The causal effect of family income on child health in the U.K.

Journal of Health Economics 36 (2014) 137–150

Contents lists available at ScienceDirect

Journal of Health Economics journal homepage: www.elsevier.com/locate/econbase

The causal effect of family income on child health in the UK夽 Daniel Kuehnle ∗ Department of Economics, Friedrich-Alexander-University Erlangen-Nuremberg, Germany

a r t i c l e

i n f o

Article history: Received 14 March 2013 Received in revised form 20 March 2014 Accepted 26 March 2014 Available online 13 April 2014 JEL classification: I1 Keywords: Child health Income gradient Instrumental variables Transmission channels United Kingdom

a b s t r a c t Recent studies examining the effect of family income on child health have been unable to account for the endogeneity of income. Using data from a British cohort study, we address this gap by exploiting exogenous variation in local labour market characteristics to instrument for family income. We estimate the causal effect of family income on different measures of child health and explore the role of potential transmission mechanisms. We find that income has a very small but significant causal effect on subjective child health and no significant effect on chronic health conditions, apart from respiratory illnesses. Using the panel structure, we show that the timing of income does not matter for young children. Moreover, our results provide further evidence that parental health does not drive a spurious relationship between family income and child health. Our study implies that financial transfers are unlikely to deliver substantial improvements in child health. © 2014 Elsevier B.V. All rights reserved.

1. Introduction The literature on the relationship between family income and child health, the so-called ‘income gradient’, has grown substantially in the last 10 years. Following the seminal contribution by Case et al. (2002), a number of studies have examined the relationship between family income and child health for different countries. These studies universally find a positive association between parent-reported child health and family income in the US, the UK, Canada, Australia, and Germany (Currie and Stabile, 2003; Currie et al., 2007; Propper et al., 2007; Khanam et al., 2009; Reinhold and Jürges, 2012; Kruk, 2012; Apouey and Geoffard, 2013). Examining the income gradient is important for a number of reasons. Several studies show that poor health in childhood correlates

夽 I am grateful for helpful comments and suggestions received by two anonymous referees, David Johnston, Guyonne Kalb, Cain Polidano, Regina T. Riphahn, Stefanie Schurer, Tony Scott, Peter Sivey, and participants at the 2012 European Society for Population Economics meeting, the Brown Bag Series at the Melbourne Institute, and the doctoral workshop at the University of Erlangen-Nuremberg. Data from the Millennium Cohort Study (MCS) were supplied by the ESRC Data Archive. Special acknowledgement is due Stephen P. Jenkins, Lucinda Platt and Jon Johnson for their help in merging in external data. ∗ Correspondence to: University of Erlangen-Nuremberg, Lange Gasse 20, 90403 Nuremberg, Germany. Tel.: +49 0911 5302 230; fax: +49 0911 5302 178. E-mail address: [email protected] http://dx.doi.org/10.1016/j.jhealeco.2014.03.011 0167-6296/© 2014 Elsevier B.V. All rights reserved.

significantly with lower educational attainment, worse health, and lower earnings in adulthood (see, for an overview, Currie, 2009). The strong correlation between child health and adult outcomes is consistent with the theoretical framework on skill development by Cunha and Heckman (2007). The model implies that health capital affects the production of other forms of human capital, such as education or health itself, through self-productivity (e.g., health generated at time t improves health at time t + 1) and dynamic complementarity (e.g., health produced at time t fosters the production of other skills at t + 1). Therefore, public policy may improve the educational, health, and labour market trajectories of children with poor health through early intervention policies that reduce inequalities in child health. Furthermore, policies that reduce the socio-economic gradient in child health may ultimately help mitigate the intergenerational transmission of poverty. Despite the recent contributions to the literature, it remains unclear whether policies that change family income are a costeffective means of improving child health. So far the literature has been unable to establish whether income has a causal effect on child health, or whether the observed relationship between family income and child health is spuriously driven by unobserved factors. The recent literature for the UK is ambiguous and reaches different conclusions as Propper et al. (2007) find that the relationship between family income and child health becomes insignificant once the authors control for maternal health, whereas Kruk (2012)

138

D. Kuehnle / Journal of Health Economics 36 (2014) 137–150

and Apouey and Geoffard (2013) find that income remains significant when controlling for parental health. However, the findings by Propper et al. (2007) suggest that public policy should address maternal health rather than family income to improve child health provided that maternal health does not pick up unobserved maternal preferences that correlate with income. The purpose of this paper is to address this gap by examining the causal effect of family income on different measures of child health in the UK. The study contributes to the literature in several ways. First and most importantly, given the lack of a natural experiment, we explicitly address the endogeneity of income with an instrumental variables (IV) approach. Income endogeneity is a large concern that could arise for three main reasons: as a result of unobserved heterogeneity, measurement error, or reverse causality. Most studies try to overcome the endogeneity problem by controlling for as many variables as possible to ‘mopup’ any observed heterogeneity (Gregg et al., 2005). However, this approach does not address unobserved heterogeneity and ignores both reverse causality and measurement error. In a novel contribution to this literature, we exploit exogenous variation in local labour market characteristics to deal with the endogeneity of income. Second, Currie (2009) points out the need for a closer examination of the transmission mechanisms that translate income into better health. Although it is unlikely that family income has a direct effect on child health, higher incomes may improve child health by relaxing families’ budget constraint on goods and services that enter a child’s production function, such as medical care, nutrition, and housing conditions. As income may systematically correlate with these (and other) unobserved factors that also determine children health, we exploit the rich data and contribute to the literature with a more detailed analysis of potential transmission mechanisms, e.g., nutrition, housing, medication. Third, the availability of panel data allows us to investigate more carefully the timing of family income. We are able to separate out the effect of current versus permanent income and are the first study to examine the effect once we instrument for income. Furthermore, we examine the effect of lagged income on current child health by instrumenting for these lagged variables. Our results show that income has a statistically significant but small causal effect on child health in the UK. A doubling of income reduces the probability of poor or fair child health by about 6 percentage points. We show that models that do not account for the endogeneity of income underestimate the size of the effect. Furthermore, we find small and mostly insignificant effects of family income on indicators of specific chronic health conditions, apart from respiratory illnesses. Using the panel structure, we show that the timing of income does not matter for young children. Our analysis provides additional evidence that parental health does not drive a spurious relationship between child health and family income. We conclude that income is not a major causal determinant of child health in the UK. All results pass a wide range of robustness checks. The paper is organised as follows. In Section 2 we briefly review the relevant literature. Section 3 presents the conceptual framework for the production of child health and discusses the empirical identification strategy. In Section 4 we describe the data and variables used in the analysis. Section 5 presents and discusses the main results. We present a variety of sensitivity tests validating the robustness of our results in Section 6. We conclude and discuss the paper’s policy and general implications in Section 7.

parent-reported child health. Typically, subjective child health is measured on an ordinal scale (where 1 = excellent, 5 = poor) and an ordered probit model is used to estimate the relationship with the explanatory variables. In their seminal study, Case et al. (2002) examine the relationship between household income and subjective child health using a variety of US datasets for children aged 0–17. The study finds a significant positive association between household income and child health that is robust to different measures of child health and income. The results also reveal an increase in the gradient with the age of the child, implying that the effects of income are larger for older children. The methodology of Case et al. (2002) was subsequently replicated and slightly modified by other studies for the US (Currie and Stabile, 2003; Chen et al., 2006; Condliffe and Link, 2008; Murasko, 2008), England (Currie et al., 2007; Propper et al., 2007; Kruk, 2012; Apouey and Geoffard, 2013), Australia (Khanam et al., 2009), and Germany (Reinhold and Jürges, 2012). These studies generally confirm the positive relationship between household income and subjective child health, controlling for a similar, or the same, set of explanatory variables.1 In particular, Currie and Stabile (2003) find a similarly sized income coefficient for Canada and thus provide evidence for a steepening income/health gradient in a country with universal health insurance. This suggests that universal access to health insurance does not reduce the income gradient or prevent the steepening of the gradient. In another US study, Chen et al. (2006) examine the association between income and various measures for child health, and explore whether the relationship changes with the age of the child. Although the authors use the same data as Case et al. (2002), the study finds no increase in the income/health gradient using subjective health. However, Case et al. (2007) show that differences in the selection of the estimation sample can largely explain the differences between the two studies. For the UK, Currie et al. (2007) use English data for children aged 2–15 and confirm the existence of an income gradient in England, although the size of the income coefficient is considerably smaller than in the US. The study finds no association between income and objective measures of health, such as professional examinations or blood test results. Furthermore, the authors show that nutritional intake and parental physical activity are influential transmission mechanisms in determining child health. The study does not find, however, that the gradient increases with the age of the child. In response, Case et al. (2008) use the same data and exploit three additional years of data to show that the income gradient increases with age in the UK as well, although the income gradient remains smaller than in the US. In another study, Propper et al. (2007) use UK data for children aged 0–7 and find that although an income gradient exists, it does not increase with the age of the child and actually disappears when the authors control for the mother’s selfrated health. The study also explores whether maternal behaviours, e.g., nutrition, time spent with child, or smoking, involved in the production of child health correlate with income, but finds that these factors are largely insignificant. Khanam et al. (2009) use Australian data for children aged 0–7 and also find that the relationship disappears once they control for subjective maternal health. In contrast, the studies by Kruk (2012) and Apouey and Geoffard (2013) both use UK data and show that the correlation between family income and child health does not disappear when including controls for parental health. More specifically, Kruk (2012) focuses on the relationship between income and the prevalence, severity, and

