|Home | About | Journals | Submit | Contact Us | Français|
Using a variety of data sets from two countries, we examine possible explanations for the relationship between education and health behaviors, known as the education gradient. We show that income, health insurance, and family background can account for about 30 percent of the gradient. Knowledge and measures of cognitive ability explain an additional 30 percent. Social networks account for another 10 percent. Our proxies for discounting, risk aversion, or the value of future do not account for any of the education gradient, and neither do personality factors such as a sense of control of oneself or over one’s life.
In 1990, a 25 year-old male college graduate could expect to live another 54 years. A high school dropout of the same age could expect to live 8 years fewer (Richards and Barry, 1998). This enormous difference in life expectancy by education is true for every demographic group, is persistent – if not increasing – over time (Kitagawa and Hauser, 1973; Elo and Preston, 1996; Meara, Richards, and Cutler, 2008), and is present in other countries (Marmot, Shipley, and Rose, 1984 (the U.K.); Mustard, et al. 1997 (Canada); Kunst and Mackenbach, 1994 (northern European countries)).1
A major reason for these differences in health outcomes is differences in health behaviors. 2 In the United States, smoking rates for the better educated are one-third the rate for the less educated. Obesity rates are half as high among the better educated (with a particularly pronounced gradient among women), as is heavy drinking. Mokdad et al. (2004) estimate that nearly half of all deaths in the United States are attributable to behavioral factors, most importantly smoking, excessive weight, and heavy alcohol intake. Any theory of health differences by education thus needs to explain differences in health behaviors by education. We search for explanations in this paper.3
In standard economic models, people choose different consumption bundles because they face different constraints (for example, income or prices differ), because they have different beliefs about the impact of their actions, or because they have different tastes. We start by showing, as others have as well, that income and price differences do not account for all of these behavioral differences. We estimate that access to material resources, such as gyms and smoking cessation methods, can account for at most 30 percent of the education gradient in health behaviors. Price differences work the other way. Many unhealthy behaviors are costly (smoking, drinking, and overeating), and evidence suggests that the less educated are more responsive to price than the better educated. As a result, we consider primarily differences in information and in tastes.
Some of the differences by education are indeed due to differences in specific factual knowledge — we estimate that knowledge of the harms of smoking and drinking accounts for about 10 percent of the education gradient in those behaviors. However, more important than specific knowledge is how one thinks. Our most striking finding, shown using US and UK data, is that a good deal of the education effect – about 20 percent – is associated with general cognitive ability. Furthermore this seems to be driven by the fact that education raises cognition which in turn improves behavior.
A lengthy literature suggests that education affects health because both are determined by individual taste differences, specifically in discounting, risk aversion, and the value of the future—which also affect health behaviors and thus health. Victor Fuchs (1982) was the first to test the theory empirically, finding limited support for it. We know that taste differences in childhood cannot explain all of the effect of schooling, since a number of studies show that exogenous variation in education influences health. For example, Lleras-Muney (2005) shows that adults affected by compulsory schooling laws when they were children are healthier than adults who left school earlier. Currie and Moretti (2003) show that women living in counties where college is more readily available have healthier babies than women living in other counties. However, education can increase the value of the future simply by raising earnings and can also change tastes.
Nevertheless, using a number of different measures of taste and health behaviors, we are unable to find a large impact of differences in discounting, value of the future, or risk aversion on the education gradient in health behaviors. Nor do we find much role for theories that stress the difficulty of translating intentions into actions, for example, that depression or lack of self control inhibits appropriate action (Salovey, Rothman, and Rodin, 1998). Such theories are uniformly unsupported in our data, with one exception: about 10 percent of the education gradient in health behaviors is a result of greater social and emotional support.
All told, we account for about two-thirds of the education gradient with information on material resources, cognition, and social interactions. However, it is worth noting that our results have several limitations. First, we lack the ability to make causal claims, especially because it is difficult to estimate models where multiple mechanisms are at play. Second, we recognize that in many cases the mechanisms we are testing require the use of proxies which can be very noisy, causing us to dismiss potentially important theories. Nevertheless we view this paper as an important systematic exploration of possible mechanisms, and as suggesting directions for future research.
