Bunching at Kink Points in the Dutch Tax System

Transcription

1 Bunching at Kink Points in the Dutch Tax System Vincent Dekker Kristina Strohmaier 13th September 2015 Abstract This paper presents new empirical evidence on taxpayers responsiveness to taxation by estimating the compensated elasticity of taxable income with respect to the net-of-tax rate in the Netherlands. Making use of the bunching approach introduced by Saez (2010), we find small, but clear evidence of bunching behaviour at all thresholds of the Dutch tax schedule with a precise estimated elasticity of at the upper threshold. In line with the literature, we find much larger estimates for women and self-employed individuals, but we can also identify bunching behaviour of male and employed individuals. We add to the bunching literature by proposing an intuitive, iterative procedure for determining the bunching window as well as the optimal counterfactual model, where we rely on the information criteria to determine the model. JEL Classification: H21, H24 Keywords: Bunching, Elasticity of Taxable Income, Optimal Bunching Window, Counterfactual Model We thank Nadja Dwenger, Nadine Riedel and Robert Stüber for invaluable comments on an earlier draft, as well as the participants of the 71st Annual Congress of the IIPF in Dublin, 2015 for helpful comments and discussion. University of Hohenheim, Institute for Economics, Chair for Public Finance. vincent.dekker@uni-hohenheim.de Ruhr University Bochum, Institute for Economics, Chair for Public Finance and Economic Policy. Kristina.Strohmaier@rub.de 1

2 1. Introduction Do individuals respond to income taxation? This central question has been posed in a vast amount of studies in the field of public economics that are concerned with the corresponding welfare consequences. The potential channels of responsiveness to taxation are manifold, but the focus of a recent strand of the literature has been on estimating the compensated elasticity of taxable income to determine the deadweight loss (see e.g. Feldstein (1995), Chetty (2009), Saez et al. (2012)). Depending on the identification strategy or data used, a wide range of estimates exists. 1 Current research focusses on explaining this variance in the estimates and explains the modest elasticities found in recent microeconometric studies by optimisation frictions, intertemporal shifting of income and differences between short- and long-term elasticities (Chetty et al. (2011), Bastani and Selin (2014), Le Maire and Schjerning (2013)). We estimate the compensated elasticity of income with respect to the net-of-tax rate in the Netherlands using the bunching approach introduced by Saez (2010). For identification one uses the clustering behaviour of individuals at kinks of a non-linear tax system 2 and measures the elasticity of taxable income by the amount of individuals that adjust their income to stay under the threshold at the kink. Because the Dutch tax schedule exhibits three such kinks, it lends itself perfectly to identify the elasticity of taxable income with the bunching approach. Moreover, we can exploit a unique data set that contains the exact taxable income of about 1% of the Dutch population observed for a time period of 2003 to It contains rich covariates such as age, gender, marital status as well as information on self-employment that we use to construct our different sub-samples. Next to implementing the bunching approach for identification of the elasticity of taxable income in the context of the Netherlands, we contribute to the literature by improving the methodology in two ways: first, we propose a simple, data-driven procedure to determine the bunching window, rather than the visual inspection that is used in other studies. We also allow this bunching window to be asymmetric around the threshold. Secondly, we rely on information criteria to select the best counterfactual model and find that contrary to the existing literature, a case-by-case selection of the order of polynomia for the counterfactual distribution improves the significance of the results. In the pooled sample from 2003 to 2013, we estimate an elasticity of taxable income with respect to the net-of-tax rate of at the highest tax threshold, which is significant 1 See Saez et al. (2012) for a comprehensive overview. 2 Kinks appear at thresholds of a tax schedule, where the marginal tax rate jumps up. 2

