Intergenerational Mobility in Self-Reported Health Status in the US

Abstract

We present estimates of intergenerational mobility in self-reported health status (SRHS) in the US using data from the PSID. We estimate that the rank-rank slope in SRHS is 0.26. We show that including both parent health and income in models of intergenerational mobility increases the explanatory power of child outcomes. We construct a monetary metric for health and then use this to combine income and health into a measure of welfare and estimate the rank-rank slope to be about 0.4 for this new measure. Finally, we document striking health mobility gaps by race, region and parent education.

1. Introduction

A large and growing multi-disciplinary literature on intergenerational mobility has emerged in recent decades primarily motivated by concerns over equality of opportunity. Most of these studies have focused on income, education or occupation. However, one key aspect of socioeconomic status, general health status, has been relatively under-studied.¹ This is unfortunate since health is an especially important component of welfare (Jones and Klenow, 2016). For example, longevity, which depends in large part on health, is clearly a powerful barometer of lifetime utility. A large literature has also highlighted how poor health early in life leads to reduced educational attainment, worse labor market outcomes, and onset of chronic disease later in life (e.g. Case et al., 2005; Aizer and Currie, 2014). In addition, health, especially at later ages is fundamental for decisions related to work, retirement, consumption, and savings (e.g. Rust and Phelan, 1997; Palumbo, 1999; French and Jones, 2017).

Studying intergenerational mobility with respect to overall health status, however, is a formidable task. First, it requires panel data containing broad-based health measures for adults in two generations, which is difficult to obtain. Second, since the concept of interest is latent, health is inherently difficult to measure. Morbidity measures or anthropometric indicators such as height, weight and BMI are typically blunt proxies for a more fundamental underlying latent variable. Third, long-run latent health status might not be revealed until relatively later in life when variability in organ function becomes more pronounced (Steves et al., 2012), chronic diseases begin to emerge, and functional abilities increasingly become impaired.

We address these issues by using the Panel Study of Income Dynamics (PSID). The PSID is the world’s longest running longitudinal dataset. It tracks individuals as they form new households and has been widely used to study intergenerational mobility.² Since 1984, the PSID began collecting information on self-reported health status (SRHS). SRHS has long been established in the epidemiology literature as a valid omnibus health measure that is highly predictive of mortality, even when compared to clinical measures such as chronic illnesses (e.g. Miilunpalo et al. 1997, Idler and Benyamini, 1997 and DeSalvo et al. 2005). SRHS has specifically been validated in the PSID using proprietary mortality files, where the measure predicts mortality even after controlling for baseline demographic characteristics (Halliday, 2014). Importantly, to our knowledge, the PSID has collected data on SRHS for longer than any other longitudinal dataset.

We construct an intergenerational sample of parents and their adult children using all available information on health status when individuals are at least 30 years old. We employ a method used by the National Center for Health Statistics to convert SRHS to a continuous measure that is akin to a quality adjusted life year or “QALY” (see Erickson et al. 1995).³ Following the income mobility literature, we use time averages of the QALY to proxy for lifetime health status. We view time averaging as a method for extracting a time invariant latent variable. The use of health reports at multiple points in time, and at different points of the lifecycle, enables us to overcome the key obstacles to studying intergenerational health mobility.

Our first mobility measure is the Intergenerational Health Association (IHA), which is the coefficient on parent health status from a regression of child health on parent health (adjusting for age). This provides a measure of the persistence in health status that is similar to the intergenerational income elasticity. We also construct a variety of rank-based measures and run intergenerational “rank-rank” regressions as popularized by Chetty et al (2014). We estimate the slope of this regression, the “rank-rank slope,” as well as measures of the expected rank for a child whose parents were at the 25^th and 75^th percentiles of the parent health distribution to provide estimates of upward and downward mobility.

Conceptually, the IHA represents the degree of intergenerational persistence in health units as defined by the QALY, whereas the rank-rank slope captures persistence in ranks. In principle, these could differ if, for example, the magnitude of differences in health are larger at some points in the health distribution than at other points. There are two benefits of using ranks. First, they allow us to compare persistence parameters across different domains of socioeconomic status such as income and health. Second, in a companion paper (Halliday et al., 2020), we show that the rank-based estimators also appear to be more robust to measurement error.

Our estimates of the IHA range from 0.20 to 0.25 with a preferred estimate of 0.23 when we pool men and women and combine both parents. This implies that an additional year of quality life among parents is associated with close to three additional months of healthy life for children. This suggests that there is a modest degree of persistence in health status in the US. We estimate the rank-rank slope (Spearman correlation) to be between 0.21 to 0.29 with a preferred estimate of 0.26. We further find that the expected rank of children starting at the 25^th percentile is about the 44^th percentile suggesting considerable upward mobility. The expected rank of children starting at the 75^th percentile is the 57^th percentile which also suggests that there is considerable downward mobility.

In our companion paper (Halliday et al., 2020), we obtain similar estimates for rank-based measures of intergenerational health mobility when using a more sophisticated model of latent health that we estimate via Bayesian methods. In contrast, we find that the estimates of the IHA may be biased down by as much as 15% using the linear methods used in this paper. However, an advantage of the linear model used here is that the methodology is very transparent and tractable and allows us to consider the interplay between income and health and to explore various aspects of heterogeneity. In any event, our overall finding of an intergenerational association of around 0.2 is roughly consistent with estimates of intergenerational associations in birth weight, longevity and mental health which are also around 0.2.⁴

Although SRHS is highly validated and widely used, it is nonetheless a subjective measure. To address this concern, we combine a set of 21 more objective, but still self-reported, health measures (e.g. arthritis, heart disease, difficulty walking) that have been collected in the PSID since 1999 to construct an alternative health index (AHI). For a subset of our sample, we compare estimates using SRHS to those using the AHI. Remarkably, we find that the results are extremely similar, further confirming that SRHS appears to be a valid measure of intergenerational health mobility.⁵

We also consider how the joint distribution of income and health evolves over a generation. We do this by adding parent income rank to the regression of child health rank on parent health rank and by adding parent health rank to the regression of child income rank on parent income rank. We show that including both dimensions of socioeconomic status of parents improves the explanatory power of these regressions, demonstrating the independent value of these measures.

Having established the independent importance of both income and health, we then create a monetary metric for health and then use this to combine income and health into one overall measure of welfare. We show that the rank-rank slope in this combined measure is about 0.43 which is higher than the association in either measure taken on its own.

The finding that income and health may capture somewhat distinct aspects of socioeconomic status is one reason why the degree of health mobility might differ from income mobility, as we find for the US. Another possible explanation is institutional factors. For example, it may be the case that the US provides greater insurance for the possibility of poor health outcomes than poor labor market outcomes. A third explanation is measurement error, we may just do a better job of measuring income than health. As more studies of both intergenerational income mobility and intergenerational health mobility are completed in various countries using other measures of health status (including administrative data) we may arrive at a better of understanding of the degree to which income mobility differs from health mobility.

We also analyze heterogeneity in health mobility by race, region and parent education level using rank-rank regressions. We generally find that the same groups that are disadvantaged with respect to income mobility are also disadvantaged with respect to health mobility. We find that Blacks experience both lower upward mobility and higher downward mobility.⁶ We also show that the gaps in expected health rank are even more pronounced when comparing individuals by their parents’ education level. This suggests that the well-known gradient in health by education levels extends to the subsequent generation. We also find that those born in the South experience both lower upward mobility and higher downward mobility.

2. Literature Review

In this section, we discuss previous related work and how our paper fits with the current literature. A growing literature has examined intergenerational mobility with respects to various measures related to health. However, most of these papers focus on highly specific aspects of health status like anthropomorphic measures or specific health-related behaviors. In contrast, we are interested in a broader measure of overall health status that can be thought of as an aggregation of measures that reflect these specific components. There are only a few studies like ours, that examine an omnibus measure of health status.

Several studies have examined intergenerational persistence in birth weight (e.g. Currie and Morretti, 2007; Black et al. 2007; Giuntella et al., 2019) which could be thought of as a precursor to health later in life. These studies have typically found relatively modest intergenerational associations. For example, Currie and Morretti (2007) estimate that the intergenerational elasticity in birth weight is in the range of 0.17–0.20. Other work on anthropomorphic measures include Bhalotra and Rawlings (2013), Classen (2010) and Akbulut-Yaksel and Kugler (2016). Classen (2010) estimates an intergenerational correlation of around 0.35 in BMI. One notable study by Johnston et al. (2013) has examined mental health and estimates an intergenerational association in mental health that is between 0.13 and 0.19.

Other studies have estimated intergenerational transmission in health-related behaviors. For example, Loreiro et al. (2010), and Darden and Gilleskie (2016) have examined smoking and Schmidt and Tauchmann (2011) have analyzed alcohol consumption. These studies have also highlighted an important gender-specific transmission to engagement in health-related behaviors. This motivates our focus on also looking at gender differences.

Among studies that have used a broader measure of health many have focused on longevity. There is a rather old literature that has estimated intergenerational correlations in life spans (see Beeton and Pearson 1901, Ahlburg 1998, and Yashin and Iachine 1997, Lach et al., 2008, Hong and Park, 2016, Kaplanis, et al. 2018, for example). This literature tends to find intergenerational correlations on the order of 0.15–0.30. However, it is important to note that these estimates require observing the completed life spans of two generations, which is typically only feasible with very long running historical administrative data sets that are not often readily available to researchers. These estimates by their very nature may not be indicative of health transmission in the modern period or for more advanced economies.

