1 Introduction

Early childhood education and care (ECEC) is widely considered an important channel to help children reach their full potential; however, a consensus is yet to emerge on whether all children benefit and the causal mechanism if any. As the period from birth to the age of five is the most crucial time for children’s overall development, or “crucial period,” ECEC attracts worldwide attention from policymakers. Following the 1965 introduction of the Head Start programme in the USA, many Organisation for Economic Co-operation and Development (OECD) and middle-income countries have established or expanded their ECEC programmes. The United Nations has also included universal ECEC among its Sustainable Development Goals (SDGs) (United Nations 2015).Footnote 1 In Japan, early childcare was improved as a pillar of post-war reconstruction and economic development, and its continued importance is reflected by the government’s decision to introduce free ECEC in 2020.Footnote 2

Besides the political realm, ECEC has long been of interest to researchers in several academic fields (Phillips and Shonkoff 2000; Knudsen et al. 2006; Barnett 2011; Currie and Almond 2011; Campbell et al. 2014). Economics studies focus on relatively short-run effects or effects of targeted programmes (e.g., Garces et al. (2002)).Footnote 3 Other recent studies discuss the effects of universal ECEC programmes using a difference-in-differences (DID) approach (Havnes and Mogstad 2011b; Herbst 2017). Some investigate which aspects of children’s ability developed during an ECEC programme affect later outcomes (Heckman et al. 2013). In Japan, high-quality ECEC facilities have attracted many scholars’ attention (Matsushima 2015; Yamaguchi et al. 2018b).

Despite numerous ECEC studies, little is known about universal ECEC’s (local) average treatment effects in the long run or the observable paths through which such programmes affect children. These issues are important for three reasons: first, DID estimates the effects on all children, but cannot separate the effects of those who change their behaviours based on the treatment from overall effects.Footnote 4 In addition to what we can learn about overall benefits from the DID approach, it is important to consider who is a potential beneficiary and how large the effect is on them when formulating policy. Second, if the aim is to identify the optimal ECEC design for maximum social benefit, it is necessary to understand the effects of a universal programme rather than a targeted programme. The latter excludes many potentially eligible people, and the effects might differ between the disadvantaged and advantaged. Finally, if we formulate a new policy, which may affect a later life stage, we need to consider long-term effects and their mechanisms. In Japan, high-quality ECEC facilities have existed since World War II, which are suitable for analysing long-term effects, but few analyses have exploited this long history because of a lack of data.

These concerns motivate the present analysis of ECEC’s long-term local average treatment effects (LATE) on children’s future outcomes, such as income and education attainment. Besides these socioeconomic outcomes, this study analyses the effects on individuals’ psychological outcomes, such as risk preference and the “Big Five” indicators, as non-cognitive abilities may affect socioeconomic outcomes (Heckman and Masterov 2007; Heckman et al. 2013; Havnes and Mogstad 2015).Footnote 5

One challenge in analysing ECEC’s effects is the existence of a potential endogeneity problem, as a parent’s choice to enrol a child in ECEC is likely to be correlated with unobservable household characteristics. Therefore, we need to exploit an exogenous shock to reveal the causal effects of enrolling in the childcare system. To overcome this problem, this study uses the universal ECEC expansion in Japan from the 1960s to the 1980s as a quasi-random shock. During this post-war recovery period of rapid economic growth, a larger labour force was needed to support growth and the government and companies urged women to work (Matsushima 2015). To support women’s labour participation, the government opened more ECEC institutions, such as kindergartens and nursery schools. ECEC institutions were opened to mitigate regional imbalances in the number of facilities. Therefore, jointly controlling for area, cohort fixed effects, and their interactions, the intensity of the expansion can be seen as quasi-random.Footnote 6Footnote 7Footnote 8 Another challenge is related to data quality. To investigate the long-term (local) average treatment effect of enrolling in ECEC, we need data on whether children went to ECEC facilities and their outcomes in adulthood. Information about both is rarely available for the same person. However, this study uses unique Japanese survey data containing both types of information. Using this data set and quasi-random variation, this study estimates ECEC’s long-term local average treatment effects.

The results show that enrolling in a universal childcare system increases the probability of college completion and leads to higher future wages and annual income. However, it does not increase working status nor working hours, which indicates that income increases because of a rise in wages. Further subsample analyses suggest that the effects of universal ECEC are concentrated among females. This reflects the gender inequality at that time in Japan, and that women were generally disadvantaged. However, there are no effects on psychological abilities measured by risk preference and the Big Five indicators. These estimates are robust to other specifications. I also conduct a control experiment where I examine the “effects” of ECEC attendance for people old enough not to have been benefited by the expansion of ECEC, following Duflo (2001). I do not find any effects for them, which suggests the exclusion restriction holds.

Next, I conduct mediation analysis to identify the mechanism driving the above results. Most effects of a childcare system on future income can be explained by an increase in the likelihood of college completion owing to childcare enrolment. Together with the finding of a positive effect on wage but neither on working status nor working hours, the results imply that ECEC increases future income through education, leading to higher wages, which is consistent with theories on human capital development (Mincer 1974; Becker 1994; Kane and Rouse 1995; Thomas 2003). This study’s findings also imply that most of the effects of ECEC can be explained by an increase in the likelihood of college completion. No other channels are uncovered from ECEC to future income. This is consistent with not finding any long-run psychological effects.

This research contributes to the literature in the following aspects. First, as discussed above, few studies discuss long-term effects of universal ECEC. In this analysis, using individual choices on whether children enrolled in ECEC, their current outcomes observed in the unique data set, and quasi-random variation on expanding universal ECEC by the Japanese government, I estimate the LATE of a universal ECEC programme.Footnote 9

Second, this study addresses a gap in the literature by revealing the mechanism behind ECEC effects. The mediation analysis distinguishes the direct effects of ECEC from its observable indirect effects through educational attainment. I find that ECEC increases the likelihood of college graduation, which, in turn, increases future income.

This study has implications for policymakers by showing the long-term positive effects of early-stage educational and childcare intervention on education attainment and income. ECEC reduces inequalities between the advantaged and disadvantaged, especially among genders given the cultural situation in Japan at that time. This implication provides governments with a good reason to expand ECEC, lower fees, and eliminate barriers to enrolment, particularly for disadvantaged children, given the finding of no negative effect for advantaged children based on the analysis of the marginal treatment effects. Although the results do not seem directly applicable to the current situation in Japan because the availability of ECEC is substantially different from that of Japan in the 1960s to the 1980s, the knowledge can be useful for countries with a gender culture similar to Japan’s, or where the government is investing in ECEC, including many developing countries, following the SDGs (United Nations 2015).

The remainder of the paper is organised as follows. Section 2 reviews the background and literature related to this research, while Section 3 explains Japan’s ECEC system. Section 4 describes the data, Section 5 explains the econometric framework, and Section 6 shows the identification strategy. Section 7 presents the results of the long-term LATE analysis. Further analysis of the mechanism is provided in Section 8. Section 9 concludes the study.

2 Prior research

Many studies have discussed ECEC’s effects in the literature of economics and education. However, most examined the effects of targeted programmes or relatively short- or medium-term effects of universal ECEC. Some examined the long-term ITT effects, but could not capture the effect on those who are treated and react to the change in universal ECEC availability. In this study, I use Japanese data to study the long-term LATE of universal ECEC on children’s income and educational attainment, in addition to investigating its mechanism.

While many studies have focused on the short- or medium-term effects of ECEC, several have also paid attention to the long-term effects.Footnote 10Footnote 11 Many studies on the long-term effects on educational outcomes or earnings have analysed Head Start, a federal matching grant programme to improve poor children’s academic and social skills and health status in the USA.Footnote 12Footnote 13 Currie and Thomas (1995) find Head Start positively impacts children’s test scores and reduces the likelihood of repeating the same grade. Similarly, Garces et al. (2002) find positive effects on children’s long-term outcomes, such as educational level and earnings when in their early 20s.

Regarding targeted programmes other than Head Start, Lavy (2018) examines a free school choice programme for disadvantaged primary school students in Israel. Although the focus is not on a preschool, he shows the long-term effects of schooling in relatively early childhood: he finds that the programme increases post-secondary education (university or college training) and the future income of programme participants when aged around 30. Heckman et al. (2013) discuss ECEC’s long-term effects on income using data from the Perry Preschool Project, an early intervention programme for disadvantaged U.S. African American youths. Although the sample size is relatively small, the results show that ECEC reduces the number of crimes, increases the probability of employment for men, and increases the duration of marriage for women. The results further show that the source of these effects is persistent, with personality skills cited as the underlying mechanism.Footnote 14 Furthermore, Heckman et al. (2010) reveal that the Perry Preschool Project’s internal rate of return is higher than the historical rate of return on standard equity of approximately 5.8% considering participants’ income, crime status, and other socioeconomic outcomes.Footnote 15 These targeted programmes are shown to have long-term positive effects on various outcomes.

