Associations of Grandparental Schooling With Adult Grandchildren's Health Status, Smoking, and Obesity

Abstract

Despite persistent schooling-related health disparities in the United States, little is known about the multigenerational effects of schooling on adult health. As expected lifespans increase, direct influences of grandparental schooling on grandchildren's health may become increasingly important. We used multigenerational data spanning 41 years from a national sample of US families to investigate associations of grandparents’ educational attainment with global health status, smoking, and obesity in their grandchildren who were aged 25–55 years in 2009. We estimated total effects of grandparental schooling and, by using marginal structural models, we estimated controlled direct effects that were independent of parents’ and participants’ schooling. Among whites, lower levels of grandparental schooling were monotonically associated with poor health status, current smoking, and obesity in adult grandchildren. There was also evidence suggesting direct effects, which was stronger for poor health status among participants whose highest-educated grandparent lived in the same state. Among blacks, the only association suggesting a total or direct effect of grandparental schooling was for smoking. Despite the relative imprecision of our estimates and possible residual bias, these results suggest that higher levels of grandparental schooling may benefit the health of grandchildren in adulthood, especially among whites. Furthermore, part of those apparent effects, especially for obesity, may not be mediated by parents’ and grandchildren's schooling.

Persistent social and health inequalities, coupled with strong intragenerational schooling-health associations and plausible mechanistic pathways linking schooling of prior generations to the health of later generations (1–12) suggest that the influences of schooling on health may extend across multiple generations. Multigenerational effects of schooling on health may be mediated in part through intervening generations. For example, grandparental schooling may influence the schooling or health of the parents, which in turn influences the health of the grandchildren. However, grandparental schooling may also influence grandchildren's health directly. In the United States, 23% of children under age 5 years were regularly cared for by a grandparent in 2005 (13), and in 2010, 5.4 million children shared a household with a grandparent (14). These grandparents may influence their grandchildren's health by modeling healthy behaviors, providing emotional support, and instructing the grandchildren on life choices. Other direct influences—such as paying for health care, providing monetary gifts, and facilitating social or institutional relationships—do not require contact, or even that the grandparent be living (15). Grandparental influences on health may be beneficial (e.g., facilitating access to health care) or harmful (e.g., smoking in the presence of the grandchild).

Previous studies have linked higher parental schooling with better physical and mental health of the offspring in adulthood (1–8, 10–12). The few studies examining effects of grandparental schooling on adult health have related higher grandparental education to lower mortality risk among Danish men (16), greater height among Polish university students (17), and a higher risk of hospitalization for an eating disorder in Swedish women (18). We used longitudinal data from the Panel Study of Income Dynamics (PSID) to estimate the effects of grandparental schooling on the health status, smoking behaviors, and body mass index (BMI) values (weight (kg)/height (m)²) of adult grandchildren. We estimated associations representing both the total effects of grandparental schooling and “controlled direct effects” through pathways not mediated by parents’ and participants’ schooling (specifically, with parents’ and participants’ schooling fixed at their observed levels) (19). Controlled direct effects may provide useful estimates of the health effects of grandparental schooling we may expect if we alter grandchildren's schooling, such as through educational policy changes.

It has been posited that the health effects of schooling may differ by race (20). The “minority poverty hypothesis” proposes a steeper schooling gradient in health among minority populations than among non-Hispanic whites because of synergistic effects of low education and discrimination. In contrast, the “diminishing returns hypothesis” proposes a shallower schooling gradient among minority populations because of the smaller income and occupational returns to education they have historically experienced (21, 22). These hypotheses can be readily extended to multigenerational effects. We tested them by investigating modification of the effects of grandparental schooling on adult health by race.

Finally, we examined whether the associations of grandparental schooling differ by the geographical proximity of the grandchild to living grandparents. We would expect the amount of direct social interaction between grandparents and grandchildren to be higher if they live closer to each other. Therefore, stronger associations between grandparental schooling and health among grandchildren whose grandparents live geographically closer would support the existence of direct effects of grandparental schooling through pathways involving social interaction.

METHODS

Study population

Data came from the PSID, a longitudinal study started in 1968 of a national sample of US families (23). Interviews were conducted annually until 1997 and biennially since then (23). An adult family member, usually the head of household, serves as the primary respondent. When a family member leaves the household (e.g., when a child grows up), his or her new household is incorporated into the sample. The study contains information on up to 3 generations of any given family (23).

Our sample comprised 10,262 PSID participants who were 25–55 years of age in the 2009 study wave. The age range was bounded to limit age variation and maximize availability of childhood socioeconomic data. We also excluded participants who 1) were not heads of households (“heads”) or wives or cohabiting female partners of heads (“wives”) (n =1,321) because relevant data for our analysis were not collected for other family members, 2) did not have at least 1 parent who was also a PSID head or wife (n = 3,901), or 3) were not non-Hispanic and of black or white race (n = 392). Participants of other racial/ethnic groups were excluded because of insufficient sample sizes with multigenerational information. These criteria yielded an analytical sample of 4,648. The additional exclusion of participants who were in school in 2009 (7% of the sample) did not affect results (20, 24–27).

