The relationship between child health, developmental …jenni.uchicago.edu/Spencer_Conference/Papers 2010/SCHUNK SalmSc… · The relationship between child health, developmental

The relationship between child health, developmental gaps, and

parental education: Evidence from administrative data.

Martin Salm Tilburg University, The Netherlands

[email protected]

Daniel Schunk University of Zurich, Switzerland

[email protected]

November 2010

Preliminary version. Please do not cite or distribute.

We are greatly obliged to the department of health services of the city of Osnabrueck and in particular to Inge Rohling, MD, for making the data available and for patiently answering endless questions about the data and details of school entrance examinations. We thank Padmaja Ayyagari, Anders Björklund, Axel Börsch-Supan, Katie Carman, Jon Christianson, Janet Currie, Ernst Fehr, Hendrik Jürges, Steven Lehrer, Keren Ladin, Petter Lundborg, Jürgen Maurer, Andrew Oswald, Matthias Parey, Steffen Reinhold, Frank Sloan, Arthur van Soest, Rainer Winkelmann, participants of the conference of the American Society of Health Economists 2008 at Duke University, as well as seminar participants at the Universities of Aachen, Innsbruck, Mannheim, Tilburg, and Zurich for many valuable suggestions. We also thank Michael Ingenhaag and Dörte Heger for excellent research assistance.

2

Abstract

We use administrative German data to examine the role of physical and mental health conditions in explaining developmental gaps between children whose parents have different educational levels. Specifically, we employ sibling fixed effect models to estimate the effect of a comprehensive list of childhood health conditions – diagnosed by government physicians – on the cognitive and verbal ability of pre-school children. We also apply decomposition methods to examine the extent to which gaps in child development can be attributed to child health conditions. While most physical health conditions have small and insignificant effects, mental health conditions, in particular hyperactivity, have a large and significant effect on development. Mental health conditions account for 14% to 36% of the gap in cognitive ability and for 23% to 24% of that in verbal ability. Thus, policies aimed at reducing disparities in child development and socioeconomic inequalities later in life should focus more on the early diagnosis and effective treatment of mental health conditions.

Keywords: Childhood health, child development, socioeconomic inequality, human

capital formation

JEL Codes: J13, I20, I12

3

1. Introduction

Substantial developmental gaps between children of different socioeconomic groups are already measurable very early in life (Currie 2005, Heckman 2006). While measures of child development, such as cognitive or verbal ability, differ based on parental income and education, they also predict earnings, employment, and risk-taking behavior in adulthood (Currie and Thomas 2001, Heckman et al. 2006). This suggests that early developmental gaps are a source of severe socioeconomic inequalities later in life, and many claim that closing these early gaps is the most effective and efficient strategy for combating poverty and inequality (e.g., Currie 2009, Heckman 2007). However, policies that aim to close this gap depend on detailed knowledge about its formation.

In our study, we examine the effect of chronic childhood health conditions on the cognitive and verbal abilities of pre-school children, and we use decomposition methods to estimate the extent to which gaps in cognitive and verbal abilities by parental education groups can be attributed to child health. Cognitive and verbal abilities are still malleable at pre-school age and could thus be adversely affected by negative health conditions (see surveys by Currie 2009 and Behrman 1996). While previous studies, such as Currie (2005) and Case et al. (2005), emphasize that child health might be an important pathway for explaining gaps in child development between different socioeconomic groups, our study is the first to provide a comprehensive quantification of the extent to which differences in child health can explain these gaps. We use unique administrative data for one German city to estimate both the share of developmental gaps attributable to differences in the prevalence of health conditions, and the share attributable to differences in the magnitude of the effect of health conditions on development.

Previous studies find a robust positive gradient between parents’ socioeconomic status and child health in the United States (Case et al. 2002), but also in countries with universal health insurance coverage such as Canada (Currie and Stabile 2003) and the United Kingdom (Currie et al. 2007). The existing literature also provides evidence for a negative correlation between child health conditions and child development (Kaestner and Corman 1995, Paxson and Schady 2007). However, a negative correlation between child health and child development does not prove a causal impact of child health on child development. This correlation could also be due to unobserved family characteristics and environmental factors, which could simultaneously influence both child health and development. In order to account for these unobserved characteristics, Currie and Stabile (2006) use sibling fixed effects to examine the effect of child mental health conditions on child development. Ding et al. (2009) use genetic markers as

4

instrumental variables in order to examine the effect of ADHD, depression, and obesity on school achievements.

Our study contributes to the existing literature on child health and child development in several ways. First, our data on child health are based on detailed examinations administered by government pediatricians during elementary school entrance medical exams. The information obtained in these exams is far more reliable than that based on survey questions administered to children’s parents. Survey information was used in almost all previous studies, but it is well known that parents often have limited knowledge about their children’s health status and that the degree of this knowledge is strongly related to parent’s socioeconomic status (see Currie 2000). Second, the fact that these exams are compulsory for all children in Germany at the age of six years gives the data an extraordinary degree of representativeness. Third, we examine a much wider range of health conditions than those typically available in previous studies. We look at the effects of health conditions such as obesity, underweight, low birth weight, ear and eye conditions, mental health conditions, asthma, and allergies; this allows us to capture the inherently multi-dimensional nature of health in our estimations. Fourth, we use sibling fixed effects models to address omitted variables bias. Finally, our study goes beyond estimating the effect of child health conditions on child development by also quantifying the extent to which developmental gaps between parental education groups can be attributed to child health.

We find that child health conditions tend to be more common among children of parents with less education, and that most physical health conditions have small and insignificant effects on cognitive and verbal abilities. In contrast, mental health conditions, in particular hyperactivity, have a large and statistically significant effect on both cognitive and verbal ability before school entry. Moreover, mental health conditions account for a substantial share of the gaps in cognitive and verbal ability between children from different parental education groups. In alternative specifications, mental health conditions account for 14% to 36% of the gap in cognitive ability and for 23% to 24% of the gap in verbal ability. Differences in the prevalence of mental health conditions and in the magnitude of the effects of mental health conditions both contribute to the observed developmental gap. Germany has a very generous system of child health prevention and almost universal health insurance coverage. Our findings suggest that in addition to diagnosing and treating physical health conditions existing medical support programs should place enhanced focus on mental health. Also, learning more about the causes of and effective treatments for mental health impairments could be very valuable.

Our study continues as follows. Section 2 describes the data. Section 3 discusses the empirical strategy and section 4 presents and discusses our estimation results for child

5

cognitive ability as a development measure. Section 5 extends our results to verbal ability, and section 6 concludes.

2. Data

Estimation sample. Our analysis is based on administrative data from the department of health services of the city of Osnabrueck, Germany. Osnabrueck is the third largest city in the German state of Lower Saxony, with a population of approximately 163,000. The data were collected during official school entrance medical examinations, which are compulsory for all children at the age of 6 years. The examinations take place in the months before children enroll in elementary school, and take between 40 and 70 minutes, during which the children are randomly assigned to one of three female government pediatricians.

Our estimation sample includes information on 4,245 school entrance medical examinations, which took place in the years 2002 to 2005.1 The estimation sample for sibling fixed effects models is based on a subsample of 947 children with at least one sibling in the estimation sample. This sample consists of 866 children with one sibling, 69 children with two siblings, and 12 children with three siblings. For both the full sample and the sibling sample, we define two subsamples based on parents’ education: in the full sample, 1,439 children have at least one parent with a college degree, while neither of the parents have a college degree for the remaining sample of 2,806 children. In the sibling sample, 321 children have at least one parent with a college degree, and neither parent has a college degree for the remaining 626 children.

