Combating child poverty through work strategies

COMBATING CHILD POVERTY

THROUGH WORK STRATEGIES

An evaluation of the effectiveness of parental employment

stimulation policies in the Netherlands

Jeroen van Veldhoven

s1149571

11-1-2019

Public Administration: Economics and Governance Supervisor: Dr. E. Suari-Andreu

Second reader: Dr. M. van Lent 21,936 words

1 CONTENTS 1. INTRODUCTION ... 2 2. THEORETICAL FRAMEWORK ... 4 2.1CONCEPTUALISATION ... 4 2.2DETERMINANTS ... 5 2.2.1 Macro-factors ... 5 2.2.2 Micro-factors ... 5 2.3PUBLIC POLICIES ... 6

2.3.1 Broad family programmes ... 7

2.3.2 Work strategies ... 8

2.3.3 Benefit strategies ... 10

3. DUTCH POLICY CONTEXT ... 11

3.1BROAD FAMILY PROGRAMMES ... 11

3.2WORK STRATEGIES ... 12 3.3BENEFIT STRATEGIES ... 14 4. METHOD ... 14 4.1RESEARCH DESIGN ... 14 4.2STATISTICAL MODEL ... 17 5. DATA ... 18

5.1HOUSEHOLD INCOME AND POVERTY DATA ... 19

5.2TREATMENT AND CONTROL GROUP ... 20

5.3USE OF THE CHILDCARE SUBSIDY ... 23

5.4MULTICOLLINEARITY AND HETEROSKEDASTICITY ... 25

6. RESULTS ... 25 6.1MAIN RESULTS ... 25 6.2ROBUSTNESS CHECKS ... 33 6.3INTERACTION EFFECTS ... 34 7. CONCLUSION ... 38 8. REFERENCES ... 40 9. APPENDIX ... 44

2

1. Introduction

Numerous scholars have conducted research on the effects of child poverty. Among other things, these studies have demonstrated that deprivation during childhood has a negative impact on health, life expectancy, education level, labour market position, financial situation, social relationships and self-image, while it also increases the likelihood of criminal behaviour and teenage pregnancy (Griggs & Walker, 2008, pp. 7–8; Vleminckx & Smeeding, 2001, p. 1; Pisu, 2012, p. 28; UNICEF, 2000, pp. 3, 12; Gregg & Machin, 2001, p. 130; Bitler et al., 2014, p. 2; Heckman & Masterov, 2007, p. 2; Stella Hoff, 2017, pp. 15–16). Additionally, this research has pointed out that child poverty has negative effects in the short-, medium-, long- and even intergenerational term, at the level of the individual, household, neighbourhood and society. Due to the enormous impact of child poverty, the economist James Heckman famously argued in favour of more public spending on young children as the productivity returns of such investments are so high that there is no ‘equity-efficiency trade-off’ (Heckman & Masterov, 2007, p. 2). In order to fully capitalise on the potential of such expenditures, it is necessary to know the impacts of different government programmes on child poverty (Van Lancker & Van Mechelen, 2015, p. 60). Despite suggestions that work strategies are the most adequate way to deal with child poverty (Pisu, 2012, pp. 16, 28; Whiteford & Adema, 2007, p. 7; Haskins & Sawhill, 2003, p. 3), the effectiveness of concrete parental employment stimulation policies in combating deprivation during childhood is still far from established as the academic field lacks causal studies seeking to estimate the effects of these measures on child poverty.

This study aims to contribute to filling the knowledge gap concerning the impact of work strategies on child poverty by investigating the following research question: how effective is parental employment stimulation in reducing child poverty? For this enquiry, child poverty is conceptualised in absolute as well as relative terms and calculated with the use of net disposable household income. The main analysis builds upon the method and findings of an earlier study by Bettendorf et al. (2015) who analysed the influence of work strategy reforms in the Netherlands on (female) labour supply and found statistical evidence for a small, positive effect. Similar to that study, this thesis applies a difference-in-difference analysis to estimate the common intention-to-treat effect of parental employment stimulation reforms in 2005, which introduced a demand-side childcare subsidy and significantly increased the earned income tax credit targeted at single parents and secondary earners in families with children (the Additional combination credit), on a dependent variable. The treatment and control group are also alike as households where the youngest child is below twelve years of age are selected as the treatment group, whereas those families in which the age of the youngest child is twelve to

3 seventeen years constitute the control group.

Differently, however, this study analyses the effect of the work strategy reforms on child poverty, which largely works through the impact on parental labour supply, rather than looking at the influence on parental work itself. Thereby, the analysis contributes to a different academic debate, while also complementing the findings of Bettendorf et al. concerning the impact of the policy reforms. Current studies into government programmes for child poverty reduction have mostly focused on cross-national analysis, using either macro- or micro-data, which makes it more difficult to estimate the exact effectiveness of different policies as they have to account for much divergence between countries and their policy regimes. This study, therefore, uses the DNB Household Survey that provides yearly panel micro-data for the Netherlands to enable more accurate estimations of the policy reforms’ effects. The decision in favour of the precision of national micro-data necessarily comes at the cost of external validity of the results. The issue of representativeness is limited, however, as the studied work strategy reforms in the Netherlands are instances of the most widely implemented parental employment stimulation policies. The findings of this study of Dutch policies are, therefore, still relevant for a broader group of highly economically developed countries with similar policies.

Noteworthy, the data used for the study is also different from Bettendorf et al. as it has a panel rather than a repeated cross-sectional structure. This panel data enables a rough investigation of the characteristics of the households that actually make use of the availability of the Childcare subsidy between 2007 and 2009. These findings are, then, used to study the possibility of interaction effects related to these attributes in the last section of the analysis. Furthermore, the results of the main analysis are tested for robustness by making adjustments to the bandwidth of the data, the operationalisation of the time dummies in the difference-in-difference analyses and by correcting for movement between the treatment and control group. The main findings of this thesis are that there are no structural, statistically significant short- or medium-term effects of the parental employment stimulation reforms on absolute and relative child poverty between 2005 and 2009 (1), but there is an isolated, statistically significant impact of the policy changes on absolute and relative child poverty in 2006 (2). Additionally, the short-term effects of the work strategy reforms regarding child poverty turn out to be particularly beneficial to couples with children compared with single-parent households (3). Even when considering the last findings, however, the general picture of the estimated effects of the government programmes on child poverty is that these are relatively insignificant, especially when considering the size of the reforms. These results are remarkable

4

as the current academic debate predominantly suggests that parental employment stimulation is the most effective way to reduce child poverty.

2. Theoretical framework

2.1 Conceptualisation

To be able to answer the research question, the concept of child poverty needs more clarification first. As with the broader phenomenon of poverty, child poverty has many dimensions and definitions (Chen & Corak, 2008, p. 538; UNICEF, 2000, p. 6; Barr, 2012, pp. 108–109). This study conceptualises poverty as a monetary, income-based phenomenon. The major benefit of this approach is that it allows for large-N comparison and consistent statistical estimates which is necessary to analyse the effectiveness of policy measures.

