Nội dung

National Poverty Center Gerald R. Ford School of Public Policy, University of Michigan www.npc.umich.edu Education and Health: Evaluating Theories and Evidence David Cutler, Harvard University Adriana Lleras‐Muney, Princeton University This paper was delivered at a National Poverty Center conference. Any opinions, findings, conclusions, or recommendations expressed in this material are those of the author(s) and do not necessarily reflect the view of the National Poverty Center or any sponsoring agency. Education and Health: Evaluating Theories and Evidence ∗ David M. Cutler and Adriana Lleras-Muney I. Introduction There is a well known large and persistent association between education and health. This relationship has been observed in many countries and time periods, and for a wide variety of health measures. 1 The differences between the more and the less educated are significant: in 1999, the ageadjusted mortality rate of high school dropouts ages 25 to 64 was more than twice as large as the mortality rate of those with some college (table 26, National Vital Statistics Reports, 2001). Substantial attention has been paid to these “health inequalities.” Gradients in health by education are now being systematically monitored in many countries (the United States includes them as part of its Healthy People 2010 goals), and countries such as the United Kingdom have target goals of reducing health disparities –specifically by education or factors correlated with education. 2 In this paper, we review what is known and not known about the relationship between education and health, in particular about the possible causal relationships between education and health and the mechanisms behind them. We then assess the extent to which education policies can or should be thought of as health policies. We note at the outset that this is a controversial topic. A number of authors have written about education-related health inequalities, and the conclusions frequently differ. To some extent, this is a result of data limitations. Many of the data sets that we and others employ use health measures that are self-reported. In addition to true differences in health, there will thus be some differences related to knowledge of existing conditions, which may itself be related to education. Also very important, however, is that work on the mechanisms underlying the link between health and education has not been conclusive. Not all relevant theories have been tested, and when they have, studies will often conflict with each other. We highlight the discrepancies as best we can. We do not resolve the differences here – that is an enormous task, and is not doable with current information. Noting the points of disagreement is 1 important in its own right, however. Along the way, we indicate where more research would be particularly valuable. II. The Relationship Between Health and Education To document the basic correlations between education and health, we estimate the following regression: H i = c + β Ei + X i δ + ε i where Hi is a measure of individual i’s health or health behavior, Ei stands for individual i’s years of completed education, Xi is a vector of individual characteristics that includes race, gender and single year of age dummies, c is a constant term and ε is the error term. The coefficient on education β (also referred to as the education gradient) is the object of interest, and it measures the effect of one more year of education on the particular measure of health. We focus on individuals ages 25 and above since they have most likely already completed their education. Education is included either in years (as in the labor literature), or using dummies for each year of education, to be as flexible as possible. We first report results for the entire sample, and then for different demographic groups. We estimate linear models for continuous variables. For dichotomous variables we estimate logit probability models and report the marginal effects. The data we employ are from various years of the National Health Interview Survey (NHIS) in the United States. 3 We use the NHIS because it has a large number of health outcomes and behaviors. Generally, results from the NHIS match other surveys with self-reports (Cutler and Glaeser 2005) and even physical assessments, though clearly there are exceptions, such as weight and height. We note possible reporting issues as we present the results. Table 1 reports the coefficient on years of schooling in explaining various measures of health. The first outcome we look at is whether an individual died within 5 years of the interview. In the NHIS this is determined by matching individual information to death certificates through the National Death Index 2 (see Appendix for more details). Then we look at gradients in the self-report of a past acute or chronic disease diagnosis. Most of these diseases are very serious (cancer or heart disease, for example), and people would certainly know if they have had been diagnosed with them (although it is possible that conditional on having the disease, the more educated are more likely to know about it. If that is the case then the gradients we report for these diseases could partially reflect differential diagnosis and knowledge—this is not the case for mortality however). Of course, since the sample is of people who are alive, differential mortality between better educated and less educated is an issue. But this would tend to reduce reported gradients, if