The Public childcare and mothers' labor supply - Evidence from a quasi-experiment
written by Stefan Bauernschuster and Martin Schlotter
Introduction
Increasing the labor market participation of females to 75 percent is part of the EU 2020 strategy. Multiple strategies have been used to achieve this goal but one strategy is certainly the increased availability of public childcare. But for policy makers it is relevant to know how large the effects of such an increase in public childcare attendence would be. The topic has gained further relevance in Germany since their are just are a few studies investigating the effects of an expansion of public childcare yet although it was a major topic in several state election campaigns and included in the talks between CDU/CSU and the SPD to form a new coalition for Germany. The agreement that half-time childcare should be free of charge for everyone was labeled as a lighthouse-project by SPD chairman Martin Schulz.
Whereas the effects of an expansion of public daycare was not investigated in detail for Germany, for other industrialized countries more evidence already exists. Del Boca (2015) shows that the estimates obtained for different countries vary a lot, ranging from very small effects estimated by Wrohlich (2005) for Germany going up to enormous effects observed by Viitanen (2005) for the United Kingdom. While the studies named here use some structural estimation techniques, more recent paper focus on quasi-experimental evidence. Cascio (2009) produces a paper which is representative for the existing literature using experimental or quasi-experimental methods. Using the introduction of free childcare for pre-school children as unexpected event in the US, Cascio (2009) measures strong responses of maternal labor supply on the introduction of free childcare in the early 1960ies. Other papers such as Fitzpatrick (2010) measuring the effect in the US using more recent data on the other hand observe much smaller and even insignificant effects of comparable measures. For Canada, Lefebvre & Merrigan (2008) exploit the introduction of universal childcare with increased subsidies in Quebec and use the remainder of Canada as a control group for their difference-in-difference approach. Also the Scandinavian countries with their highly developed welfare state have been subjected to some research measuring the effects of increased public daycare availability. Havenes & Mogstad (2011) observe that a reform in Norway in the 1970ies almost yields no effects on maternal labor supply. Although the research conducted so far has yielded very different results and do not point in any direction at first glance, Lundin et al. (2008) however observes some regularities: Most studies which report quantitatively high effects are conducted in a setting where female labor market participation is rather low (i.e. Cascio using data from the 1960ies) and especially countries with already high labor market participation of show almost no effects even in cases where full and free public childcare would be introduced (Norway, East Germany). Very few papers furthermore discuss whether such a reforms a financing themselves. A recent paper by Bettendorf et al. (2015) estimate that the costs of the introduction of a large childcare subsidy in the Netherlands are 2.6 bn Euros, which increased labor market participation by 30000 full-time equvalents and therefore Bettendorf et al.(2015) state that the reform was rather expensive. In this context the authors also report that the Dutch government in 2015 (VVD (liberal-conservative) & PvdA (Social Democrats)) at that time planned to reduce the subsidy introduced before in order to comply withthe Maastricht criteria. Furthermore the Dutch government believed that the program was ineffective which looks reasonable recognizing the results by Bettendorf et al. (2015). It must be stated that the magnitude of the results obtained for the Netherlands are not particularly high and rank between the results of Havnes & Mogstad (2011) and Lefebvre & Merrigan (2008).
Before we go into the details of the paper at hand, we will briefly discuss the results obtained for Germany. As already mentioned above Wrohlich (2005) using a structural model reports rather small effects even if public childcare would be available for everyone. Gathmann & Sass (2012) exploit a unique setting in the federal state of Thuringia where the (conservative) state government introduced a subsidy for caring children at home and not bringing them to a childcare institution. They observe, using the other Eastern German States as control group, small but significant (negative) effects on childcare attendance by this reform. Further evidence for Germany comes from Geyer et al. (2015) which use a structural model and a quasi-expermental methods for mothers with children aged three years and below. Together with new generous parental leave regualtions the authors conclude thchildcare especially in the first year after birth maternal employment increases by almost five percentage points which the authors interpret as considerably.
The Setting of the legal claim on childcare in the 1990ies
Until the early 1990ies the Western German labor market was rather traditional with low employment rates for women, especially those with children. Same was true in West Germany for childcare attendance. Children younger than four almost never attended a kindergarden at that time. To counter the demographic developments that Germany aldready faced in the early 1990ies, one of the main goals of the federal government was to increase female labor market particiaption. In this setting the conservative-liberal government of Helmut Kohl introduced a legal claim for a subsidized public childcare spot for children aged three to six (Rechtsanspruch auf einen Kindergartenplatz) in 1995. On the 1st of January 1996 the law entered into force but the German Bundesrat (the legislative chamber where the federal states are represented) demanded some changes in order to relieve the municipalities which in the end are responsible for the creation of the respective spots. This lead to a compromise which allowed the municipalities to set cut-off rules, granting only children spots which were three years old at the beginning of every kindergarten year (which is more or less equivalent to the start of the school year). Municipalities which used the rule could do so until 1998. However this rule was not manatory and not all municipalities employed it. We will later elaborate on the details and the problems of this cut-off rule.
80 percent of the costs of public childcare are subsized in Germany therefore Bauernschuster & Schlotter argue the fee parents would have to pay for childcare can be neglected in their empirical analysis. We will later come back to this assumption and will strongly argue against the claim Bauernschuster & Schlotter are expressing here.
The identification strategy
The setting described can be easily characterized as a quasiexperimental setup created by an unanticipated policy shift. With that the paper is in line with most of the national and international literature which already exists. The emloyed methods are both rather simple and applied straight forward.