The next step in defining poverty is to choose between an absolute and relative definition of the term. Absolute poverty occurs when incomes are lower than a threshold based on the capability of households to buy the necessities of living within a country (Pisu, 2012, pp. 5, 9; UNICEF, 2000, p. 6, 2007, p. 5; Bradbury & Jäntti, 2001a, p. 13; Barr, 2012, p. 109). Such a conceptualisation of child poverty does not take the rest of society into account, however, which leads to an underestimation of the problem of child poverty in economically developed countries, mainly by ignoring the role of social exclusion (Mood & Jonsson, 2016, p. 828; UNICEF, 2000, p. 6; Bradbury & Jäntti, 2001a, p. 13; Barr, 2012, p. 109). Once an economy has reached a stage in its development where it has passed a minimum standard of living, many scholars would, therefore, argue that it makes more sense to use a relative income standard (UNICEF, 2000, p. 9). At the same time, other authors point out that poverty is not a neutral concept and relative poverty may be more an indicator of income inequality than poverty as such.

This study uses both an absolute and relative poverty standard to make the findings relevant to both sides of this conceptual discussion and enable comparison between them. On the one hand, the absolute poverty standard is based upon the method of the ‘low-income threshold’ of Statistics Netherlands which uses a fixed level of purchasing power (Centraal Bureau voor de Statistiek, 2015, p. 16). On the other hand, the relative poverty indicator is computed using the approach of Peter Whiteford and Willem Adema determining the boundary at fifty per cent of median equivalised disposable household income (Whiteford & Adema, 2007, p. 10).

5 2.2 Determinants

In the academic debate about child poverty, most research has studied the extent to which different factors can account for the variation in child poverty among households and between countries without using a causal approach (Bradbury et al., 2017, pp. 3–4). These studies led to the identification of correlations between child poverty and several factors on the macro- as well as the micro-level which are now undisputed.

2.2.1 Macro-factors

On the macro-level, the research has shown that the macroconditions of a country at a given moment, such as the level of economic growth, unemployment, social protection, taxes, wages, income inequality and the size as well as structure of the population, are of major importance in explaining the extensity and intensity of child poverty (Bitler et al., 2014, p. 16; Chzhen, 2014, pp. 22–23; Bradbury et al., pp. 4, 7; Chen & Corak, 2008, p. 544; UNICEF, 2000, p. 17; Bradshaw & Finch, 2002, pp. 22–23). These macro-factors are overarching determinants, among other things shaping the parental (employment) opportunities of the time and the possibilities for income transfers by influencing the taxable income. As this study analyses policy effects at the household level, however, the focus is on micro-factors rather than these macro-determinants.

2.2.2 Micro-factors

Although it is difficult to distinguish different micro-factors clearly since they are overlapping and interrelated, there is wide consensus that certain demographic characteristics on the household level are of vital importance when explaining child poverty. A large share of the theoretical debate on child poverty has been focusing on a descriptive analysis of these demographic elements. Janet Gornick and Markus Jäntti (2011), for instance, use regressions on data from the Luxembourg Income Study, which integrates micro-data across countries in a single dataset, to estimate cross-country correlations between deprivation during childhood and these factors.

Because of such studies, it is now generally accepted wisdom that parental education (Chen & Corak, 2008, p. 541; Gornick & Jäntti, 2012, p. 564; Haskins & Sawhill, 2003, p. 3; Stella Hoff, 2017, p. 15; Tárki Social Research Institute, 2010, p. 37) and parental employment (Bitler et al., 2014, p. 2; Chen & Corak, 2008, p. 542; Gornick & Jäntti, 2012, p. 564; Rector & Hederman, 2003, p. 6; Tárki Social Research Institute, 2010, pp. 12, 48; Whiteford & Adema, 2007, p. 19) have a substantial, negative correlation with child poverty, whereas deprivation

6

during childhood tends to be more eminent in families with a higher number of children (Arcanjo et al., 2013, p. 14; Chen & Corak, 2008, p. 541; Haskins & Sawhill, 2003, p. 3; Tárki Social Research Institute, 2010, pp. 11, 41) and single-parent households (Arcanjo et al., 2013, p. 14; Bradbury & Jäntti, 1999, p. 28, 2001b, p. 79; Bradbury et al., 2017, pp. 3–4; Bradshaw & Finch, 2002, p. 28; Chen & Corak, 2008, p. 541; Gornick & Jäntti, 2011, p. 15; Immervol et al., 2001, p. 24; Tárki Social Research Institute, 2010, pp. 11, 38–39; UNICEF, 2000, p. 13; Whiteford & Adema, 2007, p. 19). Among single-parent households, deprivation during childhood is even more likely in families consisting of single mothers than those with single fathers, although the probability of child poverty for households with lone fathers is also relatively high (Chen & Corak, 2008, p. 542; Christopher et al., 2001, pp. 207, 209; Gornick & Jäntti, 2011, pp. 15–16).

2.3 Public policies

While these micro-determinants are relatively well-equipped to explain variation in static child poverty outcomes, they are not suitable for accounting for changes in deprivation during childhood over time, however (Bradshaw, 2006, p. 8).

Wen-Hao Chen and Miles Corak (2008, p. 537), in particular, have pointed this out by analysing changes in poverty rates across twelve OECD countries during the 1990s, ensuing their commitments to stimulating children’s well-being through the Convention of the Rights of the Child. After conducting a decomposition analysis using data from the Luxembourg Income Study, they find that parental age, education, the number of children and single parenthood can only explain a small part of the development in child poverty between countries as these factors change too gradually to account for sudden shocks in deprivation during childhood (Chen & Corak, 2008, pp. 551–552). In the countries that experienced a decline in poverty rates, most of the variation was either accounted for by labour market developments on the macro-level, in the form of employment opportunities and wage levels, or by government transfers (Chen & Corak, 2008, pp. 545–546).

Janet Gornick and Markus Jäntti (2012, p. 565) further underlined the importance of public policies when they conducted a quantile multivariate regression model to analyse the importance of government interventions in explaining cross-national child poverty variation. After estimating the relevant correlations for the child poverty determinants across countries in all percentiles with these regressions, they used the United States as a reference case with relatively low social transfers to see how the child poverty among households would change when they shifted to the American (policy) context. Just like Chen and Corak, these authors

7

find that demographic differences, such as household structure and parental background, can only account for a small share of the cross-national variation in poverty rates (Gornick & Jäntti, 2012, p. 566). Rather the change of the policy framework is of major importance as a change to the American system would have a huge impact on the child poverty levels of these families. 2.3.1 Broad family programmes

Due to the increasingly apparent societal effects of deprivation during childhood and relevance of public policies in determining dynamic child poverty outcomes, the focus of the recent academic debate has shifted to the effectiveness of different government programmes as measures against the phenomenon. Analysing the overall impact of public policies on households with children is relatively complex, however. The main reason for this intricacy is the wide variety of policies that affect families. Following its broad definition, family policy concerns all those government programmes that affect the socioeconomic position of households with children, without the requirement of targeting (Whiteford & Adema, 2007, p. 21; Bitler et al., 2014, p. 6).