2. Previous literature The majority of the economics literature on the income gradient examines the relationship between household income and

1 These include controls for parental education, age of the child, child cohort, family size, child’s sex, child’s ethnicity, presence of mother and father in the household, father’s employment status.


persistence of diseases. Apouey and Geoffard (2013) examine the evolution of the income-gradient and show that it remains stable between the ages of 2 and 17. However, none of these studies explicitly deal with the endogeneity of income. The only study that uses an IV approach is Doyle et al. (2007). However, the study uses ‘grandparental smoking’ as an instrumental variable for family income. The instrument validity relies upon the very strong assumption that grandparental smoking does not correlate with any other unobserved determinants of child health. Furthermore, Case et al. (2002) and Currie et al. (2007) also instrument for income using information about the job characteristics of workers in the household and find that the general results do not change. However, as Currie et al. (2007) concede, these instrumental variables are likely to fail the exclusion restriction required for a valid instrument. 3. Analytical framework and empirical strategy In line with the previous literature we specify a production function of child health that depends on the quantity and quality of inputs. Specifically, we model child health in its reduced form as follows: Hit = ˛Yit + it Hit =

+ Xitc ˇc

+ it

(2)

where we first estimate the raw correlation between the health (H) and family income (Y) of child i at time t as in Eq. (1). We then add p some basic exogenous control variables for parental (Xit ) and child c (Xit ) characteristics to form our base case specification as in Eq. (2). The base case uses a similar set of covariates as Currie et al. (2007) and includes controls for the child’s age, sex and ethnicity, household size (log), the mother’s and father’s age, parental education, and indicator variables if either of the parents work.2 ˇp and ˇc represent vectors of parameters to be estimated and it is a random error. The main econometric challenge in identifying the causal effect of income on child health results from the potential endogeneity of income. Income might be endogenous for three reasons: first, the association between income and child health might reflect a purely spurious association as an unobserved third factor, such as parental health, or time preference rates (Fuchs, 1982), could correlate with both income and health. If this third factor positively correlates with both income and child health (such as parental health), the estimated income coefficient is biased upwards. The third factor could also correlate negatively with income and positively with child health (such as a preference to provide child care at home), imposing a negative bias. Secondly, reverse causality arises if poor child health causes either parent to reduce their labour supply and consequently family income. This would impose an upward bias on the income coefficient. Third, income is likely to be measured with error which may bias the income coefficient towards zero due to attenuation bias. As a result of these opposing signs of bias, we cannot a priori assign a direction of the overall bias. We employ an IV approach to deal with the endogeneity of income. The IV approach exploits variation in income generated by a factor that, holding other things constant, only affects child health through parental income. To apply the two-stage least squares (2SLS) estimator, we need to specify the following first-stage equation for income:

Yit = Xit ıp + Xitc ıc + Zit ız + it

2

A valid instrument Zit needs to correlate with income, i.e. ız = / 0, and must be uncorrelated with the error term from Eq. (2), i.e. cov(Zit , it ) = 0. We employ a set of instruments, local labour market characteristics, that satisfy these conditions and that have been used in slightly different contexts in the literature before (see Ettner, 1996; Xu and Kaestner, 2010). In our empirical approach, we first estimate Eqs. (1) to (2) using both OLS and 2SLS to compare the coefficients. However, note that the interpretation of coefficients differs between the OLS and 2SLS estimators. Whilst OLS produces average marginal effects, 2SLS produces local average treatment effects (LATE, see Imbens and Angrist, 1994). The LATE identifies the effect of variation in income caused by changes in local employment conditions for the group of compliers. In our case, families with volatile or low-quality employment are most likely to be the group of compliers as fluctuations in the business cycle are likely to influence their incomes. In the second part of the analysis, we examine the transmission mechanisms through which family income could influence child health. We derive and classify potential mechanisms from the literature on the determinants of child health in Section 4.5. For the analysis, we estimate the following regressions: p

Hit = ˛Yit + Xit ˇp + Xitc ˇc + TM ijt j + it

(1)

p ˛Yit + Xit ˇp

p

139

(3)

For a detailed list and definition of variables used in the analysis, see Table A.1.

Hit =

p ˛Yit + Xit ˇp

+ Xitc ˇc

+

J

TM ijt j + it

(4)

(5)

j=1

We denote each group j of transmission mechanisms by a vector TM ijt and add it separately to the base case specification as in Eq. (4). This approach allows us to disentangle how much of the income effect is mediated by each group. In the final regression, Eq. (5), we control for all transmission mechanisms simultaneously. However, the interpretation of changes in the income coefficient with the inclusion of the transmission mechanisms differs between the OLS and the 2SLS estimators. Using 2SLS, we condition on the transmission mechanisms in the first stage and thus purge instrumented income of the effect of these mechanisms. This means that only the OLS estimates can be interpreted as transmission mechanisms, whereas the instrumented results provide estimates of the causal effect of income conditional on a particular set of transmission mechanisms. Furthermore, this approach indirectly performs a check for instrument validity: if the 2SLS estimates remain stable with the inclusion of the transmission mechanisms, then coefficient stability implies that the instruments are independent of these observed determinants of child health (Angrist and Pischke, 2009). 3.1. Discussion of potential threats to the IV validity For our main analysis, we instrument family income using local unemployment rates. Given their strong correlation with family incomes, local unemployment rates are particularly relevant instruments for income in the first stage. Most importantly, the identification strategy assumes that the local unemployment rate does not have a direct impact on child health other than through family income. We now discuss factors that may invalidate this assumption and how we deal with these to ensure instrument validity. First, the instrumental variables strategy is threatened if unemployment rates have a direct effect on child health. We argue that unemployment rates do not have a direct first-order effect on child health, e.g., because young children do not participate in the labour force in industrialised countries (Case et al., 2002). We are aware of concerns about this instrumental variables strategy in the context of adult health. For instance, Ettner (1996) instruments income using local unemployment rates in her study on the effect of income

140


on adult health. Kawachi et al. (2010) criticise the validity of the instrument by arguing that local unemployment rates are likely to have an independent effect on adult health. However, when child health is the dependent variable, the concerns raised by Kawachi et al. (2010) do not apply as local unemployment rates do not have a direct effect on child health. Second, given the large literature on the negative relationship between health and unemployment (Schmitz, 2011; Paul and Moser, 2009), we need to address any indirect effect that the unemployment rate may have on child health via parents’ health, physical or mental. For instance, if parents’ health correlates with local unemployment rates and if we do not control for parental health in our regressions, then our exclusion restriction fails to hold. However, we argue and show that this (potential) correlation does not invalidate our results. First, we include numerous controls for parents’ mental, physical, and self-rated health in our estimations. By removing some observed variation in parental health from the error term, we also aim to remove any potential correlation between the instrument and the error term. Although we cannot perfectly control for latent parental health, we can use standard measures for physical and mental health of both the mother and the father. These measures capture many important aspects of latent parental health and are used regularly in the health economics literature. Second, failure to capture latent parental health perfectly (e.g., caused by measurement error in the parental health variables) only poses a problem to our identification strategy if that measurement error systematically correlates with the instrumental variables. Fortunately, we have more than one instrument, so we can directly test whether the instruments correlate with unobserved latent parental health by means of the Sargan test. In Section 5 we show that we never reject the null hypothesis of a Sargan test implying that the instrumental variables do not systematically correlate with the residuals (which contain unobserved parental health) from the instrumental variables regressions. Finally, we perform a placebo-test in Section 6.1 which provides further evidence that local unemployment rates do not correlate with unobserved determinants of child health. Third, unemployment rates might also affect child health through the quality or availability of hospitals. We believe this not to be a problem in the UK as hospital funding does not depend on regional incomes but instead on the centralised National Health Service. Furthermore, our estimations include regional fixed effects that should capture regional differences in hospital infrastructure and quality. Fourth, recent studies highlighted the importance of neighbourhoods as a determinant of health (Bilger and Carrieri, 2013; Jacob et al., 2013). These studies show that social, physical, and environmental factors have a causal effect on adult health and child mortality. If unobserved neighbourhood characteristics have an effect on child health and if neighbourhoods systematically correlate with the instrumental variables, the validity of the instrumental variable is threatened. Unfortunately, the MCS does not provide information on the neighbourhoods of families. However, the data allows us to control for council housing which acts as a proxy for poor neighbourhood quality as council housing estates are typically areas of high deprivation (Atkinson and Kintrea, 2001).3 We are therefore able to control for neighbourhoods that are most likely to have a negative impact on child health. Furthermore, as with unobserved parental health, the results from the Sargan test and the placebo-regressions show that the instrumental

3

In the UK, social or public housing is known as ‘council housing’.

