Investing in disadvantaged children for common prosperity in China

Abstract

Despite rapid economic growth in China since 1978, rural-urban inequality has widened. High levels of socioeconomic inequality can have profound implications for child development and lifelong educational equity. Using a dataset containing early childhood development (ECD) outcomes of 0- to 3-year-olds (N = 9,053) from study sites in Eastern, Central, and Western China, the study finds that the risks of cognitive, language, and motor delay are, respectively, 43.2, 18.3, and 20.7 percentage points higher in rural study sites than in urban Shanghai (ps < .01). Impact evaluation of cluster-randomized experiments shows that parental training (focusing on child psychosocial stimulation and caregiver-child interaction) can improve parenting beliefs and practices (or investments) and ECD outcomes of disadvantaged rural children (p<.01). Such programs can play an important role in advancing progress toward more social equality and economic equity, the stated goals of China’s “Common Prosperity” policy.

Introduction

The introduction of the 1978 reforms in China led to four decades of sustained economic growth (Zhang 2021) and steady improvement in educational attainment (Li et al. 2017). During this same period, however, rural-urban inequality increased sharply (Rozelle and Hell 2020). As of 2016, the average urban disposable income in China was, on average, 170% higher than the average disposable income of the rural population (Zhang 2021). Furthermore, the Gini coefficient in China increased from about 0.30 in 1978 to 0.47 in 2016 (Zhang 2021). Luo and Xie (2020) and Yang and Gan (2020) show that, since the 2010s, the level of income inequality in China is among the highest in the world.

While trends in income inequality in China have long been a topic of academic scrutiny and debate, few studies have investigated the implications of socioeconomic inequalities for child development and lifelong educational equity in China. Research findings from other countries have shown that children in lower socioeconomic status (SES) families are at a higher risk of developmental and educational delays than are children from higher-SES families (Bornstein and Bradley 2002; Falk et al. 2021; Heckman and Mosso 2014; Schady et al. 2015), which may contribute to the intergenerational transmission of disadvantage (Qin, Wang, and Zhuang 2016). For example, Schady et al. (2015) finds that gaps in the cognitive ability between children from low-SES and high-SES families in Chile, Colombia, Ecuador, Nicaragua, and Peru emerge by age 3 and do not change substantively during later stages of childhood. Fernald et al. (2012) and Rubio-Codina et al. (2015) extend this research by showing that, when comparing high- and low-SES families, gaps in the cognitive skills of young children in low- and middle-income countries (LMICs) emerge as early as at 7 months of age and (in the absence of intervention) tend to persist as the children age throughout their childhood.

It is well established in the literature that the first years of life are foundational for human development (Currie and Almond 2011; Heckman 2006) and that SES differences during this sensitive and critical stage for human development may have lifelong consequences. Delays in early childhood development (ECD) have been associated with lasting impacts on a range of adult human capital outcomes, including lower cognitive functioning (Walker et al. 2011), lower incomes (Goodman, Joyce, and Smith 2011), increased risk of mental health problems (Walker et al. 2022), and decreased labor productivity (Gertler et al. 2014). Research from the United States has shown that SES gaps in cognitive and non-cognitive abilities that emerge during early childhood tend to remain constant into adulthood (i.e., the gaps remain high or low across time—Heckman 2008).

The human development literature highlights the role of the family environment as one of the key mechanisms behind the impact of family SES on ECD outcomes (Young and Hannum 2018). This means that SES inequalities in children’s outcomes originating in early childhood are to a large extent determined by inequalities in how stimulating, nurturing, and safe the children’s home environments are (Attanasio, Cattan, and Meghir 2022). Caregiver investments in a cognitively stimulating home environment that is sensitive to the needs of the child have been identified as key inputs for the cognitive and noncognitive development of young children. Furthermore, Attanasio, Cattan, and Meghir (2022) argue that, whereas investments in the home environment are strongly correlated with household SES, financial resources do not play a large role for children’s developmental opportunities per se. Empirical evidence has shown that the SES gap in caregiver investments in the family environment is, to a large extent, driven by differences in parenting beliefs, knowledge, and skills (Cunha 2015). Hence, providing low-SES households with advice, guidance, or training on effective parenting strategies has been shown to improve equity in child development and reduce intergenerational transmission of disadvantage.

To develop a scalable remediation strategy that can offset the intergenerational transmission of disadvantage in diverse cultural and socioeconomic settings,¹ evidence on the effectiveness and efficacy of promising policy interventions and their underlying mechanisms are needed. Pioneering studies from the 1970s and 1980s, with long-term follow-ups, including the Jamaican Nutrition and Cognitive Stimulation Program, found long-lasting positive effects of small-scale parenting interventions (with a focus on child psychosocial stimulation and caregiver-child interaction) on a range of adult human capital outcomes, including lifetime educational attainment (Walker et al. 2022). Such promising results incentivized researchers and policymakers in the 2010s, in particular, after the introduction of Target 4.2 of the Sustainable Development Goals, to “ensure that all girls and boys have access to quality early childhood development” by 2030 and to replicate these types of intervention studies in resource-poor settings around the world (e.g., the studies of Attanasio et al. 2014 in rural Colombia; Singla, Kumbakumba, and Aboud 2015 in Uganda; Yousafzai et al. 2014 in Pakistan). Over the past several decades, evidence from randomized controlled trials in LMICs in Africa, Asia, and Latin-America have confirmed that parental training focusing on interactive caregiver-child activities can improve ECD outcomes in resource-poor settings (Emmers et al. 2022). Therefore, global policymakers no longer debate the effectiveness of parenting programs for improving the home environment and ECD outcomes. Instead, the main focus of the debates in these countries is how quality child and family services can be delivered in cost-effective, sustainable, and at scale ways.

Evidence on the underlying mechanisms of treatment impacts can inform policymakers about ways to design cost-effective policies that can support the development of disadvantaged children at scale and advance progress toward more social equality and economic equity, the stated goals of China’s “Common Prosperity” policy. The international literature has shown that the impacts of parental training programs on ECD outcomes are partially mediated by improvements in parental investments in the quality of the early learning and the home environment (Amaro Da Costa Luz Carneiro et al. 2019; Attanasio et al. 2020). In addition, studies have shown that changes in parenting beliefs—the perceptions, attitudes, knowledge, ideas, and values of parents with regard to child rearing and child development—can positively mediate treatment impacts of parental training programs on ECD outcomes (Amaro Da Costa Luz Carneiro et al. 2019; Attanasio, Cunha, and Jervis 2019). Changes in parenting beliefs can affect ECD outcomes directly (e.g., via changes in the responsiveness of caregivers) and indirectly via changes in the nature of parental investment (e.g., improved parenting beliefs and knowledge may induce more investment in the time that parents spend on interactive reading, story-telling, and play activities with young children; Attanasio, Cunha, and Jervis 2019).

For the case of China, empirical evidence has demonstrated that skill gaps between rural and urban children can already be observed starting from an early age onwards when rural children start to lag behind urban peers in their development (e.g., Wang et al. 2019; Yue et al. 2017; Zhang et al. 2021). An important limitation of the evidence base, however, is that earlier population-level ECD assessments have focused mainly on children between 3 and 6 years of age, whereas population means for children younger than age 3 remain to be assessed (Clark et al. 2020). A systematic review and meta-analysis of studies that utilize empirical primary data that were collected in rural study sites across Central and Western China illustrated that the risk of early cognitive and language delay for rural children younger than 5 years is high: 45 and 46%, respectively (Emmers et al. 2021). Each of the studies included in the meta-analysis used standardized instruments, such as the Bayley Scales of Infant and Toddler Development (Bayley 2006) to assess ECD outcomes. In line with the guidelines of the Bayley Scales (the gold standard of ECD assessment) and definitions of the Global Research on Developmental Disabilities Collaborators (2018), developmental delay was defined as a development score of 1 or more standard deviations (SDs) below the mean of a reference population whose developmental trajectory is expected to be normal (i.e., a population of children in healthy/developed regions and did not include prematurely born, severely malnourished, or severely diseased children). Note that observational assessments, such as the Bayley Scales, are unlikely to be sensitive to SES differences. Furthermore, for the standardized, caregiver-reported ECD measures used in the studies, evidence has suggested that parents across different SES groups can provide accurate measurements of the development of children as validated by direct and observational assessments (Fernald et al. 2017). Emmers et al.’s meta-analysis, unfortunately, compares observational assessments of ECD outcomes of children under 3 years of age in Central and Western China (as well as migrant communities) only and does not examine the difference in ECD outcomes of samples either across rural and urban study sites or between inland rural China (i.e., rural samples from less-developed areas in Central and Western China) and Eastern China (i.e., historically, the more developed segment of China’s rural population; Zhang 2021). Therefore, the current study sets out to construct a unique dataset containing Bayley test scores from birth cohorts in urban areas as well as rural areas of Eastern, Central, and Western China in order to assess whether ECD outcomes of rural children (from Eastern, Central, and Western regions) are significantly different from their urban peers before they reach the age of 3.

The high prevalence of developmental delay in rural areas of China may be explained, at least partly, by the comparatively high prevalence of low-SES families in rural areas. SES differences in China have empirically been linked to differences in educational outcomes (Young and Hannum 2018). Zhang et al. (2021), for example, used large-scale field data containing early Human Capability Index (eHCI) scores of 63,559 children between 3 and 6 years of age to show that children in low-SES families and those in families with lower maternal educational attainment are, on average, 12 and 8 percentage points less likely to be developmentally on track than are children in higher-SES families, respectively. This large sample of 3- to 6-year-olds was selected using a carefully designed stratified randomized sampling strategy across four provinces in China: Shanghai (i.e., a megacity), Zhejiang (i.e., representing the most-developed region of the country in Eastern China), and Gansu and Yunnan (i.e., representing China’s least-developed regions). To the best of our knowledge, however, at-scale evidence remains to be provided of the SES differences in ECD outcomes for children under the age of 3 years in rural China. In order to address this research gap, the current study tests whether ECD outcomes differ significantly between children under age of 3 years in low-SES and high-SES households. In addition, the study tracks SES differences over age (i.e., how the SES gradient evolves for each child between the age of 6 months and the age of 3 years).

Over the past decade, parental training has been encouraged and tested as a scalable remediation strategy both to give more disadvantaged children in rural areas of China a fair start in life and to offset the intergenerational transmission of disadvantage. Evidence from 10 randomized controlled trials in rural China demonstrated that parental training programs focusing on caregiver-child interactions can improve ECD outcomes of disadvantaged children in rural areas (Emmers et al. 2021). The intervention programs were adapted to the local culture and settings in order to encourage program delivery, participation, and take-up. For example, considering that approximately 30% of the children under the age of 3 in rural China grow up in the care of a grandparent while their parents migrate to cities for work (Bai et al. 2022), grandparent caregivers in charge of the daily care of young children were encouraged to participate in the programs as well. Moreover, the underlying mechanisms of the treatment impacts of the parenting experiments on ECD outcomes remain to be investigated. In order to shed light on the underlying mechanisms, this study pools data from three cluster-randomized controlled trials that tested parental training programs in study sites in Central and Western China. The pooled data set allows us to test the aggregate treatment impacts of parental training on Bayley cognition, language, and motor scores of young children in rural China, as well as to explore the share of the treatment impacts on ECD outcomes that are mediated via changes in parenting beliefs and parental investment in caregiver-child interactions.