less educated people die more when they have any disease, and thus are not alive to report the disease. The first column includes a very basic set of controls: a full set of age dummies, race, and gender. The results (column 1) show that individuals with higher levels of education are less likely to die within 5 years. The second block of the table shows the more educated also report having lower morbidity from the most common acute and chronic diseases (heart condition, stroke hypertension, cholesterol, emphysema, diabetes, asthma attacks, ulcer). The only exceptions are cancer, chicken pox and hay fever. Differential reporting of hay fever could possibly be related to differential knowledge of disease (better educated people will be more likely to go to specialists for testing). This might be the explanation for cancer as well; skin cancer is the most common cancer, and could be subject to reporting bias. But that might not be the whole explanation. Some evidence suggests that some cancer risk factors are adverse for the better educated (as with late childbearing age and breast cancer). It may also be that better educated people are more likely to survive with cancer, or that better care for competing risks keeps the better educated alive long enough to die of cancer. Differences in chronic disease prevalence are similar. Better educated people are less likely to be hypertensive, or suffer from emphysema or diabetes. The third set of rows shows that physical and mental functioning is better for the better educated. The better educated are substantially less likely to report themselves in poor health, and less likely to report anxiety or depression. Finally, the last block 3 shows that better educated people report spending fewer days in bed or not at work due to disease, and have fewer functional limitations. The magnitude of the relationship between education and health varies across conditions, but they are generally large. An additional four years of education lowers five year mortality by 1.8 percentage points (relative to a base of 11 percent); it also reduces the risk of heart disease by 2.16 percentage points (relative to a base of 31 percent), and the risk of diabetes by 1.3 percentage points (relative to a base of 7 percent). Four more years of schooling lowers the probability of reporting in fair or poor health by 6 percentage points (the mean is 12 percent), and reduce lost days of work to sickness by 2.3 each year (relative to 5.15 on average). Although the effects of gender and race are not shown, the magnitude of 4 years of schooling is roughly comparable in size to being female or being African American. These are not trivial effects. The reasons for these associations are multi-factorial, although it is likely that these health differences are in part the result of differences in behavior across education groups. Table 2 shows the relation between education and various health risk factors: smoking, drinking, diet/exercise, use of illegal drugs, household safety, use of preventive medical care, and care for hypertension and diabetes. Overall, the results suggest very strong gradients where the better educated have healthier behaviors along virtually every margin (although some of these behaviors may also reflect differential access to care). Those with more years of schooling (we report the effects of 4 more years) are less likely to smoke (11 percentage points relative to a mean of 23 percent), to drink a lot (7 fewer days of 5 or more drinks in a year, among those who drink, of a base of 11), to be overweight or obese (5 percentage points lower obesity, compared to an average of 23 percent), or to use illegal drugs (0.6 percentage points less likely to use other illegal drugs, relative to an average of 5 percent). Interestingly, the better educated report having tried illegal drugs more frequently, but they gave them up more readily. Similarly, the better educated are more likely to exercise and to obtain preventive care such as flu shots (7 percentage points relative to an average of 31 percent), vaccines, mammograms (10 percentage points relative to an average of 54 percent), pap smears (10 percentage points relative to an average of 60 4 percent) and colonoscopies (2.4 percentage points relative to an average of 9 percent). Among those with chronic conditions such as diabetes and hypertension, the more educated are more likely to have their condition under control. Furthermore, they are more likely to use seat belts (12 percentage points more likely to always use a seat belt, compared to the average of 68 percent) and to have a house with a smoke detector (10.8 percentage points relative to an average of 79 percent) and that has been tested for radon (2.6 percentage points relative to a base of 4 percent). All of these behavioral effects are very large. It is worth noting that these health behaviors explain some, but not all of the differences in health. For example, in the famous Whitehall study of British civil servants (Marmot 1994), smoking, drinking, and other health behaviors explain only one-third of the difference in mortality between those of higher rank and those of lower rank. Although that study did not focus on educational differences, we find similar results. In the NHIS, the effect of education on mortality is reduced by 30% when controlling for exercise, smoking, drinking, seat belt use, and use of preventive care (results available upon request). This is perhaps an underestimate – one