Joint Effects of Parenting and Nutrition Status on Child Development : Evidence from Rural Cambodia

Substantial work has demonstrated that early nutrition and home environments, including the degree to which children receive cognitive stimulation and emotional support from parents, play a profound role in influencing early childhood development. Yet, less work has documented the joint influences of parenting and nutrition status on child development among children in the preschool years living in low-income countries. Using panel data on parenting, nutrition status, and early developmental outcomes of about 7,000 Cambodian preschool-age children, this paper demonstrates that inequities in early development associated with family wealth are evident at the start of preschool and increase over time. A significant share of these inequalities can be explained by differences in parental stimulation and early nutrition status. Better educated parents engage in better parental activities that stimulate children's development. However, the positive association between parental activities and child outcomes is particularly strong for non-stunted children, and parental activities can only explain about 8-14 percent of the cognitive gap between the lowest and highest wealth quintiles. The results highlight the need for integrated interventions that address both parenting and early nutrition, also suggesting that parenting interventions for the most disadvantaged families should be carefully designed and evaluated to ensure maximum effectiveness.

Substantial work has demonstrated that early nutrition and home environments, including the degree to which children receive cognitive stimulation and emotional support from parents, play a profound role in influencing early childhood development. Yet, less work has documented the joint influences of parenting and nutrition status on child development among children in the preschool years living in low-income countries. Using panel data on parenting, nutrition status, and early developmental outcomes of about 7,000 Cambodian preschool-age children, this paper demonstrates that inequities in early development associated with family wealth are evident at the start of preschool and increase over time. A significant share of these inequalities can be explained by differences in parental stimulation and early nutrition status. Better educated parents engage in better parental activities that stimulate children's development. However, the positive association between parental activities and child outcomes is particularly strong for non-stunted children, and parental activities can only explain about 8-14 percent of the cognitive gap between the lowest and highest wealth quintiles. The results highlight the need for integrated interventions that address both parenting and early nutrition, also suggesting that parenting interventions for the most disadvantaged families should be carefully designed and evaluated to ensure maximum effectiveness. This paper is a product of the Development Research Group, Development Economics. It is part of a larger effort by the World Bank to provide open access to its research and make a contribution to development policy discussions around the world. Policy Research Working Papers are also posted on the Web at http://www.worldbank.org/research. The authors may be contacted via jberkes@diw.de or dfilmer@worldbank.org.

Joint Effects of Parental Stimulation and Nutrition Status on Child Development:
Evidence from Rural Cambodia

1) Introduction and Background
Early childhood development sets the course for lifelong learning, health and well-being (Black et al., 2017). Despite the potential social capital that can be built through investments in early childhood development (e.g. Heckman, 2006), recent estimates suggest that 250 million children living in low-and middle-income countries are at risk of not fulfilling their developmental potential (Black et al., 2017).
Family poverty is associated with several conditions that negatively impact child development. Children who experience early stunting, poor health or lack of educational opportunities are less likely to achieve their developmental potential than other children (Britto et al., 2017), with a common underlying influence of family poverty. Family poverty has been shown to have a wide range of negative effects on child development in many countries, with gradients in child development due to family income evident as early as 4 months of age and growing throughout the course of childhood (Paxson and Schady, 2007;Naudeau et al., 2011, Fernald, Kariger, Hidrobo & Gertler, 2012Schady et al., 2015;World Bank 2018).
Yet, the mechanisms accounting for the correlation between family wealth and child development remain imperfectly described and likely include a range of factors, including health and nutrition as well as the quality of home environments (e.g., McLoyd, 1998;Black et al., 2017). While perinatal (in-vitro and during the first months after birth) health and nutrition conditions are known to impact later cognition (Veena, Gale, Krishnaveni, Kehoe, Srinivasan, & Fall, 2016), little is known about how parenting may help compensate for disparities in development due to poverty.