While longevity is certainly a key component of human welfare, it is a relatively blunt measure and does not measure the quality of a person’s health when they are alive. In contrast, health quality is often measured using quality adjusted life-years or QALY’s. We estimate intergenerational persistence in a measure of QALY’s derived from self-reported health status (SRHS).

Two papers written after ours, build on our methodology and estimate intergenerational persistence in QALY’s.⁷ Fletcher and Jajtner (2019) use the National Longitudinal Study of Adolescent to Adult Health or Add Health for the US and estimate the intergenerational health association (IHA) to be about 0.17. They also consider regional variation and heterogeneity by population subgroups. While Fletcher and Jajtner (2019) have a much a larger sample, their sample of adult children is considerably younger, and they use shorter time averages of health. As we discuss later in section 5.2, these aspects of measurement can attenuate the estimates. Graeber (2020) builds on our empirical approach and estimates intergenerational associations for Germany using the German Socio-Economic Panel. Graeber (2020) finds a rank-rank association of about 0.23, which is similar to what we find for the U.S. Graeber (2020) also considers nonlinearities and how associations differ based on family background.

Another recent paper by Andersen (2019) uses rich Danish administrative health records and a principal components analysis to estimate intergenerational health associations. Anderson estimates range from 0.1–0.2. An open question for future research is how comparable estimates based on administrative data are with those based on household surveys. An obvious advantage of administrative data is that it can avoid bias from self-reporting. However, administrative data on morbidities requires individuals to select into receiving care and there can be many institutional factors that can affect this propensity. In addition, collecting information on specific morbidities using ICD10 codes may not account for the severity of a given condition. For example, the severity of a diabetes or arthritis diagnosis may vary from person to person. In general, researchers must rely on measures that are based on administrative goals rather than questions that can be specifically tailored for research purposes such as activities of daily living. Another advantage of survey data is that it enables us to show health mobility varies across population subgroups as administrative data often does not collect detailed demographic information such as race, ethnicity, or other aspects of family background.

Estimating intergenerational health mobility using QALY’s based on SRHS also facilitates cross-national comparisons. The data on SRHS for two generations are available in countries such as the US, Germany, the United Kingdom and Indonesia. While studies have compared income mobility in different settings (e.g. Bratberg et al. 2017), we still know little about how health mobility varies across countries.

Finally, a separate strand of this literature has attempted to disentangle the roles of genetic and environmental factors in the intergenerational transmission of health. Thompson (2017) estimates the intergenerational persistence of asthma in the United States using data on adoptees and finds suggestive evidence that environmental factors play a relatively more important role in its transmission in poorer households. Bjorkegren, et al. (2019) use Swedish administrative data on adoptees to estimate intergenerational transmissions in life expectancy. They find most of the transmission takes place via biological parents. Notably, Classen and Thompson (2016) also found similar results for body mass index. This is another important area for future work that requires unique data.

3. Data

We use the Panel Study of Income Dynamics (PSID). The PSID is a U.S. longitudinal household survey that began in 1968 with a nationally representative sample of over 18,000 individuals living in 5,000 families. Including the original and subsequent samples, over 70,000 people have participated in the survey. Extensive information is collected on a wide range of outcomes including employment, income, wealth, childhood development, and education. Individuals in the PSID families and anyone subsequently born to or adopted by a sample person are followed over time even if they form separate family units. The unique design of the PSID allows us to link adult children to their parents across survey waves.

Starting in 1984, the PSID included questions on the health status of household heads and their spouses. Specifically, they asked, “Would you say your health in general is excellent, very good, good, fair, or poor?”⁸ This question, commonly referred to as self-reported health status (SRHS), is highly predictive of mortality even after controlling for other health measures and outperforms other objective health measures (see Miilunpalo et al. 1997; Idler and Benyamini, 1997, DeSalvo et al. 2006, and Halliday, 2014). However, as a robustness check, we supplement our analysis by constructing an alternative health index (AHI) using 21 objective self-reported health measures available in survey years beginning in 1999. Details on the AHI are described in Section III and in Appendix B.

We construct a sample of 8,115 men and women who are at least 30 years old, provide SRHS in at least one survey year, and who are matched to at least one parent who also provides SRHS at least once.⁹ We collect all values of SRHS between 1984 and 2013 for each person and, following Johnson and Schoeni (2011) who also used the PSID, convert the categorical values into a continuous measure using health utility-based scale developed for the Health and Activity Limitation Index (HALex). This approach is designed to estimate the percentage of a year that is considered to be of quality health, or a “quality adjusted life year” (QALY).¹⁰ The value ranges for each health status category are as follows: excellent is [95,100]; very good is [85,95); good is [70,85); fair is [30,70); and poor is [1,30). We assign the midpoint of the interval for each reported health category in each year and then average these values over all available years for each individual.

There are several reasons we chose to convert the ordered SRHS variable into a QALY. First, it is not so straightforward to interpret intergenerational relationships in terms of the five ordered categories and there is no established tradition for this. In contrast, converting SRHS to a continuous measure allows us to use traditional methods such as regression to the mean and ranks to better convey the intergenerational dynamics in terms comparable to the previous literature. Second, the QALY measure itself is easy to interpret since it tells us the portion of a year spent in quality health which may be easier to understand than being in “good” health or very good” health. Third, in health economics there is tremendous value in quantifying the impacts of health in economic terms by converting health into a monetary metric which we can operationalize with the QALY. Fourth, using a QALY also allows us to combine income with health and pursue a more general welfare analysis that is in the spirit of Jones and Klenow (2016).

In Figure 1, we plot the mean health status over the life cycle pooling all individuals in both generations. The dashed lines indicate the +/− 1 standard deviation in health at each age. The figure shows that health is roughly flat from age 30 to 40 but then begins to decline roughly linearly through age 80.¹¹ To address this lifecycle pattern and in order to compare individuals at different ages we also construct a regression adjusted measure of health status.¹² Moreover, the widening of the standard deviation in health status at each age suggests that there is considerably more variation in health as an individual ages. This is consistent with greater variation in organ function (Steves et al. 2012) and the rising onset of chronic diseases at later ages. This suggests that health status is more indicative of latent health at age 60 than at age 40. It is also consistent with the well-known fact that inequality in general tends to increase as cohorts age. Deaton and Paxson (1994) provide evidence for consumption; Deaton and Paxson (1998) and Halliday (2011) provide evidence for numerous health measures including SRHS.

Figure 1: Health status over life cycle.

Figure 1 plots the mean continuous health measure at each age up to 85 for the full sample. The dashed line represents +/− 1 SD. It includes all generations and genders. The mean at each age is weighted using the most recently available individual weights. The scale reflects the lower cutoffs between reported health status categories on the 0–100 HALex scale where 100 equals perfect health and zero is equivalent to death: [95,100] is excellent, [85,95) is very good, [70,85) is good, [30,70) is fair and [1,30) for poor health. The continuous health measure for each individual at a given survey year is the midpoint of the interval corresponding to their reported health category.

In addition, we collect data on total family income which includes all taxable income (e.g. earnings, interest and dividends) and cash transfers for all family members measured in 2013 dollars deflated using the CPI-U. We adjust for family size by dividing by the square root of the number of family members. We also average income over all available years. For race, we use the reported race of the child. To measure educational attainment, we use the last available report on years of completed education. Finally, region is based on the child’s most often reported region of residence before the age of 18.

For our analysis of early life influences, we use a subsample of 3,281 adults in the 2013 PSID who were also part of the Childhood Retrospective Circumstance Study (CRCS). The CRCS collects data from household heads and spouses on their childhood and young adulthood experiences. Topics include parental relationship quality, childhood health, socioeconomic status, neighborhoods, friendships, school experiences, relationship quality with parents/guardians and young adult mentoring. For some categories we create indices by taking the largest component from a principal components analysis (PCA).¹³

In Table 1 we present summary statistics for our main sample (using sampling weights). Panel A shows the characteristics of parents. The mean age is around 56 and the average of years of education is between 12 and 13. Fewer than 10 percent report that their health is excellent. On average, our sample contains about 15 years of data on health status. Panel B shows that the children are on average 38 to 39 years old with about 14 years of education. Well over half report being in very good or excellent health.