The paper is structured as follows. We first discuss the data and empirical methods. The next section presents basic facts on the relation between education and health. The next two sections discuss the role of income and prices in mediating the education-behavior link. The fourth section considers other theories about why education and health might be related: the cognition theory; the future orientation theory; and the personality theory. These theories are then tested in the next three sections. We then turn to data from the U.K. The final section concludes.
In the course of our research, we use a number of different data sets. These include the National Health Interview Survey (NHIS), the National Longitudinal Survey of Youth (NLSY), the National Survey of Midlife Development in the United States (MIDUS), the Health and Retirement Study (HRS), the Survey on Smoking (SOS), and the National Childhood Development Study (NCDS) in the U.K. We use many data sets because no single source of data has information allowing us to test all the relevant theories. For the US we have restricted our attention to the whites only because our earlier work showed larger education gradients among them (Cutler and Lleras-Muney 2008b) but the results presented here are not particularly sensitive to that choice. A lengthy data appendix discusses the surveys in more detail.
In all data sets we restrict the samples to individuals ages 25 and above (so education has been mostly completed)—but place no upper limit on age. The health behaviors we look at are self-reported. This is a limitation of our study, but we were unable to find data containing measured (rather than self reported) behaviors to test our theories.4 To the extent that biases in self reporting vary across behaviors, our use of multiple health behaviors mitigates this bias. Nevertheless it is worth noting that not much is known about whether biases in reporting vary systematically by education.
To document the effect of education on health behaviors, we estimate the following regression:
Where Hi is a health behavior of individual i, Education is measured as years of schooling in the US, and as a dummy for whether the individual passed any A level examinations in the UK.5 The basic regression controls for basic demographic characteristics (gender, age dummies and ethnicity) and all available parental background measures (which vary depending on the data we use). Ideally in this basic specification we would like to control for parent characteristics and all other variables that determine education but cannot be affected by it, such as genetic and health endowments at birth— we control for the variables that best seem to fit this criterion in each data set.6 The education gradient is given by β1, the coefficient on education, and measures the effect of schooling on behavior, which could be thought of as causal if our baseline controls were exhaustive. We discuss below whether the best specification of education is linear or non-linear.
In testing a particular theory we then re-estimate equation (1) adding a set of explanatory variables Z:
We then report, for each health measure, the percent decline in the coefficient of education from adding each set of variables, 1 - α1/β1.
Many of our health measures are binary. To allow for comparability across outcomes, we estimate all models using linear probability, but our results are not very different if we instead use a non-linear model. Thus, the coefficients are the percentage point change in the relevant outcome. Since we have many outcomes, it is helpful to summarize them in a single number. We use three methods to form a summary. First we compute the average reduction of the gradient across outcomes for those outcomes with a statistically significant gradient in the baseline specification. Of course, not all behaviors contribute equally to health outcomes. Our second summary measure weights the different behaviors by their impact on mortality. The regression model, using the 1971–75 National Health and Nutrition Examination Survey Epidemiological Follow-up Study, is described in the Appendix. For comparability reasons, the behaviors are restricted to smoking, drinking, and obesity. The summary measure is the predicted change in 10 year mortality associated with each additional year of education.7 Finally, we report the average effect of education across outcomes using the methodology described in Kling, Liebman, and Katz (2007), which weights outcomes equally after standardizing them.8
We start by presenting some basic facts relating education and health behaviors, before discussing theories linking the two. Health behaviors are asked about in a number of surveys. Probably the most complete is the National Health Interview Survey (NHIS). In order to examine as many behaviors as possible, we use data from a number of NHIS years, 1990, 1991, 1994 and 2000.9 We group health behaviors into eight groups: smoking, diet/exercise, alcohol use, illegal drugs, automobile safety, household safety, preventive care, and care for people with chronic diseases (diabetes or hypertension). Within each group, there are multiple measures of health behaviors. Because the NHIS surveys are large, our sample sizes are up to approximately 23,000.
Table 1 shows the health behaviors we analyze and the mean rates in the adult population. We do not remark upon each variable, but rather discuss a few in some depth. Current cigarette smoking is a central measure of poor health. Mokdad et al. (2004) estimate that cigarette smoking is the leading cause of preventable deaths in the country (accounting for 18 percent of all deaths). The first row shows that twenty-three percent of white adults in 2000 smoked cigarettes. The next columns relate cigarette smoking to years of education, entered linearly. We control for single year of age dummies, a dummy for females, and a dummy for Hispanic.