However, universal childcare systems may have different effects from those of targeted programmes. Baker (2011) urges caution when using targeted ECEC programmes’ results to assess universal ECEC’s effects because of significant differences in household backgrounds. Especially, the effects of targeted programmes might be shown to be much higher than those of universal programmes. Cornelissen et al. (2018) also points out the potentially higher treatment effects observed in targeted programmes.Footnote 16 A targeted programme is often used to fill the gap between the advantaged and the disadvantaged by offering the programme to the latter. For example, Head Start in the U.S. accepts children whose family’s income is lower than $26,500 per year for a four-person family in most states.Footnote 17 Given that the quality of education of Head Start can be regarded as generally high, the effects of attending Head Start will be high for these disadvantaged children (Currie and Thomas 1995; Garces et al. 2002). However, children from wealthy families can go to a private school with a better environment, which tends to be too expensive for children from poor families to enrol in. Therefore, the effects of attending Head Start would not be high, or even be negative because they might already be going to the ideal school. If Head Start became a universal programme, the effects would be averaged out, and we would not see any positive effects of it. On the other hand, since Head Start’s programme is unique and different from other programmes, it may be beneficial even for those from rich families. In this case, the effects of universal Head Start would be positive on average, and one could conclude that expanding this programme to everyone would be effective. These comparisons are feasible only when we examine a universal programme and a targeted programme. Although what I analyse in this paper is not an existing targeted programme, examining the effects of a universal programme seems important for policymaking.Footnote 18

Therefore, extending the analysis to universal systems is crucial. Some studies focus on universal childcare systems. Felfe et al. (2015) find positive effects on the cognitive ability of the children attending publicly subsidised childcare in Spain using the DID framework. They observe enhanced effects for girls and disadvantaged children. Herbst (2017) examines the ITT and treatment-on-the-treated (TT) effects of near-universal childcare in the USA. Herbst (2017) finds positive effects on college completion. Havnes and Mogstad (2011b) estimate the ITT effects of subsidised childcare on children’s long-run outcomes using Norwegian data. They find positive effects on children’s educational attainment and labour market participation and decreases in welfare dependency. The policy is most beneficial for girls and children with less-educated mothers. On the other hand, Baker et al. (2008) discuss the universal childcare system’s negative persistent effects in Quebec on family well-being after approximately ten years. Haeck et al. (2018) show that these negative effects last in the medium-term, but disappear when children become teens. Furthermore, Baker et al. (2019) report the negative effect on non-cognitive outcomes, health outcomes, and self-satisfaction. Moreover, in Bologna, Fort et al. (2020) find that the daycare system reduces the participants’ intelligence quotient. In Japan, the focus of this study, Yamaguchi et al. (2018b) examine the effects of childcare enrolment on short-term childhood outcomes by exploiting staggered childcare expansion across regions. They find that childcare improves language development among boys and reduces aggression and the symptoms of attention deficit hyperactivity disorder among children with less-educated mothers. The effects are stronger for poorer families. Although there have been many studies, as seen above, the conclusion on the effects of universal ECEC is still mixed, and estimates do not capture the long-term effects on those who changed their behaviour because of the change in universal ECEC’s availability, which is quite informative in policymaking. Furthermore, regarding the discussion on effective policies, the life-long effects must be discussed. Therefore, examining the effects about half a century after enrolling in ECEC seems important.

Before ending the literature review, it is worth reviewing (Duflo 2001) because I will follow some of her methodology, such as the control experiment. She examines the effects of school construction in Indonesia. She adopts a similar strategy to mine: she exploits the differences in the timing of construction and uses it as an instrumental variable for the attending status. She finds positive effects on the years of education and wage. What is important in the relationship between her study and mine is the way to conduct a control experiment. She investigates the “effects” of school construction for those who were older than the school-age. This means that they would not have been benefited from the school construction, so the estimation should be close to zero, if there is no other path affecting the construction of schools and educational and economic outcomes. This is useful for examining whether the exclusion restriction holds or not. I adopt this method and discuss the result in Section 7.2 in the results section.

3 Background and education system in Japan

Before explaining the ECEC system, I briefly review Japan’s entire education system. In Japan during the 1950s–1980s, which covers the focus of this study, children started compulsory schooling at six years, at an elementary school with a six-year curriculum. After graduation, children attended a three-year junior high school. Compulsory education finished at this level, and children could choose to enrol in a high school for three years, followed by a college or university for two to four years. Before this compulsory education, some children enrolled in preschools, which were of roughly two types: full-time nursery school and part-time kindergarten. Nursery schools accepted children aged from zero to five, while kindergartens were for children aged three to five.Footnote 19 However, because of the capacity constraint discussed later, many parents could not choose the facility at which they wanted to enrol their children, and many children could not even enrol in ECEC, as discussed below. In the last subsection of this section, I briefly review the gender differences in the education environment, since this paper investigates gender heterogeneity in the effects of ECEC, and knowing the difference in circumstances helps us understand the mechanism behind the effects.

Fig. 1
figure 1

Capacity and number of children enrolled in kindergartens. Each point represents a pair of capacity and the number of children enrolled in a year in a prefecture

Fig. 2
figure 2

Capacity and number of children enrolled in kindergartens. Each point represents a pair of capacity and the number of children enrolled in a year in a prefecture

3.1 Preschool systems and roles

Here, I describe Japan’s nursery school and kindergarten system and their roles.Footnote 20 I also discuss ECEC’s history after World War II. Since I use the expansion of ECEC from the 1960s to the 1980s for my analysis, I specifically focus on the 1950s to 1980s in this review.

The purposes of part-time kindergartens and full-time nursery schools differ. Kindergartens aim to develop children’s minds and physical strengths by providing a sound educational environment, while nursery schools provide care for children whose parents cannot care for them because of work or other commitments (Akabayashi and Tanaka 2013).

Despite different purposes and administration, their functions overlap in many ways. Both provide education to children, and, therefore, many consider them roughly equivalent (Akabayashi and Tanaka 2013).Footnote 21 Because of the imbalance in locations of kindergartens and nursery schools, some parents who might have wanted to send their children to nursery schools actually had to send them to kindergartens and vice versa, thereby reducing the differences (Matsushima 2015). Furthermore, both institutions are accredited by the government.Footnote 22

For accreditation, both institutions must satisfy some quality requirements set by the Child Welfare Act and School Education Act in 1947 (Ministry of Education 1979; Shakai Fukushi Jigyo Shinko Kai 1963; Yamaguchi et al. 2018b). There are also educational requirements for kindergarten teachers and caregivers in nursery schools. Typically, the minimum required educational attainment for a kindergarten teacher’s licence is two years of college or university education, and for a nursery schoolteacher, two years of high school education and one year of experience (Ministry of Education 1979; Yoshimi 2001; Yamaguchi et al. 2018b). These requirements indicate that teachers and caregivers are skilled, and that the quality of preschools can be considered high.

The teacher–student ratio seems similar between the two types of facilities. Zen-Nihon Shi You Ren Soumu Iinkai (2017) showed that a kindergarten’s average teacher-student ratio was about 25 in the 1960s to 1980s. The nursery school teacher-student ratio was similar. The Minimum Standards for Child Welfare Facilities required at least one childcare worker in a nursery school for every three infants, at least one for every six children under the age of one to three, at least one for every 20 children under the age of three to four, and at least one for every 30 children aged four or more. Together with the fact that older children were prioritised for enrolment in a nursery school, as discussed by Matsushima (2015), and the capacity was binding as discussed later, the average ratio is around 25–30.Footnote 23 Figures 1 and 2 show the relationship between capacity and number of children enrolled in kindergartens and nursery schools. The figures illustrate the lack of space for additional children in both kindergartens and nursery schools.Footnote 24Footnote 25

Regarding the ECEC fee, the average price of a nursery school was around 7,000 JPY per month in 1979 (Takayama 1982), although it changed over time across municipalities and was subsidised for low-income families.Footnote 26Footnote 27Footnote 28 Meanwhile, the fee of a kindergarten was around 8,800 JPY in 1979 (Minister’s Secretariat, Ministry of Education, Culture, Sports, Science and Technology 1980).Footnote 29 This evidence supports the outcome of similar qualities among nursery schools and kindergartens.

3.2 Expansion of ECEC

After World War II, Japan became westernised in many aspects, including ECEC (Kawai 1979). Kindergartens were established according to the 1947 School Education Law under the Ministry of Education, and nursery schools were constructed following the 1948 Child Welfare Law under the Ministry of Welfare (Tobin et al. 1991). Behind this movement lay the evolution of Japanese families from traditional extended families to nuclear families (Shwalb et al. 1992).Footnote 30 This decline in the traditional form gave mothers and ECEC more responsibility for childcare, although there was still a slight increase in the number of three-generation families in which grandparents could assume responsibility for childcare (Miyake 1989). The average number of family members decreased in this period: 5.0 in 1920–1955, 4.5 in 1960, 4.0 in 1965, and 3.1 in 1988 (Shwalb et al. 1992). The decline in the birth rate also contributed to this fall. Furthermore, given the shrinking of neighbourhood communities, Japanese children had fewer playmates around their homes, which increased the importance of childcare facilities for peer interactions (Shwalb et al. 1992; Miyake 1989). Actually, the expansion of ECEC does not seem to have replaced maternal care: in the literature, Asai et al. (2015, 2016); Yamaguchi (2017); Yamaguchi et al. (2018) report no effects of ECEC on maternal labour supply. Moreover, Asai et al. (2016) and Yamaguchi et al. (2018) find that ECEC crowded out the informal childcare provided by grandparents. Although their analyses focus on the 1990s to 2010s, given the situation where the number of family members is declining, the same seems to have occurred.Footnote 31

