Cross-sectional schooling-health associations misrepresented causal schooling effects on adult health and health-related behaviors: Evidence from the Chinese Adults Twins Survey

Abstract

Adult health outcomes and health behaviors are often associated with schooling. However, such associations do not necessarily imply that schooling has causal effects on health with the signs or magnitudes found in the cross-sectional associations. Schooling may be proxying for unobserved factors related to genetics and family background that directly affect both health and schooling. Recently several studies have used within-monozygotic (MZ) twins methods to control for unobserved factors shared by identical twins. Within-MZ estimates for developed countries are generally smaller than suggested by cross-sectional associations, consistent with positive correlations between unobserved factors that determine schooling and those that determine health. This study contributes new estimates of cross-sectional associations and within-MZ causal effects using the Chinese Adults Twins Survey, the first study of its type for developing countries. The cross-sectional estimates suggest that schooling is significantly associated with adult health-related behaviors (smoking, drinking, exercising) but not with own or spouse health outcomes (general health, mental health, overweight, chronic diseases). However, within-MZ-twins estimators change the estimates for approximately half of these health indicators, in one case declining in absolute magnitudes and becoming insignificant and in the other cases increasing in absolute magnitudes. Within-MZ estimates indicate significant pro-health effects for at least one of the indicators for own health (better mental health), own health-related behaviors (less smoking) and spouse health (less overweight).

1. Introduction

Positive associations between schooling and health are extensively documented. More-schooled individuals are more likely to have better health behaviors and outcomes (Cutler and Lleras-Muney, 2008, 2010; Albert and Davia, 2010; Zhang et al., 2010; Lange, 2011; Montez et al., 2011; Montez et al., 2012). However, these wide-spread positive schooling-health associations do not necessarily reflect beneficial effects of schooling on health. The health-schooling gradient may be in part due to unobserved endowments – such as innate physical and mental capabilities and family and community background – that affect both schooling and health. For example, people with better health-related endowments may not only have better health outcomes but may also have had more energy to learn and thus obtain more schooling. To what degree health-schooling associations actually reflect the causal effects of schooling on health is an important question for understanding both schooling returns and health determinants.

A number of recent studies examine the causal effects of schooling on health in developed countries. One set of studies employs changes in schooling policies, usually increases in the minimum age or the legally permitted grade to leave school, as instrumental variables for schooling attainment to control for endogeneity. These studies provide LATE (local average treatment effect) estimates of the impact of schooling on health for individuals close to the minimum legal schooling levels. Another set of studies uses identical twins (monozygotic, i.e., MZ) to difference out the unobserved endowments shared by these twins. These studies obtain estimates of the impacts of the differences in schooling within a pair of identical twins on their health differences at various schooling levels, not only the minimum legal schooling levels.

The results from these studies vary. Some suggest that schooling has positive causal effects on health outcomes such as long-term illness, mortality, height and weight, and little effects on health-related behaviors such as smoking and drinking (Groot and Maassen van den Brink, 2007; Park and Kang, 2008; Fujiwara and Kawachi, 2009; Sillers, 2009; Kemptner et al., 2011; Lundborg, 2013). Others conclude that there are many fewer and much smaller causal effects than in the cross-sectional associations or that there are no causal effects of schooling on health. Some of these studies imply that more-schooled people have better health mainly because they have better endowments (Arendt, 2005; Albouy and Lequien, 2009; Braakmann, 2011; Behrman et al., 2011; Amin et al., 2013; Clark and Royer, 2013; Amin et al., 2014.

A smaller number of previous studies investigate the health effects of schooling within marriage and find positive associations of own schooling with spouse health status (Egeland et al., 2002; Lipowicz, 2003; Jaffe et al., 2005; Kravdal, 2008; Li et al., 2011). We are aware, however, of only one such study that uses within-MZ-twins estimates to control for unobserved endowments. This study, using US twins, finds a number of significant cross-sectional associations between own schooling and spouse health outcomes and health-related behaviors, but they all become insignificant in within-MZ-twins estimates (Amin et al., 2014).

For developing countries, there also are numerous studies of the associations between own schooling and adult health. However, to our knowledge, there are few studies of the causal effects of schooling on health and health-related behaviors controlling for schooling endogeneity, and no previous study that uses within-MZ-twins estimates for such a purpose. The few studies of which we are aware use randomized controlled trails (RCTs), instrumental variables (IV), or regression discontinuity designs (RDD) based on “natural policy experiments”. Jensen and Lleras-Muney (2012) analyze a Dominican Republic program that gives teenage boys randomized information about returns to schooling, which induces boys to stay in school longer and improves health-related behaviors by delaying the onset of heavy drinking and reducing smoking at age 18. Duflo et al., (2012) report that a randomized Kenyan school subsidy program increases schooling but has weak impacts on sexual behaviors and STD infections. Alsan and Cutler (2013), using distance from school as an IV, find that more secondary schooling for adolescent Ugandan girls increases sexual abstinence and therefore reduces exposure to sexually transmitted diseases. Behrman (2014) finds that abolishing school fees in the mid-1990s increases schooling for girls in Malawi and Uganda; she then uses an RDD to estimate that a one-grade increase in the schooling of girls leads to a reduction of six percentage points in the probability of testing positive for HIV as an adult in Malawi and a three percentage point reduction in Uganda.

