Evidence from a Dutch tax credit
Author: Maarten Cornelis van Alphen
Supervisor: Maarten Allers
University of Groningen
January 2013
Abstract
I analyze the effect on labor market participation of the elderly worker tax credit, a fiscal measure implemented by the Dutch government in 2009 to increase labor market participation of older workers. I construct a difference-in-differences model, use data from the Labor Force Survey of Statistics Netherlands for the period 2002-2010 and find that the fiscal measure indeed increases labor market participation of older workers. Furthermore, the effect of the elderly worker tax credit on labor market participation is different between subgroups; the tax credit has especially increased the probability of participation for higher educated older workers.
Journal of Economic Literature codes:
H300 J210 J220 J260
Key words:
2
1. Introduction
In 2008, there were approximately 2.4 million people in the Netherlands above the age of 65. This amount will increase by 1.9 million to 4.3 million in 2038. By that time, the share of people above the age of 65 relative to the whole population will be around 25 percent (Garssen & Van Duin, 2007). The number of people above the age of 65 relative to people in the age range of 20-65, a common statistic to indicate the level of ageing in the economy, is expected to nearly double in the upcoming years, from 25.6 percent to 47 percent in 2011 and 2038, respectively (Garssen & Van Duin, 2007 & Statline CBS). In Figure 1, I show the average age of the workforce and the participation rate of individuals of 50 to 65 years old.
Figure 1: Average age workforce & participation rate 50-65 year olds
Source: Statistics Netherlands (CBS)
3
in participation of the control group are similar to both treatment groups. I use a linear probability model to estimate the effect of the elderly worker tax credit on the probability of participation of the treatment groups.
First, I find that the elderly worker tax credit increased the probability of participation by approximately 3 percentage points. Second, I find different effects of the elderly worker tax credit for subgroups. Higher educated persons, immigrants and singles are in general more responsive to the fiscal measure. Moreover, females are more sensitive to the elderly worker tax credit at low minimum hours of participation, whereas males react stronger to it when minimum hours of participation are higher. Third, I find different treatment effects in my first treatment group, by specifying the benefits for 62, 63 and 64 years old individuals separately. This shows that the effect of the elderly worker tax credit decreases with age as well as minimum hours of participation. Finally, I show there is a treatment effect in my second treatment group in 2010. A number of robustness checks confirm my results. I start with an introduction of the elderly worker tax credit.
1.1 Natural experiment: The elderly worker tax credit
In an attempt to face the challenges resulting from demographic changes on the labor market, in particular ageing, the Dutch government searched for effective fiscal measures to postpone retirement. In 2009, the Dutch government implemented a fiscal measure to postpone retirement of old workers. This fiscal measure, the elderly worker tax credit, gives individuals a financial incentive to continue working at older ages. At the time the fiscal measure was implemented there was a large unused potential of labor supply consisting of individuals between 62 and 65. According to the Netherlands Bureau for Economic Policy Analysis (CPB, 2008), financial incentives are assumed to be moderately effective to increase labor market participation of this age group, which increases the sustainability of the social security system in the Netherlands.
4
benefit workers who significantly participate in the labor force there is, besides age, also a minimum gross income requirement of 8860 euro to be eligible for the elderly worker tax credit. Furthermore, labor income above 54776 euro is not considered into calculations of tax subtraction that is the result of participation in the workforce. In 2009, the tax cut was calculated by multiplying the age corresponding percentage by the difference between the maximum and minimum income (Dutch Ministry of Finance, 2008). This results in the following Table with maximum tax benefits per age category:
Table 1: Tax benefits per person (in 2009)
Age 62 63 64 65
Percentage 5% 7% 10% 2%
Maximum tax credit (in €) 2296 3214 4592 918
The fiscal measure was expected to provide an incentive to continue working, as well as to work more, since the amount of the bonus increases with labor income, up to a certain level. The elderly worker tax credit was implemented on the first of January 2009 with a budget of 265 million euro in 2009 and 254 million euro in 2010. The CPB expected that the total labor force participation of individuals in the age 60-64 would increase by 0.6 percentage point (CPB, 2008). The effectiveness of the elderly worker tax credit will mainly depend on the dynamics in the labor market and the wage elasticity of elderly. In the following section I give an overview of the literature regarding labor supply of, and demand for, older workers. I discuss the labor supply of older workers in more detail than the labor demand for older workers, since the fiscal measure most likely affects behavior of individual workers on the supply side of the labor market.
2 Literature
2.1 Theoretic models of retirement behavior
In this section, I begin with a brief introduction into the economics behind the decision for elderly to work from a labor supply perspective. Then, I present certain relevant theoretical models of retirement behavior.
5
(1)
(2)
Where W is the value function, represents the expectations at time t, U is the utility function, the level of assets is denoted by A, ρ stands for the individual discount rate, full-time employed wage is denoted by w, and r equals the interest rate. On the right hand side of (1), the expectations of future wages allow for randomness in the model. To be able to derive solutions of the optimal time path, particular assumptions regarding the functional form of the utility function have to be made. L can only take on the value 1 or 0, signaling whether an individual works or not, which means that no distinction is made between the amount of hours worked. Once someone has retired, this cannot be reversed. With these assumptions it is possible to reconstruct the optimal leisure time path. The following equation represents the optimality condition for immediate retirement (Euwals et al., 2010):
Here, R represents the retirement age, T is the maximum age and the discount factor for age s is given by
. According to this equation, one should retire immediately when the expected utility of
retiring is larger than the expected utility gained from continuing to work plus future expected utility gained from retiring in period R. Estimating the parameters for this model is quite complex. However, the inclusion of simplified assumptions can make the estimation of this model less difficult. Rust and Phelan (1997) assume that households cannot borrow or save. As a result, consumption at any time t equals income at time t. Further simplifying equation (3) can be done by writing it in terms of an indirect utility function V instead of a direct utility function U. This transformation leads to the following equation:
6
The decision to retire depends on the expected indirect utility from continuing working versus the expected indirect utility from retiring at age R. Since individual income is more easily observed than individual savings or consumption, indirect utility functions are commonly used in practice to determine retirement age. Rust and Phelan (1997) find a strong relationship between Social Security, a federal insurance program in the USA that provides benefits to retired persons, the unemployed, and the disabled, and the peaks in the distribution of retirement ages at 62 and 65, the ages of early and normal eligibility for Social Security retirement benefits, respectively. This suggests that financial incentives play a role in the retirement decision of elderly.
Instead of optimizing lifetime utility, Stock & Wise (1990) use the option value model to choose the date for which immediate retirement is optimal. This retirement date is captured by the following equation:
This model makes three key assumptions. First, the retirement decision is re-evaluated every year by the participant as new information becomes available. Second, all future pension possibilities are considered by participants while contemplating retirement. Third, factors that influence the decision to retire, in the context of a defined contribution pension plan, are worker’s level of wealth, desire for leisure, current earnings, risk aversion with regard to income stability, personal discount factor, the price to annuitize and economic projection assumptions (MacDonald & Cairns, 2011). Stock & Wise (1990) observe that it is not simply the level of retirement wealth and the increment with one additional year of work that matters in the retirement decision. Instead, they find that the entire evolution of future wealth with further work is considered in the retirement decision.
