The heterogeneous impact of monetary and
emotional investment on skills in the early
childhood
Master’s thesisSupervised by Prof. dr. Chris Elbers
Esther Heesemann
M.Sc. Development Economics
University of Amsterdam
August 27, 2014
Abstract
Several ECD programs have been implemented over the years in order to ensure that also children from disadvantaged households develop their full potential of cognitive and character skills. These programs focus either on monetary or emotional investment in children. In order to analyse the impact of both kinds of investment separately at different ages, a dynamic latent factor model will be applied on a four period individual panel dataset of the Caribbean island St. Lucia. This analysis provides evidence for a positive and heterogenous impact of monetary and emotional investment and reveal the high sensitivity of skills to inputs according to age. It further confirms the persistence of skills over time and their positive interaction.
Contents
1 Introduction 3 2 Literature review 4 3 Methodology 6 4 Data 8 4.1 Background on St. Lucia . . . 84.2 The Roving Caregivers Program . . . 9
5 Identification of the latent factors 12 5.1 Measurements of monetary and emotional investment . . . 13
5.2 Measurements of skills . . . 15
6 Results 15 6.1 Cognitive skill formation . . . 17
6.2 Character skill formation . . . 18
6.3 Determinants of parental investment . . . 20
7 Robustness Check 20 8 Discussion and conclusion 25 References 27
List of Tables
1 Descriptive statistics St. Lucia and Latin America & Caribbean (LAC) . 10 2 Descriptive statistics in 2008 . . . 123 Parental investment measurements . . . 14
4 Cognitive skill formation . . . 19
5 Character skill formation . . . 21
6 Monetary investment . . . 22
7 Emotional investment . . . 22
8 Cognitive skill formation, robustness check . . . 24
9 Character skills formation, robustness check . . . 24
1 Introduction
In the recent years more and more attention was drawn to investment in the very early years of a child’s life. The economic reasoning behind it lies in the rapidly decreasing returns to investment in human capital over lifetime (Heckman 2000). Investing in the human capital formation of very young children via early childhood interventions is thus much more cost-effective than of students at higher age by for instance lowering class size or monitoring of teachers’ performance. The importance of the development of cog-nitive and socio-emotional (character) skills in the early childhood for future academic and labor market outcomes is widely acknowledged by economists and psychologists. Early test scores (together with background variables) can predict a substantial part of the variation in high school completion, wages and employment (Currie and Almond 2011). Besides, a lack of skill formation in this period of life will affect not only the individuals life by less schooling, hence earnings, but also society as a whole through a loss of work force, teenage pregnancies, smoking and delinquent behavior (Heckman, Stixruf and Urzua 2006). Hanushek and Woessmann (2012) found that a one standard deviation increase of the cognitive skill level of a country’s workforce will lift the annual growth by two percentage points.
Circumstances that hinder children from developing their full potential are often re-lated to a disadvantaged environment. A lack of cognitive stimulations, malnutrition during pregnancy and childhood and exposure to stress are the most striking risk fac-tors (Walker et al. 2007). Moreover children from poor households are likely to suffer from poor hygiene and sanitary conditions which negatively affect the children’s health status. This is again associated with deficits in cognitive skill formation and lower ed-ucational outcomes (Walker et al. 2007). The correlation of poor skill formation and poverty gives rise to large differences in capabilities even before entering primary school, depending on the socio-economic status. Limitation of early childhood development (ECD) is therefore a crucial channel of the transmission of intergenerational poverty.
Many different initiatives have taken place in developing as well as in developed coun-tries to tackle the inequality in ECD. But before implementing costly programs to sup-port disadvantaged families, it is crucial to understand the true technology of skill for-mation in the early years. How do character and cognitive skills persist over time and influence each other? What kind input will actually lead to positive outcomes? The purpose of this thesis is to shed light on this skill accumulation process. It mainly fo-cusses on the question, whether it is just learning material, books and toys (so called monetary investments) or also the parental support and child rearing methods (emo-tional investment) that matter for a proper development in the early childhood. It will be further investigating at which age skills are sensitive to parental inputs and the past skill formation.
Since skills and parental investment are not directly observable, a dynamic latent factor model, developed by Cunha and Heckman (2008), will be used. The model will
be applied to a four-period dataset of St. Lucia containing economically disadvantaged children from birth up to 7 years.1 This research will show how emotional and montary investment affect cognitive and character skills in a developing country. Furthermore, the thesis contributed to the existing literature by estimating for the first time the Cunha Heckman model in a setting of more than two periods and in its application to observa-tions of very young age.
The thesis is structured in the following way: Section 2 reviews prior literature on skill formation and early childhood development from a empirical and theoretical perspective. Section 3 explains the methodology used, followed by a description of the dataset. In Section 5, the application of the model will be explained. Section 6 shows the estimation results, which will be checked on robustness in Section 7. The conclusion of this thesis and the discussion of the outcomes will be presented in Section 8.
2 Literature review
In the last decade, economic scholars started to define human capital not only by cog-nitive skills such as for example reasoning skills but also by character skills, which refer to socio-emotional capabilities. They thereby acknowledge the importance of both di-mensions of skills for current and future success. Character skills such as self-control, self-esteem and motivation are crucial determinants of school and labor market out-comes (Shonkoff and Phillips 2000, Dunifon and Duncan 1998, Heckman, Stixrud and Urzua 2006). Duckworth and Seligman (2005) show that pupils with a high level of self-discipline outperform their impulsive fellows in academic achievement test as well as in high school admission and attendance rates. Rosen et al. (2010) reviews numerous stud-ies regarding the influence of motivation and summarizes that intrinsically motivated students are more likely to succeed academically. Analyzing long-term panel data from the UK, Carneiro et al. (2007) show that an increase of character skills by one standard deviation increases the likelihood of passing the O-levels by 2.7%.2 Heckman (2000) em-phasizes that early childhood interventions are particularly successful in fostering social instead of cognitive skills (IQ). Moreover, several long-term evaluation of ECD programs in the U.S. shows that children that were targeted by center-based interventions had a lower rate of grade retention, less need for special education and decreased delinquent behavior later on in life (Heckman 2000, Cunha et al. 2006).
ECD programs may have different objectives: improving child health in terms of height and weight, fostering cognitive development such as language and motors, or improving child behaviour in terms of character skills. The methods used to achieve these goals vary from pure cash transfers and food supplements to complete day care and parenting interventions. The link between parenting and cognitive development during childhood
1Although the data was collected in order to evaluate an ECD program, this thesis does not aim on
assessing the impact of this program.
had been established in psychological literature (Ravis, Kessenich and Morrison 2004, Tavis-LaMonda, Bornstein and Baumwell 2001, Pungello et al. 2009). They stress the positive impact of responsive behavior and child-parent interactions on language devel-opment. Intrusive behaviour, controlling and non-educative punishment is seen as a negative influence on cognition. Exposure to violence within families is found to in-crease the risk for behavioral problems including aggression and self-control (Walker et al. 2011). Parenting programs aim on encouraging child-parent interaction, such as story reading, and child rearing, and are often implemented via home-visits or commu-nity groups. Nores and Barnett (2009) distinguish between educational, nutritional and cash-transfer programs and point out the beneficial effect of interventions that include a care or stimulation components on cognitive skills. Engle et al. (2011) stresses the increased effectiveness when programs target both children and parents. Conditional cash transfer programs proved to be as well effective in fostering language development but only to a small scale. In a long-term study of an ECD program in Jamaica, Gertler et al. (2013) compare stunted and not stunted children and the effect of a combined intervention of food supplements and cognitive stimulation. They find catching-up of the stunted treatment group in term of employment status and wages compared to their control counterpart and the non-stunted group. They find no significant effect of the food supplement treatment itself, emphasizing the importance of cognitive stimulation. Investigating skill formation comes with several challenges. First of all evaluating the impact of non-randomized programs gives biased results due to sample selection. Second there is no single measurement for cognitive or character skill that fully covers all their dimensions. Third skill development in the early childhood is extremely age dependent. An intervention that showed good outcomes for two-year old does not necessarily work for five-year olds.