3 at the one-percent level. The bunching window is between -750 e and +350 e from the threshold and the counterfactual is estimated with a linear model. Our procedure produces more efficient estimates of the elasticity compared to traditional approaches from the literature, which holds across all specifications. The magnitude of our results is in line with some of the literature, as for example Chetty et al. (2011) find an elasticity below 0.02 for their full sample in Denmark, but Bastani and Selin (2014) only find an elasticity of for the full sample of Swedish tax payers. Like all studies implementing the bunching approach, we find significantly higher compensated elasticities of taxable income for women and self-employed individuals, but contrary to most studies, we can identify a positive elasticity for employed individuals as well. After giving a short theoretical motivation and commenting on our empirical identification strategy in Section 2, we will present the institutional setting and the data in Section 3. Section 4 contains the results of our analysis of bunching behaviour at kink points in the Dutch tax system. Section 5 concludes. 2. Theoretical Motivation and Estimation There are numerous ways in which individuals may potentially alter their behaviour in response to taxation. As suggested by standard microeconomic theory, the distortion of prices and wages in the economy due to taxation induces individuals to adjust their working hours or working effort. To test this prediction empirically and to quantify the responses, the amount of bunching individuals is used for the identification of the compensated elasticity of taxable income in the spirit of Feldstein (1995). This central parameter is defined as the percentage change in taxable income z due to an increase in the net-of-tax rate 1 τ of one percent: e(z) = dz z d(1 τ) 1 τ. (1) More technically spoken, the introduction of a kink in the budget set of individuals induces bunching behavior within a certain income range provided that preferences are convex and smoothly distributed in the population. For illustration, assume that there exist two periods: in the pre-reform period with a constant marginal tax rate τ 1, individual before-tax incomes z are smoothly distributed according to a smooth density distribution h 0 (z). Now, a change in the marginal tax rate from τ 1 to τ 2 with τ 1 < τ 2 is introduced at a specific earnings level k so that all individuals with z > k face a higher 3

4 marginal tax rate. 3 For the post-reform income distribution h 1 (z), this implies that all individuals in the pre-reform period with z k are not affected by the reform (meaning that they stay on the same indifference curve). Those with z > k + dz will reduce their taxable income, but they will not exactly bunch at k in response to the reform (see Saez (2010) for a more detailed illustration here). In contrast, all individuals with income z (k; k +dz] will theoretically move to z = k. This bunching behaviour induces a spike in the income density distribution and the density above k + dz shifts to k. For small tax changes, the mass of taxpayers bunching is given by: B = k+dz k h 0 (z) dz h 0 (z) dz (2) where the approximation follows from the mean value theorem for integration. Substituting dz = from (2) into (1) at z = k leads to B h 0 (z) e(k) = B(dz) k h 0 (z) log( 1 τ 1 1 τ 2 ). (3) Thus, the elasticity of taxable income is non-parametrically identified by expression (3) if and only if the derivative h 0 (z) with respect to z is continuous in z z, which means there should be no peak in the pre-reform distribution at the kink point. However, although the counterfactual pre-reform density is not observable in reality, in most applications it seems reasonable to hold on to that assumption if there are no obvious violations. 4 Note, that the expression in (3) shows that the elasticity parameter is proportional to the number of taxpayers who bunch at the kink point k. While k and log( 1 τ 1 1 τ 2 ) from (3) are directly observable policy parameters, the only thing that has to be estimated is the relative excess mass of taxpayers defined by b = B(dz). h 0 (k) We follow the estimation approach developed in Chetty et al. (2011) and estimate the counterfactual density h 0 (k) the density in absence of any kink directly by local polynomial regression. However, for plausible reasons, the empirical post-reform distribution will not have a single spike at z = k even if there is behavioral reaction in response to a tax change in the population. Therefore, the theoretical spike at k will become an empirical bunching window around k ( imperfect bunching ). Some of the 3 To abstract from any income effects, the utility function is assumed to be quasi-linear. However, simulations by Bastani and Selin (2014) show no economically significant bias of income effects on compensated elasticities, which would make it superfluous to assume such a utility function. 4 Besides, there exist some smoothness checks depending on the respective setting. Le Maire and Schjerning (2013) propose to exploit shifting of kinks over time to examine the smoothness of the density at the respective threshold in the after-shift period. Moreover, it seems plausible to assume smoothness if the distribution to the left and right of the bunching window does not show any jumps. Nevertheless, the assumption cannot be tested directly. 4