As an alternative to the option value method from Stock & Wise (1990), where variation across individuals is triggered by the inclusion of future wages in the function, Coile & Gruber (2007) use the peak value model. Coile & Gruber (2007) suggest that the option value model largely measures the effects of income dispersion instead of the effects of changes in pension schemes. The peak value model is given by the following equation:
7
Where denotes the cash flow to or from the pension fund at age s given retirement age R. The peak value, H(t), is the difference between total discounted pension wealth at its maximum expected value and its value if retirement occurs immediately (Euwals et al., 2010).
In conclusion, this section showed the theory behind the retirement decision. On the one hand, the retirement decision can be an outcome of the lifetime optimality process, where retirement is captured in the leisure time path. On the other hand, Stock & Wise (1990) and Coile & Gruber (2007) choose the retirement date for which expected utility is maximized. In section 2.2, I describe the empirical findings of Coile & Gruber (2007), who use the peak value model to explain retirement behavior. Furthermore, I present empirical findings of studies on retirement behavior and relate their findings to the effects of the elderly working tax credit.
2.2 Empirical studies
Coile & Gruber (2007) use the peak value model to analyze the impact of Social Security on retirement in the United States of America (USA), using data of the Health and Retirement Study (HRS), which contains data on demographic and job characteristics, labor force attachment, earnings history, and the features of private pension plans for a large sample of individuals near retirement age. They find that if policy proposals minimize the wealth offset to dynamic incentive change, such as raising incentives to work only at older ages, then large effects on retirement behavior can be expected (Coile & Gruber, 2007). Following this conclusion, I expect to find a significant effect of the Dutch elderly worker tax credit on retirement behavior. Furthermore, Coile & Gruber (2007) assume heterogeneity in the responsiveness to financial incentives as a result of differences in wage rate. Higher educated workers, who are assumed to earn higher wages, are expected to react weaker to financial incentives such as the elderly worker tax credit than low educated workers (Coile & Gruber, 2007). Additionally, combined earnings makes it easier for a household with two earners to smooth consumption over time compared to single earners. Therefore, I expect a stronger incentive effect on single earners than on households with two earners.
8
different patterns of labor supply of elderly can be discovered. Blau & Goodstein’s (2010) results point to a very weak contribution of a change in Social Security rules, lifetime earnings and wages in explaining changes in labor force participation of old men, suggesting a weak impact of programs like the Dutch elderly worker tax credit on labor market participation. However, a change in the educational composition of the older male population is one of the explanations for the increase in labor force participation. Blau & Goodstein (2010) show that higher-participating college attendees replaced lower-participating high school dropouts and this replacement can account for a major part of the increase of labor force participation during these years. As I will describe in section 4.3, I find no significant change in educational composition, which means that this does not explain the increase in labor force participation. On the one hand, Blau & Goodstein (2010) are not able to prove that changes in Social Security significantly influence labor supply of older men. On the other hand, they cannot reject the hypothesis that changes in Social Security do not influence labor supply of older men.
Warren and Oguzoglu (2010) explicitly focus on the role of financial incentives in the retirement decision. They use the Australian retirement income system to build a simulation model to analyze whether individuals between 55 and 70 years of age are sensitive to financial stimuli to postpone retirement. They suggest that two competing sets of financial incentives are important in the retirement decision. On the one hand, longer participation in the labor force increases the retirement income when older workers retire. On the other hand more years of work also means fewer years of retirement. At some point, the increase in retirement income as a result of longer participation is not large enough to offset the decrease in utility because of less enjoyable retirement years. Consequently, at one point they have a financial incentive to retire (Warren and Oguzoglu, 2010). Warren & Oguzoglu (2010) test the hypothesis that the strength of the financial incentives influences the probability of elderly working. According to their model, the Australian retirement system indeed provides incentives to retire early, to which men react stronger than women (Warren & Oguzoglu, 2010). This suggests that the effect of the elderly worker tax credit might be stronger for men than for women. In contrast, Evers et al. (2008), use the elasticities of 30 empirical studies performed in different countries and find that the labor supply elasticity with respect to the wage rate of women is higher than for men, which suggests a stronger reaction of labor market participation of women as a result of the elderly worker tax credit.
9
to work longer under the older pension system, the pension benefit level was not adjusted. Therefore, there was no clear incentive for workers to continue working. This changed in the new scheme, where early retirement wealth is considerably lower. This wealth or income effect provides an incentive to postpone retirement. Furthermore, under the new pension scheme the price of leisure is not equal to zero as under the older scheme. Therefore, individuals can decide whether they prefer to have more leisure at higher costs, and retire early, or they postpone retirement and pay a lower price for leisure (Euwals et al., 2010). Eventually, their model predicts that the first part of the transition to the new pension scheme has led to an average postponement of retirement by four months, while a fully completed transition is expected to lead to an increase of retirement age by nine months on average (Euwals et al., 2010). The decreased benefits of the pension scheme leads to postponement of retirement. For the effect of the elderly worker tax credit on participation this suggests that increased after tax earnings lead to earlier retirement.
Erosa et al. (2011) use the differences in tax and transfer programs across countries to forecast whether these programs influence the labor supply decision. In their qualitative experiment they replace the tax and transfer system of the United States by the tax and transfer systems of European countries. Their life-cycle model of labor supply and retirement decisions, that builds on French (2005), shows that government policies have a considerable effect on labor supply in European countries, where changes in social security rules have a stronger effect than income taxation changes (Erosa et al., 2011). This suggests that the elderly worker tax credit influences labor supply in the Netherlands.
10
postponement of retirement. For the option value to wait this increase translates into a higher reward for waiting for retirement, especially for low wage earners (Euwals & Trevisan, 2011).
From the described empirical studies I can conclude that the evidence of the effect of financial incentives on labor supply is mixed. Some studies do not find an effect of financial incentives of labor supply at all, while others find a weak effect compared to the effects of social security changes. Finally, I expect heterogeneous responses to the elderly worker tax credit between subgroups.
2.3 Labor demand for elderly workers
Whether elderly participate in the workforce is determined by labor supply of elderly as well as labor demand for older workers. Policy implementations for extending working life have no chance of success without the active support and commitment of employers. The increased number of older workers relative to young workers changed the composition of the supply side of the labor market. Therefore, on the demand side of the labor market, employers have to be innovative to anticipate on this changed composition. Besides surveys of the thoughts of employers regarding older workers, there are objective and empirically tested theories that consider the pros and cons of hiring older workers. In this section, I discuss the theoretical background as well as empirical findings of both the advantages as well as the objections of employers to hire older workers.