Another interesting aspect of the skill formation is the interaction of the skills them-selves. Children with higher character skills probably also have more ease in concen-trating and therefore learning new tasks. At the same time, high cognitive skills are likely to boost character skills such as motivation and self-esteem. The analysis of skill accumulation has therefore to take this interaction into account.
A large body of literature on human capital accumulation focusses mostly on the de-terminants of cognitive skills. For a long time economic literature mainly examined the effect only of school inputs such as reducing class size on academic achievements. But as stated above, inputs from families and households are just as important, especially in the context of early childhood development, where the exposure to school environment is still limited (Carneiro and Heckman 2003). Empirical approaches investigating the formation of cognitive skills follow contemporaneous and value-added specifications. The contemporaneous approach uses contemporaneous measures of family and school inputs to estimate cognitive skills. This implies, among other assumptions, that previous invest-ment in human capital has no effect on current achieveinvest-ment (Todd and Wolpin 2003). But since human capital development is an accumulative process, past investment does
matter greatly for today’s cognitive skill formation. This raises doubt of the feasibility of the contemporaneous model specification. The alternative value-added approach adds a lagged value of the dependent variable to the model, assuming that it contains all potentially unobservable information from the past.
Lack of sufficient information is an other common constraint when estimating skill development. Using a dynamic framework requires detailed longitude data with both household and child characteristics, school inputs and outcomes. The lack of the com-bination of the two types of dataset makes the application of the model challenging. Furthermore unobservable initial endowment of children plays an important role for the skill accumulation and may lead to endogeneity bias when correlated with other input factors. The same problem appears when input factors such as parenting or home envi-ronment are not measured directly or inaccurately.
Cunha and Heckman (2008) therefore develop a dynamic latent factor model on the basis of Todd and Wolpin (2003). They treat not only investment but also skills as latent factors that can be identified by a sufficient amount of measurements that are intercor-related only by the latent factor itself. The model falls in the category of structural equation models that are widely used in psychology. A detailed description of the model will be given in Section 4. The aim of the Cunha-Heckman model is to examine how skills interact and promote future skill formation, and how parental investment interferes in the process. Analysing two periods of dataset of six to twelve year old girls from the United States they come to the following results:
• cognitive and character skills are persistent over time (self-productivity),
• character skills affect the accumulation of the next period’s cognitive skills but not vice versa (cross-productivity),
• parental investment has an effect on both types of skills.
Helmers and Patnam (2011) analyse skill formation in India for eight to twelve year-olds using the same methodology. Their findings confirm the self-productivity of skills overtime and the positive impact of parental investment. Moreover they find evidence for cross-productivity not only from character to cognitive skills but also in the other direction.
3 Methodology
This research makes use of the dynamic latent factor model, developed by Cunha and Heckman in 2008, in order to identify the influence of monetary and emotional investment on the skill development. The following paragraphs will briefly introduce the empirical model.
The original model contains three latent factors: cognitive skills θtC, character skills θNt and parental investment θIt depending on period t.3 I will split the investment factor into two, differentiating between purely monetary θtM and purely emotional investment θEt in children. This results in a four latent factor model. Separating emotional from monetary investment allows for a heterogeneous impact of parental investment on cog-nitive and character skills.
Given that skills and investment are not directly observable, confirmatory factor anal-ysis will be used to estimate those latent factors in each period. There are mkt measure-ments for each of the four latent factors k in period t. The measurement system can be written as the following:
Yj,tk = µkj,t+ αkj,tθtk+ kj,t (1) where j = (1, . . . , mkt), k = (C, N, M, E), t = (1, . . . , T )
αkj,t is the factor loading of measurement j and skill/investment k. The equation system connects the latent factors to its observable proxies. The measurement error is captured in the term kj,t. In case of classical measurement errors, the error terms are independent across children and over time, and independent across the measurements of the same skill. The dynamic process of skill formation and investment is expressed as θkt = γ0,tk + γN,tk θt−1N + γC,tk θCt−1+ γM,tk θt−1M + γE,tk θEt−1+ γX,tk Xt+ ηtk (2)
where k = (C, N ), and
θkt = γ0,tk + γN,tk θNt−1+ γC,tk θCt−1+ γk,tk θt−1k + γX,tk Xt+ ηkt (3)
where k = (M, E).
In equation (2) the possible self- and cross-productivity of skills, and the impact of both dimensions of parental investment emerges. Self-productivity will be captured by γC,tC and γN,tN , which is the effect of the skill achievement in the period t − 1 on the skill level in period t. γNC,t and γN,tC show potential cross-productivity. Equation (3) captures the persistence of parental investment over time and the possibly skewed allocation of resources according to the past skill levels. In contrast to the dynamic skill equations, the dynamic investment equations are assumed to be only affected by their own lagged investment. In other words, no ”cross-productivity” of investment is allowed in this model. There are two reason for this assumption. First, such a correlation is arguably not likely: parents probably do not relate their monetary investment decision on their own emotional investment in the past period or vice versa. Second, several severe er-rors occur when running the extended model in the statistical program. Note that as a consequence of this restriction, the three latent factor model is no longer nested in the four factor model. The former is thus not a constraint version of the latter where the
correlation of monetary and emotional investment would be fixed to one. The vector Xt contains relevant covariates for each observation in the period of skill assessment.
The error terms ηt+1k are assumed to be independent between individuals and indepen-dent across time within individuals. A correlation between ηt+1C and ηNt+1 is possible and likely. ηt+1k is assumed to be normally distributed with mean zero and variance of Pt+1. The dataset provides sufficient information for estimating a four periods model,
hence four dynamic equations per skill factor, denoted by t = (2, 3, 4, 7) according to the approximate age of the children, will be estimated. The coefficients of interest γN,tk , γC,tk , γM,tk and γE,tk , are estimated by Maximum Likelihood.
The model can be identified as soon as there are two measurements per latent factor available:
Cov(Y1,t−1k , Y1,tk) = Cov(θt−1k , θkt) (4) Cov(Yj,t−1k , Y1,tk) = αj,t−1k Cov(θkt−1, θtk) (5) Cov(Y1,t−1k , Yj,tk) = αj,tk Cov(θt−1k , θkt) (6) The factor loadings αj,t−1k are identified by dividing the equation (5) and (6) by equa-tion (4). By dividing equaequa-tion (1) by the corresponding factor loadings, the joint distri-bution of the latent variables is calculated.
Cov(Y1,tk, Yj,tk)/αkj,t = V ar(θkt) (7) Cov(Y1,tk, Y1,τl ) = Cov(θtk, θτl) (8) Cov(Yi,tk, Yj,τl )/αki,tαljτ = Cov(θkt, θlτ) (9) For a complete identification, the number of free parameters has to be further reduced. There are two ways of doing so. Cunha and Heckman (2008) set the factor loadings of the first measurement of each latent factor in each period (αk1,t) equal to 1. The factor loadings of all remaining measurements can then be interpreted in relation to that. Another method is to standardize the distribution of the latent factors to mean zero and unit variance, wherefore µkj,t in equation (1) will be zero. The latter will be used for this analysis because it allows a better interpretability of the results.4
4 Data
4.1 Background on St. Lucia
A dataset from St. Lucia will be used for the analysis. St. Lucia is a small island of 617 km2 located in the eastern Caribbean. According to the World Bank, it belongs to the group of upper middle income countries with a GDP per capital of US$ 6,003 (in 2012, constant 2005 US$, PPP). The economy depends mainly on the export of agricultural
4
The St. Lucia data does not provide the same measurements of the latent factors in each period. The interpretation when setting αk
products as bananas and coconuts, and tourism. Due to its size, St. Lucia is very vul-nerable to external shocks. During the data collection, the country was suffering from the consequences of the financial crisis and the domestic and agricultural damage caused by Hurricane Tomas in 2010. Table 1 compares socio-economic indicators of St. Lucia to other developing countries in the region. It reveals a high unemployment rates especially for the youth in St. Lucia and an augmented rate of undernourishment. Enrolment in primary and secondary school is close to 100% even though declining over time. In the past thirty years, St. Lucia has shown remarkable progress in human development. Average years of schooling increased by 2.4 years and life expectancy at birth by 5.7 years. All together, St. Lucia belongs now to the high human development category, even though they slightly lack behind the region’s average (Malik 2013). Despite the positive picture that is given by these numbers, many children are still growing up in disadvantaged environment. A UNICEF report of 2006 found that more than half of children in St. Lucia are ”at risk”.5 The greatest risks originate from food insecurity and chronic diseases of a parent (UNICEF 2006). Furthermore child rearing methods in the Caribbean differ greatly from the approach used in western European countries with physical punishment as a common reaction to disobedience.