In this study, using MZ twins data from urban China for the year 2002, we report both cross-sectional and within-MZ-twins estimates of the effects of one’s own schooling on one’s own health outcomes and health-related behaviors, as well as on spouse health outcomes. Cross-sectional estimates indicate that: (1) Own schooling has no significant associations with self-reported health status (SRH), mental health, being overweight and the number of chronic diseases; (2) own schooling is negatively associated with packs of cigarettes consumed and drinking alcohol and positively associated with exercising – and is thereby associated with improved health-related behaviors; (3) own schooling is not significantly associated with spouse global health status, the spouse being overweight, or the number of chronic diseases of the spouse.

However, within-MZ-twins causal estimates indicate that schooling has protective effects on some of one’s own and spouse health outcomes and own health-related behaviors that differ from the cross-sectional associations: (1) Schooling improves significantly own mental health (and at the 0.10 level, own self-reported health and reduces the number of chronic diseases) in contrast to insignificant cross-sectional associations; (2) schooling reduces smoking as in the cross-sectional associations but no longer significantly affects drinking or exercise as in the cross-sectional estimates; and (3) schooling reduces the probability of the spouse being overweight in contrast to the cross-sectional insignificant association. The majority of the causal estimates are larger in absolute magnitudes than the cross-sectional associations.

2. Methods

2.1. Data

The Chinese Adults Twins Survey (CATS) was conducted by one author of this study, Zhang, and the Urban Survey Unit (USU) of the National Bureau of Statistics (NBS). CATS was undertaken in 2002 in five Chinese cities with current population of approximately 35 million inhabitants: Chengdu, Chongqing, Harbin, Hefei, and Wuhan. CATS covered substantial socioeconomic information, building on previous twins questionnaires used in the US and elsewhere. Adult twins between the ages of 18 and 65 were identified by local statistical bureaus through channels that included colleagues, friends, relatives, newspaper advertising, neighborhood notices, neighborhood management committees, and household records from local public security bureaus. These channels permitted an approximately equal probability for contacting all twins in these cities, so the sample was approximately representative. Questionnaires were completed through face-to-face interviews. The survey was conducted with considerable care. Zhang made several site checks and closely supervised and monitored data input processes.

CATS was the first Chinese, and probably Asian, adult socioeconomic twin dataset. Completed questionnaires were collected from 3002 individuals, of whom 2996 were twins and six were triplets. Among these individuals were 914 complete pairs of identical twins (1828 individuals) for which both twins responded that they had identical hair color, looks, gender, and age. The data were assessed to be of high quality and were used in a number of published studies (see Li et al., 2007, 2010, 2012; Huang et al., 2009; and Rosenzweig and Zhang, 2013). The descriptive statistics for key variables of CATS are similar to those in the Urban Household Survey (UHS) conducted in 2002 by the National Bureau of Statistics of China (Appendix Table A1).

2.1.1. Own and spouse health outcomes

(1) Self-reported health (SRH): Respondents were asked to evaluate their own and their spouses’ global health; “How is your current health status? 1 means very good, 2 means good, 3 means just so–so, and 4 and 5 mean poor and very poor, respectively.” Because only 4.6% of interviewees reported poor and very poor SRH, our variable is 0 (not in good health) if the answers were 3, 4, and 5 and 1 otherwise (in good health). (2) Overweight: Respondents were asked to report their own and their spouses’ current heights in centimeters and weights in kilograms, which were used to calculate body mass indices (BMI). “Overweight” is 1 if BMI was ≥25.0 and 0 otherwise. (3) Number of chronic diseases: The interviewees were asked about their own and their spouses’ chronic diseases: “Do you have the following symptoms?” The 12 symptoms/conditions included: hemicranias, allergy to pollen, skin rash, hearing damage, hypertension, neurasthenia, problems triggered by drinking, cardiomyopathy, neck injury, dorsum injury, arm injury and leg injury. The number of chronic diseases is the sum of positive responses. (4) Mental health: The respondents were asked, “Do you have the following emotions: Sadness, Fear, Indignation, and Disgust? 1 means often have, 2 means occasionally have, 3 means seldom have, and 4 means never have.” Our mental health index weights these emotions equally.

2.1.2. Own health-related behaviors

Three continuous variables are used based on the answers to: “How many packs of cigarettes do you smoke per day?” “How many days do you drink per week?” and “How many days do you exercise at least 30 min per month?”