cares about the length of time smoked, the specific cigarettes smoked, the number of puffs taken, and the like. But absent measurement error in behaviors, the result implies that there must be unobserved health behaviors that also contribute to health differences, or alternatively, that the more educated might be healthier due to reasons/behaviors that are not known to be health improving. Equally important, we do not understand why the more educated make larger investments in their health; we return to this in the next sections. The relationship between education and health shows up across countries as well. Figure 1 shows the simple correlation between average education (using the well-known Barro-Lee international data) and life expectancy (without any additional controls). As average education increases, life expectancy improves, although the returns appear to be larger for poorer countries. The same is true within countries as well. The more educated are more likely to live longer not just in the US, but also in Canada (Mustard, et al. 1997), Israel (Manor, et al. 1999) and both Western and Eastern Europe, 4 including Russia (Shkolnikov, et al. 1998). This relationship has also been documented 5 in developing countries, such as Bangladesh (Hurt, et al. 2004), Korea (Khang et al. 2004), and China (Liang, et al. 2000). In most cases, however, education is not associated with lower cancer mortality. Heterogeneous effects. The basic correlations we just described do not fully describe important aspects of the relationship between education and health. For example, it is important to know whether the returns to schooling are constant for every additional year of school, regardless of the initial level of schooling, or whether the benefits from say primary schooling exceed those from higher education. To better understand the shape of the relationship between education and health, we estimate non-parametric models that include a dummy variable for each year of schooling as explanatory variables (rather than years of education as a continuous variable as in Tables 1 and 2), and include the same basic demographic controls we included previously. Figure 2 plots the estimated effects for a number of health and health behaviors. We chose four representative health measures (mortality, SRHS, depression and functional limitations) and four measures of behaviors that cover a range of different areas: smoking is an addictive behavior that is known to adversely affect health and has potentially an important social component; colorectal screening is preventive but may be related to access to health care; wearing a seat belt is also preventive but not monetarily costly; and lastly smoke detectors at home, which picks up general safety. Although the estimates are noisy (some education categories have very few observations), they show that for many outcomes, there are returns beyond high school completion (12 years of schooling). Education matters for health not just because of basic reading and writing skills. For some outcomes, the relationship between years of schooling and health appears to be linear (see mortality, colorectal screenings and smoke detectors). For other outcomes, such as functional limitations, smoking and obesity, the relationship is non-linear, with an increased effect of an additional year of school only for people who are better educated. In all cases, however, the relationship between education and health is roughly linear after 10 years of school; we do not see large evidence of sheepskin effects in health – that is, there does not appear to be an additional health benefit associated with the 6 completion of a degree, beyond what would be expected given the number of years of schooling (although for some outcomes such as SHRS and functional limitations there may be a small effect of high school graduation). In contrast, there are clear sheepskin effects on wages, for example see Tyler, Murnane and Willet (2000). Subject to the possibility of small effects that we cannot measure accurately, (e.g., the product of the sheepskin effect in wages and the impact of income on health may be small), this allows us to reject the idea that the health returns to education (the health benefit associated with one more year of schooling) are driven by the labor market returns to education This also implies that there may be substantial health returns to education policies that promote college attendance. The effects of education on health and health behaviors also differ along other dimensions. These effects vary significantly for individuals of different ages. Figure 3 shows the coefficients of education estimated by single year of age. Some of these education gradients (mostly those related to behaviors) fall continuously with age (smoking, seat belt use, smoke detector); whereas others increase with age until middle ages, and then start to fall (functional limitations, depression and colorectal screening). In all cases, however, we find that the effect of education starts to fall sometime between ages 50 and 60. Other studies have also documented smaller effects of education for older ages on mortality (Elo and Preston 1996). Interestingly, some studies also find that the health differences associated with income also diminish after middle age (Smith 2005), though this is not true in all studies (Wolfson 1993). Some of the decline in the education gradient after age 50 must certainly be due to the selective survival of the more