The Instrumental Variable Approach
The cut-off described above is the key mechanism that Bauernschuster & Schlotter exploit for their Instrumental Variable (IV) approach. The authors argue that a child being above and below the cut-off is a sort of exogenous variation which influences the probabilty of attending a kindergarten significantly and influences maternal labor supply through no other channel than the attendance of the kindergarten. With this setting they follow Angrist & Krueger (1991) on their famous example to measure the effects of schooling. This means we will have a variable \(Z_i\) indicating whether the youngest child of a mother was 36 or older (\(Z_{i\ }=\ 1\)) or younger (\(Z_i=0\)) when the school year started. This leads to the First-Stage-Regression
\(D_{i} = \alpha + \delta Z_{i} + \beta X_{i} + \epsilon_{i}\) (1)
In this regression \(D_i\) indicates whether the youngest child attends a kindergarten or not , \(Z\) is our cut-off dummy and \(X\) is a selection of observable characteristics which are included to control for characteristics such as maternal eduaction. \(\epsilon\) is - usually - the error term. The paper now follows the standard two-stages-least squares (2-SLS) procedure to obtain the causal estimate, the authors are interested in.
\(Y_{i} = \eta + \tau \hat{D_{i}} + \varphi X_{i} +\upsilon_{i}\) (2)
\(Y_i\) is a variable expressing the labor market status of the mother (could be extensive margin, meaning a dummy indicating whether the mother works or intensive margin reporting the number of hours the mother has worked), \(\hat{D_{i}}\) are the predicted values of \(D_i\) from the first stage regression , \(X\) is again a set of control variables and \(v_i\) is the error term. If \(Y_i\) is a dummy than (2) can be interpreted as Linear Probability Model (LPM).
Now it is necessary to shed light on what this IV approach actually would measure: It should be clear from the description of the cut-off mechanism that only a very particular group is targeted by this instrument and therefore Bauernschuster & Schlotter obtain a very specific Local Average Treatment Effect (LATE). So only complier households (meaning those who would without the claim don't bring their child to the kindergarten but do so after the introduction of the claim) are captured by the estimate of the authors. So neither households which already have their child in the childcare institution even without the claim (which are more than 55%!) nor the parents that would leave their kids at home are captured by the IV approach of this paper. Furthermore the cut-off mechanism is imprecise in various regards. The authors have no idea which municipalities actually applied the cut-off rule until when and furthermore other municipalities were not forced to apply it, so technically the authors can only observe the effects for municipalities they cannot name.
The Difference-in-Differnece Approach
To further validatethe results from the IV approach, Bauernschuster & Schlotter also make use of a Difference in Difference (DiD) appoach. Now they compare the development of those mothers targeted by the reform to three different control groups with the following equation:
\(Y_{i} = \alpha + \beta T_{i} + \gamma D_{i} + \delta (T_{i} * D_{i}) + \epsilon_{i}\) (3)
where \(Y_i\) is again the labor market outcome of the mother, \(T_i\) a dummy whether the observation is from 2001 (\(T_i\ =\ 1\)) of from 1996 (\(T_i=0\)), \(D_i\) a dummy whether the person belonged to the treatment group (\(D_i=1\)) or the control group (\(D_i\ =0\)), and \(\epsilon\) the error term. The parameter of interest in this case is the \(\delta\) because the interaction term shows the effect over time that can't be attributed to a general time trend but to the particular change the treatment group experienced. This approach relies on the common-trends asssumptions which states that the evolvment of the outcome variables over time in treatment and control group is similar (controling for differnces in observables).
Compared with the LATE from the IV approach Bauernschuster & Schlotter obtain now an Average Treatment Effect of the Treated (ATT), which is much more general than the LATE from before because now the authors have broader definition of treatment group namely all those mothers with their children which became elgible through the reform. However some questions arise when we take a closer look to the control groups the two reseachers take for their DiD approach: Since the claim was introduced on the federal level, other regions could not serve as control group. Therefore they take mothers where the youngest child is 11 or 12, females without children aged 29 to 36 and to all females between 18 and 60. This idea is not completely new in case a reform is implemented on a nation wide basis (check i.e. Bettendorf et al. (2015)) but other studies with a control group within the same country like Lefebvre & Merrigan (2008) or Gathmann & Sass (2012) have a control group which is more likely to fulfill the common-trends assumption.
Data Sources and Descriptive Statistics
Data Sources
Both approaches use micro data from mothers which youngest child is between three and four when the reform was implemented in the late 1990ies. The best potential data sources for this purpose and period in Germany is the Sozioökonomisches Panel (SOEP) which is an annual penal data survey with a rich set of observables to control. Furthermore it is possible to infer in which month a child was born, which is necessary to calculate whether the child was three years old when the school year started and therefore determine the value of the instrument \(Z_i\) . Problems the SOEP could have are that it potentially has not enough observations but to deal with that, Bauernschuster & Schlotter use the German Microcensus, one of the largest regular surveys in the world, which also include multiple covariates but - and this is crucial for the IV approach - not the birth month of the youngest child. Therefore the instrument can not be calculated properly and the microcensus data set can only used for the DiD approach. Given the fact that SOEP data set at least potentially contains information about the county of residence (if you have access to SOEP remote that should be no problem at all, and even the humbled author of this paper review has access to SOEP remote) it is questionable why they were not able to identify the municipalities applied the cut-off rule. The authors themselves remain silent on the matter but could be criticized for that.
Descriptive Statistics
It might be interesting before one sees the estimation results to have a look at the descriptive statistics for some of the variables which are important for the discussion of the results. Of most interest may be the development of the kindergarten attendance rates for children of various age groups (namely 3,4,5 and 6 years old) during the 1990ies and the early 2000nds displayed in Figure 1.