Measures

Schooling information for the participants’ grandparents (G1) was retrospectively reported by the participants’ parents (G2) in categories ranging from 0 (“no education/could not read or write)” to 8 (“graduate work/professional degree”) (Web Appendix 1, available at http://aje.oxfordjournals.org/). We combined information on all 4 grandparents to create a set of indicators for a single measure of the highest schooling category ever reported for any grandparent (i.e., less than a high school degree, high school degree, some college, or college degree). Models alternatively using indicators for average grandparental schooling produced similar results, as did models using only grandmothers’ or grandfathers’ schooling. Schooling of the participants’ parents (G2) was retrospectively reported by participant (G3) heads on behalf of themselves and wives using the same categories. We created a single ordinal measure ranging from 0 to 8 of the highest schooling category ever reported for either parent. We modeled participant (G3) schooling information, drawn from the 2009 study wave and reported in years of completed schooling ranging from 1 to 17, as an ordinal variable. Alternative specifications of G2 and G3 schooling, including allowing for nonlinear associations with G3 health, did not affect results.

The outcomes were participant global health status in 2009, current smoking, and obesity. In most cases, information was provided by heads on behalf of themselves and wives. We dichotomized the 5-point health status measure into excellent, very good, or good (all termed “good”) versus fair or poor (both termed “poor”) (28–30). The PSID smoking variables pertain specifically to cigarettes. Participants were categorized as current smokers, with former and never smokers combined into a referent group. We calculated each participant's BMI value on the basis of reported height and weight and then categorized participants with BMI values of 30 or higher as obese (31).

Figure 1 shows the assumed causal structure we used to guide variable selection for adjusted models. To estimate the total effect of grandparental schooling, we estimated models adjusted for the following variables (potential confounders referred to herein as C1): age, sex, head versus wife status (among women), and whether the participant ever lived in the same state as any grandparent during the PSID. Participants were categorized as ever living in the same state as a grandparent if 1) both the participant and a grandparent were present in the PSID and living in the same state during any study wave, or 2) a parent of the participant reported having a parent living in the same state as the participant in a special module of the 1988 study wave.

Figure 1.

Causal diagram of relationships between grandparents’ (G1), parents’ (G2), and grandchildren's (G3) schooling and grandchildren's health in adulthood. The long-dashed arrow represents the direct effect of grandparental schooling on grandchildren's health. The dotted arrows represent confounders of the mediator-outcome associations that are affected by grandparental schooling. U represents unmeasured common causes of G1 schooling and C1. Although they are shown separately to facilitate interpretation of the diagram, potential confounders C1, C2, and C3 are not mutually exclusive. In our models of grandchildren's health, C1 includes participant sex, race, age, head versus wife status (“head” refers to a head of household; “wife” refers to a wife or cohabiting female partner of a head), and presence of a grandparent in the same state. C2 includes C1 factors, as well as years of parents’ births and an indicator for at least 1 parent being poor while growing up. C3 includes C2 factors, as well as mother's age and marital status at the time of the participant's birth, whether the participant lived with both parents growing up, whether the participant reported being poor while growing up, average family income-to-poverty ratio when the participant was less than 18 years of age, parental health status, parental obesity, whether a parent smoked when the participant was a child, and participant health status during childhood. Panel Study of Income Dynamics, 2009.

When estimating the direct effect of grandparental (G1) schooling independent of parents’ (G2) and participants’ (G3) schooling, we additionally included confounders of associations between these mediators and participant health outcomes (C2 and C3 in Figure 1). Potential confounders of the association between G2 schooling and G3 health included parents’ reports of poverty in childhood (as an indicator variable) and the parents’ years of birth. Additional potential confounders of the association between G3 schooling and G3 health included having an unmarried or teen mother at birth, as well as childhood circumstances (poverty, living with both natural parents, and health status). We also included the following relevant measures of parental health: either parent ever reporting fair or poor health, a parent smoking while the participant was a child, and either parent ever being obese. These data were collected prospectively from the parents; therefore, they reflect only study waves in which this information was collected (i.e., 1984–2009 for health status and 1986 and 1999–2009 for obesity and smoking). We supplemented the prospective parental smoking information with the question, “Did your/her parents smoke during your/her childhood?”, which was asked of heads on behalf of themselves and wives in 2007 and 2009.