Commissioned by the European Commission, the social research institute Tárki analysed the effect of family policies on child poverty in EU-countries in the broad sense. Using the EUROMOD micro-simulation approach in combination with EU-SILC micro-data, the research institute estimated the impact of a broad range of social transfers, such as old-age, sickness, housing and unemployment benefits. They were unable to fully account for the influence of the tax income side of the fiscal policy system (e.g. the role of earned income tax credits), however (Tárki Social Research Institute, 2010, p. 75). The resulting estimates showed that child-contingent transfers generally had a slightly larger impact than non-child-contingent policies, despite a lot of variation between the European countries (Tárki Social Research Institute, 2010, pp. 78–79).

Although such a comprehensive approach can shed new light on the debate, the approach generally comes at the cost of the precision of the estimated policy effects due to the limited resources of research projects. Like most of the literature, this study, therefore, focuses on the narrow definition of family policy which only considers those socioeconomic policies directly targeted at families with children. These public policies generally follow an indirect work strategy or a direct benefit approach.

8 2.3.2 Work strategies

Currently, most studies suggest that parental employment stimulation is the most effective way to reduce child poverty (Pisu, 2012, pp. 16, 28; Whiteford & Adema, 2007, p. 7; Haskins & Sawhill, 2003, p. 3). This strategy of combating child poverty through parental work is indirect as it aims to stimulate parental employment as a mediating variable, which, in turn, has to reduce child poverty. The interest in this approach largely builds upon the cross-national finding that market incomes are better able to explain variation between countries in low family incomes than social transfers (Bradbury et al., 2017, p. 4; Pisu, 2012, p. 28; Tárki Social Research Institute, 2010, p. 43; UNICEF, 2000, p. 13; Whiteford & Adema, 2007, p. 29).

Whiteford and Adema (2007, p. 29), for instance, used OECD Income Distribution Study data to simulate how child poverty rates would change under different employment assumptions. Concretely, they analysed how household incomes would transform if countries had the same joblessness rates as Portugal, the third-best performing OECD country at the time, and the same share of couples with two wage earners as Denmark, also the third-best performing in this regard. Their estimates showed that both changes would have had a reasonable effect on child poverty depending on the rank of the analysed country in the order of joblessness rates and share of couples with two wage earners. The analysis builds upon a ceteris paribus assumption which seems problematic, however, due to the complex indirect effect of the parental employment stimulation measure (Whiteford & Adema, 2007, pp. 29–30).

When authors move past the simplistic market income and social transfer dichotomy by trying to provide estimates of the effectiveness of already implemented work strategies, research becomes more complicated, however. The first reason is that there are many different work strategies. Still, the child-contingent employment stimulation transfers generally come down to two major government programmes: targeted earned income tax credits (EITCs) and childcare subsidies. Firstly, EITCs decrease the amount of income tax that is subtracted from market income for certain target groups such as (single) parents (Bettendorf et al., 2014, p. 49; Barr, 2012, p. 201). The logic behind this policy is that the tax deduction for parents of children increases market income after taxes which makes it more attractive for parents to work (more). Secondly, childcare subsidies aim to stimulate parental employment by reducing the costs of outsourcing the care for children (Berden & Kok, 2009, pp. 5, 38).

The second reason is that the effects of these parental employment stimulation measures have proven to be more complex in practice which can be illustrated with the research on changes in EITCs and childcare subsidies in the Dutch policy context (Blommesteijn, 2014, p.

9

17). Firstly, specific empirical studies on the (enlargement of the) EITC schemes in the Dutch policy context using regression discontinuity and difference-in-difference designs have not been able to find significant labour supply effects (Bettendorf et al., 2014, p. 57; Jansen, 2015, p. 17), while Bettendorf et al. (2015, pp. 117–118) found a small common effect of the introduction of a demand-side childcare subsidy and expansion of the EITC schemes between 2004 and 2009. Secondly, empirical research on the effectiveness of the introduction of the demand-side Childcare subsidy in 2005 has found mixed results on (female) labour supply (Roeters & Bucx, 2018, pp. 51–52). Whereas most of the studies indicate that the childcare subsidy was one of the more effective measures for stimulating labour supply (Bettendorf et al., 2015, pp. 117–118; Centraal Planbureau, 2016, pp. 60–61; Jongen et al., 2014, p. 27), empirical analysis of the cutbacks of the same childcare policy by the government does not show that it led to a restriction of labour supply despite a considerable fall in demand (Blommesteijn, 2014, p. 17; Sociaal-Economische Raad, 2016, p. 126). For the same cutbacks, however, government surveys in the Netherlands indicate that mothers reduced their labour supply after the cutbacks (Portegijs et al., pp. 121–122).

In the case of the childcare subsidy, several factors may explain the mixed results of these studies. Firstly, subsidising parents that already used the same amount of childcare and pay less for this service after the introduction of the policy, makes the instrument less effective regarding labour supply as people stick to the same working behaviour (Berden & Kok, 2009, p. 6). Secondly, households with more resources tend to use childcare more often but are relatively less affected by the working incentives. Thirdly, the benefit may move parents to replace informal childcare with formal childcare as formal childcare has become relatively cheaper (Centraal Planbureau, 2016, pp. 60–61; Berden & Kok, 2009, p. 7). Interestingly, this factor plays no role in the case of other working strategies such as EITCs. Fourthly, parents may spend the additional free time in other ways than for more work (Berden & Kok, 2009, p. 7). Fifthly, in a very indirect way, the increase in taxes required for childcare subsidies may also provide a disincentive to labour participation when these taxes are raised from market incomes (Berden & Kok, 2009, p. 7).

As the work strategy aims to reduce child poverty through parental employment, the studies into the effects of these measures on (female) labour supply are also relevant for the impact of childcare subsidies on child poverty. The effects of parental employment stimulation measures on (female) labour supply are not exactly similar to those on child poverty, however, despite the mediatory role of parental work. With the introduction of the Childcare subsidy in the Netherlands, for example, the small labour supply effects that were found by Bettendorf et

10

al. (2015) may also have been caused by the fact that general participation levels in the Netherlands were already quite high in the period before the reform (Akgündüz & Plantenga, 2015, p. 16). In this case, the policy might have had a larger effect on child poverty than on parental employment, as the instrument could have significantly reduced the costs of childcare for households while only limitedly affecting their labour supply. This study aims to address this difference as it analyses the (largely) indirect effect of work strategies on child poverty rather than focusing on parental work as the outcome.