Impacts of parenting and family wealth on child development
Defining the mechanisms that link family assets to child development is critical for developing effective interventions. One such mechanism is parenting. Parenting quality has been shown to have strong and long-term impacts on child development (Bradley & Corwyn, 2002; see Bornstein, 2005, for a complete review). Parenting has been posited to be a central mediator of the relationship between family poverty and negative child outcomes (Blair & Raver, 2010).
Parenting quality may be lower among families living in poverty due to the stress of living without adequate economic resources and the concomitant characteristics of poverty, such as low levels of parental education, which in turn contributes to the negative impacts of poverty on children. Thus, a key hypothesis is that parenting quality mediates the association between family poverty and child development.
Several studies have examined the hypothesis that undernutrition arises in part from poorquality parenting, particularly mediated through lack of maternal responsiveness. Results have been mixed. Maternal depression and other common mental disorders in the first two years after a child's birth have been associated with child underweight and stunting (Nyugen et al., 2013;Surkan, Kennedy, Hurley & Black, 2011). This effect may be mediated through lack of engaged and stimulating parenting which in turn influences child feeding and growth, but sources of this association are not yet clear. A recent meta-analysis of interventions focused on improving "responsive feeding" yielded few studies documenting positive impacts of responsive feeding, which suggests that responsive parenting is not a primary determinant of undernutrition (Bentley, 4 Wasser & Creed-Kanashiro, 2011). In sum, while existing literature suggests that maternal mental disorders and undernutrition may be associated, the evidence supporting a general association between parenting quality and undernutrition is mixed, suggesting that parenting and undernutrition may influence child development through mostly distinct pathways.
Parenting quality clearly has important impacts on child development overall. Both child development and economic theories suggest two sources of variation in parenting on child development: first, that different types of parenting may affect different domains of child development and, second, that parenting may have variable effects on child development depending on child characteristics and environmental conditions.

Role of different types of parenting on child development.
In interactions with young children, parents engage in multiple forms of teaching and support. Parenting is the primary mechanism by which children's health and nutrition, as well as overall growth and development are facilitated (Phillips & Shonkoff, 2000). Emotionally responsive parenting refers to parents' abilities to respond sensitively to children both verbally and physically, while cognitively stimulating parenting refers to parents' stimulation of learning through use of language, introduction of new games and ideas, as well as exposure to high-quality learning environments (Bornstein & Putnick, 2012).
Existing research suggests some degree of differentiation between domains of parenting: parents who are emotionally responsive are not necessarily cognitively stimulating; in addition, there is some degree of differentiation demonstrating that emotionally-responsive parenting is especially important for social/emotional development, while cognitively-stimulating parenting promotes young children's learning and development. "Negative parenting," or reliance on harsh, punitive physical and verbal parenting, relates to disruptions in social/emotional development and also delayed cognitive development (Chang, Schwartz, Dodge & McBride-Chang, 2003;Hughes et al., 2006). Maternal education and, specifically, executive function (Cuevas et al., 2014) show relations with the development of children's executive function. An overall tendency for parents to engage in more emotionally-supportive parenting than cognitively stimulating parenting has been documented across countries (Bornstein & Putnick, 2012).

Defining differential impacts of parenting on child development. Recent work has
documented that environmental conditions may influence children's development differently.
Describing the interaction between the child and their environment has emerged as a seminal element of understanding child development, especially in relation to the interaction between biological and environmental factors (Ellis, Boyce, Belsky, Bakersmans-Kranenburg, & van Ijzendoorn, 2011). Recent research suggests that children's responsiveness to environmental conditions, and thus reactivity to both positive and negative stimuli, is not uniform across children (e.g., Belsky, Bakersmans-Kranenburg, & van Ijzendoorn, 2007). While most of this work has been conducted in high-income countries, the concept of studying child-environment interactions may also apply to children who have experienced extreme environmental deprivation resulting in negative health outcomes such as stunting. Economic analyses have also suggested that children with very low cognitive and emotional resources may respond less to parenting, based on the theory that low vectors of skill development persist over time, even when environmental resources are enhanced (Heckman, Stixrud & Urzua, 2006). Little work to date has examined this interaction in the context of environmental conditions that characterize child development in low-and middle-income countries.