2.1.3. Schooling

Interviewees were asked how many years they completed in each schooling level, which are summed to represent the total years of schooling. Respondents were also asked, “How many years did your spouse receive formal education since primary school?” Further, they were asked about the schooling of their twin sibling, the responses to which we use to control for random measurement error in reporting schooling (see below).

2.1.4. Birth weight

Interviewees were asked their birth weight.

2.2. Modeling and estimation

The objective of this study is to estimate the causal effects of one’s schooling on one’s own health and health-related behaviors and on the health of the spouse, using within-MZ-twins estimators. We begin with the following linear equation for the mth health outcome (or health-related behavior) for adult i from parental family j.

$$H_{ij}^m={\beta }_0^m+{\beta }_1^m{S}_{ij}+{\mathit{X}}_{\mathit{ij}}{\mathit{\beta }}_2^{\mathit{m}}+{\mathit{\theta }}_{\mathit{j}}{\mathit{\beta }}_3^{\mathit{m}}+{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\beta }}_4^{\mathit{m}}+u_{ij}^m$$ (1)

where $H_{ij}^m=\text{health}\text{outcome}(\text{behavior})$, S_ij = schooling, X_ij = vector of observed individual characteristics (i.e., age, gender, birth weight, and regional dummies), θ_j = vector of unobserved endowments (including genes and early family background) common to all adults from jth parental family, α_ij = vector of unobserved individual-specific endowments of ith adult from jth parental family and $u_{ij}^m$ = random health shock not correlated with other right-side variables in Equation (1). As noted, the symbols in bold are vectors. We are interested in estimating ${\beta }_1^m$, the effect of one’s own schooling on the mth own health outcome (behavior). We include birth weight in X_ij because birth weights summarize health/nutritional experiences from conception to birth, as well as the possible induced parental investments in schooling and health in response to birth outcomes. Birth weights may differ between a pair of identical twins exogenously (due to different placement relative to the placenta). The relevant parameters ( ${\beta }_0^m,{\beta }_1^m,{\mathit{\beta }}_2^{\mathit{m}},{\mathit{\beta }}_3^{\mathit{m}},{\mathit{\beta }}_4^{\mathit{m}}$) and random disturbance term $u_{ij}^m$ in general differ among different health outcomes and health-related behaviors, for which reason the superscript m refers to the mth health outcome or health-related behavior. In addition, it is important that the common and individual-specific endowments (θ_j,α_ij) are vectors with multiple dimensions that have coefficient parameters ( ${\mathit{\beta }}_3^{\mathit{m}},{\mathit{\beta }}_4^{\mathit{m}}$) that are dependent on the health outcome (behavior) under study. This allows the components of the endowments, for example, to affect SRH differently than they affect mental health, being overweight, or number of chronic diseases. One important implication of there being multiple components of endowments with different weights for different health outcomes and behaviors is that a priori we expect different effects of controlling for endowments across different health outcomes and behaviors.

To estimate the relationship between spouse health and own schooling within marriage, we use a parallel specification, but $H_{ij}^m$ represents spouse health status; ${\mathit{X}}_{\mathit{ij}}^{\mathit{m}}$ includes spouse age, gender, schooling and regional dummies; and ${\beta }_1^m$ is the effect of own schooling on spouse mth health outcome net of assortative mating on schooling.

To illustrate the challenges in obtaining unbiased estimates of ${\beta }_1^m$, consider the following specification that determines own schooling:

$$S_{ij}={\gamma }_0+{\mathit{X}}_{\mathit{ij}}{\mathit{\gamma }}_1+{\mathit{\theta }}_{\mathit{j}}{\mathit{\gamma }}_2+{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\gamma }}_3+{\mathit{\alpha }}_{\mathit{kj}}{\mathit{\gamma }}_4+v_{ij}$$ (2)

Own schooling is hypothesized to be determined by the vector X_ij of observed individual characteristics (because some of the elements of ${\mathit{\beta }}_2^{\mathit{m}}$ and of γ₁ may be zero, the individual characteristics that affect schooling may differ from those that affect health); the vector θ_j of unobserved endowments common to adults from the jth parental family; the vector α_ij of individual-specific endowments of the ith adult from the jth parental family; the vector α_kj of individual-specific endowments of the kth adult from the jth parental family because of intrahousehold allocations among all children in the family and a random schooling shock v_ij that was not correlated with other right-side variables in Equation (2). (The basic points below hold with a two-child family; however, if there are more children than two in the family, α_kj is a vector including the individual-specific endowments of all other children in the family (Becker and Tomes, 1976; Behrman et al., 1982, 1995).)

In cross-sectional estimates of Equation (1), the estimated schooling coefficient includes the true effect of schooling ${\beta }_1^m$ plus effects of any correlated unobserved variables that are included in the compound disturbance term ${\mathit{\theta }}_{\mathit{j}}{\mathit{\beta }}_3^{\mathit{m}}+{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\beta }}_4^{\mathit{m}}+u_{ij}^m$. The common family endowment θ_j and individual-specific endowment α_ij in this compound disturbance term both determine schooling in relation (2), which implies that schooling in general is correlated with the compound disturbance term in relation (1) and thus the estimated coefficient of ${\beta }_1^m$ is biased.

