41 most likely is not enough to prove this relationship. Furthermore, the position of the government appears not to matter. This actually is very surprising, as the political left want to spend more on social securities than the political right, because of their preference to level income across the country (Huber, Ragin, & Stephens, 1993). The same reasoning as with the budget deficit applies. Because governments tend to be in the political middle due to coalition forming, it may be hard to unveil the real preferences of parties when it comes to social securities. Furthermore, it may take more than four years of ruling to truly alter the social security spending. Because elections are held every four years, this is hard to achieve for politicians.

Conclusions

In this thesis the existence of a political budget cycle in The Netherlands was examined by testing for correlation between both the budget deficit and the spending on social securities and an election dummy, using lags and leads. No convincing evidence for the existence of a political budget cycle was found. This is for part in line with theory, which stated the existence of a political budget cycle in The Netherlands would be hard to prove. It therefore is in line with theory that it can’t be found that the budget deficit increases during election time. Theory however also suggested, that spending on social securities would be raised. For this as well no evidence is found. This is not exactly in line with theory, but failure to prove the existence doesn’t mean the political budget cycle doesn’t exist as the betas were as expected but not significant. A longer time-series could have changed this, as this tends to decrease the standard error. The most important advantages of the methodology being used are the transparency and easiness to interpret. By testing for stationarity, heteroscedasticity and autocorrelation in the error-term and also reporting the outcomes, it is easier to replicate the research. Also, by using linear regression, the results are easy to interpret. Main disadvantages of the research method used are the small N, making it harder to find evidence for a political budget cycle. There is however good theoretical ground for choosing a short time-series because of pillarization in The Netherlands before the 1970’s. Using data at the municipal level or using quarterly data could have avoided this problem. Quarterly data could pose a problem, as the budget deficit is calculated per year. Also, mayors in The Netherlands are not elected but appointed, so using a political budget cycle wouldn’t make any sense (Andeweg

& Irwin, 2009). For this reasoning, when using a time-series regression in The Netherlands, using annual data from 1971-2014 is the best way to conduct the research. Another disadvantage is that the used regression method is relatively simple. A more sophisticated regression model, such as an ARIMA regression, could have yielded more precise results. Contribution to the topic is that the country-specific regression used, has never been done before in the Netherlands. Therefore, every result would be a contribution to the topic. Theory from panel studies already suggested the existence of a political budget would be hard to prove in developed countries with proportional elections, such as the Netherlands. Testing the political budget cycle in only the Netherlands indeed resulted in no convincing evidence for a political budget cycle. A conclusion therefore is that The Netherlands is quite typical for a developed country with proportional elections when it comes to applying a political budget cycle. Suggestion for further research could be divided into three proposals. The first is repeating this research in about twenty to 25 years. This would significantly increase the N, making the estimations more precise. Secondly, another way of conducting the same research could be used. For example, splitting the budget deficit in tax income and expenses and testing for change in this separately. This could give more convincing results. Also the composition of the government spending could be investigated. Lastly, it could be researched whether the usage of a political budget cycle helps the incumbent gain votes. This has however never been researched properly. The research mentioned in this paper by Ueyama (2013) only tested whether or not more economic growth yielded more votes, it didn’t mention the political budget cycle separately. The results of such a research could give further explanations for the results of this research. If the practice of the political budget cycle would prove to not yield more votes, it would be even more less of a surprise that the existence of a political budget cycle in The Netherlands cannot be found.

Bibliographies Andeweg, R., & Irwin, G. (2009). Governance and politics of the Netherlands. New York: NY: Palgrave Macmillan. Brender, A. (2003). The effect of fiscal performance on local government election results in Israel: 1989–1998. Journal of Public Economics , 87 (9), 2187-2205. Brender, A., & Drazen, A. (2008). How Do Budget De cits and Economic Growth Affect Reelection Prospects? Evidence from a Large Panel of Countries . American Economic Review , 98 (5), 2203-2220. Brender, A., & Drazen, A. (2005). Political budget cycles in new versus established democracies. Journal of Monetary Economics , 52 (7), 1271-1295. Brender, A., & Drazen, A. (2007). Political implications of fiscal performance in OECD countries. Fiscal Indicators, Banca D'Italia, , 957-982. CBS. (2015, 12 17). CBS Statline - table explanation. Retrieved 01 07, 2016, from statline.cbs.nl: http://statline.cbs.nl/StatWeb/publication/?VW=T&DM=SLEN&PA=80479ENG&D1=4-5,10-13&D2=0&D3=0&D4=(l-25)-l&HD=100712-1629&LA=EN&HDR=T&STB=G1,G2,G3 Drazen, A., & Eslava, M. (2005). Electoral manipulation via expenditure composition: theory and evidence. (No. w11085) . EBRD. (2011). The intangible transition: support for markets and democracy after the crisis. In EBRD, Crisis and Transition: The People's Perspective (pp. 61-75). Brussels: European Bank for Reconstruction and Development. Eslava, M. (2005). Political Budget Cycles Or Voters As Fiscal Conservatives? Evidence From Colombia. DOCUMENTOS CEDE 003343 . Eurostat. (2015). General government deficit/surplus - % of GDP and million EUR. Retrieved 01 07, 2016, from http://ec.europa.eu/eurostat: http://ec.europa.eu/eurostat/tgm/web/table/description.jsp Huber, E., Ragin, C., & Stephens, J. D. (1993). Social Democracy, Christian Democracy, Constitutional Structure, and the Welfare State. American Journal of Sociology , 99 (3), 711-749. Laney, L., & Willett, T. (1983). Presidential Politics, Budget Deficits, and Monetary Policy in the United States; 1960-1976 . Public Choice , 40 (1), 53-69. Matsusaka, J. G. (2005). Direct Democracy Works. The Journal of Economic Perspectives , 19 (2), 185-206. Ober, J. (2007). The original meaning of “democracy”: Capacity to do things, not majority rule. Stanford University, Political Sciences. Princeton/Stanford Working Papers in Classics. Persson, T., & Tabellini, G. (2003, March). Do electoral cycles differ across political systems? . IGIER Working Paper No. 232 .