This study primarily makes three contributions to the literature. First, the study provides evidence of the rural-urban gap in ECD outcomes that is more representative for the whole of China than earlier studies. To be more specific, the current study is the first study that pools data on ECD outcomes of young children from study sites in rural and urban areas in all three major geographical regions: Eastern, Central, and Western China. We are able to accomplish this by using a unique dataset with ECD outcomes of 0- to 3-year-olds (N = 9,053) from rural study sites in Eastern, Central, and Western China as well as from urban Shanghai. Second, the study quantifies the SES gradient in ECD outcomes and tracks its age pattern (i.e., how the gradient evolves between the age of 6 months and 3 years). To the best of our knowledge, no earlier study has provided evidence of the SES gradient in skills for children under age 3 in China. In particular, no earlier study provided evidence of age patterns in the SES gradient via regular tracking of a birth cohort between the ages of 6 months and 3 years. Third, this is the first study that conducts a mediation analysis of the mechanisms behind treatment impacts on ECD outcomes of parental training programs in rural study sites in China.

The remainder of this paper is structured as follows. Section “Background: the importance of ECD for common prosperity” focuses on the importance of equality and equity in the opportunity of ECD for Common Prosperity. The data and methodology are described in Sections “Data” and “Methods,” respectively. Section “Results” presents the findings on the rural-urban gap in ECD outcomes, the SES gradient in ECD outcomes over time (between the ages of 6 and 30 months), the impacts of cluster-randomized parenting experiments on ECD outcomes, and the mediating role of parenting beliefs and parental investments. Section “Discussion and Conclusion” provides a discussion and concludes.

Background: the importance of ECD for common prosperity

China’s government has announced its intention to steer China on a path of people-centered development toward more social equality and economic equity in the context of the “Common Prosperity” policy. This path is markedly different from China’s developmental trajectory during the past couple of decades (Wang and Peach 2019). During the beginning of the reform and opening-up period, one of the slogans advocated by Deng Xiaoping was, “Let some people get rich first” (Naughton 1993). After decades of steady economic growth and institutional development, China achieved several notable milestones: China became the second largest economy in the world after the United States in 2010; China completed the eradication of extreme poverty in 2020; and China announced that the country had achieved its first Centenary Goal of becoming a moderately prosperous society in 2021 as the nation celebrated the 100th anniversary of the founding of the Communist Party (Kakwani et al. 2022). Despite achieving these milestones (specifically, China is now a moderately prosperous society without extreme poverty), the nation’s GDP per capita remains below the average GDP per capita of the world (World Bank 2022), and regional and rural-urban inequalities have widened (Zhang 2021; Rozelle and Hell 2020).

Facing these remaining challenges, China’s central leadership has announced that it is no longer acceptable to allow some people to fall too far behind (Kakwani et al. 2022). In its most high-profile recognition of the problem, since 2021, China’s top leaders have given prominence to the phrase Common Prosperity, which is used to describe a policy shift that is intended to curtail income inequality and excessive wealth accumulation by individuals. For example, during the opening speech to the 20th National Congress of the Communist Party of China on October 16, 2022, General Secretary Xi Jinping stated:

Achieving Common Prosperity is a defining feature of socialism with Chinese characteristics and involves a long historical process. The immutable goal of our modernization drive is to meet the people’s aspirations for a better life. We will endeavor to maintain and promote social fairness and justice, bring prosperity to all, and prevent polarization. (p. 18)

Equality and equity of opportunity are the foundations of the Common Prosperity policy. A high degree of inequality of opportunity in China is due to the historic legacies of the country. According to their policy announcements, the current leadership appears to recognize that a large share of public investment has been channeled toward urban areas in the post-reform era at the dispense of rural areas. As a result of these systematic trends in investment and policy efforts, rural-urban inequality explains over 70% of income inequality in China today (Afridi, Li, and Ren 2015). Further, if the explanation in Xi Jinping’s statement is valid and that rural residents are faced with a substantial inequality of educational opportunity that stems from developmental differences during early childhood is true, then it follows that providing young children with a fair start in life is crucial to address the “principal contradiction.. between unbalanced and inadequate development and the people’s ever-growing needs for a better life” (Kakwani et al. 2022, 29) that China’s society faces today.

Data

The dataset contains information on 9,053 unique caregiver-child dyads from rural areas of Central and Western China (n = 6,205) and rural and urban areas in the provincial-level municipality of Shanghai (n = 2,848; Table 1). The Shanghai Birth Cohort data were collected between 2015 and 2018 by the Shanghai Key Laboratory of Children’s Environmental Health and the Shanghai Children’s Medical Center at Shanghai Jiaotong University. The sample contains 2,145 and 703 Shanghai residents and their young children living in urban and rural areas of Shanghai, respectively (see Table 1; for a more extensive description of the sampling strategy and data collection procedures, please refer to Zhang et al. 2019). The birth cohort data from Central and Western China were collected by an international research collaboration between the Rural Education Action Program and local partners in China between 2013 and 2020. For the samples from Central and Western China, repeated rounds of observations were collected for each child in the context of ongoing intervention studies (see Section “Impact evaluation and mediation analysis” for further details). The subjects of the Shanghai Birth Cohort were not enrolled in a randomized controlled trial that included an intervention to the children assigned to the treatment arm. The pooled sample contains a total of 14,807 “untreated” observations (i.e., baseline or control group observations).

Table 1. Data sources.

No.	Source	Location	Sample size	No. of waves	No. of untreated observations
1	Shanghai Birth Cohort	Urban areas in Shanghai	2,145	1	2,145
		Rural areas in Shanghai (Chongming)	703	1	703
2	Birth cohorts from rural areas of Central and Western China	Rural areas of Shaanxi Province	1,801	4	5,641
		Rural areas of Shaanxi Province	2,929	3	3,996
		Rural areas of Yunnan and Hebei Provinces	449	3	815
		Rural areas of Yunnan Province	1,026	2	1,490
Total			9,053		14,807

Note. The birth cohort samples from rural areas in Central and Western China were collected in the context of cluster-randomized controlled trials. Untreated observations are the baseline and follow-up assessments of caregiver-child dyads assigned to the control group.

Information on ECD outcomes and child and household characteristics were collected for each observation in the pooled dataset. In addition, for the observations from Central and Western China, we collected measures of parenting beliefs (e.g., belief that reading books is important for child development) and parental investments in child psychosocial stimulation (e.g., engagement of children in reading, storytelling, and interactive play). For the sample from Shanghai, no secondary outcome measures of parenting beliefs and parental investment were collected.

ECD outcomes

The ECD outcomes for each of the observations in the pooled sample were previously assessed via behavioral assessments using the first or third edition of the Bayley Scales of Infant and Toddler Development (Bayley 2006). The Bayley scale is generally considered the gold standard for measuring ECD outcomes (Del Rosario et al. 2021). The first edition of the Bayley (Bayley-I) test contains two scales for the assessment of 0- to 30-month-olds in the domains of cognition and motor development (Bayley 1933). The third edition of the Bayley (Bayley-III) test contains scales for the observational assessment of 1- to 42-month-olds in the developmental domains of cognition; (expressive and receptive) language; and (fine and gross) motor skills. The Bayley Scales were translated in Mandarin Chinese and validated for a healthy Chinese sample (Xu et al. 2011; Yi 1995). The observational assessments require the child to engage in a series of increasingly difficult tasks. If the child passes a task, the item receives a score of 1. Otherwise, the item receives a score of 0. When the child fails to perform six consecutive tasks, the test ends, and the raw score is found by the aggregation of the raw item scores. Studies have shown that the versions of the Bayley tests that are used in China have good reliability and validity (Hua et al. 2019; Yi 1995; Yue et al. 2019). The Bayley tests were all administered one-to-one via behavioral assessment by trained enumerators. Prior to test administration, the enumerators completed one week of intensive training, including 2.5 days of experiential learning in the field.

For use in the analysis of this study, the Bayley raw scores were converted into three intuitive ECD indicators. First, we converted the Bayley raw scores into Bayley composite scores, following the scoring guidelines in the Bayley manual (Bayley 2006; Yi 1995). The composite scores have a mean of 100, an SD of 15, and a range of 40–160. Second, we used the composite scores to construct indicators of developmental delay. We define developmental delay as a composite score of 1 or more SDs below the mean of a reference population whose developmental trajectory is expected to be normal (i.e., a child population in a developed region who was not born prematurely, severely malnourished, or severely diseased; Emmers et al. 2021).² This approach is in line with the usage guidelines of the Bayley test, which conventionally defines children with a score of more than 1 SD below the normative sample mean as mildly delayed in their development (Bayley 2006).

Third, with a view to comparability with treatment impact estimates reported in published studies, we convert the Bayley raw scores to standard normally distributed z-scores. Considering that Bayley raw scores increase with age, we eliminate the age effect by internally standardizing raw scores within age (month) groups. This is done by first estimating age-conditional means and SDs using non-parametric regression and then using these estimated statistics to compute age-adjusted internal z-scores for each subscale. This non-parametric standardization procedure has the advantage that it is less sensitive to outliers and small sample sizes within age categories than are parametric procedures (Rubio-Codina et al. 2016).

Parenting beliefs

In addition to measures of child development outcomes, the research team collected measures of parenting beliefs and parental investments in child psychosocial stimulating activities for the rural samples from Central and Western China (Table 1). These measures were collected using household interview questionnaires administered to the household’s main caregivers, whereby the main caregiver was defined as the person who is responsible for the daily nutrition and care of the child. To measure parenting beliefs, we asked a series of questions related to the caregiver’s beliefs and attitudes toward parenting, including whether the caregiver thinks it is important to play interactively or read with the child. This set of questions was also used to assess parenting knowledge and beliefs in earlier published studies (e.g., Luo, Emmers, et al. 2019; Sylvia et al. 2022).

Parental investment

To measure parental investment in a cognitively stimulating home environment, we administered a subset of items extracted from the Family Care Indicators scale (FCI; Hamadani et al. 2010). Specifically, we assessed the number of picture books at the home and caregiver engagement in four at-home, play activities with the child: reading books, telling stories, singing songs, and playing with toys. Previous studies demonstrated that the FCI provides a reliable measure of parenting and the home environment in developing countries (Hamadani et al. 2010). The FCI has been adapted to the Chinese language/context and has been used in earlier studies across rural China (Wang et al. 2022). We then estimated a dedicated measurement system to relate the observed measures of parenting beliefs and parental investment to their corresponding latent factors using factor analysis (Appendix A). Finally, we internally standardized the factor scores within the age (in months) group of the child to obtain standard normally distributed factor z-scores that are uncorrelated with the age of the children (Rubio-Codina et al. 2016).