2.3.3 Benefit strategies

Contrary to parental employment stimulation, the reasoning behind benefit strategies is relatively straightforward (Whiteford & Adema, 2007, pp. 7–8; Adam et al., 2006, p. 1; Solera, 2001, p. 461). When defining child poverty as a household income below a certain threshold, it seems most effective to directly target this shortage of income with a transfer via the tax system (Pisu, 2012, p. 28; Haskins & Sawhill, 2003, p. 1). The actual effects of different income benefits in combating (child) poverty are still widely disputed, however, and the estimation of their impact heavily influenced by the way in which they are measured (Haskins & Sawhill, 2003, p. 5). Concerning these policies, the most important distinction is between means-tested and non-means-tested or universal benefits (Immervol et al., 2001, p. 407; Solera, 2001, p. 461; Tárki Social Research Institute, 2010, pp. 119–120; Bradshaw & Finch, 2002, p. 49). These transfers can, then, either occur directly via a deposit or indirectly via an income tax break (Solera, 2001, p. 461).

Critics of income transfers as effective child poverty reduction mechanisms point to the classic ‘trade-off between work incentives and income redistribution’ (Pisu, 2012, p. 26; Adam et al., 2006, p. 1; Haskins & Sawhill, 2003, p. 5), while proponents state that the unique character of child poverty makes that there is no such trade-off (Heckman & Masterov, 2007, p. 2). As the causal mechanism between income benefits and child poverty is more direct than in the case of work strategies, predictions of the effects of these transfer programmes on child poverty have been more reliable. These studies generally use simulations such as the well-known EUROMOD model. Stuart Adam et al., for example, use TAXBEN, a micro-simulation tax and benefit model and data from the Family Resources Survey to estimate the effects that certain tax and benefit reforms in Britain would have on child poverty rates (Adam et al., 2006, p. 25). Ambiguously, they find that means-tested benefits are the most appropriate income benefit to address poverty in the poorest families but are also the most significant impediment on parental work incentives (Adam et al., 2006, p. 36) which is in line with the academic

11

consensus. When poorly targeted, means-tested income transfers are, therefore, in direct conflict with work strategies.

3. Dutch policy context

Interestingly, all of these strategies are present in the Dutch policy context. When looking at these government programmes in the Netherlands, it is important to know that the Dutch economy has recently been characterised by relatively high labour participation. Many families have two wage earners, and the share of households with children without employment has only been a few per cent since the turn of the millennium (Tárki Social Research Institute, 2010, pp. 15, 157–158; Bettendorf et al., 2015, p. 114). At the same time, these (mostly female) second wage earners generally work part-time, meaning there is a gender disparity in work intensity rather than the participation rate.

3.1 Broad family programmes

The high labour participation guarantees that the share of household income earned via the market is generally relatively high. Still, the Dutch policy context consists of a wide variety of policies that affect families with children to deal with their higher poverty risk. Typical of this governmental system is a relatively low degree of horizontal redistribution from those without off-spring to families with children, although this may be underestimated due to the significant role of in-kind benefits and support services (Tárki Social Research Institute, 2010, pp. 13, 75). Meanwhile, there is a relatively high degree of vertical redistribution between households with children that are well-off and those families in risk of poverty (Tárki Social Research Institute, 2010, pp. 13, 75). Non-child-contingent social transfers are generally highly concentrated on households with an income below sixty per cent of the household median (Tárki Social Research Institute, 2010, p. 157; Blommesteijn, 2014, p. 18).1 In addition to these schemes,

1 _{The main example of such policies during the time frame of this study was the social assistance programme for}

people of working age who are not considered disabled (WWB). Besides, a separate programme was started in 2004 for the self-employed with similar requirements (Bbz). Furthermore, there was general disability scheme (WAO before 2006, afterwards WIA), a scheme for young people who are considered (partly) disabled and have no history on the labour market (Wajong) and an instrument for older people of working age that are considered (partly) disabled (IOAW and IOAZ). Also, there is relatively generous sick pay (Ziektewet) and unemployment insurance (WW), which was complemented in 2008 with a separate scheme for people getting unemployed at an older age (IOW). Additionally, there were widely assigned benefits for housing in the rented sector (Huursubsidie until 2005, thereafter Huurtoeslag) and healthcare (Zorgtoeslag, introduced in 2005).

12

there is a wide variety of child-contingent instruments. With few exceptions,2 these narrow family policies fall into two categories: work and benefit strategies.

3.2 Work strategies

The theoretical framework demonstrated that two parental employment stimulation policies play a major role in the academic debate: targeted earned income tax credits and childcare subsidies. Interestingly, the Netherlands uses both approaches and has intensified its spending on these programmes over time as the rise of single-parent households in the Netherlands, who tend to work less, has made such policy instruments ever more salient (Bettendorf et al., 2014, p. 49). Since this study analyses the common effect of both work strategies between 2005 and 2009, it is necessary to discuss their development within this time frame in more detail.

Firstly, the Netherlands has had a relatively complex policy system when it comes to targeted earned income tax credits. During the study period, there were four EITCs of interest: the Combination credit, Additional combination credit, Additional credit for single parents and the Income dependent combination credit. For all these credits, only parents with a market income above a certain, relatively low threshold were eligible (Blommesteijn, 2014, p. 18; Jansen, 2015, p. 5). In the first place, the Combination credit (Combinatiekorting) was introduced in 2001 which was specifically targeted at increasing employment for families with children under twelve (Jansen, 2015, p. 5). In the second place, the Additional credit for single parents (Aanvullende alleenstaande ouderkorting) was also implemented in 2001 as a separate scheme next to the previously mentioned Combination credit (Bettendorf et al., 2014, p. 49; Jansen, 2015, p. 10). Only single-parent families were eligible for this scheme, while the youngest child in the household had to be below 12 years old, which became 16 years in 2002. In the third place, these EITCs were complemented by the Additional combination credit (Aanvullende combinatiekorting) in 2004 which targeted single parents and secondary earners in families with children under twelve.

Of these three EITCs, primary income earners were, therefore, only eligible for the Combination credit. Although the common size of the Combination credit and Additional combination credit increased, from 0.7 billion in 2004 to 1.3 billion in 2009, the relative share of the schemes for which primary earners were eligible gradually reduced and finally became zero in 2009 when the Income dependent combination credit (Inkomensafhankelijke combinatiekorting) replaced both combination credits (Jansen, 2015, p. 7; Bettendorf et al.,

2 _{The most important exceptions during the time frame of this study were a maternity leave scheme, a specific care}

allowance for parents with disabled children in need a lot of care (TOG) and a relatively ungenerous system of parental leave complemented by a special tax credit.

13

2014, pp. 114–115). In the same period, the size of the Additional credit for single parents remained roughly the same. Consequently, a study of the work strategy reforms between 2005 and 2009, de facto, analyses the structural, substantial increase in the size of the Additional combination credit when considering changes in EITCs.