4.2 The Roving Caregivers Program
The data for this research was collected in the context of the Roving Caregivers Pro-gram (RCP) in St. Lucia. The proPro-gram has been implemented in 2006 with the purpose of supporting the development of children with disadvantaged background. The inter-vention consists of two home-visits per week from so called ”Rovers” to families with children under three years of age. The Rovers are RCP facilitators that live in the program communities. The home-visits aim at improving parenting practices by the transmission of knowledge about developmental milestones and age-appropriate child rearing methods. A second component of the RCP is a monthly community meeting for parents. For the program to target the families most in need, it is implemented only in communities with above average poverty rates and low access to center-based care. For the purpose of the impact evaluation, the program was extended in 2007 to two new regions on the island.6 The evaluation design followed a quasi-experimental setup with eight treatment and seven comparable, matched control communities. Within the
5
This refers to the risk that one or more of their human rights is being violated.
6Previous work on this data includes, besides a range of impact evaluations on cognitive and character
skills, a study on health status and on parenting behavior. The impact assessment one year after the implementation shows a positive impact on cognitive skills for children that entered the program at a young age. After three years, the positive impact has faded out completely. The authors did not find any significant impact of the treatment on the formation of character skills (Janssens, Rosemberg, van Spijk 2008, Janssens, Groot Bruinderink, van der Gaag 2010, Janssens et al. 2011). Since previous impact evaluations could not find a mayor impact of the program, the separation of the sample in treatment and control group will only be used as a covariate of the main analysis. A better understanding of the determinants of skill formation could help to explain the prior findings of the impact evaluations.
T able 1: Descriptiv e statistics St. Lucia and Latin America & Caribb ean (LA C) Indicator Coun try / Region 2004 2005 2006 2007 2008 2009 2010 2011 2012 GDP p er capita St. Lucia 5666 5493 5923 5923 6145 6067 6016 6029 6003 (constan t 2005 US$) LA C 4638 4782 4987 5202 5350 5206 5447 5617 5713 GDP gro wth St. Lucia 8.4 -1.9 9.3 1.5 5.3 0.1 0.4 1.3 0.5 (ann ual %) LA C 5.9 4.4 5.6 5.6 4.1 -1.6 5.8 4.3 2.9 Life exp ectancy at birth, St. Lucia 72.6 73.0 73.4 73.7 74.0 74.2 74.4 74.6 74.7 total (y ears) LA C 72.5 72.7 73.0 73.2 73.5 73.7 74.0 74.2 74.5 Undernourishmen t St. Lucia 12.1 12.2 11.3 11.7 11.7 12.8 13.7 13.8 12.2 (% of p opulation) LA C 11.8 11.5 11.3 10.8 10.4 9.9 9.6 9.4 9.3 Mortalit y rate, infan t St. Lucia 15.3 15.5 15.8 15.8 15.9 15.7 15.5 15.2 14.9 (p er 1,000 liv e births) LA C 22.4 21.4 20.4 19.6 18.9 18.2 17.9 16.8 16.3 P opulation ages 0-14 St. Lucia 29.2 28.5 27.7 27.0 26.4 25.8 25.3 24.8 24.3 (% of total) LA C 30.8 30.4 30.0 29.6 29.2 28.8 28.4 28.0 27.6 Sc ho ol enrolmen t, St. Lucia 102.1 103.1 107.4 100.3 96.9 94.6 92.3 90.2 87.4 primary (% gross) LA C 118.7 117.8 118.1 117.2 116.3 115.8 114.3 112.9 .. Sc ho ol enrolmen t, St. Lucia 73.7 77.8 82.3 88.0 93.7 95.4 95.2 94.6 91.1 secondary (% gross) LA C 87.4 87.9 88.8 87.8 88.9 89.6 90.2 90.2 .. Unemplo ymen t, St. Lucia 21.0 18.7 16.0 14.0 15.7 18.1 20.6 .. .. total (% of total lab or force) LA C 8.5 8.1 7.3 6.9 6.5 7.5 .. 6.7 .. Unemplo ymen t, y outh St. Lucia 40.8 38.7 31.7 27.5 .. .. .. .. (% of total lab or force ages 15-24) LA C 16.4 15.7 15.3 14.4 13.6 15.6 .. 14.5 .. Notes: LA C, as d efined here, only consists of the dev eloping coun tries of the region. Source: W orld Bank database
program communities, all children are eligible to join as long as they are not yet enrolled in a daycare center or preschool. As a further condition to participation, the caregiver has to be present during the home-visits. This is mostly the mother, but occasionally the father, grandmother or an aunt when the mother is absent or full-time employed. Children and their caregivers can enrol immediately after birth, drop out of the program upon enrolment in center-based care and graduate from the program at 36 months of age. Data collection included household surveys, cognitive and non-cognitive child assess-ments, anthropometric measures, and home environment and parenting assessments. The baseline survey was conducted between July and November 2006, with follow-up surveys in 2008, 2009, 2010 and 2013. In total, 91% of children in the control commu-nities but only 76% of children in treatment commucommu-nities completed both the survey and the assessments at baseline. Therefore, a second baseline was conducted in Spring 2007. The intervention in the expansion communities started in January 2007 onwards. Children born after July 1st 2006 are not included in the sample but they could enrol in the ongoing program. The sample in 2006 included 461 children.
Also after the intervention, the participating households in control and treatment com-munities were interviewed to assess short and medium term impacts of the home-visits. Together with the surveys, the assessment of the children were continued to be con-ducted to document their progress. The survey items and assessment tools changed over time in order to adapt to the children’s age. Most importantly the home assessment, which provides the majority of the investment measurements, only started in the second survey round in 2008. For this reason, only the data from this year onwards can be used for this research. Another challenging aspect of the St. Lucia dataset is the difference in age of children in the same survey year. For example during the survey round of 2008, children were between 21 and 46 month old.
Since this thesis aims on identifying the impact of emotional and monetary investment at different ages, the sample needed to be reduced to children of a similar age. The new sample therefore contains only observations between 21 and 39 months, or children of 2 years + 3 months. This diminishes the sample size to 284 observations in 2008, 276 in 2009, 273 in 2010 and 265 in 2013.7 Table 2 summarizes the main characteristics of the new sample.
The descriptive statistics reveal the successful targeting of the poor households by the program. Interestingly although 80% were below the poverty line, only 58% of the families announced to have had money problems in the past. Besides from the financial side, stunting and underweight, measured by length/height and weight for age, are often used as indicators for a disadvantaged environment in the early childhood since it
mea-7Due to the suspect of cheating, the character skill assessment data collected by two interviewers in
2010 and one interviewer in 2013 had to be omitted in the analysis. The test scores of 62 observations in 2010 and another 40 in 2013 were therefore replaced with missing values.