variables do not correlate with the unobserved determinants of child health (which include neighbourhoods). Finally, location (and hence local unemployment rates) may not be exogenous as families may decide to move to regions that suit the parents’ preferences for work (e.g., based on low/high unemployment rates). However, the opportunity costs of moving are generally higher for families with school-aged children, which we examine in our analysis, compared to families without children or with children who have already left school (Long, 1972; Mincer, 1977). The descriptive evidence for our sample confirms our hypothesis about low mobility amongst families with school-aged children.4 In addition to these considerations, we empirically test whether residential sorting biases the estimates by re-estimating the model only for those individuals who never move.5 Based on these considerations, we argue that we can treat local unemployment rates as exogenous to the child health equation. 4. Data and variable definitions We use British data from the Millennium Cohort Study (MCS), a longitudinal cohort study representative of all children born in the UK between 2000 and 2001. The MCS provides a rich collection of information on the child’s development, the socioeconomic context and the family’s circumstances. The survey is a disproportionately stratified clustered sample and provides detailed information capturing the emotional, physical, cognitive, and health-related development of the child. Parents were first asked to subjectively assess their child’s health in wave 3 of the MCS. We therefore use data from waves 3 and 4 (collected mainly in 2006 and 2008) when the children were aged 4–6 and 6–8, respectively, and supplement these with information merged in from earlier waves. Our final estimation sample includes 11,728 children from wave 3 and 10,641 children from wave 4. 4.1. Measuring income We construct several measures of household income that are consistent with definitions used by Case et al. (2002) and Currie et al. (2007). We derive these measures from the questionnaire which asks the respondent to which of 19 income bands the household belongs. The MCS defines net household income as the combined annual income from all sources after deductions (e.g., tax, national insurance). We first convert the banded income variable into a pseudo-continuous measure by assigning the mid-point of each income interval to the respondent. We then convert all prices to 2005 prices using the monthly consumer price index and take the natural logarithm to capture potential income nonlinearities. Average annual net family income in 2006 and 2008 is £ 29,154 and £ 31,610, respectively.6 As our main regressor we use (logged) permanent household income that we calculated as the mean of current income across all available waves. Using permanent income carries a number of advantages. First, averaging reduces measurement error in income. Permanent income also reduces the potential bias exerted by reverse causality as temporary reductions in family income due to poor child health affect permanent income less than current income. Furthermore, Case et al. (2002) hypothesise that families

4 In our study, about 16.2% and 9.8% of families moved at waves 3 and 4, respectively. Of those who moved, the vast majority (89.5%) stayed within the same region. 5 Our estimations produce almost identical results for the entire sample and for the subsample of those who never move. Results are available upon request. 6 We applied sampling weights to these figures to account for the sampling design.


0

20

40

Percentage

40 20

Percentage

60

60

80

141

Poor

Fair

Good

25−50%

M C S M 3 C S4

C S M 3 C S4

M

C S M 3 C S4

M

C S M 3 C S4

M

M

C S M 3 C S4

0

Lowest 25%

Very good

Excellent Good

Excellent

Fig. 1. Distribution of child health status – by wave.

make health investment decisions on the basis of long-run average income rather than current income. We test this hypothesis using measures of both current and permanent income in Section 6.3. As a robustness check, to reflect household composition, we apply the equivalence scales suggested by Gravelle and Sutton (2003) to generate measures of permanent and current equivalised income.7 4.2. Child health For comparability with the previous literature, we use parentreported child health as the main dependent variable. In the MCS, the primary carer rates the child’s health according to the following ordinal scale: 1 = poor, 2 = fair, 3 = good, 4 = very good, 5 = excellent.8 Child health is distributed as shown in Fig. 1 with a mean of 4.3 and standard deviation of 0.85 in wave 3 (4.5 and 0.77 in wave 4, respectively). The graph shows that the majority of children are in excellent or very good health. However, health clearly differs by income groups as seen in Fig. 2 which illustrates that children from poorer families are consistently in worse health than richer children. Given the ordinal ranking of the child health variable and to be consistent with the previous literature, we generate a binary indicator equal to 1 if the child is reported to be in poor or fair health. Only 4% and 2.9% of children are in poor or fair health at waves 3 and 4, respectively. In addition to subjective child health, the MCS provides information on specific conditions that are classified using the International Classification of Diseases (ICD-10). We use the information to generate indicator variables for 13 specific health conditions.9 The most prevalent chronic conditions are respiratory (7.7%) and skin diseases (3.7%)

7 Gravelle and Sutton (2003) define equivalised incomes as follows: equivalised income = √ householdincome .

50−75%

Highest 25%

Very good Fair/poor

Fig. 2. Distribution of child health status – by income quartiles.

4.3. Base case specification The base case specification contains a basic set of controls for family and child characteristics. In terms of family characteristics, we control for the household size (log) and whether the natural father is resident in the household. We include indicator variables for the child’s sex, ethnic background, age, and height. Controls for the mother’s characteristics include age, height, and education.10 For education, we construct four indicator variables for the highest level of educational attainment. Although education may be endogenous with respect to child health, failure to control for education will bias the income coefficient upwards due to the strong correlation between education and income.11 We also control for a set of regional fixed effects. 4.4. Parental health We add controls for parental health to the base case specification for two main reasons. First, parental health is an important direct and indirect determinant of child health (Kelly and Bartley, 2010). The relationship between parental and child health may arise for three main reasons: first, parental health may have a direct effect on child health. Second, the health of parents and their children may be linked through genetics. Third, unobserved factors at the family level may create a spurious association between child and parental health. Thus, parental health should enter a child’s health production function as an explanatory variable that has an independent effect on child health. However, another interpretation is that parental health may be partially determined by income and should therefore be considered a transmission mechanism. Given the ambiguous interpretation and the need to control for parental health for the three reasons cited earlier, we do not classify parental health as a transmission mechanism (although it may partially act as one) and instead add it separately to the base case

(adults+0.5∗children)

8

We reversed the coding of the variable to facilitate the interpretation of coefficients. 9 These specific conditions include endocrine nutritional and metabolic diseases, mental and behavioural disorders, diseases of the nervous system, diseases of the eye and adnexa, diseases of the ear and mastoid process, diseases of the blood and circulatory system, diseases of the respiratory system, diseases of the digestive system, diseases of skin and subcutaneous tissue, diseases of the musculoskeletal system and connective tissue, diseases of the genitourinary system, congenital malformations deformations and chromosomal abnormalities, and any other conditions.

10 We add height of the mother as an additional explanatory variable because adult height has been shown to be a good proxy for the intra-uterine nutritional environment which affects cognitive development and ability, see e.g.,Case and Paxson (2008) and Heineck (2009). 11 However, this does not pose any problems to our analysis of the income effect as most mothers (94.5%) had attained their highest level of education by the time of the first survey. Hence, education is not determined by income and, importantly, does not lead to over-controlling.

142


specification.12 Secondly, we include parental health to investigate whether parental health is itself an unobservable factor that may account for the relationship between family income and child health. The MCS provides a number of standard measures for parental physical and mental health that we use in our regressions. We generate indicator variables for whether either parent currently has a long-standing illness, and whether this long-standing illness is limiting. To capture mental health, we control for whether either parent has ever been diagnosed with depression or serious anxiety by a doctor. We also control for the parents’ body-mass index (BMI). 4.5. Transmission mechanisms 4.5.1. Housing Housing conditions may have a direct effect on child health as several studies show that children living in damp houses have an increased risk of suffering from respiratory illnesses (Peat et al., 2007; Andriessen et al., 1998). We therefore generate an indicator variable whether the house suffers from dampness. To serve as a proxy for quality of the house and the local neighbourhood, we additionally control for whether the family lives in a council house. 4.5.2. Foetal-origins hypothesis We construct variables to address the implications of the foetalorigins hypothesis. This hypothesis argues that maternal behaviour (such as drinking and smoking) and nutritional intake during pregnancy not only have immediate effects in-utero, but generate long-lasting health consequences for the child (Barker, 1998; Ravelli et al., 1998). Economic studies provide strong empirical support for this hypothesis (Almond and Currie, 2011). To address these mechanisms, we control for the child’s birth-weight and whether any problems occurred during childbirth or the first week of the child’s life. We also include an indicator variable for whether the child was breastfed as mothers who breastfeed may have behaved differently during pregnancy, e.g. in terms of child health investments, compared to mothers who do not breastfeed. 4.5.3. Nutrition Consistent with the idea that higher incomes relax budget constraints on the purchase of higher quality foods, leading to better child health, Currie et al. (2007) show that nutrition is an influential transmission mechanism. The questionnaire provides information about whether the child snacks on different types of foods between meals. Factor analysis of these 13 categories reveals that these variables load onto two factors that we interpret as savoury and sweet foods. We therefore generate two factors scores that capture the child’s consumption of savoury or sweet foods between meals. We also control for the amount of fruit consumed daily by the child (ranging from zero to three or more portions). Furthermore, we control for whether the child has breakfast less than five times per week to capture potential nutritional deficiencies.13

12 It is worthwhile pointing out that the relationship between parental health and family income shares many of the ambiguities that relate to the association between child health and income, e.g., unobserved third factors. 13 Whilst we interpret these variables as measures of nutritional aspects and/or quality, another interpretation could be that these variables reflect unobserved parental characteristics such as time preference rates, productivity, or even health. Whilst we cannot separate out these different interpretations, we show in the results section that this ambiguity does not invalidate our IV estimates. Instead, even if we accept the alternative interpretation for the nutrition variables, the results presented in Table 1 support our identification strategy as they imply that