Nutrition and child development
A large body of work demonstrates strong associations between early nutrition status and child development (Prado & Dewey, 2014). Children with poor early nutrition fail to reach expected linear growth milestones, leading to stunting, long-term cognitive and motor deficits, and eventually worse economic outcomes (e.g., Galasso, Wagstaff, Naudeau & Shekar, 2017;Sudfield et al., 2015). Some work similarly demonstrates that cognitive stimulation can ameliorate some of the effects of poor early nutrition on child development, especially when coupled with parenting interventions (Paxson & Schady, 2007;Grantham-McGregor, Powell, Walker & Hines, 1993). Most evidence to date has been produced through intervention studies, with less work conducted using large-scale samples, where effects may be more representative of typical interactions between parenting and nutrition status.
The present study examines the emergence of wealth gradients in child development and the joint roles of nutrition status and parenting in explaining the gaps between high-and lowwealth children. We had three hypotheses: first, we expected to find gradients in child development based on family wealth in a low-income country. Second, based on existing parenting literature, we expected to find that there are different types of parenting, such as cognitively-stimulating and emotionally-responsive, with unique implications for child development, and that parenting quality is associated with both family wealth and child development. Third, we hypothesized that parenting and nutrition status would jointly influence child development, and that parenting could have different impacts on child development depending on nutrition status. In sum, we hypothesized that family assets would exert a strong and consistent impact on child development, but that parenting quality and nutrition status could account for a portion of the variance attributable to family wealth, demonstrated by a reduction in the impact of family assets on child development when parenting and nutrition status are 7 included in the model. Our analysis is informed by two rounds of data, collected from a large sample of children and their caregivers in rural Cambodia.

2) Methods
We investigated our hypotheses using data from a large panel of children in rural Cambodia. We rely on a unique panel data set of almost 7,000 children (

Participants
During the baseline data collection exercise in 2016, a survey firm 1 used an adapted version of the EPI walk to select 6,934 eligible households in 305 villages. 2 Each sampled household included children aged between 24 and 59 months. 3 Approximately 11 months later, 1 Angkor Research©, under the close supervision of the Principal Investigators. 2 EPI refers to the Expanded Programme on Immunization of the World Health Organization (see e.g. Henderson & Sundaresan (1982)). 3 The study was approved by the institutional review board of the Paris School of Economics and the National Ethics Committee for Health Research by the Cambodian Ministry of Health. The data were collected as part of a randomized controlled trial investigating the effects of preschools on school readiness of children in rural Cambodia. Our main outcomes for this study, collected in 2017, stem from a first follow-up on the trial intervention. Between baseline and the first follow-up, community preschools were constructed in some of the sample villages. An awareness campaign was conducted to stimulate enrollment into the new preschools. Further, a home-based program, in which trained core-mothers address a range of topics around parenting, was implemented. These 8 the same survey firm revisited 93% of the sampled households. 4 The follow-up sample, which forms the main sample for the analysis in this study, includes 6,452 households with 6,914 tested children at the age 3.91), of 2 ( 184), 3 ( 2271), 4 ( 2459), 5 ( 1965) and 6 ( 35). The average number of followed up households per village is 21.2 and the average number of eligible children per household is 1.07. Using the follow-up sample instead of baseline allows to control for baseline test scores in regressions and to observe outcomes during an age period which is critical for the development of language and early numeracy skills.
Upon arrival at households' home, the enumerator first identified the main caregiver and the eligible children, 5 and administered the household and caregiver surveys as well as completed the anthropometric measures and development tests. The child's testing session lasted an average of 31 minutes ( 50 30).