Table 2: Descriptive statistics in 2008
Variable Obs Mean Std. Dev. Min Max
Age (in month) 284 30.05 5.18 21 39
Male 284 0.46 0.50 0 1
Birth weight (in ounces) 272 112.22 20.63 50 184 Length/Height for age (z-score) 271 -0.04 2.25 -12.14 5.87 Weight for age (z-score) 265 -0.07 1.17 -2.54 2.90
Household size 284 5.61 2.47 1 14
Number of children 284 3.23 1.65 1 8
Household head employed 279 0.72 0.45 0 1
Household head female 282 0.43 0.50 0 1
Mother’s years of education 258 9.46 2.69 0 17 Below poverty line of 2006 277 0.80 0.40 0 1 No money problems past month 284 0.42 0.49 0 1 Child slapped last week 234 0.67 0.47 0 1 Child beaten last week 280 0.49 0.50 0 1
Treatment received 284 0.39 0.49 0 1
Treatment offered 284 0.46 0.50 0 1
Notes: Poverty line in 2006 was EC$ 423.83.
sures the short and long term impact of malnutrition. Both means are slightly negative suggesting that underweight and stunting is indeed prevalent in the sample. Physical punishment is a common disciplinary practise in the Caribbean which is confirmed by 49% and 67% of the children been beaten and slapped in the past week. Table 1 also shows that not all families that were offered to participate in the home-visits did accept the offer. From 46% of the sample living in the treatment area, only 39% participated eventually in the program.
5 Identification of the latent factors
One of the strong characteristics of the dynamic latent factor model is that there is no need for data on the same measurements in all periods. Various instruments, even of different scale, can be used to identify the latent skill and investment variables. This specific feature of the model makes it possible to analyse datasets with changing as-sessments and survey items over time. When selecting appropriate measurements, it is crucial to make sure that the indicators only respond to one of the four latent factors. Wealth indicators such as household income for instance might certainly be considered as a reasonable measurement of monetary investment. But at the same time, wealth might be a result of a high level of parental cognitive skills, which is likely to be cor-related to children’s cognitive skills. Household wealth is therefore not an appropriate
proxy for monetary investment. Using maternal education as measurement for cognitive skills would cause similar problems. Moreover, it can be assumed that wealth status and parental education are mostly constant over time. Using these time invariant variables would not capture the changes in the latent factor over time. In order to control for the existing heterogeneity in the sample, such such variables will be used as covariates.
5.1 Measurements of monetary and emotional investment
Parental investment is a very broad concept that comprises monetary investment in children such as providing toys, paying to tuition fees and health care but also the emo-tional support. Encouraging children in their actions, playing together, using educative instead of physical punishments are all considered as good parenting practices and are as much part of the parental investment as any dollar spent. In order to investigate the separate effect of monetary and emotional investment in children, two latent investment factors will be identified by a variety of proxies taken from the questionnaires and skill assessments.
The survey rounds of 2008 - 2013 include the Home Observation for Measurement of the Environment (HOME) tool designed by Caldwell and Bradley (1984). This tool rates the home environment in terms of learning material, parenting, family integration etc. This scale can be used to identify emotional as well as monetary investment. The HOME items change over the years in order to adapt to the particular needs of children at different years of life. For this research, the items were newly divided into categories in order to separate emotional from monetary investment. Further, a new category called ”Toy” was created. Beside from the HOME tool, monthly expenditure on daycare is used to identify monetary investment in 2010. By then the vast majority of children already attended preschool or kindergarten and spendings for daycare were positive for all but one household. In 2013 this item was no longer part of the survey, possibly due to the fact that children already entered primary school.
In addition to the components of the HOME tool, emotional investment is further identified by ”Child-parent interaction” and ”Violence”. Child-parent interaction re-flects daily family activities such as singing or storytelling. The measurement counts the positive interactions between caregiver and child and has a range from 0 to 20. ”Vio-lence” measures how often children were beaten, slapped or shouted at in the past week. It varies from 0 (”none of the three”, ”never”) to 15 (”all of the three”, ”always”) and is a negative measurement.
Table 3 reports the descriptive statistics of the monetary and emotional investment proxies. Monetary investment is estimated by the measurements Toy and Environment (and Daycare Expenditure in 2010). The remaining measurements are used for the identification of emotional investment. For a detailed documentation of the items used for each category, see Appendix B and Janssens and Rosemberg (2009).
Table 3: Parental investment measurements
2008 Obs. Mean St. Dev. Min Max
Responsitivity 284 6.48 2.08 0.00 9.00 Variety 284 3.00 0.96 0.00 4.00 Acceptance 284 5.19 1.04 1.00 6.00 Violence 279 6.31 2.72 0.00 15.00 Child-Parent Interaction 283 6.19 2.32 1.00 18.00 Toy 284 7.61 2.63 2.00 11.00 Environment 284 3.39 1.72 0.00 7.00 2009 Responsitivity 276 6.27 1.86 0.00 8.00 Variety 276 6.19 1.79 0.00 9.00 Acceptance 276 2.20 0.63 0.00 3.00 Academic Stimulation 276 4.47 1.44 0.00 6.00 Language Stimulation 276 4.28 0.99 0.00 5.00 Modeling 276 2.54 1.43 0.00 5.00 Violence 265 6.37 2.72 3.00 15.00 Toy 276 3.58 2.41 0.00 8.00 Environment 276 6.84 2.65 0.00 11.00 2010 Variety 273 4.96 1.42 0.00 7.00 Acceptance 273 0.25 0.43 0.00 1.00 Academic Stimulation 273 5.36 1.08 0.00 6.00 Language Stimulation 273 3.51 0.67 0.00 4.00 Violence 272 6.15 2.48 3.00 15.00 Toy 273 3.57 2.59 0.00 8.00 Environment 273 6.25 2.90 0.00 11.00
Daycare Expenditure (per month) 248 188.11 142.78 0.00 800.00 2013 Responsitivity 265 7.00 2.41 0.00 10.00 Emotional Climate 265 2.98 1.66 0.00 8.00 Encouragement 265 4.81 1.71 0.00 7.00 Enrichment 265 1.86 1.38 0.00 6.00 Companionship 265 3.86 1.46 0.00 6.00 Integration 265 2.74 1.43 0.00 4.00 Toy 265 2.66 1.69 0.00 8.00 Environment 265 4.55 1.96 0.00 8.00
5.2 Measurements of skills
Given the different stages of development, the methods to capture the skills of the chil-dren differ between the years. For the first three survey rounds, cognitive skills were quantified by the Mullen Scales of Early Learning Test (MSEL). This test consists of five parts: gross motor functions, fine motor functions, visual reception, receptive language, and expressive language. It can be applied to children between 0 − 68 months of age. The scores are age-standardized (based on a North-American reference population) with an average of 50 points and a standard deviation of 15. From 2010 on, the Peabody Vocabulary Picture Test (PPVT) was used since children had grown too old for the MSEL. The PPVT measures receptive language and verbal abilities of children. It is standardized with mean 100 and standard deviation 15 according to the age of the test taker. Unfortunately this test only gives one single score and is not decomposable as the MSEL. But for cognitive skills to be identified in 2010, at least two measurements are needed. For this purpose, a relative PPVT score is created which accounts for the deviation from the mean score of the reference population. The obvious correlation to the initial score is controlled for in the regression. In 2013, the Raven Test was added to the survey. The Raven Test captures reasoning skills independently of reading and writing abilities. It is hence a valuable complement to the PPVT outcomes. The Raven test is a commonly used tool to assess general intelligence from the age of five on.
In 2008 and 2009, character skills were measured by the Vineland Social-Emotional Early Childhood Scales (SEEC). This tool can be used for children up to 71 months and is divided into three subcategories (interpersonal relationships, play and leisure time scale and coping skills) with standardized scores with mean 100 and standard deviation of 15. In 2010, it was switched to the Child Behavior Checklist (CBCL). Both formats are caregiver reported measures with the purpose to assess the socio-economic behavior of children. The major difference between the two tests is that SEEC measures positive and desirable outcomes, such as the ability to interact with others, whereas CBCL captures negative patterns, such as anxiety or aggression. In 2013, the Self-Efficacy Questionnaire for Children (SEQ-C) was applied as a additional measure of character skills. The SEQ-C reports the ability to deal with social challenges in their peer group.
6 Results
The following section presents the results of skill formation analysis. For the estimation of the model, the ready-made R package lavaan was used.8 The exact equations are described in Appendix A. Since the latent factors are standardized to mean zero and unity standard deviation, the results can be interpreted in standard deviations (SD). Table 4 and 5 show the regression outputs of the dynamic skill estimation for four pe-riods. The subscripts t of the dependent variables θtk present the period in which the latent factor k is measured and equals approximately the age of the child at that time.