Each year of education is associated with a 3.0 percentage point lower probability of smoking. Put another way, a college grad is 12 percentage points less likely to smoke than a high school grad. Given that smoking is associated with 6 years shorter life expectancy (Cutler et al., 2000), this difference is immense.
Entering education linearly may not be right. One might imagine that some base level of education is important, and that additional education beyond that level would not reduce smoking. That is not correct, however. The first part of Figure 1 shows the relationship between exact years of education and smoking: the figure reports the marginal effect of an additional year of education for each level of education, estimated using a logit model. If anything, the story is the opposite of the ‘base education’ hypothesis; the impact of education is greater at higher levels of education, rather than lower levels of education (although there are few observations at the lower end of the education distribution and thus these estimates are imprecise). Overall the relationship appears to be linear above 10 years of schooling for all of the outcomes in Figure 1.
Next to smoking, obesity is the leading behavioral cause of death. While all measures of excess weight are correlated, we focus particularly on obesity (defined as a Body Mass Index or BMI equal to or greater than 30). Twenty-two percent of the population in 2000 self-reported themselves to be obese.10 This too is negatively related to education; each year of additional schooling reduces the probability of being obese by 1.4 percent (Table 1). The shape by exact year of education is similar to that for smoking (Figure 1). Obesity declines particularly rapidly for people with more than 12 years of education.
Heavy drinking is similarly harmful to health. We focus on the probability that the person is a heavy drinker – defined as having an average of 5 or more drinks when a person drinks. Eight percent of people are heavy drinkers. Each additional year of education lowers this by 1.8 percent. Interestingly the better educated are more likely to drink but less likely to drink heavily.
Self-reported use of illegal drugs is relatively low; only 2 to 8 percent of people report using such drugs in the past year. Recent use of illegal drugs is generally unrelated to education (at least for marijuana and cocaine). But better educated people report they are more likely to have ever tried these drugs. Better educated people seem better at quitting bad habits, or at controlling their consumption. This shows up in cigarette smoking as well, where the gradient in current smoking is somewhat greater than the gradient in ever smoking.
Automobile safety is positively related to education; better educated people wear seat belts much more regularly than less educated people. The mean rate of always wearing a seat belt is 69 percent; each year of education adds 3.3 percent to the rate. The analysis of seat belt use is particularly interesting. Putting on a seat belt is as close to costless as a health behavior comes. Further, knowledge of the harms of non-seat belt use is also very high. But the gradient in health behaviors is still extremely large.
Household safety is similarly related to education. Better educated people keep dangerous objects such as handguns safe and know what to do when something does happen (for example, they know the poison control phone number).
Better educated people engage in more preventive and risk control behavior. Better educated women get mammograms and pap smears more regularly, better educated men and women get colorectal screening and other tests, and better educated people are more likely to get flu shots. Among those with hypertension, the better educated are more likely to have their blood pressure under control. Services involving medical care are the least clear of our education gradients to examine, since access to health care matters for receipt of these services. We thus focus more on the other behaviors. But, these data are worth remarking on because it does not appear that access to medical care is the big driver. Controlling for receipt of health insurance does not diminish these gradients to any large extent (the education coefficient on receipt of a mammogram is reduced by only 18 percent, for example, if we control for insurance in addition to age and ethnicity alone). This is consistent with the Rand Health Insurance Experiment (Newhouse et al., 1993); making medical care free increases use, but even when care is free, there is still significant underuse. Seeing a doctor may be like wearing a seat belt; it is something that better educated people do more regularly.
Table 1 makes clear that education is associated with an enormous range of positive health behaviors, the majority of health behaviors that we explore. The average predicted 10 year mortality rate is 11 percent, shown in the last row of the Table. Relative to this average, our results suggest that every year of education lowers the mortality risk by 0.3 percentage points, or 24 percent, through reduction in risky behaviors (drinking, smoking, and weight).