Outcome variables. Measures for child development are important predictors of achievement later in life (Heckman 2006). Children in our sample are approximately 6 years old and cognitive skill measures – such as raw IQ scores – are still malleable at this young age (see, e.g., Duyme et al. 1999). Thus, they could potentially be affected by health conditions. Moreover, children in our sample have not yet been to school, and schooling experiences have not yet affected developmental outcomes. We use two developmental measures as outcome variables. Our first outcome variable is based on Raven's Colored Progressive Matrices Test (CPM), a test developed specifically for children that measures abstract nonverbal reasoning ability (Raven et al. 1998). The test has been used successfully in a variety of cultural groups and is considered “culture-free” (Reynolds and Kamphaus 2003). It consists of a sequence of colored patterns; the child being tested must choose a missing pattern from a number of choices.2 The original test

1 Since the department of health services of the city of Osnabrueck stopped recording sociodemographic information in 2006, our data end in 2005. The Appendix describes further details about the data collection, missing variables, and the imputation of some missing observations. 2 All patterns are presented as large pictures (see Rohling 2002 for details), meaning that vision impairments will not confound CPM-test measures.

6

consists of 3 scales involving 12 items each. The school entrance exam in Osnabrueck includes 14 items from the test, 4 from the first scale, 4 from the second, and 6 from the third. The CPM score is the sum of the 14 items, i.e. 14 is the highest possible score. Figure 1 shows the cumulative frequency distribution of CPM scores by parental education. CPM scores tend to be higher across the entire distribution for children of college educated parents compared to children of less educated parents. Table 1 shows that the mean score for children of college educated parents is more than a third of a standard deviation higher than the mean score for children of less educated parents (11.37 versus 10.6); this difference is highly significant. Differences in CPM scores between children of college educated parents and less educated parents are slightly, and insignificantly, larger for the sibling sample (11.56 versus 10.49).

Our second outcome variable is a binary measure of verbal ability. This variable is based on the pediatricians’ assessment whether a child’s verbal ability corresponds to age level. The measurement is derived from a number of tests during which children are asked to describe what they see on a series of pictures, to repeat actual and imaginary words and sentences, and to talk freely about their favorite games and activities. The pediatrician assesses the accuracy of pronunciation and grammar as well as overall verbal ability. Table 1 shows that verbal ability is judged to be at age level for 88.4% of children with college educated parents, as compared to 78.7% for children of less educated parents. The difference in these shares is statistically highly significant, both in the full sample and the sibling sample.

Health variables. During the school enrollment examination, the pediatricians focus specifically on chronic health conditions, since only these conditions are relevant for the enrollment decision. Examinations were postponed for children with an acute illness that could influence their test results. Our analysis uses detailed information on all chronic health conditions available in the data. Information on health conditions comes from three sources: medical examinations during the school entrance examinations, children’s official vaccination and health records, and information provided by parents. During the examination, pediatricians measure the children’s height and weight and examine their visual and hearing abilities using standardized testing devices such as audiometers. Procedures for measuring eyesight and hearing are age adapted in a way that they are not confounded with verbal or cognitive skills.

Children above the 85th percentile of the body-mass-index distribution in our sample are classified as overweight, while those below the 5th percentile of the body-mass-index distribution in the full sample are classified as underweight.3 A binary variable on ear conditions takes on the value one if the examination with an audiometer

3 These definitions follow standard classifications from the U.S. Department of Health and Human Services (Center for Disease Control and Prevention, 2007).

7

reveals the child’s hearing abilities to be below the normal range. A binary variable on eye conditions takes on the value one if there are limitations in visual ability.4 Information on birth weight comes from the mandatory children’s health certificate that the attending pediatrician completes at the child’s birth. We create a binary variable for children with low birth weight (less than 2500 grams).5

Information on asthma and allergies is based on answers provided by parents in a questionnaire, on conversations with parents, on the observation of children, and on stethoscopic examinations of lung function during the exam. We create a binary variable for children who suffer from asthma. Information on allergies is collected in the same way as for asthma.6

The pediatricians use three sources to elicit overall assessment of mental health conditions. The first source of information is the pediatrician’s observations during the 50 to 70 minutes of interaction with the child during the examination. The second source is a form that parents are asked to fill out before the child’s examination; this form includes questions on the child’s behavior in everyday life. The third source is the Strengths and Difficulties Questionnaire (SDQ), which parents are asked to fill out in a separate room during their child’s examination. The SDQ is a standardized questionnaire about child mental health, developed in England and specifically designed for children between ages four and eleven (Goodman et al. 1998, Goodman 2001). It has been officially translated into over 40 languages, and the German version has been systematically evaluated (Klasen et al. 2000, Woerner et al. 2004). Parents are reassured that their answers will not influence the physicians’ school enrollment recommendation.

In alternative specifications, we use two different measures for mental health. The first measure is derived from the pediatrician’s overall assessment on whether a child has a mental health condition or not. This assessment is based on all three pieces of information described above, and it is likely to be more objective than parents’ answers, since different parents might have different standards in evaluating their child’s behaviors. The second measure is based on SDQ scores, in which parents rate 20 different attributes about their children’s behavior as being not true, partially true, or entirely true. The 20 attributes are subdivided into four subscales that allow us to create four binary variables identifying children with emotional problems, hyperactivity, conduct problems, and peer problems (Goodman 2001). Our specification based on SDQ

4 Exams testing vision are taken with glasses and include a test for the ability to see in three dimensions. Children who cannot distinguish red and green colors but who otherwise perform normally on all other visual tests are not classified as visually impaired. 5 This definition follows the International Statistical Classification of Diseases and Related Health Problems (World Health Organization 2007). 6 The definition of allergies includes allergic rhinitis and eczema.

8

scores provides a robustness check for the results based on physician’s assessments and also allows disentangling the effect of different types of mental health conditions.

Table 1 shows summary statistics on the prevalence of health conditions among children in our sample. 14.8% of the children in the full sample are overweight and 4.7% are underweight. 5.7% were born with low birth weight, 11.0% are affected by ear problems, and 19.3% by eye problems. 18.2% have a mental health condition. 6.4% of the children have asthma, and 3.5% suffer from allergies. Table 1 shows further that many health conditions are significantly more common among children of less educated parents than among those of college educated parents. This relationship holds and is statistically significant for being overweight (17.7% vs. 9.1%), low birth weight (6.3% vs. 4.5%), and for asthma (6.8% vs. 5.4%). We also find significant differences between parental education groups for mental conditions (20.7% vs. 13.2%), and all four dimensions of mental health measured by the SDQ-test confirm this (hyperactivity: 18.5% vs. 9.3%, emotional problems: 12.5% vs. 10.0%, peer problems: 9.5% vs. 5.8% conduct problems: 7.3% vs. 4.0%). The prevalence of most health conditions in the sibling sample is similar to the full sample. However, children in the sibling sample are less likely to be overweight (11.9% versus 14.8%) and much more likely to be low birth weight children (10.0% versus 5.7%). The high share of low birth weight children in the sibling sample can largely be explained by the twins in this sample, who tend to have considerably lower birth weight.

Socioeconomic and demographic variables. We also include information on socioeconomic and demographic characteristics of children and their parents. We control for children’s age, gender, preschool attendance (3+ years vs. less than 3 years), and type of family (living with both parents vs. other types). We also control for birth order by including the corresponding three dummy variables (fourth or later child is omitted category), and the number of siblings (one sibling is omitted category). These variables refer to the total number of siblings, not just those included in the sample. We further control for the mother's age at the time of the examination, for children who have at least one parent in full time employment, and for children with ethnic origins in Turkey, in Eastern European countries, or other foreign countries. Sample means and standard deviations for all included variables are listed in Table 1.