The extent and even the sign of this bias, however, may vary across health outcomes and health-related behaviors. To illustrate, consider the case in which the common family endowment θ_j and individual-specific endowment α_ij both have two components, the first of which is positively associated with schooling (e.g., greater innate intellectual ability leads directly both to more schooling and better health) and the second of which is inversely associated with schooling (e.g., worse innate health leads to poorer health but induces more focus on intellectual rather than physical activities and therefore more schooling). If the weights on the different components of θ_j and α_ij in relation (1) (i.e., the components of ${\mathit{\beta }}_3^{\mathit{m}}$ and ${\mathit{\beta }}_4^{\mathit{m}}$) are much greater for the first than the second component for a particular health outcome, the weighted impact of the endowments that determine that health outcome ( ${\mathit{\theta }}_{\mathit{j}}{\mathit{\beta }}_3^{\mathit{m}}+{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\beta }}_4^{\mathit{m}}$) is positively correlated with the weighted impact of the endowments that determine schooling (θ_jγ₂+α_ijγ₃). In this case, the cross-sectional estimates of the absolute magnitude of schooling impacts on the mth health outcome or behavior are biased upwards, as reported in most of the previous literature using the within-MZ estimators summarized above. However, if the weights on the different components of θ_j and α_ij in relation (1) (i.e., the components of ${\mathit{\beta }}_3^{\mathit{m}}$ and ${\mathit{\beta }}_4^{\mathit{m}}$) are much greater for the second than the first component for a different health outcome or behavior, the weighted impact of the endowments that determine that health outcome ( ${\mathit{\theta }}_{\mathit{j}}{\mathit{\beta }}_3^{\mathit{m}}+{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\beta }}_4^{\mathit{m}}$) is negatively correlated with the weighted impact of the endowments that determine schooling (θ_jγ₂+α_ijγ₃). In this case, the cross-sectional estimates of the absolute magnitude of schooling impacts on this health outcome or behavior are biased downwards, as reported in some previous studies (e.g., Behrman and Rosenzweig, 2004; Behrman et al., 2014). Between these two cases is the possibility that for some other health outcomes and behaviors, there is no bias in the estimate for ${\mathit{\beta }}_1^{\mathit{m}}$.

Finally, schooling is likely measured with error. Classical random measurement error causes biases toward zero in estimated schooling coefficients ( ${\beta }_1^m$).

Within-siblings estimates (that is, family fixed effects, but we write them here as within-siblings estimates because this makes it more transparent how the unobserved family endowments are controlled) have been proposed to control for unobserved common family endowments:

$$\mathrm{\Delta }{H}_{ij}^m=\mathrm{\Delta }{\beta }_0^m+{\beta }_1^m\mathrm{\Delta }{S}_{ij}+\mathrm{\Delta }{\mathit{X}}_{\mathit{ij}}{\mathit{\beta }}_2^{\mathit{m}}+\mathrm{\Delta }{\mathit{\theta }}_{\mathit{j}}{\mathit{\beta }}_3^{\mathit{m}}+\mathrm{\Delta }{\mathit{\alpha }}_{\mathit{ij}}{\mathit{\beta }}_4^{\mathit{m}}+\mathrm{\Delta }{u}_{ij}^m$$ (3)

where Δ refers to the difference between the values of indicated variables for the ith and those for the kth adult sibling in the jth parental family (i.e., $\mathrm{\Delta }{H}_{ij}^m=H_{ij}^m-H_{kj}^m$). Within-siblings estimates control for common family endowments because Δθ_j = 0. In general, the individual-specific endowments differ among siblings due to individual-specific differences in genetic endowments, so Δα_ij ≠ 0, which means that the within-siblings estimates of the schooling coefficients are still biased. However, for MZ twins, Δα_ij = 0 because they had identical genetic endowments at conception. Therefore, within-MZ-twins estimates eliminate biases that are likely to contaminate cross-sectional estimates of relation (1) due to either unobserved common or individual-specific endowments. Because they control for unobserved family factors, they also control for the possibility that being raised in families with twins is different from growing up in families without twins.

Within-MZ-twins estimates (like any fixed-effects estimates) also exacerbate attenuation toward zero due to classical random measurement error. We adopt the procedure suggested and implemented by Ashenfelter and Krueger (1994) and Behrman et al. (1994), using reports on respondent schooling by others to instrument own schooling reports under the assumption that measurement errors in own and others’ reports are not correlated. We use co-twin reports to instrument own-schooling reports for fixed-effects instrumental variable estimates (FE-IV). We also use the same procedure to control for random measurement error in our estimates of relation (1) (LEVEL-IV) so that the cross-sectional LEVEL-IV estimates and the FE-IV estimates can be compared.