We have examined the education gradient in health behaviors using other data sets as well. Some of these results are presented later in the paper. In each case, there are large education differences across a variety of health behaviors and for somewhat different samples. Education differences in health behaviors are not specific to the United States. They are apparent in the U.K. as well. As documented later in the paper (Appendix Table 3), we analyze a sample of British men and women at ages 41–42. People who passed the A levels are 15 percent less likely to smoke than those who did not pass. Additionally those that passed A levels are 6 percent less likely to be obese, and are 3 percent less likely to be heavy drinkers.
An obvious difference between better educated and less educated people is resources. Better educated people earn more than less educated people, and these differences in earnings could affect health. There are two channels for this. First, higher income allows people to purchase goods that improve health, for example health insurance. In addition, higher income increases steady-state consumption, and thus raises the utility of living to an older age. We focus here on the impact of current income as a whole, and consider specifically the value of the future in a later section.
A number of studies suggest that both education and income are each associated with better health. Thus, it is clear that income does not account for all of the education relationship. But for our purposes, the magnitude of the covariance is important. We examine this by adding income to our basic regressions in Table 1. The NHIS asks about income in 9 categories (13 in 2000). We include dummy variables for each income bracket. There are endogeneity issues with income. Current income might be low because a person is sick, rather than the reverse—although the endogeneity problem is less clear for behaviors than for health. Nevertheless, we can interpret these variables as a sensitivity test for the potential role of income as a mediating factor.
The second columns in Table 1 report regressions including family income. Adding income accounts for some of the education effect. For example, the coefficient on years of education in the current smoking equation falls by 26 percent. The coefficient on body mass index falls by 16 percent (roughly the same as the fall in the coefficients on overweight and obese), and the coefficient on heavy drinking falls by 12 percent. The average decline (for outcomes with a significant gradient at baseline) is 12 percent. The mortality-weighted average is a decline of 24 percent. It is worth noting that our income measure includes both permanent and transitory income and further is measured with error. Thus, the reduction in education coefficients we observe might be too small.
The NHIS contains a number of other measures of economic status beyond current income, including major activity (whether individual is working, at home, in school, etc), whether the person is covered by health insurance,11 geographic measures (region and urban location), family size, and marital status. These variables are likely to determine permanent income and in principle can be affected by educational attainment.
As with income, each of these variables may be endogenous. Sicker people (or those with poor risky behaviors) may be more or less likely to get insurance, depending on the operation of public and private insurance markets. In each case, the coefficients on those variables may not be the ‘true effect’, and furthermore, including these variables may bias the coefficient of education. Still, the results are an important sensitivity test: the results are suggestive about what the largest effect of “resources” broadly construed may be.
The last column in Table 1 adds these additional economic controls to the regressions (in addition to income). As a group, these variables do not add much beyond income. The additional reduction in the education coefficient is 7 percent in the smoking regression, 11 percent for obesity, and 1 percent for heavy drinking. All told, the effect of material resources in the NHIS accounts for 20 to 30 percent of the education effect. 12 The reduction of 20–30 percent may be an underestimate of the true effect, because characteristics like permanent income are measured with error, or an overstatement, because we control for variables that are themselves influenced by education.
The NHIS does not have measures of wealth or family background. Further, measures of income in the NHIS are underreported, as in many surveys. To obtain better estimates of the possible effect of resources on the education gradient (beyond background), we repeated our analysis using the Health and Retirement Study, a sample of older adults. The economic data in the HRS are generally believed to be extremely accurate and HRS has family information as well, although only four health behaviors are asked about: smoking, diet/exercise, drinking, and preventive care.
Table 2 shows the HRS results. The first column shows results controlling for demographics and a large set of socioeconomic background measures: a dummy for father alive, father's age (current or at death), dummy for mother alive, mother's age (current or at death), father's education, mother's education, religion, self reported SES at age 16, self reported health at age 16, and dad's occupation at age 16. The HRS data show similar gradients to the NHIS data, though in some cases they are smaller. For example, smoking declines by 2 percentage points with each year of education, compared with 3 percentage points in the NHIS. In part, this reduction results from the fact we have added more extensive background controls as thus would be expected. If we used only the same basic demographics available in the NHIS, we would still find somewhat smaller gradients in the HRS (available upon request). Lower coefficients might also be due to selective mortality: lower educated individuals die younger and thus are less likely to be in the HRS. Although we do not know the reason, our finding that education gradients are smaller for older individuals has been noted elsewhere (see Cutler and Lleras-Muney 2008a for references).