11
Empirical studies acknowledge the first hypothesis that higher human capital leads to higher income. According to the American economist Thurow (1975), human capital and income are related to each other, although not as direct as the human capital theory presumes. Thurow (1975) states that this relationship does not necessarily holds over the entire cycle of a career. He argues that there is an implicit contract between employers and employees describing the relationship between productivity and income during the career of the employee. From this implicit contract follows that during the first part of his career, the employee gets rewarded below his productivity level and during the second part of his career, he gets rewarded above his productivity level (Lazear, 1999). This system of rewarding leads to an increase in income with age. This finding is in contrast with the theoretical thought that income decreases when the amount of human capital declines. Therefore, it is more expensive for employers and firms to hire old workers compared to young workers (Van Dalen et al., 2007).
Cardoso et al., (2011) find that older workers wage gains lag behind their productivity gains. They use a panel of Portuguese private sector firms over a period of 20 years and find that the age-productivity profile of the firm increases until the age interval between 50-54, while the age-wage profile remains rather flat after the age interval between 25-29. This suggests that older workers are underpaid and therefore stands in contrast with the view that older workers are not attractive for firms.
Ilmakunnas & Ilmakunnas (2011) find that, in the light of diversity at the workplace, age diversity is positively related to total factor productivity at firm level. Per 1 year increase in the standard deviation of age, total factor productivity increases with 1.5%. Additionally, one extra year of education increases total factor productivity by 6% on firm level. These findings suggest that it is important for employers to have both young and old workers, which potentially lowers the earlier mentioned obstacle on the demand side of the labor market.
12
countries that demographic developments will affect the future labor force. Especially in the Netherlands, employers expect that labor costs will rise, while only 7 percent of the Dutch employers expect a rise in productivity. Furthermore, it is remarkable to notice that there are so few training facilities for old workers, since the lack of training possibilities makes it difficult for them to increase productivity and stay valuable for employers. In the Netherlands, old workers are only rarely considered as a solution to labor market shortages. After (partly) disabled workers, women and immigrants they are considered as an option by only a few employers. In contrast, the employers expect a growing discrepancy between rising labor costs and declining productivity (Van Dalen et al., 2007).
I conclude that on the demand side, there are arguments in favor of and against the effectiveness of the elderly worker tax credit on labor market participation. There is no consensus on the question whether older workers are overpaid or underpaid. This seems to differ between countries. However, the subjective analysis under employers suggests that there are significant obstacles to hiring workers from demand side perspective. Therefore, I cannot reject the idea that the effectiveness of the elderly worker tax credit is weakened by demand side characteristics.
2.4 Conclusions
There are arguments for a positive and a negative effect of the elderly worker tax credit on retirement behavior. First, increasing net income may lead to a disincentive to continue working. For example, increasing retirement wealth may work as an incentive for early retirement. This wealth or income effect possibly also influences the labor supply response of old workers to the elderly worker tax credit. In contrast, there are also several arguments in favor of the effectiveness of the elderly worker tax credit to increase labor market participation. Opposite to the income effect, there is also a price or substitution effect that influences the retirement decision. With the increase in income, the opportunity cost of leisure increases and the worker will substitute leisure hours for work hours. Following this substitution effect, I suggest that a rise in net income as a result of the elderly worker tax credit will lead to a higher participation rate and postponement of retirement.
13
I present the research framework to find out whether the substitution effect of the elderly workers tax credit dominates the income effect.
3 Research framework
The goal of this paper is to answer the following research question:
Has the elderly worker tax credit increased labor market participation of elderly aged 61 to 65 years old in the
Netherlands in 2009 and 2010?
Furthermore, I analyze whether the effectiveness of the fiscal measure is different between subgroups, for example low versus high educated persons or (single) males versus (single) females. These differences in effectiveness of the fiscal measure can give insights in responses to fiscal measures for certain groups, which can help in the future when new fiscal measures are implemented. Therefore, I answer the following sub-question:
Is the labor market participation response to the elderly worker tax credit different for subgroups?
I consider differences in education, gender, marital status and ethnicity.
Finally, I analyze whether there were other regulations or fiscal measures around the time of the implementation of the elderly worker tax credit that could have had an effect on labor supply of 55-65 year old workers. Other regulations could bias my results on the effect of the fiscal measure on labor market participation. Accordingly, I answer the following sub-question:
Were there other fiscal measures implemented around 2009 that could have affected labor supply of elderly
workers?
4 Research design
4.1 Methodology
14
2010. When I notice a common pre-reform trend in labor market participation, I am more significantly able to explain a possible change in this trend as a result of the implementation of the elderly worker tax credit.
In line with Bettendorf et al. (2012), I estimate the effects on labor market participation as a result of the policy reform using a difference-in-differences approach. This method identifies the effect of a policy change using two types of comparisons. First, I measure the treatment effect on a treatment group before and after the policy change. However, there are other factors that might influence behavior of the treatment group. To correct for these other factors, I compare the treatment effect on the treatment group, 61-65 year old individuals, before and after the policy change with the change in participation rate of a control group, 55-58 year old individuals, before and after the policy change. I select the age category 55-58 as my control group, since I also expect an anticipation effect as a result of the elderly worker tax credit on individuals of 59 and 60 years old. Therefore, I consider the individuals of 59 and 60 years old as my second treatment group and in the robustness checks section I also present possible treatment effects for this group. The main treatment group consists of individuals eligible, or close to the eligibility age, for receiving the elderly worker tax credit and the control group ideally consists of individuals similar to the treatment group except for the fact that they are not eligible for receiving benefits from the policy reform. Additionally, I assume that an increase in labor market participation of the treatment group will not have a direct influence on labor market participation of the control group. In Figure 2, I illustrate the ideal situation for a difference-in-difference approach:
Figure 2. Difference-in-difference example
15
similar trends. After the reform the slope of the control group increases by, for example, changes in the state of the economy. Additionally, the slope of the treatment group increased in a similar way, keeping the difference between both lines constant. In the absence of a treatment effect, the participation rate of the treatment group would eventually reach point B. However, when a treatment effect occurs, the line of the treatment group will develop similar to the middle line that begins at the point where the reform is implemented by the government, and participation rate will reach point A. I now measure the actual treatment effect by the difference between participation at point A and participation at point B. Since point A lies above point B, the treatment effect on participation is positive in this example.
When determining the control group it is important that the control group and the treatment group have a common trend in participation rate before the policy reform. Furthermore, in the absence of the policy reform the trend for both the control group and the treatment group should be similar. Although I cannot test this last assumption for the period after the reform, I can test whether this assumption holds in the pre-reform period, since I have data from several years before the implementation of the reform. I find that the hypothesis of a common trend cannot be rejected. Besides these tests, I also estimate placebo treatment effects in the period before the reform as a robustness check.