Secondly, the other major part of the parental employment stimulation agenda in this period was the introduction of the Childcare subsidy in 2005 (Blommesteijn, 2014, p. 17). As new subsidy on formal childcare, the Childcare subsidy (Kinderopvangtoeslag) established a turnaround in the approach to childcare stimulation in the Netherlands (Berden & Kok, 2009, p. 1; Sociaal-Economische Raad, 2016, pp. 53, 55). Whereas previous policies, mostly enacted by local governments and employers, had solely focused on encouraging the supply of childcare, this public subsidy was implemented to lower costs on the demand-side (Roeters & Bucx, 2018, p. 42; Bettendorf et al., 2015, p. 114). The reform particularly lowered the costs of childcare in formal centres that were previously unsubsidised from the supply-side due to diverging policies by local authorities and employers. Additionally, guest parent care, which is ‘small-scale care at the home of the guest parent or at the home of the children’, was now also eligible for the benefit (Bettendorf et al., 2015, p. 114).

The conditions of the instrument quickly became relatively generous as the degree of subsidy significantly increased in 2006 and 2007, reducing the average parental fee from 37 per cent in 2005 to 18 per cent in 2007 (Bettendorf et al., 2015, pp. 112, 115). At the time, the Childcare subsidy had two major conditions: all parents in a household had to work in order to be eligible, and the youngest child in the family had to be not yet enrolled in secondary school which generally occurred at the age of twelve (Bettendorf et al., 2015, pp. 114, 116). The height of the benefit was dependent upon several other factors. Firstly, the degree of subsidy by the government was inversely related with the market income of the household (Roeters & Bucx, 2018, p. 42; Sociaal-Economische Raad, 2016, p. 47). The benefit, therefore, actively targeted those families that were prone to child poverty. Secondly, the hourly rates and the number of hours eligible for the benefit were maximised (Bettendorf et al., 2015, p. 112; Sociaal-Economische Raad, 2016, p. 47). Thirdly, the government only subsidised families with a household income under a certain threshold. In 2007, this boundary increased substantially to 130,000 euro, so that nearly all households were eligible for the Childcare subsidy (Berden & Kok, 2009, pp. 6, 19). Whereas 19.5 per cent of children were in formal daycare in 2006, the same group constituted only 6.8 per cent in 2007 (Berden & Kok, 2009, p. 19). Unsurprisingly, the expansion of the scope and generosity of the Dutch childcare policy was accompanied by a

14

sharp upsurge in its costs, from one billion euros in 2005 to three billion in 2009 (Bettendorf et al., 2015, pp. 112, 114; Blommesteijn, 2014, p. 17).

3.3 Benefit strategies

In the same period, the Dutch government also used several benefit strategies to combat child poverty. Of these programmes, the General Child benefit act (Algemene Kinderbijslagwet) is the most widely known as it is the traditional family policy of the Netherlands (Blommesteijn, 2014, p. 20; Van Daalen, 2002, pp. 300–301). In the ’90s, this instrument became a relatively modest, universal income transfer dependent upon the number of children in a household and their age (Blommesteijn, 2014, p. 20; Bradshaw & Finch, 2002, p. 52; Van Daalen, 2002, pp. 307–308). From 2001 onwards, a means-tested income transfer programme has been accompanying this measure which has changed significantly over time. The means-tested benefit started as an income tax break called the Child tax credit (Kinderkorting)(Caminada, 2007, pp. 2–3). In 2008, the means-tested Child subsidy (Kindertoeslag) was used to bridge the old system of the Child tax credit with the new Child-related budget (Kindgebonden Budget; KGB) system which has been in place since 2009 (Boer, 2018, p. 300). The height of the deposit of the Child-related budget has been dependent upon the number of children in a household as well as the family’s income (Blommesteijn, 2014, p. 20).

4. Method

4.1 Research design

This thesis focuses on the influence of work strategies on child poverty, however, studying the dependent variable both in absolute and relative terms as was mentioned in the conceptualisation section. Furthermore, the impact of the work strategies on deprivation during childhood is analysed by estimating the common effect of parental employment stimulation reforms in the Netherlands between 2005 and 2009. Concretely, these policy changes consisted of a substantial increase in the size of the Additional combination credit as well as the introduction (and later enlargement) of the Childcare subsidy. As the academic consensus suggests that parental employment stimulation programmes are effective instruments in combating child poverty, this study hypothesises that the work strategy reforms have a significant, negative common effect on child poverty, both in absolute and relative terms.

The study by Leon J.H. Bettendorf, Egbert L.W. Jongen and Paul Muller (2015, pp. 113, 116) which estimated the common effect of the same parental employment stimulation agenda on (female) labour supply is the main inspiration for the methodology of this thesis. Similar to that paper, the treatment group of this study consists of households where the youngest child is

15

below twelve years old, while families in which the age of the youngest child is twelve to seventeen years serve as the control group. As a result, families that solely contain children older than seventeen or none at all are left aside. Such a divide between the treatment and control group makes sense as most children are twelve years old when they go to secondary school while the youngest child in the household has to be not yet enrolled in secondary school in order for the family to be eligible for the Childcare subsidy (Bettendorf et al., 2015, pp. 115– 116). These groups are used to conduct a difference-in-difference analysis which compares the pre- and post-reform outcomes of the treatment and control group to determine the policy impact or fixed group effects, thereby blocking the influence of fixed time effects with the use of the control group (Angrist & Pischke, 2009, p. 227; Bettendorf et al., 2015, p. 115).

Like Bettendorf et al. (2015, p. 118), the analysis only estimates the aggregate, intention-to-treat impact of the Childcare subsidy (instead of the actual treatment effect) since the data quality regarding the usage of the benefit in the first years after its introduction is insufficient for an analysis of an actual, individual treatment effect (Angrist & Pischke, 2009, pp. 227–228). Building upon the earlier research by Bettendorf et al., this study aims to estimate the common effect of the same policy changes on child poverty, which (largely) takes place through parental work, rather than looking at the impact on labour supply as an outcome variable itself. Similarly, however, the analysis distinguishes between the short- (2005-2007) and medium-term (2008-2009) when analysing these effects. Additionally, separate difference-in-difference analyses are run to test for interaction effects of the policy reforms related to parental education, single parenthood and single motherhood.

The central assumption of the difference-in-difference technique is that the treatment and control group follow common trends when leaving the intervention and its effect aside (Angrist & Pischke, 2009, p. 230). The trends between the groups should, therefore, be similar in the period before the policy reforms of the parental stimulation agenda. Bettendorf et al. (2015, p. 116) used two placebo dummies to test the common trends assumption, making a distinction between the periods 2000-2002 and 2003-2004. For the first time frame, the placebo dummies partly turned out significant (Bettendorf et al., 2015, p. 119), which may be caused by the introduction and reforms of the Additional credit for single parents and Combination credit in this period as these EITCs were directly targeted at families with children under twelve (Bettendorf et al., 2014, p. 49; Jansen, 2015, p. 10). Due to these substantial policy changes, this time frame is not a suitable period for a placebo test. The placebo dummy for 2003 and 2004, directly before the policy reforms examined in this study, is an appropriate test for the common trends assumption, however, and is therefore included in the analysis. The dummy is

16

1 in 2003 and 2004 while being 0 in the other years in the study period. When the dummy turns out insignificant, this suggests that the common trends assumption holds which is a prerequisite for the difference-in-difference analysis.