Child and household characteristics

Information on basic child characteristics (e.g., sex, age, gestational age at birth) and caregiver characteristics (e.g., age, educational attainment) also was collected for the full sample. Finally, information on household SES was collected. For individuals in the Shanghai sample, survey respondents were asked to rate their socioeconomic situation on a 5-point Likert scale, for which 1 indicates a very comfortable life and 5 indicates that they have a difficult time financially. Scores of 1, 2, 3, and 4/5 corresponds to very well off, fairly well off, fairly poor, and very poor, respectively. For the rural samples from Central and Western China, the survey team collected information on the presence of a number of household assets at home. A household asset index was constructed based on six indicators: access to air conditioning, flush toilet, boiler, car, TV, and internet. Similar SES indices have been used extensively in the medical, demographic, nutritional, and economics literatures (e.g., Rubio-Codina, Attanasio, et al. 2016; Schady et al. 2015). A household is defined as very (fairly) poor if the household has access to one to three (four to six) of these household assets.

Methods

Descriptive analysis of rural-urban and SES differences

We conduct the complete statistical analysis with Stata/MP, version 17.0. First, we summarized demographic characteristics and ECD outcomes for the rural and urban subsamples. Then, we compared the means of the demographic characteristics and ECD outcomes of children from rural study sites in Eastern, Central, and Western China with the respective means of their urban counterparts. Two-tailed t-tests were used to test for the significance of the differences in means.

Second, in line with the methodology of earlier studies (e.g., Fernald et al. 2012; Rubio-Codina et al. 2015; Schady et al. 2015), we provided descriptive evidence of the SES gradient in ECD outcomes by comparing ECD scores across different SES groups. Specifically, we compared ECD outcomes across households with high versus low access to household assets and different levels of maternal educational attainment. We used two-tailed t-tests to determine the significance of the differences in means. In addition, using a panel dataset with four waves of ECD scores (n = 1,801; Table 1), we tracked the evolution of this SES gradients of children between the ages of 6 and 30 months.

Impact evaluation and mediation analysis

For treatment impact estimation and the mediation analysis, we pooled data from three cluster-randomized parenting experiments in rural study sites in Central and Western China. The study sample from Yunnan Province (see bottom row of Table 1) was excluded for the impact evaluation and mediation analysis, because of the lack of secondary outcome measures of parenting beliefs and parental investment (as described in Sections “Parenting beliefs” and “Parental investment”). First, we describe the three randomized parenting experiments from which data are pooled for the subsequent part of the analysis. We then explain the estimation strategy for the assessment of treatment impacts on primary (i.e., ECD outcomes) and secondary outcomes (i.e., parenting beliefs and parental investment) and the mediation analysis.

Intervention design and procedures

The pooled dataset contains data from three cluster-randomized intervention studies in rural study sites in Central and Western China. Taken together, the three randomized controlled trials cover 274 villages in 28 nationally designated poverty counties in rural Shaanxi, Yunnan, and Hebei Provinces (Appendix Figure B.1). In each study, treatment assignment was randomized at the village level to reduce the risk of contamination (among families within a village) across experimental groups. The impact evaluation sample includes 2,800 children with complete records of ECD scores at baseline and follow-up. Appendix B contains a detailed description of the study design, participants, randomization, masking, and procedures.

During the intervention period, all children and caregivers who resided in clusters of the intervention group were enrolled in a parental training program that focused on child psychosocial stimulation that was delivered at home or in a child center at a central location in the community. In the first home-based programs in rural Shaanxi, caregiver-child dyads were invited to participate in 24 weekly training sessions over a 6-month period between November 2014 and April 2015. In the second home-based program in Yunnan and Hebei Province, caregiver-child dyads were enrolled in 24 bi-weekly training sessions over a one-year period between October 2015 and September 2016. In the center-based program in Shaanxi, caregiver-child dyads were invited to participate in 48 weekly training sessions over a period of one year. During each of these parental training sessions, community health workers (paid and managed by the research team) trained caregivers on interactive caregiver-child activities (e.g., reading, storytelling, interactive playing with toys). Training sessions were designed to last for one hour.

All interactive caregiver-child activities were stage-based and fully scripted in a curriculum. This curriculum was based loosely on the curriculum developed for a pioneering parenting intervention in Kingston, Jamaica, and adapted to the rural China setting by local child development psychologists and ECD experts (Sylvia et al. 2021). Each activity focused on child development in one of four skill domains: cognition, language, motor, or social-emotional development. Two new activities were introduced during each training session (i.e., parents were instructed on practicing these activities with their children) such that caregivers would encounter at least one new activity to stimulate development in each of the four targeted skill domains each month. In addition to developmental activities, the curriculum also included one module with basic information on child health and nutrition that was delivered at the end of each session.

Training sessions were delivered by community health workers in the service of China’s National Health Commission. Before the start of the intervention, each of the community health workers had to complete a one-week intensive training course on basic theories of ECD and on communication, coaching, and counselling techniques. The training program consisted of five days of classroom-based instruction by ECD experts in combination with two days of field practice. Fully scripting the curriculum obviated the need for more extensive training of community health workers. Upon completion of the training course, each of the community health workers received a copy of the fully scripted ECD curriculum and a toolkit (which included a growth chart, development check list, counseling materials, and user handbook). To monitor service quality, a carefully designed supervisory system was set up that involved tracking the completion of training sessions (via phone, via mobile app, or by managers at the child centers), unannounced supervisory visits with feedback, and phone interviews with caregivers for feedback. Throughout the program, trainers received periodic refresher training by phone (in the home visitation programs) or in person (in the center-based program).

Empirical estimation strategy

First, we estimated the impact of the treatment assignment on the primary (i.e., ECD outcomes) and the secondary outcomes of interest (i.e., parenting beliefs and parental investment). Provided that treatment assignment and attrition were random, a comparison of the means of outcomes between treatment arms resulted in unbiased estimates of the treatment effect. We used ordinary least squares regression to estimate intention-to-treat (ITT) effects with the following ANOVA regressions specification: (1) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{\alpha }_2{Y}_{\mathit{\text{ij}}0}+{\alpha }_3{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1},$$(1) where $Y_{\mathit{\text{ij}}1}$ is a primary or secondary outcome measure for child i in village j at the end of intervention period; $T_j$ is a dummy variable that indicates treatment assignment of village j; $Y_{\mathit{\text{ij}}0}$ is the outcome measure for child i at baseline; and $X_{\mathit{\text{ij}}1}$ is a vector of covariates that includes children’s sex, age (in months), an indicator for premature birth status, and the relationship between the child and the primary caregiver. ${\pi }_s$ is a set of province, study, and cohort fixed effects. We adjusted standard errors for clustering at the village level, using the cluster-correlated Huber-White estimator.

Second, we conducted a mediation analysis to explore the underlying mechanisms of the treatment impacts on ECD outcomes. The goal in a mediation analysis is to decompose the total treatment effect on the primary outcome into the indirect and direct effects (Keele, Tingley, and Yamamoto 2015). Because we were estimating ITT effects, we decomposed the average total ITT effect (${\overline{\tau }}_{\mathit{\text{ITT}}}$) into a mediated ITT effect (${\overline{\mu }}_{\mathit{\text{ITT}}}$) and an unmediated ITT effect (${\overline{\upsilon }}_{\mathit{\text{ITT}}}$). ${\overline{\mu }}_{\mathit{\text{ITT}}}$ measures the extent to which the effect of the parenting intervention on skill development is transmitted through parenting beliefs and/or parental investment. ${\overline{\upsilon }}_{\mathit{\text{ITT}}}$ reflects the proportion of the average ITT effect that results from all other treatment-induced changes, both observed and unobserved. Following the methodology of Heckman, Pinto, and Savelyev (2013), we first estimated the following augmented outcome equation, for which we added the mediation channels of interest (i.e., the secondary outcomes) to Equation (1)(1) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{\alpha }_2{Y}_{\mathit{\text{ij}}0}+{\alpha }_3{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1},$$(1) : (2) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{{\alpha }_2{M}_{\mathit{\text{ij}}1}+\alpha }_3{Y}_{\mathit{\text{ij}}0}+{\alpha }_4{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1}$$(2) where $Y_{\mathit{\text{ij}}1}$ is the ECD outcome of child i in village j at the end of the intervention period, and $M_{\mathit{\text{ij}}1}$ is the mediating outcome for child i in village j. In all of our specifications, we included a set of covariates $(X_{\mathit{\text{ij}}1)}$ to account for shocks that potentially affect both the ECD outcomes and our mediating outcomes. We first added one mediator to estimate the indirect effect of this particular mediator. Next, we included both mediators (i.e., parenting knowledge and parental investment) simultaneously to examine the total indirect effect of the two mediators.

To attribute a causal interpretation to the mediation analysis, we needed to assume that the mediating variables are exogenous with respect to ECD outcomes, conditional on other observed covariates and treatment assignment. It is possible, however, that there were unobserved variables that are affected by the intervention and, in turn, affected ECD outcomes and the secondary outcomes of interest, which would lead to biased estimates of ${\overline{\mu }}_{\mathit{\text{ITT}}}$. To address this issue, we follow the methodology proposed by Oster (2019) to assess the sensitivity of ${\overline{\mu }}_{\mathit{\text{ITT}}}$ to the inclusion of additional controls. We calculate a statistic known as Oster’s $\delta$ for each ${\overline{\mu }}_{\mathit{\text{ITT}}}.$ This statistic captures the ratio between the selection bias introduced by unobservable variables and the selection bias introduced by the observable controls that would be required such that the true value of ${\overline{\mu }}_{\mathit{\text{ITT}}}$ would be statistically insignificant from zero. This means, for example, that a value of Oster’s $\delta$ = 1 or $\delta$ = 2 suggests, respectively, that the explanatory power of unobserved confounders would need to be at least as high or twice as high as the explanatory power of the observed controls to produce an insignificant ${\mu }_{\mathit{\text{ITT}}}$. As a rule, values of Oster’s $\delta$ larger than 1 indicate that ${\overline{\mu }}_{\mathit{\text{ITT}}}$ is unlikely to be biased due to unobserved heterogeneity.

Finally, to allow for multi-stage mediation effects, we conducted a multi-stage mediation analysis using structural equation modelling (Hayes 2009). In particular, we aimed to test whether the parental training intervention led to an improvement in parenting beliefs, which, in turn, led to an improvement in parental investment and, as a result, improved ECD outcomes. We used structural equation modelling to simultaneously estimate three equations: (a) an augmented Equation (2)(2) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{{\alpha }_2{M}_{\mathit{\text{ij}}1}+\alpha }_3{Y}_{\mathit{\text{ij}}0}+{\alpha }_4{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1}$$(2) to estimate the impact of treatment assignment and the mediators on the ECD outcome of interest; (b) a mediation Equation (1)(1) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{\alpha }_2{Y}_{\mathit{\text{ij}}0}+{\alpha }_3{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1},$$(1) to predict the impact of the treatment assignment and parental investment on parenting beliefs; and (c) a mediation Equation (1)(1) $$Y_{\mathit{\text{ij}}1}={\alpha }_0+{\alpha }_1{T}_j+{\alpha }_2{Y}_{\mathit{\text{ij}}0}+{\alpha }_3{X}_{\mathit{\text{ij}}1}+{\pi }_s+{\epsilon }_{\mathit{\text{ij}}1},$$(1) to predict the impact of treatment assignment on parental investment. In this model, the total indirect effect consists of three indirect effects: the indirect effect through parenting beliefs; indirect effect through parental investment; and indirect effect through parenting beliefs and parental investment. We used a bootstrapping approach with 1,000 replications to construct bias-corrected 95% confidence intervals for the estimated effects (Preacher and Hayes 2008).