3. Determinants of child development and empirical specification

Our aim is to estimate the effects of child health conditions on child development. Outcome variables are the CPM score as a measure of cognitive ability and a binary indicator for verbal ability. A growing body of work suggests that socioeconomic and health conditions in early childhood have fundamental effects on development (e.g., Heckman 2007). In line with this work, we assume that cognitive and verbal ability are

determined by a human capital production function which links child development to child health conditions, socioeconomic conditions, and other input factors:

9

)( , ,ic ic ic icy f H X u= (1) Here, icy is the outcome variable; is a vector of chronic child health conditions; icH icX is a vector of observed socioeconomic and demographic characteristics; and scalar represents unobserved characteristics.

icu{1.. }c C∈ indexes children in a given family, and

indexes families. {1i N∈ .. }The years before school entry are critical for the development of cognitive and

verbal abilities. Cognitive ability is still malleable at this age (e.g., Duyme et al. 1999), and chronic health conditions can interfere with the process of cognitive and verbal development in various ways (Currie 2009). For example, pain, stress, and fatigue reduce the ability to concentrate and to learn. They can crowd out activities that are beneficial to the regular cognitive stimulation necessary for development. Also, illness can change relationships with parents and others in ways that might disturb cognitive and verbal development. Finally, some health conditions can have direct negative impacts on outcome variables such as hearing impairments on verbal ability. Our study only considers chronic health conditions which can adversely affect child development for a prolonged period.

In addition to the child's chronic health characteristics, equation (1) also includes socio-economic and demographic characteristics as further determinants of child human capital (see Table 1). Previous studies show a strong correlation between child outcomes and socioeconomic characteristics such as parental education and parental employment (Carneiro et al. 2007, Ruhm 2004). Child outcomes are also known to vary with demographic characteristics such as birth order, number of siblings, gender, age, and ethnic background. The socioeconomic and demographic characteristics included in our estimations can be viewed as proxy variables for different levels of initial endowment and different levels of parental investment in child human capital. Overall, variables on socio-economic and demographic characteristics fall in two categories: variables for characteristics that are constant for all siblings in the same family, such as ethnic origin, and variables that vary between siblings, like gender or birth order, for example.

The model in equation (1) also accounts for unobserved determinants of child outcomes uic. These unobserved determinants can be decomposed into a family-specific component – e.g. parenting style – and into a component that is specific to the child. Hence, our model assumes that the unobserved component can be expressed as the sum of two parts:

icu

ic i icu a ε= + (2) This follows standard notation for panel data models; scalar is a family-specific com-ponent for family i, and scalar

ia

icε is a child-specific component for child c in family i.

Finally, equation (1) also allows for interactions between the effect of health conditions and the effects of socioeconomic and demographic characteristics on child outcomes. The effect of health conditions on child outcomes can vary according to parental education, for example, because the ability to compensate for the negative consequences of health conditions could depend on parental background. In our study, we stratify our sample by parental education and estimate separate models for the sample of children with college educated parents, for the sample of children with less than college educated parents, and for the pooled sample of all children. For each group, we estimate both linear models without family fixed effects and linear models which include family fixed effects. In all cases our estimation allows for robust standard errors that are clustered at the family level. The estimation equation for the model without family fixed effects is:

' 'ic edu ic edu ic edu icy H Xμ β γ= + + + u (3)

Child outcome icy is the dependent variable; child health conditions and socioeconomic and demographic characteristics

icH

icX are explanatory variables. βedu and γedu are vectors of estimation coefficients. eduμ is the intercept. Subscript edu refers to the estimation samples which are defined by parental levels of education. βedu is the coefficient of interest and represents the effect of health conditions on child outcomes. We can estimate βedu consistently with OLS under the following exogeneity assumption:

[ | , ] [ | , ]ic ic ic i ic ic icE u H X E a H X 0ε= + = (4) This assumption could be violated, however, if there are unobserved family-

specific characteristics, such as parenting style, which influence both health conditions and outcome variables. This is the standard panel data situation when there are reasons to suspect that unobserved group-specific characteristics iα could be correlated with explanatory variables. Therefore, we also estimate sibling fixed effects models, which control for unobserved family-specific characteristics by examining the effect of differences between siblings’ health on differences in siblings’ outcome variables. The fixed effects model is:

' 'ic ic edu ic edu i icy H Z aβ γ= + + +ε (5)

Vector icZ represents child-specific variables such as gender and birth-order. eduβ is the effect of child health conditions on outcome variables. This effect can be estimated consistently by fixed effects estimation if the following strong exogeneity assumption holds: 1 2 1 2[ | ,.., , , ,.., ] 0ic i iC i i iCE H H H Z Z Z,iε = (6)

This condition makes no assumptions about the family-specific unobserved component iα . The child-specific unobserved component icε , however, must be uncorrelated not only with the child’s own explanatory variables and icH icZ , but also with the sibling’s explanatory variables. In the following paragraphs, we discuss whether the strong exogeneity assumption is plausible in the context of our study such that βedu

10

can be estimated consistently. Specifically, we start by discussing three causes of potential violations of the strong exogeneity assumption, and we then present a statistical test for this assumption.

First, a violation of the strong exogeneity assumption could arise if siblings face systematic differences in available resources or parental treatment that affect both their health and development. These differences could occur, for example, if parents smoke in the presence of one child (thus increasing the risk of asthma), but not in the presence of the other, and if, at the same time, they would play memory games with only one child. Such differential treatment could simultaneously affect and icH icε , thereby violating the strong exogeneity assumption. However, we consider such differential treatment unlikely in our sample. All siblings in our data are very close in age, implying that the family's socioeconomic situation they experienced during childhood was similar. In our analysis, we also control for gender and birth order, two factors which could explain differential treatment. Furthermore, there is evidence that parents treat children equally when dividing wealth and other resources, or even allocate more resources to a disadvantaged child (Behrman and Rosenzweig 2004, Kuenemund et al. 2005). It is possible that parents invest more time in children with chronic health conditions, or that a sibling’s negative health condition has an adverse impact on child outcomes. Such compensating behavior, however, would only bias our estimates downward.

A second cause of violation of the strong exogeneity assumption could be reverse causality. For example, if poor cognitive ability caused health problems, this would imply a negative correlation between and icH icε . However, parents rather than children diagnose needs for medical or psychological services and initiate treatments for children at the ages examined in our study, and parents play the key role in the supervision of these treatments. Moreover, our estimations only include children who have not yet attended school. Even for mental health conditions, an effect of cognitive ability is to be expected only after children have been confronted with their low school performance (Currie and Stabile 2006). Thus, it is very unlikely that low cognitive ability causes chronic health conditions. Finally, we can think of only rare cases in which health and cognitive ability have mutual determinants, for example severe forms of birth asphyxia.7

A third reason for violation of the strong exogeneity assumption could be measurement errors in explanatory variables. In this case, observed health conditions measure true health conditions only with errors, and true health conditions are correlated with

icH

icH icH

icε , causing a downward bias of estimation coefficients for health

11

7 The best indicator for birth asphyxia is the pH-Value of the umbilical cord. Unfortunately, however, umbilical pH-values were only collected for a subsample of 3,400 children. To investigate the effect of birth asphyxia, we have also estimated our models on this subsample, including indicators for abnormally low values for the pH-value in the umbilical cord (1%, 2.5%, 5%, 10%, 15%, 20, and 25% bottom percentile) or a continuous variable for the pH-value. We find that the corresponding coefficient is always insignificant (p>0.95) and its inclusion has no effect on any of the other included coefficients.

conditions. Sibling fixed effects estimation can exacerbate such a bias by increasing noise due to potentially strong correlations between siblings and reducing exogenous variation. Specifically, Griliches (1979) and Bound and Solon (1999) show that measurement errors can result in substantial downward bias of coefficients in models with sibling fixed effects. However, biases arising from this source seem much less of a problem in our study than in existing studies for the following two reasons. First, there is substantial variation in health conditions between siblings, and second, our measures for health conditions are derived from administrative data, which were collected by experienced physicians in highly standardized procedures. There is much less scope for measurement error in our health variables than in previous studies on child health, which use data based on survey questions, and particularly less than in the studies discussed by Griliches (1979) and Bound and Solon (1999), which estimate the returns to education on wages.