Measures
We developed and adapted the battery of child tests in close cooperation with the researchers of the Measuring Early Learning Quality and Outcomes project (MELQO; UNESCO, 2017) and the local survey firm. The survey firm recruited interviewers based on their familiarity with data collection and their experience with young children. To ensure that children and parents correctly understood the questionnaire and that the instruments were reliable, the research team pretested every instrument at least three times before collecting data on the interventions were randomized between villages and still at an early stage during the time of the first follow-up. We therefore ignore the interventions in this analysis. All our regression estimates are robust to the inclusion of treatment group variables to account for the randomized treatment.
9 sampled villages. The survey firm translated the questionnaires into Khmer and an independent third party back-translated them into English. The research team participated in the interviewer training conducted by the local firm's fieldwork manager. Field staff was organized into six groups, each comprising of four interviewers, one supervisor and one field editor. All supervisors had several years of data collection experience in Cambodia and were responsible for household sampling and quality control procedures. Editors supported supervisors doing spot-checks, interviewer observations, and conducted independent re-interviews of at least 20% of interviews in each village. 6 Household survey. The household survey includes information about family structure and socio-economic background. We constructed a dwelling quality index and household assets index. 7 Both indexes were then standardized and summed up to create a wealth index. We use quintiles and quartiles of this index to analyze the cognitive gap in rural Cambodia. To construct a binary measurement of household wealth, we define low living standard as belonging to the lowest 2 quintiles of the wealth index. We use household size 4.48, 1.68 and the binary measure of wealth 0.40 as measures of socio-economic background at the household level.
Caregiver survey. The educational background of the caregiver is measured by years of schooling 3.48, 3.15 . For better interpretability, we use a binary indicator for completion of primary school (6 years) .266 in our regressions. Caregiver non-verbal reasoning is measured using a set of 12 Raven's Progressive Matrices (Carpenter et al. 1990;Raven 2000). This test was also administered at baseline. When administered to the same caregiver again 11 months later, non-verbal reasoning scores showed a correlation of 0.46.
To account for the multidimensionality of parent-child interaction (Bornstein, 2012), we developed a questionnaire that covers multiple aspects of parenting and that fits well the Cambodian context. Caregivers answered 25 items related to regular activities, parent-child bonding, punitive behavior and different child-rearing methods. Possible responses ranged from 1 (never or almost never) to 5 (every day or almost every day). The factor structure of these parenting questions is reported in the Results section.

3) Results
The analyses plan was designed to test the three main research questions: 1) are there gradients in child development based on family wealth for children three to five years of age; 2) does parenting show unique components and if so, how do they differentially affect child development, and how is parenting associated with family wealth and child development; 3) how do parenting and nutrition status jointly contribute to child development, and do they account for a portion of the variance in child development due to wealth gradients? We are also interested in looking across these three questions to estimate the proportion of child development accounted for by variations in family wealth, parenting quality, and nutrition status.

Are there gradients in child development based on family wealth for children aged three
to five years? We start our analysis by looking at the wealth gradient for our four child development measures in Table 1  which means that measurement error may contribute to dampening the size of the gap.
The last three rows (age 3, age 4 and age 5) of Table 1 and column 'No controls' clearly indicate that the wealth gaps significantly widen with time. Between 3 and 5 years old, the gap increases, especially for early numeracy and language. From a small 0.11 SD at age 3 in early numeracy (0.31 in language), the gap reaches a large magnitude of 0.87 SD in early numeracy (1.12 in language) by age 5. This result is consistent with the fact that the 3 to 5 years of age period is critical for both competences. The gap increases to a lesser extent in executive function from 0.37 to 0.54 SD. The gap for socio-emotional problems increases from -0.18 to -0.26 SD.
To better visualize the differences in child development by socio-economic status over age, we used separate nonparametric regressions for the lowest and highest quintiles of the SES index ( Figure 1). Gaps in executive function, language and early numeracy skill are large and already evident at the age of 3, the earliest enrollment age for public preschools in Cambodia.
The gradient in executive function skills stabilizes at the age of 4, with a size of approximately 0.5 standard deviations. This is in stark contrast to gradients in language and early numeracy skills. The initial gap in language at the age of three is comparable to executive function.
However, the gap continues to grow until primary school-age to a magnitude of 0.9 SD (early numeracy) and 1.1 (language) standard deviations. This gap is of substantial magnitude and equivalent to a delay of 14-18 months (a child progressing of .75 SD per year in our sample).