Note that there is a three years gap between the last periods. The interpretation of the estimates of θ4k is therefore ambiguous. It might either describe the direct effect of the latent factor in 2010 on the depending variable on 2013 or the indirect effect through the skill formation in the three years between the two periods. The tables present the results of estimating the skill formation process for a model with three and four latent variables. Specification (1) gives the estimates of the model described in above with monetary and emotional investment as separate latent factors, which will be referred to as the preferred specification. It tests the response of skills to both investment types separately and thus allows for heterogeneity in the impact. The results of estimating the original model with three latent factors are presented as specification (2) for comparison. The one-dimensional parental investment is captured by the variable θt−1I .
Both specifications control for a range of household and child characteristics, namely gender, child’s age, birth weight, household size, number of children in the household, mother’s education, employment status of the household head, whether the household had any financial problems in the past. It further explores a possible RCP treatment effect via a dummy variable indicating whether the household lives in a treated com-munity. Including maternal education, employment status and the financial situation may contribute to ensure a consistent estimate without omitted variable bias. Richer household can spend more easily money for their children which then might translate into a higher skill level. At the same time, children of a richer household might be already born with a higher initial endowment of skills if the financial situation of the household is correlated with the skill level of the parents. The gender covariate gives insights in the heterogeneity in skill formation of male and female. Birth weight acts as health indicator for mother’s health during pregnancy. The covariates also intend to include the initial endowment of children in the first period.
The estimation results show a positive impact of monetary investment on both skills and of emotional investment on character skills. They also support the evidence for the self-productivity of skills. Cross-productivity is found in both direction in the first trans-mission. Before discussing the results in more detail, note that besides ”Modelling” and ”Emotional Climate”, all investment measurements have highly significant factor load-ings with the expected sign. For the skill equations, all factor loadload-ings but those of the SEQ-C are significant. Comparing the model fits of specification (2) to (1) by the differences in chi square shows a significantly higher fit of the three latent factor model. This indicates that the three-latent-factor-model is indeed not nested in the four-latent-factor-model in which case the chi square would have been significantly lower or equal the chi square of the four-latent-factor-model.9
9
For (2) to be nested in (1), the three-latent-factor-model would represent a restricted version of the four-latent-factor-model that constraints to correlation between emotional and monetary investment to one. Relaxing this constrain should consequently lead to at least as good of a fit as the three latent factor model.
6.1 Cognitive skill formation
Table 4 reports the dynamics of the cognitive skill formation process. The first columns for each specification show the skill formation at the age of two. Given that the analy-sis starts at this point the skills in this period are only shaped by the covariates. The following columns demonstrate persistence and interactions of skills and investment. In-dependently of the specification, the self-enforcement of the cognitive skills is evident throughout all periods. The impact varies from 0.56 to 0.83 SD and is significant at the 1% level. An increase of cognitive skills by 1 SD at age two translates into an advance of 0.31 SD at age seven, if only the self-enforcing channel is considered.10 Between age two and three, cognitive skills rise significantly by 0.20 SD for each SD increment of char-acter skills in the previous period. This cross-productivity is however not observable in the later periods. The analysis of parental investment in specification (1) underlines the different effect of individual parental inputs. While the estimate of monetary investment on cognitive skills is only significant at the age of four (with a size of 0.17), the impact of emotional investment shows to be significant only at the age of seven. The magnitude of the estimate of θE4 is with 0.25 slightly higher than the estimate of θ3M. The magnitude is comparable to the cross-productivity but far lower than the self-productivity. The investment coefficients from specification (2) are significant and positive for all periods. This confirms the sensitivity of skills at the age of four and seven, but is not in line with the insignificant estimates at the age of three. The differences of these findings could be explained by the larger number of measurements available for θI
t which reduce the
measurement errors of the latent factor. The higher magnitude might depict the joint impact of emotional and monetary investment on cognitive skills.
The estimates of the covariates suggest significant heterogeneity in sex and in parental education level. The estimate of the gender dummy indicates that girls develop faster cognitive skills than boys when they are two and three years old. Later on, the effect is reversed. The education level of the mother has a overall positive and significant impact on skill formation. One more year of education raises the cognitive skill level of their children in the following period by between 0.05 and 0.13 SD. This relationship can be interpreted in several ways. If ability is a genetic property which is transmitted from mother to child, and schooling is a appropriate indicator for such genetic or epigenetic mechanisms, then children from highly educated mothers are likely to also have higher abilities and therefore higher cognitive skills. The other side of the nature versus nur-ture discussion emphasises that an educated mother appreciates education more than an uneducated mother, and is more able to support their children in the skill development process. Hence the higher mothers’ education, the higher the cognitive skill acquisition of children.
The estimation outcome for the treatment dummy shows no significant differences in skill formation between control and treatment communities. Note that this does not necessarily imply that the entire RCP had no impact on cognitive skills accumulation.
The treatment dummy only captures, if anything, the ”intention-to-treat” which is not equal to the actual treatment effect since some households in the treatment communities refused to participate in the program and the corresponding surveys.11 Thus a random allocation of the treatment could not have been guaranteed. This becomes problematic for the feasibility of the analysis if the participation in assessments and surveys differed in treatment and control group. If for instances, only those households in the treatment communities participated in the RCP that already had problems with child rearing, then the negative selection bias would underestimate the ”intention-to-treat” effect. A similar selection bias might occur if a different subset of household answered the surveys. Given the much lower participation rate in the treatment communities (86% at baseline) than in the control communities (95% at baseline), such a bias is likely. Furthermore, the channel through which the treatment is supposed to affect the skill formation is emotional and monetary investment. But in this model, investment is used as an explanatory factor itself. Hence the ”intention-to-treat” effect would only capture the direct impact that does not run through improved investment in children. For these reasons, the insignificant coefficient of the treatment dummy has to be interpreted with caution.
6.2 Character skill formation
The formation of character skills displayed in Table 5 shows clear evidence for self-productivity over time. The persistence of character skills decreases from 0.63 at age three, to 0.25 at age four and 0.27 at age seven. These estimates of self-productivity for character skills are much lower than those for cognitive skills: An increase of 1 SD at age two translates only into an advance of 0.04 SD at age seven, fixing all other channels. The estimate of θC2 is significantly positive, indicating cross-productivity between period 2 and 3. The impact of θM3 is positive and significant at the 5% level, as it is for cognitive skill. Both types of skills at this age benefit significantly from toys and a good envi-ronment in the past period. Emotional investment affects character skills in all periods. At age three and seven, the level of character skills increases by about 0.3 SD for SD increment in emotional investment in the previous period. Surprisingly the coefficient of θE
3 is negative, however only significant at the 10% level. Before drawing conclusion,
the robustness of the result will be analysed in the following section. Specification (2) confirms the positive impact found for period 4 and 7 but gives an insignificant estimate for θI
3.
Most of the included control variables do not influence character skill significantly. The ”intention-to-treat” effect is surprisingly negative at age two and seven at a signif-icance level of 5%, even though the aim of the RCP was clearly the opposite.12 This indicates that growing up in treatment communities actually reduces level of character skill by approximately 0.32 SD compared to control communities. This effect is
observ-11
For more details see Janssens and Rosemberg (2013) and AIID reports
12Changing the ”intention-to-treat” dummy for a dummy indicating if treatment was actually received
did not affect the findings of this research. Moreover, the coefficients showed same direction and significance level as the alternative approach used in specification (1).