Besides this main issue, some other methodological concerns require attention. The first is the possibility of anticipation of the intervention. As the significant increases in the subsidisation of childcare were only established in 2006 and 2007 while being unknown at the time of policy introduction in 2005, the anticipation of the reduction in the effective childcare fee by households is unlikely (Bettendorf et al., 2015, p. 116). Additionally, the placebo dummy of 2003 and 2004 tests for this issue as it should be significant when substantial anticipation of the policy reforms takes place. A second issue may be that the characteristics of the treatment group related to the outcome variable would change during the time frame. This problem would occur when there would be a change in fertility rates between 2005 and 2009. On the basis of administrative data, Bettendorf et al. (2015, p. 116) showed that there was no such change in fertility rates in the Netherlands during the study period, however. Therefore, this potential caveat is no issue in this paper which studies the same groups in the same country over the same time frame. Thirdly, heteroskedasticity may also be a concern in this research design. Therefore, a Breusch-Pagan test is conducted to test the assumption of homoskedasticity that is required for the use of regular standard errors (Angrist & Pischke, 2009, p. 46). If the test results in a rejection of the assumption of homoskedasticity, this study uses robust standard errors instead. Fourthly, movement between the treatment and control group may be an issue when it significantly changes the characteristics of both groups. In the data section, the demographic attributes of the movement group are compared with the static treatment and control group. This analysis indicates that the movement between the groups may be problematic. The issue is, therefore, addressed in the robustness checks. In the first place, the estimates are tested for the potential problem of movement by running an alternative difference-in-difference specification using the method of Bettendorf et al. In order to guarantee that there was no overlap between the treatment and control group after the policy intervention, they decreased the age range of the control group to households where the age of the youngest child is either 16 or 17. This approach, however, does not account for families getting another child which still moves them from the control to the treatment group. In the second place, the robustness of the estimates is, therefore, also checked by completely excluding the movement group from the analysis resulting in estimates which are solely based on the static treatment and control group. Fifthly, the chosen time frames of the difference-in-difference variables and data have to be checked on their robustness. For this test, the placebo, short-term and medium-run time

17

dummies are transformed into year dummies and corresponding year effects as part of the robustness checks. Furthermore, the effects are estimated with the use of another data bandwidth. In this robustness check, the difference-in-difference analyses are conducted on data from 2003 to 2009 rather than from 2001 to 2009.

In general, the robustness of the estimates is also tested in the main analysis itself by looking at the effects on four distinct poverty outcome variables and by conducting an extensive analysis per dependent variable which estimates the results of the difference-in-difference analysis for a wide variety of model specifications (Angrist & Pischke, 2009, p. 245).

4.2 Statistical model

A basic statistical model underlies this analysis:

𝑌𝑖𝑔𝑡 = 𝛼 + 𝛾𝑇𝑔+ 𝜅𝑃𝑡+ 𝜆𝑆𝑡+ 𝜌𝑀𝑡+ 𝜋(𝑇𝑔∗ 𝑃𝑡) + 𝜎(𝑇𝑔∗ 𝑆𝑡) + 𝜇(𝑇𝑔∗ 𝑀𝑡) + 𝛽𝑋𝑖+ 𝜀𝑖𝑔𝑡

In this equation, the subscript i identifies the individual household, while g indicates the group and t the time period. Starting at the left side of the equation, child poverty functions as the dependent variable Y. On the right side, α refers to the intercept of the control group, while T is a treatment group dummy variable which is 1 when a household is in the treatment group and is 0 when the family is in the control group. γ, then, reflects the effect of being in the treatment group in comparison with the control group excluding the policy reforms, also known as the group fixed effects. After that, the following three variables are time dummies reflecting the different periods under study. P is the time dummy for the placebo test which is 1 in 2003 as well as 2004, and 0 in the other years. Similarly, S and M are the time dummy variables for respectively the short- (2005, 2006 and 2007) and medium-term (2008 and 2009). The Greek symbols next to these variables, κ, λ and ρ, point to the corresponding time fixed effects in these time frames.

The symbols of major interest, however, are the aggregate intention-to-treat effects: π, σ and μ. More specifically, π is the placebo effect, whereas σ and μ reflect the intention-to-treat impact of parental employment stimulation on child poverty on the short and medium run. The statistical hypothesis of π is that there is no significant effect which would be in line with the common trends assumption. Consistent with the theoretical hypothesis of this study, the statistical hypothesis of σ and μ is that they are significant and negative as a higher degree of parental employment stimulation should reduce child poverty. As Bettendorf et al. (2015, p. 119) found that the medium-term effects of the reforms on labour supply were larger than the short-term influences, the expectation is that μ turns out higher than σ. Finally, ε symbolises

18

the residuals in the regression model, while X is a vector containing the control variables and β resembles their effects.

The major micro-determinants of child poverty in the academic debate serve as these control variables. Following the theoretical framework, parental education, single parenthood, single motherhood, the number of children and the age of the household head are, therefore, selected as covariates.3 Parental work is excluded, however, as it is an alternative outcome variable rather than a control variable that is expected to mediate between child poverty and parental employment stimulation, thereby interfering with the effect under analysis (Angrist & Pischke, 2009, p. 64). Contrary to the intention-to-treat effects of the treatment dummies, the analysis estimates direct effects in the case of the control variables because of linked individual household data between the covariates and poverty rates.

5. Data

For the estimation of the policy reforms’ effects, this study makes use of the DNB Household Survey (DHS) for the years 2001 to 2009. This dataset provides micro-data on a panel of more than 5000 households in the Netherlands. As the survey works with waves containing data on the previous year, the years of analysis correspond with the waves of the dataset between 2002 and 2010. The decision in favour of national micro-data is made to optimise the precision of the estimates. Since there is much variation between the socioeconomic regimes of different countries, estimating relative effects with cross-national data is difficult and problematic (Bradbury et al., 2017, p. 3). The choice for precision through national data comes at the cost of the external validity of the analysis, however. Although the analysis remains relevant to a wider scope of economically highly developed countries, the implications of the estimates for these countries should, therefore, be interpreted with more caution.

Following the structure of the treatment and control group, only families with at least one adult and one child below 18 years old are relevant to this analysis and are, therefore, kept in the dataset. Households with more than five adults are seen as outliers and not taken into account. After these adjustments, there are at most 4749 observations per variable at the

3 _{See Table A.1 and A.2 in the Appendix for the operationalisation and summary statistics of all the variables. As}

these tables indicate, parental health variables were also considered in the analysis. These variables are left out of the reported results, however, for four reasons. Firstly, they were relatively irrelevant in the academic debate on child poverty determinants when compared with the other covariates. Secondly, the differences between the treatment and control group for these variables were comparatively modest. Thirdly, the inclusion of the parental health variables had a very small effect on the estimates of the other variables, while they also proved relatively unable to explain child poverty outcomes. Fourthly, the significantly lower number of observations for these variables makes comparisons between models that include and exclude the factor of parental health problematic and tables that include these variables, therefore, confusing.