3. Results

3.1. Descriptive statistics

The top panel of Table 1 displays twins’ general characteristics: 56% male, an average age of 37, and 11.3 years of schooling. The middle panel gives various health measures: (1) 59% reported that their global health status is good or very good; (2) 16% are overweight; (3) the mean mental health index is 2.9, reflecting that approximately 30% had emotions of sadness, fear, indignation, and disgust, 58.7% seldom had those emotions, and the rest never had them; and (4) the average number of chronic diseases is 0.65. The health-related average behaviors are 0.27 packs of cigarettes smoked per day, 0.98 days per week of alcohol consumption, and 4.7 days of exercise per month.

Table 1.

Descriptive statistics.

	Number of observations	Mean	S.D.
Twins’ Characteristics
Schooling, years	1744	11.26	2.94
Age, years	1744	37.18	10.20
Male	1744	0.56	0.50
Birth weight, kg	1744	2.42	0.57
Married	1744	0.72	0.45
Labor income, yuan per month	1744	573.20	622.53
Twins’ Health Outcomes
Self-reported health status	1744	0.59	0.49
Mental health	1730	2.91	0.59
Overweight	1738	0.16	0.36
Number of chronic diseases	1720	0.65	1.17
Twins’ Health-Related Behaviors
Packs of cigarettes per day	1742	0.27	0.41
Days of drinking per week	1664	0.98	2.02
Days of exercise per month	1716	4.67	8.27
Spouse Characteristics (Reported by Twins)
Spouse schooling, years	1094	10.79	2.99
Spouse age, years	1094	41.41	8.65
Spouse health status, reported by twins	1094	0.56	0.50
Spouse overweight	1094	0.15	0.35
Number of chronic diseases of spouse	1058	0.51	1.01

The bottom panel presents the characteristics of the spouses of the twins. Twins’ spouses have an average age of 41.4 years and an average of 10.8 years of schooling. The twins reported that 55.6% of their spouses are in good health and 14.7% are overweight; the average number of chronic diseases is 0.51.

To effectively implement within-MZ-twins estimators, there needs to be some within-twin-pair schooling variation. Table 2 indicates that 44.6% of twin pairs have schooling differences: 12.8% have a one-year difference, 11.4% have a two-year difference, 12.7% have a three-year difference and 7.7% have a difference of four or more years. This distribution suggests that within-twin-pair variation in schooling is adequate. To provide a perspective for reference, compulsory schooling laws in the US increased schooling by only 0.05 grades (Lleras-Muney, 2005), and the well-known Mexican PROGRESA program increased schooling by 0.7 grades (Schultz, 2004).

Table 2.

Differences in the years of schooling within MZ-Twins pairs.

Differences in schooling attainment, years	Number of twin pairs	Percentages
0	483	55.4
1	112	12.8
2	99	11.4
3	111	12.7
4	36	4.1
>4	31	3.6

There also has to be sufficient within-twin-pair variations for health outcomes and health-related behaviors of interest. Table 3 presents the means, standard deviations and percentages of no differences between twins. The standard deviations are generally relatively large. The percentages with no difference range from 28% (mental health) to 85.6% (overweight) for twin’s own health, from 69.6% (days of exercise per month) to 76.7% (days of alcohol consumption per week) for twins’ health-related behaviors, and from 63.3% (number of chronic diseases) to 76.1% (overweight) for spouse health outcomes. Therefore, there appears to be sufficient within-MZ-twin variations to support the analyses, but there is less variation for certain health outcomes, such as being overweight.

Table 3.

Descriptive statistics: within-MZ-Twins variations.

	Mean	S.D.	Percentages of no differences within twin pairs
Twin’s Health Outcomes
Self-reported health status	−0.0011	0.50	75.1
Mental health	−0.0098	0.59	28.0
Overweight	−0.0173	0.38	85.6
Number of chronic diseases	−0.0570	1.31	60.1
Twins’ Health-Related Behaviors
Packs of cigarettes per day	−0.0109	0.31	75.0
Days of drinking per week	−0.0108	1.59	76.7
Days of exercise per month	−0.4831	8.30	69.6
Spouse Schooling & Health Outcomes (Reported by Twins)
Spouse schooling	0.0238	3.00	33.8
Spouse health status	0.0000	0.56	68.6
Spouse overweight	0.0530	0.49	76.1
Number of chronic diseases of spouse	−0.0378	1.24	63.3
Twins’ Other Socioeconomic Status
Birth weight	−0.0044	0.34	24.2
Labor income, yuan per month	25.6819	671.85	29.1
Married	0.0023	0.43	81.9

3.2. LEVEL-IV and FE-IV estimates

Table 4 provides the LEVEL-IV estimates of Equation (1) in column 1 and the FE-IV estimates of Equation (3) in column 2; in both cases, the co-twins’ reports of own schooling are used to control for measurement error (as indicated by “IV”; Appendix Table A2 displays the cross-sectional and FE estimates that are not controlled for measurement error, which are very similar in their implications to the IV estimates though with some slight differences in details). The standard errors in parentheses are robust, with the level estimates clustered at the family level. The first panel refers to own-health outcomes, the second panel to own-health-related behaviors, the third panel to spouse health outcomes, and the fourth panel to auxiliary outcomes for which schooling impacts are often considered important in the literature (labor income, marital status, spouse schooling).