Table 1:

Summary Statistics

A. Parents
		Father	Mother
		(1)	(2)
Age		56.72	56.17
		(10.48)	(11.04)
Years of Education		12.96	12.51
		(3.10)	(2.68)
Total Family Income (2013 Dollars)		59405.94	50318.97
		(50472.40)	(45929.52)
Overall Health Status		77.37	75.73
		(17.08)	(16.60)
Excellent		7%	4%
Very Good		35%	30%
Good		34%	39%
Fair		22%	25%
Poor		2%	2%
Years of Health Measurement (Min=1, Max=22)		14.81	15.48
Number of Observations		5,440	7,721
Number of Observations (CRCS)		2,425	3,151
B. Children
	All	Sons	Daughters
	(3)	(4)	(5)
Age	38.54	38.68	38.41
	(6.02)	(6.09)	(5.94)
Years of Education	13.96	13.85	14.06
	(2.25)	(2.30)	(2.20)
Total Family Income (2013 Dollars)	54303.96	56973.13	51636.98
	(46086.89)	(45849.59)	(46174.67)
Overall Health Status	82.60	83.38	81.83
	(13.50)	(13.51)	(13.44)
Excellent	11%	13%	9%
Very Good	44%	45%	43%
Good	32%	30%	34%
Fair	12%	11%	14%
Poor	1%	1%	1%
Race
White	83%	85%	81%
Black	14%	13%	16%
Other	3%	3%	3%
Childhood Region
Northeast	22.4%	22.1%	22.7%
North Central	28.1%	28.7%	27.6%
South	31.8%	31.7%	32.0%
West	17.4%	17.2%	17.6%
Alaska and Hawaii	0.1%	0.1%	0.0%
Foreign Country	0.2%	0.3%	0.1%
Years of Health Measurement (Min=1, Max=22)	8.7	8.5	8.8
Number of Observations	8,115	3,828	4,287
Number of Observations (CRCS)	3,281	1,407	1,874
C. CRCS Variables
	All	Sons	Daughters
	(6)	(7)	(8)
Family Socioeconomic Background
SES Index Age 0–5	0.00	0.01	−0.01
	(1.00)	(1.01)	(1.00)
SES Index Age 6–12	0.00	0.01	−0.01
	(1.00)	(1.00)	(1.00)
SES Index Age 13–16	0.00	0.05	−0.04
	(1.00)	(0.95)	(1.04)
Neighborhood Quality Index	0.00	−0.02	0.02
	(1.00)	(1.02)	(0.98)
Childhood Health
Child Health Index	0.00	0.08	−0.07
	(1.00)	(0.93)	(1.05)
Underweight at 13	0.06	0.06	0.06
	(0.24)	(0.23)	(0.24)
Overweight at 13	0.17	0.22	0.12
	(0.38)	(0.42)	(0.33)
Obese at 13	0.12	0.14	0.09
	(0.32)	(0.35)	(0.29)
Childhood Stability
# Times Moved in Childhood	1.04	1.06	1.02
	(1.81)	(1.82)	(1.80)
# Schools Attended Before 17	3.35	3.26	3.42
	(1.71)	(1.70)	(1.71)
Parents Satisfied with Relationship	0.72	0.75	0.70
	(0.45)	(0.43)	(0.46)
Parents Ever Divorced	0.13	0.13	0.14
	(0.34)	(0.33)	(0.34)
School Experience
# Times Repeat School Grade	0.13	0.17	0.10
	(0.45)	(0.44)	(0.45)
School Experience Index Age 6–12	0.00	−0.15	0.13
	(1.00)	(1.00)	(0.98)
School Experience Index Age 13–16	0.00	−0.13	0.11
	(1.00)	(1.02)	(0.97)
Childhood Relationship
Friendship Quality Index Age 6–12	0.00	0.01	−0.01
	(1.00)	(0.95)	(1.04)
Friendship Quality Index Age 13–16	0.00	0.03	−0.03
	(1.00)	(0.97)	(1.03)
Relationship with Mother Quality Index	0.00	0.11	−0.10
	(1.00)	(0.90)	(1.07)
Relationship with Father Quality Index	0.00	0.02	−0.02
	(1.00)	(0.97)	(1.03)
Had Mentor Age 17–30	0.65	0.63	0.67
	(0.48)	(0.48)	(0.47)

Table 1 provides descriptive statistics of the data. Panel A and B reports the summary statistics for the main sample from the 1984–2013 survey years of the Panel Study of Income Dynamics (PSID). This sample includes only individuals who are matched to at least one parent. Across both generations, only individuals with at least one health status observation measured at age 30 and older are included. Panel C reports the summary statistics for the individuals in the child generation who were also part of the 2013 Childhood Retrospective Circumstance Study (CRCS). Age refers to the mean time-averaged age of the individual at the time of all available health observations. Years of education is the mean total years of education attained reported at most recently available survey. Total family income reported in 2013 dollars is the mean time-averaged total family income, which includes all taxable income and cash transfers for all family members after adjusting for family size and inflation. Overall health status is the time-averaged of all available health observations after converting the ordinal health status into continuous units on a 0–100 scale. The categories of health status (excellent, very good, good, fair, poor) are the percentage of individuals whose time-averaged overall health status is in that category according to the HALex scale. Years of health measurement refers to the mean number of total years of health observations for each individual. The race categories refer to the percentage of the sample that identifies with that race in most recently available survey. Childhood region categories refer to percentage of the sample that grew up in that region, defined as the modal region in which the household is surveyed before the child is 18. For CRCS variables (Panel C), all index variables are reported in original units and are constructed using PCA across the full CRCS sample. Details on the index construction and all other CRCS variables can be found in Appendix A. Standard deviations are reported in parentheses. All reported means and standard deviations are weighted using the most recently available individual sampling weight. For the CRCS variables, means and standard deviations are weighted using the individual CRCS sampling weight.

In Panel C, we report summary statistics for the CRCS sample. We report the statistics for indices in standardized units.¹⁴ We break down the CRCS childhood experience variables into the following categories: family socioeconomic background, childhood health, childhood stability, school experience, and childhood relationships. See Appendix A for more detail.

4. Methodology

Intergenerational Health Association (IHA)

Many studies in the income mobility literature have estimated the intergenerational elasticity or “IGE”. We start by creating an analogous measure, which we refer to as the intergenerational health association (IHA). The IHA is based on estimating the following regression:

$$y_{\mathit{\text{1i}}}=\alpha +\beta y_{\mathit{\text{0i}}}+\gamma {\mathit{X}}_{\mathit{i}}+{\epsilon }_i$$ (1)

where conceptually, y_1i represents the lifetime health of the child in family i, and y_0i is the lifetime health of one or both of the parents. The vector X is a set of controls and includes the quadratic age terms for both the parent(s) and the child. The parameter β provides a measure of intergenerational persistence and 1 – β is a measure of mobility. In our case, y measures the percentage of a healthy life year in which a value of 100 denotes one year in perfect health and 0 denotes a health state that is viewed as equivalent to death. If, for example, β is 0.2, this implies that if the difference in health between two families in the parent’s generation was 10 percent of a healthy year, then we would expect the difference in health to be only about 2 percent of a healthy year in the children’s generation. In this case, most health differences between families dissipate in a generation, so that the rate of regression to the mean is quite high. In contrast, if β is 0.8, we would consider health to be highly persistent so that there is low degree of mobility. Our preferred estimates combine the health status of both parents (when available) by using an average of the time averages of each parent and using just a single parent’s health measure when only one parent is linked to a child. We also report gender-specific estimates because as noted in the literature review, previous studies on health-related behaviors showed a gender-specificity in the degree of intergenerational transmission. Standard errors are clustered by family.

Rank Mobility Measures

While the IHA, like the IGE, is useful for characterizing the rate of regression to the mean in one simple parameter, it is not ideal for all purposes. In particular, when comparing subgroups of the population (i.e. differences by race and region) relative to a common distribution, one may prefer to use rank-based measures (Mazumder, 2016). Rank-based measures are also better suited for distinguishing upward and downward mobility patterns. We calculate the percentile rank of age-adjusted health separately for each gender in each generation.¹⁵ In addition to percentile ranks for each parent, we also construct a “both parents” measure that uses all available health observations from both parents and combines them into a single rank.¹⁶ Similarly, we also construct an “all children” rank that pools together the age-adjusted child health measures for sons and daughters. We then estimate regressions¹⁷ of the following form:

$$r_{\mathit{\text{1i}}}=\alpha +\rho r_{\mathit{\text{0i}}}+{\epsilon }_i$$ (2)

where r₁ and r₀ now represent the percentile rank of health in each respective generation. In this case, ρ provides an estimate of persistence in rank position and 1− ρ provides a measure of positional mobility. We will often refer to ρ, which is equivalent to the Spearman correlation as the “rank-rank slope.” In principle, β and ρ can differ. It could be for example, that if the health distribution becomes more compressed in the child distribution than it was in the parent generation, then a given amount of rank mobility could be more consequential in terms of health as measured by years of quality life.

In addition to estimates of rank persistence, we use the rank-rank regression framework to calculate expected ranks at the 25^th and 75^th percentile. These estimates at “p25” and “p75” convey information about “directional” (upward or downward) mobility for a typical child coming from lower and higher health families.¹⁸ For example, if the expected health rank of individuals coming from the 25^th percentile is the 45th percentile then this would suggest upward mobility of about 20 percentiles.¹⁹ We also construct a parallel set of income rank measures.²⁰ When analyzing subgroups (e.g. region, race, education), we calculate ranks based on the full population enabling us to make mobility comparisons with respect to the national distribution.²¹

Alternative Health Index (AHI)

As a robustness check, we develop an alternative health index (AHI) that is constructed from more objective self-reported health measures that are only available in survey years after 1999.²² In total, we compile 21 indicators of adverse mental and physical health conditions that take on the value of 1 if the individual has the health condition and 0 otherwise. These measures, of course are not perfect. The measures do not necessarily reflect the severity of the condition and individuals may have to select into receiving medical care in order to receive an official diagnosis. Details on the individual conditions can be found in Appendix B. We construct a simple index using the fraction of the conditions that the individual does not have so that a higher index value will indicate better health. We then take the time average of an age-adjusted AHI over all available years between 1999 and 2013 for each individual. We can then compare estimates of intergenerational health associations and rank-rank slopes based on the AHI to a similar set of estimates based on our health measure where we use the identical sample of individuals and restrict our SRHS data to reports from 1999–2013.