T able 4: Cognitiv e skill formation F our laten t factors Three laten t factors (1) (2) θ C 2 θ C 3 θ C 4 θ C 7 θ C 2 θ C 3 θ C 4 θ C 7 θ C t− 1 0.670*** 0.555*** 0.831*** 0.657*** 0.536*** 0.767*** (0.126) (0.117) (0.202) (0.123) (0.116) (0.171) θ N t− 1 0.201** 0.001 0.115 0.191** -0.016 0.119 (0.094) (0.086) (0.134) (0.093) (0.076) (0.128) θ M t− 1 0.145 0.172** 0.072 (0.113) (0.076) (0.083) θ E t− 1 0.074 0.004 0.247* (0.135) (0.065) (0.139) θ I t− 1 0.207** 0.131** 0.321*** (0.093) (0.053) (0.102) Male -0.589*** -0.316* 0.298* 0.499** -0.587*** -0.328* 0.275* 0.489** (0.155) (0.177) (0.181) (0.238) (0. 154) (0.174) (0.164) (0.223) Age in mon th 0.001 0.008 -0.073*** 0.073** 0.001 0.008 -0.066*** 0.064** (0.016) (0.017) (0.023) (0.029) (0 .015) (0.016) (0.020) (0.026) Mother’s y ears 0.108*** 0.054* 0.100** 0.127** 0.107*** 0.058* 0.109*** 0.127*** of education (0.030) (0.031) (0.039) (0.051) (0.030) (0.031) (0.036) (0.044) Birth w eigh t 0.009** 0.006 0.005 0.004 0.009** 0.007 0.003 0.005 (0.004) (0.004) (0.004) (0.005) (0 .004) (0.004) (0.004) (0.005) Household size -0.050 0.149** -0.016 0.048 -0.047 0.147** -0.008 0.029 (0.039) (0.059) (0.052) (0.079) (0.039) (0.057) (0.047) (0.075) Num b er of children -0.098* -0.191** -0.008 -0.142 -0.102* -0.187** -0.016 -0.118 (0.059) (0.086) (0.078) (0.127) (0.059) (0.086) (0.071) (0.120) No money problems -0.059 0.304* -0.367** 0.455* -0.046 0.294* -0.328* 0.398* (0.165) (0.181) (0.186) (0.248) (0.165) (0.179) (0.168) (0.232) Household head 0.103 0.101 0.218 -0.516 0.098 0.106 0.210 -0.504** emplo y e d (0.155) (0.166) (0.169) (0.264) (0.155) (0.164) (0.154) (0.249) T reatmen t offered -0.141 -0.015 0.057 0.148 -0.143 -0.003 -0.016 0.116 (0.149) (0.168) (0.170) (0.230) (0.149) (0.167) (0.147) (0.213) Notes: Standard errors in paren theses. *: significan t at 10% lev el, **: significan t at 5% le v el, ** *: significan t at 1% lev el.
able in both model specifications at a similar magnitude and significance level. However as stated previously, the regression equation tests the impact of parental investment and intention-to-treat separately. So any indirect impact of the treatment via investment is shown by the estimates θt−1E or θIt−1, which will be discussed in the following section.
6.3 Determinants of parental investment
As discussed previously, investment does not only influence skill formation but also the investment decision might be influenced by cognitive and character skills of the children. This impact is captured in the coefficients γC,tE , γC,tE , γN,tM and γN,tE . Furthermore the per-sistence of investment over time will be analysed. The estimates are presented in Table 6 and 7.
The results show that monetary investment responses positively to cognitive skill level at the age of two and to character skill level at the age of three. Emotional investment on the other hand is even more affected by the past skill level showing significantly positive estimates of γEC,tin period 3 and 7 and of γN,tE for the periods of 4 and 7. The estimation outcomes also show strong persistence of both types of investment overtime. This is not surprising given the fact that monetary investment is measured in toys and housing environment which by nature have a rather persistent character.
As argued above, the treatment aims at enhancing skill development via increased parental investment. If the intervention was successful, the coefficient of the treatment dummy (only reported in Appendix C) should be positive and significant. But this is only the case for monetary investment at the age of four, and only at a significance level of 5%. Instead, it reveals a highly significant but negative impact on the level of emotional investment at the age of seven and of monetary investment at the age of three. This results might emerge either from a failure of randomization or of the treatment. The matching of treatment and control communities was only based on socio-economic, observable characteristics. But, if parents in treatment areas independently of the treatment invested less than parents in control areas, the treatment might have enhanced investment significantly but would not be shown in the regression output due to the initial lag. On the other hand, it might just be the case that the RCP did no succeed in improving parenting or at least not in means that are captured here. Estimating the model with a dummy for treatment received instead of treatment offered gives mainly the same results.
7 Robustness Check
This section addresses concerns regarding the underlying assumption and the model specification in order to support the feasibility of the findings. The results of the ro-bustness checks are reported in Table 8 and 9. Before going into detail of the roro-bustness analysis, it is worth noting, since not reported in the main table, that all additional
T able 5: Character skill formation F our laten t factors Three laten t factors (1) (2) θ N 2 θ N 3 θ N 4 θ N 7 θ N 2 θ N 3 θ N 4 θ N 7 θ C t− 1 0.351*** -0.105 0.136* 0.330*** -0.058 0.108 (0.111) (0.085) (0.075) (0.108) (0.080) (0.069) θ N t− 1 0.634*** 0.251*** 0.271*** 0.623*** 0.209** 0.226*** (0.121) (0.086) (0.086) (0.119) (0.082) (0.085) θ M t− 1 0.166 0.138** 0.001 (0.119) (0.069) (0.045) θ E t− 1 0.266* -0.132* 0.306*** (0.152) (0.073) (0.092) θ I t− 1 0.349*** 0.016 0.242*** (0.103) (0.051) (0.067) Male -0.168 -0.057 -0.164 -0.038 -0.167 -0.108 -0.144 -0.039 (0.141) (0.187) (0.174) (0.156) (0.140) (0.180) (0.168) (0.155) Age in mon th -0.036*** -0. 0 21 -0.005 0.006 -0.037*** -0.020 -0.004 0.007 (0.014) (0.017) (0.016) (0.015) (0.014) (0.017) (0.016) (0.015) Mother’s y ears 0.006 0.018 0.029 0.014 0.009 0.024 0.039 -0.006 of education (0.027) (0.032) (0.036) (0.033) (0.027) (0.031) (0.035) (0.030) Birth w eigh t 0.000 -0.003 0.000 -0.003 0.000 -0.001 -0.001 -0.003 (0.003) (0.005) (0.004) (0.004) (0.003) (0.004) (0.004) (0.004) Household size 0.023 0.125** -0.053 -0.059 0.023 0.111* -0.028 -0.065 (0.036) (0.060) (0.051) (0.055) (0.036) (0.058) (0.049) (0.054) Num b er of children -0.052 -0.093 0.126 0.182** -0.051 -0.085 0.118 0.185*** (0.056) (0.089) (0.079) (0.091) (0.056) (0.088) (0.077) (0.090) No money problems -0.103 0.286 0.193 0.107 -0.100 0.315* 0.187 0.099 (0.153) (0.189) (0.185) (0.164) (0.153) (0.186) (0.181) (0.163) Household head 0.097 0.183 -0.263 -0.058 0.095 0.188 -0.219 -0.081 emplo y e d (0.145) (0.172) (0.167) (0.169) (0.145) (0.169) (0.164) (0.167) T reatmen t offered -0.324** 0.006 -0.024 -0.318** -0.334** 0.038 -0.031 -0.327** (0.140) (0.175) (0.165) (0.155) (0.140) (0.171) (0.158) (0.152) Notes: Standard errors in paren theses. *: significan t at 10% lev el, **: significan t at 5% le v el, ** *: significan t at 1% lev el.
Table 6: Monetary investment θM 3 θM4 θM7 θCt−1 0.392** -0.102 0.114 (0.172) (0.203) (0.104) θN t−1 -0.106 0.430* -0.130 (0.141) (0.244) (0.129) θM t−1 1.488*** 1.130** 0.310** (0.345) (0.553) (0.151)
Table 7: Emotional investment
θE 3 θE4 θE7 θCt−1 0.303** 0.068 0.233** (0.152) (0.085) (0.093) θN t−1 0.085 0.246*** 0.209* (0.129) (0.089) (0.119) θE t−1 1.391*** 0.192** 0.267** (0.506) (0.075) (0.105)
Notes: Standard errors in parentheses. The same set of covariates is used as in specification (1). See Appendix C for the full outcomes table.