In the middle columns of the table, we include economic controls: labor force status, total family income, family size, assets, major activity, region, MSA, and marital status. The reduction in the education coefficient ranges from 0 percent for flu shots to 25 percent for current drinking. The average reduction in the education effect is 20 percent, and the mortality-weighted reduction is 17 percent.
In total, therefore, we estimate that material resources account for about 20 percent of the impact of higher education on health behaviors, assuming that all our measures can be thought of as material resources. This matches what we find in other data sets as well (see below). With the understanding that this estimate is likely too high (because of endogeneity), we conclude that there is a large share of the education effect still to be explained.
Differences in prices or in response to prices are a second potential reason for education-related differences in health behaviors. This shows up most clearly in behaviors involving the medical system. In surveys, lower income people regularly report that time and money are major impediments to seeking medical care.13 Even given health insurance, out-of-pocket costs may be greater for the poor than for the rich – for example, their insurance might be less generous. Time prices to access care may be higher as well, if for example travel time is higher for the less educated.
A consideration of the behaviors in Table 1 suggests that price differences are unlikely to be the major explanation, however. While interacting with medical care or joining a gym costs money, other health-promoting behaviors save money: smoking, drinking, and overeating all cost more than their health-improving alternatives. It is possible that the better educated are more responsive to price than the less educated, explaining why they smoke less and are less obese. But that would not explain the findings for other behaviors which are costly but still show a favorable education gradient: having a radon detector or a smoke detector, for example. Still other behaviors have essentially no money or time cost, but still display very strong gradients: wearing a seat belt, for example.
More detailed analysis of the cigarette example shows that consideration of prices exacerbates the education differences. A number of studies show that less educated people have more elastic cigarette demand than do better educated people.14 Prices of cigarettes have increased substantially over time. Gruber (2001) shows that cigarette prices more than doubled in real terms between 1954 and 1999; counting the payments from tobacco companies to state governments enacted as part of the Master Settlement Agreement, real cigarette taxes are now at their highest level in the post-war era. Yet over the same time period, smoking rates among the better educated fell more than half, and smoking rates among the less educated declined by only one-third. For these reasons, we do not attribute any of the education gradient in health behaviors to prices.15
The next theory we explore is that education differences in behavior result from differences in what people know. Some information is almost always learned in school (advanced mathematics, for example). Other information could be more available to educated individuals because they read more. Still other information may be freely distributed, but believed more by the better educated. Most health information is of the latter type. Everyone has access to it, but not everyone internalizes it.
The possible importance of information is demonstrated by differences in how people learn about health news. Half of people with a high school degree or less get their information from a doctor, compared to one-third of those with at least some college.16 In contrast, 49 percent of people with some college report receiving their most useful health information from books, newspapers, or magazines, compared to 18 percent among the less educated.
The 1990 NHIS asks people 12 questions about the health risks of smoking and 7 questions about drinking (see the Data Appendix). In the smoking section, respondents were asked whether smoking increased the chances of getting several diseases (emphysema, bladder cancer, cancer of the larynx or voice box, cancer of the esophagus, chronic bronchitis and lung cancer). For those under 45, the survey also asked respondents if smoking increased the chances of miscarriage, stillbirth, premature birth and low birth weight; and also whether they knew that smoking increases the risk of stroke for women using birth control. In the heart disease module individuals were asked if smoking increases chances of heart disease. Similarly, respondents were asked whether heavy drinking increased one’s chances of getting throat cancer, cirrhosis of the liver, and cancer of the mouth. For those under 45, the survey also asked respondents if heavy drinking increased the chances of miscarriage, mental retardation, low birth weight and birth defects.
These questions are important, though they do suffer a (typical) flaw – the answer in each case is yes. Still, not everyone knows this. Table 3 shows the share of questions that the average person answered correctly, separated by education group. About three-quarters of people do not answer all questions correctly (not reported in the table). This seems low, but the answers are much better on common conditions. For example, 96 percent of people believe that smoking is related to lung cancer, and 92 percent believe it is related to heart disease. On average, individuals get 81 percent of smoking questions correct and 67 percent of drinking questions correct. There are some differences in responses by education, but often these are not that large. For example, 91 percent of high school dropouts report that smoking causes lung cancer, compared to 97 percent of those with a college degree. For heart disease, there is a bigger difference: 84 percent of high school dropouts versus 96 percent of the college educated believe smoking is related to heart disease.