Does parenting show unique components, and associations with family wealth and child development?
We conducted explanatory factor analysis in order to determine if the parenting scores could be grouped by distinct dimensions of parenting. Sampling adequacy was supported by a Kaiser-Meyer-Olkin statistic of 0.868 (Kaiser, 1974), confirming a compact pattern of correlations between the 25 parenting items. Bartlett's test for sphericity rejected the null hypothesis of zero covariance, p<.001. Parallel analysis suggested that there are 11 factors, equaling the number of factors with eigenvalue greater than zero. We followed a similar procedure as Jones et al. (2017) to reduce the number of factors retained. We rejected the large number of factors as only 3 of them had more than 1 item with a factor loading greater than 0.3 and all but 3 factors had poor reliability .6 . We gradually reduced the number of factors until all factors had reliability of at least .6 and at least two items with loadings greater than 0.3. With this procedure we obtained three parenting factors, which are summarized in Table 2.
The three identified parenting factors capture three dimensions of parenting. First, "cognitive parenting" measures how parents actively interact with the child in ways that are likely to develop cognitive competencies: games, reading books and playing with toys or objects.
The second dimension, "socio-emotional parenting" captures the emotional support and responsiveness of the child and the parent. Third, "negative parenting", in which higher values indicate more negative parenting, refers to the degree to which parents use harsh or punitive statements to discipline the child. It is important to note that the constructs of emotional support and responsiveness and the use of harsh or punitive statements are not two extremes of the same dimension; instead, in our sample, they appear to be unique constructs.
To assess the extent to which the three parenting scores capture meaningful differences in parenting behavior, we correlate them with other family characteristics. Table 3 indicates that parenting quality is strongly correlated with socio-economic characteristics of the household and of the caregiver: wealthier and more educated families tend to adopt parenting behavior that is Cognitive parenting is significantly positively correlated with scores on the child tests, negative parenting is significantly negatively correlated with these scores, and the direction of the effect of socio-emotional parenting is ambiguous in this correlation analysis. While socioemotional parenting is negatively correlated with socio-emotional problems -0.04, .001 , it is also negatively correlated with the executive function score -0.03, .040 . Since parenting is significantly associated with several other household characteristics, isolating the independent role of parenting is important. We show below that after controlling for household and caregiver characteristics and even baseline test scores, the effect of parenting remains significant and that socio-emotional parenting does not negatively affect executive function skills. To investigate further the variation accounted for by parenting and other factors, as well as to test the hypothesis that the wealth gap is partially mediated by parenting and nutrition, in Table 4, we use four specifications in a hierarchical regression model. In a first step (column (1)), we regress the dependent variable on child age and child gender to determine how much of the total variation is explained by exogenous characteristics. The R-squareds in these regressions vary from 2.1% for socio-emotional problems, to 38% for early numeracy and approximately 43% for executive function and language skills. Note that the three child-administered test scores increased by approximately 0.72 to 0.78 SD by year of age. This means that the cognitive gap that separates a child from the bottom quartile and the top quartile (between 0.42 and 0.60 SD, see Table 1) corresponds to between 60 and 80 percent of one year of average development.

What are the joint roles of parenting and nutrition in explaining variation in
Girls perform significantly better than boys across all domains (9-12% of a SD). The gender difference is particularly pronounced for executive function. Of our three parenting measures, cognitive parenting had the strongest effect on the three cognitive scores. A one SD increase in our measure of cognitive parenting is associated with a 0.092-0.098 SD increase in language and early numeracy scores, an increase that is comparable to half to two-thirds of the negative effect of stunting, low living standard, or low caregiver 20 education. Based on the magnitudes of these associations, to compensate for the cumulative detrimental effect of low caregiver education, low living standard and stunting, the cognitive parenting behavior score would need to move from the bottom 19 percentile to the top 95 percentile (3 SD), i.e. from a parent with poor cognitive parenting behavior to one of the best parents in our sample. Given this large magnitude, it is perhaps unreasonable to expect that cognitive stimulation alone might be able to make up for the effects of negative shocks or a negative environment.
Socio-emotional parenting affects executive function skills positively but does not affect the other cognitive competences. Higher scores on socio-emotional parenting are also significantly associated with less socio-emotional problems. While smaller in size than for cognitive parenting, the effect of negative parenting is significantly negative for all cognitive competences. The size of the relationship is especially large for the socio-emotional problem score (0.346 SD).
To test if nutrition also explains additional variance we performed a similar exercise by adding it to the model in column 4 instead of the parenting variables. Stunting cannot explain variance in socio-emotional problems Δ .000 , similar to Table 1. It explains a larger share of variance in executive function than our measures of parenting do Δ .008 , a similar share for language Δ .008 and a slightly smaller share for early numeracy Δ .006 .
The coefficients of household and caregiver variables change only marginally when stunting is added to the model. Thus, while the relationships with outcomes are significant-equivalent to about 3 months of cognitive development, for example-stunting seems to be a weaker mediator of caregiver education than parenting behaviors are. 10 We explore additional associations and mediating factors in Table 5. First, we include baseline test scores collected 11 months prior to the main outcomes of this study (which would also have the effect of controlling for unobserved factors already captured in those scores).
While the magnitudes fall, the relationships with parenting typically remain significant, suggesting that the effect of parenting is cumulative (and not "one-shot"). 11 We also hypothesized that parenting may have differential effects on child development, depending on children's nutrition status. To test the hypothesis that the association of cognitive parenting and child outcomes is weaker for stunted children, we looked in column 3 of Table 5 at the effect of parenting, separately for stunted and non-stunted children and with baseline controls. We focused on cognitive parenting because of the notable role of nutrition status in cognitive development; we thus hypothesized that cognitive parenting and subsequent child development would be especially sensitive to nutrition status. While no significant interactions between parenting and child nutrition status were observed for executive function and socioemotional problems, effects of cognitive parenting differed by stunting for language and early numeracy outcomes. The magnitude of the cognitive parenting effect is about 35% weaker for effects on language and 54% weaker for effects on early numeracy, showing that the positive effect of cognitive parenting is weaker for stunted children. We repeated the same exercise 10 The fact that wealth gradients remain large even after controlling for anthropometrics and parenting should be interpreted with care. As our measure of parenting is only self-reported and based on recall, it is possible that it does not capture the full dynamic of parenting throughout the children's life time. Our measure of parenting could for instance be imperfectly correlated with past parenting (during pregnancy or breastfeeding). Our measure of parenting may also imperfectly measure the quality of parenting. Longitudinal and external measures of parenting would be necessary to enhance our measures. 11 Note that in Table 5 we control for both parenting and stunting. The effects of each are relatively unchanged.