5 reasons for this are uncertainty in income due to random income components (income volatility) or the inability to perfectly adjust labour supply in case of contracted hours constraints from the employer (see Chetty et al. (2011)). Additionally, to account for changing thresholds over time when using taxable income from several years, we need to re-centre the income variable by calculating the difference between taxable income and the respective threshold before pooling the data. Individuals are then grouped into bins of length δ, with Z j being the midpoint of the distance interval. Thus, Z j is the absolute distance between income bin j and the threshold k where k differs over years (e.g., due to inflation adjustments). To estimate the counterfactual distribution, Chetty et al. (2011) propose a two-step procedure using a local polynomial regression while excluding all observations within the bunching window [ R; R]: Ñ j = q β i Z i + i=0 R i= R γ i I[Z j = i] + ɛ j (4) with N j denoting the number of individuals in income bin j and q denoting the order of the polynomial. I is an indicator function equal to one if the bin point lies within the bunching window. ɛ j denotes the error of the polynomial regression. A initial simple estimate of the number of bunching individuals is B = R R (N j ˆN j). However, this estimate would overestimate the excess mass as the integration constraint is not satisfied. The area under the counterfactual density would not sum to one as the observations within the bunching window (with an expected higher density) are left out from the regression. Therefore, the counterfactual density has to be shifted to the upper right for those observations that lie to the right of the bunching threshold. ˆB is thus equally distributed on the total right tail of the income distribution following B N j (1 + I[j > R] R+1 N ) = j q R β i Z i + γ i I[Z j = i] + ɛ j. (5) i=0 i= R To get the relative excess mass ˆb, the estimated bunching mass has to be related to the (average) height of the counterfactual density at the kink in the following way ˆb = ˆB R R ˆN j 2R+1. (6) Along with the policy parameters k, expressed in terms of δ (Bastani and Selin (2014)), and log( 1 τ 1 1 τ 2 ), the compensated elasticity of taxable income with respect to the net-oftax-rate can be estimated by inserting all values into (3). With regard to the calculation of standard errors, we decided to use a parametric residual bootstrap for our analysis. 5

6 Our estimation procedure extends the existing literature in two ways. First, we do not exogenously use a 7th order polynomial model as counterfactual. 5 We exploit information criteria to select which model is best suited as counterfactual model in each specification, thus making the choice of the counterfactual endogenous. Because of the large sample size, we prefer Schwarz s Bayesian Information Criterion (BIC) that has a better punishing mechanism for high N, although the Akaike Information Criterion (AIC) delivers similar results. We specify several models ranging from a simple linear model to a 7th order polynomial model and then evaluate each model using the BIC. By implementing this procedure, we find in most cases that the simple linear regression is the best approximation for the counterfactual. Second, we rely on the data at hand to determine the bunching window. It should optimally comprise of all individuals that adjust their taxable income as a reaction to the tax change at the threshold. The bunching window should not be too small, thereby omitting some taxpayers that bunch at the kink, nor should it be too large, which would bias the results downward, because non-bunchers are included as well. The existing literature implements symmetric bunching windows around the kink with varying sizes that are determined by graphical inspection. We propose to use a possibly asymmetric bunching window with an endogenously determined size. The argument for using a symmetric bunching window is that of imperfect control over taxable income and misjudgments by the individuals, which implies that some individuals who would want to bunch remain on the right side of the threshold (see e.g., Saez (2010)). We acknowledge this, but it is debateable how big this error in judgement really is. Using a bin width of 100 e, which will be our baseline specification 6, the bin located exactly at the kink ranges from -50 e to +50 e and therefore already accounts for those individuals who erroneously report slightly too much income or are not able to adjust their labour supply. In order to determine the optimal bunching window, we iterate through the following step-wise procedure: 1. Set an excluded region. 2. Run a linear regression through all data bins outside the excluded region and predict the values for Z. 3. Compute a confidence interval around the prediction. 4. Subsequent bin midpoints outside the confidence interval comprise the bunching window. 5 This is done by Chetty et al. (2011) and subsequently by many other authors. 6 This is a rather small bin size compared to the 1,000 DKR used in Chetty et al. (2011) or the 1,000 SEK used by Bastani and Selin (2014), but for the sake of precision, it should be preferable to use a small bin width, as long as there are enough observations. Our results are robust to larger bin sizes. 6