In order to have unbiased estimates of the treatment effect, the change in participation as a result of the reform should be exogenous. It would be problematic when individuals anticipated on the policy reform and adapted their behavior in advance, or when government was anticipating a change in behavior when they decided to implement the new law. Then it would be unjustified to link the change in participation rate to the fiscal measure (Bettendorf et al., 2012). Eventually, when using the difference-in-differences analysis it is necessary to control for changes in composition in both the treatment group as well as the control groups. I overcome the problem of changes in composition in both groups by controlling for individual characteristics, such as ethnicity, education, gender and marital status and I compare these characteristics of both groups before and after the reform.
4.2 Model
16 A standard linear probability model can be written as:
(7)
Where Y is the dependent dummy variable. The expected value of y is then:
(8)
Accordingly, Y is a binary variable taking on the values 0 and 1. Hence:
(9)
Therefore, the probability that Y=1 is the same as the expected value of Y. This finding leads to:
(10)
This formula generalizes the interpretation of the coefficients that result from the difference-in-difference estimation. Hence, reflects how the probability of Y=1 changes when changes.
There are two main problems that can arise when using linear probability models. First, since the dependent variable is a binary variable with value 0 or 1, it would be problematic if the estimated probabilities fall outside the [0,1] interval (Hill et al. 2008). I find that all predicted values are larger than 0, while 6.5% of the predicted values are larger than 1. Accordingly, in Table A.9 in the appendix I show the outcome of a probit test, which should be interpreted with caution, since the trend test gives unsatisfying results.
Second, by definition, the linear probability model will produce heteroskedasticity in the residual variance (Hill et al. 2008); the variance of the error term varies from one observation to another. To overcome this problem, I include robust standard errors (Bertrand et al., 2004). Finally, I use sample weights to weight all regressions so the outcome of the model will be the population average effect (Cameron & Trivedi, 2005).
17
I regress participation status on year dummies ( , group dummies ( ), individual characteristics ( and interaction dummies for years after the reform .
The year dummies control for a similar trend between treatment group and control group through the years. The group dummies control for the constant difference in participation between treatment group and control group. The individual characteristics are included in the model to control for changes in group composition through the years. The interaction dummy for the first treatment group has value 1 when an individual is between 61 and 65 years old and the year variable is equal to 2009 or 2010, specified per year.
4.3 Data
I use data from the Dutch Labor Force Survey (Enquête Beroepsbevolking) of Statistics Netherlands. I have repeated cross-sections for the period 2002-2010 of this annual survey, which includes approximately 80,000 respondents per year. I select individuals in the age category of 55 to 65 years old. This is the age range around the eligibility age for the elderly worker tax credit.
18 Table 2: Relevant variables
Variable Description Value
Year20xx Survey year 2002-2010
Part2 Participation dummy =1 if hours worked>11 Part3 Participation dummy =1 if hours worked>20 Part4 Participation dummy =1 if hours worked>30
Education Level of education 0=low 1=high
Group Treatment or control 0=55-58 1=59-60 2=61-65
Couple Marital status 0=single 1=couple
Ethnicity Ethnicity 0=immigrant 1=native
Gender Dummy for gender 1=female 2=male
Note:Year20xx corresponds to year dummies for 2002-2010.
For the difference-in-differences method it is crucial that the composition of the treatment and control group do not differ significantly before and after the implementation date of the policy reform. Therefore, I show the descriptive statistics of all relevant variables of the treatment group and compare them with the relevant variables of the control group for the period before and after the reform, see Table 3 & 4.
Table 3: Descriptive statistics before and after the reform
Treatment group Control group
2002-2008 2009-2010 2002-2008 2009-2010
Variables Mean Mean Mean Mean
Part2 0.76 (0.430) 0.80 (0.400) 0.92 (0.271) 0.94 (0.240) Part3 0.55 (0.497) 0.61 (0.490) 0.78 (0.418) 0.80 (0.240) Part4 0.43 (0.495) 0.47 (0.499) 0.65 (0.478) 0.66 (0.475) Age 62.44 (1.36) 62.40 (1.29) 56.40 (1.11) 56.41 (1.11) Gender 1.35 (0.478) 1.37 (0.486) 1.39 (0.490) 1.42 (0.494) Education 0.33 (0.470) 0.36 (0.480) 0.31 (0.464) 0.34 (0.475) Marital status 0.83 (0.375) 0.82 (0.385) 0.86 (0.345) 0.85 (0.357) Ethnicity 0.89 (0.314) 0.88 (0.325) 0.89 (0.311) 0.89 (0.316) Observations 37479 16696 124806 36861
19 Table 4: Differences in individual characteristics
Differences Normalized differences 2002-2008 2009-2010 2002-2008 2009-2010 Variables Treat-Control Treat-Control Treat-Control Treat-Control
Part2 -0.16 -0.14 -0.32 -0.30 Part3 -0.22 -0.19 -0.34 -0.31 Part4 -0.22 -0.19 -0.32 -0.27 Age 6.04 5.98 3.44 3.51 Gender -0.04 -0.06 -0.06 -0.08 Education 0.02 0.02 0.02 0.02 Marital status -0.03 -0.03 -0.06 -0.06 Ethnicity 0.00 -0.01 -0.01 -0.02
Note: Bold signals that the difference is significant.
The main treatment group consists of individuals aged 61 to 65 years old, the second treatment group consists of 59-60 years old and the control group consists of individuals aged 55 to 58 years old. In the next section, I will elaborate further on the choice of the control group.
From Table 4 follows that, in general, the differences in composition between the treatment group and control group are insignificant before as well as after the reform. Logically, large differences in age do appear, since age is the variable that defines the control and treatment group. Although there are some differences in characteristics between the groups, these differences are insignificant and rather stable over time. For example, I find no change in sign of the differences before and after the reform. Therefore, I can conclude that the comparison of the composition of both groups gives no reason to assume that changes in individual characteristics will bias my estimation results.
20
Figure 3: Participation rate(>11 hours) before and after the reform
Note: Year on the horizontal axis, participation rate on the vertical axis
In Figure 3 the participation rates, for individuals working 12 hours or more, of both the first treatment and the control group are shown over time. In the appendix, I present similar figures for workers whose minimum hours of participation are higher. The solid vertical line marks the implementation date of the policy reform. From this figure, I suggest the existence of an upward trend in participation rates for both groups and a rather constant difference between treatment and control group in 2004-2008. Both figures in the appendix show a similar pattern. Furthermore, the assumption of equal trends, which is important for difference-in-difference estimates, seems to be valid. However, I give a more formal validation of this assumption in Table 5, were I show the results of the trend test. Between 2002 and 2004 I notice a different trend in participation rate between both groups. Therefore, I include trend tests for 2004-2008 and 2006-2008 since I expect a more equal trend in these years.