4.5.4. Parental smoking A substantial body of evidence demonstrates the negative effects of parental smoking, both pre- and post-natal, on child health (Cook and Strachan, 1999; Gilliland et al., 2001). We therefore separately control for the number of cigarettes smoked by the mother and partner, and whether anyone smokes when the child is present in the same room. 4.5.5. Medications Family income may allow families to purchase medications that improve their children’s health. In the UK, most patients do not have to pay for prescription drugs as 88% of prescribed items are paid for by the NHS.14 Unfortunately, the MCS does not provide any information on out-of-pocket expenses for medication taken by the child. To capture any potential income effect that is mediated through medical drugs, we include a binary indicator that is equal to 1 if the child regularly takes any medication. 4.5.6. Other pathways Finally, we extend the literature by including proxy-variables that capture a child’s behavioural difficulties. For instance, families with lower incomes might be more likely to have children with behavioural difficulties which are likely to have a negative impact on children’s (mental) health. Indeed, a recent study demonstrates a significant link between the psychosomatic problems and bullying at school (Gini and Pozzoli, 2009). We therefore include indicator variables if the child is being bullied at school; if the child bullies other children at school; and whether the child has irregular bedtimes. 4.6. Instrumental variables To implement our identification strategy, we merge in external information for the local labour market characteristics, which are available at the level of local authorities, provided by the British Office for National Statistics (ONS).15,16 We can exploit a total of 406 local authorities according to the 2011 classification.17 For the analysis we use the following three instruments: the local unemployment rate; information on the number of business start-ups and closures; and the average regional household income. For the main analysis, we employ the local unemployment rate as the main instrumental variable (IV1). Based on the business registrations data, we generate a business growth variable.18 To perform a number of robustness checks in Section 6.1, we generate and use the following sets of additional instrumental variables: 1 IV2: lagged local unemployment rate (by one year); 2 IV3: average local unemployment rate (averaged over available waves);

the instrumental variables are independent of time preference rates or maternal productivity. 14 http://www.politics.co.uk/reference/nhs-prescription-charges, last accessed on 24/10/2013. 15 We downloaded the data from the official website: http://www. nomisweb.co.uk on 06/07/2012. 16 The ONS defines local authorities as ‘[. . .] the lower tier of local government [. . .] including non-metropolitan districts, metropolitan districts, unitary authorities and London boroughs in England; unitary authorities in Wales; council areas in Scotland’. 17 For a map, see: http://www.ons.gov.uk/ons/guide-method/geography/ beginner-s-guide/maps/all-local-authorities-in-the-uk-effective-at-31stdecember-2011.pdf. Last accessed on 24/10/2013. 18 Since the ONS collected the business registrations data only until 2007, we are required to use the 2007 values for the interviews conducted in 2008.


3 IV4: business growth rate and regional GDI (gross domestic income per head); 4 IV5: local unemployment rate, business growth rate, and regional GDI; 5 IV6: average local unemployment rate, business growth rate, and regional GDI. Table A.2 provides full summary statistics for variables used in the analysis. 5. Empirical analyses

143

different proxy variables for parental and child health. For example, Propper et al. (2007) do not use the same dependent variable but instead whether the child belongs to the top quintile by number of symptoms or poor health. The study also differs in the definition of family income as they use information on financial hardship as a measure of low-income.22 Although Khanam et al. (2009) use a similar set of control variables as our study, they use subjective maternal health, which highly correlates with subjective child health, as the proxy for parental health. We further explore whether subjective maternal health explains the difference in Section 5.3. Furthermore, both studies examine a broader age group (0–7 years) compared to our study (4–8 years).23

5.1. Does income have a causal effect? 5.2. Transmission mechanisms We first estimate the causal effect of family income on the probability of a child being in poor or fair health. Table 1 summarises the estimated income coefficients from models with different control variables for the pooled sample of children and for the two separate age groups. First we focus on Panel A and discuss the pooled sample not accounting for the endogeneity of income in column 1. Once we control for the base case variables, the estimated coefficient shows that a doubling of income reduces the probability of a child being in poor/fair health by 1.5 percentage points (pp). Our marginal effects that to do not account for the endogeneity of income are very similar to those presented in Currie et al. (2007).19 Column 2 presents the results when we instrument for income using the local unemployment rate. The results from Panel A show that income has a significant causal effect that is about four times (−6pp) larger than the OLS estimate. The diagnostic tests for the base case specification confirm the relevance of the instruments. The instrumental variables strongly correlate with family income (F = 160.8) and the model passes the underidentification test (p = 0.000).20 As found in other studies that instrument household income (Doyle et al., 2007; Maurin, 2002), the effect of income is greater for the IV estimates although less precise. Given the strength of the instrument and the large difference between the OLS and IV estimates, the comparison implies that previous studies using OLS and banded income variables may suffer from relatively large endogeneity biases. The results from estimating separate models by age of the child are presented in columns 3–6. For the two age groups, our results are consistent with findings by Currie et al. (2007) and Apouey and Geoffard (2013) who find no evidence of an age gradient in income for the UK. Although the point estimates may suggest that income has a stronger impact on the health of younger children, our statistical tests show that the coefficients for the two age groups are never significantly different.21 Including controls for parental health, the OLS estimate remains highly significant and reduces slightly. This finding contrasts with results by Propper et al. (2007) and Khanam et al. (2009) who find no significant association between income and child health with controls for maternal health. Instead, and consistent with Kruk (2012) and Apouey and Geoffard (2013), our OLS results show that parental health does not drive the relationship between child health and family income. Many factors might explain the different results between our studies, including

19 For the probability of being in the bottom three categories, their study reports a decline in the probability of about 3 pp. We find a decline of 4 pp, see Table A.4. 20 Table A.3 provides results from diagnostic tests such as the first-stage F statistics and the p-value from an underidentification test. Under the null hypothesis of the underidentification test, the model is underidentified. A rejection of the Null thus implies that the instrumental variable is relevant and correlated with the endogenous variable. 21 We estimated the model for the pooled data and performed a t-test on the interaction term between income and age. Results are available upon request.

Panel B in Table 1 presents the results once we control separately, and jointly, for the impact of each set of transmission mechanisms in addition to the base case variables. The regressors included in each additional group are listed in Table A.1. All instrumented models pass the first-stage F-test and the underidentification test (see Table A.3 for details). As shown in column 1, the addition of transmission mechanisms reduces the coefficient on income only marginally. When we control for all transmission mechanisms jointly, the income coefficient (−0.012) remains highly significant. Our results demonstrate that these transmission mechanisms, in particular housing, medications, and ‘other pathways’ are important determinants of child health but still do not fully explain the income gradient. Column 2 presents the results once we instrument for income using the local unemployment rate. The key finding is that adding the transmission mechanisms to the base case specification hardly affects the size and statistical significance of the income coefficient. The robustness of the income coefficient confirms the independence of the instrumental variables with respect to the control variables. The point estimate for income obtained from the most extensive (−0.056) and the base case specification (−0.060) are almost identical. The results demonstrate that income has small a causal effect on child health where a doubling of income reduces the probability of a child being in poor health by about 6 pp. Columns 3–6 in Panel B present the results for regressions by age of the child. The findings from the discussion of the pooled sample largely apply to these subgroups as well. Consistent with Apouey and Geoffard (2013), we find that the correlation between income and child health is stable over time. Hence, we pool the age groups for the rest of the paper. 5.3. What are we missing? Extended models of child health. We now explore whether three additional factors reduce the income effect: the presence of a chronic condition in the previous period; the presence of a current chronic condition; and the mother’s subjective health.24 Although these variables are

22 However, the authors mention that the results are robust to using information from a banded income variable provided by the survey. 23 We also explored the robustness of the results with respect to unobserved heterogeneity by estimating a variety of random-effects models. Both the GLS and IV (using the same instrument as in Table 1) random-effects models, that we specified with and without group means of time-varying covariates (Mundlak, 1978), produce almost identical results and are available upon request. 24 Note that if we interpret ‘medications’ alternatively as ‘condition severity’ that is not captured by reported health status, then we should include medications in this section of the analysis. The data does not allow us to distinguish between these two interpretations, which ultimately only affect the interpretation of the medications variable and not the validity of our approach.