A Welfare Measure that Combines Income and Health

Finally, we construct a measure of welfare that combines both income and health into one overall measure of welfare.²³ In order to do this we first construct a monetary metric for health. We follow Finkelstein et al. (2019) who take the consensus value of a statistical life year to be $100,000 based on Cutler (2004). Therefore, our monetary metric of health assigns a value of $100,000 to each quality adjusted life year (QALY) in each generation. We then sum this value with our income measure to create our welfare measure. We then convert this to ranks and estimate a regression analogous to (2) using our combined measured. We also estimate a version of (1) using log of the combined measure of health and income.

5. Results

We first present our main estimates of intergenerational health mobility in Section 5.1. We then consider the robustness of these baseline results to measurement issues in Section 5.2. We next explore the joint distribution of health and income mobility in Section 5.3 and our combined welfare measure in Section 5.4. We then measure how health mobility differs across subgroups of the population in Section 5.5. In Section 5.6, we provide suggestive evidence on how health mobilty has evolved over the past decades. Finally, in Section 5.7, we consider potential mechanisms that can explain our results on intergenerational health mobility.

5.1. Intergenerational Health Mobility

Basic Descriptive Patterns

Before presenting our main estimates, we start by showing some simple associations between parent and child health in Appendix Table A.1 that are easy to interpret. We convert the time averages of our continuous health measure for each individual back into the original five SRHS categories using the scale described in Section II. We find that if both parents (or one parent in the case of single parent families) are in at least good health, then children are 10.9 percentage points more likely to report being in at least good health compared to children whose parents were not in good health.²⁴ This differs somewhat by gender. Sons are 11.8 percentage points more like to be in good health when their parents are in good health compared to a 9.9 percentage estimate for daughters.

We explore this association along two further dimensions in Table 2. First, we separately examine the health associations of children with mothers versus fathers. Second, we investigate how associations differ among parents by different categories of the SRHS variable: good, very good and excellent health. We find that relative to having a mother in fair or poor health, having a mother in exactly good health increases the likelihood that a child will be in at least good health by 10.9 percentage points (column 1). Having a mother in very good or excellent health increases the association even further to about 16 percentage points. The estimates are fairly similar for sons and daughters as shown in columns (2) and (3). Columns 4 through 6 show the comparable estimates when we examine the estimates for fathers’ health on all kids, sons, and daughters. Compared to mothers, there appears to be a slightly lower association between fathers and children.

Table 2:

Probability of child in at least good health conditioned on mother or father’s health status

	Mother’s Health			Father’s Health
	All	Sons	Daughters	All	Sons	Daughters
	(1)	(2)	(3)	(4)	(5)	(6)
Mother’s Health Excellent	0.159	0.179	0.136
	(0.0249)	(0.0269)	(0.0414)
Mother’s Health Very Good	0.152	0.144	0.160
	(0.0163)	(0.0223)	(0.0222)
Mother’s Health Good	0.109	0.0955	0.121
	(0.0166)	(0.0232)	(0.0220)
Father’s Health Excellent				0.145	0.166	0.122
				(0.0222)	(0.0299)	(0.0326)
Father’s Health Very Good				0.122	0.123	0.120
				(0.0184)	(0.0279)	(0.0229)
Father’s Health Good				0.106	0.107	0.103
				(0.0181)	(0.0268)	(0.0236)
Constant	0.571	0.517	0.632	0.344	0.599	0.103
	(0.205)	(0.289)	(0.287)	(0.263)	(0.306)	(0.361)
Observations	7,606	3,600	4,006	5,376	2,596	2,780
R-squared	0.048	0.052	0.046	0.039	0.037	0.043
Y-mean	0.871	0.884	0.859	0.895	0.900	0.890

Each column of Table 2 reports the coefficients and standard errors from a weighted regression using sampling weights of the most recently available individual weights for the child. The dependent variable for all specifications is an indicator variable that takes on the value of 1 (and 0 otherwise) if the child’s time-averaged continuous health measure is in good, very good or excellent health according to the HALex scale. The omitted category for all regressions is parent (mother or father) health in poor or fair health. All specifications include as controls the quadratic age terms of the parent (mother or father) and quadratic age terms of the child. Age for both generations are defined as the time-averaged age of the individual at the time of all available health observations. Columns 1 and 4 report the results using all children. Columns 2 and 5 report the results using sons only. Columns 3 and 6 report the results using daughters only. Y-mean refers to the weighted mean of the dependent variable within the regression sample. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation.

Estimates of Intergenerational Health Mobility

In Table 3 we show the estimates of the intergenerational health association (IHA) for various parent-child groups. We find that when we combine both parents’ health for the pooled sample of sons and daughters (column 1) we obtain an estimate of 0.23. In terms of years of quality life, the estimate implies that for every additional year of quality life the parents have, the child, on average, is expected to have almost three additional months (23% of a year) of healthy life. This is higher than either using only mother’s health (0.20) or using father’s health (0.17).²⁵ Note that the estimates that combine the health status of both parents necessarily take averages over a larger number of health measures. So, the estimates in the third row that combine both parents may be higher because they do a better job extracting the “signal” from the health measures of the parents. We find roughly similar patterns if we look either at sons (column 2) or daughters (column 3). Both sons and daughters’ health are more strongly associated with mother’s health than with father’s health and the highest estimates arise when pooling both parents’ health. The associations appear to be slightly higher from parents to daughters than sons.

Table 3:

Intergenerational health associations by parent-child samples

	All Children	Sons	Daughters
	(1)	(2)	(3)
Mother’s Health Only	0.204	0.200	0.206
	(0.019)	(0.023)	(0.025)
Father’s Health Only	0.172	0.165	0.181
	(0.017)	(0.023)	(0.025)
Both Parents’ Health	0.229	0.218	0.238
	(0.020)	(0.024)	(0.025)
Y-Mean	73.2	73.46	72.94
Observations	7987	3763	4224

Each cell of Table 3 reports the coefficient and standard error on the parent health measure from a separate regression. The regressions are weighted using sampling weights of the most recently available individual weights for the child. The dependent variable for all specifications is the child’s time-averaged continuous health measure. The main explanatory variable for specifications using mother’s health or father’s health is the parent’s time-averaged continuous health measure. For regressions using both parents’ health, the parent health measure is the average of the mother’s and father’s health if available. Otherwise, only one parent’s health measure is used. All specifications include as controls the quadratic age terms of the parent (mother or father) and quadratic age terms of the child. Age for both generations is defined as the time-averaged age of the individual at the time of all available health observations. In specifications using both parents’ health, quadratic age terms of the mother and father are included separately. If the individual is missing health observations from one of the parents, the quadratic age terms for that parent is replaced with a 0. Two indicator variables, one for mother and one for father, are included that take on the value of 1 (and 0 otherwise) if that parent is missing. Column 1 reports the results using all children. Column 2 reports the results using sons only. Column 3 reports the results using daughters only. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation. Y-mean refers to the weighted mean of the dependent variable within the regression sample using both parents health for that column. Observations is the number of observations in the regression sample using both parents health for that column.

In a companion paper, Halliday et al. (2020), we estimate a non-linear model of SRHS in which health is determined by a latent variable. Using Bayesian econometric techniques, we find somewhat higher estimates of the IHA. For example, when looking at the transmission from all parents to all children, we obtained an IHA of 0.29 (see Appendix Table A.2).

We next turn to estimates of rank mobility. In Figure 2A, we show the binned scatterplot of expected health rank at every percentile of the parent health distribution for all children in our sample. The relationship is almost linear. For every 10-percentile rank increase of the parents, the child is expected to be 2.61 percentiles higher in the health distribution of their own generation. In panel A of Table 4, we show estimates of the rank-rank slopes, expected rank at 25^th and 75^th percentile for all parent-child combinations. Estimates for the rank-rank slopes for the different subsamples range from 0.21 to 0.29. In our companion paper, Halliday et al. (2020) we found virtually identical rank-rank slopes when we directly modeled SRHS in a latent variable framework. Similar to the intergenerational health association results, we find the strongest association between mothers and daughters. We also find that the associations are larger when using mothers than when using fathers. Both sons and daughters have similar expected ranks when the mother (or father) is at the 25th percentile of the parent health distribution with estimates ranging from the 44^th to the 47^th percentile. Expected rank at the 75^th percentile is also similar across the samples, ranging from 56 to 60^th percentile, though, daughters appear to experience less downward mobility than sons.

Figure 2: Health and income rank mobility using both parents’ health for all children.

Figure 2 Panel A plots the mean child health percentile rank at each percentile of the parent health distribution using both parents’ health for all children. Panel B plots the mean child income percentile rank at each percentile of the parent income distribution using both parents’ income for all children. The red line in each graph is the estimated regression line from the weighted bivariate regression of child rank on parent rank. The rank-rank slope is the coefficient on parent income percentile. The expected rank at the 25th (or 75th) percentile is the predicted rank from the rank-rank specification for a child with a parent at the 25th (or 75th) percentile of the parent health or income rank distribution. Health percentile ranks are constructed from the age-adjusted health measure and are ranked separately for each generation. Income percentile ranks are constructed from the time-averaged total family income adjusted for age, family size and inflation and are ranked separately for each generation. All means and regressions are weighted using the most recently available individual sampling weights of the child. Standard errors for the regression coefficients (in parentheses) are robust to heteroskedasticity and within-family correlation.