19

household level over the whole study period.4 Although there is a slight downward trend, the yearly number of observations is relatively similar across the time frame.

5.1 Household income and poverty data

For measuring the child poverty outcomes, this thesis uses the DHS data on net disposable household income. If there is no direct data on net disposable household income for a certain household, the values are based on imputation by the DHS researchers using the broad variety of other questions in the survey related to the income of households to reconstruct this indicator. While children are the units of analysis in this study, households, therefore, constitute the units of observation. Relying on household income to measure child poverty rather than the individual resources of children assumes that the resources of the family are equally shared with the children in the household (Corak et al., p. 11; Immervol et al., 2001, p. 12). This presupposition has shown to be reasonable, as household income strongly correlates with the living conditions of the children in the family.

Building upon this assumption, there is an instance of child poverty when household income is under the child poverty threshold. As this study analyses absolute as well as relative poverty, there are two operationalisations of this poverty line. Firstly, absolute poverty in the Netherlands is measured using the method of its national statistics bureau Statistics Netherlands (Centraal Bureau voor de Statistiek; CBS). Even when using an absolute indicator, it makes the most sense to select a poverty line that fits the society under study. This study, therefore, uses the national absolute income standard established by the Dutch national statistics bureau to measure child poverty rather than an international poverty line which is unadjusted for the Dutch socioeconomic context. The national poverty boundary of Statistics Netherlands is called the ‘low-income threshold’ (lage-inkomensgrens) and determined yearly based upon the purchasing power of the relatively generous social assistance level in 1979 corrected for inflation (Centraal Bureau voor de Statistiek, 2015, p. 16).

The poverty indicator uses equivalence scales to account for differences in household composition (Centraal Bureau voor de Statistiek, 2015, p. 15). The single-person household is the reference category and, therefore, has an equivalence scale of 1. In the base year 2000, the low-income threshold for a single-person household was 9,249 euros. Using the Consumer Price Index of Statistics Netherlands, the poverty lines of single-person households in the consecutive years are calculated, rounded up to whole euros.5 After computing the low-income

4 _{See Table A.2 in the Appendix for the number of observations per variable.}

20

threshold for single-person households, the multiplication of this poverty threshold with the equivalence factor of a specific family composition results in the poverty line for that household type. The number of adults and children in the family together determine this equivalence factor. There are at most five adults in a household in the data for this study, while the maximum number of children in a family is seven. Up to households with a maximum of four adults and four children, Statistics Netherlands has published exact equivalence scales. For additional adults and children, respectively 0,19 and 0,15 is added on top of these figures.6 After multiplication with the equivalence scale, the resulting low-income threshold for every household is used to calculate the absolute poverty indicators.

Secondly, the relative poverty threshold of this study follows the definition of Whiteford and Adema. In their OECD report on child poverty reduction, they define the relative poverty threshold as fifty per cent of the national median equivalised disposable household income (Whiteford & Adema, 2007, p. 10). Single-person households once again have an equivalence scale of 1 which means their poverty lines are precisely at fifty per cent of national median disposable household income. The computation of the poverty thresholds for different household compositions is conducted with the same method and equivalence scales as the absolute poverty threshold.

After the determination of the poverty line per household, these boundaries are used to create dependent variables for the statistical analysis. In the first place, the child poverty lines are used to create a headcount dummy which is 1 when the household is under the poverty boundary and 0 when this is not the case. The overall mean of this dummy is known as the headcount index (Haughton & Khandker, 2009, pp. 68–69) and is particularly useful when analysing the extensity of child poverty in the Netherlands. The second method also pays attention to the intensity of poverty by looking at the distance below the poverty line as a percentage of this threshold. When the household income of a family is equal to or larger than the poverty threshold, the value of the variable for this household is 0. The average of this indicator is called the poverty gap index (Haughton & Khandker, 2009, p. 70; Whiteford & Adema, 2007, pp. 10–11). As both indicators are created for each poverty line, the study makes use of four child poverty outcome variables.

5.2 Treatment and control group

Importantly, the trends in the pre-intervention period for these four child poverty variables have to be equal between the treatment and control group to conduct a reliable

21

difference analysis. A visual inspection of the development of the variables per group gives an indication whether such common trends exist. Figure 1, therefore, demonstrates the trends of the treatment and control group for the poverty variables.

Figure 1 Trend of treatment and control group for poverty variables

In the top left corner, the graph for the absolute child poverty headcount shows remarkably similar trends between the treatment and control group except for a small dissimilarity between 2001 and 2002. The same goes for the headcount graph of relative child poverty that shows a stable, harmonious development at a significantly higher ratio. Notwithstanding a small difference from 2001 to 2002, the trends in the figure are impressively alike. The trends of the poverty gap variables show more volatility than those of the headcount ratio, however. When looking at the development of the absolute child poverty gap in the bottom left corner, the general trends of the treatment and control group are still relatively alike. In 2004, there is a remarkable disruption in the trend of the treatment group, however. Interestingly, there is no such volatility between 2003 and 2005 for the relative child poverty gap which suggests it is closely related to the absolute poverty threshold. Despite a shock in 2002, the trends for the relative child poverty gap seem to be relatively similar.

22

that the development of the treatment and control group are relatively similar. As there seems to be a remarkable shock in 2004 for the absolute child poverty gap, however, this finding is far from conclusive and needs further examination. Another way of looking at the similarity is by investigating the demographic characteristics of both groups in more detail. Table 1 shows the summary statistics of these variables specified for the treatment and control group.

Table 1 Demographic characteristics treatment and control group7

Treatment group Control group

Mean Std.

Dev.

Min. Max. Obs. Mean Std.

Dev.

Min. Max. Obs.

Higher education head 0.392 0.488 0 1 3432 0.358 0.480 0 1 1313

Higher education partner 0.319 0.466 0 1 3249 0.205 0.404 0 1 1134

Single parent 0.054 0.226 0 1 3436 0.134 0.341 0 1 1313

Single mother 0.045 0.208 0 1 3436 0.104 0.305 0 1 1313

Number of children 2.073 0.885 1 7 3436 2.026 0.733 1 5 1313

Birth year head 1967.417 6.646 1931 1983 3436 1957.238 5.437 1924 1977 1313

The table demonstrates that the treatment and control group are relatively similar when it comes to the educational attainment of the household head and the number of children. There are also variables with greater differences between both groups, however, especially for the educational level of the partner, single parenthood, single motherhood and the birth year of the household head. Of these, the difference in the birth year of the household head is common sense as the control group by definition consists of people with older children.