7 In general, the excluded region can be set arbitrarily, however we propose to iterate through different combinations of upper and lower bounds of the excluded region to check for the sensitivity of the bunching window to the choice of the excluded region. The choice of the appropriate confidence interval is also at the researchers discretion, where higher confidence levels tend to lead to a smaller bunching window. Depending on the setting and the data, this will lead to more conservative estimates of the elasticity. A graphical intuition for the procedure is given in figure 1. The bins around the threshold that have a higher actual number of taxpayers than predicted (coloured in red in figure 1) are then used as bunching window. Figure 1: Data-Driven Procedure to Determine Bunching Window 0 Distance to Threshold Notes: This figure shows the bin midpoints as well as the the fitted values of a linear regression. The grey confidence band is calculated with the standard errors of the point prediction. Here, four bin midpoints lie outside and therefore determine the relevant (asymmetric) bunching window. 3. Institutional Setting and Data The Dutch tax system is an individualised progressive tax system that taxes all residents on their worldwide income. Each individual is treated as an own entity and taxed separately on their taxable income with the exception of just a very few joint components (e.g. children s allowances). Moreover, the system is characterised by a clear separation of labour and capital income. There are three different categories of income according to the source of income generation. Each of these three income boxes has its own tax schedule and its own definition of taxable income. Box 1 includes the income from employment, 7

8 business profits and further regular cash payments such as income from owner-occupied dwellings or any kind of social transfers. Box 2 income includes dividends and other capital gains derived from substantial shareholdings. Any other income from savings and investments is taxed in income Box 3. Whilst Box 1 income is taxed at progressive rates that jump up at certain thresholds and thus creating kinks in the tax schedule, Box 2 and Box 3 income are subject to a flat tax, which was 25% and 30% respectively in It is worth noting that losses from different boxes cannot be offset against each other. The focus of our analysis is on Box 1 income, because it contains the income from labour and the tax schedule exhibits kinks that we can use for identification. It should be emphasised that for employees the employer withholds income tax from Box 1 as a wage tax, which can be seen as a prepayment credited against the final tax amount at the end of the year. This third-party reporting is important for the interpretation of the results as systematic tax evasion, which is one possibility to adjust taxable income, becomes more difficult. Income falling into Box 1 minus personal deductions is taxed at progressive tax rates. The tax schedule consists of four tax brackets with increasing marginal tax rates. Figure 2 shows the development of the marginal tax rates over the period from 2001 until The two lower tax rates do also include a social security contribution of around 33%. 7 Due to the social security contributions that apply in the second but not in the third tax bracket, marginal tax rates in the two brackets are similar. However, there is a large jump in marginal tax rates from 42% to 52% for the last bracket. In the considered time period, the income thresholds have been adjusted upwards by a weighted derived consumer price index to account for inflation and to avoid the cold progression phenomena. Figure 3 shows the development of the thresholds over the sample period 2001 to Although we analyse all thresholds, we focus on the third threshold for several reasons: First, at the lower thresholds, the argument could be made that income is very insensitive to tax changes, because people need their full income for a living and are less likely to adjust their labour supply for tax saving purposes. Consequently, individuals with a high income are more prone to adjust their taxable income in the light of a kinked tax schedule. Secondly, the jump in the marginal tax rate is substantial at the third threshold, which implies that the opportunity costs of working an additional hour to obtain income which is taxed at the higher tax rate are significantly increased. Based on this line of reasoning, an individual located at or around the third threshold could be 7 This should not be confused with employee social security contributions which is solely withheld for salaries. 8

9 Figure 2: Development of Marginal Tax Rates Notes: The figure depicts the changes in the marginal tax rates in the Netherlands from 2001 to Figure 3: Development of Tax Thresholds Notes: The figure depicts the changes in the threshold values in the Netherlands from 2001 to Threshold 1 is between MTR 1 and MTR 2, Threshold 2 between MTR 2 and MTR 3 and Threshold 3 between MTR 3 and MTR 4. 9