Table 4:

Health and income rank mobility by parent-child samples

A. Health Rank Mobility
	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile	Observations
	(1)	(2)	(3)	(4)
Mother-Son	0.243	44.72	56.847	3564
	(0.025)	(0.933)	(0.979)
Mother-Daughter	0.287	44.137	58.472	3960
	(0.022)	(0.827)	(0.900)
Father-Son	0.212	47.116	57.706	2520
	(0.028)	(1.113)	(1.071)
Father-Daughter	0.251	47.426	60.001	2689
	(0.025)	(0.992)	(0.995)
Both Parents-All Children	0.261	44.342	57.402	7937
	(0.017)	(0.644)	(0.688)
B. Income Rank Mobility
	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile	Observations
	(5)	(6)	(7)	(8)
Mother-Son	0.447	39.508	61.872	3564
	(0.024)	(0.900)	(0.951)
Mother-Daughter	0.473	39.935	63.58	3960
	(0.021)	(0.771)	(0.882)
Father-Son	0.406	43.495	63.785	2520
	(0.029)	(1.102)	(1.098)
Father-Daughter	0.417	44.284	65.129	2689
	(0.024)	(0.943)	(0.987)
Both Parents-All Children	0.393	40.766	60.439	7937
	(0.018)	(0.684)	(0.690)

Each row of Table 4 reports the rank-rank slope, expected ranks at the 25th and 75th health (Panel A) or income (Panel B) percentile and number of observations for each parent-child sample. The rank-rank slope is the coefficient on parent health or income percentile from the bivariate regression of child rank on parent rank. The expected rank at the 25th (or 75th) percentile is the predicted rank from the rank-rank specification for a child with a parent at the 25th (or 75th) percentile of the parent health or income rank distribution. All regressions are weighted using the most recently available sampling weight of the child. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation.

Income mobility estimates for the identical samples are shown in panel B. The corresponding binned scatterplot for the full sample is shown in Figure 2B. The estimates for the rank-rank slopes range from 0.41 to 0.47, implying greater persistence in income rank than in health rank.²⁶ Comparing the estimates for the expected rank at p25 and p75 in panels A and B shows that there is less upward mobility from the bottom, and less downward mobility from the top, when using income compared to using health.

Alternative Health Index (AHI)

In Table 5, we compare mobility estimates based on self-reported health status (SRHS) to estimates based on the Alternative Health Index (AHI) using identical samples.²⁷ Whether we focus on the IHA or rank mobility estimates, we find that the estimates are remarkably similar across the two measures; we find no evidence of systematic downward bias in using SRHS which might have been expected if SRHS contained greater measurement error than the AHI. For example, the IHA estimates range from 0.091 to 0.199 when using SRHS and range from between 0.092 to 0.184 when using the AHI. Note that the estimates in this table tend to be lower than the estimates of health persistence in the previous two tables because they employ shorter averages of the health measures, which is a point that we will come back to shortly. This is not surprising given that we find that the two measures are highly correlated, with correlation coefficients ranging from 0.66 to 0.76 depending on the generation we use. Overall, this analysis suggests that SRHS is at least as informative of latent health as measures based on questions about very specific health conditions.

Table 5:

Health mobility measures using alternative health index (1999–2013 sample)

A. Intergenerational Health Associations
	Post-1999 Self-Reported Health Status			Alternative Health Index
	All Children	Sons	Daughters	All Children	Sons	Daughters
	(1)	(2)	(3)	(4)	(5)	(6)
Mother’s Health Only	0.171	0.162	0.179	0.171	0.156	0.184
	(0.017)	(0.025)	(0.022)	(0.015)	(0.021)	(0.022)
Father’s Health Only	0.114	0.091	0.14	0.094	0.092	0.094
	(0.017)	(0.021)	(0.025)	(0.017)	(0.020)	(0.026)
Both Parents’ Health	0.179	0.157	0.199	0.165	0.157	0.171
	(0.017)	(0.025)	(0.022)	(0.016)	(0.021)	(0.024)
Y-Mean	69.85	69.84	69.86	0.85	0.85	0.84
Observations	5162	2415	2747	5162	2415	2747
B. Rank Mobility
	Post-1999 Self-Reported Health Status			Alternative Health Index
	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile
	(7)	(8)	(9)	(10)	(11)	(12)
Mother-Son	0.188	45.946	55.351	0.243	44.373	56.528
	(0.027)	(1.054)	(1.042)	(0.026)	(1.051)	(0.964)
Mother-Daughter	0.258	44.29	57.212	0.244	44.701	56.915
	(0.025)	(0.961)	(0.952)	(0.025)	(0.986)	(0.958)
Father-Son	0.142	49.656	56.732	0.169	47.995	56.432
	(0.030)	(1.168)	(1.187)	(0.030)	(1.204)	(1.086)
Father-Daughter	0.219	47.591	58.523	0.145	47.999	55.267
	(0.029)	(1.165)	(1.094)	(0.030)	(1.157)	(1.170)
Both Parents-All Children	0.212	45.505	56.092	0.227	45.065	56.398
	(0.019)	(0.718)	(0.741)	(0.018)	(0.716)	(0.695)

Table 5 reports the intergenerational health associations and rank-rank slopes using only individuals with health observations at age 30 and older from 1999–2013. The Post-1999 Self-Reported Health Status is time-averaged continuous health measure analogous to baseline health measure using only data from survey years 1999–2013. The Alternative Health Index is the time-averaged fraction of 21 adverse health conditions that the individual does not have. Details on the Alternative Health Index is provided in Appendix B. Each cell of Panel A reports the coefficient and standard error on the parent health measure from a weighted regression of child health on parent health. Specifications in Columns 1 to 3 use the Post-1999 Self-Reported Health Status as the health measure for both parent and child generations. Columns 4 to 6 use the Alternative Health Index as the health measure for both parent and child generations. Y-mean refers to the weighted mean of the dependent variable within the regression sample using both parents’ health for that column. Observations is the number of observations in the regression sample using both parents’ health for that column. See notes to Table 3 for additional details on the intergenerational health association specifications. Each row of Panel B reports the rank-rank slope, expected ranks at the parent 25th and 75th health percentile and number of observations each parent-child sample. Columns 7 to 9 use the Post-1999 Self-Reported Health Status to construct percentile ranks for both parent and child generation separately for each gender. Columns 10 to 12 use the Alternative Health Index to construct percentile ranks for each parent and child generation separately for each gender. See notes to Table 4 for additional details on rank-rank specifications. All regressions are weighted using the most recently available sampling weight of the child. Standard errors for all regressions (in parentheses) are robust to heteroskedasticity and within-family correlation.

5.2. Measurement Issues

Attenuation Biases and Time Averaging

Prior research has emphasized the importance of addressing measurement error/transitory shocks and lifecycle biases in producing accurate estimates of intergenerational associations in lifetime income (e.g. Jenkins, 1987; Solon, 1992; Mazumder, 2005; Grawe, 2006; Haider and Solon, 2006; Mazumder, 2016). Longer time averages of parent income have been shown to reduce attenuation bias. We analyze whether this is also the case in the context of health by following the same approach. In order to avoid having the composition of the sample change as we use longer-time averages, we hold the sample size fixed by requiring parents to report health in some minimum number of years (either 5, 7, 10 or 15 years). In all cases we keep the time average of the child’s health measure fixed by using all available years.

The results for the rank-rank slopes of using longer time averages of parent health for daughters are shown in panels A and B of Figure 3 and the analogous figures for sons are shown in panels A and B Appendix Figure A7. We find that increasing the number of years in the time average of either the father or the mother leads to progressively higher estimates that plateaus once we have a time average of about 8 years. For example, in Panel A of Figure 3 when using the sample where mothers’ health status was observed for at least 15 years, the estimates of the rank-rank slope between mothers and daughters increase from 0.21 to 0.31 as we increase the length of the time average from 1 year to 8 years. The increase is even larger when we examine the father-daughter relationship in Panel B, where the estimates increase from 0.17 to 0.29 over the same range. Although the attenuation bias appears to differ somewhat by parent-child combination, we find that time averaging is important for rank-rank slopes in health. This contrasts with the case of income where Mazumder (2016) and Nybom and Stuhler (2016) have found that rank-rank slopes are more robust to measurement error and transitory fluctuations.

Figure 3: Robustness of rank-rank slopes.

Figure 3 evaluates the robustness of the estimates of rank-rank slopes to attenuation and life cycle biases. Panels A and B plot the rank-rank slopes using varying time averages of mother (Panel A) and fathers (Panel B) health within fixed samples of daughters with parents with at least 5, 7, 10, or 15 years of health observations. The number of observations for each fixed sample is reported in parentheses. Panels C and D plot the rank-rank slopes using child’s health observations within the 10-year age bins and all available parent health observations over age of 30. In all specifications in Figure 3, the rank-rank slopes are estimated from weighted bivariate regressions of child health rank on parent health rank using the most recently available individual sampling weights of the child. Age adjustment and percentile ranks are done separately for each alternative parent and child health measure.

In Appendix Figure A.6, panels A and B, we plot analogous graphs for the IHA where we estimate the models pooling sons and daughters. We see a similar pattern of rising estimates as we increase the length of parent time averages. When looking at the association between fathers and children in panel B of Figure A6, the increase in the IHA nearly doubles from 0.11 when using a single year of fathers’ health to 0.20 when using a ten-year average (for the sample where fathers’ health is observed for at least 15 years). These findings suggest that as is the case with income and occupation (Mazumder and Acosta, 2015) it is critical to use long time averages of health status to measure the IHA. These findings are also broadly consistent with the results in Halliday et al. (2020) which directly estimates the IHA using the categorical SRHS data in a latent variable econometric framework rather than first converting the SRHS to QALY and then using time averages.