In the paragraphs above, we argued in favor of maintaining the strong exogeneity assumption 1 2 1 2[ | , ,.., , , ,.., ] 0ic i i iC i i iCE H H H Z Z Zε = . In addition to verbal reasoning, we check the validity of this assumption with a statistical test suggested by Wooldridge (2002, p. 285) which is based on the following equation:

2 1 2 1 2 1 2 1( ) ' ( ) ' (i i i i edu i i edu i iy y H H Z Z )β γ ε ε− = − + − + − (7) This equation is derived from the fixed effect model in equation (5): by taking differences between siblings, the family-specific unobserved component is eliminated, and the difference in the child-specific unobserved components

ia

i2i 1ε ε− can be isolated. Thus, under the strong exogeneity assumption, the error term in equation (7) should not be related to health conditions or . Consequently, we can test the strong exogeneity assumption by adding (or , respectively) to the equation and by applying an F-test for the significance of (or , respectively). If the coefficients are jointly not significantly different from zero, we can maintain the hypothesis that our sibling fixed effects model consistently estimates the effect of child health conditions on child outcomes.

1iH

1iH

2iH

2

2iH1iH iH

Having estimated the sibling fixed-effects specifications separately for all three samples, we proceed by examining how much of the developmental gap can be attributed to differences in the prevalence and severity of negative health conditions. For this purpose we use decomposition analysis, similar to Oaxaca (1973). We first calculate the part of developmental gaps which can be attributed to differences in the prevalence of health conditions. For this calculation, we keep the magnitude of the effect of health conditions on child development as well as all non-health factors constant between education groups. The prevalence effect is then calculated as follows:

poolednocollegecollege HH-effectprevalence β’)( -= (8) Here, collegeH is the vector of mean values for health conditions for the sample with college educated parents, while nocollegeH is the corresponding vector for the sample with

12

less educated parents. is the vector of estimation coefficients for the reference group. Following Neumark (1988) we take estimation coefficients for the pooled sample as the reference group.

pooledβ

We also calculate the extent to which developmental gaps can be attributed to differences in the magnitude of the effect of health conditions on child development, i.e. the severity effect. The magnitude of the effect of health conditions for the sample with college educated parents is measured by collegeβ , while the magnitude of this effect for the sample with less educated parents is measured by nocollegeβ . To calculate the impact of differences in the magnitude of health effects, we keep the prevalence of health conditions as well as all non-health factors constant between parental education groups. The severity effect is then calculated as follows:

)(’nocollegecollegepooledHffectseverity-e ββ -= . (9)

The vector pooledH refers to the prevalence of health conditions for the pooled sample. The sum of the prevalence effect and the severity effect constitutes the total effect of child health conditions on the developmental gap.

4. Results

4.1. Estimation results for CPM score

Estimation results for the effect of physician-diagnosed health conditions on cognitive ability are shown in Table 2. Columns 1 to 3 of Table 2 present results for OLS estimation without family fixed effects. Column 1 shows results for the pooled sample, which includes both children with highly educated parents and children with less educated parents. We find a strong and significantly negative association between most health conditions and lower cognitive ability. The negative association is strongest for mental health conditions. CPM scores for children with a mental health condition are reduced by 1.07 points, equivalent to about half of a standard deviation. Column 2 shows results for the sample with college educated parents and column 3 shows results for the sample with less than college educated parents. Estimation coefficients for both physical and mental health conditions tend to be more negative for the sample with less educated parents and the coefficient for mental health conditions is large and significant for all samples.

Columns 4 to 6 of Table 2 show fixed effects estimation results, which control for unobserved family characteristics by comparing cognitive ability scores for siblings with and without health conditions. The sample is restricted to children with at least one sibling in our estimation sample and includes 947 observations in the pooled sample. For the pooled sample, none of the coefficients for physical health conditions is statistically significant. Further, coefficients for physical health conditions tend to be less negative for

13

14

fixed-effect estimation than for OLS estimation, suggesting that OLS coefficients for physical health conditions might be considerably biased and should not be interpreted as causal effects.8 In contrast, mental health conditions have a large and significant negative impact on cognitive ability; they reduce CPM scores by around 1.2 points. The size of this coefficient is similar for fixed effect estimation and for OLS estimation, suggesting no further bias in the OLS estimates for mental health conditions. The effects of mental health conditions are large and significant not only for the pooled sample, but also for the sample with college educated parents (column 5), and even stronger for the sample with less educated parents (column 6). Moreover, we also find that asthma has a large and marginally significant effect on cognitive ability for children of less educated parents, while its effect is insignificant and close to zero for the sample with college educated parents. Asthma can severely limit children’s activity if not treated adequately; however, if well controlled by the parents, asthma attacks can be prevented, and it should have little or no direct effect on cognitive development (Currie 2005). These findings suggest that college educated parents are better able to mitigate the negative consequences of certain child health conditions. F-tests for joint significance of the health variables support this interpretation. They reveal that health conditions have a strongly negative effect on cognitive skills for the pooled sample (p=0.01) and for children of less educated parents (p=0.00), but they are insignificant for the sample of college educated children (p=0.10).

The fixed effects estimators rely on the strong exogeneity assumption that the child specific error term is not related to observed characteristics. In order to test for the validity of this assumption, we employ the test statistic described in Section 3 and find no evidence that the strong exogeneity assumption is violated (p>0.30 in all three samples).

Our main finding is that mental health conditions have strong and significantly negative effects on cognitive ability in all three samples. We thus explore the effect of mental health conditions further by investigating an alternative specification which includes binary indicators for hyperactivity, emotional problems, peer problems, and conduct disorders based on SDQ subscales (see Table 3). In our alternative specification, these four indicators replace the variable for overall physician diagnosed mental health problems, and they allow us to determine which types of mental health conditions have the strongest effect on cognitive ability.

Columns 1 to 3 of Table 3 show OLS estimation results for the alternative specification. The estimation coefficients for all health conditions other than the mental conditions are similar to the OLS specifications shown in Table 2. Columns 4 to 6 of Table 3 show fixed effect estimation results. As before, there is no evidence that the

8 Fixed effects coefficients are significantly different from coefficients based on OLS estimations (p<0.01).

15

strong exogeneity assumption for the fixed effect estimation might be violated (p>0.40 in all three samples), and an F-test shows that health variables are jointly significant for the sample of less educated parents (see Table 3). We also find that coefficients for models with family fixed effects are significantly different from coefficients in OLS models (p<0.01 in all samples). Focusing on specific health conditions, we find for the pooled sample (column 4), that hyperactivity, emotional problems, and peer problems all have a significantly negative effect on cognition, with the strongest effect for hyperactivity. For children with college educated parents, none of the mental health variables have a significant effect (column 5), but hyperactivity significantly reduces cognitive ability (column 6) for the sample with less educated parents. Estimation coefficients for all other health conditions are not significant.

Our results are in line with previous findings that hyperactivity-related disorders at a young age have large effects on test scores of older youth (e.g. Currie and Stabile 2006, Currie et al. 2010), and we show that these effects tend to be larger than those of a wide range of physical health conditions. Moreover, we provide evidence that specific types of mental problems other than hyperactivity, such as emotional problems and peer problems, also have negative, albeit somewhat weaker effects on cognitive ability. Emotional problems and peer problems are characteristic of introverted children who tend to be unhappy and often cry as well as of children without friends and with few social interactions. The correlation between emotional and peer problems is much stronger than between all other SDQ subscales in our data. This suggests that social isolation, introversion, and unhappiness often coincide in children and are associated with less exposure to cognitive learning experiences in social interactions with friends. In contrast, conduct problems – i.e. children who do not obey their parents – do not seem to have a negative effect on cognitive ability.