144


Table 1 The effect of permanent income on the probability of poor or fair child health. Pooled sample

Controls Panel A: Raw income Base case (BC) BC + parental health Panel B: Base case plus controls for Housing Foetal origins Nutrition Parental smoking Medications Other pathways All Panel C: Base case plus controls for Lagged chronic condition Child chronic condition Subjective maternal health All N

Ages 4–6

Ages 6–8

OLS (1)

IV (2)

OLS (3)

IV (4)

OLS (5)

IV (6)

−0.030** (0.002) −0.019** (0.003) −0.016** (0.003)

−0.055** (0.008) −0.060† (0.032) −0.061† (0.031)

−0.032** (0.003) −0.022** (0.005) −0.019** (0.005)

−0.053** (0.011) −0.085* (0.041) −0.085* (0.041)

−0.027** (0.003) −0.015** (0.004) −0.012** (0.004)

−0.044** (0.009) −0.034 (0.039) −0.034 (0.039)

−0.017** (0.003) −0.018** (0.003) −0.018** (0.003) −0.018** (0.003) −0.017** (0.003) −0.016** (0.003) −0.012** (0.003)

−0.059† (0.033) −0.062† (0.032) −0.059† (0.032) −0.060† (0.032) −0.056† (0.030) −0.058† (0.032) −0.056† (0.032)

−0.021** (0.005) −0.022** (0.005) −0.022** (0.005) −0.022** (0.005) −0.020** (0.004) −0.019** (0.005) −0.015** (0.005)

−0.086* (0.042) −0.088* (0.041) −0.085* (0.041) −0.086* (0.041) −0.078* (0.040) −0.083* (0.041) −0.076† (0.041)

−0.013** (0.004) −0.014** (0.004) −0.014** (0.004) −0.014** (0.004) −0.014** (0.004) −0.013** (0.004) −0.009* (0.004)

−0.032 (0.041) −0.035 (0.040) −0.032 (0.039) −0.034 (0.039) −0.033 (0.038) −0.033 (0.039) −0.032 (0.040)

−0.017** (0.003) −0.016** (0.003) −0.013** (0.003) −0.010** (0.003)

−0.059† (0.031) −0.053† (0.029) −0.058† (0.031) −0.049 (0.031)

−0.022** (0.005) −0.018** (0.004) −0.016** (0.005) −0.011* (0.004)

−0.087* (0.041) −0.073† (0.039) −0.087* (0.041) −0.067† (0.040)

−0.013** (0.004) −0.013** (0.004) −0.010* (0.004) −0.008* (0.004)

−0.030 (0.038) −0.029 (0.036) −0.029 (0.039) −0.024 (0.039)

22,369

22,369

11,728

11,728

10,641

10,641

Notes: Dependent variable: if child is in poor or fair health. Each coefficient is obtained from a separate regression of a linear probability model. Variables included in each specification are listed in Table A.1. IV uses local unemployment rate as instrumental variable. Cluster-robust standard errors in parentheses. ** p < 0.01, * p < 0.05, † p < 0.1. First-stage F values and underidentification test statistics for each regression are listed in Table A.3.

endogenous with respect to child health, we seek to examine whether the inclusion affects the estimated income coefficient. First, we want to rule out the possibility that the mother’s current assessment of her child’s health depends on the child’s history of chronic illnesses. Second, the evaluation of subjective health may differ conditional on whether the child currently has a chronic illness. For instance, whether a child currently suffers from asthma is likely to affect a mother’s evaluation of her child’s health. Finally, a mother’s assessment of her child’s health may reflect her own subjective health status (or the other way around due to reverse causality). Results appear in Panel C of Table 1. We first discuss the OLS results for the pooled sample. With the addition of the lagged chronic condition, the OLS coefficient reduces slightly from −0.019 to −0.017 and remains highly significant. Next, we control for the presence of a current chronic condition. The income coefficient reduces to −0.016 and remains significant. If the presence of a chronic condition fully explained a child’s health status, the income coefficient should reduce to nearzero. The results thus imply that the subjective health assessment captures more information about a child’s health apart from the presence of a chronic condition. Finally, we control for subjective maternal health. The interpretation of this variable is ambiguous as it captures a mix of mental and physical health, chronic and non-chronic conditions. Further, it could also reflect variations in how parents base their ratings standards for their child’s health. The addition of this single variable results in the largest reduction of the income coefficient to −0.013, which is highly significant.

Controlling for all variables used so far jointly, income still is significantly associated with child health. Hence, we do not find that maternal health drives a spurious relationship between child health and family income. Next we discuss the instrumented results. As before, the instrumented income coefficients are larger than the OLS coefficients. Second, controlling for each set of regressors individually, the instrumented income coefficients remain significant. However, the coefficients are not significantly different from each other and the key message remains the same – income has a small but significant causal effect on child health even when conditioning on these additional health covariates. The combined evidence from our OLS and IV results show that income has a causal effect on child health that is not driven by one of these three factors. 5.4. Effect on chronic health conditions Having examined the effect of income on subjective child health, we now analyse the effect of income on specific chronic health conditions classified according to the ICD-10. If family income only has a small effect on subjective child health, we expect to see a small, if any, effect of income on the incidence of having a chronic condition. Although the ICD-10 classification allows a detailed breakdown of diseases into different conditions, the proportion of children affected by these conditions is often fairly low (see first column of Table 2). We follow the same methodology as earlier and present the results in Table 2.


145

Table 2 The effect of permanent income on the incidence of specific chronic health conditions. %withcondition

Base case

Endocrine

0.38

Mental

1.04

Nervous

0.58

Eye

1.03

Ear

1.77

Blood

0.42

Respiratory

7.77

Digestive

0.94

Skin

3.73

Musculoskeletal

0.61

Genitourinary

0.56

Congenital

0.90

Any other condition

1.81

BC + parental health

All jointly

OLS

IV

OLS

IV

OLS

IV

−0.001 (0.001) 0.002 (0.002) 0.005** (0.002) 0.001 (0.002) −0.002 (0.002) −0.001 (0.001) −0.023** (0.005) 0.000 (0.002) −0.003 (0.003) −0.001 (0.001) −0.001 (0.001) 0.001 (0.002) −0.001 (0.002)

−0.011 (0.013) 0.006 (0.016) 0.013 (0.015) −0.009 (0.015) 0.000 (0.022) 0.004 (0.009) −0.042 (0.045) 0.018 (0.014) 0.010 (0.031) 0.015 (0.011) −0.003 (0.012) −0.014 (0.016) 0.017 (0.021)

0.000 (0.001) 0.003 (0.002) 0.005** (0.002) 0.001 (0.002) −0.001 (0.002) −0.001 (0.001) −0.018** (0.005) 0.001 (0.002) −0.001 (0.003) −0.001 (0.001) −0.001 (0.001) 0.001 (0.002) 0.000 (0.002)

−0.012 (0.013) 0.006 (0.016) 0.013 (0.015) −0.011 (0.015) 0.000 (0.022) 0.005 (0.009) −0.049 (0.045) 0.018 (0.014) 0.007 (0.031) 0.015 (0.011) −0.003 (0.012) −0.015 (0.016) 0.015 (0.021)

0.000 (0.001) 0.004* (0.002) 0.005** (0.002) 0.001 (0.002) −0.001 (0.002) −0.001 (0.001) −0.013** (0.004) 0.001 (0.002) −0.001 (0.003) 0.000 (0.001) 0.000 (0.001) 0.001 (0.002) 0.001 (0.002)

−0.012 (0.014) 0.009 (0.016) 0.013 (0.015) −0.013 (0.016) −0.001 (0.023) 0.005 (0.010) −0.036 (0.041) 0.019 (0.015) 0.010 (0.031) 0.016 (0.012) −0.002 (0.012) −0.019 (0.017) 0.015 (0.022)

N − 22, 369 Notes: Each coefficient represents a separate regression. Variables included in each specification are listed in Table A.1. IV uses local unemployment rate as instrumental variable. Cluster-robust standard errors in parentheses. ** p < 0.01, * p < 0.05, † p < 0.1.

The OLS results in Table 2 show that income significantly correlates with a child having a respiratory illness given the base case specification. A doubling of family income reduces the probability of a child having a respiratory condition by 2.3 pp. Apart from respiratory illnesses, income correlates significantly only with the prevalence of diseases of the nervous system, but the coefficient is very small (0.005). Once we control for parental health, the association between family income and respiratory illnesses decreases but is still significant (−0.018). Once we control for all transmission mechanisms jointly, income is no longer significantly associated with any chronic health conditions apart from respiratory illnesses (−0.013). The results show that the negative effect on respiratory

Table 3 Instrument validity – joint significance of instruments.

IV1 IV2 IV3 IV4 IV5 IV6

Birth weight

Gestational age

Problems during/after birth

0.857 0.867 0.870 0.308 0.496 0.492

0.863 0.950 0.912 0.262 0.436 0.437

0.335 0.437 0.394 0.555 0.631 0.647

N − 22, 369 Notes: This table reports p-values on the joint significance of instruments from separate regressions of child health indicators on the base case variables (see Table A.1 for details), excluding family income, and the instruments using the pooled sample. Birth weight is measured in kilograms, gestational age in days, and problems during birth or in the first week of the child’s life is captured with a binary indicator variable. Cluster-robust standard errors were used in all regressions. Instruments used: IV1: local unemployment rate. IV2: lagged local unemployment rate (t − 1). IV3: average local unemployment rate. IV4: business growth rate and regional GDI (gross domestic income per head). IV5: local unemployment rate, business growth rate, and regional GDI. IV6: average local unemployment rate, business growth rate, and regional GDI.

diseases is partially explained by the transmission mechanisms examined, e.g., parental health, housing, nutrition. We confirm these findings once we instrument for income. Given the base case specification, the effect of income on most specific health conditions is zero and insignificant, apart from respiratory illnesses. A doubling of income reduces the probability of respiratory illnesses by 4.2 pp, but the coefficient is insignificant. That we do not find a significant effect on the incidence of a longstanding condition is consistent with our previous section and the results by Currie et al. (2007) who find no association between family income and objective measures of child health. However, a common problem with the IV estimator relates to an increase in standard errors, especially in small samples. We therefore take the cautious interpretation that income may have a negative effect on the occurrence of respiratory diseases, but no effect on any of the other specific chronic conditions.