Table 4 examines how important knowledge differences are for smoking and drinking. The first columns in the table show the gradient in poor behaviors associated with education when controlling for socioeconomic factors and income but not knowledge. The coefficients are roughly similar to those reported in the last specification of Table 1, although from a decade earlier.
As the next columns show, people who answer more smoking questions correctly are less likely to smoke. Indeed, answering all questions correctly eliminates smoking. Similarly, people who answer drinking questions correctly are less likely to drink heavily. But knowledge has only a modest impact on the education gradient in smoking and little impact on the gradient in drinking. The coefficient on years of education in explaining current smoking declines by 17 percent with the knowledge questions included, while the coefficient for drinking is essentially unaffected. The average reduction is between 5 and 18 percent, depending on the metric. These results thus suggest that specific knowledge is a source, but not the major source, of differences in smoking and drinking. These results are in line with those found by Meara (2001) and interestingly with those reported by Kenkel (1991), who attempted to account for the possibility that health knowledge is endogenous.17
Cognitive dissonance suggests an important caveat to these findings: individuals may differ in the extent to which they report they know about what is harmful as a function of their habits (for example smokers might report they don’t know as much). In the case of smoking Viscusi (1992) suggests that both smokers and non-smokers vastly overestimate the risks of smoking (though other studies find different results, see Schoenbaum 1997 for example). Most importantly here, it is not known whether these biases differ by education.
One potential concern about the knowledge questions is that we do not know the extent to which the answers reflect the depth of individuals’ beliefs. People may know what the correct answer is without believing it that strongly. For decades, tobacco producers sought to portray the issue of smoking and cancer as an unresolved debate, rather than a scientific fact. This might have had a greater impact on the beliefs of the less educated, for whom the methods of science are less clear.18
We have only a single piece of evidence along these lines. We examined self-reported questions from the Motor Vehicle Occupant Safety Survey (MVOSS), which asks people about the value of wearing a seat belt (results available upon request).19 Respondents are asked to strongly agree, somewhat agree, somewhat disagree, or strongly disagree with two questions about seat belt use: “If I were in an accident, I would want to have my seat belt on,” and “Seat belts are just as likely to harm you as help you.” A claim that seat belts harm people in an accident is commonly expressed by those who oppose mandatory seat belt legislation, somewhat akin to the ‘debate’ about the harms of tobacco.
Answers to the question about wanting a seat belt in an accident are uniformly high; 89 to 97 percent of people strongly or somewhat agree that they would want a seat belt on if they were in an accident. But there is still residual doubt about the value of a seat belt that is much more common among the less educated. Fifty-five percent of people with less than a high school degree strongly or somewhat agree that seat belts are just as likely to harm as help them, compared to only 17 percent of those with a college degree.20 These patterns suggest that superficially, individuals of all education levels have received the main public health message that one should wear a seat belt, and they report as much when asked. But uneducated individuals seem less certain of the validity of that information, and that becomes clear when the questions are asked slightly differently. Furthermore, we can “explain” a larger share of the effect of education on seat belt use when we include these alternative measures of “depth of knowledge”.
We cannot further examine this possibility here. We simply note that our results suggest that providing factual information alone may not be sufficient to make individuals change their behavior, and that differences in information alone are not sufficient to explain much of the education gradient in health behavior.
The tobacco and seat belt examples suggest that information processing, more than (or in addition to) exposure to knowledge, may be the key to explaining education gradients in behaviors. Similar arguments have been made to explain why education raises earnings in the labor market. Nelson and Phelps (1966) first hypothesized that “education is especially important to those functions requiring adaption to change” and that “the rate of return to education is greater the more technologically progressive is the economy.” This was echoed by Schultz (1975), who proposed that education enhances individuals’ “ability to deal with disequilibria” and Rosenzweig (1995), who argued that education improves individuals’ ability to “decipher” that information. All of these ideas can easily be applied in the context of health behaviors.