22
without baseline controls in column 2 of Table 6 and observed similar interaction effect sizes.
However, the interaction term for the language outcome is only significant in Table 5 where baseline test scores were added to the model. 12 While the cognitive parenting x stunting interaction is statistically significant for our language and early numeracy outcomes, this finding can also be explained by an unobserved heterogeneous effect by parenting quality, since parenting quality is correlated with stunting. As a proxy for parenting quality, we therefore chose caregiver education, which is also negatively correlated with stunting. We then added a cognitive parenting x caregiver education to the models in the last columns of Table 5 and Table 6 to see if the coefficient of the cognitive parenting x stunting interaction changes. The inclusion of caregiver education, and its interaction with nutrition status, did not change the significance of the previous interaction. We thus conclude that cognitive parenting has a stronger impact on some domains of child development for non-stunted vs. stunted children.

4) Discussion
The objective of this study was to quantify the role of parenting and nutrition in explaining the cognitive gap that we observed between children from low and high wealth backgrounds. This gap exists in Cambodia and is large: the difference between the highest and 12 We also tested if our results are robust to different cutoff points for the stunting indicator. As Perumal et al. (2018) point out, there is no biological basis for the commonly used cutoff at -2 SD in the height-for-age score (HAZ) and children slightly below and above the cutoff might be equally affected by undesirable health outcomes due to growth impairments. Using a binary indicator equal to one if HAZ<-1.5 M=0.592 does yield smaller and insignificant estimates for the cognitive parenting-stunted interaction term of column 3 in Table 5 for language -0.0233, .183 and early numeracy development -0.0165, .396 . Yet, using a lower cutoff score of HAZ<-2.5 M=0.201 yields slightly larger and significant interaction effects on language -0.0311, .057 and early numeracy development -0.0433, .014 than found in Table 5. The same pattern is observed without baseline control variables. This demonstrates that the weaker association of cognitive parenting and test scores is predominantly driven by children in the lower tail of the HAZ distribution.
lowest quintiles amounts to about 7-9 months of child development. For pre-academic competencies (language and early numeracy) the gap widens significantly between age 3 to 5: confirming that this period is crucial for development of early numeracy and language skills and is a time when gaps emerge and may become entrenched.
We document four main findings. First, building on previous research on socio-economic gradients in developing countries, we confirm that a substantial cognitive gap exists between children from different backgrounds in rural Cambodia (confirming earlier findings in Naudeau et al., 2011). In our sample, 0.60 SD separate the children in the bottom quartile of a wealth distribution from the children in the top quartile (language test). The gap appears larger for cognitive competences (executive function, language, early numeracy) than for noncognitive/socio-emotional outcomes (Strengths and Difficulties Questionnaire). While significant and large, these gaps in rural Cambodia are at the low end of those documented in rural Latin America by Schady et al. (2015) (between 0.57 to 1.12 SD on a similar language test).
Second, the gap in child development due to family wealth widens with age (e.g. Paxson & Schady, 2007;Fernald et al., 2012). Between three and five years of age, the gap especially widens in language (from 0.31 SD at age 3 to 1.12 SD at age 5) and early numeracy, but less for executive function (from 0.37 SD to 0.54 SD). It also increases strongly for socio-emotional problems. The fact that the executive function gap is already present at age 3 and only increases by 46% between age 3 and 5 suggests that before age 3, children are already experiencing substantial growth in executive function that is affected by family socio-economic status.