Fig. 3
figure 3

Mean of the proportion of children enrolled in ECEC at age four across regions over time. The variables are weighted by the population of the cohort in each prefecture. There are variations across regions over time. Some values exceed one because I calculate the proportion of capacity in nursery schools at each age using the capacities in nursery schools across ages and the proportion of each age at the national level. Regarding area divisions, see Table 13 and Fig. 10 in Appendix A

Fig. 4
figure 4

Change in the proportion of children enrolled in ECEC at age four from 1961 to 1984. I use Stata command maptile with the option jpn_pref created by Chigusa Okamoto. See http://www.crepe.e.u-tokyo.ac.jp/en/materials/maptile.html for more details (Last access: 1st June 2022). This is a preliminary use, and I thank her for permitting its application

Fig. 5
figure 5

Proportion of children enrolled in ECEC at age four by gender. The proportion is calculated by taking the average of the dummy variable indicating whether a child went to a kindergarten or nursery school at the age of four (1 if enrolled) in the main data

Fig. 6
figure 6

Proportion of young adults enrolled in higher education over time. This number includes people who attended two-year colleges. The data source is the School Basic Survey collected by the Ministry of Education, Culture, Sports, Science, and Technology (MEXT). This includes all age groups

Fig. 7
figure 7

Nominal wages of men and women over time in Japan. This figure is from The Japan Institute for Labour Policy and Training (2020), with the labels translated by the author from Japanese into English

When the post-war fallout stabilised, the number of births increased dramatically (i.e., a baby boom) and this increased people’s interest in ECEC. Simultaneously, women’s social status was rising. Although the Japanese people have traditionally valued mothers’ participation in childcare, more Japanese women entered the workforce, particularly after World War II, thus driving the demand for childcare systems (Shwalb et al. 1992). In the middle of the 1950s, Japan’s economic growth surged, and female labour demand increased dramatically (Matsushima 2015). At this time, there were much fewer capacities of ECEC facilities than the number of children whose parents wanted them to enrol in an ECEC facility. This increase in the demand for ECEC happened all over Japan.Footnote 32 The government tried to increase the ECEC facilities, but the problem was not solved, and the size of the budget and the numbers of teachers and facilities were still too low over the 1950s, although the number of facilities themselves increased over this period. Therefore, the government prioritised areas where the number of facilities was relatively smaller compared to the population. In 1956, the Kindergarten Establishment Standards was promulgated and it said the expansion of kindergartens should be based on the imbalance of the number of facilities. In 1957, the government decided to offer subsidies for constructing nursery schools in areas where the number was not yet sufficient (Kousei Shou Jidou Kyoku 1959; Matsushima 2015). The government used the measure of the number of nursery schools per 1,000 people and that of kindergartens per 10,000 people. They tried to capture the imbalance in ECEC and prioritise the area where the resources were scarce (Matsushima 2015). I examine the determinants of the ECEC expansion quantitatively, following the analysis by Cornelissen et al. (2018).Footnote 33

As a result of prioritisation, there were large differences in the number of ECEC facilities (e.g., nursery schools and kindergartens) at that time among areas (Matsushima 2015). Moreover, the supply of ECEC was binding—even in areas with more ECEC facilities—and government policy was necessary to meet future demand (Matsushima 2015). Thus, the Japanese government increased the number of kindergartens and nursery schools. First, the government increased the number of kindergartens under the First Kindergarten Education Revival Plan in 1963. Then, in 1966, it decided to expand nursery schools by adopting the First Nursery School Emergency Maintenance Five-year Plan. However, the expansion was insufficient to meet the demand. Therefore, the government adopted new plans, including the Second Kindergarten Education Revival Plan and the Second Nursery School Emergency Maintenance Five-year Plan in 1971. Because of these expansions, the childcare system’s enrolment rate increased dramatically (Matsushima 2015). However, there were large regional differences in the expansion rate and, consequently, the enrolment rate differed, as shown in Figs. 3 and 4. Figure 3 shows the mean of the sum of the rate of enrolment in kindergartens and the ratio of capacity to the population in nursery schools at age four across all Japanese regions from 1960 to 1989. Besides the difference in initial values, there are differences in the slopes of the lines, indicating that the rates of increase differed regionally. Figure 4 shows the changes in the rate of enrolment in ECEC at age four from 1961 to 1984.Footnote 34 There is no systematic pattern in the expansion rate. Therefore, I exploit these variations. The ministries in question tried separating kindergartens from nursery schools based on their purposes, namely, education for kindergartens and social welfare for nursery schools (Takada 2014; Okada 2014). However, as their functions overlapped in many ways, the two entities’ roles gradually became similar (Akabayashi and Tanaka 2013; Shwalb et al. 1992). Therefore, I analyse the effects of ECEC rather than those of kindergartens and nursery schools separately.Footnote 35

The variations in ECEC expansion look plausible based on the above discussion. However, this would fail to work well if exclusion restriction did not hold. One of the largest concerns would be the case where some specific areas were eager to invest in education. If this were the case, expansion of ECEC and that of higher education, including universities and colleges, would be highly correlated. Also, the quality of education in these areas might be higher. These violate the assumption of exclusion restrictions. However, in my data, this is not likely to happen: the correlation between the expansion in ECEC and that of universities and colleges is 0.061. Hence, this potential threat seems minimal. I also check whether exclusion restriction holds by conducting a control experiment. See Section 7.2 in the result section for details. Furthermore, I examine the determinant of ECEC expansion quantitatively. Although the government prioritised the areas where the capacity was scarce, the exclusion restriction would fail if there were a correlation between the growth in the ECEC availability and important labour market characteristics, such as the fraction of female labour supply. As discussed in Section 6, I follow Cornelissen et al. (2018) and investigate the factors and confirm that the government increased the availability of ECEC based on the initial situation.

3.3 Gender inequality in the education system

Since this paper examines the heterogeneous effects of ECEC concerning gender, it is worth noting briefly the gender difference in the education system in the 1960s to 1980s in Japan. In Japan, at that time, there existed considerable gender inequality. Although ECEC availability was similar between boys and girls, as shown in Fig. 5, other situations differed significantly. Regarding higher education, men’s enrolment rates were much higher than those of women in the 1950s to 1980s, as shown in Fig. 6. This partly reflected the atmosphere at that time—that, as women should stay home to raise the children and do housework, they do not necessarily require higher education. This culture was also reflected in the working environment. Women’s employment rate at around 50% was far lower than that of men at about 80%, according to Census. Furthermore, as Fig. 7 illustrates, wages are much lower for women than men. Based on the Basic Survey on Wage Structure collected by the Ministry of Health, Labour and Welfare, the income ratio of women to men was approximately 0.6 in the mid-1980s. This indicates large gender inequality in both the educational and working environments.

4 Data

To examine the long-term effects of ECEC expansion, I use two types of data: individual-level survey data and prefecture-level administrative data sets.Footnote 36

I use survey data from the Preference Parameters Study of Osaka University, which are collected annually since 2003, to create a panel data set.Footnote 37 In 2004, 2006, and 2009, Osaka University added new observations to the original panel data to compensate for the small sample size owing to the survey’s attrition. This data set contains samples of individuals aged 20 to 69 years; the questionnaire’s response rate is 60–70% for new samples and 70–95% for repeated surveys. I use data collected in 2009 and 2012 and focus on people born in 1960 to 1989. The survey in 2009 includes a retrospective question on whether respondents went to a kindergarten or nursery school when they were at each age from zero to five. It elicits information on their socioeconomic outcomes, including annual income, working status, working hours, wages, health-related behaviours such as smoking and drinking, education histories, parents’ education levels, and parents’ ages when they became adults; their subjective family financial background when they were young; and the prefectures in which they lived when they were aged 15 years.Footnote 38

I omit observations if they do not have data on the ECEC enrolment at age four and its propensity score. Then, we have 2,647 observations remaining, which is 91% of the total observations in the data set of those born between 1960 and 1989. To maximise the sample size, if there are missing data in the control variables, I substitute them with 0 and make the dummy variable, indicating the substitution. In all regressions, I control for the dummy variables. Therefore, there is no contamination owing to missing data. While all remaining observations have educational information, some do not have data on income, and the number of remaining samples is 1,915. Therefore, I examine if there is any systematic tendency between ECEC enrolment and missing data in income; if there is, the main analysis would be biased. However, based on Table 15 in Appendix A, I cannot observe such a tendency for the age four analysis. This implies the estimation is not biased owing to attrition.

The 2012 data contain information on socio-emotional outcomes, such as the Big Five.Footnote 39 Since the collection years differ between the main data set (in 2009) and the data set for Big Five (in 2012), the problem of attrition arises. Table 15 shows the result of examining whether there are systematic differences between children who answered Big Five outcomes and those who did not in terms of ECEC enrolment. In the main analysis, I focus on the analysis of ECEC enrolment at age four. Hence, this result implies there is no attrition problem. As the ECEC enrolment data are retrospective, it is necessary to discuss the extent to which people could remember their enrolment status accurately. To address this problem, I compare the enrolment rate of the national data with that of the survey data. Table 16 in Appendix A presents the analysis results. These numbers are similar, confirming that using retrospective data is a suitable approach.