Results

Rural-urban differences in demographics and ECD outcomes

Table 2 provides a summary of the demographic characteristics of the pooled sample. We observe that 48.0% of the 0- to 3-year-olds in the study sample are male; 6.1% of the children were born prematurely (i.e., born before 37 weeks of pregnancy); 62.7% of the children were first born; and 37.3% had an older sibling. Mothers were, on average, 27.7 years old at the time of survey administration, and 42.1, 22.8, and 5.0% of the mothers graduated from senior high school, college or university, and postgraduate education, respectively. Finally, 24.5% of the households were relatively well off (in terms of family assets), while the remaining 75.5% were relatively poor.

Table 2. Demographic characteristics of the study sample (N = 9,053).

	Shanghai (including Chongming; n = 2,848)	Rural study sites in Central and Western China (n = 6,205)	Overall (N = 9,053)	Difference between (1) and (2)
	(1)	(2)	(3)
Characteristic	n (%) or M ± SD	n (%) or M ± SD	n (%) or M ± SD	p-Value
Child
Child sex^a
Male	1,481 (51.7%)	3,232 (52.1%)	4,713 (52.0%)	.903
Female	1,384 (48.3%)	2,973 (47.9%)	4,357 (48.0%)
Preterm birth (<37 weeks of gestation)^a
No	2,644 (94.9%)	5,796 (93.4%)	8,440 (93.9%)	<.001
Yes	142 (5.1%)	409 (6.6%)	551 (6.1%)
Child is first born (n = 7,549)^a
Yes	2,093 (85.6%)	2,639 (51.7%)	4,732 (62.7%)	<.001
No	351 (14.4%)	2466 (48.3%)	2,817 (37.3%)
Household
Maternal age (n = 8,653)^b	28.7 ± 3.7	27.3 ± 5.0	27.7 ± 4.6	< 001
Maternal education level (in years) (n = 8,564)^a
Junior high school or below (≤ 9 years)	48 (1.9%)	4,909 (80.3%)	4,957 (57.9%)	<.001
Senior high school (10–12 years)	783 (32.0%)	871 (14.2%)	1654 (19.3%)
College (13–16 years)	1,304 (53.2%)	223 (3.6%)	1,527 (17.8%)
Postgraduate education (≥16 years)	316 (12.9%)	110 (1.8%)	426 (5.0%)
Self-reported socioeconomic status (n = 8,564)^a
Very well off	459 (19.4%)	0 (0%)	459 (5.4%)	<.001
Fairly well off	1,635 (69.1%)	0 (0%)	1,635 (19.1%)
Fairly poor	248 (10.5%)	1,745 (71.8%)	1,993 (23.3%)
Very poor	26 (1.1%)	4,445 (28.2%)	4,471 (52.2%)

Note. ^aMeans compared by chi-square test; ^bMeans compared by Student’s t-test.

Table 2 also provides a comparison of summary statistics between Shanghai and the rural sample from Central and Western China. Children from Shanghai were, respectively, 1.5 and 33.9 percentage points less likely to be born prematurely and to have an older sibling than were their rural counterparts from Central and Western China (ps<.01). Shanghai mothers were, on average, 1.4 years older (p<.01). Finally, mothers from Shanghai were 78.5, 60.7, and 11.1 percentage points more likely to have graduated from senior high school, college or university, and postgraduate education than were mothers from the rural study sites in Central and Western China, respectively (ps<.01).

Table 3 presents the differences in means (and SDs) between the rural and urban in our China samples. A comparison of the means in Columns 1 and 3 demonstrates that Bayley cognition, language, and motor composite scores are, on average, 31.3, 5.7, and 11.5 points higher, respectively, for the children from the Shanghai urban sample than for the rural sample from Central and Western China (ps<.01). The table also provides evidence of the differences in developmental outcomes between urban children from Shanghai and rural children from Chongming, a peri-urban county in the municipality of Shanghai (Columns 1 and 2). The results show that, respectively, the cognition, language, and motor scores of children from urban Shanghai are 32.3, 16.8, and 11.7 points higher than the scores of their same-age peers from Chongming (ps<.01). Interestingly, the findings show that cognition and motor scores do not differ significantly between children from Chongming, a rural area in Eastern China, and children from rural areas in Central and Western China (p > .10). Language scores are 11.1 points lower for children from Chongming as compared to children from rural study sites in Central and Western China (ps<.01).

Table 3. Rural-urban differences in ECD outcomes.

	Urban study site in Eastern China (Shanghai)		Rural study site in Eastern China (Chongming)		Rural study sites in Central and Western China
	(1)		(2)		(3)
Bayley score	n	M ± SD	n	M ± SD	n	M ± SD	Difference in means (1)–(2)	Difference in means (1)–(3)	Difference in means (2)–(3)
Bayley cognition composite score	2,145	124.6 ± 19.3	703	92.3 ± 11.8	11,942	93.3 ± 16.8	32.3***	31.3***	−1.0
Bayley language composite score	2,144	101.3 ± 15.3	703	84.5 ± 9.8	6,292	95.6 ± 13.9	16.8***	5.7***	−11.1***
Bayley motor composite score	2,144	110.7 ± 15.2	703	99.0 ± 9.5	11,909	99.2 ± 18.7	11.7***	11.5***	−0.2

* p<.10; **p<.05; ***p<.01.

Figure 1 illustrates the prevalence of cognitive, language, and motor delay for the rural and urban samples in the study. For the Shanghai urban sample, the data show that the prevalence of cognitive and motor delay is low (<10%), whereas the prevalence of language delay is 35.6%. For the rural study sample, we find that the prevalence of delay is significantly higher (ps<0.01). The prevalence of cognition delay is 56.1 and 46.7% for the rural samples from Chongming and Central and Western China, respectively. The prevalence of language delay is 86.2 and 50.2%, and the prevalence of motor delay is 25.8 and 26.9% for the respective rural samples. With an overall prevalence of cognition, language, and motor delay of 47.7, 53.9, and 26.8% in the pooled rural sample,⁴ we find that the prevalence of cognition, language, and motor delay is, respectively, 43.2, 18.3, and 20.7 percentage points higher in the rural sample than in the urban sample (ps<.01).

Figure 1. Prevalence of delay (n = 14,807).

Based on a subsample of the overall dataset (n = 1,801), which contains repeated observations of Bayley-I cognition and motor scores from rural areas in Central/Western China, Figure 2 presents trends in mean ECD scores and the prevalence of delay as the children age between 6 and 29 months. With regard to cognition scores, the left-hand side of Figure 2 demonstrates that the mean Bayley cognition scores of children deteriorate during the first years of life (i.e., between the ages of 6 and 11 months and 24 and 29 months); the prevalence of cognitive delay increases from 19.9 to 57.0%. In contrast, with regard to motor scores, we observe an improvement in the composite scores and a decrease in the prevalence of delay from 32.1 to 17.1% during this early stage of life.

Figure 2. Mean ECD scores and prevalence of delay over age (in rural sample from Western China; n = 1,801).

SES gradient in ECD outcomes

Table 4 presents differences in ECD outcomes between subgroups based on maternal educational attainment and household SES. The results show that ECD outcomes in the rural and urban samples are consistently higher for children with mothers who completed more years of education and in households with a higher SES. Each of the differences in the means of the ECD outcomes between subgroups, based on maternal educational attainment and household SES, is significant at the 5% level.⁵ These results indicate that maternal educational attainment and household SES are important predictors of ECD outcomes.

Table 4. Differences in ECD outcomes based on child and household characteristics.

	Bayley cognition composite score		Bayley language composite score		Bayley motor composite score
Characteristic	Shanghai (including Chongming)	Rural study sites in Central and Western China	Shanghai (including Chongming)	Rural study sites in Central and Western China	Shanghai (including Chongming)	Rural study sites in Central and Western China
Child sex
Boys	114.9 ± 22.5	87.2 ± 18.1	95.0 ± 15.6	97.6 ± 12.1	106.4 ± 14.5	104.1 ± 17.0
Girls	118.7 ± 22.4	89.8 ± 18.4	99.6 ± 15.8	101.6 ± 13.0	109.4 ± 15.1	106.0 ± 16.8
p-Value of difference in means	<.001	<.001	<.001	<.001	<.001	.002
Preterm birth (<37 weeks of gestation)
No	116.7 ± 22.4	88.8 ± 18.2	97.3 ± 15.9	99.7 ± 12.8	108.0 ± 15.0	105.3 ± 16.9
Yes	118.0 ± 23.3	84.5 ± 19.0	94.9 ± 16.3	96.2 ± 9.9	106.4 ± 15.3	101.0 ± 17.0
p-value of difference in means	.508	.001	.077	.015	.233	<.001
Maternal educational attainment
Postgraduate (≥16 years)	125.1 ± 19.8	100.5 ± 15.3	102.7 ± 15.1	109.2 ± 15.7	111.0 ± 16.1	110.1 ± 14.1
College (13–16 years)	119.7 ± 21.6	96.8 ± 15.2	99.2 ± 15.8	105.3 ± 11.1	109.0 ± 15.1	110.8 ± 16.5
Senior high school (10–12 years)	107.1 ± 22.2	94.1 ± 16.7	91.5 ± 14.8	102.4 ± 14.3	104.8 ± 14.2	107.4 ± 16.3
Junior high school or below (≤9 years)	94.8 ± 17.3	87.2 ± 18.2	81.3 ± 12.0	98.5 ± 12.1	99.5 ± 11.5	104.2 ± 17.1
p-Value of difference in means	<.001	<.001	<.001	<.001	<.001	.003
Self-reported socioeconomic status (SES)
Very well off	118.8 ± 22.5		99.9 ± 16.5		110.2 ± 15.6
Fairly well off	115.7 ± 22.5		96.7 ± 15.6		107.5 ± 14.8
Fairly poor	113.5 ± 22.7	94.7 ± 15.7	94.4 ± 16.1	102.1 ± 14.0	105.2 ± 14.7	108.2 ± 15.9
Very poor	104.2 ± 21.1	86.5 ± 18.6	87.8 ± 17.5	98.3 ± 11.9	101.8 ± 12.2	104.0 ± 17.2
p-Value of difference in means	<.001	<.001	<.001	<.001	<.001	<.001

The results presented in Figure 3 add to these findings by showing that inequalities in ECD outcomes tend to persist as children age and may even persist over generations. Figure 3(a) shows that children who belong to the bottom (top) of the skill distribution at age 6–11 months, in general, still belong to the top (bottom) of the skill distribution at age 24–29 months (ps<.01).⁶ Hence, in the absence of any intervention, on average, disadvantaged children who suffer from an early delay in development are unlikely to catch up with their more advantaged peers. Figures 3(b,c) also provide evidence of the intergenerational transmission of developmental delay. If mothers completed fewer years of education or households have lower SES, then their children are more likely to have lower ECD scores at age 6–11 months (ps<.01). Moreover, these inequalities persist at least until the children reach 24–29 months of age.