Analysis

To handle missing information, we conducted multiple imputation with 25 imputations using the sequential regression method with IVEware software and simultaneously includied all analytical variables, as well as additional variables predictive of missingness patterns (32, 33). We assessed bivariable associations of covariates with grandparental schooling and the health outcomes using unadjusted multinomial logistic and logistic regression, respectively.

Estimation of the controlled direct effect of grandparental schooling (i.e., the long-dashed line in Figure 1) by adjusting for parents’ and participants’ schooling and the variables in C1, C2, and C3 (19, 34, 35) can produce biased estimates if the exposure affects confounders of the mediator-outcome associations (the dotted lines in Figure 1) (19, 34). This situation is likely in our context. For example, parental poverty during childhood is affected by grandparental schooling and may confound associations between parents’ schooling and participants’ health. Therefore, we estimated controlled direct effects using marginal structural models (MSMs) to account for confounders using inverse probability weighting rather than adjustment (19, 36, 37).

We first estimated a weight w_i for each observation i of the form

$$w_i=\prod _{g=1}^3\frac{\text{Pr}[S_g=s_{gi}|{\mathit{S}}_{\mathit{g}-1}={\mathit{s}}_{(\mathit{g}-1)\mathit{i}},\text{race}]}{\text{Pr}[S_g=s_{gi}|{\mathit{S}}_{\mathit{g}-1}={\mathit{s}}_{(\mathit{g}-1)\mathit{i}},{\mathit{C}}_{\mathit{g}}={\mathit{c}}_{\mathit{g}i},\text{race}]},$$

where g is the generation (G1, G2, or G3), S_g is the schooling of that generation, S_g−1 is the schooling of the previous generation(s) (S₀ = 0), and C_g is the appropriate set of confounders (C1, C2, or C3 in Figure 2). We estimated the numerator and denominator probabilities by dividing each generation's schooling into 4 categories (less than high school, high school, some college, college degree) and using multinomial logistic regression to estimate the conditional probability of the observed category.

Figure 2.

Predicted probabilities of A) poor health status, B) obesity, and C) current smoking by grandparental schooling and race. White columns represent white participants; gray columns represent black participants. Probabilities were estimated using marginal structural models adjusted for race, parents' schooling, and participant's schooling. Parents’ and participants’ schooling were both set to a high school degree and allowed to vary by race. Panel Study of Income Dynamics, 2009.

We incorporated the weights into models estimating associations of grandparental schooling with participants’ health. Because the relatively high prevalence of our outcomes would make odds ratios poor estimates of prevalence ratios, we estimated prevalence ratios with robust standard errors using Poisson regression (38, 39) as follows:

$$\begin{array}{rl}\text{ln[}P(Y_i=1|S_1& =s_{1i},S_2=s_{2i},S_3=s_{3i},\text{race)}]\\ & ={\beta }_0+{\beta }_1{\mathit{S}}_1+{\beta }_2{S}_2+{\beta }_3{S}_3+\text{race}+{\beta }_4{\mathit{S}}_1\\ & \times \text{race}+{\beta }_5{S}_2\times \text{race}+{\beta }_6{S}_3\times \text{race,}\end{array}$$

where Y_i is the outcome (poor health status, current smoking, or obesity) for person i. We used product terms between grandparental schooling indicators (S₁) and race to test the minority poverty and diminishing returns hypotheses by treating race as a modifier of the effect of grandparental schooling. We used the household identifier from the 1968 baseline PSID wave to account for correlated observations between family members (39). We tested product terms between grandparental schooling and parents’ and participants’ schooling. There was some evidence of interaction only for obesity among whites, but given the large number of terms tested and the lack of consistent evidence, we omitted the interaction terms from final models. For comparison, we also estimated direct effects in unweighted models adjusted for all confounders.

To assess whether the controlled direct effect of grandparental schooling differed by geographical proximity, we estimated MSMs, subcategorizing the indicators of grandparental schooling by whether the most educated grandparent lived in the same state. Because the additional parameters made the models less stable, we combined the “high school” and “some college” categories of grandparental schooling for this analysis.

RESULTS

Ninety-nine percent of participants had lifespans that overlapped with at least 1 grandparent, and 44% had at least 1 grandparent in the PSID sample. Missing outcome and covariate information was rare, with the exception of childhood family income and health status, which were each missing in 19% of the sample. Black participants were more likely not to have a father present in the PSID (41% vs. 10%); as a corollary, paternal grandparent information was imputed for 44% of black participants. Variable distributions in the original and imputed samples were very similar (Web Appendix 2). In sensitivity analyses, models incorporating only observed (i.e., not imputed) measures of grandparental schooling produced very similar estimates.