Tables 1 and 2 show summary statistics for enrolment rates across ages, participants’ socioeconomic outcomes and characteristics, and the Big Five indexes. Table 1 shows that the ECEC enrolment rate is increasing as the age increases, which is consistent with the fact that the government prioritised the enrolment of children at an older age. In the analysis, I focus on ECEC enrolling status at age four because of its wide variation.Footnote 40Footnote 41 78% of the observations went to ECEC in my sample. Table 2 shows the average of outcomes and other characteristics. I also show the difference between the values of the control group (children without ECEC) and the treatment group (children with ECEC) with the t-test results. We can see some differences between them in their education levels, for example.

Table 1 Summary statistics in enrolment status
Table 2 Summary statistics for socioeconomic outcomes and big five components

Potential caveats to the data are worth noting. Although the questionnaire used is unique, the sample size is smaller than that of recent empirical works. Therefore, estimated standard errors are larger, and, consequently, the estimations are less accurate, although the data are nationally representative. Moreover, the data are not panel data and contain subjective records, which could reduce accuracy, although I discuss these problems later.Footnote 42

I also use administrative data sets collected by some Japanese ministries to construct the instrumental variables (IV) for the main and mediation analyses, and to check validity. First, I use Census for 1960–1990 collected by the Ministry of Internal Affairs and Communications every five years. Particularly, I use the information on the population of each cohort born from 1960 to 1985 to create the variable on the fraction of people enrolled in ECEC, in addition to that of a college for the mediation analysis. Because the data are for only every five years, I interpolate the in-between data to create a panel data set.Footnote 43Footnote 44 I also use data on the population of women aged 20–35 years and the number of employed women in the same age range, when they are likely to quit their job and give birth. These data are used in robustness checks to test whether demand for ECEC led to its supply. Moreover, I use the information of each prefecture on population, the number of middle-educated and high-educated people, and the fraction of female labour force for the analysis on the determinant of the ECEC expansion. This analysis is done together with the data on the the mean wages from the data set called the Historical Statistics of Japan, collected by the Ministry of Internal Affairs and Communications.

Second, I use the Survey of Social Welfare Institutions collected by the Ministry of Health, Labour and Welfare. This data set describes nursery schools’ capacity in each prefecture each year. I use data for 1960–1989. However, these data do not include full information on the children’s age. Thus, I cannot separate their ages completely. I use the data set containing the ratio of the ages of children enrolled in a nursery school at the country level. I assume that the ratio is the same across prefectures, and I use it to calculate the capacity for each age in each prefecture.

Third, the analysis is based on the School Basic Survey data set collected by the Ministry of Education, Culture, Sports, Science, and Technology (MEXT). These data contain information on Japanese schools, including the number of children in kindergartens at each age in each prefecture. I use the data for 1960–1989. This survey is also used to calculate the college enrolment rate. I use the number of children enrolled instead of capacity because data are available for the former but not the latter. This variable can be used as an IV, and its validity is discussed in the section on the identification strategy.

Fourth, to check the validity of the IV, the discussion on identification is partly based on the National Survey on Migration in the Annual Population and Social Security Surveys collected by the National Institute of Population and Social Security Research. I use data on the migration rate and reasons for migration collected in 1996.

Finally, for mediation analysis, I use the list of colleges collected by MEXT, which contains data on the capacity of national, public, and private colleges. I use the data for 1980–2005, which corresponds to people at the age of 18 who were born between 1962 and 1987. I use these data to create the college capacity variable for the IV for college completion.

5 Empirical model

To examine the effect of ECEC on people’s outcomes in adulthood, this study uses the following model to regress these outcomes on enrolment in the childcare system:

$$\begin{aligned} y_{ijT} = d_{ijt}\gamma + \varvec{x}_{it}^{\prime } \varvec{\beta } + f(\mu _j, \nu _t) + \varepsilon _{ijT}, \end{aligned}$$

where \(y_{ijT}\) is the outcome of individual i who lived in prefecture j at age 15, which is measured in year T; \(d_{ijt}\) is the ECEC enrolling status dummy (enrolled = 1) of individual i in prefecture j at year t (at age four); \(\varvec{x}_{it}\) is individual i’s characteristics at year t (at age four), including time-invariant variables, such as gender; \(\mu _j\) is the prefecture fixed effects; \(\nu _t\) is the cohort fixed effects; \(f(\mu _j, \nu _t)\) is a function of fixed effects; and \(\varepsilon _{ijT}\) is an error term.Footnote 45

The controls can include other variables, such as parents’ education levels, parents’ ages, number of siblings, and gender dummy depending on the specification. As stated above, if any variable is missing, I substitute it with 0 and include the dummy variable indicating this substitution as another control. I assume that people did not move from where they were born until age 15. I discuss this point in detail in the subsection on the identification strategy. If \(y_{ijT}\) is a binary variable, such as an indicator of college completion, the model is a linear probability model.Footnote 46

However, there may be omitted variable bias because parents’ decisions to enrol their children in ECEC could be highly correlated with their unobservable household characteristics. Therefore, I adopt an IV analysis and obtain a two-stage least squares estimator.

The IV here is the propensity score of enrolment in ECEC (i.e., probability of enrolment in ECEC), constructed as follows. The validity of this analysis using the number of children enrolled is discussed in the next section. First, define the fraction of children enrolled in ECEC, \(r_{jt}\) as

$$\begin{aligned} r_{jt} := \frac{n_{jt}}{N_{jt}}, \end{aligned}$$

where \(n_{jt}\) is the number of children enrolled in ECEC, a nursery school or a kindergarten, in prefecture j in cohort t, which is almost the same as ECEC capacity when the capacity is binding, and \(N_{jt}\) is the population of prefecture j in cohort t. One potential problem is that data on the capacity of kindergartens and nursery schools for children at age four are not available. However, regarding kindergartens, the number of children enrolled at each age is available. Therefore, given that the capacity constraints are binding based on Fig. 1, in the main analysis, I use the number of children enrolled in kindergartens at the age of four. For nursery schools, since age-level data on both the capacity and number of children enrolled are not available, I construct the capacity ratio of children at age four by multiplying the total capacity by the ratio of the number of children enrolled at age four with the total number of children enrolled in a nursery school. I assume that this ratio is the same across Japan over time. However, this assumption is not crucial because I also control for prefecture fixed effects and cohort fixed effects in the probit model below. Hence, any measurement error among prefectures and cohorts is absorbed by them.Footnote 47

Given this proportion \(r_{jt}\), I estimate the propensity score using a probit model:

$$\begin{aligned} d_{ijt}^*= & {} \varvec{x}_{0i}^{\prime }\varvec{\lambda } + r_{jt}\phi + g(\eta _j, \psi _t) + e_{ijt}, \\ d_{ijt}= & {} \left\{ \begin{array}{ll} 1 &{} \text {if } d_{ijt}^* > 0 \\ 0 &{} \text {otherwise}, \end{array}\right. \end{aligned}$$

where \(d_{ijt}\) is defined above, that is, the ECEC enrolment status dummy (enrolled = 1) of individual i in prefecture j at year t (at age four); \(d_{ijt}^*\) is its continuous latent variable; \(\varvec{x}_{0i}\) is individual i’s gender; \(\eta _j\) is the prefecture fixed effects; \(\psi _t\) is the cohort fixed effects; and \(e_{ijt}\) is an error term, which is assumed to be independent and identically distributed (i.i.d.) over a normal distribution with mean 0 and variance 1. \(g(\eta _j, \psi _t)\) is a function of fixed effects.Footnote 48 Define \(\hat{z}_{ijt}\) as the probability of enrolment in ECEC for individual i in prefecture j in cohort t:

$$\begin{aligned} \hat{z}_{ijt}\equiv & {} \Pr \left( d_{ijt} = 1 | \varvec{x}_{0i}, r_{jt}, \eta _j, \psi _t \right) \nonumber \\ {}= & {} \Phi \left( \varvec{x}_{0i}^{\prime }\varvec{\hat{\lambda }} + r_{jt}\hat{\phi } + g(\hat{\eta }_j, \hat{\psi }_t) \right) , \end{aligned}$$

where \(\Phi (\cdot )\) is the standard normal cumulative distribution function, and the hats over the variables show their estimates.

Although \(r_{jt}\) itself can be used as an IV, I use \(\hat{z}_{ijt}\) as the IV in my main analysis instead, because it provides a more efficient estimation (Wooldridge 2010). Accordingly, in this study, the LATE is the average user’s LATE over the marginal treatment effects.Footnote 49Footnote 50

Based on the above models, I consider the first- and second-stage regressions as

$$\begin{aligned} d_{ijt}= & {} \hat{z}_{ijt}\theta + \varvec{x}_{it}^{\prime }\varvec{\kappa } + f(\mu _j, \nu _t) + \epsilon _{ijt},\end{aligned}$$
$$\begin{aligned} y_{ijT}= & {} \hat{d}_{ijt}\tilde{\gamma } + \varvec{x}_{it}^{\prime }\tilde{\varvec{\beta }} + f(\mu _j, \nu _t) + u_{ijT}, \end{aligned}$$

where \(\epsilon _{ijt}\) and \(u_{iT}\) are error terms independent of the other controls, \(\omega _{jt}\) is prefecture-specific time trends, and \(\hat{d}_{ijt}\) is the fitted value of \(\hat{d}_{ijt}\) based on Eq. (4). The parameter of interest, \(\tilde{\gamma }\), is estimated using two-stage least squares (2SLS) based on Eqs. (1), (4), and (5).Footnote 51 If ECEC has a positive effect on outcome Y, coefficient \(\tilde{\gamma }\) is positive.Footnote 52

Moreover, I show the results based on the reduced-form analysis, which is based on the following regression model:

$$\begin{aligned} y_{ijT} = \hat{z}_{ijt}\delta + \varvec{x}_{it}'\varvec{\varphi } + f(\mu _j, \nu _t) + \zeta _{ijT}, \end{aligned}$$

where \(\zeta _{ijT}\) is an error term independent of the other controls.