Table 4.

Twins’ schooling attainment and health outcomes, health-related behaviors, spouse health outcomes, and other outcomes.

Dependent variables	LEVEL-IVa	FE-IVb	Observations
1. Twins’ Own Health Outcomes
Self-reported health statusc	0.007 (0.005)	0.018* (0.010)	1744
Mental healthd	0.003 (0.006)	0.035*** (0.012)	1730
Overweighte	0.001 (0.004)	0.005 (0.008)	1738
Number of chronic diseases	−0.007 (0.010)	−0.053* (0.027)	1720
2. Twins’ Own Health-Related Behaviors
Packs of cigarettes per day	−0.020*** (0.004)	−0.0173*** (0.006)	1742
Days of drinking per week	−0.053*** (0.020)	0.028 (0.035)	1664
Days of exercise per month	0.210*** (0.075)	0.208 (0.178)	1716
3. Twins’ Spouse Health Outcomes (Reported by Twins)
Spouse health statusc	0.008 (0.006)	−0.001 (0.015)	1094
Spouse overweighte	−0.007 (0.005)	−0.032** (0.013)	1094
Number of chronic diseases of spouse	−0.005 (0.011)	−0.030 (0.033)	1058
4. Twins’ Other Socioeconomic Outcomes
Log labor income	0.087*** (0.006)	0.032** (0.015)	944
Spouse schooling	0.547*** (0.035)	0.329*** (0.077)	1094
Married	0.002 (0.004)	0.006 (0.009)	1746

Notes: Standard errors in parentheses were robust to heteroscedasticity and for LEVEL-IV clustered at the family level.

^*significant at 10%;

^**significant at 5%;

^***significant at 1%.

^aLEVEL-IV denotes cross-sectional estimation with twin’s schooling corrected for measurement error, where schooling reported by twin 1 (S¹₁S¹₂) is used as the regressor, and schooling reported by twin 2 (S²₁S²₂) as the IV. Regressions in Panel 1 and 2 also include twins’ ages, gender, birth weight and regional dummies. Regressions in Panel 3 also include spouse age, gender and schooling, twins’ birth weight and regional dummies.

^bFE-IV denotes within-MZ-twins estimation with twin’s schooling corrected for the measurement error, where S¹₁–S¹₂ is used as the regressor and S²₁–S²₂ as the IV.

^cSelf-reported health = 0 if respondents reported that they were not in good health and 1 otherwise.

^dMental health index weighted equally the following emotions: sadness, fear, indignation, and disgust; measured as never have to often have.

^eOverweight = 1 if BMI≧25.0 and 0 otherwise.

3.3. Own-health outcomes (Table 4, panel 1)

LEVEL-IV associations of schooling with various health outcomes (column 1) indicate that more schooling is not significantly associated with the four health outcomes. However, the lack of these significant associations of schooling with own-health outcomes does not necessarily reflect the lack of causal effects but may be due in part or entirely to unobserved endowments that affect both schooling (Equation (2)) and health outcomes (Equation (1)). FE-IV estimates are conducted to control for such endowments as in Equation (3) (column 2, Table 4).

The FE-IV schooling coefficient estimates (column 2) for three of the four health outcomes (the exception is being overweight) are significantly positive (at the 0.10 level) for SHR, negative (at the 0.10 level) for number of chronic diseases, and positive (at the 0.05 level, used hereafter) for mental health. These results indicate that schooling has desirable causal effects on these health dimensions. The absolute magnitudes of the FE-IV schooling coefficient estimates with the SRH, mental health and number of chronic disease outcomes are more than twice as large as the LEVEL-IV estimates. This implies that people who completed college (16 grades) in comparison with people who completed junior high school (9 grades) have a 12.6% higher probability of having good health, a 0.41 standard deviation improvement in mental health and 0.37 fewer chronic diseases.

3.4. Own-health-related behaviors (Table 4, panel 2)

Schooling predicts better health-related behaviors. LEVEL-IV schooling coefficients for smoking and drinking are significantly negative, and for exercising, they are significantly positive. An additional year of schooling predicts that per year respondents smoke 7.3 fewer packs of cigarettes, drink 2.8 days less, and exercise 2.5 more days.

For packs of cigarettes smoked per day, the FE-IV estimates of causal impacts of schooling are significant and approximately the same as the LEVEL-IV estimates. In contrast, for days of drinking per week and days of exercise per month, they become insignificant – approximately the same magnitude as the LEVEL-IV estimates for days of exercise per month but the opposite sign for days of drinking per week.