4.2. Decomposition analysis

Descriptive statistics in Table 1 show that there is a significant gap in children’s cognitive abilities between parental education groups. The total difference in mean CPM scores between parental education groups is 1.057 points for the sibling sample or about half a standard deviation. While most physical health conditions have no significant effect on cognitive ability, mental health conditions have a significantly negative effect which differs according to parental education. These findings raise the question as to which share of the observed gap in cognitive ability can be attributed to mental health conditions. We use decomposition analysis to answer this question. Table 4 shows the results for the decomposition analysis for the baseline fixed effect specification (Panel A)

16

and for the alternative fixed effect specification with the four mental health variables based on SDQ subscales (Panel B).9

The first column of Table 4 shows the prevalence effect. The prevalence effect defines the share of the cognitive ability gap that differences in the prevalence of health conditions can explain. It accounts for 10.9% of the gap in cognitive ability based on the analysis with the physician diagnosed mental health variable in Panel A, and for 9.8% of this gap based on the analysis with the four mental health conditions based on SDQ scales in Panel B. Even if we focus only on hyperactivity problems, the prevalence effect accounts for 3.4% of the gap. 95% confidence intervals, calculated based on 200 bootstrap replications, reveal that the prevalence effects are significantly different from zero for all specifications. The second column of Table 4 shows results for the severity effect. The severity effect explains which share of the cognitive ability gap can be accounted for by differences in the magnitude of the effect of health conditions on cognitive ability. It accounts for 2.9% of the gap in cognitive development in Panel A and for 22.8% of the gap in Panel B. Hyperactivity alone accounts for 18% of the gap in cognitive development. Severity effects are measured somewhat less precisely and – except for the decomposition of the hyperactivity condition – they are not significant.

The third column shows the total effect – i.e. the sum of the prevalence effect and the severity effect – of child health conditions on cognitive ability. In total, mental health conditions account for 13.5% of the gap in cognitive ability in Panel A and for 36% in Panel B. Hyperactivity alone accounts for 23.5% of the cognitive ability gap. This lends further support to our observation that hyperactivity, out of all the four considered mental health conditions, has the strongest effect on cognitive ability. The corresponding confidence intervals show that the estimated total health effects are significant for Panel B and almost significant for Panel A.

Overall, our findings from the decomposition analysis suggest that mental health conditions are a very important factor in explaining gaps in cognitive ability by parental education. In a back-of-the-envelope calculation, Currie (2005) suggested that child health might explain around 12% of the cognitive ability gap between white and black children in the United States at school entry. Compared to our results for mental health alone, this looks like a conservative estimate.

5. Extensions

In the previous section, we focused on cognitive ability as a measure for child development. In this section, we show that similar findings are obtained from extending our study to verbal ability as an outcome variable. Our estimation results for the effect of

9 Further decomposition results for physical health conditions and for non-health characteristics are shown in tables A8 and A9 of the appendix.

17

health conditions on verbal ability are shown in Tables 5 and 6. Based on estimation with sibling fixed effects, and as in our findings in the analysis of CPM scores, only mental conditions have a significantly negative impact on verbal ability among all health condi-tions (Table 5). Furthermore, among mental health conditions, hyperactivity-related dis-orders have the most robust effects on verbal ability (Table 6). We find no evidence that the strong exogeneity assumption for the fixed effect estimation might be violated in any of the specifications (p>0.10 in all samples). We also find that coefficients for models with family fixed effects are significantly different from coefficients in OLS models (p<0.01 in all samples). Table 7 reports the decomposition analysis for verbal ability. Based on the baseline specification, the total effect of child mental health conditions on verbal ability is estimated to account for 23.1% of the ability gap (Panel A), while our alternative specification finds the total effect of mental health conditions to account for 24% (panel B). As in our analysis of the CPM score, hyperactivity alone accounts for most of the mental health effect (panel B). Similar to our previous findings, the prevalen-ce effect is also estimated to be significant; it accounts for 8.3% of the total gap in both specifications. Overall, the analysis for verbal ability corroborates the finding that mental health conditions are an important factor for explaining gaps in child development.

6. Conclusions

There are substantial developmental gaps between children of different socioeconomic groups. Existing studies, e.g. the large literature on the black-white test score gap (e.g., Currie 2000, Fryer and Levitt 2004), has demonstrated that early developmental gaps are a source of severe social and economic inequalities later in life. Policies that intend to close these gaps depend on knowledge about the underlying pathways behind its formation. In this paper, we use unique administrative data from German elementary school medical entrance examinations in order to examine the role of child health in explaining developmental gaps between children of different parental education groups. Our data contain professionally measured information on child health conditions such as obesity, underweight, low birth weight, ear problems, eye problems, mental conditions, asthma, and allergies. This allows us to capture the inherently multi-dimensional nature of health in our estimations.

We find large differences in child cognitive and verbal ability by parental education groups, and we also find that both physical and mental health conditions are more common among children of less educated parents. While most physical health conditions have small and insignificant effects, mental health conditions, in particular hyperactivity, have a large and significant effect on child development. This effect is large across parental education groups, but it is strongest for children whose parents have less education. Mental health conditions account for a substantial share of developmental

18

gaps: based on estimation with sibling fixed effects, we find that mental health conditions account for 14% to 36% of the gap in cognitive ability and for 23% to 24% of the gap in verbal ability. 10% to 11% of the gap in cognitive ability as well as 8% of the gap in verbal ability can be attributed to differences in the prevalence of health conditions, and the remainder can be attributed to differences in the magnitude of the effects.

Our findings have several policy implications. They suggest that the German child health and prevention system tends to diagnose and treat physical health conditions rather adequately, meaning that they do not have negative effects on children’s cognitive and verbal abilities. The German health system is characterized by almost universal health insurance coverage and a focus on child health and prevention programs. 99.8% of the German population are enrolled in mandatory health insurance and those who are not enrolled are mostly the very rich (German Federal Statistical Office 2004). Since 1971, there has been a child health prevention program which involves nine pediatric examinations between birth and the age of five years. The German health insurance system pays for all examinations as well as subsequent treatment costs – without any deductible. Attendance at the first seven of these examinations is well above 90%, and attendance of the ninth examinations is still about 80% (Schubert et al. 2004).

However, while our findings indicate adequate treatment of physical health conditions, they suggest that mental health conditions might remain a problem. First, mental health conditions explain a substantial share of developmental gaps between children of different parental education groups. In view of the institutional facts described above, this finding cannot only be attributed to differences in access to care by parental education groups. It implies that there could be a need for further policy support to remedy the impact of family background on the ability to address mental health conditions appropriately. Second, mental health conditions have large and significant effects on child development for all parental education groups. Hence, it may well be that the health care system neither diagnoses nor treats mental health conditions adequately. A better understanding of the causes and effective treatments of mental health conditions could have large future benefits. This could include health training for preschool supervisors and parents of small children, which enable them to better recognize and address children’s mental health problems.

In summary, our findings confirm that it is important for policy to better understand the lifecycle of skill and health formation, and we suggest that policies aimed at reducing disparities in child development should also aim at reducing disparities in mental health with a focus on children from disadvantaged families.

19

References

Behrman, J. (1996): The impact of health and nutrition on education. The World Bank Research Observer, 11, 23-37.