Figure 3. Mean ECD scores over age by subgroup (in rural sample from Western China; n = 1,801).

Treatment impacts of parental training on primary and secondary outcomes

We conducted two-tailed t-tests to check whether outcomes were balanced across the treatment and control groups at baseline. First, we conducted a balance test for the full baseline sample (Appendix Table B.1 in Appendix B). According to the data, there were no significant differences between the baseline treatment and control subsamples (ps>.10). Second, we tested the balance between the subsample of stayers (i.e., the subsample that did not attrite by the time of follow-up data collection). The results in Appendix Table B.2 show that the characteristics of stayers are balanced across the treatment/controls arms as well. In addition, an omnibus test for the joint significance of all baseline characteristics confirms that treatment assignment cannot be predicted by the observed covariates for either the full sample or the subsample of stayers (ps>.10). Finally, attrition is insignificantly correlated with treatment status (p>.10). Hence, in this study, the randomization (and/or attrition bias) are unlikely to affect estimated ITT effect.

Figure 4 depicts the estimated ITT effects on primary and secondary outcomes. Parental training improves child cognition z-scores, on average, by 0.20 SD (p<.01). In addition, parental training has a small but significant and positive impact on child language and motor z-scores of 0.09 SD and 0.08 SD, respectively (ps<.10). With regard to the secondary outcomes of interest, the ITT effect estimates illustrate that parental training, on average, increases both parenting beliefs and parental investment by 0.17 SD and 0.26 SD, respectively (ps<.01).

Figure 4. Summary of regression analyses testing for treatment impacts of parental training program on ECD outcomes, parenting beliefs, and parental investment. Note. This figure presents ITT effects of cluster-randomized parenting experiments on ECD outcomes, parenting beliefs, and parental investment. The dependent variables are expressed in standardized z-scores (with a mean of 0 and SD of 1). The ITT effects are estimated using an ANCOVA specification. We control for the outcome variable of interest at baseline, control covariates, and study, cohort, and province fixed effects. Standard errors are clustered at the village-level.

Mediated and unmediated impacts

Table 5 presents the findings of the mediation analysis. In Column 1, we report the ITT effect on child cognition, language, and motor z-scores, controlling for a set of baseline child and household characteristics. In Columns 2 and 3, we report our estimations of direct ITT effects and the impact of each of the mediators. Finally, we estimate the direct ITT effects and the impacts of both of the mediators simultaneously, which we report in Column 4.

Table 5. Mediation analysis of impacts of parental training programs on ECD outcomes (in rural areas of Central and Western China).

	Bayley cognition raw z-score				Bayley language raw z-score				Bayley motor raw z-score
Variable	(1)	(2)	(3)	(4)	(1)	(2)	(3)	(4)	(1)	(2)	(3)	(4)
Treatment assignment (1= treatment; 0 = control)	0.20***	0.18***	0.16***	0.15***	0.09*	0.07	0.05	0.04	0.08*	0.06	0.06	0.06
	(0.05)	(0.05)	(0.05)	(0.05)	(0.05)	(0.05)	(0.05)	(0.05)	(0.04)	(0.04)	(0.04)	(0.04)
	[0.10–0.30]	[0.08–0.27]	[0.07–0.26]	[0.06–0.25]	[−0.01–0.18]	[−0.03–0.16]	[−0.04–0.14]	[−0.05–0.13]	[−0.01–0.16]	[−0.02–0.15]	[−0.02–0.15]	[−0.03–0.14]
Parenting beliefs factor z-score		0.12***		0.09***		0.11***		0.08***		0.06***		0.05**
		(0.02)		(0.02)		(0.02)		(0.02)		(0.02)		(0.02)
		[0.07–0.16]		[0.05–0.13]		[0.08–0.15]		[0.05–0.12]		[0.02–0.10]		[0.01–0.10]
Parental investment factor z-score			0.12***	0.11***			0.15***	0.13***			0.04**	0.03*
			(0.02)	(0.02)			(0.02)	(0.02)			(0.02)	(0.02)
			[0.09–0.16]	[0.07–0.14]			[0.12–0.18]	[0.10–0.17]			[0.01–0.07]	[-0.00-–0.06]
No. of observations	2,800	2,800	2,800	2,800	2,569	2,569	2,569	2,569	2,800	2,800	2,800	2,800
Adjusted R^2	0.12	0.13	0.14	0.14	0.13	0.14	0.15	0.15	0.08	0.08	0.08	0.08
Oster’s $\delta$		3.77	8.79			5.58	8.74			9.18	6.50

Note. In all regressions, we control for children’s sex, children’s age (in months), whether the child was born prematurely, whether the mother was the primary caregiver, parental age (in months), and province, cohort, and study fixed effects. All standard errors (in brackets) are clustered at the village-level; 95% confidence intervals are reported in square brackets. Oster’s $\delta$ is reported under the condition that the treatment effect is zero. We assume $R_{\mathit{\text{max}}}=1.3R^2$, as suggested by OOster (2019).

* p<.10; **p<.05; ***p<.01

The results of this mediation analysis show that a significant share of the treatment impact is mediated via parenting beliefs and parental investment. After controlling for treatment assignment to the parental training program, we find that the belief and investment factor z-scores significantly predict the cognition, language, and motor z-scores of young children (ps<.10). When our analysis included the mediator outcomes in the regression specification, the estimated direct impact of the parenting program on Bayley cognition, language, and motor scores were reduced, respectively, from 0.20 SD to 0.15 SD (ps<.01), from 0.09 SD (p<.10) to 0.04 SD (p>.10), and from 0.08 SD (p<.10) to 0.06 SD (p>.10). Considering that the estimates for the Oster’s δ of each regression specification range between 3.77 and 9.18, it can be concluded that there is evidence that the estimated coefficients of the mediated impacts are robust to unobserved heterogeneity.

Table 6 shows the decomposition of ${\tau }_{\mathit{\text{ITT}}}$ into ${\mu }_{\mathit{\text{ITT}}}$ and ${\upsilon }_{\mathit{\text{ITT}}}$. The mediated effect is further disaggregated in the share of the ${\tau }_{\mathit{\text{ITT}}}$ that is mediated via parenting beliefs, parental investment, and the multi-stage mediation effect via parenting beliefs and parental investment. As shown in Column 1 of Table 6, in total, 22.3% of the ITT effect on child cognition z-scores is mediated via the secondary outcomes of interest; 8.2, 12.1, and 2.0% of ${\mu }_{\mathit{\text{ITT}}}$ are attributable to the mediating role of parenting beliefs, parental investment, and the multi-stage mediation effect, respectively. For language development, we find that, in total, 53.0% of ${\tau }_{\mathit{\text{ITT}}}$ is mediated via the secondary outcomes of interest (Column 2); 16.5, 30.8, and 5.7% of ${\mu }_{\mathit{\text{ITT}}}$ are attributable to the mediating role of parenting beliefs, parental investment, and the multi-stage mediation effect, respectively. Finally, for motor scores, we find that, in total, 23.2% of ${\tau }_{\mathit{\text{ITT}}}$ is mediated via the secondary outcomes of interest (Column 2); 12.7, 9.0, and 1.5% of ${\mu }_{\mathit{\text{ITT}}}$ are attributable to the mediating role of parenting beliefs, parental investment, and the multi-stage mediation effect, respectively.

Table 6. Breakdown of total program impacts on ECD outcomes.

		Bayley cognition z-score	Bayley language z-score	Bayley motor z-score
Effect	Notation	(1)	(2)	(3)
Total mediated ITT effect	${\mu }_{\mathit{\text{ITT}}}$	0.044	0.045	0.017
		[0.029–0.062]	[0.030–0.062]	[0.006–0.029]
ITT effect mediated through parenting beliefs		0.016	0.014	0.01
		[0.007–0.027]	[0.007–0.026]	[0.003–0.020]
ITT effect mediated through parental investment		0.024	0.026	0.007
		[0.014–0.036]	[0.015-0.040]	[-0.001–0.015]
ITT effect mediated through parenting beliefs and parental investment		0.004	0.005	0.001
		[0.002–0.007]	[0.003–0.008]	[-0.000–0.003]
Unmediated ITT effect	${\upsilon }_{\mathit{\text{ITT}}}$	0.151	0.04	0.057
		[0.079–0.221]	[-0.032–0.118]	[-0.007–0.132]
Total ITT effect	${\tau }_{\mathit{\text{ITT}}}$	0.195	0.085	0.075
		[0.123–0.266]	[0.012–0.163]	[0.008–0.146]
% Of ${\tau }_{\mathit{\text{ITT}}}$ mediated		22.3%	53.0%	23.2%
		[20.8–23.9%]	[51.0–54.9%]	[21.6–24.8%]
% Of ${\tau }_{\mathit{\text{ITT}}}$ mediated through parenting beliefs		8.2%	16.5%	12.7%
		[7.2–9.3%]	[15.1–18.0%]	[11.5–14.0%]
% Of ${\tau }_{\mathit{\text{ITT}}}$ mediated through parental investment		12.1%	30.8%	9.0%
		[10.9–13.3%]	[29.0–32.7%]	[7.9–10.1%]
% Of ${\tau }_{\mathit{\text{ITT}}}$ mediated through parenting beliefs and parental investment		2.0%	5.7%	1.5%
		[1.5–2.6%]	[4.8–6.6%]	[1.1–2.1%]
No. of observations		2,800	2,569	2,800

Discussion and conclusion

In this study, we show that rural-urban differences in educational opportunities in China emerge early in life. One major finding is that, even before the age of 3, children in rural study sites across rural areas in Eastern, Central, and Western China are already lagging behind in their development as compared to their urban counterparts.^7,8 On the one hand, we find that, overall, urban children in Shanghai are developmentally on track with a prevalence of cognition delay of only 4.5%. The prevalence of motor skill delays (6.1%) also demonstrates that urban children do not suffer from developmental delays. Only levels of language skills (35.6%) are high, as compared to international norms. The healthy levels of development of urban children in the sample are in line with the research of ZZhang et al. (2021), who found that only 14% of all children in Shanghai were at risk of any developmental delay. On the other hand, our findings show that rural children are, respectively, 43.2, 18.3, and 20.7 percentage points more likely to be delayed in their cognition, language, and motor development than are their urban peers in Shanghai.