The average age in 2009 among the participants (G3) was 39 years; regardless of whether the person was still alive, the average age in 2009 among parents (G2) was 67 years and among grandparents (G1) was 86 years. Schooling levels were higher in each successive generation and were lower among blacks than whites in each generation (Tables 1 and 2). Black participants were more likely to have poor health status or to be obese and were less likely to be current smokers. Higher levels of grandparental schooling were associated with younger age, higher parents’ and participants’ schooling, better financial conditions across generations, and better health outcomes (Tables 1 and 2); these associations were generally stronger among whites than among blacks. More highly educated participants were less likely to report poor health status, as were participants with a grandparent who lived in the same state (Table 3). Obesity was more prevalent among women and was inversely related to parents’ and participants’ schooling among whites. Smoking was more prevalent among men and participants with low schooling.

Table 1.

Characteristics by Level of Grandparental Schooling Among 2,790 White Participants, Panel Study of Income Dynamics, 2009^a

Characteristic	Total		Grandparents' Maximum Educational Level								P Valueb
			< High School		High School		Some College		College Degree
	%	Median	%	Median	%	Median	%	Median	%	Median
Grandparents' educational level (for all participants)			15		33		26		26
Wife (vs. head)c	30		31		40		38		34		0.01
Age, years		37		45		40		35		33	<0.001
Female sex	53		49		54		54		50		0.20
Marital status											<0.001
Married	64		63		69		62		61
Never married	21		13		15		24		30
Divorced/separated/widowed	15		24		17		15		9
Income-to-poverty ratio		4.3		3.7		4.3		4.3		4.6	0.03
Participant's years of completed schooling		14		12		13		14		15	<0.001
Parents’ maximum educational level											<0.001
<High school	7		22		6		4		1
High school	28		45		35		21		15
Some college	26		18		30		32		21
Bachelor's degree or higher	39		14		30		42		63
Grandparent ever lived in same state	76		63		77		78		79		<0.001
Grandparent with maximum schooling ever lived in same state	58		63		61		54		56		0.08
Average income-to-poverty ratio when aged <18 years		3.1		2.4		3.0		3.3		3.8	<0.001
Poor while growing up	22		27		19		24		22		0.03
At least 1 parent poor while growing up	49		74		55		45		31		<0.001
Mother's year of birth		1946		1937		1944		1948		1948	<0.001
Father's year of birth		1943		1934		1941		1945		1947	<0.001
Mother unmarried at participant's birth	6		6		5		8		3		0.04
Lived with both natural parents most of time until age 16 years	73		76		76		67		72		0.003
Mother aged <20 years at participant's birth	9		13		9		10		6		0.003
Parent smoked during childhood	61		69		64		64		51		<0.001
Parent ever reported fair/poor health status (in 1984–2009)	72		84		77		70		62		<0.001
Parent ever obese (in 1986 or 1999–2009)	62		57		64		62		61		0.87
Ever reported poor health at ages 0–16 years	4		5		4		5		4		0.71
Poor health status	10		14		10		9		6		<0.001
Obese	27		36		29		26		21		<0.001
Current smoker	24		27		25		27		18		0.004

^a Uses imputed data.

^b From unadjusted multinomial logistic regression with clustering by 1968 Panel Study of Income Dynamics family.

^c “Head” refers to a head of household; “wife” refers to a wife or cohabiting female partner of a head.

Table 2.

Characteristics by Level of Grandparental Schooling Among 1,858 Black Participants, Panel Study of Income Dynamics, 2009^a

Characteristic	Total		Grandparents' Maximum Educational Level								P Valueb
			< High School		High School		Some College		College Degree
	%	Median	%	Median	%	Median	%	Median	%	Median
Grandparents' educational level (for all participants)			40		35		17		8
Wife (vs. head)c	30		21		20		18		22		0.81
Age, years		39		44		36		34		33	<0.001
Female sex	63		64		62		62		67		0.73
Marital status											0.008
Married	31		34		30		27		28
Never married	44		38		46		54		48
Divorced/separated/widowed	25		28		24		20		24
Income-to-poverty ratio		2.3		2.3		2.3		2.2		2.3	0.77
Participant's years of completed schooling		12		12		12		12		13	0.36
Parents’ maximum educational level											<0.001
<High school	19		30		14		10		10
High school	40		44		42		33		26
Some college	26		17		30		36		31
Bachelor's degree or higher	15		9		14		20		34
Grandparent ever lived in same state	72		64		75		79		83		<0.001
Grandparent with maximum schooling ever lived in same state	55		64		52		43		50		0.001
Average income-to-poverty ratio when aged <18 years		1.3		1.2		1.4		1.5		1.5	0.003
Poor while growing up	35		39		32		33		29		0.05
At least 1 parent poor while growing up	70		81		64		58		63		<0.001
Mother's year of birth		1947		1939		1951		1951		1951	<0.001
Father's year of birth		1938		1934		1940		1943		1945	<0.001
Mother unmarried at participant's birth	41		37		40		52		45		0.02
Lived with both natural parents most of time until age 16 years	52		59		51		40		45		<0.001
Mother aged <20 years at participant's birth	22		16		24		28		29		0.001
Parent smoked during childhood	58		54		63		59		62		0.10
Parent ever reported fair/poor health status (in 1984–2009)	85		90		83		81		79		0.06
Parent ever obese (in 1986 or 1999–2009)	66		67		65		66		63		0.96
Ever reported poor health at ages 0–16 years	6		6		6		5		4		0.84
Poor health status	18		19		19		16		15		0.55
Obese	42		45		41		39		38		0.29
Current smoker	27		28		26		26		25		0.91

^a Uses imputed data.