*: significant at 10% level, **: significant at 5% level, ***: significant at 1% level.
models give a statistically significant and negative coefficient of the treatment dummy on character skills at the age three. The effect has a range of -0.35 and -0.28.
For specification (3), only time invariant covariates (age, gender, birth weight and treatment dummy) were included in the regression. Dynamic adaptations of the control variables to either the treatment or the skill development might cause endogeneity. Ex-cluding all time varying covariates avoids this potential threat. The outcome confirms the self-productivity for both skills and the cross-productivity in period 3. Furthermore, it shows a positive impact of monetary investment in period 4 which is in line with the findings of the preferred specification. Also when applying specification (3), emotional investment does not observably influence cognitive skill formation, whereas it has sig-nificant positive effect on character skills in period 3 and 7. Most importantly, this first robustness check does not support the negative estimate of θE3 on the character skill formation at the age of four. At this point it should be mentioned that excluding covari-ates might not necessarily reduce endogeneity bias because the newly excluded variable itself might now cause omitted variable bias. The estimates of the reduced specification should therefore be interpreted with caution.
Specification (4) addresses the issue that appeared in previous work with the dataset: the sensitivity of the results to the sample construction. So far the total sample was reduced to children that were between 21 and 39 months old by day of the examination in 2008. The second robustness check now tests the preferred specification (1) on a even smaller sample which includes only those individual that are truly two years old in 2008, thus between 24 and 36 months old at the time of the survey. This reduces the dataset to 204 observations in 2008. The findings of self-productivity are once again the same as in Table 4 and 5 and so is cross-productivity for character skills, and the impact of monetary investment. Besides the insignificant estimate of θ3E on character skills, the co-efficients of emotional investment also supports the findings of the preferred specification. All models tested so far assume classical measurement errors. This assumption is
crucial to ensure that the joint use of several measurements reduces (or eliminates) the measurement error that a single measurement has. But if the measurement errors are not distributed independently, using several measurements might still fail to give an unbiased estimate. For this analysis, the single measurements are mainly components of one assessment. Still they were treated as independent observations which this is not necessarily true since the tests were conducted by the same interviewer and at the same time. Assuming classical measurement errors is therefore a questionable approach. For this reason, column (5) in Table 8 and 9 report the estimation results allowing for correlation between the measurements of the same test. Significance level and magni-tude of the estimates for cognitive skills are of similar size as in specification (1). For character skills, assuming non-classical measurement errors raises the size of almost all the estimates. Nevertheless, the coefficients of θC4, θE2 and θ3E are no longer significant due to increased standard errors.
Overall, the robustness checks fully confirm the self-productivity of skills over time. It also gives evidence for cross-productivity of skills between age two and three. The impact of parental investment is more sensitive to the different model specifications. Still, the positive impact of monetary investment on both types of skills at the age of four and of emotional investment on character skills at the age of seven was found to be robust.
T able 8: Cognitiv e skill formation, robustness chec k Time in v arian t co v ariates Reduced sample size Non-classical me asuremen t errors (3) (4) (5) θ C 3 θ C 4 θ C 7 θ C 3 θ C 4 θ C 7 θ C 3 θ C 4 θ C 7 θ C t− 1 0.745*** 0.615*** 0.949*** 0.605*** 0.444*** 0.791*** 0.637*** 0.550*** 0.809*** (0.130) (0.122) (0.250) (0.141) (0.127) (0.237) (0.137) (0.123) (0.197) θ N t− 1 0.229** -0.027 0.079 0.150 -0.080 0.131 0.259*** 0.009 0.096 (0.095) (0.096) (0.145) (0.103) (0.083) (0.155) (0.097) (0.095) (0.141) θ M t− 1 0.353*** 0.308*** 0.103 0.207 0.163** 0.114 0.154 0.170** 0.059 (0.129) (0.103) (0.093) (0.139) (0.079) (0.10) (0.118) (0.074) (0.078) θ E t− 1 0.065 0.004 0.222 0.023 0.021 0.197 0.061 -0.007 0.236* (0.131) (0.057) (0.143) (0.162) (0.059) (0.144) (0.142) (0.069) (0.136) Notes: Standard erro rs in paren theses. F or sp ecification (3) only age, gender, birth w eigh t and treatmen t w ere used as con trol v ariables. F or sp ecification (4) and (5) the usual set of con trol v ariables w as used. *: significan t at 10% lev el, **: significan t at 5% le v el, ** *: significan t at 1% lev el. T able 9: Character skills formation, robustness chec k Time in v arian t co v ariates Reduced sample size Non-classical me asur e men t errors (3) (4) (5) θ N 3 θ N 4 θ N 7 θ N 3 θ N 4 θ N 7 θ N 3 θ N 4 θ N 7 θ C t− 1 0.402*** -0.110 0.123 0.252* -0.057 0.257*** 0.392** -0.163 0.122 (0.111) (0.084) (0.077) (0.132) (0.101) (0.090) (0.164) (0.106) (0.082) θ N t− 1 0.645*** 0.224*** 0.253*** 0.680*** 0.232** 0.300*** 0.780*** 0.311*** 0.426*** (0.120) (0.083) (0.087) (0.143) (0.10) (0.108) (0.195) (0.116) (0.119) θ M t− 1 0.335*** 0.155** 0.019 0.206 0.186** -0.070 0.353* 0.179** -0.054 (0.129) (0.068) (0.045) (0.152) (0.092) (0.059) (0.188) (0.086) (0.055) θ E t− 1 0.286** -0.100 0.262*** 0.368* -0.158 0.452*** 0.286 -0.107 0.343*** (0.138) (0.067) (0.089) (0.190) (0.104) (0.121) (0.196) (0.084) (0.107) Notes: Standard erro rs in paren theses. F or sp ecification (3) only age, gender, birth w eigh t and treatmen t w ere used as con trol v ariables. F or sp ecification (4) and (5) the usual set of con trol v ariables w as used. *: significan t at 10% lev el, **: significan t at 5% le v el, ** *: significan t at 1% lev el.
8 Discussion and conclusion
The goal of this thesis was threefold: First to investigate the individual impact of mon-etary and emotional investment on cognitive and character skill, second to reveal the critical periods in the early childhood for parental investments and third to provide fur-ther evidence for the self- and cross-productivity of skills. The robustness check confirms a partial positive effect of providing toys, books and an appropriate environment. At the age of four, monetary investment raises the cognitive skill level by 0.17 SD and the character skill level by 0.14 SD. Emotional investment, which is measured among others by emotional climate and family integration, increases character skills significantly by 0.31 SD at the age of seven. No robustly significant impact on cognitive skill was found. The estimation results confirm the persistence of skills found by Cunha and Heckman (2008) and Helmers and Pradhan (2011). The self-productivity of cognitive skills peaks during the transmission from four to seven years of age with a coefficient of 0.83. Es-timates of the earlier periods reach a magnitude of 0.56 - 0.67. The self-productivity of past character skills is particularly high in the transmission from two to three years with a magnitude of 0.63 SD. In the later periods, the effect is shrinking to 0.25 - 0.27 SD and is clearly lower than the the impact of cognitive skills. Cross-productivity was found from two to three years, with a magnitude of 0.20 SD from character to cognitive skills and 0.35 SD in the other direction. This finding stresses the age sensitivity in the skill formation process and underlines the importance of analysing a multi-period model. The combination of self-productivity of skills and the positive impact of emotional and monetary investment has important implications for the effectiveness of ECD in-terventions. Although parental investment is only effective in certain periods, it is still capable of setting the self-enforcing process in motion and through this channel influence the skill levels substantially over time. A 1 SD rise in monetary investment in period 3 translates into a 0.14 SD rise in cognitive skills and a 0.04 SD rise in character skills in period 7 thanks to self-productivity of skills.