With regard to language development, we observe that, although language delay in urban China was high compared to international norms, the prevalence of language delay in rural China is even higher. Interestingly, the prevalence of language delay of rural sample children is more prevalent in the peri-urban sample from Chongming County in Eastern China than is the prevalence of language delay among sample children from rural areas in Central and Western China. Such a finding can perhaps be explained by earlier research that shows that the home language environment of young children in peri-urban areas tends to be less stimulating, with lower levels of child-directed speech than in the home language environment in remote rural areas; this explanation suggests that such patterns of delay may be due to more stringent time constraints and a higher risk of mental health issues of peri-urban caregivers as compared to caregivers in more remote rural communities (Feng et al. 2023; Ma et al. 2023).

sThe significantly higher rates of delay among rural children, compared to the rates of delay among urban children, are not surprising, given the income differentials in China between the relatively well-off families in China’s cities and the relatively poorer families in rural China. In 2021, the average disposable income per capita in urban Chinese households was approximately 2.5 times higher than in rural households (i.e., the disposable income per capita amounted to CNY 47,412 in urban households as compared to CNY 18,931 in rural households—National Bureau of Statistics of China 2022).⁹ The finding that children growing up in lower-SES families are at an elevated risk of developmental delay as compared to children in higher-SES families is in line with the global literature showing that the elevated risk of exposure to inadequate nutrition and poor environments in low-SES households can cause developmental delay (Bornstein and Bradley 2002; Falk et al. 2021; Heckman and Mosso 2014; Schady et al. 2015).

The study also provides evidence of the persistence of inequalities in ECD outcomes over age and across generations. In this regard, when comparing high and low-SES families, gaps in developmental outcomes of young children emerge as early as 6 months of age and do not change substantively by age 2.5. These findings are in line with evidence from other LMICs, including Chile, Colombia, Ecuador, India, Indonesia, Nicaragua, Peru, and Senegal (Fernald et al. 2012; Rubio-Codina et al. 2015; Schady et al. 2015). Considering that children who grow up in rural, lower-SES families are at a higher risk of developmental delay than are children from higher-SES families, we conclude that the results provide evidence of the intergenerational transmission of disadvantage in China (Bornstein and Bradley 2002; Falk et al. 2021; Heckman and Mosso 2014; Schady et al. 2015).

Despite the fact that the current environment in rural China is associated with high levels of developmental delays, research shows that parental training programs can function as a promising strategy to support the developmental potential of at-risk populations of young children in rural communities in China. Based on evaluations of randomized experiments, the literature demonstrates that parental training programs implemented in rural study sites in Central and Western China, on average, can improve child cognition, language, and motor development by 0.20 SD, 0.09 SD, and 0.08 SD, respectively. The finding of a significant positive treatment impact on child cognition of 0.20 SD is in line with the findings of earlier studies from rural China, including those of Luo, Emmers, et al. (2019), Sylvia et al. (2021), and Sylvia et al. (2022), who report treatment impacts of 0.11 SD to 0.28 SD on child cognition. Although the Luo et al. and Sylvia et al. studies do not detect evidence of significant treatment impacts on language or motor development, the large size of our pooled evaluation sample (n = 2,800) does allow for the detection of the relatively small, but positive and significant, impacts on language and motor development. Further, the detected impacts of 0.17 SD and 0.26 SD on parenting knowledge and parental investment, respectively, also are in line with impact sizes reported in the literature (Emmers et al. 2021).

In addition, the current study, by investigating underlying mechanisms of the positive treatment impacts on cognition, language, and motor development, can inform policymakers about how the design of cost-effective parental training programs can support the development of disadvantaged children at scale and advance progress toward Common Prosperity. For each of these three developmental delays, the findings show that 22.3–53.0% of the impact is mediated via parenting beliefs in the importance of ECD and the early home environment as well as parental investments in a stimulating home environment. The overall mediated impact consists of three parts: the directly mediated impact via an increase in parenting beliefs, directly mediated impact via an increase in parental investment, and indirectly mediated impact via an increase in parenting beliefs and a subsequent increase in parental investment. Finally, it is noteworthy that the finding that the mediators play a more important role in language development (i.e., 53.0% of the impact on language z-scores) than for cognition or motor development (i.e., 22.3 and 23.2% of the impacts on development in the respective domains) is likely due to the measurement of the secondary outcomes. Respectively, two and four out of five input measures that were used to construct the parenting beliefs and parental investment factor z-scores were related to the home language environment (Appendix Table A.2). Therefore, we believe that it is more likely that our estimates underestimate rather than overestimate the mediated impacts. We recommend that future research gather more detailed information on parenting beliefs and parental investment related to specific developmental domains (e.g., cognitive or motor development) to produce a more highly refined mediation analysis of domain-specific beliefs and investments.

We acknowledge a number of limitations to this study. First, the study took place in sites in four Chinese provinces. Although it is possible that the results may differ in other regions, we are confident that the evidence provided in this study on the rural-urban gap in ECD outcomes is the most representative evidence available for the whole of China. In addition, the current study explores the mediating role of parenting beliefs and parental investments for impacts of parental training programs on the ECD outcomes of the sample children. We acknowledge that alternative mediating pathways may exist (e.g., through disciplining practices and/or through the mental health of caregivers). Due to a lack of secondary outcome measures for the sample from Shanghai, the current study is unable to compare mediating pathways across urban and rural contexts. Future studies are needed to investigate a more diverse set of mediating pathways.

In conclusion, the evidence shows that early developmental delays are prevalent across rural children in study samples from Eastern, Central, and Western China, whereas urban children in urban Shanghai are developmentally on track. Moreover, we find that these developmental inequalities tend to persist over time (as children age from 6 to 30 months) and most likely will persist across generations. Providing young children with a fair start in life is foundational to making progress toward Common Prosperity. Parental training programs that focus on psychosocial stimulation are a promising tool to improve children’s early cognitive, language, and motor development. The impacts are mediated via gains in parenting beliefs and parental investment in a stimulating home environment. To bring parental training programs to scale and deliver them as regular child and family services, further research is needed to determine which program features are key to improving cost-effectiveness, inclusiveness, and scalability of programs. For example, future studies are needed to investigate which program components (e.g., one-to-one interactive reading activities versus group-based play activities) or mediating parental investments (e.g., playing games of peekaboo versus making pegboard puzzles) can be more effective at stimulating development in certain human capital domains.

Acknowledgments

We are grateful for the insightful comments we received from Prof. Dr. Fan Wang during the Penn CSCC-NYU Shanghai CASER Workshop on Inequalities and Child Development on November 18–19, 2022.

Disclosure statement

No potential conflict of interest was reported by the author(s).

Additional information

Notes on contributors

Dorien Emmers

Dorien Emmers is a postdoctoral research scholar of the Chinese Studies Research Group of KU Leuven in Belgium. She is an affiliated researcher of the Stanford Center on China’s Economy and Institutions and LICOS Centre for Institutions and Economic Performance at KU Leuven. Her research interests focus on the economics and measurement of human capital formation and social mobility. She has been involved in the design and evaluation of early childhood development interventions in rural China. She is currently working on the observation-based measurement of early skill formation and the adaptive design of parenting experiments. Dr. Emmers completed her PhD in Economics at KU Leuven LICOS with a doctoral dissertation on “The Experimental Economics of Parenting: Evidence from Rural China.” Before joining the KU Leuven Chinese Studies Research Group, she was a postdoctoral research scholar at the Stanford Center on China’s Economy and Institutions.

Hui Wang

Hui Wang is an Associate Professor at the Shanghai Jiao Tong University School of Medicine. Prof. Dr. Wang obtained her PhD from the University of Hong Kong in 2016. She is an epidemiologist and her main research focuses on two areas: (1) investigating the effect of adverse environmental exposures in early life on neurodevelopmental outcomes in offspring using large population-based health databases; (2) examining the neonatal factors that may influence the risk of neurodevelopmental disorders or cardiovascular diseases in childhood and adulthood. She is also a principal investigator of the “Shanghai Birth Cohort” and has wide experience in birth cohort design and implementation of fieldwork.

Lei Wang

Lei Wang is a professor at the International Business School of Shaanxi Normal University. Prof. Wang received her PhD from Xi’an Jiaotong University in 2006. Her research focuses on the development of finance and economy in China. Prof. Wang’s current research evaluates the impact of parenting interventions on child developmental outcomes in rural China. She has conducted and evaluated large-scale intervention programs and policies to improve early childhood development in rural China. Her research has been published in a variety of high-quality academic journals, including World Development, Journal of Comparative Economics, Agricultural and Resource Economics, Journal of Development Economics, China & World Economy, and China Agricultural Economic Review.

Siqi Zhang

Siqi Zhang is a postdoctoral research scholar at the Stanford Center on China’s Economy and Institutions. She received her PhD in Economics from Shaanxi Normal University in 2021. Her research interests are in the fields of development economics and public policy. Her research focuses, in particular, on the economics of education and human capital in rural China. She has extensive experience with managing data collection and fieldwork. Using data from large-scale randomized controlled trials, she has evaluated the impacts of parental training interventions on child development outcomes and the impacts of a micronutrient fortification intervention on anemia among young children in rural China. Her current research focuses on the long-term persistence of the intervention impacts of a six-month early childhood home-visiting program on child development outcomes at school age.

Yiwei Qian

Yiwei Qian is an associate professor at the Southwestern University of Finance and Economics. Prof. Qian’s research interests involve the fields of development economics, labor economics, and the economics of social networks. In particular, his research focuses on human capital challenges in low-income contexts, especially rural China, and related econometrics topics in program evaluation. He has been involved in a series of randomized controlled trials to study the effective delivery of early childhood interventions at scale in rural China. Yiwei completed his Ph.D. in Economics at the University of Southern California in 2021. Before joining the Southwestern University of Finance and Economics, Yiwei Qian was a joint postdoctoral research scholar of the Stanford Center on China’s Economy and Institutions and the Stanford Graduate School of Business.

Jinsong Zhang

Jinsong Zhang is a psychiatrist, the chief physician and Associate Professor in Xinhua Hospital, which is affiliated to the School of Medicine of Shanghai Jiao Tong University. She is the Director of the Department of Medical Psychology of Xinhua Hospital. She received her BS and MS from the Health Science Center of Peking University and her PhD in Psychology from East China Normal University. She currently holds a position as a board member of several professional organizations in China, including the position of Vice Director of the Psychotherapy and Psycho-consultation Committee of the Chinese Mental Health Association, the Developmental Behavior Committee of the Chinese Medical Association, and a position as a one of the Vice Presidents of the Shanghai Mental Health Society. Her main research interests focus on children’s development and mental health disorders, such as child temperament, self-regulation, ADHD, ODD/CD, and early intervention for crisis and trauma. Her research and clinical work on ADHD with ODD/CD cover a wide age range from preschool to adulthood. She has published nearly 100 papers in Chinese and English academic journals.