^b From unadjusted multinomial logistic regression with clustering by 1968 Panel Study of Income Dynamics family.

^c “Head” refers to a head of household; “wife” refers to a wife or cohabiting female partner of a head.

Table 3.

Poor Health Outcomes by Category of Selected Predictors by Race, Panel Study of Income Dynamics, 2009^a

Predictor	White						Black
	Poor Health Status		Obese		Current Smoker		Poor Health Status		Obese		Current Smoker
	%	P Valueb	%	P Valueb	%	P Valueb	%	P Valueb	%	P Valueb	%	P Valueb
Health outcome (for all participants)	9		27		24		18		42		27
PSID sample member		0.35		<0.001		<0.001		0.81		0.02		<0.001
Head of household	10		30		28		18		41		30
Wife	9		23		17		18		47		16
Age range, years		<0.001		0.003		0.008		<0.001		0.08		0.03
25–34	7		24		28		15		39		28
35–44	10		30		21		14		42		21
45–55	13		29		21		25		45		31
Sex		0.21		0.04		<0.001		0.24		<0.001		<0.001
Female	10		29		21		19		46		24
Male	9		25		27		17		36		33
Marital status		<0.001		0.27		<0.001		0.02		0.10		<0.001
Married	7		27		17		14		46		18
Never married	11		25		34		17		40		32
Divorced/separated/widowed	16		30		38		23		42		30
Income-to-poverty ratio		<0.001		0.02		<0.001		<0.001		0.09		<0.001
<1	28		29		52		29		40		41
1–1.9	22		34		42		23		45		34
2–4.9	9		28		25		13		44		20
≥5	4		25		15		8		35		13
Participant's maximum educational attainment		<0.001		<0.001		<0.001		<0.001		0.19		<0.001
<High school	26		35		50		26		38		53
High school	13		33		36		19		43		29
Some college	9		30		25		17		45		21
Bachelor's degree or higher	4		19		9		10		38		12
Parents’ maximum educational attainment		<0.001		<0.001		<0.001		<0.001		0.19		0.04
<12 grades	24		39		38		27		45		30
12 grades (high school)	9		31		28		17		44		28
>12 grades, no bachelor's degree	10		30		23		13		38		27
Bachelor's degree or higher	7		20		20		16		42		20
Grandparent ever lived in same state		0.01		0.92		0.82		<0.001		0.41		0.19
No	12		27		24		23		40		29
Yes	9		27		24		16		43		26
Grandparent with maximum schooling ever lived in same state		0.002		0.66		0.47		0.003		0.47		0.40
No	12		28		23		21		41		28
Yes	8		27		25		15		43		26
Average income-to-poverty ratio when aged <18 years		<0.001		<0.001		<0.001		0.26		0.18		0.33
<1	24		36		49		21		40		27
1–1.9	14		36		34		16		45		28
2–4.9	9		27		22		18		40		25
≥5	4		17		18		16		40		14
Poor while growing up		0.04		0.03		0.09		0.04		0.56		0.29
No	9		26		23		17		43		26
Yes	12		31		27		21		41		29
At least 1 parent poor while growing up		<0.001		0.18		0.08		0.06		0.82		0.20
No	7		26		23		15		42		29
Yes	12		29		26		19		42		26
Mother unmarried when born		0.004		0.001		0.001		0.34		0.18		0.14
No	9		26		23		19		43		26
Yes	17		41		36		17		40		29
Lived with both natural parents most of time until age 16 years		0.03		0.53		<0.001		0.51		0.44		0.19
No	12		28		31		19		41		29
Yes	9		27		21		17		43		26
Mother aged <20 years at participant's birth		0.31		0.001		0.03		0.54		0.58		0.85
No	9		26		24		18		42		27
Yes	11		37		30		17		43		27
Parent smoked during childhood		0.002		0.002		<0.001		0.33		0.99		0.08
No	7		24		18		17		42		25
Yes	11		30		28		19		42		29
Parent ever reported fair/poor health status (in 1984–2009)		0.44		0.45		0.94		0.16		0.10		0.77
No	7		24		23		12		35		28
Yes	11		28		25		19		43		27
Parent ever obese (in 1986 or 1999–2009)		0.45		0.13		0.87		0.68		<0.001		0.50
No	8		23		24		17		35		26
Yes	10		30		24		18		46		28
Ever reported poor health at ages 0–16 years		<0.001		0.12		0.006		0.002		0.15		0.47
No	9		30		24		17		42		27
Yes	24		34		36		29		50		30
Poor health status				<0.001		<0.001				<0.001		<0.001
No			25		23				39		25
Yes			46		37				54		36
Obese		<0.001				0.07		<0.001				<0.001
No	7				25		14				31
Yes	16				21		23				22
Current smoker		<0.001		0.07				<0.001		<0.001
No	8		28				16		45
Yes	15		24				24		34