10 more likely to strategically manipulate its taxable income in order to avoid falling into the higher tax bracket. Also, analysing the top tax threshold is in line with previous literature (e.g. Chetty et al. (2011)). The data used in this paper comes from the Inkomstenpanelonderzoek (IPO) from Statistics Netherlands and covers the years 2001 to For our main specification, we will exclude the years 2001 and 2002, because we suspect to pick up after effects of the major tax reform in the Netherlands in The aim of Statistics Netherlands is to provide a picture of the distribution of income in the Netherlands. This micro dataset contains individual level data on all sources of income that an individual might have, as well as a very detailed account of possible deductions from the tax base. The dataset is a panel of individuals, which are followed over time. The panel is updated with spouse and other randomly selected individuals in every period to account for people who are no longer observable. Most importantly for this study, Statistics Netherlands provides the taxable income not only in total, but also split into the three boxes of the Dutch tax system. As we are only interested in taxable income from labour, we will rely on information on Box 1 taxable income. In addition, the dataset provides demographic characteristics, which we exploit to study heterogeneity in the bunching behaviour of different socio-economic groups. We use information on self employed individuals, who are theoretically more prone to bunching due to the lower costs of adjusting their taxable income. Furthermore we distinguish between genders and show results for married individuals. We exclude students as well as all people that receive any (governmental) transfers, because they could distort the results from our bunching analysis due to the fact that most of them receive transfers of similar hight, thus creating an artificial bunching point. Because the tax system changes for individuals aged 65 and over, we exclude them from our estimation as well as individuals below the age of 18 (European Tax Handbook 2005). Furthermore, we only keep individuals that have a positive reported taxable income in box 1. For our main specification, we exclude the years 2001 and 2002, because we suspect to pick up after effects of the major tax reform in the Netherlands in Table 1 shows descriptive statistics of our data. Self employed, gender and marital status are dummy variables indicating self employment, being a man or being married respectively. Our sample seems to be evenly balanced with respect to gender, which enhances the possibility to apply the obtained results to the population. The sample consists of 14% self employed individuals 9, 55% are male and the share of married 8 The tax reform changed the thresholds and marginal tax rates substantially. 9 This includes CEO s that are also major shareholders of their firm, because they have the possibility to decide on their own salary and are flexible in adjusting it. Thus, they have the same possibility 10

11 Table 1: Descriptive Statistics of the Pooled Sample Mean Number of Observations 1,219,572 Standard Deviation Share of Self Employed Share of Male Share of Married Notes: The table shows the mean values and the standard deviations for the variables. The samples are all restricted to individuals between the age of 18 and 64 that earn a positive income and do not receive any transfers. individuals is 65%. Note that these descriptive statistics are for the pooled sample, which consists of all years from 2003 until 2013 pooled together. 4. Empirical Results Figure 4 shows the income distribution for the most recent year of our sample. The data is collapsed into income bins of 100 e for our entire analysis. 10 The threshold incomes of 2013 are indicated by a vertical red line. The figure gives a first hint of bunching behaviour at the first and third threshold. Note that the change in the marginal tax rate at the second threshold is merely 1.08 % and so the incentive for adjusting taxable income is small Bunching at the Upper Threshold The change in the tax rate is largest at the upper threshold of the Dutch tax system with %, which would imply that bunching behaviour should be more pronounced here. We report the results for our pooled sample from 2003 to 2013 in figure 5. We estimate an excess mass (b) of taxpayers at the threshold of 2.4, which corresponds to 240 % more individuals being at the threshold than would have been in the absence of the tax change. This confirms the clear graphical bunching evidence from the figure. This excess mass implies an elasticity of taxable income with respect to the net-of-tax rate of 0.024, which is statistically significant for all usual significance levels. Quantitatof adjusting their taxable income as self employed individuals. 10 Our results are not driven by the selection of this binwidth. For higher binwidths (200 and 400), the bunching window would be smaller and similar estimates are obtained. 11

12 Figure 4: Sample Income Distribution of Notes: This figure shows the sample distribution of income below 70,000 e in the Netherlands for The data is collapsed into 100 e bins. The vertical red lines represent the first, second and third threshold of the Dutch tax system respectively. ively, a decrease in the net-of-tax rate by 10 % induces a reduction of taxable income by 0.24 %. One concern that could arise is that when pooling the data, we observe many individuals more than once. If for example a contract is signed for several years, which specifies that the salary moves along with the threshold, we would attribute bunching behaviour in every period to this individual, although the behavioural decision was made only once. This could possibly lead to an overestimation of the excess mass. To hedge against this, we randomly kept one observation per individual an reestimated the excess mass in the pooled sample. The excess mass drops to 2.02, which is in support of this hypothesis, but this could also be driven by the smaller sample size. We expect the bunching behaviour to vary with employment status and sex. On the one hand, self-employed individuals have better possibilities to adjust their taxable income and are therefore more prone to bunching. On the other hand, women, who are often second earners, are also more sensitive to changes in taxation. The results are shown in figure 6. Our results confirm the hypotheses, as the excess mass for the self-employed nearly doubles to 4.42 compared to the baseline analysis of the pooled sample. Contrary to many other studies, we also find a significant excess mass for employed individuals of 1.77, so the baseline result is not purely driven by self-employed individuals. This can be seen as an indication that collusion between the employer and the employee is present and contracts are specifically designed to achieve a taxable income at the threshold. 12