3.5. Spouse health (Table 4, panel 3)

LEVEL-IV estimates (column 1) indicate that more own schooling has no significant associations with the three spouse health outcomes that we investigate. However, as for own health and health-related behaviors, own schooling may be proxying in part for endowments in the LEVEL-IV estimates. Two of the three FE-IV coefficient estimates are insignificant, as are the LEVEL-IV estimates. The third is the FE-IV coefficient estimate of own schooling for spouse being overweight, which differs substantially from the insignificantly negative LEVEL-IV estimate by becoming significant and approximately four times as large in absolute magnitude.

3.6. Labor income, being married and spouse schooling (Table 4, panel 4)

Previous within-MZ studies have focused primarily on outcomes other than health, in particular labor income but also marital status and spouse schooling (Ashenfelter and Krueger, 1994; Behrman et al., 1994; Bound and Solon, 1999; Li et al., 2012). In the CATS sample, there are strong significant associations between schooling and both labor income and spouse schooling, though not for being married, in the LEVEL-IV estimates. However, the FE-IV coefficient estimates of the causal effects of schooling on labor income and spouse schooling are only approximately half as large as the LEVEL-IV associations. This suggests that in the LEVEL-IV estimates, schooling is proxying in part for unobserved characteristics (e.g., abilities, motivations, and personalities) that are positively correlated with schooling, labor income and spouse schooling. Nevertheless, the FE-IV estimates indicate significantly positive effects of schooling on both labor income and the schooling level of the spouse with whom he/she mates in the marriage market.

4. Discussion

The purpose of this paper is to investigate whether there are significant associations between schooling and adult health in urban China using the 2002 CATS data and how such associations differ, if at all, from the causal effects obtained from FE-IV estimates. To our knowledge, we contribute the first estimates using the within-MZ-twins estimation strategy to investigate causal schooling impacts on own-health outcomes and health-related behaviors for a developing economy. We contribute further with one of the first two estimates to our knowledge (together with contemporaneous estimates for the US in Amin et al., 2014) of cross-spouse schooling effects using within-MZ estimators to control for unobserved endowments.

The LEVEL-IV estimates in Table 4 suggest that there are certain limited significant associations between own schooling and indicators of own health, own health-related behaviors, and spouse health. In particular, significant LEVEL-IV coefficients are obtained only for the health behaviors. No significant associations are found for any of the own or spouse health outcomes.

Comparing the relative magnitudes and significance between LEVEL-IV and FE-IV estimates for schooling-health relations in Table 4 indicates four cases: (1) Associations between schooling and health outcomes are similar to FE-IV estimates both in magnitude and significance, so controlling for endowments does not lead to a different interpretation regarding significance or, if significant, the magnitudes (e.g., overweight, cigarettes, spouse general health status, and spouse number of chronic diseases); (2) the association between schooling and health outcomes is similar to the FE-IV estimate in magnitude but with a difference in significance so that the LEVEL-IV estimate appears significant but not the FE-IV estimate (e.g., exercising); (3) the FE-IV estimate is insignificant and much smaller than and opposite in sign from the significant LEVEL-IV estimate (e.g., drinking); and (4) the FE-IV estimates suggest greater effects on health than the LEVEL-IV estimates (e.g., self-reported health, mental health, number of chronic diseases, and spouse being overweight).

For half of the estimates (cases 1 and 2 in the paragraph above), there is no difference in the basic inference about the magnitude, although for the one estimate in case 2, the FE-IV estimate appears insignificant in contrast to the significant LEVEL-IV estimate. For the other half of the estimates, the LEVEL-IV estimates of associations result in misleading inferences about the magnitudes of the causal effects indicated by the FE-IV estimates. The differences, moreover, are in both directions. For one health-related behavior (case 3), the FE-IV estimates indicate smaller effects of schooling than the LEVEL-IV estimates, and for four health outcomes (case 4), larger effects are indicated. Therefore, for one health indicator, the LEVEL-IV estimates overstate the absolute magnitudes of the FE-IV estimates of the causal effects of schooling, and for others understate them. These results are consistent with the presence of different effects (parameter values) for different components of endowments that affect different indicators of health outcomes and behaviors as presented in Equation (1). They also suggest that for some health indicators, the relevant endowments that improve a particular health outcome are negatively correlated with the endowments that positively affect schooling (and vice versa), as illustrated after Equation (2) (also see Appendix Table A3).

As noted above, the most common results of previous within-MZ-twins studies and of the estimates presented in Panel 4 of Table 4 for labor income and spouse schooling suggest positive correlations between endowments that affect the outcomes of interest, including health, and those that affect schooling. In this study, however, we find that for case 4 mentioned above, the LEVEL-IV estimates substantially understate the FE-IV estimated schooling impacts on health and health-related behaviors. One should thus not take for granted that cross-sectional level estimates are always upward-biased. As discussed after Equation (2) above, the biases due to not controlling for unobserved endowments can be in either direction.