6. Robustness checks 6.1. Instrument validity and robustness Our identification strategy rests on the fundamental assumption that the instrument does not correlate with the unobserved determinants of child health. We have already provided strong evidence against this threat to identification by successively adding the transmission mechanisms to the base case specification in Table 1. The small changes in the estimated income coefficient provide strong evidence in support of the argument that the local unemployment rates do not systematically correlate with unobserved determinants of child health. We also perform an instrument placebo-test, that we adapt from Frijters et al. (2009), to evaluate whether the instrument correlates with the unobserved determinants of child health. Suppose our IV was invalid because it strongly correlates with an unobserved

146


Table 4 Sensitivity analysis – using different instruments (2SLS) with permanent income. Pooled sample. Controls

IV1

IV2

IV3

IV4

IV5

IV6

Raw income

−0.055** (0.008) −0.060† (0.032) −0.061† (0.031)

−0.050** (0.008) −0.059† (0.030) −0.061* (0.030)

−0.051** (0.008) −0.060* (0.030) −0.062* (0.030)

−0.045** (0.016) −0.060 (0.039) −0.058 (0.039)

−0.055** (0.007) −0.061* (0.027) −0.061* (0.027)

−0.051** (0.007) −0.060* (0.026) −0.060* (0.026)

160.480 0.000 166.472 0.000

161.501 0.000 167.689 0.000

87.985 0.000 44.865 0.039 0.844

209.758 0.000 73.392 0.039 0.981

219.319 0.000 76.671 0.039 0.981

−0.059† (0.033) −0.062† (0.032) −0.059† (0.032) −0.060† (0.032) −0.056† (0.030) −0.058† (0.032)

−0.059† (0.032) −0.062* (0.031) −0.058† (0.031) −0.060* (0.030) −0.055† (0.029) −0.057† (0.030)

−0.060† (0.032) −0.062* (0.030) −0.059† (0.031) −0.061* (0.030) −0.056† (0.029) −0.059† (0.030)

−0.059 (0.038) −0.064 (0.039) −0.060 (0.039) −0.060 (0.039) −0.046 (0.038) −0.059 (0.039)

−0.060* (0.028) −0.063* (0.027) −0.060* (0.027) −0.061* (0.027) −0.053* (0.026) −0.059* (0.027)

−0.059* (0.027) −0.063* (0.026) −0.059* (0.026) −0.060* (0.026) −0.053* (0.025) −0.058* (0.026)

All

−0.056† (0.032)

−0.056† (0.031)

−0.057† (0.031)

−0.046 (0.037)

−0.053* (0.027)

−0.053* (0.026)

N

22,369

22,369

22,369

22,326

22,326

22,326

Base case (BC) BC + parental health

Diagnostic tests for base case specification: Underidentification statistic 154.978 0.000 Underidentification p-value First-stage F 160.868 Sargan statistic 0.000 Sargan p-value Base case plus controls for Housing Foetal origins Nutrition Parental smoking Medications Other pathways

Notes: Each coefficient represents a separate regression. Variables included in each specification are listed in Table A.1. Underidentification test = Kleibergen-Paap test; First-stage F refers to F test of excluded instruments in the first stage. Cluster-robust standard errors in parentheses. ** p < 0.01, * p < 0.05, † p < 0.1. Instruments used: IV1: local unemployment rate. IV2: lagged local unemployment rate (t − 1). IV3: average local unemployment rate. IV4: business growth rate and regional GDI (gross domestic income per head). IV5: local unemployment rate, business growth rate, and regional GDI. IV6: average local unemployment rate, business growth rate, and regional GDI.

factor, for example, maternal ability. Then the IV is likely to significantly correlate with earlier measures of child health, such as a child’s birth weight, gestational age, or whether any problems occurred during child birth or the first week of a child’s life. To examine this possibility, we run a placebo-test by regressing these measures for early child health on the base case variables, excluding family income, and all sets of instrumental variables (described in Section 4.6). The instruments should be significant if they systematically correlate with the unobserved determinants of child health. Table 3 presents the p-values from the F-tests for joint significance of the local labour market characteristics from the placebo-regressions. The p-values show that none of the instruments significantly correlate with any early measures of child health. These results provide further evidence for instrument validity by demonstrating that the instrumental variables, i.e. local labour market characteristics, do not correlate with unobserved determinants of child health. Furthermore, we examine the robustness of the results to using different sets of instruments. In addition to the current unemployment rate (IV1), we use all potential sets of instruments. All of these instruments should correlate strongly with family income satisfying instrument relevance. For these instrumental variables we apply the same reasons for instrument validity as in Section 3.1. For each set of instruments, we perform a variety of diagnostic tests. Whenever the first stage is overidentified, we run the Sargan test to assess the validity of the instruments. Under the null hypothesis of the test, the instruments do not correlate with the unobserved determinants of child health. A rejection of the null hypothesis thus implies that the instruments are invalid. The results are presented in Table 4. Reassuringly, the newly estimated coefficients are highly robust to the choice of instruments. All instruments correlate strongly

with income as indicated by the first-stage F-statistics. Furthermore, all overidentified models pass the Sargan test providing further evidence that the instruments do not correlate with the unobserved determinants of child health. The estimated income coefficients for the base case specification range from −0.059 to −0.061 and are all significantly different from zero (apart from IV4). Not surprisingly, the size of the estimated income coefficient differs slightly between the specifications, as each specification uses different instruments and identifies different income variations. Yet, despite these differences, the general pattern and magnitude of estimated coefficients does not change with the inclusion of the transmission mechanisms. The findings for alternative instruments clearly confirm the earlier finding that income has a small but significant causal effect on child health.25 6.2. Reverse causality In addition to measurement error and unobserved heterogeneity, reverse causality may cause income endogeneity. This arises if parents have lower earnings because they reduce their labour supply to look after their ill child. Although we partially reduce this potential bias by using permanent income, we follow Currie et al. (2007) and re-estimate the model for permanent income after excluding all children with a limiting long-standing

25 As an additional check, we re-estimated the overidentified models by limited information maximum likelihood (LIML) because 2SLS is not an unbiased estimator. Asymptotically, LIML has the same distribution as 2SLS, but LIML is medianunbiased and less biased in finite samples (Davidson and MacKinnon, 2004). The results clearly show that estimated coefficients obtained by LIML are almost identical to the coefficients from Table 4, confirming the robustness of the 2SLS results for the overidentified models. The results are provided in Table A.6.


147

Table 5 Sensitivity: using different definitions of income. Current income OLS

Equivalised income IV

Current income IV (4)

OLS (5)

IV (6)

−0.054** (0.008) −0.069† (0.037) −0.070† (0.036)

−0.032** (0.002) −0.019** (0.003) −0.016** (0.003)

−0.054** (0.008) −0.059† (0.031) −0.061† (0.031)

Controls

(1)

(2)

OLS (3)

No controls

−0.022** (0.002) −0.010** (0.003) −0.009** (0.003)

−0.054** (0.008) −0.070† (0.037) −0.071† (0.037)

−0.024** (0.002) −0.010** (0.003) −0.009** (0.003)

Base case (BC) BC + parental health Underidentification statistic Underidentification p-value First-stage F Base case plus controls for Housing

84.004 0.000 85.295

Permanent income

85.962 0.000 87.328

158.074 0.000 164.339

−0.009** (0.003) −0.010** (0.003) −0.010** (0.003) −0.010** (0.003) −0.009** (0.003) −0.009** (0.003)

−0.069† (0.039) −0.072† (0.037) −0.069† (0.038) −0.071† (0.037) −0.065† (0.036) −0.068† (0.037)

−0.009** (0.003) −0.010** (0.003) −0.010** (0.003) −0.010** (0.003) −0.009** (0.003) −0.009** (0.003)

−0.068† (0.038) −0.071† (0.037) −0.068† (0.037) −0.070† (0.037) −0.064† (0.035) −0.067† (0.037)

−0.017** (0.003) −0.018** (0.003) −0.019** (0.003) −0.018** (0.003) −0.017** (0.003) −0.016** (0.003)

−0.059† (0.033) −0.061† (0.032) −0.058† (0.032) −0.060† (0.031) −0.055† (0.030) −0.058† (0.031)

All

−0.007** (0.003)

−0.064† (0.037)

−0.007* (0.003)

−0.063† (0.036)

−0.012** (0.003)

−0.055† (0.032)

N

22,369

22,369

22,369

22,369

22,369

22,369

Foetal origins Nutrition Parental smoking Medications Other pathways

Notes: Each coefficient represents a separate regression. Variables included in each specification are listed in Table A.1. Underidentification test = Kleibergen-Paap test; Firststage F refers to F test of excluded instruments in the first stage. IV uses local unemployment rate as instrumental variable. Cluster-robust standard errors in parentheses. ** p < 0.01, * p < 0.05, † p < 0.1.