Table 5: Trend test participation rate (>11 hours).
Period 2002-2008 2004-2008 2006-2008
Trend treatment Group 0.008*** (0.001) 0.007*** (0.002) 0.011*** (0.004) Trend control group 0.006*** (0.001) 0.008*** (0.001) 0.011*** (0.002)
Observations 195512 160898 102617
P-Value equal trends 0.327 0.477 0.917
21
I conduct the trend test as a pooled regression on participation for the pre-reform period, including group dummies and all individual characteristics, for both treatment as well as control group. Moreover, I include a linear trend that interacts with the group dummy to obtain an estimate of the long-term trend of the specific group, while I control for the observable characteristics and constant remaining differences captured by the group dummies, following Bettendorf et al. (2012). I use the Wald test to test for joint significance of the trend coefficients. From Table 5, I conclude that the trends of both groups are quite similar and the hypothesis of equal trends cannot be rejected. Besides the pre-reform trend for the period 2002-2008, I also include the pre-reform trend of the period 2006-2008, to show that the policy change concerning early retirement around 2006, which I describe in section 5.3, does not bias my results in a direct way.
5 Results
5.1 Basic results
I present the estimation results of the basic regression model that I describe in chapter 4. I estimate three different regressions for three different minimum hours of participation borders (>11, >20 and >30). The participation dummy that I use in the basic regression has, depending on which minimum hours of participation border dummy I estimate, value 1 for individuals who work 12, 21 or 31 hours or more and 0 otherwise. In Table 6, I present the results of the separate regressions. To begin with, I estimate the effect of the reform on the participation dummy following equation (11), with separate treatment effects for the year 2009 and 2010, including year and group dummies. Additionally, I include control variables for education, gender, marital status and ethnicity. I show the results in Table 6 and 7. In Table A.1 in the appendix I present the complete results of the regression, including year and group dummies and individual characteristics.
Table 6: Basic regression results without controls, sample 2002-2010 Variable Treatment effect
2009 2010 R-Squared Part2 (>11 hours) 0.036*** (0.006) 0.046*** (0.006) 0.039 Part3 (>20 hours) 0.031*** (0.007) 0.048*** (0.007) 0.037 Part4 (>30 hours) 0.045*** (0.008) 0.043*** (0.008) 0.030 Observations 262971
22 Table 7: Basic regression results with controls, sample 2002-2010
Variable Treatment effect
2009 2010 R-Squared Part2 (>11 hours) 0.031*** (0.006) 0.039*** (0.006) 0.099 Part3 (>20 hours) 0.023*** (0.007) 0.035*** (0.007) 0.212 Part4 (>30hours) 0.034*** (0.007) 0.025*** (0.007) 0.304 Observations 262971
Note: Robust standard errors in parentheses, *** denotes significant at 1% level. Individual characteristic and year and group dummies are included but not reported.
Both estimations show positive and significant treatment effects for all participation dummies in the years after the reform. I consider the basic regression with controls as my benchmark for the rest of my paper. According to my results, the elderly worker tax credit increased the probability of participation by approximately 3 percentage points.
5.2 Robustness checks
In this section I present some tests of the robustness of my results. First, I show the treatment effects for subgroups and find that some are significantly different from the benchmark treatment effects. Second, I show that there is also a positive treatment effect for individuals, specified by age, that already work 11 hours or more. Finally, I show that there is a small anticipation effect of individuals that are near the eligibility age for the elderly worker tax credit.
23 5.2.1 Subgroups
After analyzing the treatment effects for all individuals categorized by age groups, it is interesting to specify the treatment effects for subgroups. In this section I present the results to test my sub-question whether the effect of the elderly worker tax credit was different for subgroups. I specify the treatment effects by gender, level of education, ethnicity and marital status. I present the results for subgroups to check which individual characteristics have a significant influence on the treatment effect. I first show for which subgroups the treatment and control group are correctly chosen, meaning that pre-reform trends are similar. To test the hypothesis for equal trends, I perform a Wald test on the trends for treatment and control group of subgroups with individuals working more than 11 hours. For the other minimum hours of participation borders, I check the equal pre-reform trends with a placebo test. I show the outcome of this placebo test in Table A.5 & Table A.6 of the Appendix.
Table 8: Trend test subgroups sample 2002-2008
Trend
treatment group Trend control group
P-value equal trends Observations Low educated 0.008*** (0.002) 0.006*** (0.001) 0.200 177174 High educated 0.007*** (0.002) 0.007*** (0.001) 0.923 85797 Domestic 0.007*** (0.002) 0.006*** (0.001) 0.963 234183 Immigrant 0.015*** (0.004) 0.004*** (0.001) 0.011 28788 Male 0.004*** (0.002) 0.002*** (0.000) 0.284 121110 Female 0.016*** (0.003) 0.012*** (0.001) 0.140 74402 Single 0.008** (0.004) 0.009***(0.002) 0.710 28727 Couple 0.008*** (0.002) 0.006*** (0.001) 0.240 166785 Single Male 0.001 (0.004) 0.010*** (0.002) 0.050 12381 Single Female 0.016*** (0.005) 0.008*** (0.003) 0.190 16346
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level.
Individual characteristics and year and group dummies are included but not reported. Bold coefficients signal equal pre-reform trends.
24 Table 9: Treatment effects for subgroups
Participation
>11 hours Treatment 2009 Treatment 2010 Observations Low educated 0.031*** (0.007) 0.025*** (0.007) 177174 High educated 0.032*** (0.010) 0.068*** (0.009) 85797 Domestic 0.030*** (0.006) 0.032*** (0.006) 234183 Immigrant 0.035** (0.016) 0.077*** (0.015) 28788 Male 0.029*** (0.006) 0.032*** (0.006) 161473 Female 0.042*** (0.012) 0.058*** (0.011) 101498 Single 0.032** (0.015) 0.059*** (0.014) 39521 Couple 0.030*** (0.006) 0.033*** (0.06) 223450 Single Male 0.016 (0.018) 0.025 (0.019) 17089 Single Female 0.041* (0.021) 0.079*** (0.020) 22432 Participation
>20 hours Treatment 2009 Treatment 2010 Observations
Low educated 0.019** (0.009) 0.014 (0.009) 177174 High educated 0.035*** (0.012) 0.077*** (0.012) 85797 Domestic 0.021*** (0.008) 0.031*** (0.008) 234183 Immigrant 0.031 (0.021) 0.042** (0.020) 28788 Male 0.033*** (0.009) 0.041*** (0.008) 161473 Female 0.014 (0.012) 0.037*** (0.012) 101498 Single 0.032** (0.015) 0.060*** (0.014) 39521 Couple 0.030*** (0.006) 0.034*** (0.002) 223450 Single Male 0.090*** (0.026) 0.095*** (0.027) 17089 Single Female 0.054** (0.024) 0.029 (0.024) 22432 Participation
>30 hours Treatment 2009 Treatment 2010 Observations
Low educated 0.034*** (0.009) 0.007 (0.009) 177174 High educated 0.040*** (0.012) 0.066*** (0.012) 85797 Domestic 0.033*** (0.008) 0.016** (0.008) 234183 Immigrant 0.042 (0.022) 0.078*** (0.021) 28788 Male 0.053*** (0.009) 0.033*** (0.009) 161473 Female 0.003 (0.010) 0.014 (0.010) 101498 Single 0.014 (0.018) 0.039** (0.018) 39521 Couple 0.038*** (0.008) 0.022*** (0.008) 223450 Single Male 0.030 (0.029) 0.066** (0.029) 17089 Single Female 0.006 (0.022) 0.020 (0.024) 22432
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level.