Life Cycle Bias

We next consider how the estimates differ depending on the age at which health is measured. For each parent and child, we take an average of all available years in the following age bins: 30–39, 40–49, 50–59, and 60–69. For the rank-based estimates, percentile ranks are calculated separately for each age bin.

We start by focusing on how the age range of children affects estimates of health persistence. The life-cycle bias stemming from child age of measurement has been a key focus of the income mobility literature (e.g. Haider and Solon, 2006). In panels C and D of Figure 3 we show how estimates of the rank-rank slope vary as we increase the age range of sons or daughters, while using all available years of information irrespective of the age of parents. When we look at rank-rank slopes with respect to mothers’ health in panel C of Figure 3, we see a mixed pattern. For sons, the estimates rise from 0.21, when sons are between 30 and 39, to 0.27 when sons are between 60 and 69. However, this increase is not statistically significant. The estimates of the rank-rank slope between mothers and daughters in contrast is very flat with estimates of 0.27 for 3 of the 4 age ranges of daughters. The estimate is slightly smaller at 0.25 when daughters are between 60 and 69. In panel D of Figure 3, when we use fathers’ health, the results appear to be more consistent between sons and daughters. In this case, the rank-rank slopes are much higher when children are between the ages of 60 and 69 with point estimates of 0.30 for sons and 0.35 for daughters. However, these estimates are quite noisy.

We explore the effects of child age on estimates of the IHA in panels E and F of Appendix Figure A.6. In these figures we pool sons and daughters together, so the estimates are a little bit less noisy. We find that there is again some suggestive evidence that estimates rise as we increase the age at which children’s health is measured. This is particularly true in the case of the IHA between fathers and children, where the estimates are as low as 0.14 when children are between 30 and 39 but rise as high as 0.39 when children are between 60 and 69. Looking across the rank-rank estimates and the IHA, there appears to be suggestive evidence that higher estimates are obtained when children’s health is measured at a later age. This pattern is most notable when we use fathers’ health as our measure in the parent generation.

We next turn to an analysis of how parent age affects estimates of health persistence. The results for the rank-rank slope are shown in panels C and D of Appendix Figure A.7. We find that estimates appear to be slightly lower when parents’ health is measured between the ages of 30 and 39 but roughly constant at later ages. The results for the IHA are shown in panels C and D of Appendix Figure A.6. Interestingly here the estimates appear to decline as parents’ health is measured at later ages.

There are two general points worth making here. First, the standard errors are generally too large to find statistically significant differences across these different age groupings. Second, when we restrict the sample based on the ages in one generation, we may mechanically also alter the age composition of the other generation.²⁸

In Appendix Table A.4 we show how the rank-rank slope and IHA estimates vary across all combinations of child and parent age. This also allows us to address the issue of compositional bias. These estimates tend to be even noisier so we are hesitant to draw any firm conclusions about how the age structure of the data may affect our estimates. Nevertheless, there is some suggestive evidence that lifecycle biases may be present and that the highest estimates are obtained when both parents’ and children’s health are measured later in the lifecycle after the age of 50. This stands in contrast to the income mobility literature where IGE estimates tend to have the least bias when parents and children’s income are measured closer to mid-career.²⁹

Sample Attrition and Death

A common concern with using longitudinal survey data is sample attrition. If attrition is non-random, then estimates based on the available sample may be non-representative of the population of interest. There is mixed evidence of this with respect to the PSID. For example, Schoeni and Wiemers (2015) find that measures of intergenerational income persistence in the PSID are likely to be downward biased. Fitzgerald (2011) finds that attrition does not affect intergenerational persistence estimates and also extends the analysis to measures of health. Our main method of addressing attrition is to use the sampling weights provided by the PSID. We use the last available survey weight for the child as we expect this to have adjusted for all of the factors until the latest time period that could have led to sample attrition. To address the robustness of our estimates to the choice of weights, in Appendix Table A.15 we compare our main estimates to those without using any weights, to using the first available weight, and to using the average weight. It does not appear that our choice of weights leads to vastly different estimates.

Of particular concern is the possibility that a parent’s health observation could be missing because of mortality. In our main estimates we only use the recorded health for those individuals who are alive.³⁰ To address this we build an alternative sample in which we truncate parents’ age at 70.³¹ For families in which (at least) one parent survives past age 70, we simply computed the average of QALY’s up until age 70. For families in which both parents died prior to age 70, we imputed zeros for the number of years prior to age 70 that both parents were dead as the QALY’s for these years. Note that the QALY for death typically is zero. We then averaged the QALY’s through age 70 as before.

The results are shown in Appendix Table A.13. The first two rows show the estimates for this alternative subsample. Our estimate of the IHA is 0.220 and of the rank-rank slope is 0.276 for this subsample. When we impute missing values due to deaths with zeroes, the estimates are lowered to 0.219 and 0.240, respectively. Therefore, we do not think the failure to account for mortality significantly alters our conclusions.

5.3. Interplay between Income and Health

In this section, we examine the joint evolution of health and income across generations. Such an analysis provides a richer understanding of the unique roles of these two aspects of parental socioeconomic status in explaining child outcomes. For example, does parent income play a role distinct from parent health, in explaining child health? Similarly, is there a role for parent health in explaining child income conditional on parent income? To address this, we re-estimate a version of (2) for the pooled sample of sons and daughters where we now include the parent income rank as an explanatory variable in addition to parent health rank.³² Similarly, we run a version of (2) where our dependent variable is child income rank and where we include both parent income rank and parent health rank as right hand side variables. These results are shown in Table 6 and Figure 4.

Table 6:

Interplay of Health and Income Mobility

	Child Health Rank			Child Income Rank
	(1)	(2)	(3)	(4)	(5)	(6)
Parent Health Rank	0.261		0.207	0.272		0.126
	(0.017)		(0.019)	(0.018)		(0.019)
Parent Income Rank		0.224	0.125		0.393	0.333
		(0.019)	(0.020)		(0.018)	(0.020)
Constant	37.812	39.313	34.260	37.314	30.929	27.839
	(0.973)	(1.107)	(1.116)	(1.030)	(1.037)	(1.128)
Observations	7937	7937	7937	7937	7937	7937
R2	0.075	0.050	0.087	0.080	0.154	0.168

Table 6 reports estimates from regressing child health rank (columns 1–3) or child income rank (columns 4–6) on parent health rank and parent income rank. The parent health rank is constructed from the age-adjusted both parents health measure. The child health rank is constructed from the pooled age-adjusted child health measure for sons and daughters. All regressions are weighted using the most recently available sampling weight of the child. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation.

Figure 4: Interplay of health and income rank mobility using both parents’ health for all children.

Figure 4a plots the expected child health rank on parents’ health rank after residualizing for parents’ income rank. Figure 4b plots the expected child health rank on parents’ income rank after residualizing for parents’ health rank. Figure 4c plots the expected child income rank on parents’ health rank after residualizing for parents’ income rank. Figure 4d plots the expected child income rank on parents’ income rank after residualizing for parents’ health rank.

Columns (1) and (5) show the same univariate relationships that we showed in the bottom rows of the panels in Table 4, where the rank-rank slope in health is 0.26 and in income is 0.39. In columns (2) and (4) we show the cross-relationships. The rank-rank slope between parent income and child health is 0.22 and between parent health and child income is 0.27. These cross-relationship associations are quite large and reflect in part the correlations between parent income and health. In columns (3) and (6) we put both parent variables together into the equation simultaneously. The cross-relationships decline and are now both about 0.125. The rank-rank slopes conditional on including the other parent variable both decline by about 0.05 to 0.06 relative to the univariate model. For example, including parent income rank lowers the rank-rank slope in health to 0.207 from 0.261. However, what is important to notice is that R-squared from the models that include both parent variables rise by about 10 to 15 percent.

These results suggest that parent income and parent health are capturing somewhat distinct aspects of parent status that matter for children’s outcomes. The inclusion of both parent income and health meaningfully increases the explanatory power of the intergenerational models. Thus, we clearly learn more by including both measures than by simply using one.

5.4. Intergenerational Mobility in a Welfare Measure that Integrates Income and Health

In Table 7 we present our results when we combine income and health together using our monetary metric for health described in the previous section. We consider this integrated measure as a useful proxy for a broader measure of welfare. In column (1) we simply take logs of this measure for both children and parents and produce an elasticity, while in column (2) we calculate percentile ranks in each generation and estimate the rank-rank slope. The elasticity in our “welfare” measure is 0.37 and our rank-rank slope estimate is 0.43.

Table 7:

Intergenerational Mobility in “Welfare”, Integrating Income and Health

	(1)	(2)
	Log(Welfare)	Welfare Rank
Log(Parents’ Welfare)	0.368
	(0.020)
Parents’ Welfare Rank		0.429
		(0.017)
Constant	7.484	29.572
	(0.239)	(0.968)
Observations	7987	7937
R2	0.183	0.189

All regressions are weighted using the most recently available sampling weight of the child. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation.

The estimate for the rank-rank slope in our welfare proxy is higher than the analogous estimate when we only focus on health (0.26) or income (0.39). We speculate that the estimate would be even higher if we had an ideal sample covering the entire lifetime income stream of both parents and children (Mazumder, 2016) so we do not view this estimate as the final word on persistence in overall welfare. Nevertheless, these results suggest that integrating both income and health yields more comprehensive estimates of intergenerational mobility than looking at only one domain.