13 Figure 5: Bunching at the Third Threshold - Pooled Sample b = e = se = Notes: This figure shows bunching at the third threshold for the pooled sample from 2003 to The data is collapsed into 100 e bins. The bunching window is between -750 e and +350 e and the counterfactual model is linear. The running variable is distance to the threshold. In order to calculate an elasticity, a weighted average threshold value is used. The weights are number of taxpayers exactly at the threshold in each year. This holds for all figures. Standard errors are calculated with a parametric residual bootstrap procedure. The bottom two graphs in figure 6 are clear evidence in favour of the gender difference hypothesis, as the excess mass of women at the third threshold is 5.42, whilst the excess mass for men is only For self-employed women, the excess mass rises to 7.17 (not depicted) and is still significant, but due to the very small sample size, this result should be viewed with caution. We estimated the excess mass of taxpayers and the elasticity of taxable income at the third threshold for all years separately. This hedges against the concern that we use a weighted average threshold in the pooled sample to obtain our elasticity estimate. The results are shown in the appendix. One striking observation is that the bunching behaviour of individuals increases and becomes more precise over time. We ascribe this to learning effects as the taxpayers become more familiar with the tax system and see this as evidence for a behavioural response to taxation. In the year 2002 we can still observe delayed effects from the major tax reform of 2000 and therefore, the bunching behaviour is fuzzy and small. It then increased in the subsequent years until the excess mass reached a level of around 2, corresponding to an elasticity of

14 Figure 6: Bunching at the Third Threshold - Subsamples b = e = se = b = e = se = b = e = se = b = e = se = Notes: The figures show bunching at the third threshold from 2003 to 2013 for different subsamples. The bunching window for the employed sample is between -550 e and +250 e and the counterfactual model is linear. The bunching window for the self-employed sample is between -850 e and +450 e and the counterfactual is a second order polynomial model. The bunching window for the male sample is between -650 e and +350 e and the counterfactual model is linear. The bunching window for the female sample is between -850 e and +450 e and the counterfactual model is linear. Note that the number of observations varies between the subsamples, which should be take into account for the graphical comparison of the bunching behaviour. 14

15 4.2. Bunching at the Other Thresholds The focus of this paper has been on the top tax threshold, but a case could be made for bunching at the other thresholds of the Dutch tax system as well. Our a priori hypothesis is that we should observe limited bunching at the other two thresholds. At the second threshold, the change in the tax rate is very small, especially in the more recent years, as was shown in figure 2. At the first threshold, the income levels are quite low, which would suggest that most individuals need their full income for a living and should react insensitive to changes in the marginal tax rate. 11 Figure 7 shows the graphs for the pooled sample. Figure 7: Bunching at the First and Second Threshold - Pooled Sample Pooled Sample, Threshold Distance to Threshold Pooled Sample, Threshold Distance to Threshold 2 Notes: The figures show bunching at the first and second threshold for the pooled sample from 2003 to Due to the varying tax changes at these thresholds, no excess mass is reported. Surprisingly, we observe bunching behaviour of individuals at both thresholds. We cannot display exact estimates for the excess mass or the elasticities at these thresholds, because we have changing tax differences over time on top of the changing threshold values. Taking the average of the single-year estimations delivers an excess mass of 1.55 at the first threshold and 0.47 at the second threshold. Especially at the first threshold, the estimated average excess mass is comparable to the single-year average excess mass at the third threshold, which is A possible explanation for this could be that many second or part-time earners could realise an income around this level. Given that they are not dependent on this income, these individuals could be less constraint than we assumed. Therefore, we should run the analysis for the first threshold separately for married and unmarried individuals to test this hypothesis, but this is left to future research Bear in mind that only income above the threshold is taxed at the higher tax rate and therefore, more pre-tax income will lead to more disposable income. 12 A short analysis on a subsample of three years delivered first evidence in favour of this hypothesis, as 15