Individual characteristics and year and group dummies are included but not reported. Bold coefficients signal significant differences.
25
treatment effects between comparable subgroups. I can interpret most of the differences in treatment effects in 2010 since they are in general significantly different between comparable subgroups.
For all minimum hours of participation borders, the treatment effect in 2010 between low and high educated individuals is significantly different. From this result, I can conclude that the elderly worker tax credit has a larger effect on the probability of participation for higher educated individuals compared to lower educated individuals, while the treatment effect on high educated individuals is also significantly higher than the benchmark treatment effect. On average, lower educated individuals have physically more demanding jobs, which may cause that they are less sensitive to fiscal stimuli in their old age. Another possible explanation for the smaller treatment effect of low educated is the minimum gross wage bound to be eligible for receiving the elderly worker tax credit. Therefore, the received tax credit is smaller relative to individuals who earn a gross wage closer to the upper wage bound of the measure.
Except for individuals working more than 20 hours, the probability that immigrant workers participate in the workforce is more sensitive to fiscal stimuli than that for native workers and the treatment effect for immigrants is significantly higher than the benchmark treatment effect.
For males and females I find different treatment effects for different minimal hours of participation. For a lower minimum hours of participation border, females seem to be more sensitive to financial stimuli to participate in the working force. In contrast, for higher hours of participation borders, males seem to be more sensitive to financial stimuli in their decision to work or not.
I do not find strong participation differences between other comparable subgroups for both years. However, in 2010, singles respond stronger to the financial stimulation than couples. Furthermore, the average participation rate for females is structurally lower than for males, especially for individuals that work more. For 2008, I present this weighted average participation rate for females and males in Table 10.
Table 10: Mean participation rates
Part2(>11 hours) Part3 (>20 hours) Part4 (>30 hours)
Males 0.94 (0.230) 0.87 (0.335) 0.79 (0.405)
Females 0.80 (0.396) 0.52 (0.500) 0.28 (0.450)
Single males 0.95 (0.222) 0.86 (0.351) 0.76 (0.425)
Single females 0.85 (0.353) 0.65 (0.476) 0.44 (0.497)
Observations 262971
26
The participation rate for females is on average lower than for males, and declines relatively fast with the amount of working hours. With this Table, I can compare the relative treatment effects between males and females. For males the relative treatment effects for individuals working more than 11, 20 or 30 hours are respectively 3.4%, 4.7% and 4.2% in 2010. In contrast, for females the relative treatment effects for the three minimum hour borders of participations are 7.3%, 7.2% and 5% in 2010. According to my results, the probability of participation is more sensitive for females than for males. These results are in line with the findings of Evers et al. (2008).
By regressing for couples and singles separately, I try to address whether the primary earner is more sensitive for the tax credit than the average individual, since a single worker is by definition the primary earner. As I show in Table 9, marital status has a significant treatment effect and in 2010 this effect is significantly different between couples and singles. Unfortunately, in my dataset there is no information on whether a married individual is the primary earner or not so I cannot compare the treatment effect of the primary single earner versus the primary or secondary earner of a couple. However, by specifying the regression for single males and single females separately, including an interaction dummy for gender and marital status, I can compare the sensitivity of a female single earner versus a male single earner. To clarify, females working 12 hours or more are considerably more sensitive to financial stimuli in 2010. Additionally, males working more than 20 or 30 hours are significantly more sensitive to financial stimulate than females in 2010. To compare the treatment effects of single males and females, I show the relative treatment effects by dividing the treatment effect by the mean participation rate using Table 10. For single males, the relative treatment effects for individuals working more than 11, 20 or 30 hours are respectively, 2.6%, 11% and 8.7%, compared to 9.3%, 4.5% and 4.5% for females in 2010. To summarize, specifying the sample to singles leads to a slightly different pattern in sensitivity for financial stimuli between males and females for higher amount of working hours, where single males are more sensitive to financial stimuli to work than females.
27
larger than the substitution effect for females working more than 20 or 30 hours. For a single man that works more than 20 or 30 hours, I find strong positive and, over time, increasing coefficients of the treatment effect.
5.2.2 Age specific treatment effects
Until now, I have used the difference-in-differences method to measure differences in treatment effect between a treatment and control group in 2009 and 2010. However, now I analyze the treatment effect from another perspective, namely the development of the participation rate related to age between different years. Figure 4 shows this relationship between participation rate and age for several years.
Figure 4: Participation rate (>11) per age for different years
28
worker tax credit is larger for individuals of 62 or 63 years old compared to individuals of 64 years old. I replace the interaction dummy for an extra variable in my basic regression that represents the tax benefits from the reform. Following Table 1, the variable has value 5, 7 and 10 for respectively 62, 63 and 64 year old individuals in 2009 and 2010 that already work more than 11 hours. I included the condition of working 11 hours or more in the variable since an individual has to earn at least 8860 euro on a yearly basis to be eligible for the elderly worker tax credit.
Table 11: Effect of the benefits per age category
Part2(>11 hours) Part3 (>20 hours) Part4 (>30 hours) Benefit 62 0.055*** (0.001) 0.043*** (0.002) 0.031*** (0.002) Benefit 63 0.039*** (0.001) 0.023*** (0.002) 0.015*** (0.002) Benefit 64 0.026*** (0.001) 0.011*** (0.001) 0.005*** (0.001)
Observations 262971
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level. Individual characteristics and year dummies are included but not reported.
The coefficients in Table 11 explain the effect of the benefit on the probability of labor market participation. The effect of the benefit on the probability of participation decreases by age and by the amount of working hours. This pattern is remarkable, since the benefits from the elderly worker tax credit increases with age until the age 64. From this Table, I conclude that the effect of financial stimuli on the probability of participation decreases with age. In Table A.7 & A.8 in the appendix I present the results of an alternative test to estimate the age specific effect of the elderly worker tax credit.