6 Identification issues

Here, I discuss this study’s empirical strategy. If I examine the causal effect of ECEC using ordinary least squares (OLS) as Eq. (1), the estimator is likely to be biased because there may be a correlation between unobserved household characteristics and parents’ decisions on children’s ECEC enrolment. Therefore, an IV estimation is a possible way to estimate the parameters of interest consistently.

As an IV, I exploit the change in the expansion of capacity and number of children enrolled in nursery schools and kindergartens (see Section 3) at the prefecture level during the 1960s to 1980s. I then construct the propensity score for whether a respondent went to an ECEC facility. Therefore, the key assumption of my identification is that variations in expansion after controlling for households’ observable characteristics are quasi-random.

Table 3 Correlation between change in enrolment rate and change in birth rate in previous period
Table 4 Exogeneity check
Table 5 Determinants of the ECEC Expansion

A potential problem is that the extension of ECEC is not quasi-random. One possible scenario is that expansion of the number of spaces in the childcare system is correlated with some unobservable prefectural characteristics and some time trends. To address this problem, I control for area fixed effects, cohort fixed effects, and their interactions in the regression. Moreover, I use a prefecture-level variation in ECEC availability so the estimate would be biased if there were household heterogeneity within the same prefecture. To eliminate this potential bias, I control for observable household characteristics, such as parents’ education levels, parents’ ages, and the number of siblings. Parents’ education levels are often used to identify disadvantaged children. For mothers with a low education level, childcare use reduces stress and increases well-being (Yamaguchi et al. 2018b). Therefore, parents’ education level may matter and should be controlled for. Similarly, if parents are young, they might lack knowledge about raising their children; thus, their children might be disadvantaged. Hence, it may be important to control for this factor. Finally, the effect may differ if a child has siblings: If there are many siblings, parents might be unable to spend sufficient time with each child and choose to use ECEC. Therefore, I control for these variables. Household income may matter; however, data on household income when children were aged four years are unavailable. Instead, I use data on categorised subjective wealth at age 15 years. However, this control may be of low quality, since this variable may be affected by the choice of preschool. Therefore, in the main analysis, I do not control for subjective wealth. However, I do so in the robustness check in Section 7.2. I make a dummy variable for whether the individual’s subjective wealth was above or below average, assuming that respondents remember their family status at least roughly and that their family background did not change significantly.Footnote 53 The robustness check shows that the results do not change.

I also check whether the supply constraints were likely to be binding. As discussed above, Figs. 1 and 2 illustrate the lack of space for additional children in both kindergartens and nursery schools.Footnote 54Footnote 55 Furthermore, Table 3 shows the regression results for growth in the number of children enrolled in ECEC on the change in the number of births in the previous period at the prefecture level. This shows that, although the government might have increased ECEC availability when demand was high, coefficients of the explanatory variables are less than one, implying that ECEC was not available to all children (i.e., the supply constraint was binding).Footnote 56

Furthermore, Niimi (2002) states that local governments have difficulty finding private companies to operate a childcare system owing to the high initial costs. Moreover, many children did not attend kindergartens because of binding capacity or lack of nearby kindergartens (Minister of Ministry of Education’s Secretariat Survey Division 1972; Motoki and Yamanishi 2009; Matsushima 2015). The government recognised the lack of ECEC in Japan in their discussion in the Diet and a committee (House of Councillors, The National Diet of Japan 1967b, a).

These facts suggest that the supply-side constraints were binding, and, therefore, the increase in enrolment rate can be a good candidate of an instrumental variable.

I further examine whether increases in demand for ECEC drove its supply at the prefecture level. If this were true, the expansion of ECEC could be endogenous, which would lead to a bias in the estimation. First, as I explained above, the government expanded ECEC facilities to support the female labour supply in Japan. However, the places where the government increased ECEC facilities do not seem to be chosen because of the demand. According to Matsushima (2015), the expansion had been done to mitigate the imbalances in the supply of ECEC facilities. More concretely, the government tried to fill this imbalance and prioritise areas based on the ratio of the capacity of ECEC facilities to the population of children, or the number of facilities per population (Matsushima 2015).Footnote 57 Together with the fact that I control for area and cohort fixed effects in the estimation, although this is not perfectly done because of the collinearity, this largely mitigates the potential endogeneity problems. Besides, this hypothesis is partially testable, so I analyse whether the demand of ECEC drove its supply. The most important factor behind the ECEC expansion is changes in the female labour supply. An increase in female labour supply at age 20–35, which corresponds to a woman’s typical childbearing age, might increase demand for ECEC, thus inducing ECEC expansion. Therefore, I check whether an increase in female labour force participation in each prefecture can predict a future increase in ECEC capacity and the number of children enrolling. If so, demand would have driven supply. Specifically, I run the regression of the difference in ECEC availability on the difference in the female employment rate in the previous periods. Table 4 shows the results. I vary the time difference between the ECEC expansion and measurement timing of female labour supply, as it may take time to open a new ECEC facility and the supply may take time to reflect demand. However, the coefficients are not significant. This finding implies that an increase in female labour supply cannot predict future ECEC expansion. Thus, it is likely that demand did not drive supply.Footnote 58

Besides, as mentioned in Section 3, I check quantitatively whether the ECEC expansion was based on the initial coverage rate of the ECEC, i.e. whether the government prioritised the construction of ECEC facilities where the coverage rate was low. If there were a correlation between the growth in the ECEC availability and important labour market characteristics, such as the fraction of female labour supply, the exclusion restriction would fail. Following the analysis of Cornelissen et al. (2018), I regress the growth in the ECEC enrolment rate of the age of four from 1960 to 1985 on the coverage rate of ECEC of the age of four in 1960 and other covariates in 1960. Table 5 shows the result.Footnote 59 It confirms that ECEC expantion was based on the initial coverage rate and shows that other seemingly important factors, including the fraction of female labour supply, education level, population, and the mean wage, did not affect the expansion.Footnote 60Footnote 61

Fig. 8
figure 8

Female enrolment rate in a university across regions in Japan. This figure is from Ueyama (2011) with labels translated by the author from Japanese into English. The table does not include women who attended a two-year college

Fig. 9
figure 9

Propensity score of enrolling in ECEC at age four

Another potential threat to the violation of the exclusion restriction would be a correlation between a prefecture’s investment in ECEC and that in higher education. As discussed in Section 3, if a prefecture is eager to invest in education, its quality of education might be high. That is, the expansions of ECEC and later education, such as universities and colleges, might be correlated. If so, the assumption of the exclusion restriction would be violated. However, in my data, the correlation between the ECEC expansion in enrolment/capacity and college expansion in capacity is 0.061. Hence, this potential threat seems minimal.Footnote 62

Moreover, I check whether the differences in education levels across areas could explain the speed of ECEC expansion. As discussed earlier, the education level of parents, especially mothers, might matter for children’s enrolling in ECEC. Figure 8 from Ueyama (2011) shows the fraction of women who enrolled in a university: there are almost no differences in the enrolment rate and its expansion speed before the 1990s, implying that education level does not drive ECEC supply.

In the model section, I assume that people did not move from their place of birth until age 15. This assumption might seem extreme: I discuss its validity and the possible bias if the assumption were to be violated (i.e., if many people moved between the ages of 4 and 15). There are two types of movements: across prefectures and within the same prefecture. If there are many observations of the first type, the IV regression estimator would be biased. If movement occurred randomly, there would be attenuation bias. Suppose that these movements were not random. Note that the IV for enrolling status is the propensity score of ECEC enrolling, implying that its coefficient in the first stage should be positive. In this case, the bias might be upward if people moved from a prefecture with a high enrolment rate to a prefecture with a low enrolment rate, or vice versa. However, according to the National Survey on Migration in the Annual Population and Social Security Survey collected in 1996, over 80% of people remained in their birth prefecture for 15 years. Although this does not correspond exactly to the main survey’s data years and does not include only data on children at the aged four, the overall data imply that this problem is not severe enough to invalidate my estimation.

The second case matters if people moved to seek ECEC availability. As discussed earlier, the supply constraint of enrolment/capacity in the childcare system seems to have been binding. Therefore, it is unlikely that people migrated to seek ECEC availability. Accordingly, the assumption on non-migration is likely to hold.

As I use the propensity score, it is necessary to check whether the assumption of common support holds. Figure 9 shows the probability density distribution of the propensity score for people with and without enrolling in ECEC, and the assumption is likely to hold.Footnote 63 I also assume the monotonicity assumption for the IVs, as I use the propensity score as an IV.

7 Results

7.1 Main results

Here, I discuss the estimation results of the OLS regressions, reduced-form analysis, and IV analysis discussed in Section 5. As stated, I focus on enrolment at age four and show the results of the effects of enrolment in the entire childcare system rather than nursery schools and kindergartens separately.Footnote 64 I drop observations that do not answer questions of interest. See Appendix A for the sample selection. However, there are no systematic attrition patterns.