Yet another unexpected finding of this thesis is that parents invest more financial and emotional effort in children that are already better off in term of skills. This relationship reinforces the impact of parental investment: if parents invest more in a very early pe-riod, this will directly boost skill formation in the next pepe-riod, which will consequently ”attract” even more parental investment later in life. The other side of the coin is that children with a disadvantaged start may not receive the special care they need to catch up with their peers. External interventions could help to fill this gap. By fostering or even replacing the lack of parental investment, inequality in human capital formation will be reduced. The high degree of persistence of especially monetary investment sug-gests that even short term interventions will be effective in the long run. It also loosens the implications of the age-sensitivity of skills: if investment is persistent over time, the timing of investment will not matter greatly.
All the results of this analysis base on a reduced sample, but prior research showed strikingly different results depending on the sample cut (Elbers et al. forthcoming). In this previous work, the total sample was split according to the birth date into a young and an old cohort and showed significantly different results for the overlapping trans-mission period 3 - 4.13 The intent to replicate the four-latent-factor-model to the same two samples failed because the factor loadings of monetary investment measurements were not statistically significant. This indicates that the measurements used are not significantly affected by the latent factor. A feasible interpretation of the coefficients of interest is therefore impossible and the external validity can not be tested further using this sample. Additionally, the insignificant factor loading raise doubts about the reliability of the measurements.
Another drawback of this analysis is the interpretation of the estimation outcomes. To which extend investment in the early years affects real life outcomes in the future is difficult to quantify. Cunha and Heckman used wages and school completion as anchor variables in order to give a meaningful interpretation to the estimates. But due to the young age of the children, neither data on school outcomes nor any information on em-ployment status or wages is yet available. Hopefully future survey rounds will fill this gap and allow a more precise interpretation of the estimation outcomes.
13
The younger cohort included a similar subset of children as used for this analysis, the older cohort contained children of about 3, 4, 5, and 8 years of age.
References
Caldwell, B. M., Bradley, R. H., et al. (1984). Home observation for measurement of the environment. University of Arkansas at Little Rock Little Rock.
Carneiro, P., Crawford, C., & Goodman, A. (2007). The impact of early cognitive and non-cognitive skills on later outcomes. Centre for Economics of Education. Carneiro, P., & Heckman, J. (2003). Human capital policy. In J. J. Heckman, A. B.
Krueger, and B. M. Friedman (Eds.), Inequality in America: What Role for Human Capital Policies? Cambridge, MA: MIT Press.
Cunha, F., & Heckman, J. J. (2008). Formulating, identifying and estimating the tech-nology of cognitive and noncognitive skill formation. Journal of Human Resources, 43 (4), 738–782.
Cunha, F., Heckman, J. J., Lochner, L., & Masterov, D. V. (2006). Interpreting the evidence on life cycle skill formation. Handbook of the Economics of Education, 1 , 697–812.
Currie, J., & Almond, D. (2011). Human capital development before age five. Handbook of labor economics, 4 , 1315–1486.
Duckworth, A. L., & Seligman, M. E. (2005). Self-discipline outdoes iq in predicting academic performance of adolescents. Psychological science, 16 (12), 939–944. Dunifon, R., & Duncan, G. J. (1998). Long-run effects of motivation on labor-market
success. Social Psychology Quarterly, 33–48.
Elbers, C., Janssens, W., Gustafsson-Wright, E., & Wiegand, M. (in press). Early Childhood Development: The Dynamic Relationship between Parenting and Skill Formation in the Caribbean. AIID Research Series , Amsterdam Institute for International Development .
Engle, P. L., Fernald, L. C., Alderman, H., Behrman, J., O’Gara, C., Yousafzai, A., . . . others (2011). Strategies for reducing inequalities and improving developmental outcomes for young children in low-income and middle-income countries. The Lancet , 378 (9799), 1339–1353.
Gertler, P., Heckman, J., Pinto, R., Zanolini, A., Vermeersch, C., Walker, S., . . . Grantham-McGregor, S. (2013). Labor market returns to early childhood stim-ulation: A 20-year followup to an experimental intervention in Jamaica (Tech. Rep.). National Bureau of Economic Research.
Grantham-McGregor, S., Cheung, Y. B., Cueto, S., Glewwe, P., Richter, L., & Strupp, B. (2007). Developmental potential in the first 5 years for children in developing countries. The Lancet , 369 (9555), 60–70.
Hanushek, E. A., & Woessmann, L. (2012). Do better schools lead to more growth? Cognitive skills, economic outcomes, and causation. Journal of Economic Growth, 17 (4), 267–321.
Heckman, J. J. (2000). Policies to foster human capital. Research in economics, 54 (1), 3–56.
Heckman, J. J., Stixrud, J., & Urzua, S. (2006). The effects of cognitive and noncognitive abilities on labor market outcomes and social behavior (Tech. Rep.). National Bureau of Economic Research.
Helmers, C., & Patnam, M. (2011). The formation and evolution of childhood skill acquisition: Evidence from India. Journal of Development Economics, 95 (2), 252–266.
Janssens, W., Groot Bruinderink, M., & Van der Gaag, J. A. (2010). RCP Impact Report St Lucia 2009. AIID Research Series 110-01/1, Amsterdam Institute for International Development .
Janssens, W., Gustafsson-Wright, E., Tanovic, Z., Groot Bruinderink, M., & Van der Gaag, J. . (2011). RCP impact evaluation St. Lucia. Impact results 2010. AIID Research Series, Amsterdam Institute for International Development .
Janssens, W., & Rosemberg, C. (2009). Parenting behavior and child outcomes 2008. AIID Research Series , Amsterdam Institute for International Development . Janssens, W., & Rosemberg, C. (2013). The impact of a Caribbean home-visiting child
development program on cognitive skills. Economics of Education Review . Janssens, W., Rosemberg, C., & Van Spijk, J. (2009). RCP Impact Report St
Lu-cia 2008. AIID Research Series 09-01/1, Amsterdam Institute for International Development .
Malik, K. (2013). Human Development Report 2013. The Rise of the South: Human Progress in a Diverse World. UNDP-HDRO Human Development Reports. Nores, M., & Barnett, W. S. (2010). Benefits of early childhood interventions across
the world:(Under) Investing in the very young. Economics of Education Review , 29 (2), 271–282.
Pungello, E. P., Iruka, I. U., Dotterer, A. M., Mills-Koonce, R., & Reznick, J. S. (2009). The effects of socioeconomic status, race, and parenting on language development in early childhood. Developmental Psychology , 45 (2), 544.
Raviv, T., Kessenich, M., & Morrison, F. J. (2004). A mediational model of the associa-tion between socioeconomic status and three-year-old language abilities: The role of parenting factors. Early Childhood Research Quarterly , 19 (4), 528–547.
Rosen, J. A., Glennie, E. J., Dalton, B. W., Lennon, J. M., & Bozick, R. N. (2010). Noncognitive skills in the classroom: New perspectives on educational research. RTI Press publication No. BK-0004-1009.Research Triangle Park, NC.
Rosseel, Y. (2012). lavaan: An r package for structural equation modeling. Journal of Statistical Software, 48 (2), 1–36.
Shonkoff, J. P., & Phillips, D. A. (2000). From neurons to neighborhoods: The science of early childhood development. National Academies Press.
Tamis-LeMonda, C. S., Bornstein, M. H., & Baumwell, L. (2001). Maternal responsive-ness and children’s achievement of language milestones. Child development , 72 (3), 748–767.
Todd, P. E., & Wolpin, K. I. (2003). On the specification and estimation of the pro-duction function for cognitive achievement. The Economic Journal , 113 (485), F3–F33.
UNICEF. (2006). A study of child vulnerability in Barbados, St. Lucia and St. Vincent & the Grenadines. Christ Church, Barbados: UNICEF Office for Barbados and the Eastern Caribbean.
Walker, S. P., Wachs, T. D., Grantham-McGregor, S., Black, M. M., Nelson, C. A., Huffman, S. L., . . . others (2011). Inequality in early childhood: risk and protective factors for early child development. The Lancet , 378 (9799), 1325–1338.
Walker, S. P., Wachs, T. D., Meeks Gardner, J., Lozoff, B., Wasserman, G. A., Pollitt, E., & Carter, J. A. (2007). Child development: risk factors for adverse outcomes in developing countries. The lancet , 369 (9556), 145–157.