5.2.3 Anticipation effect
I analyze the effect of the elderly worker tax credit on the probability of labor market participation for individuals of 61 to 65 years old. However, I also notice a possible treatment effect for individuals of 59 and 60 years old, although they are not eligible for the elderly worker tax credit. I conduct a similar trend test as I do in section 4.3 and I cannot reject the hypothesis of equal trends (Table 12). To estimate the anticipation effect I include an extra interaction dummy and transform equation (11) into the following equation:
29
I show the coefficients of the interaction dummies for the second treatment group in Table 13.
Table 12: Trend test second treatment group (>11 hours).
Period 2002-2008 2004-2008 2006-2008
Trend second treatment group 0.004*** (0.001) 0.006*** (0.001) 0.008** (0.003) Trend control group 0.006*** (0.001) 0.008*** (0.001) 0.011*** (0.002)
Observations 195512 160898 102617
P-Value equal trends 0.134 0.280 0.382
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level.. Individual characteristics and group dummies are included but not reported.
Table 13: Treatment effects of 59 and 60 years old, sample 2002-2010 Variable Treatment effect
2009 2010 Part2 (>11 hours) 0.007 (0.005) 0.019*** (0.005) Part3 (>20 hours) 0.003 (0.007) 0.022*** (0.007) Part4 (>30 hours) 0.008 (0.007) 0.020*** (0.007) Observations 262971
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level. Individual characteristics and year and group dummies are included but not reported
The positive and significant treatment effect for 59 and 60 year old individuals in 2010 signals that there is a possible anticipation effect. A possible explanation for this is that individuals already adjust their labor market participation knowing that they will receive a tax benefit in the future.
5.3 Other fiscal measures
According to Euwals et al. (2009), there was one other major policy reform, “de wet aanpassing fiscal behandeling VUT/prepensioen en introductie levensloopregeling” (law VPL), between 2002 and 2010 that had an effect on the labor market participation rate of elderly in that period.
30
schemes into less generous actuarially fair schemes (Euwals et al., 2009). Initially, the VUT early retirement schemes were implemented to allow firms to lay off older workers to save jobs for the young when the Netherlands went through a severe economic crisis in the 1980s. However, the VUT early retirement scheme became too costly and therefore the Dutch government changed the early retirement conditions in 1997.
Euwals et al. (2010) show that retirement is postponed by the transition of the retirement schemes starting in 1997. By implementing the law VPL in 2006, the Dutch government abolished fiscal measures that facilitated retirement before the age of 65, which accelerated the transition of early retirement schemes. According to Euwals et al. (2009), the abolishment of the actuarially unfair early retirement scheme (VUT) increased the participation rate of 60-64 year olds by 25%-points. However, the Dutch government decided that individuals aged 55 years and older at the 1st of January 2005 kept their fiscal facilitations to retire early. At the time the elderly worker tax credit was implemented, January 2009, the minimum age for eligibility of the actuarially unfair retirement scheme was 59 years. Therefore, both of my treatment groups, 59-60 and 61-65 year olds, are not directly affected by the law VPL, since they remain able to use the fiscal facilitations to retire early. In contrast, I expect that my control group, 55-58 year olds, is influenced by the law VPL. Older workers most likely postpone their retirement age due to the abolishment of the implicit tax on working at higher ages. Although the retirement age of my control group is not directly relevant in my analysis, there could be an anticipation effect of my control group to increase participation due to the decreased incentives to retire early. Therefore, the law VPL could have caused an increase of participation rate of individuals in my control group. Despite this possible increase of participation rate, I still find a significant relevant increase in participation rate between my treatment and control groups.
31
law VPL on participation rate of the treatment group between 2006 and 2008, then the placebo treatment coefficients should be significant, which is not the case.
6 Conclusion
The current ageing of the workforce in the Netherlands and the resulting extra costs of the pension system increase the relevance for the Dutch government to increase the labor market participation of elderly. In 2009, the Dutch government implemented the elderly worker tax credit to provide an incentive for elderly to postpone retirement. I estimate a linear probability model to analyze whether the fiscal measure increased the labor force participation of elderly. Furthermore, I estimate whether the reaction to the fiscal measure was different for subgroups. Finally, I analyze whether there is an anticipation effect for age groups near the eligibility age of the fiscal measure. I use data from the Labor Force Survey of Statistics Netherlands for the period 2002-2010.
I find that the elderly worker tax credit increased the labor market participation of elderly from 61 to 65 years by approximately 3 percentage points. Furthermore, the effects of the elderly worker tax credit are significantly different for subgroups. Especially the probability of participation of high educated shows a considerable 7 percentage point increase in 2010 as a results of the elderly worker tax credit. Finally, I find a small anticipation effect for age groups near the eligibility age of the elderly worker tax credit in 2010. I show that an investment of 265 million in 2009 and 254 million in 2010 considerably increased participation of elderly on the labor market. My findings are interesting for public policy makers. Especially the identification of the treatment effects for the different groups is helpful for specific policy making.
32
33
7 References
Bertrand, M., E. Duflo & S. Mullainathan (2004). “How much should we trust differences-in-differences estimates?”. The Quarterly Journal of Economics, MIT Press, vol. 119(1), pp. 249-275.
Bettendorf, L., E. Jongen & P. Muller (2012). “Childcare subsidies and labour supply: Evidence from a large Dutch reform”. CPB Netherlands Bureau for Economic Policy Analysis Discussion paper 217.
Blau, D. & R. Goodstein (2010). “Can social security explain trends in labor force participation of older men in the United States”. The journal of Human Resources, 45(2), pp. 328-363..
Brown, C. (2006). “The role of conventional retirement age in retirement decisions”. Working Paper WP 2006-120, Michigan Retirement Research Center.
Cameron, A. & P. Trivedi (2005). “Microeconometrics: methods and applications”. Cambridge University Press.
Cardoso, A., P. Guimaraes & J. Varejao (2011). “Are older workers worthy of their pay? An empirical investigation of age-productivity and age-wage nexuses”. De Economist, 159(2), pp. 95-111.
Centraal Bureau voor de Statistiek (1987-01-01), Enquête Beroepsbevolking - EBB - jaargangen 2002 t/m 2010 Persistent Identifier: urn:nbn:nl:ui:13-sk6-fmg.
Coile, C. & J. Gruber (2007). “Future social security entitlements and the retirement decision”. The review of Economics and Statistics, 89(2) pp. 234-246.
CPB notitie (2008). “Effecten van pensioenbeleid”
34
Erosa, A., L. Fuster & G. Kambourov (2012). “Labor supply and government programs: A cross-country analysis”. Journal of Monetary Economics 59, pp. 84-107.
Euwals, R. & E. Trevisan (2011). “Early retirement and Financial Incentives: Differences between High and Low Wage Earners”. Netspar discussion paper No. 12/2011-105.
Euwals, R., D. van Vuuren, & R. Wolthoff (2010). “Early retirement behaviour in the Netherlands”. De Economist, Vol. 158, pp. 209-236.