As shown in Table 23 in Appendix D, the first stage results are sufficiently strong for IV regressions, including those of the subsample analyses, regardless of the specification. As stated in Section 6, the assumption of the exclusion restriction is likely to hold. Therefore, I use this IV.

Table 6 shows the OLS results based on Eq. (1); those of the reduced form based on Eq. (6); and those of the two-stage least squares estimators based on Equations (1), (4), and (5). As shown in the table, enrolling in ECEC increases children’s future income by approximately 44% and raises college completion rate by 38 percentage points (both statistically significant).Footnote 65Footnote 66

Table 6 Effects of ECEC at Age 4
Table 7 Heterogenous effects of ECEC at age 4

The value of 44% may seem excessive. However, compared with studies showing the short- and long-term effects of ECEC, the estimate is quite reasonable. According to Yamaguchi et al. (2018b), ECEC develops children’s language skills by 0.7 standard deviations and reduces the tendency of inattention and hyperactivity by 0.4. I conduct the same analysis with the normalised measure of income with mean 0 and standard deviation 1, and I find that ECEC increases the income by 0.5 standard deviations. Furthermore, Heckman et al. (2013) report that the Perry Preschool Project increases the income of the enrolled children 1.5 times by age 27 compared with those who did not enrol. My estimation is quite comparable to theirs. Moreover, Havnes and Mogstad (2011b) report the long-term effects of subsidised childcare on years of schooling; their estimated TT effect shows an additional 0.35 years of education per childcare place. This is smaller than my estimation. One possible explanation is that their expansion of ECEC was from 10% to 28% of the enrolment rate, which is smaller than my variation; therefore, mine might include more disadvantaged children who benefit more from ECEC. An alternative explanation is that they estimate TT effects, while I estimate LATE. The former includes those who did not go to ECEC in the treatment group and vice versa. Therefore, given that ECEC has positive effects, the estimates may be smaller.

The OLS estimators are smaller than IV estimators. This implies that unobserved household characteristics are negatively correlated with enrolling in ECEC. One possible explanation is that relatively disadvantaged families rely on ECEC because these parents are willing to work to earn. Moreover, ECEC facilities’ environment may be better than the home environment because of the minimum requirements required to operate ECEC. As discussed later in the subsection of the robustness check, the result of reduced form shows a similar tendency, although the estimation is smaller because of the specification of the reduced form Angrist and Pischke (2008), and the point estimation is comparable to the result of Havnes and Mogstad (2011b).

I further conduct analyses of the effects of ECEC on children’s working status (i.e., whether they work), working hours, and wages. The result is in Table 17 in Appendix B.Footnote 67 The table shows that there is no effect on the working status and working hours, but there is on the wage. According to this, the interpretation of the effects of ECEC is that if a child enrols in ECEC, that child is more likely to finish college and earn 44% more through an increase in wages. This result is consistent with human capital accumulation theory (Mincer 1974; Becker 1994; Kane and Rouse 1995; Thomas 2003).Footnote 68Footnote 69

Table 7 shows the results of the subsample analyses focusing on gender heterogeneity. This table shows no effects on income and college completion likelihood for men. On the contrary, there is a large and statistically significant effect of enrolment in the childcare system on future income and college completion probability for women. This means that most of the effects in Table 6 come from those on women’s outcomes. As discussed earlier, women’s enrolment rate in higher education is much lower than that of men (Fig. 6). Moreover, the wage gap is very large (Fig. 7). This implies that there was considerable room for increasing the enrolment rate and wages.

As discussed in Section 3, although the enrolment rates in ECEC were similar in both genders, women were disadvantaged in enrolling in higher education and wages in adulthood. It seems that these disadvantages at least partially reflected the cultural perspective that a woman should stay at home and care for her family. Given this situation, the result can be interpreted as that the effects are stronger for disadvantaged children. This is consistent with the analysis of the short-term effects for disadvantaged children by Yamaguchi et al. (2018b), although they define disadvantaged children as those with less-educated mothers.Footnote 70 These results imply that ECEC can reduce the inequality between the advantaged and disadvantaged by raising the outcomes for disadvantaged children to a greater extent. This also implies that ECEC enrolment could break the intergenerational poverty chain because this early childhood intervention reduces the inequality in their adulthood. The finding that the effects for women are larger is consistent with the results of the shrinking wage gap owing to ECEC reported by Havnes and Mogstad (2011b).

Furthermore, I investigate the marginal treatment effects. Figure 11 in Appendix D shows that the children who are likely to enrol in ECEC in terms of the unobserved heterogeneity benefit from ECEC more. The effect is significantly positive for around 40% of the children, while there is no negative impact uncovered for other children.

These results are robust based on the robustness check discussed in the next subsection.

7.2 Robustness checks

Here, I examine whether the results shown in the previous subsection are robust, by conducting further analyses. The analyses include another type of estimation, other specifications of the propensity score estimation, a slightly different sample, and other specifications of the main regression equations. The results are summarised in Table 8.Footnote 71

Table 8 Robustness Checks
Table 9 Control experiment

First, I conduct an analysis based on the reduced-form approach, which requires fewer assumptions.Footnote 72 If one assumption for the LATE identification used in the main analysis, which is the variation in ECEC availability, is exogenous, the reduced-form approach is also valid. This is the case even if any additional assumptions on the LATE identification were to fail, such as any correlations between parental characteristics and prefecture-level exogenous variations in the rate of expansion. Column (1) in Table 8 presents the results from the reduced-form approach. The results show a lower point estimation, as discussed by Angrist and Pischke (2008) and predicted in Section 1, but have a similar result and consistent with the main results.Footnote 73

Next, I use two additional specifications of the propensity score estimation to estimate the same effects. In the main analysis, I use a model with controls for gender, prefecture, area fixed effects, and a standard error based on the prefecture level. Another specification is based on the Probit model, but the controls are different. Since the IV must be a function of exogenous variables, there is an infinite number of candidates. Here, I control for the cohort (every five years) fixed effects, area fixed effects and their interactions, whereas I use cohort fixed effects and prefecture fixed effects in the main analysis. Another specification is based on the linear probability model instead of the Probit model. Column (2) of Table 8 shows the result with different fixed effects, and the results from the linear probability model are in Column (3). Table 28 reports coefficients of the regressions and predictions of the propensity scores under different specifications (see Appendix A). Based on these analyses, the results are robust.

Third, I omit observations in the domain when either of the propensity scores is not positive. This ensures that the analysis is based on a common support. The result shown in Column (4) of Table 8 is also robust.

Fourth, I add more controls to the main regression equation. In the main analysis, I already include parental education levels, number of siblings, and parental age. However, other variables can also represent household background, such as the retrospective subjective wealth index (0 is the poorest, and 10 is the wealthiest) at age 15. While this categorisation can change depending on the treatment status, it may be an informative indicator of household wealth. Therefore, I create a dummy variable equal to one if the index is less than five, and zero otherwise. This is because it is less likely to change depending on the treatment status than the original 11-scale index. Other variables could be considered, such as prefectures in which the individuals lived at age 15 and the age of children. However, the inclusion of these variables in the regression may cause a problem because, as noted in Section 6, the identification relies on the variation across prefectures over time. Therefore, it is not possible to control for both age and prefecture simultaneously. Given that the cohort fixed effects and area fixed effects are already controlled for, additionally controlling for age (and its polynomial) and prefecture might dramatically reduce the sample size in each age-prefecture cell (some of them have only one observation), markedly deteriorating the estimator’s accuracy. Therefore, I show the regression results including age only. I conduct an analysis with the variables on age and its polynomial up to three, where each cell has more observations than cells with variables on prefectures. The result in Column (5) of Table 8 is robust.

Fifth, similar to Duflo (2001) discussed in Section 2, I control for ECEC intensity, which is the initial capacity/enrolment of ECEC of each prefecture, and the interaction of this and cohort fixed effects. I do so because a potential threat of the identification is that children with initially more availability in ECEC could enrol in ECEC, as the prefecture was eager to invest in ECEC, which is correlated with other unobservable characteristics. I check this by controlling the initial availability and interactions between these and cohort fixed effects. However, as shown in Column (6), the coefficient is comparable, implying that this type of eagerness may not matter much.Footnote 74

Sixth, I omit samples who enrolled in ECEC at age three to examine the marginal effects of ECEC at age four. In addition, I omit samples who went to a nursery school or kindergarten at age two and endogenise the choice of enrolment at age three. I then estimate the effects of enrolment at ages three and four separately using a similar IV for enrolment at age three. Columns (7) and (8) of Table 8 report results of the former and latter analyses, respectively, which support the results’ robustness.

Seventh, I use different clusters from those in the main analysis. In the main analysis, I use clusters at the level of the area, cohort, and their interaction. Column (9) of Table 8 presents results based on the prefecture-level standard error, which is only slightly higher than the standard error in the main analysis.