Abbreviation: PSID, Panel Study of Income Dynamics.

^a Uses imputed data.

^b From unadjusted logistic regression with clustering by 1968 PSID family.

There was a graded estimated total effect of higher levels of grandparental schooling on health status among whites (P for trend = 0.005): prevalence ratios of poor health status compared with less than a high school degree were 0.82 (95% confidence interval (CI): 0.58, 1.15) for high school degree, 0.75 (95% CI: 0.51, 1.11) for some college, and 0.54 (95% CI: 0.36, 0.82) for college degree (Table 4). This association was not evident among blacks (prevalence ratio (PR) = 1.11, 95% CI: 0.84, 1.46) for high school; PR = 1.01 (95% CI: 0.69, 1.49) for some college; and PR = 0.97 (95% CI: 0.60, 1.56) for college degree; P for trend = 0.97). Among whites, the controlled direct effects of grandparental schooling estimated using MSMs remained graded but were attenuated relative to the estimated total effects and were consistent with the null hypothesis (PR = 0.84 (95% CI: 0.57, 1.22) for high school; PR = 0.81 (95% CI: 0.52, 1.27) for some college; and PR = 0.71 (95% CI: 0.44, 1.15) for college dgree; P for trend = 0.20). The direct effects estimated using unweighted adjusted regression were smaller than the MSM estimates.

Table 4.

Prevalence Ratios of Poor Health Status, Obesity, and Current Smoking by Grandparental Schooling (vs. Less Than High School) and Race, Panel Study of Income Dynamics, 2009

Estimate Type	Educational Level						P for Trenda
	High School		Some College		College Degree
	PR	95% CI	PR	95% CI	PR	95% CI
Poor Health Status
White
Total effectb	0.82	0.58, 1.15	0.75	0.51, 1.11	0.54	0.36, 0.82	0.005
MSM direct effectc,d	0.84	0.57, 1.22	0.81	0.52, 1.27	0.71	0.44, 1.15	0.20
Adjusted direct effecte	0.91	0.64, 1.29	0.90	0.59, 1.37	0.76	0.49, 1.20	0.28
Black
Total effectb	1.11	0.84, 1.46	1.01	0.69, 1.49	0.97	0.60, 1.56	0.97
MSM direct effectc,d	0.98	0.73, 1.32	1.15	0.78, 1.68	1.09	0.68, 1.74	0.52
Adjusted direct effecte	1.16	0.87, 1.54	1.06	0.71, 1.57	1.03	0.63, 1.69	0.72
Obesity
White
Total effectb	0.86	0.70, 1.06	0.80	0.64, 1.00	0.65	0.51, 0.82	<0.001
MSM direct effectc,d	0.85	0.68, 1.05	0.81	0.64, 1.03	0.73	0.56, 0.94	0.02
Adjusted direct effecte	0.92	0.74, 1.15	0.90	0.71, 1.15	0.79	0.61, 1.02	0.08
Black
Total effectb	0.93	0.77, 1.12	0.94	0.73, 1.21	0.92	0.68, 1.23	0.47
MSM direct effectc,d	0.91	0.74, 1.11	0.92	0.68, 1.23	1.02	0.74, 1.39	0.77
Adjusted direct effecte	0.92	0.76, 1.11	0.92	0.71, 1.20	0.91	0.67, 1.24	0.43
Current Smoking
White
Total effectb	0.98	0.78, 1.24	1.00	0.79, 1.28	0.71	0.55, 0.92	0.01
MSM direct effectc,d	0.95	0.75, 1.19	0.99	0.77, 1.28	0.78	0.58, 1.04	0.16
Adjusted direct effecte	1.07	0.84, 1.36	1.13	0.87, 1.45	0.88	0.66, 1.17	0.41
Black
Total effectb	0.86	0.69, 1.08	0.84	0.63, 1.12	0.81	0.56, 1.17	0.13
MSM direct effectc,d	0.97	0.75, 1.25	0.89	0.62, 1.27	0.71	0.45, 1.13	0.19
Adjusted direct effecte	0.87	0.69, 1.09	0.85	0.63, 1.15	0.84	0.57, 1.22	0.41

Abbreviations: CI, confidence interval; MSM, marginal structural model; PR, prevalence ratio.