These results are basically robust with very similar patterns of estimates with certain variations: excluding 3.6% of the observations for which the schooling difference between twins exceeded four years to determine if the results are sensitive to outliers (Appendix Table A4); excluding 0.5% of the observations for which one or both twins are still in school (Appendix Table A5); excluding birth weight from the specification (Appendix Table A6); and limiting the observations for each health category shown in Table 4 to a common subsample (Table A7). If relations are allowed to differ for females versus males, the FE-IV schooling coefficients are significantly larger for males for own overweight and exercising and significantly smaller for cigarettes and drinking (Appendix Table A8). If categorical schooling categories are used instead of a continuous measure, the FE schooling coefficients suggest that: vocational high school is particularly important for mental health; college and above is particularly important for cigarettes and spouse being overweight; and senior high school is also important for spouse being overweight (Appendix Tables A9–A12).

4.1. Limitations

First, the data permit measuring health outcomes and health-related behaviors for a limited set of indicators, not including some that would be very interesting (e.g., mortality; clinical measurements such as blood pressure, hemoglobin level and inflammation; and dietary intakes). Moreover, the available indicators may be subject to measurement error. If any measurement error is random, it does not bias the estimates though it may increase imprecision. If measurement error is systematic, such as would be the case if more-schooled respondents report greater health problems for a given objective health status, the absolute values of the schooling-health associations are downward biased in both the cross-sectional and within-MZ estimates. However, the comparisons between the cross-sectional and within-MZ-twins estimates are still informative. Second, we do not explore a possible bias in the FE-IV estimates emphasized by Bound and Solon (1999): the correlation between the random disturbance terms in the health Equation (1) and the schooling Equation (2) due to environmental shocks that affected both schooling attainment and health outcomes/behaviors. If these disturbance terms are positively correlated, the FE-IV estimates (as well as the LEVEL-IV estimates) give upper bounds on the absolute magnitudes of causal schooling effects (and vice versa with negative correlations) as demonstrated in Kohler et al. (2011). Third, we are not able to identify the channels through which schooling affects health, though the estimates in Panel 4 of Table 4 are consistent with two channels being through income and spouse schooling. Fourth, our approach is partial equilibrium. If schooling were to change substantially, there might be market adjustments that change the returns to schooling in a great deal of activities, including those related to health. Fifth, twins may be treated differently than singletons, and certainly twins differ in some observed characteristics such as the distribution of birth weights (Behrman and Rosenzweig, 2004; Rosenzweig and Zhang, 2013), for which we control in our estimates. However, the FE-IV estimates explicitly control for the additive impacts of whatever might be different about being twins, so our estimates are not biased by the first-order effects of using twins. Sixth, previous studies find that parents respond to differences in child endowments (Behrman et al., 1994) and if those responses affect both schooling and health directly, schooling may be proxying in part for such responses in the FE-IV estimates. Of course, this is part of the basic problem in the usual LEVEL estimates, and the strength of the within-MZ-twins estimates is to control for genetic and other child endowments at conception. In addition, we control for the in utero differences through birth weight. However, it is possible that the within-MZ-twins differencing does not control for all endowments, so there is some residual confounding despite the strength of the within-MZ approach, together with birth weights, for controlling for endowments. Seventh, if twins adjust their schooling in response to the schooling of their co-twin, no bias in our FE-IV estimates results; however, if they adjust their health in response to the health of their co-twin, our FE-IV estimates of schooling effects are downward (upward) biased if they positively (negatively) imitate their co-twins (Kohler et al., 2011).

5. Conclusions

Despite these limitations, this study contributes to the understanding of the causal effects of schooling on health in the most populous developing country, China. It applies for the first time for a developing country the within-MZ-twins method to estimate schooling effects on health controlling for unobserved endowments. Our results indicate that in this context, schooling has significant causal effects on one’s own mental health and smoking behavior and on the spouse being overweight – and, at the 0.10 level, on one’s own self-reported health and number of chronic diseases. However, schooling has no beneficial effects, even at the 0.10 level, on being overweight, drinking and exercise, as well as on spouse global health and number of chronic diseases. These estimated causal effects of schooling on own and spouse health differ from the associations in a number of important ways: In one case, they indicate smaller schooling effects, but more often, they indicate greater schooling effects. Therefore, if there are externalities associated with adult health or if certain aspects of adult health are viewed as basic rights, policies subsidizing schooling may be warranted by our estimates; however, these policies would likely be different from those that would be inferred from the cross-sectional associations of schooling with adult health outcomes and related behaviors. In addition, a number of health outcomes and health-related behaviors for individuals, such as those in our sample, are not likely to be affected by increased schooling.

Appendix A. Supplementary data

Supplementary data related to this article can be found at http://dx.doi.org/10.1016/j.socscimed.2014.11.011.