Eighth, as discussed earlier, the data on income are categorised, and I use class values of each category in the main analysis.Footnote 75 In the analysis, I use each category’s median as the income variable. I assume that the difference between individuals’ original income and reported value (median value of the category) is i.i.d. Although the standard errors become larger, as shown in Table 6, the estimator remains positive and significant, and enrolment in the childcare system is likely to positively affect income. As a robustness check, I also conduct an interval analysis, which considers this data-coding problem (Cameron and Trivedi 2005). Since the objective function is the IV regression, not the OLS, I adopt the control function approach to eliminate potential endogeneity (Cameron and Trivedi 2005). The result in Column (10) of Table 8 supports the main results’ robustness.

Ninth, in some cases, kindergartens might accept more students than the capacity (Fig. 1. So I conducted an analysis by dropping such observations for children who went to a kindergarten. The result shown in Column (11) implies that the magnitude of the effects is comparable, although it becomes insignificant partly because of the smaller sample size. On the other hand, there were several cases where the capacity was larger than the enrolment, while the capacity was actually binding (Matsushima 2015). This likely happened because the timings of the report of enrolment and capacity are different owing to the construction of statistics, which is the basis of my data. In some cases, children left before the time of counting the number of children enrolled. I omitted samples who went to a kindergarten whose capacity exceeded the number of children enrolled. The result is in Column (12). This made the sample size very small and the estimation imprecise, leading to an insignificant estimation. However, I test whether the estimate is different from the value of the point estimate of the main analysis and cannot reject the hypothesis that they are different, which supports that this problem does not seem to matter.

Finally, I conduct a control experiment following Duflo (2001), which I discussed in Section 2. Particularly, I conduct a reduced-form analysis wherein the main independent variable is the ECEC expansion three years after the relevant period for the cohort. Therefore, this expansion was not beneficial for those cohorts, and the coefficient should be zero.Footnote 76 Table 9 reports the results, showing that the coefficients are not statistically different from zero. This result supports the exclusion restriction and the validity of the main result in this section.

8 Potential mechanisms

8.1 Mediation analysis

Here, I analyse the extent to which the effects on future income can be explained by the increase in college completion induced by enrolment in childcare systems. I conduct a mediation analysis for this purpose. This analysis is common in the psychological literature and some social science fields (Baron and Kenny 1986; Imai et al. 2010; Hayes 2018). In the basic model of mediation analysis, if my treatment is endogenous, two IVs would be needed: for the treatment variable and the mediating variable. Therefore, I create an additional IV. The regression in the simple mediation analysis is as follows:

$$\begin{aligned} y_{ijT} = \varvec{x}_{it}^{\prime } \varvec{\beta } + d_{ijt}\gamma + m_{ijt^{\prime }}\alpha + f(\mu _j, \nu _t) + \varepsilon _{ijT}, \end{aligned}$$

where \(y_{ijT}\) is the outcome at T, such as logarithm of income; \(d_{ijt}\) is a dummy for the treatment, such as childcare enrolment; and \(m_{ijt^{\prime }}\) is a mediating variable, such as college completion, where \(t^{\prime } = t + 14\).Footnote 77Footnote 78Footnote 79 I denote \(d_{ijt}\) as childcare enrolment (enrolled = 1) and \(m_{ijt'}\) as college completion (completed = 1). \(\gamma\) in Eq. (7) captures the direct effects. Neither \(d_{ijt}\) nor \(m_{ijt^{\prime }}\) is exogenous, and, therefore, I need two IVs to estimate the coefficients consistently. I assume that the effect is homogeneous. That is, I do not permit treatment–mediator interaction effects and, thus, heterogeneity in direct and indirect effects. This follows Burgess et al. (2014). As discussed in Footnote 78, I find no such effect. However, I loosen this assumption later by partly using a reduced-form approach.Footnote 80

If most ECEC effects on income could be explained by college completion owing to childcare when young, the coefficient of \(d_{ijt}\), \(\gamma\), would become insignificant once \(m_{ijt'}\) is inserted into the model. Otherwise, there would be channels from childcare to income other than the path of college completion. This might include the effects on non-cognitive abilities. However, when comparing coefficients, this method requires caution. The coefficients in an IV analysis are often interpreted as LATE. If I were to use two IVs, the support could be different. Here, I focus on people who change their behaviours based on the IVs.

As in the previous analysis, I assume that people did not move to another prefecture after age 15, which is also asked about in the survey. However, according to the National Survey on Migration in the Annual Population and Social Security Surveys collected in 1996, around 85% of people did not move from their prefecture after age 15. Therefore, this assumption is unlikely to be violated.

First, I discuss how I develop the additional IV. As in the main analysis, I construct the IV as the ratio of the capacity of colleges to the population in cohort t in prefecture j, \(r^c_{jt}\):

$$\begin{aligned} r^c_{jt} = \frac{n^c_{jt}}{N^c_{jt}}, \end{aligned}$$

where \(n^c_{jt}\) is the capacity of colleges in prefecture j in cohort t, and \(N^c_{jt}\) is the population in prefecture j in cohort t at age 18.

Similar to the construction of the IV for childcare enrolment, I estimate the propensity score using a probit model:

$$\begin{aligned} m_{ijt}^*= & {} \varvec{x}_{0i}'\varvec{\lambda }^c + r^c_{jt}\phi ^c + h(\eta ^c_j, \psi ^c_t) + e^c_{ijt}, \\ m_{ijt}= & {} {\left\{ \begin{array}{ll} 1 &{} \text {if } m_{ijt}^* > 0 \\ 0 &{} \text {otherwise}, \end{array}\right. } \end{aligned}$$

eqnarray where \(\varvec{x}_{0i}\) is individual i’s gender; \(\eta _j\) is the prefecture fixed effect; \(\psi _t\) is the cohort fixed effect; h is a function of the fixed effects; and \(e^c_{ijt}\) is an error term, which is assumed to be i.i.d. over the normal distribution with mean 0 and variance 1. \(h(\eta ^c_j, \psi ^c_t)\) is a function of fixed effects. \(\varvec{x}_{0i}\) can include other variables, such as mothers’ and fathers’ education levels and subjective wealth level at age 15, depending on the specification. Define \(\hat{m}_{ijt}\) as the probability of enrolment in ECEC for individual i in prefecture j in cohort t:

$$\begin{aligned} \hat{m}_{ijt}\equiv & {} \Pr \left( m_{ijt} = 1 | \varvec{x}_{0i}, r^c_{jt}, \eta ^c_j, \psi ^c_t \right) \nonumber \\ {}= & {} \Phi \left( \varvec{x}_{0i}'\varvec{\hat{\lambda }}^c + r^c_{jt}\hat{\phi }^c + h\left( \hat{\eta }^c_j, \hat{\psi }^c_t\right) \right) , \end{aligned}$$

where \(\Phi (\cdot )\) is the standard normal cumulative distribution function.

Following the discussion on the IV for childcare enrolment, although \(r^c_{jt}\) can be an IV, I use \(\hat{m}_{ijt}\) as the IV in my main analysis, because it provides a more efficient estimation (Wooldridge 2010).

8.2 Identification and result of the analysis of the mechanisms of ECEC’s long-term effects

For the mediation analysis, I construct a new IV: ratio of college capacity to the cohort’s population at age 18. Because of data restrictions, I use data on the cohorts born between 1962 and 1985. In Japan, most people enter college at age 18. Thus, I focus on the period from 1980 to 2003. Before the Private Schools Act was reformed in 1974, private universities required prior approval from the government to increase college capacity (Kikuchi 2017). In addition, before 2003, the School Education Act required that all universities obtain government approval before opening and closing departments, with private universities also needing approval to change capacity at the department level. The permission process included a one- or two-year screening and investigative process, such as interviews by the MEXT based on the Standards for Establishment of Universities. Hence, changing capacity was difficult (Kikuchi 2017). Therefore, I use the proportion of the cohort’s population capacity as the other IV.

Table 10 Effects of ECEC at Age 4: Mediation analysis
Table 11 Heterogeneous effects of ECEC at Age 4: Mediation analysis

As Tables 24 and 25 in Appendix D show, IVs for the first stage result are not jointly weak regardless of the specification. Based on these valid IVs, Table 10 shows the second stage results. Columns (1) and (3) show the results of the same analysis in Table 6, except that I omit samples without information on the propensity score for college enrolment.Footnote 81 As presented in Table 10, the effect of college completion is positive and statistically significant, which is consistent with the literature (Mincer 1974; Becker 1994; Kane and Rouse 1995; Thomas 2003). However, once I control for the effect of college completion, ECEC’s effect on income disappears, although the estimates remain positive in the specification without controls.Footnote 82 This is true for the heterogeneous effects, as shown in Table 11. As discussed earlier, women were disadvantaged compared to men in terms of the likelihood of college completion and wage. Because there is room for them to improve, most effects partly come from those of women.

These results imply that most of the effects of childcare systems on future annual income can be explained by the increase in college completion, especially for female children. There are also no other channels from ECEC enrolment to future income. That is, increases in non-cognitive abilities by ECEC are unlikely to increase future income directly, as the discussed literature shows. This seems consistent with the finding that there are no long-term effects on psychological skills measured by risk preference and the Big Five (Tables 18 and 19) in Appendix B. Together with the finding on working hours, the mechanism behind the increase in annual income likely is as follows: enrolment in ECEC increases the college completion rate, and children who participated in ECEC accumulate human capital at colleges or universities, which increases their wages, which leads to an increase in their annual income.