5.5. Mobility by Subpopulations

We now use our rank mobility measures to describe how health mobility varies across different subgroups of the population. For this analysis, we pooled sons and daughters and combined the health of both parents.³³ Figure 5 plots the predicted percentile of the child’s health rank at each percentile of the parent’s health rank from rank-rank regressions. In Table 8, we report the associated rank-rank slopes and the expected ranks at 25^th and 75^th percentile by childhood race, region and parent’s education level. For comparisons of subgroups we focus attention on the conditionally expected ranks since this tells us how groups differ with respect to the overall distribution.³⁴

Figure 5: Health rank mobility by race, region and education.

Figure 5 plots estimated regression lines from the weighted bivariate regressions of child rank on parent rank by race, childhood region, and education using both parents’ health for all children. Race refers to the reported race of the child. Region refers to the region the child grew up in, defined as the modal region in which the household is surveyed before the child is 18. Education refers to the highest level attained by at least one of the parents in the most recently available survey. The rank-rank slope, denoted by β, is the coefficient on parent health percentile. The expected rank at the 25th (or 75th) percentile, denoted by p25(p75), is the predicted rank from the rank-rank specification for a child with parents at the 25th (or 75th) percentile of the parent health rank distribution. Health percentile ranks are constructed from the age-adjusted baseline health measure and are ranked separately by generation. All regressions are weighted using the most recently available individual sampling weights of the child. Standard errors for the regression coefficients (in parentheses) are robust to heteroskedasticity and within-family correlation.

Table 8:

Health and income rank mobility by race, region, and education

	Health Mobility			Income Mobility
	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile	Rank-Rank Slope	Expected Rank at 25th Percentile	Expected Rank at 75th Percentile	Observations
	(1)	(2)	(3)	(4)	(5)	(6)	(7)
Race
White	0.243	46.501	58.665	0.352	44.499	62.096	4555
	(0.021)	(0.806)	(0.733)	(0.020)	(0.815)	(0.716)
Black	0.130	36.849	43.337	0.265	27.957	41.226	3139
	(0.034)	(1.039)	(1.780)	(0.058)	(1.358)	(2.502)
Test of Equality P-Value	0.004	0.000	0.000	0.157	0.000	0.000
Region
Northeast	0.250	45.781	58.306	0.367	46.588	64.926	1073
	(0.041)	(1.573)	(1.534)	(0.044)	(1.758)	(1.521)
North Central	0.230	45.805	57.297	0.381	41.675	60.743	1896
	(0.033)	(1.225)	(1.225)	(0.033)	(1.167)	(1.339)
South	0.254	42.137	54.835	0.408	37.255	57.651	3181
	(0.031)	(1.087)	(1.357)	(0.030)	(1.134)	(1.232)
West	0.276	43.864	57.646	0.344	41.323	58.51	1020
	(0.044)	(1.854)	(1.554)	(0.044)	(1.752)	(1.591)
Test of Equality P-Value	0.864	0.095	0.321	0.649	0.000	0.002
Education
Less than HS	0.204	36.925	47.114	0.261	30.269	43.313	2245
	(0.046)	(1.209)	(2.548)	(0.048)	(1.037)	(2.657)
HS Degree	0.197	45.596	55.447	0.26	43.708	56.721	4206
	(0.023)	(0.807)	(0.939)	(0.026)	(0.892)	(0.989)
College Degree	0.202	51.801	61.891	0.3	51.623	66.648	1471
	(0.042)	(2.005)	(1.063)	(0.039)	(2.012)	(0.961)
Test of Equality P-Value	0.989	0.000	0.000	0.678	0.000	0.000

Each row of Table 8 reports the rank-rank slope, expected ranks at the 25th and 75th health (Columns 1–3) or income (Columns 4–6) percentile and number of observations (Column 7) by subgroups for all children. The parent health (income) rank is constructed from the age-adjusted both parents health (income) measure. The child health (income) rank is constructed from the pooled age-adjusted child health (income) measure for sons and daughters. Race refers to the reported race of the child. Region refers to the region the child grew up in, defined as the modal region in which the household is surveyed before the child is 18. Education refers to the highest level of education attained by at least one of the parents in the most recently available survey. All regressions are weighted using the most recently available sampling weight of the child. Standard errors for the regressions (in parentheses) are robust to heteroskedasticity and within-family correlation. P-values from F-tests on the equality of the rank-rank slopes, expected ranks at the 25th and 75th percentiles within each category (region, race, or education) are reported.

We begin by documenting that health mobility also differs substantially by race. We find that blacks experience both lower upward mobility from the bottom and higher downward mobility from the top. Panel A of Figure 5 shows that at every point in the parent health distribution, the expected rank of black children is substantially lower than the expected rank of white children. While whites with parents at the 25^th health percentile are expected to reach the 47^th percentile, blacks with parents at the same health percentile are expected to reach only the 37^th percentile. This mobility gap continues to increase throughout the parent rank distribution with blacks expected to experience higher rates of downward mobility than whites. The expected rank at the 75^th percentile is almost 15 percentiles lower than for whites.

The racial gaps in income mobility are even more pronounced.³⁵ While whites with parents at the 25^th percentile of the income distribution are expected to reach the 45^th percentile, blacks are only expected to reach the 28^th percentile, nearly 17 percentiles lower. Therefore, black-white difference in expected rank at the 25^th percentile in income (in absolute value) is therefore 7 percentiles more than the black-white difference in expected rank at the 25^th percentile in health. In panel A of Appendix Figure A.17, we plot these black-white “mobility gaps” in health and income throughout the parent distributions of health and income.

We further explore whether these racial gaps also differ by gender. In Figure 6, we visually depict the rank-rank slopes and expected rank at the 25^th percentile for different race by gender groups using bar graphs. The top two panels use a pooled sample of sons and daughters for a baseline comparison. Panel A shows that the rank-rank slope is higher for whites than blacks. This means that rank persistence is greater for whites and hence mobility by this metric is lower for whites than blacks. Panel B shows what we previously discussed, namely that the expected rank at the 25^th percentile is lower for blacks than whites for both income and health.

Figure 6: Health and income rank mobility by race and gender (using both parents’ health).

Figure 6 plots the rank-rank slope coefficients and expected rank at the 25th percentile from the weighted bivariate regressions of child rank on parent rank by race and gender. All panels use both parent’s health or income measure. Race refers to the reported race of the child. The outcome variable for income rank-rank specifications is household income or individual labor income, as specified. The rank-rank slope is the coefficient on parent health percentile. Health percentile ranks are constructed from the age-adjusted health measure and are ranked separately for each generation. Family income percentile ranks are constructed from the time-averaged total family income adjusted for age, family size and inflation and are ranked separately for each generation. Individual income percentile ranks are constructed from time-averaged total individual labor income adjusted for age and inflation. All regressions are weighted using the most recently available individual sampling weights of the child. Standard errors for the regression coefficients are robust to heteroskedasticity and within-family correlation.

In Panels C and D we now break out these gaps separately for sons and daughters. For the rank-rank slopes in household income, we continue to see that rank persistence is much higher for white men than black men but that the racial differences in rank persistence are small for daughters. However, the rank-rank slopes in health continue to be larger for whites than blacks irrespective of gender. In the case of expected ranks at the 25^th percentile, the racial gaps remain fairly pronounced irrespective of gender or whether the outcome is health or household income. This suggests that blacks are in all cases disadvantaged when it comes to upward mobility from the bottom. In panels E and F we repeat the analysis but use individual income rather than household income. This exercise highlights the notable finding previously shown in Chetty et al. (2018) that the disadvantage in upward mobility in expected rank at the 25^th percentile in individual income is now virtually eliminated for black women. Notably, there is still a large upward mobility disadvantage for black men.

We next turn to differences in health mobility across the regions of the United States in which the child grew up. The most striking finding is that growing up in the South is associated with a lower expected rank throughout the parent distribution. This echoes the findings from Chetty et al. (2014) and Davis and Mazumder (2018) who find similar patterns with income mobility and discuss many of the factors that could explain these regional gaps. Table 8 shows that the expected health rank for a child who grew up in the South with parents at the 25^th percentile, is the 42^nd percentile, the lowest of the four regions (Table 6, column 2). In comparison, children that grew up in the Northeast and North Central are expected to be at the 46^th percentile. Downward mobility is also highest among children growing up in the South. A child from the South with a parent at the 75^th percentile in the health distribution has an expected rank of the 55^th percentile, which is lower than in all other regions.³⁶ The disparity between regions in health mobility, however, is not as great as it is for income mobility (Table 8 columns 5 and 6), once again highlighting distinctions between subgroup patterns in health and income mobility.