The existing literature provides some suggestions that cognitive ability is related to education gradients. For example more educated people are better able to use complex technologies/treatments than less educated individuals. Goldman and Smith (2002) document that the more educated are more likely to comply with HIV and diabetes treatments, which are extremely demanding. Rosenzweig and Schultz (1989) similarly show that contraceptive success rates are identical for all women for “easy” contraception methods such as the pill, but the rhythm method is much more effective among educated women. The more educated appear to be better at learning. Lleras-Muney and Lichtenberg (2005) find that, controlling for insurance, the more educated are more likely to use drugs more recently approved by the FDA, but this is only true for individuals who repeatedly purchase drugs for a given condition, so for those who have an opportunity to learn. Similarly Lakdawalla and Goldman (2005) and Case, Fertig and Paxson (2005) find that the health gradient is larger for chronic diseases, where learning is possible, than for acute diseases.
To examine the possibility that cognitive ability lies behind the education gradient in behavior, we turn to measures of general cognition.21 The NLSY administered the Armed Services Vocational Aptitude Battery (ASVAB) to all participants in 1979. The ASVAB is the basis for the Armed Forces Qualification Test (AFQT) but it contains many more dimensions than are scored in the AFQT. We include the test results for all 10 subjects, namely science, arithmetic, mathematical reasoning, word knowledge, paragraph comprehension, coding speed, numeric operations speed, auto and shop information, mechanical competence, and electronic information.22 Table 3 shows that those with a college degree or more scored much higher in the AFQT (73rd percentile on average) compared to high school dropouts (18th percentile).
Table 5 shows the relation between education, ASVAB scores, and a variety of health behaviors (smoking, diet, exercise, alcohol consumption, illegal drug use and preventive care). We use behaviors from relatively recent survey years, 1998 or 2002. The respondents thus range in age from the mid-30s to the mid-40s. Mean rates of favorable and poor health behaviors are shown in the first column; these percentages are close to those for the NHIS, particularly when restricted to the same ages.
We first document education gradients and the effects of economic resources in this sample. The first column shows the impact of education on behavior including only demographic and family background controls. The impact of education on behavior is large, often times larger than the NHIS. For example, each year of education is associated with a 4.9 percent lower probability of smoking and a 1.6 percent lower chance of being obese. The next column includes economic resources. There is generally a significant impact of these variables on the education gradient. Using the mortality weights noted above we estimate that 12 percent of the education gradient in mortality is explained with economic controls (alternative averages yield similar results).
The third column includes the individual ASVAB scores, in addition to the income and family background. The additional impact of these controls is substantial, though it varies by outcome. ASVAB scores account for an additional 15 percent of the education gradient in smoking, 9 percent of the gradient in obesity, and 10 percent of the gradient in heavy drinking. The overall average reduction varies depending on whether the illegal drug use variable is included or not. Including test scores exacerbates the education gradients in illegal drug use. It is not clear why this is the case, and is not true with the British data (discussed below).23 We also find that adding cognition increases the education gradient in preventive care. The reduction is about 20 percent without those variables but near zero (or negative) with those variables. Using the mortality weights, ASVAB scores explain 15 percent of the education effect. A central concern about these results is causality: is cognitive ability affected by education, or does cognitive ability lead people to become more educated? We return to this in Section IX.
While the estimates differ across specifications, our overall summary is that together knowledge and cognition account for 5 to 30 percent of the education gradient in behaviors, although cognition measures tend to increase education gradients in illegal drug use and preventive care, a puzzle which we do not resolve here.
The most common economic explanation for different behaviors is tastes. In our framework, tastes take the form of differences in discount rates, the value of the future, or risk aversion. The source of differences in utility functions is not clear. Education may lead people to have lower discount rates (Becker and Mulligan, 1997): for example if education raises future income, individuals have an incentive to invest in lowering their discount rate. Education may also lead people be more risk averse. Alternatively, education may itself be the product of differences in utility functions (Fuchs, 1982), which may be distributed randomly, may be inherited, or may be a product of the early childhood environment.
Some preliminary evidence suggests that differences in utility functions cannot be the primary explanation for differences in health behaviors. Were the difference in health behaviors driven by fixed aspects of individuals, we would expect that health behaviors would be highly correlated across individuals: people who care about their health would maximize longevity in all ways. However, while almost all health behaviors are related to education, these behaviors are not particularly highly correlated at the individual level. Cutler and Glaeser (2005) show that the correlation between different health behaviors is generally about 0.1. Still, we can investigate this hypothesis more directly.