Behrman, J. and M. Rosenzweig (2004): Parental allocations to children: New evidence on bequest differences among siblings. Review of Economics and Statistics, 86, 637-640.

Bound, J. and G. Solon (1999): Double trouble: on the value of twin-based estimation of the return to schooling. Economics of Education Review, 18, 169-182.

Carneiro, P., C. Meghir, and M. Parey (2007): Maternal education, home environments and the development of children and adolescents, IZA Discussion Paper 3072.

Case, A., D. Lubotsky, and C. Paxson (2002): Economic status and health in childhood: The origins of the gradient. American Economic Review, 92, 1308-1334.

Case, A., A. Fertig, and C. Paxson (2005): The Lasting Impact of Childhood Health and Circumstance. Journal of Health Economics, 2005, 24, 365-389.

Center for Disease Control and Prevention (2007): U.S. Department of Health and Human Services. http://www.cdc.gov/nccdphp/dnpa/bmi/ childrens_BMI/about_childrens_BMI.htm#How%20is%20BMI%20calculated; accessed May 7, 2007.

Currie, J. (2000): Child health in developed countries. In: Handbook of Health Economics, Vol. 1B, edited by A. Culyer and J. Newhouse. North Holland.

Currie, J. (2005): Health disparities and gaps in school readiness, in: The Future of Children: School Readiness: Closing Racial and Ethnic Gaps. 15, 1, 117-138.

Currie, J. (2009): Healthy, Wealthy and Wise: Socioeconomic Status, Poor Health in Childhood, and Human Capital Development. Journal of Economic Literature. 47, 1, 87-122.

Currie, J. and M. Stabile (2003): Socioeconomic status and child health: Why is the relationship stronger for older children? American Economic Review, 93, 1813 – 1823.

Currie, A., M. Shields, and S. Price (2007): Is the child health / family income gradient universal? Evidence from England. Journal of Health Economics, 26, 212-232.

Currie, J. and M. Stabile (2006): Child Mental Health and Human Capital Accumulation: the Case of ADHD. Journal of Health Economics, 25, 6, 1094-1118.

Currie, J. and D. Thomas, (2001): Early Test Scores, Socioeconomic Status, School Quality and Future Outcomes. Research in Labor Economics. 20, 103-132.

Currie, J., Stabile, M., Manivong, P., and L. Roos (2010): Child Health and Young Adult Outcomes. Forthcoming: Journal of Human Resources.

20

Ding, W., S. Lehrer, J. Rosenquist, and J. Audrain-McGovern (2009): The impact of poor health on academic performance: new evidence using genetic markers. Journal of Health Economics, 28, 3, 578-597.

Duyme, M., A. Dumaret, and S. Tomkiewicz (1999): How can we boost IQs of ‘dull children’? A late adoption study. Proceedings of the National Academy of Sciences. 96, 8790-8794.

Fryer, R. G. and S. Levitt (2004): Understanding the Black-White Test Score Gap in the First Two Years of School. Review of Economics and Statistics, 86, 2, 447-464.

German Federal Statistical Office (2004): Krankenversicherung in Deutschland. Sonderauswertung des Mikrozensus 2003. Wiesbaden.

Goodman, R. (2001): Psychometric properties of the Strengths and Difficulties Questionnaire (SDQ). Journal of the American Academy of Child and Adolescent Psychiatry, 40, 1337-1345.

Goodman, R., H. Meltzer, and V. Bailey (1998): The Strengths and Difficulties questionnaire: A pilot study on the validity of the self-report version. European Child & Adolescent Psychiatry, 7, 125-130.

Griliches, Z. (1979): Sibling models and data in economics: beginnings of a survey. Journal of Political Economy, 87, S37-S64.

Heckman, J. (2006): Skill Formation and the Economics of Investing in Disadvantaged Children. Science, 312, 1900 – 1902.

Heckman, J. (2007): The economics, technology, and neuroscience of human capability formation. Proceedings of the National Academy of Sciences. 104, 13250–13255.

Heckman, J., J. Stixrud, and S. Urzua (2006): The Effects of Cognitive and Noncognitive Abilities on Labor Market Outcomes and Social Behavior. Journal of Labor Economics, 24, 3, 411-482.

Klasen, H., W. Woerner, D. Wolke, R. Meyer, S. Overmeyer, W. Kaschnitz, A. Rothenberger, R. Goodman (2000): Comparing the German Versions of the Strengths and Difficulties Questionnaire (SDQ-Deu) and the Child Behavior Checklist. European Child & Adolescent Psychiatry, 9, 271-276.

Kaestner, R. and H. Corman (1995): The impact of child health and family inputs on child cognitive development. NBER Working paper 5357.

Kuenemund, H., A. Motel-Klingebiel, and M. Kohli (2005): Do Inter-generational Transfers From Elderly Parents Increase Social Inequality Among Their Middle-Aged Children? Evidence from the German Aging Survey. The Journals of Gerontology Series B: Psychological Sciences and Social Sciences, 60, S30-S36.

Neumark, D. (1988): Employers’ discriminatory behaviour and the estimation of wage discrimination. Journal of Human Resources, 23, 279-295.

21

Oaxaca, R. (1973): Male-female wage differentials in urban labor markets. International Economic Review, 14, 693 – 709.

Paxson, C. and N. Schady (2007): Cognitive development among young children in Ecuador. Journal of Human Resources, 42, 49-84.

Raven, J., J.C. Raven, and J.H. Court (1998): Couloured Progressive Matrices. Oxford Psychologists Press.

Reynolds, C.R. and R.W. Kamphaus, Eds. (2003): Handbook of Psychological and Educational Assessment of Children: Intelligence, Aptitude, and Achievement (2nd edition). Guilford Press.

Rohling, I. (2002): Gesundheit und Entwicklungsstand der Osnabruecker Schulanfaenger. Stadt Osnabrueck, Fachbereich Soziales und Gesundheit.

Ruhm, C. R. (2004): Parental Employment and Child Cognitive Development. Journal of Human Resources, 39, 1, 155-192.

Schubert, I. and K. Horch (2001): Schwerpunktbericht der Gesundheits-berichterstattung des Bundes. Gesundheit von Kindern und Jugendlichen. Verlag Robert-Koch-Institut, Berlin.

Wooldridge, J. (2002): Econometric Analysis of cross section and panel data. MIT Press.

Woerner, W., A. Becker and A. Rothenberger (2004): Normative data and scale properties of the German parent SDQ. European Child & Adolescent Psychiatry, 13, ii3-ii10.

World Health Organization (2007): International Statistical Classification of Diseases and Related Health Problems – Version for 2007. http://www.who.int/classifications/apps/icd/icd10online/?gp05.htm+p05, accessed May 7, 2007.

22

Table 1: Descriptive statistics

Full sample Sibling sample

Variable

Pooled sample

Parents college

Parents no college ≠ sign. Pooled

sample Parents college

Parents no college ≠ sign.