Yunting Zhang

Yunting Zhang is an Associate Professor at the Shanghai Jiao Tong University School of Medicine. Prof. Dr. Zhang has made important contributions to research on child development in China, with a special focus on child development during the first years of life. Her research has been published in leading academic journals, including in The Lancet, JAMA Pediatrics, BMJ Open, Plos One, International Journal of Epidemiology, BMC Public Health, BMC Pediatrics, and International Journal of Endocrinology.

Jun Zhang

Jun Zhang is K. C. Wong Chair Professor at the Shanghai Jiao Tong University, the PI of the Shanghai Birth Cohort, and Chair at the Department of Maternal and Child Health, Shanghai Jiao Tong University School of Public Health. His research interests cover reproductive and perinatal epidemiology, and the effects of environmental exposure in early life on maternal and child health. He received his MB from Fudan University, China, certified by USMLE, and his PhD from University of North Carolina at Chapel Hill, USA. He has published over 400 peer reviewed papers in journals like New England Journal of Medicine, Lancet, and JAMA. Prof. Zhang was an Assistant Professor at Mount Sinai School of Medicine in New York, an Investigator and then a tenured Senior Investigator at the National Institutes of Health, U.S. Currently, he is a committee member of the International Federation of Gynecology and Obstetrics (FIGO).

Fan Jiang

Fan Jiang is the Chair of the Deans Council of Shanghai Jiao Tong University School of Medicine. Prof. Dr. Jiang has served as Vice Dean of the Shanghai Jiao Tong University School of Medicine since February 2018. Prof. Dr. Jiang is also a Senior Fellow of the Equity Initiative of the China Medical Board and is directing the China team of the “For Our Children” project, supported by the China Medical Board. In addition to Prof. Dr. Jiang’s participation in the prominent Equity Initiative of the China Medical Board, she has been involved in extensive research on early childhood development in rural areas of China. As a developmental and behavioral pediatrician with a research focus on the social determinants of child health and early development, she has devoted her career to studying the development of children and adolescents. Her research has been published in a wide range of leading journals that include The Lancet, JAMA Pediatrics, Pediatrics, Int J Epidemiol, and Sleep. As a physician, researcher, and educator, Prof. Dr. Jiang has contributed significantly to the healthy development of children and adolescents.

Scott Rozelle

Scott Rozelle is the co-director of the Stanford Center on China’s Economy and Institutions. He also is a Senior Fellow of the Freeman Spogli Institute for International Studies at Stanford University and the Stanford Institute for Economic Policy Research. Prof. Rozelle’s research focuses almost exclusively on identifying interventions for improving the human capital of rural communities in China. He received his BS from the University of California, Berkeley, and his MS and PhD from Cornell University. His research has been published in top academic journals, including Science, Nature, American Economic Review, and Journal of Economic Literature. Previously, Prof. Rozelle was a professor at the University of California, Davis, and an assistant professor in Stanford’s Food Research Institute and Department of Economics. He currently is a member of several organizations, including the American Economic Association, International Association for Agricultural Economists, and Association for Asian Studies.

Notes

1 Caregiver-child interaction and communication appear to be universal aspects of effective parenting and healthy child development. However, it is also known that parents and children can interact and communicate in certain ways that are adapted to and consistent with their cultural context (Bornstein 2012).

2 For the norm-referenced Chinese version of the Bayley-I test with a mean of 100 and SD of 15, the conventional cut-off score of 85 is used. In the absence of a Chinese norming sample, Bayley-III cut-off scores are defined based on the means (SDs) of Bayley-III composite scores for healthily developed populations that have been reported in the international literature (i.e., the mean (SD) is expected to be 105 (9.6), 109 (12.3), and 107(14) for the cognitive, language, and motor scales, respectively; Wang et al. 2019). Hence, the Bayley-III cut-off points for moderate cognitive, language, and motor delay are 95.4, 96.7, and 93, respectively.

4 By calculation of the sample-size weighted average of the prevalence of delay among young rural children in Chongming and areas in Central and Western China, we obtain an aggregate measure of the prevalence of delay across all rural study sites.

5 In addition, Table 4 shows that ECD outcomes are consistently higher for young girls than for young boys (p_s<.01). Cognition, language, and motor scores are higher for term-born children than for preterm-born children in the rural samples from Central and Western China (p_s<.05). These differences in subgroup means for term versus preterm-born children are insignificant for the Shanghai sample.

6 Note that the gap in cognitive skill development between the bottom quartile of children and the remainder of the sample narrow between the age of 6–11 months and 12–17 months. This early, small, but noticeable catch-up effect may be caused by a micronutrient fortification program that provided multiple micronutrient powders for young children to half of the caregivers in the sample. As reported in LLuo et al. (2017), the micronutrient fortification program that was initiated between the age of 6–11 months had a small positive impact on child cognitive development, but this positive impact faded out afterwards (i.e., when the sample children were 18 and 24 months).

7 The reported risks of cognitive, language, and motor delay of 46.7, 50.2, and 25.8%, respectively, that are reported for rural children from Central and Western China are in line with the risks of delay reported in earlier studies from rural study sites in Central and Western China. For example, Wang et al. (2019) report a 49, 52, and 30% risk of cognition, language, and motor delay, respectively; and Emmers et al. (2021) report a 45 and 46% risk of cognition and language delay, respectively.

8 Earlier evidence on the risk of developmental delay in urban areas of China is scarce. One study used the Bayley-III cognition scale to assess child cognitive development of 1,444 infants and toddlers in middle-sized Chinese cities (Hua et al. 2019). In another study, WWang et al. (2023) administered the Bayley-III cognition, language, and motor scales to 70 young children in Xi’an. The findings of the HHua et al. (2019) and WWang et al. (2023) studies confirm that children in urban areas of China resemble a healthily developed population with a prevalence of delay of approximately 15% or lower.

9 For the study regions of the current study, 2021 statistical yearbook data show that the average disposable income per capita for urban Shanghainese households was CNY 51,295, while the average disposable income per capita in Chongming and rural areas of Yunnan, Hebei, and Shaanxi was CNY 33,140, CNY 14,197, CNY 18,179, and CNY 14,745, respectively (National Bureau of Statistics of China 2022).

Appendix A.

The measurement system

Appendix Part A describes the dimensionality reduction techniques and measurement systems for parenting beliefs and parental investment in interactive caregiver-child parenting activities. We first use exploratory factor analysis (EFA) to determine the number of factors we need to extract from our list of measures of parental beliefs and parental investments. We use Cattell’s scree plot (Cattell 1961) and Horn’s parallel analysis (Horn 1965) to determine the number of factors to be retained (see Appendix Table A.1 and Appendix Figure A.1).

Both factor selection methods indicate that we should use a one-factor measurement model for parenting beliefs and parental investment. We then estimate a dedicated measurement system. The factor loadings are reported in Appendix Table A.2. First, we find that the five measures of parenting beliefs load positively on the parenting beliefs factor. Second, the five measures of parental investment in interactive caregiver-child activities such as storytelling and play with toys load positively on the parental investment factor. All measures load relatively strongly on the latent factors hence we retain all measures and estimate means and factor loadings. We next predict factor scores for parenting beliefs and parental investment. Finally, we internally standardize the factor scores within the age (in months) group of the child to obtain standard normally distributed factor z-scores that are uncorrelated with children’s age.

Figure A.1. Scree plots of the eigenvalues of the principal components of parenting beliefs and parental investment.

Table A.1. Exploratory factor analysis to determine the number of latent factors.

Variable	Cattell’s scree plot	Horn’s parallel analysis
Parenting beliefs	1	1
Parental investment	1	1

Table A.2. Estimated factor loadings and communality.

Factor	Factor loading	Signal-to-noise ratio
Parenting beliefs factor
Acknowledges importance of play for child development	1	0.21
Knows how to play with child	1.30	0.35
Acknowledges importance of reading for child development	1.13	0.27
Knows how to read books to child	1.27	0.37
Acknowledges importance of stimulative home environment	0.58	0.12
Parental investment factor
Uses toys to play with child on a daily frequency	1	0.10
Tells stories to child on a daily frequency	1.88	0.64
Sings songs to child on a daily frequency	1.10	0.13
Reads books to child on a daily frequency	1.40	0.55
No. of picture books for children at home	7.28	0.08

Note. We normalize the factor loading of the first measure for each factor to 1. The signal-to-noise ratio is the extent to which the variance in each of the outcome measures is driven by signal relative to noise.

Appendix B.

Cluster-randomized parenting experiments in rural areas of Central and Western China

B.1. Study design

For treatment impact estimation and the causal mediation analysis, the study pools data from three clustered RCTs in rural study sites in Central and Western China. Taken all together, the three RCTs cover 274 villages in 28 nationally designated poverty counties in rural Shaanxi, Yunnan, and Hebei Provinces (see Appendix Figure B.1). In each study, treatment assignment was randomized at the village level to reduce the risk of contamination (among families within a village) across experimental groups.

The home visitation program in Shaanxi Province was implemented for the duration of 6 months from November 2014 until April 2015. Baseline data were collected in October 2014 and follow-up data were collected in May 2015. The remaining two intervention programs were implemented for the duration of one year. The home visitation program in Yunnan and Hebei started in October 2015. Baseline and follow-up data were collected in September and October 2015 and 2016, respectively. The center-based program in Shaanxi Province was implemented in two waves. Baseline and follow-up data for the first wave were collected in August 2016 and 2017, respectively. Baseline and follow-up data for the second wave were collected one year later. Approval for the studies was obtained from the provincial offices of the National Health Commission (NHC) in Shaanxi, Yunnan, and Hebei Provinces and from the ethics committee of Stanford University (USA, number 25734) on 26 October 2012 and renewed annually.

Figure B.1. Trial profiles of the three intervention studies. Note: In the home visitation program in Shaanxi Province, 79 caregiver-child dyads in treatment group villages were randomly assigned to control at the individual level. As shown in Sylvia et al. (2021), no evidence was detected of spillover effects of the parenting intervention on child development or parenting outcomes of the untreated households in treatment group villages. In addition, in the study in Panel A, the Griffith Mental Development Scales were administered instead of the Bayley scales to assess ECD outcomes of n = 300 children. This subsample is excluded from the evaluation sample.

B.2. Participants

Sample households were selected based on a list of all registered births in the selected clusters provided by the local office of the NHC. First, in the home-based program in Shaanxi, all children aged 18–30 months on 1 November 2014 were enrolled in the study together with their primary caregivers. Second, in the home-based program in Yunnan and Hebei, all children aged 6–18 months on 15 September 2015 were enrolled in the study together with their primary caregivers. Third, in the center-based program in Shaanxi, all children aged 6–24 months on August 2016 and August 2017 were enrolled in the study together with their primary caregivers. On average, 14 children were enrolled in each village cluster. The resulting baseline sample includes 3,643 caregiver-child dyads from 274 villages. Complete records of test item scores of the Bayley Scales of Infant and Toddler Development (i.e., the main outcome of interest in this study) were collected for 3,571 infants and toddlers. Due to internal migration, 14.4% of the caregiver-child dyads attritted by the time of follow-up data collection. As a consequence, follow-up data were collected for 3,119 children living in 273 villages. As described in detail in Sylvia et al. (2021), the Griffith Mental Development Scales were administered instead of the Bayley scales to n = 300 children in the home visitation program in Shaanxi Province that were over 30 months of age at endline. This subsample is excluded from the evaluation sample. Complete records of Bayley item scores were collected for 2,800 children at follow-up and included in the impact evaluation sample.