16 At the second threshold, the bunching behaviour is largely driven by early periods of the sample, where the jump in the marginal tax rate was still noticeable. To further investigate the possibility of adjustment costs that would deny bunching behaviour, we can exploit the fact that in 2009, the jump in the marginal tax rate at the second threshold was zero. The incentive to adjust taxable income below the threshold vanished and therefore, higher income was strictly preferable. After that, the adjustment costs were higher than the marginal gain of lowering taxable income and we would not observe bunching anymore. This is depicted in figure 8. We find clear evidence for a behavioural response of individuals to taxation. In 2008, we could still observe a small excess mass of 0.65, but in 2009, where the incentive to adjust taxable income vanished, we estimate an excess mass of merely The fact that the excess mass does not rise to the same level as before 2009 even though the tax change was similar in 2010 and 2011 compared to 2008, is mild support for the hypothesis of adjustment or entry costs. Comparing the estimates from the first, second and third threshold, we conclude that the behavioural responses to taxation are heterogeneous and depend on the location of taxable income within the income distribution. This is an aspect that deserves a more thorough discussion in the future, although it is beyond the scope of this paper to shed further light on this matter. the excess mass for married individuals was nearly twice as high as for unmarried individuals. This suggests that the result at the first threshold is mainly driven by unconstrained second earners. 16

17 Figure 8: Bunching at the Second Threshold b = 0.65 b = 0.04 b = 0.23 b = 0.19 Notes: The figures show bunching at the second threshold for the years from 2008 to The bunching window for 2008 is between -50 e and +250 e and the counterfactual is a second order polynomial model. The bunching window for 2009 is between -50 e and +50 e and the counterfactual model is linear. The bunching window for 2010 is between -50 e and +50 e and the counterfactual is a third order polynomial model. The bunching window for 2011 is between -50 e and +50 e and the counterfactual is a second order polynomial model. Because of the zero tax change in 2009, we cannot calculate an elasticity for Therefore, only the estimated excess mass is reported in the graphs. 17

18 5. Concluding Remarks In this paper we have estimated the (compensated) elasticity of taxable income with respect to the net-of-tax rate in the Netherlands. Using a unique data set from Statistics Netherlands for the years 2001 to 20013, we exploit bunching behaviour at kink points of the Dutch tax schedule. The data set used in this study has the important advantage that it gives detailed information on taxable income. Our preferred specification for the upper threshold yields an excess mass of 2.4 with an estimated elasticity of With an excess mass of 4.42 for the self-employed and 5.24 for women, the estimates are in line with the third-party hypothesis and further suggest that women are more responsive to taxation. Compared to previous studies estimating the elasticity of taxable income, we find rather small behavioural responses to taxation. However, our findings are quantitatively similar to recent studies exploiting bunching with one exception. While Chetty et al. (2011) for Denmark or Bastani and Selin (2014) for Sweden find a nearly zero elasticity for wage earners, we find a small, yet statistically significant estimate of 0.02 for wage earners the Netherlands. A cross country comparison of elasticities might be difficult due to different institutional features (Bastani and Selin (2014)). To the best our knowledge, there is only one study in the Netherlands estimating the elasticity of taxable income so far that uses the large tax reform in 2000 as a natural experiment. For workers with high labour income, Jongen and Stoel (2013) find an elasticity of up to 0.46, which is much larger than our estimate. However, our data and methodology suggest a clear learning behaviour after the reform implying that the estimated elasticity increases in the years after reform. We have also contributed methodologically in two ways. Elasticities derived with the bunching approach heavily rely on both the estimated counterfactual density and the determination of the bunching window. To improve the reliability of the bunching estimation, we first propose to chose the counterfactual model based on the information criteria. Second, we implemented an intuitive purely data-driven procedure to find an optimal bunching window, that is neither too small (which would lead to an underestimation of the elasticity) nor too large. Starting with a large bunching window, all subsequent bin points lying outside the 95% confidence interval around the regression fit identify the bunching window. Applying these extensions to our data, we find elasticities that are marginally smaller, yet statistically more significant. In addition, our determination of the bunching window suggest the window to be asymmetric with considerably more excess mass to the left than to the right of the threshold. This is very intuitive in the sense that if people do not bunch perfectly, they are more likely to overadjust their income and hence are further to the left of the kink than underadjust their taxable in- 18