To control for other dissimilarities, the demographic variables function as covariates in the difference-in-difference analysis, however. As the visual inspection of the data also showed some volatility in the general development of the treatment and control group, the common trends assumption is also statistically tested using a placebo time dummy in the difference-in-difference analysis. Interestingly, some of the observed disruptions in the trends may also be caused by the effect of movement between the treatment and control group in the panel. This movement may occur in both directions. Either people in the control group can get a baby which would move them to treatment group, or the youngest child in a household in the treatment group may reach the age of 12 which leads to a change of group in the opposite direction. Such movement can be problematic when the moving households are not typical of their group and, therefore, change the composition of the treatment and control group.

Table A.5 in the Appendix demonstrates the summary statistics of households that are only part of the treatment group, families that are just part of the control group and those that

23

move between the treatment and control group during the study period. As families can move both ways, it is reasonable to expect that the means of the movement group for the variables are in between the averages of the stable treatment and control group unless the movement is conditional upon the demographic characteristic. The table indicates that the movement group’s mean of single parenthood, single motherhood and birth year of the household head is in line with these expectations. There is no such alignment, however, for parental education and the number of children. For the latter variable, this makes sense as people who move from the control to the treatment group need to have at least two children which may increase the mean of the movement group. The lower educational level of the movement group, both regarding the household head and his or her partner, can be problematic, however, as the literature indicates that parental education is one of the most important determinants of household income and child poverty. Nevertheless, the change because of the movement seems to be modest when comparing the mean of the parental education dummies of the treatment and control group in Table A.5 with those in Table 1.

Still, the issue of movement needs further attention in the analysis to ensure robust results. For that reason, this study pays attention to this potential caveat in three ways. Firstly, the factor of education, both of the household head and his or her partner, is used as a control variable in the difference-in-difference analysis which makes certain that dissimilarities in educational attainment are taken into account. Secondly, the control group is modified in accordance with the approach of Bettendorf et al. (2015, p. 119) to only include households where the youngest child is 16 or 17 years old which makes sure that there is movement from the treatment to the control group after the policy introduction in 2005 (the data used ends four years afterwards in 2009). This method only offers a partial solution, however, as it only works to control for movement from the treatment to the control group but not the other way around since people can still move in this direction by getting a child. Thirdly, therefore, several additional difference-in-difference analyses are run as part of the robustness check which completely exclude the dynamic group and, therefore, only look at the stable treatment and control group.

5.3 Use of the Childcare subsidy

In addition to this divergence from the method of Bettendorf et al., this study also works with different data as it has a panel rather than a repeated cross-sections structure. Interestingly, this dataset enables a rough investigation of the use of the Childcare subsidy. Dissimilar to the effect of the EITC where people with specific characteristics certainly receive treatment, this study analyses the influence of the Childcare subsidy as an intention-to-treat effect as it studies the

24

impact of the availability of cheaper Childcare instead of the effect of actual use of the benefit on child poverty outcomes. Still, it is relevant to investigate whether people with certain demographic characteristics are likely to avail themselves of the policy. Although there is no data on the use during the first years of the programme, an analysis of the limited Childcare subsidy usage from 2007 to 2009 gives an indication of the relevant characteristics of people who take advantage of the programme which has not been possible in earlier studies where such individual data was absent. Table 2 demonstrates the attributes of people in the treatment group who did and did not make use of the Childcare subsidy between 2007 and 2009.

Table 2 Characteristics of the treatment group by use of the Childcare subsidy between 2007 and 2009

Use Childcare subsidy Not use Childcare subsidy

Mean Std.

Dev.

Min. Max. Obs. Mean Std.

Dev.

Min. Max. Obs.

Household income 39,377 17,931 11,144 116,566 167 35,176 18,954 4,373 188,838 778 Higher education head 0.551 0.499 0 1 167 0.407 0.492 0 1 778 Higher education partner 0.524 0.501 0 1 147 0.269 0.444 0 1 714 Single parent 0.120 0.326 0 1 167 0.082 0.275 0 1 778 Single mother 0.114 0.318 0 1 167 0.060 0.238 0 1 778 Number of children 1.946 0.688 1 4 167 2.162 0.862 1 5 778 Birth year head 1970.503 5.205 1953 1982 167 1964.702 8.129 1938 1983 778

Notes: The values of household income in this table are rounded up to whole euros.

Remarkably, the table shows that there are reasonable differences between both groups. Those households that make use of the Childcare subsidy tend to earn more, have higher parental education, include more single parents, contain more single mothers, have fewer children and a younger household head. The fact that these families earn a higher income despite being single more often and significantly younger is quite remarkable as it contradicts the general pattern of the data. This finding may be caused, however, by the vast difference in parental education, especially when looking at the partner in the household. Although the quality of the Childcare subsidy data in the survey is not high enough to ensure the robustness of these findings on their own, they provide a motive for an extension of the difference-in-difference analysis. Looking at interaction effects based upon the deviating characteristics of the group that actually used the policy between 2007 and 2009, respectively parental higher education, single parenthood and single motherhood, gives a deeper understanding of the policy impact. These extensions of the analysis are, therefore, carried out at the end of this study.

25 5.4 Multicollinearity and heteroskedasticity

Two more potential issues require attention first, however: multicollinearity and heteroskedasticity. To test for multicollinearity, an analysis is conducted on the correlations between the selected covariates. This enquiry indicates that the correlation between the educational attainment of the partner in the household and single parenthood as well as single motherhood is too large for these variables to be included simultaneously in the same difference-in-difference model. These covariates are, therefore, addressed in separate models of the difference-in-difference analyses. Additionally, the homoskedasticity assumption underlying the statistical method is tested with the use of the Breusch-Pagan test. Across all four dependent child poverty variables, the application of this Breusch-Pagan test leads to highly significant results which means that the hypothesis of homoskedasticity has to be rejected. This study copes with this heteroskedasticity complication by using robust standard errors in the difference-in-difference analyses.

6. Results

6.1 Main results

Now using robust standard errors, the difference-in-difference analyses are conducted following an additive specification structure. Tables 3 to 6 present the results of these difference-in-difference analyses. The first two models of the difference-in-difference analyses are used to measure the policy impact and test the common trends assumption by solely looking at fixed effects without controlling for other variables. In the other models, covariates are inserted to test the robustness of these findings. Due to multicollinearity, education of the partner in the household and single partnerhood as well as single motherhood, have to be addressed in separate models. Model 3, therefore, only includes the parental education variables, the number of children and the birth year of the household head as covariates, whereas the partner education dummy is replaced with the single parenthood and single motherhood variables in model 4.

Model 5, then, constitutes a repetition of the placebo test with the covariate specification of model 4. Choosing between the partner education dummy and single parenthood as well as single motherhood for the control variables in the placebo test model is difficult as both parental education and single parenthood are considered important determinants of child poverty, while the summary statistics of the partner education, single parenthood and single motherhood dummies all show dissimilarities between the treatment and control group. Since the educational attainment of the partner in the household rather than the general factor of parental education is deemed less important than single parenthood and single motherhood from a