CPM score (0-14) 10.857 11.368 10.596 *** 10.850 11.555 10.489 *** (2.172) (2.012) (2.204) (2.203) (1.877) (2.271) Verbal ability at 0.820 0.884 0.787 *** 0.800 0.872 0.763 *** age level (0.385) (0.321) (0.410) (0.400) (0.335) (0.426) Overweight 0.148 0.091 0.177 *** 0.119 0.062 0.149 *** (0.355) (0.288) (0.382) (0.324) (0.242) (0.356) Underweight 0.047 0.048 0.046 0.052 0.053 0.051 (0.211) (0.214) (0.210) (0.222) (0.224) (0.220) Low birth weight 0.057 0.045 0.063 ** 0.100 0.081 0.110 (0.231) (0.208) (0.243) (0.301) (0.273) (0.313) Ear condition 0.110 0.117 0.106 0.116 0.143 0.102 * (0.313) (0.321) (0.308) (0.321) (0.351) (0.303) Eye condition 0.193 0.186 0.196 0.196 0.162 0.214 * (0.395) (0.389) (0.397) (0.398) (0.369) (0.411) Asthma 0.064 0.054 0.068 * 0.071 0.075 0.069 (0.244) (0.227) (0.253) (0.257) (0.263) (0.253) Allergy 0.035 0.033 0.036 0.039 0.047 0.035 (0.183) (0.178) (0.185) (0.194) (0.211) (0.184) Mental condition 0.182 0.132 0.207 *** 0.191 0.128 0.224 *** (0.386) (0.339) (0.406) (0.393) (0.334) (0.417) Hyperactivity 0.154 0.093 0.185 *** 0.131 0.097 0.149 ** (0.361) (0.291) (0.389) (0.338) (0.296) (0.356) Emotional 0.117 0.100 0.125 ** 0.127 0.100 0.141 Symptoms (0.321) (0.300) (0.331) (0.333) (0.300) (0.348) Peer problems 0.082 0.058 0.095 *** 0.070 0.040 0.086 *** (0.275) (0.233) (0.293) (0.255) (0.197) (0.281) Conduct problems 0.062 0.040 0.073 *** 0.071 0.062 0.075 (0.241) (0.197) (0.260) (0.257) (0.242) (0.264) Age 6.188 6.169 6.198 *** 6.198 6.189 6.203 (0.303) (0.283) (0.312) (0.324) (0.304) (0.334) Female 0.483 0.482 0.483 0.503 0.502 0.503 (0.500) (0.500) (0.500) (0.500) (0.501) (0.500) Preschool 3+ 0.796 0.871 0.757 *** 0.799 0.907 0.744 *** Years (0.403) (0.336) (0.429) (0.401) (0.292) (0.437) Single child 0.209 0.193 0.217 * 0 0 0 (0.406) (0.395) (0.412) - - - 3+ children 0.288 0.278 0.293 0.445 0.433 0.451 (0.453) (0.448) (0.455) (0.497) (0.496) (0.498) Observations 4,245 1,439 2,806 947 321 626

Standard deviations in parentheses Differences by parental education group are significant * at 10%; ** at 5%; *** at 1%

23

Table 1 (cont.): Descriptive statistics.

Full sample Sibling sample

Variable

Pooled sample

Parents college

Parents no college ≠ sign. Pooled

sample Parents college

Parents no college ≠ sign.

First child 0.504 0.510 0.501 0.377 0.371 0.380 (0.500) (0.500) (0.500) (0.485) (0.484) (0.486) Second child 0.347 0.359 0.341 0.440 0.470 0.425 (0.476) (0.480) (0.474) (0.497) (0.499) (0.495) Third child 0.1053 0.104 0.106 0.125 0.131 0.1214 (0.307) (0.305) (0.308) (0.330) (0.338) (0.327) Lives with both 0.825 0.892 0.790 *** 0.893 0.956 0.861 *** Parents (0.380) (0.310) (0.407) (0.309) (0.205) (0.346) Parent full time 0.804 0.897 0.756 *** 0.822 0.913 0.775 *** Employed (0.397) (0.305) (0.429) (0.383) (0.283) (0.418) Age of mother 36.010 38.261 34.855 *** 35.895 38.377 34.623 *** (5.166) (4.419) (5.142) (4.794) (4.013) (4.665) Turkish origin 0.080 0.040 0.101 *** 0.111 0.050 0.142 *** (0.272) (0.195) (0.301) (0.314) (0.218) (0.350) Eastern European 0.120 0.084 0.138 *** 0.083 0.053 0.099 ** Origin (0.325) (0.278) (0.345) (0.277) (0.224) (0.299) Other foreigners 0.043 0.021 0.055 *** 0.034 0.003 0.050 *** (0.203) (0.143) (0.227) (0.181) (0.056) (0.217) Observations 4,245 1,439 2,806 947 321 626

Standard deviations in parentheses Differences by parental education group are significant * at 10%; ** at 5%; *** at 1%

24

Table 2: Baseline regression models for the effect of health conditions on cognitive ability (CPM score).

Dependent variable: CPM score for cognitive ability (0-14)

OLS (Full sample) Sibling fixed effects (Sibling sample)

Pooled Sample

Parents college

Parents no college

Pooled Sample

Parents college

Parents no college

(1) (2) (3) (4) (5) (6)

Overweight -0.316*** -0.113 -0.300*** 0.067 0.430 0.067 (0.093) (0.172) (0.108) (0.383) (0.563) (0.437) Underweight -0.069 0.067 -0.197 -0.195 0.616 -0.813 (0.169) (0.212) (0.236) (0.405) (0.579) (0.585) Low birth weight -0.779*** -1.066*** -0.599*** -0.306 -0.052 0.484 (0.178) (0.354) (0.202) (0.408) (0.652) (0.463) Ear condition -0.167 -0.078 -0.252* 0.166 0.236 0.253 (0.103) (0.159) (0.133) (0.251) (0.313) (0.395) Eye condition -0.230*** -0.238* -0.221** 0.173 -0.052 0.237 (0.085) (0.144) (0.105) (0.184) (0.335) (0.244) Asthma -0.493*** 0.083 -0.699*** -0.176 -0.088 -0.825* (0.164) (0.194) (0.209) (0.380) (0.467) (0.475) Allergy 0.079 0.019 0.121 0.015 -0.313 -0.201 (0.155) (0.245) (0.197) (0.455) (0.375) (0.466) Mental condition -1.067*** -0.941*** -1.047*** -1.195*** -0.996*** -1.140*** (0.104) (0.179) (0.125) (0.287) (0.342) (0.393) F-test health variables

p=0.00 p=0.00 p=0.00 p=0.01 p=0.10 p=0.00

Test: OLS=FE p=0.00 p=0.00 p=0.00 - - - Test: Strong exo-geneity assumption

- - - p=0.35 p=0.93 p=0.30

Observations 4,245 1,439 2,806 947 321 626 R-squared 0.12 0.11 0.11 0.11 0.15 0.13 Parentheses show robust standard errors, clustered at household level. In the OLS-regression, all demographic and socioeconomic characteristics listed in Table 1 are included, but not shown. In the sibling fixed effects regression, all demographic and socioeconomic characteristics listed in Table 1 that are not identical for siblings (”age”, ”female”, ”preschool visit 3+ years”, ”first child”, ”second child”, ”third child”, ”lives with both parents”, and ”parent full time employed”) are included, but not shown. * significant at 10%; ** significant at 5%; *** significant at 1%

25

Table 3: Regression models for the effect of health conditions on cognitive ability (CPM score). Child mental health measured with parental questionnaire.