Third, we confirm that cognitive-stimulating and emotionally-responsive parenting is one of the important explanatory factors of cognitive differences due to family wealth. Our measures of parenting behaviors are strongly correlated with children's performance, accounting for about 24 8-14% of the gap in cognitive competences observed among 3 and 5 year olds. While the explanatory power of parenting is robust, the magnitude of the parental effect is relatively small in comparison with other explanatory factors: We show that a large 1 SD improvement in the main parental competences--an improvement beyond the scope of most parenting programs-would only compensate about a half of the effect of stunting.
Fourth, we provide evidence that the positive effects of parenting may be stronger for children who are not stunted. This is an important finding for two reasons. First, it demonstrates the importance of designing and testing integrated program models of child development that anticipate the complex interactions that will arise between child characteristics and environmental stimuli. It is unlikely that "one size fits all" (Shonkoff, Radner & Foote, 2017).
Second, this interaction -albeit a small one, and not consistent across all domains of parenting or child development -adds to the growing body of research documenting the variable effects of environments on child development, by adding undernutrition as a condition that may moderate the effects of later environmental stimuli on child development.
Encouraging stimulating and responsive parenting and improving children's educational environments are key to reducing cognitive inequality and improving children's performance during that critical time when the cognitive gap widens. Yet, to fully compensate for initial inequality, improvements in the measured dimensions of parenting are not a panacea: other factors (nutrition, poverty level) are important. Poverty and undernutrition must be targeted, along with parenting, and program designers should expect that the effects of various interventions may vary based on children's nutrition status, especially if undernutrition is already evident. As with all elements of early development, prevention and early intervention are more effective than amelioration -and in this case, that may pertain to addressing undernutrition 25 before conception and during pregnancy. At a minimum, the findings from this study confirm that a combination of interventions may be needed to effectively support young children's development in low-income countries. The wealth gradient in child development is well-documented in many countries. We show that parenting behaviors explain a share of this-but other factors matter as well. More work is needed to examine the broader range of environmental conditions that contribute to this gap, and perhaps more critically, how they interact with one another. Such analyses would help guide the design of effective interventions which would exploit complementarities between parenting, nutrition and other factors for the promotion of child development in challenging contexts.    Notes. Factor loadings >.3 in bold are the items of the three parenting composite scores. Possible responses ranged from 1 (never or almost never) to 5 (every day or almost every day).
33 Table 3: Pairwise correlations between follow-up child test scores, child, household and caregiver control variables.
(2) in column (4). Two-tailed p-values for t-tests and F-tests using robust standard errors clustered on the village level. *p < .1 **p < .05 ***p < .01 Table 5: Results of hierarchical regression analysis predicting different domains of child development using baseline controls, a cognitive parenting x stunted and a cognitive parenting x caregiver education interaction term.
Notes. Changes in the R² (R² and F-tests for joint significance of additional variables F refer to a comparison of (2) vs. (1) in column (2), (3) vs.
(3) in column (4). Two-tailed p-values for t-tests and F-tests using robust standard errors clustered on the village level. *p < .1 **p < .05 ***p < .01 Table 6: Results of hierarchical regression analysis predicting different domains of child development using a cognitive parenting x stunted and cognitive parenting x caregiver education interaction term.
Notes. Changes in the R² (R² and F-tests for joint significance of additional variables F refer to a comparison of (2) vs. (1) in column (2) and (3) vs.