educated (Lynch 2003). There may also be additional cohort effects—education may have become more important for younger cohorts. Or education may simply matter less after retirement, with stable incomes and universal insurance coverage. It is difficult to separate these effects. There are important differences by gender as well. Table 3 shows the impact of education for men and women (the second and third columns), blacks and whites (the fourth and fifth columns), and rich and poor (the sixth and seventh columns). The table reports whether the marginal effect of education is significantly different for the two groups, as well as the effect of one more year of education as a percentage of the mean level for the group (to account for the fact that different groups may have different 7 baselines). In more than half the cases, education has a statistically indistinguishable effect for men and women. In some cases, education has a greater impact for women (depression and obesity, for example). In other cases, the effect is bigger for men (mortality and heavy drinking). Whether these differences result from biology or behavior is not known. In the next two columns we compare gradients for whites and blacks. Again, the coefficients are similar most of the time. Where they differ, education gradients are larger for whites than for blacks (with the exception of smoke detectors), although the effects are closer when the effects are rescaled as a percentage of the mean. One possible explanation is that the quality of education is lower for blacks than for whites, though we have no direct evidence on this. These findings are also consistent with lower returns to education on wages among blacks. Lastly, we examine whether education matters more for those with low family incomes (incomes below $20,000)—although we note here that because education affects income, and health may determine income, it is more difficult to interpret these results. In most cases we examine, education matters more among the non-poor than among the poor. This suggests that income and education are complementary in the production of health. This would be the case if, for example, education allows people to know about particular new treatments and income allows them to purchase the treatment. The results by race and income together suggest that socio-economic advantages are complementary (or cumulative). They also suggest that interactions between education and other variables may be important. The education gradient over time. Education gradients in mortality appear to be increasing in both the United States (Pappas, et al. 1993) 5 and Europe (Mackenbach, et al. 2003), (Kunst, et al. 2002). As a result, even though life expectancy is improving for all, the differences in life expectancy between college educated and others have become larger. Other measures of health confirm these findings. For example, Goesling (2005) finds that there has been an increase in the effect of education on self-reported health since 1982. Looking at the same period, Schoeni et al (2005) find that although disability rates in the US have fallen, they have fallen more among the educated. The gradient in some health behaviors is also 8 increasing: there were very small differences in smoking rates between education groups prior to the Surgeon General Report in 1964, but these differences are substantial today (Pamuk, et al. 1998; figure 35). Although compositional changes could be driving the observed differences – educational attainment has increased enormously over time – the results suggest that health inequalities could continue rising. Spillovers across people. It is well known that maternal education is strongly associated with infant and child health, both in the US and in developing countries (for developing countries see Strauss and Thomas 1995, for the US see Meara 2001, or Currie and Moretti, 2003). More educated mothers are less likely to have low or very low birth weight babies, and their babies are less likely to die within their first year of life. These effects persist well into adulthood: Case, Fertig and Paxson (2005) find that mother’s education predicts self reported health at age 42. Recent research further suggests that more educated children have an effect on the health of their parents: Field (2005) finds that parents of individuals who obtained more schooling were subsequently more likely to stop smoking. It is also possible that having an educated spouse positively affects health. For example, Egeland, (2002) and Bosma et al (1995) find that even controlling for own education, those who are married to more educated spouses have lower mortality rates (although this finding is not universal, for example see Suarez and Barrett-Connor 1984). Having a more educated spouse is also associated with better health and health behaviors such as smoking and excessive drinking (Monden, 2003). Of course it is difficult to know whether this relationship is driven by assortative mating or whether it reflects a causal effect. III. Is the effect of education on health causal? In a very broad sense, there are three possible reasons for the link between health and education. One possibility is that poor health leads to low levels of schooling. Another possibility is that increasing education improves health. And lastly there may be third factors that increase both schooling and health. It is important for policy to understand how much of the observed correlation between education and 9