Persson, T., & Tabellini, G. (2001). Political Institutions and Policy Outcomes: What Are the Stylized Facts? No. 459. Rogoff, K. (1987). Equilibrium Political Budget Cycles. NBER Working Paper Series. Cambridge: National Bureau of Economic Research. Shi, M., & Svensson, J. (2006). Political budget cycles: Do they differ across countries and why? . Journal of Public Economics , 90 (8), 1367-1389. Stock, J. H., & Watson, M. W. (2015a). Assessing Studies Based on Multiple Regression. In J. H. Stock, & M. W. Watson, Introduction to Econometrics (pp. 361-395). Essex: Pearson Education Limited. Stock, J. H., & Watson, M. W. (2015b). Regression Analysis of Economic Time Series Data. In J. H. Stock, & M. W. Watson, Introduction to Econometrics (pp. 568-717). Essex: Pearson Education Limited. Tavares, J. (2004). Does right or left matter? Cabinets, credibility and fiscal adjustments. Journal of Public Economics , 2447 – 2468. Ueyama, K. (2013). Economic Voting in the Netherlands: Voter Behavior in a Multiparty Democracy. Veiga, L., & Veiga, F. (2007a). Does opportunism pay off? Economic Letters , 96 (2), 177-182. Veiga, L., & Veiga, F. (2007b). Political Business Cycles at the Municipal Level. Public Choice , 131 (1/2), 45-64. Vliegenthart, R. (2014). Moving up. Applying aggregate level time series analysis in the study of media coverage. Quality & Quantity , 48 (5), 2427-2445.

Appendices

A) The different dates of elections along with the ruling of governments from 1971 on. There were no elections for Van Agt III because CDA and D66 decided to continue their ruling without the majority of seats. Election date

Ruling from Ruling to Prime-minister Involved parties Position cabinet Left-Right 26/03/1971 06/07/1971 10/05/1973 Biesheuvel KVP, ARP, CHU, VVD en DS'70 -18,25 29/11/1972 11/05/1973 18/12/1977 Den Uyl PvdA, PPR, D66, KVP en ARP -31,6 25/05/1977 19/12/1977 10/09/1981 Van Agt I CDA, VVD -2,5 26/05/1981 11/09/1981 28/05/1982 Van Agt II CDA, PvdA, D66 -23,5

No elections 29/05/1982 03/11/1982 Van Agt III CDA, D66 -20

08/09/1982 04/11/1982 13/07/1986 Lubbers I CDA, VVD 2,8 21/05/1986 14/07/1986 06/11/1989 Lubbers II CDA, VVD -0,5 06/09/1989 07/11/1989 21/08/1994 Lubbers III CDA, PvdA -14,4 03/05/1994 22/08/1994 02/08/1998 Kok I PvdA, VVD, D66 4,7 06/05/1998 03/08/1998 21/07/2002 Kok II PvdA, VVD, D66 -10 15/05/2002 22/07/2002 26/05/2003 Balkenende I CDA, VVD, LPF 10,2 22/01/2003 27/05/2003 06/07/2006 Balkenende II CDA, VVD, D66 12,1

No elections 07/07/2006 21/02/2007 Balkenende III CDA, VVD 12,4

22/11/2006 22/02/2007 13/10/2010 Balkenende IV CDA, PvdA, CU 3,1 09/06/2010 14/10/2010 04/11/2012 Rutte I VVD, PvdA, PVV 11,1 12/09/2012 06/11/2012 ….. Rutte II VVD, PvdA 8,6

B) The variables and their transformations to stationarity. Once a variable is stationary, there are no further transformations conducted.