condition. This approach assumes that the labour supply of mothers will not be affected by the health of children that are not limited by any chronic conditions. The results from the ordered probit regressions presented in Table A.5 produce very similar estimates between the two samples. Although the coefficients change slightly after dropping the children, they are not statistically different from the full sample. Consistent with Kruk (2012), we conclude that reverse causality may lead to a small upward bias which, however, does not invalidate our main results. 6.3. Permanent or current income? So far we examined the effect of permanent income on child health, but what if parents base their health investments on current rather than permanent income? Conversely, if families base their investments on long-run income, as discussed in Section 4.1, then the timing of income should not be important and the coefficients on current and permanent income on current child health should not be different. Separately examining the effect of current versus permanent income allows us to answer this question. We therefore repeat the previous analysis for the pooled sample using current instead of permanent income. The results are presented in Table 5. We first discuss the OLS results in the first column. First, the estimated coefficients are smaller than those using permanent income (see Table 1), probably due to measurement error although we cannot be certain. Given the base case specification, a doubling of current income results in a 1 pp reduction in the probability that a child is in poor or fair health. That the estimate for current income is not statistically different from permanent income (−1.9 pp) supports the hypothesis that the timing of income does not matter in the UK. Second, adding the transmission mechanisms to

the base case, we observe the same pattern as with permanent income – the income coefficient decreases, but remains significant in all specifications. Even with all controls in place, a doubling of current income significantly reduces the probability of poor or fair child health by 0.7 pp. These results show that our general conclusions for permanent income are not driven by the definition of income. Column 2 presents the results instrumenting for income. The diagnostic tests show that the instrument correlates strongly with current income (F = 85.3). Again, the instrumented coefficients are larger than the coefficients obtained by OLS. For the base case specification, a doubling of income reduces the probability poor or fair health by 7 pp. This coefficient is not statistically significantly different from the instrumented coefficient for permanent income (6 pp). Our results are consistent with the idea that long-run average income rather than current income determines child health. The finding holds when we add controls for the transmission mechanisms. Furthermore, the IV estimate for current income is robust to the inclusion of the transmission mechanisms. Controlling for all covariates jointly, a doubling of current income reduces the probability of poor or fair health by 6.4 pp. For permanent income, the effect of the same specification is 5.6 pp. Our results, consistent with Apouey and Geoffard (2013), show that the timing of income appears not to matter in the UK. As a further robustness check, columns 3–6 demonstrate that these results are robust to the equivalisation of income, i.e. whether we take into account household size and composition. The coefficients for equivalised and non-equivalised measures of income are almost identical. Furthermore, the panel structure of the MCS allows us to investigate the issue from a different angle by examining the effects of past income on the current probability of poor child health. For

148


Table 6 The effect of income at different ages of the child. Panel A: OLS results Ages 4–6 (Log) current income at

(1)

Survey 1 (Age 9 months)

−0.015** (0.004)

Ages 6–8 (2)

(3)

(4)

(5) −0.009** (0.003)

−0.010** (0.004)

−0.010* (0.004) −0.001 (0.004) −0.007 (0.005)

−0.006* (0.003)

Survey 2 (Ages 2–4) Survey 3 (Ages 4–6)

(6)

(7)

10,882

9801

12,092

8908

(9)

−0.009* (0.004)

−0.001 (0.004) −0.008* (0.004) −0.002 (0.004) −0.006 (0.006)

10,895

7489

−0.009** (0.003) −0.009** (0.003)

Survey 4 (Ages 6–8) N

(8)

9869

8939

9678

Panel B: 2SLS results Ages 4–6 (Log) Current income at

(1)

Survey 1 (Age 9 months)

−0.025 (0.030)

Ages 6–8 (2)

(3)

(4)

−0.085 (0.052)

Survey 2 (Ages 2–4)

(5)

(6)

(7)

−0.014 (0.025) −0.018 (0.039) −0.128* (0.055)

Survey 3 (Ages 4–6)

−0.050 (0.044) −0.062 (0.046)

Survey 4 (Ages 6–8) N

10,882

9800

12,091

9869

8938

9677

10894

Underidentification statistic Underidentification p-value First-stage F

103.958 0.000 106.630

33.761 0.000 34.062

54.910 0.000 55.742

99.637 0.000 102.716

33.752 0.000 34.209

42.707 0.000 43.346

50.619 0.000 50.759

Notes: Each coefficient represents a separate regression. In addition to income, each regression controls for the covariates from the base case specification (see Table A.1). In Panel B, we use the local unemployment rate as the instrumental variable. Underidentification test = Kleibergen-Paap test; First-stage F refers to F test of excluded instruments in the first stage. Cluster-robust standard errors in parentheses. ** p < 0.01, * p < 0.05, † p < 0.1.

instance, for children aged 4–6 (6–8) we can exploit two (three) lagged values of income. In our empirical approach, we estimate LPMs for the probability of poor child health (separately by age group) using covariates from the base case specification and add controls for (current) income at different points in the child’s life. The results are shown in Table 6. Panel A presents the results for the OLS estimations, Panel B for the IV estimations. The OLS results in panel A show that lagged forms of income significantly correlate with the probability of poor child health. The coefficients on current income and previous income are, however, never significantly different from each other. This indicates a similar level of importance of income at different points in time. For instance, as seen in Panel A, the effect of income at 9 months (survey 1) on current health for children aged 4–6 is −0.015, which is very similar to the effect of current income (−0.010). The same finding applies to the instrumented results, although the coefficients are again larger in size and estimated less precisely. We confirm this pattern for older children, where the estimated coefficients on lagged income (waves 1–3) are all identical (−0.009). Given that the effect of current income is not different from the effect of permanent income, these findings lend further support to the hypothesis that both permanent (i.e. long-run) and current income have a similar effect on child health in the UK. 7. Conclusion The present study examines the causal effect of family income on child health in the UK using an instrumental variables approach. We exploit exogenous variation in local labour market characteristics to instrument for income. The study extends the previous

literature that has universally confirmed a positive association between parent-reported child health and family income. However, whether income has a causal effect on child health or whether the relationship is driven spuriously by unobserved variables is still unclear. In this study, we employ different sets of instrumental variables to identify the causal effect of family income on different measures of child health. Furthermore, we exploit the rich socioeconomic data from the Millennium Cohort Study to examine the transmission mechanisms through which income could affect child health. The results provide novel evidence that income has a significant but very small causal effect on subjective child health in the UK. A doubling family income reduces the probability of poor or fair child health by about 6 percentage points. The results from various specifications show that the difference between the OLS and IV estimates imply fairly large endogeneity biases for the OLS estimates. Furthermore, and consistent with other studies for the UK, we do not find evidence of a steepening income gradient with the age of the child. We also examine the transmission mechanisms through which family income could influence child health, e.g., housing, nutrition, medications, or parental smoking. We separately add each group of transmission mechanisms to the base case specification to identify how much of the income effect is mediated via this group. We find that housing, nutrition, medications, parental smoking, and other pathways (behavioural problems of the child), are all influential transmission mechanisms. In contrast to Propper et al. (2007) and Khanam et al. (2009), but consistent with Kruk (2012), we do not find that the relationship becomes insignificant once we control for parental health. Furthermore, income remains


significantly associated with subjective child health with the addition of all transmission mechanisms. These results are confirmed by the IV estimates. Adding potential transmission mechanisms hardly changes the instrumented income coefficient. This provides an implicit robustness test for the exogeneity of the instrument with regard to these control variables. The study also investigates the effect of family income on a set of specific chronic health conditions, e.g., respiratory diseases, skin diseases, or mental conditions. We find that income correlates significantly only with the occurrence of respiratory problems, and the effect is larger but insignificant when using the IV estimator. This is consistent with previous research (Currie et al., 2007) who have provided evidence that income does not correlate with objective measures of child health. However, due to the larger standard errors associated with an IV estimator, we take the cautious interpretation that income may have a negative effect on the occurrence of respiratory diseases, but no effect on any of the other specific chronic conditions. The negative effect on respiratory diseases is partially explained by the transmission mechanisms examined, e.g., parental health, housing, parental smoking, and nutrition. Furthermore, exploiting the panel structure and information on both current and permanent income, we conclude that the timing of income does not matter in the UK. This is consistent with the hypothesis that families base their investment decisions on longrun rather than current income. We show that when we instrument for income, the coefficients on permanent and current income become almost identical. Clearly, the validity of our results depend on the validity of the instrumental variables. We provide a number of robustness checks that provide strong empirical support for the identifying assumption that local labour market characteristics do not correlate with the unobserved determinants of child health, conditional on the base case variables. Furthermore, the causal effect that we identify is a local average treatment effect. Whilst we cannot characterise the group of compliers whose family income is changed by the instrument, we can hypothesise that changes in local unemployment rates are most likely to affect families with volatile or low-quality employment that is sensitive to changes in the business cycle. The effect we identify may therefore be of particular policy interest as low-quality employment also correlates with lower incomes and worse child health. Our study implies that financial transfers to households are unlikely to improve child health meaningfully. For instance, we estimate that a transfer on the order of 10–30% of net household income reduces the probability of poor child health by about 0.6–1.8 percentage points. We therefore argue that financial transfers are unlikely to deliver substantial improvements in child health. To aid the formulation of appropriate policies that are likely to deliver improvements in child health, future research should address other potential means of improving child health, e.g., parental health, housing, quality of medical care, or nutrition. Studies that exploit exogenous variation in any of these factors, especially parental mental health, would substantially enhance our understanding of the causal determinants of child health and assist the design of appropriate intervention policies. Methodologically, our identification strategy may prove useful in other research on child development. In particular, using local labour market characteristics as instruments for parental employment-related characteristics may allow the identification of other causal effects of interest. Given the substantial bias we found for the OLS estimates, future research in the context of child health should explicitly address the endogeneity of income and could use our identification strategy to asses whether previous conclusions about the ‘prevalence’ and ‘severity’ effects of income hold once income endogeneity is accounted for.

