Data analysis

We first examine the prevalence of ECD delays reported in studies across rural communities in China. We define developmental delay as a cognitive, language or social-emotional development score of 1 or more SDs below the mean of a reference population whose developmental trajectory is expected to be normal (ie, a child population in developed regions who was not prematurely born, severely malnourished, or severely diseased; see online supplemental appendix 2 for more detail). This is in line with the guidelines of the BSID (the gold standard of ECD measurement), which conventionally defines children with a score of more than 1 SD below the normative sample mean as mildly delayed in their development.²¹ Using 1 SD below the normative mean as a cut-off to capture all severities of developmental impairment is also in line with definitions from the Global Research on Developmental Disabilities Collaborators and the American Association on Intellectual and Developmental Disabilities.²² Many academic studies have used 1 SD below the normative mean as their cut-off as well.^{23 24} We report and compare the shares of each study sample who is delayed in cognitive, language and social-emotional development. We conduct a meta-analysis of the prevalence of delay by pooling prevalence of delays across studies using a DerSimonian and Laird random-effects model.

In addition to presenting risks of delay among our sampled children, we also investigate parental engagement in stimulating caregiver practices in rural China. We use indicators of the FCI survey, a validated survey instrument developed by UNICEF to evaluate the quality of the home environment (see table 1A).^{25 26} We report on the share of parents in each study who engaged in interactive reading, storytelling or singing in the days prior to the survey. These three indicators of the FCI survey are selected because they are most commonly reported in the literature on cognitively stimulating parenting practices in rural China. In addition, we compute aggregate shares of rural caregivers in China who engage in stimulating parenting practices. We conduct a meta-analysis of the prevalence of certain parenting practices by pooling the prevalence of parenting practices across studies using a DerSimonian and Laird random-effects model.

Summary of included studies

*Modified FCI assessed parenting practices by the primary caregiver in the day before the survey.

†MICS/FCI assessed parenting practices by the child’s mother/father in the three days before the survey.

‡Wang et al (2019) use a pooled sample, including both original survey data (not previously reported) and survey data from previously reported studies, also included in this review. For our analysis, we include only the original data reported in Wang et al (2019). The remaining observations are reported in Luo et al (2019), Emmers et al (2020) and Zhong et al (2017).

§ECD centres were installed at centrally located places in intervention communities. Caregivers could decide how often they wanted to frequent parenting centres and/or participate in group reading or play activities at the centres.

¶Wang et al (2020) evaluates the persistence of treatment effects 2.5 years after program completion.

ASQ, Ages and Stages Questionnaires; BSID, Bayley Scales of Infant Development; C, ECD centre; ECD, early childhood development; FCI, family care indicators; G, group sessions; H, home; N, nutrition; O, one-on-one sessions; P, play area; RCT, randomised controlled trial; S, psychosocial stimulation.

Third, for each of the impact evaluation studies identified by our systematic search, we present the impact on children’s cognitive, language and social-emotional development. To facilitate the comparison of treatment impacts across studies, we again conduct a frequentist meta-analysis of the estimated treatment impacts expressed in standardised mean differences, using a DerSimonian and Laird random-effects model. Of the studies that yield significant impacts on one or more indicators of child development, we then identified features common to all or most of the interventions to draw policy lessons for implementers and future researchers.

We use the risk of bias tool for prevalence studies developed by Hoy et al to assess the risk of bias for prevalence studies of developmental delay and parenting practices.²⁷ We use the Effective Public Health Practice Project quality assessment tool to assess the risk of bias for RCTs.²⁸ Two reviewers independently rated the risk of bias of each study. Any disagreement was resolved by consensus with a third member of the review team. We then synthesised data in tabular formats. Further, we graded the overall certainty of evidence on the effectiveness of parental training programmes using the Grading of Recommendations Assessment, Development and Evaluation (GRADE) approach.

Finally, we assess heterogeneity in risk of delay and treatment impacts across studies by programme curricula and delivery mode, measurement tool and geographical location. We use frequentist meta-analysis to assess heterogeneity in prevalence rates of delay and estimated treatment impacts expressed in standardised mean differences based on a DerSimonian and Laird random-effects model. The statistical analysis was conducted with Stata version 16.1.