19 come. Our modifications are thus a valuable contribution to the literature as they allow for a more precise calculation of the excess mass at the kink, which should improve the relevance and validity of the inferred elasticities of taxable income. Overall, our empirical results show, that the Dutch population responds to taxation and adjust their taxable income. One of the reasons for the small elasticity of taxable income could be that we have not controlled for firm responses due to a lack of data. Especially at the third threshold, where the tax rate does not change over our sample period, it could be that firms offer contracts that exactly locate individuals at the kink in the tax system. Closely related to this is the aspect of collusion of the employee with the firm. An interesting question is, whether wage income is substituted by amenities, such as a company car. If such contracts are offered, the observed bunching effect might not be solely due to behavioural responses to the change in the net-of-tax rate. Evasion is another possible channel of consideration, especially for incomes that are not third-party reported. It is left to our future research to shed light into this matter. 19

20 References Bastani, S. and H. k. Selin (2014): Bunching and non-bunching at kink points of the Swedish tax schedule, Journal of Public Economics, 109, Chetty, R. (2009): Is the Taxable Income Elasticity Sufficient to Calculate Deadweight Loss? The Implications of Evasion and Avoidance, American Economic Journal: Economic Policy, 1, Chetty, R., J. N. Friedman, T. Olsen, and L. Pistaferri (2011): Adjustment Costs, Firm Responses, and Micro vs. Macro Labor Supply Elasticities: Evidence from Danish Tax Records. The quarterly journal of economics, 126, Feldstein, M. (1995): The Effect of Marginal Tax Rates on Taxable Income: A Panel Study of the 1986 Tax Reform Act, Journal of Political Economy, 103, Jongen, E. L. W. and M. Stoel (2013): Estimating the Elasticity of Taxable Labour Income in the Netherlands, CPB Background Document. Le Maire, D. and B. Schjerning (2013): Tax bunching, income shifting and selfemployment, Journal of Public Economics, 107, Saez, E. (2010): Do Taxpayers Bunch at Kink Points? American Economic Journal: Economic Policy, 2, Saez, E., J. Slemrod, and S. H. Giertz (2012): The Elasticity of Taxable Income with Respect to Marginal Tax Rates: A Critical Review, Journal of Economic Literature, 50, pp

21 A. Appendix Figure 9: Bunching at the Third Threshold - Single Years , Threshold , Threshold , Threshold b = 0.42, e = , se = b = 0.41, e = , se = b = 1.35, e = , se = , Threshold , Threshold , Threshold b = 1.98, e = , se = b = 1.80, e = , se = b = 1.94, e = , se = , Threshold , Threshold , Threshold b = 1.94, e = , se = b = 2.67, e = , se = b = 1.67, e = , se = , Threshold , Threshold , Threshold b = 2.61, e = , se = b = 2.27, e = , se = b = 2.03, e = , se =

22 Table 2: Bunching Window and Counterfactual for figure 9 Year Bunching Window Counterfactual Model e to +50 e linear e to +50 e linear e to +250 e linear e to +50 e Third-Order Polynomial e to +50 e linear e to +50 e linear e to +150 e linear e to +150 e linear e to +150 e linear e to +250 e Second-Order Polynomial e to +250 e linear e to +450 e linear Notes: The tabel shows the specifications used to obtain the estimates in figure 9. 22

23 Table 3: Comparison of Different Model Specifications (1) (2) (3) Sample b e se t b e se t b e se t Pooled Sample Employed Self-Employed Men Women Notes: (1) represents our model with the endogenous bunching window and the counterfactual model, which is determined by the BIC. (2) show the results for a symmetric bunching window going from -750 e to +750 e and using a 7th order polynomial counterfactual model. (3) show the results for a symmetric bunching window going from -350 e to +350 e and using a 7th order polynomial counterfactual model. b is the estimated excess mass, e the corresponding elasticity, se the standard error obtained from a parametric residual bootstrap procedure and t is a t-value, obtained by dividing the elasticity by the standard error.