149

Appendix A. Supplementary Data Supplementary data associated with this article can be found, in the online version, at http://dx.doi.org/10.1016/ j.jhealeco.2014.03.011. References Almond, D., Currie, J., 2011. Killing me softly: the fetal origins hypothesis. J. Econ. Perspect. 25 (3), 153–172. Andriessen, J., Brunekreef, B., Roemer, W., et al., 1998. Home dampness and respiratory health status in European children. Clin. Exp. Allergy 28 (10), 1191–1200. Angrist, J., Pischke, J., 2009. Mostly Harmless Econometrics. Princeton University Press, Princeton, NJ. Apouey, B., Geoffard, P.-Y., 2013. Family income and child health in the UK. J. Health Econ. 32 (4), 715–727. Atkinson, R., Kintrea, K., 2001. Disentangling area effects: evidence from deprived and non-deprived neighbourhoods. Urban Stud. 38 (12), 2277–2298. Barker, D., 1998. Mothers, Babies and Health in Later Life, 2nd ed. Churchill Livingstone, Edinburgh, UK. Bilger, M., Carrieri, V., 2013. Health in the cities: when the neighborhood matters more than income. J. Health Econ. 32 (1), 1–11. Case, A., Lee, D., Paxson, C., 2008. The income gradient in children’s health: a comment on Currie, Shields and Wheatley Price. J. Health Econ. 27 (3), 801–807. Case, A., Lubotsky, D., Paxson, C., 2002. Economic status and health in childhood: the origins of the gradient. Am. Econ. Rev. 92 (5), 1308–1334. Case, A., Paxson, C., 2008. Stature and status: height, ability, and labor market outcomes. J. Pol. Econ. 116 (3). Case, A., Paxson, C., Vogl, T., 2007. Socioeconomic status and health in childhood: a comment on Chen, Martin and Matthews. Soc. Sci. Med. 64 (4), 757–761. Chen, E., Martin, A.D., Matthews, K.A., 2006. Socioeconomic status and health: do gradients differ within childhood and adolescence? Soc. Sci. Med. 62 (9), 2161–2170. Condliffe, S., Link, C.R., 2008. The relationship between economic status and child health: evidence from the United States. Am. Econ. Rev. 98 (4), 1605–1618. Cook, D., Strachan, D., 1999. Summary of effects of parental smoking on the respiratory health of children and implications for research. Thorax 54 (4), 357–366. Cunha, F., Heckman, J., 2007. The technology of skill formation. Am. Econ. Rev. 97 (2), 31–47. Currie, A., Shields, M.A., Price, S.W., 2007. The child health/family income gradient: evidence from England. J. Health Econ. 26 (2), 213–232. Currie, J., 2009. Healthy, wealthy, and wise: socioeconomic status, poor health in childhood, and human capital development. J. Econ. Lit. 47 (1), 87–122. Currie, J., Stabile, M., 2003. Socioeconomic status and child health: why is the relationship stronger for older children? Am. Econ. Rev. 93 (5), 1813–1823. Davidson, R., MacKinnon, J., 2004. Econometric Theory and Methods. Oxford Univ. Press, New York, NY. Doyle, O., Harmon, C., Walker, I., 2007. The impact of parental income and education on child health, Further evidence for England. Working Papers 200706. Geary Institute, University College Dublin. Ettner, S.L., 1996. New evidence on the relationship between income and health. J. Health Econ. 15 (1), 67–85. Frijters, P., Johnston, D.W., Shah, M., Shields, M.A., 2009. To work or not to work? Child development and maternal labor supply. Am. Econ. J.: Appl. Econ. 1 (3), 97–110. Fuchs, V.R., 1982. Time preference and health: an exploratory study. In: Economic Aspects of Health, NBER Chapters,. National Bureau of Economic Research, Inc, pp. 93–120. Gilliland, F., Li, Y., Peters, J., 2001. Effects of maternal smoking during pregnancy and environmental tobacco smoke on asthma and wheezing in children. Am. J. Respir. Crit. Care Med. 163 (2), 429–436. Gini, G., Pozzoli, T., 2009. Association between bullying and psychosomatic problems: a meta-analysis. Pediatrics 123 (3), 1059–1065. Gravelle, H., Sutton, M., 2003. Income related inequalities in self assessed health in Britain: 1979–1995. J. Epidemiol. Community Health 57 (2), 125–129. Gregg, P., Washbrook, E., Propper, C., Burgess, S., 2005. The effects of a mother’s return to work decision on child development in the UK. Econ. J. 115 (501), F48–F80. Heineck, G., 2009. Too tall to be smart? the relationship between height and cognitive abilities. Econ. Lett. 105 (1), 78–80. Imbens, G., Angrist, J., 1994. Identification and estimation of local average treatment effects. Econometrica 62 (2), 467–475. Jacob, B.A., Ludwig, J., Miller, D.L., 2013. The effects of housing and neighborhood conditions on child mortality. J. Health Econ. 32 (1), 195–206. Kawachi, I., Adler, N.E., Dow, W.H., 2010. Money, schooling, and health: mechanisms and causal evidence. Ann. N.Y. Acad. Sci. 1186 (1), 56–68. Kelly, Y., Bartley, M., 2010. Parental and child health. In: Children of the 21st century: the first five years 2, pp. 249–272. Khanam, R., Nghiem, H.S., Connelly, L.B., 2009. Child health and the income gradient: evidence from Australia. J. Health Econ. 28 (4), 805–817. Kruk, K.E., 2012. Parental income and the dynamics of health inequality in early childhood evidence from the UK. Health Econ. 22 (10), 1199–1214.

150


Long, L., 1972. The influence of number and ages of children on residential mobility. Demography 9 (3), 371–382. Maurin, E., 2002. The impact of parental income on early schooling transitions: a re-examination using data over three generations. J. Public Econ. 85 (3), 301–332. Mincer, J., 1977. Family Migration Decisions. National Bureau of Economic Research Cambridge, MA, USA. Mundlak, Y., 1978. On the pooling of time series and cross section data. Econometrica 46 (1), 69–85. Murasko, J.E., 2008. An evaluation of the age-profile in the relationship between household income and the health of children in the United States. J. Health Econ. 27 (6), 1489–1502. Paul, K.I., Moser, K., 2009. Unemployment impairs mental health: meta-analyses. J. Vocat. Behav. 74 (3), 264–282.

Peat, J., Dickerson, J., Li, J., 2007. Effects of damp and mould in the home on respiratory health: a review of the literature. Allergy 53 (2), 120–128. Propper, C., Rigg, J., Burgess, S., 2007. Child health: evidence on the roles of family income and maternal mental health from a UK birth cohort. Health Econ. 16 (11), 1245–1269. Ravelli, A., van der Meulen, J., Michels, R., Osmond, C., Barker, D., Hales, C., Bleker, O., et al., 1998. Glucose tolerance in adults after prenatal exposure to famine. Lancet 351 (9097), 173–177. Reinhold, S., Jürges, H., 2012. Parental income and child health in Germany. Health Econ. 21 (5), 562–579. Schmitz, H., 2011. Why are the unemployed in worse health? The causal effect of unemployment on health. Labour Econ. 18 (1), 71–78. Xu, X., Kaestner, R., 2010. The business cycle and health behaviors. Technical report. National Bureau of Economic Research.

Child health and access to health care in France: Evidence on the role of family income.

Associations between family structure change and child behavior problems: the moderating effect of family income.

The effect of maternal and child health and family planning services on mortality: is prevention enough?

Effect of the new GP contract on child health clinics.

Family structure and child health.

Income gradients in oral health according to child age.

Child poverty and deprivation in the UK.

Maternal mental health, and child growth and development, in four low-income and middle-income countries.

Family employment and child socioemotional behaviour: longitudinal findings from the UK Millennium Cohort Study.

The child in renal failure. Emotional impact of treatment on the child and his family.

Child care in the new towns in the UK.

Poverty and child health in the UK: using evidence for action.

Job loss and unmet health care needs in the economic recession: different associations by family income.

Smoothing out the transition of care between maternity and child and family health services: perspectives of child and family health nurses and midwives'.

Impact of pertussis on the afflicted child and family.

Data on Income inequality in Germany, France, Italy, Spain, the UK, and other affluent nations, 2012.

The effects of maternal depression on child mental health problems based on gender of the child.

Neighborhood Effect Heterogeneity by Family Income and Developmental Period.

Family income dynamics, early childhood education and care, and early child behavior problems in Norway.

Professional paradigms in child and family health programs.

Income inequality and child maltreatment in the United States.

The psychosomatic family in child psychiatry.

The partial effect of income on suicide is always negative.

Relative age effect on Nobel laureates in the UK.