Dependent variable: CPM score for cognitive ability (0-14)


Pooled Sample

Parents college

Parents no college

Pooled Sample

Parents college

Parents no college

(1) (2) (3) (4) (5) (6)

Overweight -0.317*** -0.104 -0.300*** 0.079 0.188 0.138 (0.093) (0.174) (0.108) (0.393) (0.573) (0.462) Underweight -0.135 0.028 -0.288 -0.350 0.511 -0.877* (0.170) (0.210) (0.237) (0.375) (0.589) (0.488) Low birth weight -0.787*** -1.070*** -0.601*** -0.350 -0.387 0.524 (0.173) (0.343) (0.199) (0.401) (0.650) (0.481) Ear condition -0.221** -0.122 -0.316** 0.131 0.009 0.193 (0.103) (0.163) (0.132) (0.248) (0.333) (0.391) Eye condition -0.242*** -0.230* -0.244** 0.134 -0.181 0.197 (0.087) (0.146) (0.108) (0.189) (0.328) (0.250) Asthma -0.503*** 0.153 -0.752*** 0.090 0.188 -0.620 (0.164) (0.198) (0.208) (0.355) (0.464) (0.450) Allergy 0.082 -0.036 0.141 -0.029 -0.672 0.013 (0.162) (0.251) (0.207) (0.412) (0.375) (0.481) Hyperactivity -0.744*** -0.606** -0.734*** -0.687** 0.543 -1.076*** (0.115) (0.247) (0.129) (0.297) (0.526) (0.350) Emot. symptoms -0.244** -0.153 -0.286* -0.604* -0.519 -0.492 (0.124) (0.198) (0.153) (0.337) (0.472) (0.435) Peer problems -0.277* -0.578* -0.133 -0.996* 0.434 -0.958 (0.161) (0.345) (0.182) (0.532) (0.689) (0.656) Conduct problems 0.197 0.131 0.225 0.171 -0.132 0.283 (0.145) (0.285) (0.170) (0.424) (0.498) (0.544) F-test health variables

p=0.00 p=0.03 p=0.00 p=0.17 p=0.66 p=0.03

Test: OLS=FE p=0.00 p=0.00 p=0.00 - - - Test: Strong exogeneity assumption

- - - p=0.80 p=0.40 p=0.79


26

Table 4: Decomposition analysis of the effect of child health conditions on cognitive ability (CPM score). Estimated from sibling fixed effects specifications. Total difference in CPM between ‘parents college’ and ‘parents no college’: 1.057 [100.0%]

Prevalence effect Severity effect Total effect

Panel A: Decomposition based on estimation with pediatrician’s mental health measure.

Mental health condition 0.115 0.027 0.142 (0.046, 0.190) (-0.165, 0.236) (-0.085, 0.358) [10.9%] [2.6%] [13.5%]

Panel B: Decomposition based on estimation with parental SDQ results as mental health measure.

All four mental health variables 0.104 0.277 0.381 (0.019, 0.196) (-0.062, 0.586) (0.016, 0.682) [9.8%] [26.2%] [36.0%] Only hyperactivity 0.036 0.212 0.248 (0.000, 0.089) (0.053, 0.388) (0.062, 0.440) [3.4%] [20.1%] [23.5%] Parentheses include 95% confidence intervals. Square brackets include share of total difference in average cognitive ability by parental education, which can be explained by health effects.

27

Table 5: Linear probability models for the effect of health conditions on verbal ability.

Dependent variable: verbal ability at age level


Pooled Sample

Parents college

Parents no college

Pooled Sample

Parents college

Parents no college

(1) (2) (3) (4) (5) (6)

Overweight -0.023 0.032 -0.031 -0.038 0.055 -0.065 (0.017) (0.031) (0.021) (0.062) (0.143) (0.073) Underweight 0.025 0.032 0.017 0.110 -0.025 0.082 (0.026) (0.031) (0.036) (0.093) (0.109) (0.136) Low birth weight -0.056** -0.047 -0.051 -0.021 0.072 0.021 (0.028) (0.044) (0.035) (0.074) (0.107) (0.089) Ear condition -0.025 -0.020 -0.032 -0.039 -0.061 -0.072 (0.020) (0.028) (0.027) (0.058) (0.081) (0.083) Eye condition -0.007 -0.003 -0.009 -0.028 0.053 -0.073 (0.015) (0.022) (0.020) (0.039) (0.059) (0.055) Asthma -0.086*** -0.043 -0.098*** -0.094 -0.118 -0.120 (0.027) (0.043) (0.034) (0.078) (0.118) (0.098) Allergy 0.003 -0.074 0.039 0.034 -0.011 -0.046 (0.032) (0.057) (0.038) (0.091) (0.217) (0.092) Mental condition -0.017 0.004 -0.018 -0.097** -0.017 -0.100* (0.016) (0.025) (0.020) (0.049) (0.094) (0.059) F-test health variables

p=0.01 p=0.61 p=0.02 p=0.26 p=0.89 p=0.18


- - - p=0.24 p=0.55 p=0.32


28

Table 6: Linear probability models for the effect of health conditions on verbal ability. Child mental health measured with parental questionnaire.

Dependent variable: verbal ability at age level


Pooled Sample

Parents college

Parents no college

Pooled Sample

Parents college

Parents no college

(1) (2) (3) (4) (5) (6)

Overweight -0.025 0.032 -0.032 -0.035 0.055 -0.057 (0.017) (0.031) (0.021) (0.062) (0.151) (0.071) Underweight 0.023 0.030 0.015 0.101 -0.031 0.074 (0.026) (0.031) (0.036) (0.092) (0.110) (0.133) Low birth weight -0.054** -0.048 -0.049 -0.017 0.070 0.031 (0.027) (0.044) (0.034) (0.074) (0.107) (0.092) Ear condition -0.025 -0.021 -0.032 -0.044 -0.073 -0.078 (0.020) (0.028) (0.027) (0.056) (0.082) (0.083) Eye condition -0.006 -0.002 -0.008 -0.027 0.049 -0.073 (0.015) (0.021) (0.020) (0.039) (0.063) (0.056) Asthma -0.085*** -0.043 -0.098*** -0.078 -0.114 -0.109 (0.028) (0.042) (0.034) (0.078) (0.117) (0.099) Allergy 0.003 -0.074 0.038 0.035 -0.019 -0.029 (0.032) (0.056) (0.039) (0.089) (0.215) (0.090) Hyperactivity -0.026* 0.012 -0.031 -0.086* 0.044 -0.101* (0.015) (0.030) (0.022) (0.046) (0.064) (0.058) Emot. symptoms 0.012 0.007 0.013 -0.002 0.003 -0.007 (0.019) (0.028) (0.025) (0.055) (0.049) (0.059) Peer problems -0.023 -0.042 -0.018 -0.088 -0.105 -0.074 (0.025) (0.044) (0.031) (0.078) (0.097) (0.099) Conduct problems -0.028 -0.053 -0.016 0.006 -0.011 -0.006 (0.029) (0.054) (0.034) (0.077) (0.115) (0.112) F-test health variables

p=0.01 p=0.67 p=0.04 p=0.33 p=0.92 p=0.39


- - - p=0.37 p=0.10 p=0.22


29

Table 7: Decomposition analysis of the effect of child health conditions on verbal ability. Estimated from sibling fixed effects specifications. Total difference in verbal ability between ‘parents college’ and ‘parents no college’: 0.108 [100.0%]

Prevalence effect Severity effect Total effect

Panel A: Decomposition based on estimation with pediatrician’s mental health measure.

Mental health condition 0.009 0.016 0.025 (0.000, 0.021) (-0.024, 0.068) (-0.019, 0.078) [8.3%] [14.8%] [23.1%]

Panel B: Decomposition based on estimation with parental SDQ results as mental health measure.

All four mental health variables 0.009 0.018 0.026 (-0.001, 0.020) (-0.021, 0.046) (-0.020, 0.058) [8.3%] [16.7%] [24.0%] Only hyperactivity 0.004 0.019 0.023 (0.000, 0.013) (-0.005, 0.042) (-0.003, .050) [3.7%] [17.6%] [21.3%] Parentheses include 95% confidence intervals. Square brackets include share of total difference in average verbal ability by parental education, which can be explained by health effects.

Figure 1: Cumulative frequency distribution of cognitive ability (CPM score) for the full sample by parental education.

0

20

40

60

80

100

0 1 2 3 4 5 6 7 8 9 10 11 12 13 14

Cum

ulat

ive

freq

uenc

y [in

%]

CPM score

Parents collegeParents no college

30

Documents

The relationship between child health, developmental …jenni.uchicago.edu/Spencer_Conference/Papers 2010/SCHUNK SalmSc… · The relationship between child health, developmental