As a robustness check, I conduct an additional analysis wherein I do not impose the assumption of homogeneity. Instead, I use the propensity score for attending college as a control as follows:

$$\begin{aligned} y_{ijT} = \varvec{x}_{it}^{\prime }\varvec{\beta } + d_{ijt}\gamma + \hat{m}_{ijt^{\prime }}\alpha +f( \mu _j , \nu _t ) + \varepsilon _{ijT}. \end{aligned}$$

This partial reduced-form approach does not require assumptions on the IV for college graduation except for its exogeneity, as discussed before. The result in Table 26 in Appendix D is similar to the result shown above.

One important point that might violate the story above is that ECEC can increase mothers’ labour force participation and income significantly enough to allow children to go to college. To discuss this possibility, I conduct the following analyses.

First, I discuss briefly again whether there is any pre-trend on labour force participation Second, I examine whether there is any difference in labour force participation after introducing ECEC. Finally, I discuss whether there is any effect on their income based on subjective recall data.

If there is any systematic pre-trend in the female labour force supply before ECEC expanded, this could capture the potential supply after ECEC expanded. However, we do not observe this based on Table 4, and this is not likely to have happened.

Then, I examine whether there is an effect of ECEC expansion on the change in female labour force participation, using a similar method of examining the pre-trend of female labour force participation: I run a regression of percentage change in female employment at time \(t+j-1\) to \(t+j\) (\(j = 1, ... ,6\)) on the growth in the number of four-year-old children enrolled in ECEC at time \(t-1\) to t. Given that the expansion of ECEC can be seen as quasi-random, the change in female employment after ECEC expansion can be considered causal.Footnote 83

The result is shown in Table 21 in Appendix B. Although I find a positive effect, that is, parents were likely to work around one to four percentage points more using ECEC, the coefficient seems too small to have an economic interpretation in this study’s context: a coefficient of 0.04 implies that if ECEC availability increased by 100 percentage points, female labour force participation would increase by four percentage points. Note that, based on the two figures on capacity and enrolment, the situation was that if ECEC was available, parents used it. Given this, an overall 4% increase in income is not likely to generate a large enough difference in disposable income to explain the income effect of the ECEC effect on children’s outcomes. Even if we assume the effect is cumulative for six years, it is less than 10% of all parents. This still seems too small to explain that the effects are coming from the income effect.

The literature stated that the effect of ECEC on parents’ (especially maternal) working choice is mixed: Childcare increased maternal labour supply in some studies in the USA in the 1980s (Gelbach 2002), Argentina in the 1990s (Berlinski and Galiani 2007), Canada in the 1990s to the early 2000s (Lefebvre and Merrigan 2008; Baker et al. 2008), Spain in the 1980s to the 1990s (Nollenberger and Rodríguez-Planas 2015), and Germany in the 2000s to 2010s (Busse and Gathmann 2020). However, some reported no effect in the USA from 1950 to 1990 (Cascio 2009), the 1990s (Fitzpatrick 2010), and 1980 to 2000 (Fitzpatrick 2012); France in the 1990s (Goux and Maurin 2010); and Norway in the 1960s to 2000s (Havnes and Mogstad 2011a). In Japan, on which this study focuses, Asai et al. (2015, 2016) and Yamaguchi (2017) investigate it in the 1990s to 2010s and Yamaguchi et al. (2018) investigate it in the 2010s, but they did not find effects on maternal labour supply. Based on the literature, the effect seems small, and the role of parents’ working status and income effect induced by this seems limited.Footnote 84

Finally, I analysed the effect of ECEC on children’s subjective wealth at age 15, which are the only available data on income when they were young.Footnote 85 As the results show in Table 22 in Appendix B, we do not see any effects of ECEC on their households’ income, although we need to be cautious because of the noise in the measure. This seems to be consistent with the finding that there is little change in female labour force participation: the increase in income seems too small to conclude that there is an income effect.

8.3 Mechanisms behind the long-term effects of ECEC

Here, I discuss potential mechanisms that drive the results presented in the previous sections based on the literature, although some cited studies considered different timings than mine, and so we must be cautious. First, in the short run, Yamaguchi et al. (2018b) find that ECEC positively affects language development and reduces symptoms of inattention, hyperactivity, and aggression among disadvantaged children.Footnote 86 This implies that disadvantaged children receiving ECEC in my sample are likely to have improved these abilities during enrolment.

Chetty et al. (2011) contend that students taught by more experienced teachers in kindergarten earn more in the future. They found that licensed childcare with more experienced teachers in terms of childcare raises children’s abilities. Berlinski et al. (2009) state that attending preschool increases test scores in elementary school and improves students’ self-control (e.g., better school attendance and discipline, and improved attention and efforts). This improvement can be seen in ECEC in Japan too. Although many studies found that the effects on cognitive abilities fade quickly (Heckman and Masterov 2007; Heckman et al. 2013; Havnes and Mogstad 2015), these effects could last during the early stages of elementary schools and might be sufficient to push children into better school cohorts. Children’s improved non-cognitive abilities can also make a difference, and the effects on non-cognitive abilities are likely to last longer (Heckman and Masterov 2007; Heckman et al. 2013; Havnes and Mogstad 2015), despite some fluctuations (Roberts et al. 2006; Borghans et al. 2008; Roberts and DelVecchio 2000). The traits fostered in childhood matter to the outcomes in adulthood (Conti et al. 2016). Improvements in self-control could lead to college completion. Consequently, a higher education level leads to higher wages (Mincer 1974; Becker 1994; Kane and Rouse 1995; Thomas 2003), which boosts income.

9 Conclusion

This study estimated the long-term effects of ECEC participation on college completion, adult income, and personality traits. The main analysis was based on an IV regression because there is a potential threat of endogeneity; parental choice to enrol children in ECEC may be correlated with unobservable household characteristics. To address this issue, I use a policy change that affected childcare enrolment. Specifically, the analysis used the quasi-random expansion of ECEC from the 1960s to 1980s in Japan owing to the spike in demand under Japan’s high rate of economic growth. This rate of expansion differed regionally, and supply-side constraints were generally binding, allowing me to treat this variation as quasi-random.

The results show that enrolment in Japan’s childcare system at age four increases children’s future income and wages by age 50 as well as the likelihood of college completion. The subsample analysis reveals that the long-term effects of ECEC are mainly triggered by changes in females’ college completion rates and adult income. In other words, female children benefited more from ECEC over the long run. Based on the analysis of the marginal treatment effects, there is no negative effect uncovered for any children in terms of unobserved heterogeneity. There seem to be no long-term effects on non-cognitive abilities measured using the Big Five indexes. Furthermore, I examined the mechanism behind the effects by conducting mediation analysis. As a result, I find that most of the effects on income can be explained by the rise in the likelihood of college completion, which in turn leads to higher wages, and, therefore, leads to higher income.

This research contributes to the literature as well as policymaking. Few previous studies have discussed the long-term LATE of universal childcare systems as opposed to the shorter-term effects, or effects of targeted ECEC programmes. Besides, few studies have discussed the mechanisms behind this phenomenon. This paper contributes to the literature by showing the long-term effects of a universal ECEC programme and its mechanism. Furthermore, knowing the long-term positive effects of an early-stage intervention on income is useful for policymaking. If a government allocates a fixed budget for one cohort over its lifetime, my results might imply that it may be more efficient to limit the allocation of funds to them when they are young rather than when they become adults, only because doing so would boost skills, such as their cognitive and non-cognitive abilities in childhood, and increase their future income. Spending resources on older cohorts may lead only to consumption without any economic growth. This research exemplifies the usefulness of ECEC by showing its effects. ECEC reduces inequalities between the advantaged and disadvantaged by benefiting the latter and not imposing a direct non-monetary cost on the former. I also find that ECEC increases the likelihood of college completion and raises future income for women. These results imply that ECEC can reduce inequalities and encourage governments to eliminate any barriers to enrolment, such as childcare fees, particularly for disadvantaged children, given the finding that there is no effect and cost for advantaged children. The results also imply that ECEC enrolment could break the intergenerational poverty chain because the early childhood intervention reduces the inequality in their adulthood. Regarding gender inequality, these results are informative for countries or communities where women have culturally and historically been relegated to home production or part-time work. This research is also useful for discussing the expansion of ECEC in developing countries, as described in Target 4.2 of the United Nation’s SDGs (United Nations 2015).Footnote 87 This is because this analysis is based on the universal (i.e., not targeted) ECEC expansion of 1960–1989 in Japan, a period during which the country experienced rapid development. Some might think that an alternative policy can be more efficient, such as parental leave when parents, and especially mothers, are supposed to spend more time with their children. However, as Yamaguchi et al. (2018b) discuss, using childcare is beneficial for improving parenting quality and mothers’ subjective well-being, in addition to reducing stress among low-educated mothers, since the parenting quality of low-educated mothers is reported as low. Therefore, a childcare system is beneficial for both parents and children from disadvantaged households.

The main caveat of this research comes from the data. Although the questionnaire used is unique, the number of observations is smaller than that of recent empirical papers. Therefore, the estimated standard errors are larger, leading to less accurate estimations than otherwise. The data are also not panel data and contain retrospective or subjective records, which could reduce accuracy. Overcoming these caveats is left for future research. Another potential caveat is that the result cannot be directly used to improve the current situation in Japan, because the availability of ECEC is substantially different. However, the implication stated here can be useful in developing countries where the government is investing in ECEC under the SDGs (United Nations 2015).