Euwals, R. & C. Folmer (2009). “Arbeidsaanbod en gewerkte uren tot 2040: een beleidsvrij scenario”. CPB Netherlands Bureau for Economic Policy Analysis Memorandum 225.
Euwals, R., R. de Mooij & D. van Vuuren (2009). "Rethinking retirement; from participation towards allocation”. CPB Special Publication 80, CPB Netherlands Bureau for Economic Policy Analysis.
Evers, M., R. De Mooij & D. Van Vuuren (2008). “The wage elasticity of labour supply: A synthesis of empirical estimates”. De Economist Vol. 156(1) pp. 25-43.
French, E., (2005). “The effects of health, wealth and wages on labour supply and retirement behavior”. Review of Economic Studies, 72(2), pp. 395-427.
Garssen, J. & C. van Duin (2007). “Grijze druk zal verdubbelen”. Sociaaleconomische trends, 2e kwartaal 2007, pp. 23–26.
Hill R., W.E. Griffiths & G.C. Lim (2008). “Principles of econometrics”. Third edition pp. 197-225.
Ilmakunnas, P. & S. Ilmakunnas (2011). “Diversity at the workplace: Whom does it benefit”? De Economist, 159, pp. 223-255.
35
program evaluation”. Journal of Economic Literature, 47(1), pp. 5-86.
Lazear, E.P. (1999). “Personnel economics: past lessons and future directions”. NBER Working paper No. 6957.
MacDonald, B. & A. Cairns (2011). “Three retirement decision models for defined contribution pension plan members: A simulation study”. Elsevier, 48 (1), pp. 1-18.
Rust, J. & C. Phelan (1997). "How Social Security and Medicare Affect Retirement Behavior in a World of Incomplete Markets". Econometrica, 65, pp. 781-831.
Stock, J. H. & D. A. Wise (1990). “Pensions, the option value of Work, and Retirement”. Econometrica, 58(5), pp. 1151-1180.
Thurow, L.C. (1975). “Generating inequality: mechanisms of distribution in the USA”. New York: Basic Books.
Van Dalen, H., K. Henkens & J. Schippers (2007). “Oudere werknemers door de lens van de werkgever”. Nederlands interdisciplinair demografisch instituut, rapport no. 74.
Van Dalen, H., K. Henkens & J. Schippers (2009). “Dealing with older workers in Europe: a comparative survey of employers’ attitudes and actions”. Journal of European Social Policy, (19), pp. 47.
36
8 Appendix
A.1 Complementary figures and tables
Figure A.1: Participation rate(>20 hours) before and after the reform
Note: Year on the horizontal axis, participation rate on the vertical axis.
Figure A.2: Participation rate(>30 hours) before and after the reform
37 Table A.1: Full output Linear Probability Model
Variable Part2 Part3 Part4
Constant 0.828*** (0.005) 0.575*** (0.006) 0.381*** (0.006) Year2003 0.009* (0.005) 0.011* (0.006) 0.005 (0.006) Year2004 0.012*** (0.004) 0.020*** (0.005) 0.015*** (0.005) Year2005 0.009* (0.004) 0.021*** (0.005) 0.017*** (0.005) Year2006 0.017*** (0.004) 0.023*** (0.005) 0.008 (0.005) Year2007 0.028*** (0.004) 0.039*** (0.005) 0.012** (0.005) Year2008 0.038*** (0.004) 0.049*** (0.005) 0.025*** (0.005) Year2009 0.045*** (0.004) 0.064*** (0.005) 0.042*** (0.005) Year2010 0.050*** (0.004) 0.067*** (0.005) 0.044*** (0.005) Group1 -0.047*** (0.002) -0.070*** (0.002) -0.075*** (0.002) Group2 -0.182*** (0.003) -0.250*** (0.003) -0.252*** (0.003) Education 0.031*** (0.001) 0.067*** (0.002) 0.034*** (0.002) Ethinicity -0.022*** (0.002) -0.042*** (0.003) -0.041*** (0.003) Couple -0.028*** (0.002) -0.065*** (0.003) -0.066*** (0.003) Gender 0.162*** (0.002) 0.381*** (0.002) 0.532*** (0.002) Treatment09 0.031*** (0.006) 0.023*** (0.007) 0.034*** (0.007) Treatment10 0.039*** (0.006) 0.035*** (0.007) 0.025*** (0.007) Observations 262971 262971 262971 R-squared 0.100 0.212 0.304
Note:Standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level, * denotes significant at 10% level.
A.2 Robustness checks
Smaller sample
38 Table A.2: Basic regression results, sample 2004-2010
Variable Treatment effect
2009 2010 Part2 (>11 hours) 0.028*** (0.006) 0.036*** (0.006) Part3 (>20 hours) 0.019** (0.007) 0.031*** (0.007) Part4 (>30 hours) 0.028*** (0.007) 0.018** (0.007) Observations 262971
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level. Individual characteristics and year and group dummies are included but not reported.
The treatment effects for the smaller sample are somewhat lower than the basic sample, however they are not significantly different.
Placebo
39 Table A.3: Placebo treatment effects
Variable Part2(>11 hours) Part3(>20 hours) Part4(>30 hours)
Placebo 2006-2008 0.002 (0.006) 0.010 (0.006) 0.014* (0.006)
Treatment 2009 0.032*** (0.007) 0.028*** (0.008) 0.041*** (0.008)
Treatment 2010 0.040*** (0.007) 0.040*** (0.008) 0.032*** (0.008)
Observations 262971
Note: Robust standard errors in parentheses, *** denotes significant at 1% level, ** denotes significant at 5% level. Individual characteristics and year and group dummies are included but not reported.
From this table, I can conclude that for individuals working 12 hours or more or 21 hours or more I do not find a significant effect for the placebo years, which is in line with my expectation that the elderly worker tax credit has influenced participation rate of individuals from 2009 onwards. Moreover, the choice of my control group is partly justified by the results of the placebo tests in Table A.3, since significant coefficients for the placebo treatment dummies would possibly signal a difference in autonomous trend between treatment and control group. Finally, I test whether the reported placebo effects are significantly different from the actual treatment effects. In Table A.4 I present the p-values of the Wald test whether the coefficients for placebo effects are equal to the coefficients for treatment effects. A p-value below 5% signals that the placebo and treatment effects are significantly different.
Table A.4: P-values equal coefficients test
Part2 Part3 Part4
Placebo = Treatment 2009 0.0% 1.9% 0.0%
Placebo = Treatment 2010 0.0% 0.0% 1.8%
Observations 262971
40 Placebo test subgroups
With the placebo test I want to confirm that I use a proper control group to estimate the treatment effects per subgroup. I estimate my regular linear probability model for subgroups and include a placebo treatment dummy that has value 1 for individuals in the treatment group between 2006 and 2008. I show the estimation results in Table A.6 & A.7. For most of the subgroups the placebo effect is insignificant and significantly different from the treatment effect. However, I have some remarks.