^a P value for coefficient when 4 categories of grandparental schooling are modeled as a single ordinal variable.

^b Adjusted for sex, race, age, head versus wife status (“head” refers to a head of household, and “wife” refers to a wife or cohabiting female partner of a head), and ever having a grandparent living in the same state.

^c MSM adjusted for race, parent schooling, and participant schooling.

^d P values for joint test of race interaction terms in model with categories of grandparental schooling and for single interaction term in trend model (i.e., single ordinal variable for grandparental schooling (not shown)), respectively, were 0.54 and 0.02 for poor health status, 0.47 and 0.22 for obesity, and 0.91 and 0.78 for current smoking.

^e Adjusted for sex, race, age, head versus wife status, ever having a grandparent living in the same state, parents’ schooling, participants’ schooling, having a parent who grew up poor, whether participant reported growing up poor, average family income-to-poverty ratio before age 18 years, unmarried mother at participant's birth, teenaged mother, whether lived with both parents growing up, parental health status, parental obesity, parental smoking during participant's childhood, and childhood health status.

The results for obesity followed the same pattern as those for poor health status (Table 4). Among whites, the total effect estimates showed a graded inverse association between grandparental schooling and obesity (P for trend < 0.001). Estimated direct effects were slightly smaller and were more attenuated when using unweighted adjusted models than when using MSMs. The MSM direct-effect prevalence ratios among whites were 0.85 (95% CI: 0.68, 1.05) for high school, 0.81 (95% CI: 0.64, 1.03) for some college, and 0.73 (95% CI: 0.56, 0.94) for college degree (P for trend = 0.02). As with health status, the estimated effects of grandparental schooling on obesity among blacks were minimal.

Among whites, only having a grandparent with a college degree was inversely associated with current smoking (Table 4; estimated total effect PR = 0.71 (95% CI: 0.55, 0.92); MSM PR = 0.78 (95% CI: 0.58, 1.04)). Unlike the other outcomes, associations for smoking were more consistently graded among blacks than among whites, although 95% confidence intervals spanned the null; estimated total-effect prevalence ratios were 0.86 (95% CI: 0.69, 1.08) for high school, 0.84 (95% CI: 0.63, 1.12) for some college, and 0.81 (95% CI: 0.56, 1.17) for a college degree (P for trend = 0.13). Direct-effect MSM prevalence ratios among blacks were 0.97 (95% CI: 0.75, 1.25) for high school, 0.89 (95% CI: 0.62, 1.27) for some college, and 0.71 (95% CI: 0.45, 1.13) for a college degree (P for trend = 0.19).

Figure 2 shows predicted probabilities for the 3 outcomes by categories of grandparental schooling, calculated from the direct-effect MSM with parents’ and participants’ schooling set to a high school degree. Probabilities of poor health status and obesity were higher among blacks, and the grandparental schooling gradient evident among whites was absent. In contrast, the predicted probabilities of current smoking were similar between the racial groups, and in both groups, the probability of smoking was lower among participants who had a grandparent with a college degree.

Figure 3 shows prevalence ratios for MSM estimates of controlled direct effects of each category of grandparental schooling (compared with less than high school) stratified by whether the most educated grandparent ever lived in the same state as the participant. Among both blacks and whites, the inverse association between grandparental schooling and poor health status was more pronounced when the most educated grandparent lived in the same state (Figure 3A). In fact, point estimates among blacks were in the unexpected direction when the most educated grandparent lived in another state, although estimates were imprecise. The pattern of results was similar for obesity, but differences based on geographical proximity were smaller (Figure 3B). For smoking, estimates for having a grandparent with a college degree were larger in magnitude when the grandparent lived in a different state, especially among blacks (Figure 3C).

Figure 3.

Direct-effect prevalence ratios of A) poor health status, B) obesity, and C) current smoking for grandparental schooling (vs. less than high school (HS)) by whether the highest-educated grandparent ever lived in the same state as the participant and race. White markers represent the highest-educated grandparent never having lived in the same state as the participant; black markers represent the highest-educated grandparent having lived in the same state. From marginal structural models adjusted for race, parents’ schooling, and participants’ schooling. P values for the null hypothesis that the estimates are equal for living in same or different state and for for maximal educational levels of high school/some college or college degree, respectively, are A) 0.01 and 0.15 among whites and 0.02 and 0.43 among blacks; B) 0.08 and 0.88 among whites and 0.45 and 0.21 among blacks; and C) 0.82 and 0.74 among whites and 0.96 and 0.12 among blacks. Panel Study of Income Dynamics, 2009.