Variables (x) Stationarity x Stationarity of ln.(x) Stationarity of d.(x) Budget deficit as % of GDP 0.0591* X X Social securities as % of GDP 0.8483 0.8512 0.0036*** Election dummy 0.000*** X X Economic growth per capita 0.000*** X X Unemployment rate 0.2977 0.0012*** X Inflation 0.3784 0.0494** X Debt to GDP ratio 0.7650 0.7668 0.0008*** Interest% Dutch government has to pay 0.9286 0.7184 0.0000*** Interest payment as % of GDP 0.9220 0.6352 0.0001*** Position of the government 0.1773 0.4161 0.0000*** *Stationair at the 10% level ** Stationair at the 5% level *** Stationair at the 1% level

C) Dependent variable = Budget deficit as percentage of GDP. Linear regression with robust standard errors is used, unless indicated otherwise Variable/reg no. 1 (homosceda stic only SE) 2 (homoscedasti c only SE) 3 4 5 6 7 8 Constant -.2964934 (1.191701) -1.039504 (1.160574) -2.529055 (2.425312) -1.113422 (.9952743) -1.262736 (1.131428) -.522388 (.8903721) -.4405246 (1.2651) -2.511601 (1.612845) Election dummy (t-1) .8027157 (.6628422) .7241598 (.821938) .6329692 (.5704922) .7511683 (.6156218) .5343256 (.5843948) Election dummy (t) .3557138 (.599533) .5913893 (.650071) 1.120369 (.8235394) .7028006 (.4909001) .7191972 (.5178925) .7099613 (.508198) Election dummy (t+1) .5704787 (.7068072) 916666 (.7201298) .762196 (.6021058) .7788222 (.6143508) .8351855 (.638718) Corrected election dummy (t-1) -.1948421 (1.012674) Corrected election dummy (t) .2535881 (.9003912) Corrected election dummy (t+1) -.5753031 (.9949427) Expected years to election = 3 -.0140431 (.8652254) Expected years to election = 2 -1.028819 (.8675003) Expected years to election = 1 .3735304 (.7846706) Expected years to election = 0 .1975592 (.6498257) Budget deficit (t-1) .519527*** (.1525103) .5120218*** (.1572571) .3740113 (.3119432) .4409109** (.1733297) .4390271 (.1793107) .4105271* (.2052091) .4394893** (.1803952) .4634223** (.1638512) Budget deficit (t-2) .0359228 (.2034548) Economic growth per capita (t-1) -.0603841 (.1588774) .0704204 (.2280531) Economic growth per capita (t-2) .1162932 (.1856649) ln (Unemployment rate (t)) 1.058966 (.7750694) 1.166109 (.7956391) 4.362841* (2.170669) (.5708898) 1.305232* 1.361847 (.6235734) 5.079426* (2.147615) 4.171097** (2.028244) 1.800292*** (.7367557)

rate (t-1)) (1.987812) (1.563741) (1.545242) ln(Inflation (t)) .8750933 (1.393044) .5267726 (.5770855) ln(Inflation (t-1)) -.5082227 (1.215004) Δ Position of government (t) -.003683 (.0340023) -.0234464 (.0266545) .0116013 (.0150643) Δ% of GDP spent on interest (t) -.0351197 (.0404847) -.0368635 (.0440748) .6849831* (.4774783) .9226229** (.3715752) .5668901 (.3933099) .5419579 (.3388436) 1.13912*** (.4034869) Δ % of GDP spent on interest (t-1) .0775134 (.038881) .0240754 (.0347341) .0471106 (.0323914) Δ Annual interest rate on Dutch bonds (t) -29.43694 (19.83198) .5039066 (.2927098) -24.4207 (16.63105) -23.06795 (14.16091) Δ EMU debt/GDP (t) .0662331 (.0917298) .1493977 (.1081784) .1978825 (.1138357) .0726421 (.0852153) .0990175 (.083572) Δ EMU debt/GDP (t-1) .1210987 (.1953285) .0479556 (.1352489) .0541127 (.1305368) N 43 42 38 42 42 42 42 42 R² 0.5751 0.6733 0.6277 0.6200 0.6515 0.6690 0.5922

Adj. R² 0.4556 0.4876 n/a n/a n/a n/a n/a n/a

AIC 178.1 177.2 165.49 168.51 171.37 170.2 172.73 163.9 BP for heteroscedasticity p-value 0.4402 0.8319 0.2071 0.7974 0.7721 0.0877 0.0510 0.0962 P-value joint hypotheses test on election variables 0.6333 0.5785 0.7236 0.4606 0.4399 0.6838 0.5960 0.4059 * Significant at the 10% level ** Significant at the 5% level *** Significant at the 1% level