Lastly, we explore examine differences by parent education level where we again find significant differences. Children whose parents are at the 25^th percentile but have a college degree are expected to be at the 52^nd percentile, but those with parents without a high school degree are only expected to attain the 37^th percentile. This striking disparity is evident throughout the parent health distribution (Panel C of Figure 5). This highlights that the well-known disparity in health by education level also persists to the next generation when looking at offspring health. One explanation is that more educated parents have access to resources that can improve their children’s health regardless of their own health status (Case et al. 2002). Overall, our analysis suggests that the same groups that face disadvantages with respect to income mobility are also experiencing lower mobility with respect to health.³⁷

5.6. Trends in Health Mobility

We next examine trends in health mobility for three groups of cohorts born between: 1950–1959; 1960–1969; and 1970–1979. We start with trends in the IHA which are displayed in Appendix Figure A.18. For this analysis, we use only health observations from age 30 to 40 for children and from age 40 to 70 for parents.³⁸ Appendix Figure A.18 shows an increase in the intergenerational health association from 0.18 to 0.26 between the birth cohorts born in the 1950s and the 1970s. This increase appears for both the son and daughter subsamples. However, the magnitude of the increase and its statistical significance are somewhat sensitive to the choice of ages used to measure parent health. In Appendix Table A.10, we find that the across cohort change is smaller and not statistically significant when we restrict the samples to measure parent health between the ages of 50 and 70.

We also investigate how rank mobility differs by birth cohort. In Appendix Figure A.19 we plot the rank-rank slopes (Panel A) and expected health ranks at the 25^th and 75^th percentile (Panel B) for the three cohorts.³⁹ Unlike the IHA, we find more limited evidence of an increase in rank-rank slopes. While the point estimate increased from 0.23 to 0.27 for the full sample, this change is not statistically significant. When we examine the expected ranks at the 25^th and 75^th percentile, we do find suggestive evidence that upward mobility from the bottom declined and that downward mobility from the top has increased for more recent cohorts. In Appendix Table A.11, we examine the changes in rank mobility across the different parent-child types. We find evidence of significant changes in rank persistence and upward mobility from the bottom between fathers and sons.

Overall, we believe this constitutes modestly suggestive evidence of a decline in intergenerational health mobility for more recent cohorts. This finding is potentially consistent with growing evidence of a decline in intergenerational income mobility (e.g. Aaronson and Mazumder, 2008; Davis and Mazumder, 2017) and a decline in intergenerational educational mobility reported by Hilger (2017). Nevertheless, since the most recent cohorts (born since 1970) are still relatively young, future work may be able to better substantiate whether a change in health mobility has taken place.

5.7. The Role of Childhood Circumstances

Finally, we consider how childhood circumstances affect health mobility by using a rich set of covariates on childhood circumstances available in the PSID’s Childhood Retrospective Circumstance Study (CRCS). We begin with estimates of the IHA from a pooled sample of sons and daughters in which we combine both parents’ health. The results are depicted graphically in Figure 7 using a dot for the point estimate and horizontal lines for the 95 percent confidence interval. The baseline IHA estimate for this sample is 0.241. In Panel A, we then control for different sets or “categories” of control variables. When we include a set of measures of socioeconomic status (e.g. parent years of education, family income, child race and various indices of SES), the IHA falls to 0.169. This finding that SES can account for a significant share (29%) of the intergenerational association in health is consistent with previous studies including Currie and Moretti (2007). If instead of SES controls, we include a set of childhood health measures the IHA estimate falls to just 0.221. Controlling only for measures of childhood stability, school experience, or childhood relationships appears to have little effect on the IHA. Using all the variables together lowers the IHA to 0.154. This accounts for 36% of the unconditional IHA. Finally, panel B, depicts the associated estimates when controlling for one variable at a time, rather than using whole categories.

Figure 7: Effect of childhood factors on intergenerational health associations.

Figure 7 shows how the baseline intergenerational health association is attributable to various childhood factors for the sample of individuals in the child generation who were also part of the 2014 Childhood Retrospective Circumstance Study (CRCS). Panel A plots the intergenerational health associations as groups of childhood factors are added to the baseline regression of child’s health measure on parent’s health measure. Family SES Background includes mother’s years of education, father’s years of education, family income, SES Index Age 0–5, SES Index Age 6–12, SES Age 13–16, Neighborhood Quality Index, and controls for race of child (white, black or other). Childhood Health includes Child Health Index, Underweight at 13, Overweight at 13, and Obese at 13. Childhood Stability includes number of times moved in childhood, number of schools attended before 17, if parents were satisfied with their relationship, and if parents ever divorced. School Experience includes number of times repeat school grade, School Experience Index Age 6–12, School Experience Age 13–16. Childhood Relationship includes Friendship Quality Index Age 6–12, Friendship Quality Index Age 13–16, Relationship with Mother Quality Index, Relationship with Father Quality Index, and having a mentor at age 17–30. Panel B plots the intergenerational health associations as individual childhood factors are added to the baseline regression. The dependent variable for all specifications is the child’s time-averaged continuous health measure. The parent health measure is the average of the mother’s and father’s health if available. Otherwise, only one parent’s health measure is used. All specifications include as controls the quadratic age terms of the mother, father and child, and missing indicators for mother and father. Age for both generations is defined as the time-averaged age of the individual at the time of health observations. The red dashed lines denote the baseline intergenerational health association. Additional details on the CRCS variables can be found in Appendix A. All regressions are weighted using individual CRCS sampling weights of the child. 95% confidence intervals are shown calculated using standard errors that are robust to heteroskedasticity and within-family correlation.

In Figure 8, we do an analogous breakdown of the rank-rank slope and find very similar patterns. Our baseline estimate of 0.292 falls to 0.232 if we control for family SES background variables and 0.223 if we include all of our controls. Thus, we can account for 24 percent of the rank-rank persistence.

Figure 8: Effect of childhood factors on rank-rank slopes.

Figure 8 shows how the baseline rank-rank slope is attributable to various childhood factors for the sample of individuals in the child generation who were also part of the 2014 Childhood Retrospective Circumstance Study (CRCS). Panel A plots the rank-rank slopes as groups of childhood factors are added to the baseline bivariate regression of child’s health rank on parent’s health rank. Family SES Background includes mother’s years of education, father’s years of education, family income, SES Index Age 0–5, SES Index Age 6–12, SES Age 13–16, Neighborhood Quality Index, and controls for race of child (white, black or other). Childhood Health includes Child Health Index, Underweight at 13, Overweight at 13, and Obese at 13. Childhood Stability includes number of times moved in childhood, number of schools attended before 17, if parents were satisfied with their relationship, and if parents ever divorced. School Experience includes number of times repeat school grade, School Experience Index Age 6–12, School Experience Age 13–16. Childhood Relationship includes Friendship Quality Index Age 6–12, Friendship Quality Index Age 13–16, Relationship with Mother Quality Index, Relationship with Father Quality Index, and having a mentor at age 17–30. Panel B plots the rank-rank slopes as individual childhood factors are added to the baseline regression. Health percentile ranks are constructed from the age-adjusted health measure and are ranked separately by gender within each generation. The red dashed lines denote the baseline rank-rank slope. Additional details on the CRCS variables can be found in Appendix A. All regressions are weighted using individual CRCS sampling weights of the child. 95% confidence intervals are shown calculated using standard errors that are robust to heteroskedasticity and within-family correlation.

We caution that these decompositions are difficult to interpret in the absence of a structural model as some of these factors might be considered pre-determined and exogenous, but others are clearly endogenous choices that are affected by parent and child health. We simply take these results as useful for a first-pass descriptive analysis. Future researchers may consider structural models to better understand the mechanisms or utilize credible research designs that are able to identify causal channels.

In a separate exercise we also attempt to examine the role of access to health insurance. Access to insurance could be a factor in reducing health persistence and increasing mobility. Consistent with this, Appendix Table A.12 shows that accounting for access to health insurance does appear to lower health persistence and increases the expected ranks of children, but of course insurance choices are endogenous. Therefore, we leave a formal exploration of this mechanism for future research that may be able to exploit more credible research designs with larger samples.

6. Conclusion

Given the rise of inequality and associated concerns about unequal opportunity, studies of intergenerational mobility have received growing attention. Most studies have focused primarily on income, education, or occupation. However, until recently few studies have considered broad-based measures of health despite its central importance to welfare. To fill this void, we provide estimates of intergenerational mobility with respect to a broad-based measure of lifetime health in the US by using repeated measures of quality adjusted life years (QALY) based on self-reported health status. We find that the rank-rank slope in health is about 0.26 suggesting a low degree of intergenerational persistence in health and therefore a relatively high degree of intergenerational health mobility.

We also consider how the joint distribution of income and health evolves over a generation by including both income and health in the parent generation in our rank-based models. Including both dimensions increases the explanatory power of children’s outcomes. Hence, health appears to capture a somewhat distinct dimension of socioeconomic status than income.

We further attempt to push the literature forward by integrating income and health into an overall measure of welfare by converting health to a monetary metric. We find the rank-rank slope in this overall measure of welfare to be 0.43 in our sample which is higher than when we consider either income or health separately. We suspect that in samples better designed to capture lifetime income, we would obtain even higher estimates, but we leave that for future research. Future studies may wish to consider how other elements of socioeconomic status might also be integrated into broader measures of welfare.

Finally, we also document important differences in intergenerational health mobility by race, race by gender, region, and parent education levels. For example, we find that blacks experience significantly less upward mobility and significantly higher downward mobility than whites. However, the racial gap in health mobility is smaller than the analogous racial mobility gaps in income. We also find significant difference by parent education level suggesting that the well-known socioeconomic gradient in health persists intergenerationally. Overall, it is clear that the same groups that experience lower intergenerational income mobility also experience lower intergenerational health mobility.

Highlights:

• The rank-rank slope in self-reported health status is 0.26

• Including both parent health and income in models of intergenerational mobility increases the explanatory power of child outcomes.

• The rank-rank slope in a welfare measure that combines income and health is about 0.4

• There are striking health mobility gaps by race, region and parent education.