DISCUSSION

In a national sample of US blacks and whites who were 25–55 years of age in 2009, higher levels of grandparental schooling were consistently related in a graded manner with better health status, less current smoking, and less obesity among whites. There was also evidence among whites suggesting direct effects of grandparental schooling that were not mediated by parents’ and participants’ schooling, particularly for obesity. This evidence was stronger for poor health status among participants whose most highly educated grandparent lived in the same state. Among blacks, the only association suggesting a total or direct effect of grandparental schooling was for smoking. Despite the relative imprecision of our estimates and possible residual bias, these results are consistent with those of limited past research linking higher levels of grandparental schooling to better health of grandchildren and may have implications for research documenting multigenerational family histories of socioeconomically patterned diseases (16, 17, 40, 41).

The stronger associations we observed among whites for health status and obesity support the diminishing returns hypothesis rather than the minority poverty hypothesis and mirror findings in other national samples (20, 42, 43). In contrast, our estimates for smoking were similar among black participants and white participants. This may reflect mechanistic differences in how grandparental schooling relates to health across different outcomes or temporal differences in race-specific prevalence rates of different outcomes. Currently in the United States, poor health status and overweight are more prevalent among blacks than whites, whereas smoking prevalence is approximately equal between the groups (15). Early in the 20th century, whites were more likely to smoke than blacks, but this pattern reversed in midcentury; whites have also historically been more likely to quit smoking than blacks (44). The results here may reflect differences in both smoking initiation and cessation (44).

The differences we observed depending on whether the most educated grandparent lived in the same state as the participant provide preliminary evidence that social contact with grandparents—facilitated by geographical proximity—may play a role in how grandparental schooling influences grandchildren's health. For health status and obesity, the estimated health effects of grandparental schooling were larger when the most educated grandparent lived in the same state as the participant, whereas the opposite was true for smoking. One possibility is that, given secular changes in education-related smoking gradients, college-educated grandparents living in the same state may have, in fact, been more likely to smoke in the presence of their grandchildren, leading to a dampening of the beneficial influence of their schooling. Pertinent limitations of our analysis included the proxy measure of social contact and lack of information about the duration of contact with grandparents and grandparental health and health behaviors.

Additional research is needed to identify mechanisms specific to different outcomes. Such research may also further address how grandparents influence grandchildren's health when, like most US grandparents, they do not live in the same household as their grandchildren; much past research has focused on grandparents who live with or serve as primary caretakers of their grandchildren (45). More generally, it may be fruitful to consider MSMs in future research addressing life-course socioeconomic influences, where the problem they address—variables that are affected both by the exposure and confounders of mediator-outcome associations—is common (19, 36).

Our measure of maximum level of grandparental schooling implied several simplifying assumptions that warrant further investigation; it did not account for differences by lineage, sex, or other aspects of grandparental schooling, such as minimum schooling or variation between grandparents (46). Neither did it account for whether grandparents were related by birth or the possible presence of additional grandparent figures, such as after a remarriage. Approximately 3% of our sample had at least 1 adoptive parent in the PSID. Our test of interactions between grandparental schooling and parents’ and participants’ schooling may have been insufficiently powered, which is a challenge because the rarest combinations—those with large intergenerational schooling differences—may be particularly important for health (47). Our analyses were also subject to assumptions of no residual confounding, correct model specification, and positivity (19, 48).

The reliance on self- and proxy-reported data, along with substantial missing exposure data, may have resulted in misclassification bias, particularly with respect to the measures of grandparental and parental schooling. Distributions of health measures in the PSID are similar to national estimates (49, 50). Although multiple imputation cannot address possible bias induced when missingness is associated with unmeasured data (i.e., we could not correct for bias when the missing data mechanism is missing not at random), it can minimize the bias induced when missingness is associated only with measured data (i.e., when the missing data mechanism is missing at random) (51). Measurement error may have also contributed to the imprecision of our estimates. Another concern is attrition (52), which has been substantial despite high per-wave response rates (23), although attrition bias in models of adult socioeconomic and health outcomes in the PSID may be limited (52). We had limited ability to address birth cohort differences, which may be important because of strong secular trends in schooling, obesity, and smoking. We addressed this by restricting the age range of our sample and accounting for the parents’ birth years; grandparents’ birth year information was missing for the majority of observations. In a sensitivity analysis, overall associations between higher levels of grandparental schooling and better health outcomes persisted when we estimated separate models for participants aged 25–34, 35–44, and 45–55 years.

The identification of multigenerational effects of schooling on health may have important intervention and policy implications. If our health is affected not only by our own schooling but also by that of our parents and grandparents, the benefits we gain depend not only on our own educational opportunities but on those conferred to previous generations. Therefore, policies to reduce inequities in education may also serve to reduce health disparities in future generations. By the same token, failure to reduce current inequities in educational opportunities may contribute to health disparities in future generations.