The caregivers of all children gave oral informed consent before enrollment and baseline data collection. Using Zelen’s randomized consent design, we sought participant consent conditional on the assigned treatment status. All caregivers were asked to give their informed consent for data collection. Prior to the start of the intervention, caregivers assigned to the intervention group were additionally asked to give consent for participation in the intervention.

B.3. Randomization and masking

After baseline data collection, the clusters were randomly assigned to control and treatment within each study using a computerized random number generator. In the home visitation programs in Shaanxi, the randomization procedure was stratified by county and study cohort. In the home visitation program in Yunnan and Hebei and the center-based program, the clusters were randomly assigned to control and treatment within township-level and county-level strata, respectively. Caregivers were unaware that they were involved in an experiment and control group households were unaware of the intervention.

B.4. Procedures

All children and caregivers residing in clusters of the intervention group were enrolled for a home-based or center-based parental training program focusing on child psychosocial stimulation. In the home-based programs in Shaanxi and in Yunnan and Hebei, caregiver-child dyads were invited to participate in 24 (weekly or bi-weekly) training sessions over the duration of six months or one year, respectively. In the center-based program in Shaanxi, caregiver-child dyads were invited to participate in 48 weekly training sessions over the duration of one year. During each of these parental training sessions, community health workers (paid and managed by the research team) trained caregivers on interactive caregiver-child activities (e.g., storytelling or play with toys). Training sessions were designed to last for one hour.

All interactive caregiver-child activities were stage-based and fully scripted in an ECD curriculum. This curriculum was based loosely on the curriculum developed for a pioneering parenting intervention in Kingston, Jamaica and adapted to Chinese settings by local child development psychologists and ECD experts (Sylvia et al. 2021). Each activity focused on one of the following four skill domains: cognition, language, motor, or social-emotional development. During each training session two new activities were introduced (i.e., parents were instructed on practicing these activities with their children) such that caregivers would encounter at least one new activity to stimulate development in each of the four targeted skill domains each month. In addition to developmental activities, the curriculum also included one module with basic information on child health and nutrition that was delivered at the end of each session.

The parental training sessions were delivered at home or at a parenting center. On the one hand, in the home visitation programs, the parental trainers scheduled a suitable day and time with each of the primary caregiver to frequent the home of the household and deliver the training session. The parental trainer would bring the required counselling materials and leave toys and books with the household such that caregivers could practice the activities with the child during the week. On the other hand, in the center-based program, a parenting center was installed at a central location in the community (e.g., in a repurposed school building, cultural center, or office space) provided by the village committee. The parenting centers were designed to be open five hours per day, six days per week (Sylvia et al. 2022). Each of the parenting centers contained a room for one-on-one parental training sessions, a library with standard sets of toys and children’s books, and a large child-friendly open play area. In addition to attending one-on-one parental training sessions, caregivers were encouraged to bring their children to the parenting centers during open hours to play ad libitum in the open play area, but they were not allowed to leave their children alone in the parenting centers (i.e., this was not day care).

Training sessions were delivered by community health workers in the service of the NHC. To provide the community health workers with the necessary knowledge, skills, and tools to instruct caregivers on interactive caregiver-child reading and play activities, the community health workers received a copy of the fully-scripted ECD curriculum and an ECD toolkit (including a growth chart, development check list, counselling materials, and user handbook). In addition, before the start of the intervention, each of the community health workers had to complete a one-week intensive training course on basic theories of ECD and on communication, coaching, and counselling techniques. The training program consisted of five days of classroom-based instruction by ECD experts in combination with two days of field practice. Fully-scripting the curriculum obviated the need for more extensive training of community health workers, which significantly reduced the overall costs of program delivery. Throughout the program, trainers received periodic refresher training by phone (in the home visitation programs) or in person (in the center-based program) on curriculum activities, which would vary according to the ages of children to whom they were assigned.

To monitor service quality, a carefully designed supervisory system was set up. First, the compliance of the community health workers was tracked by phone (in the home-based program in Shaanxi), by mobile app (in the home-based program in Yunnan and Hebei), or by the manager of each parenting center (in the center-based program). Second, county-level or township-level officials of the NHC randomly conducted unannounced supervisory visits of a training session of each instructor and provided feedback and reinforcement training. Third, phone interviews were conducted to invite caregivers to give feedback on the quality of home visits. All intervention households received phone calls on a revolving or random basis. Course content of the reinforcement training sessions was adjusted according to caregiver feedback.

B.5. Balance

At baseline, 149 and 163 villages (with n = 1,798 and n = 1,658 caregiver-child dyads) were randomly assigned to the treatment and control group, respectively. Table B.1 presents summary statistics on children and household characteristics. Just over half (51.6%) of the children in our sample were male. No more than 3.8% of the sample belonged to an ethnic minority group other than the Han ethnicity. The children were, on average, 14 months old at baseline. Additionally, 47.5% of the children were first born and had no older siblings. Approximately 4.4% of the children were born with a birthweight below 2.5 kg, while 5.3% of them were born before 37 weeks of gestation.

More than three-quarters (79.4%) of the mothers were living at home and taking care of the child at the time of baseline data collection. The mothers of the sample children were on average 27.9 years old and had completed 9.3 years of education. This implies that the average level of educational attainment was no higher than junior high school. In the sample, 11.1% of the families were recipients of social security support packages.

In order to ensure that the randomization procedure was successful, we conducted two-tailed t-tests to check whether the child and household characteristics were balanced across the treatment and control groups at baseline. First, we conducted a balance test for the full baseline sample (see right column of Table B.1). We detect no significant differences in the child and household characteristics of the treatment and control group (p_s > .10). Second, we tested the balance between the subsample of stayers (i.e., the subsample that did not attrite by the time of follow-up data collection, n = 2,800 caregiver-child dyads). The results in Appendix Table B.2 show that the characteristics of stayers are balanced across the treatment arms as well. In addition, an omnibus test for the joint significance of all baseline characteristics confirmed that treatment assignment cannot be predicted by the observed covariates for either the full sample or the subsample of stayers (p_s > .10). Hence, we find no evidence indicating that the control and treatment groups were unbalanced at baseline.

Table B.1. Baseline characteristics by study arm.

	Full sample	Control	Treatment	(2) vs. (3), p-Value
	(1)	(2)	(3)
Panel A: Child characteristics
Child is male (1 = yes; 0 = no)	0.516	0.527	0.505	.133
	(0.008)	(0.012)	(0.009)
Age (in months)	14.029	14.173	13.897	.700
	(0.127)	(0.203)	(0.153)
Child belongs to ethnic minority (1 = yes; 0 = no)	0.038	0.034	0.041	.753
	(0.009)	(0.012)	(0.014)
Child is firstborn (1 = yes; 0 = no)	0.475	0.470	0.480	.491
	(0.011)	(0.016)	(0.014)
Prematurely born (1 = yes; 0 = no)	0.053	0.049	0.056	.312
	(0.004)	(0.005)	(0.005)
Birthweight <2.5 kg (1 = yes; 0 = no)	0.044	0.044	0.044	.959
	(0.004)	(0.006)	(0.005)
Bayley cognition raw z-score	−0.023	−0.019	−0.026	.731
	(0.026)	(0.035)	(0.039)
Bayley language raw z-score	0.000	−0.017	0.016	.632
	(0.028)	(0.035)	(0.042)
Bayley motor raw z-score	−0.002	0.018	−0.020	.451
	(0.030)	(0.038)	(0.045)
Panel B: Household characteristics
Mother at home (1 = yes; 0 = no)	0.794	0.803	0.786	.321
	(0.009)	(0.013)	(0.013)
Maternal age (in years)	27.882	27.833	27.927	.920
	(0.122)	(0.175)	(0.170)
Maternal education (in years)	9.350	9.341	9.357	.884
	(0.079)	(0.117)	(0.106)
Household SES score	2.866	2.751	2.972	.322
	(0.093)	(0.137)	(0.126)
Recipient of social security support packages (‘dibao’; 1 = yes; 0 = no)	0.111	0.111	0.112	.692
	(0.007)	(0.009)	(0.010)
No. of observations	3,456	1,658	1,798

Notes: p-Values account for clustering at the village level and for stratification at the cohort level. An omnibus test across all baseline covariates shows that the two groups are well balanced at baseline. We test this by regressing the indicator of treatment assignment on the complete set of baseline characteristics and testing whether the coefficients are jointly equal to zero (p-value = .482).

Table B.2. Baseline characteristics of stayers by study arm.

	Full sample	Control	Treatment	(2) vs. (3), p-Value
	(1)	(2)	(3)
Panel A: Child characteristics
Child is male (1 = yes; 0 = no)	0.506	0.523	0.492	.076
	(0.009)	(0.014)	(0.012)
Age (in months)	14.205	14.361	14.067	.733
	(0.139)	(0.229)	(0.162)
Child belongs to ethnic minority (1 = yes; 0 = no)	0.042	0.042	0.042	.945
	(0.010)	(0.015)	(0.014)
Child is firstborn (1 = yes; 0 = no)	0.461	0.459	0.463	.703
	(0.012)	(0.018)	(0.015)
Prematurely born (1 = yes; 0 = no)	0.053	0.052	0.054	.702
	(0.004)	(0.007)	(0.006)
Birthweight <2.5 kg (1 = yes; 0 = no)	0.044	0.045	0.044	.922
	(0.004)	(0.006)	(0.006)
Bayley cognition raw z-score	−0.023	−0.016	−0.029	.605
	(0.026)	(0.037)	(0.036)
Bayley language raw z-score	0.018	0.008	0.027	.851
	(0.029)	(0.037)	(0.043)
Bayley motor raw z-score	0.003	0.018	−0.010	.570
	(0.030)	(0.042)	(0.043)
Panel B: Household characteristics
Mother at home (1 = yes; 0 = no)	0.806	0.822	0.792	.101
	(0.009)	(0.012)	(0.014)
Maternal age (in years)	27.990	27.915	28.055	.838
	(0.133)	(0.187)	(0.188)
Maternal education (in years)	9.284	9.267	9.298	.922
	(0.082)	(0.123)	(0.110)
Household SES score	2.851	2.737	2.951	.420
	(0.094)	(0.139)	(0.126)
Recipient of social security support packages (‘dibao’; 1 = yes; 0 = no)	0.113	0.115	0.112	.839
	(0.007)	(0.010)	(0.010)
No. of observations	2,800	1,312	1,488

Notes: p-Values account for clustering at the village level and for stratification at the cohort level. An omnibus test across all baseline covariates shows that the two groups are well balanced at baseline. We test this by regressing the indicator of treatment assignment on the complete set of baseline characteristics and testing whether the coefficients are jointly equal to zero (p-value = .585).