We start first with the value of the future. Probably the best measures of discounting and of the value of the future come from the National Survey of Midlife Development in the United States, or MIDUS, a sample of people aged 25–74 in the mid-1990s.24 MIDUS has several measures of the value of the future. In an overall summary question about future expectations, individuals are asked “Looking ahead ten years into the future, what do you expect your life overall will be like at that time?”25 The same question is asked about current situation, which we include as well. There are some questions that can also be used as proxies for discount rates. Individuals were asked whether they agreed with the following statement: “I live one day at a time and don't really think about the future.” We code those who strongly disagree as being able to plan for the future. Theory suggests that that people with higher future utilities or who are able to plan will invest more in health, and possibly that there will be an interaction between the two (those who value the future and are good at planning will invest even more in health).
Table 3 shows summary measures of these variables by education. High school dropouts are indeed less future oriented than those with more than a college degree, but there appears to be no difference between high school graduates and those with some college only. The more educated are equally satisfied with their current life as the least educated, and those with some college report the lowest current satisfaction. The relationship between education and future satisfaction is also not linear, being the highest among the college educated, followed by high school graduates, those with some college and high school dropouts. Although these satisfaction measures are not very highly correlated with education, Figure 2 shows that the ratio of future to current satisfaction is monotonically increasing in education—the more educated value the future more relative to the present.26
MIDUS asks about some measures of health, though not as many as dedicated health surveys. It includes smoking and weight, though not alcohol consumption. Questions are also asked about general health behavior, illegal drug use, and receipt of preventive care.
Table 6 shows results from the MIDUS survey. The first columns report means of the independent variables. Where we can compare, the means are close to the NHIS. Using just demographic and family background measures as controls (the first column of regression coefficients) the education coefficients are also similar, if anything slightly larger. Each year of education reduces smoking by 3.5 percent and obesity by 1.6 percent.
The next columns show the impact of including economic resources. The impact is somewhat lower than the NHIS and NLSY. On average, 11 percent of education differences in behavior are attributable to economic resources.
The next column includes measures of current and future life satisfaction, the ability to plan for the future, and the interaction of planning and future life satisfaction, in addition to economic resources.27 There is no significant impact of these variables on education gradients. Indeed, in some cases the addition of these variables actually increases the effect of education. For the major outcomes we consider, smoking and obesity, the changes are 2 percent or less.
The measures of discount rates in the MIDUS are not ideal. Indeed, it is not entirely clear that there is a single measure of discounting that applies to all settings. To investigate whether there is variation in the appropriate measure, we use data from the Survey on Smoking (SOS), a sample of 663 individuals between 50 and 70 years of age.28 The SOS asks a variety of discounting questions (discussed below). The drawback of the SOS is the sample size and lack of many health questions (in addition to the fact that the sample is not nationally representative). For these reasons, we can only relate education to two outcomes – current smoker and obesity.
Table 7 shows the basic gradients in smoking and obesity in this sample. Education significantly lowers the likelihood of smoking and of being obese. Controlling for income (a dummy is used for each income category) lowers the smoking gradient by 9 percent and the obesity gradient by 21 percent.
We then look at the effect of adding various financial discounting measures. For our first measure of financial discounting, we use responses to 4 questions of the form “would you rather win (lose) $x now or $y a year from now?” The mean responses to these questions by education level are reported in Table 3. On average, individuals are very impatient (64% prefer $1000 now to $1500 in a year), and more so when the stakes are small (80% prefer $20 now to $30 in a year). When the questions refer to losing amounts, individuals are very impatient, but less than for gains. More importantly, for all the questions, more educated individuals are on average more patient (with the exception of the last question) as predicted by Fuchs. However, Table 7 shows that adding these discounting questions as regressors increases the magnitude of the coefficient on education for smoking and has no effect on obesity.
A second measure of discounting is the planning horizon that people use. Respondents were asked “in planning your savings and spending, which of the following time periods is most important to you and your family? (choices are “the next few months, the next year, the next few years, the next 5–10 years, longer than 10 years”). The answers were converted into numbers using the middle of the category. Table 3 shows that more educated individuals have longer planning horizons. Controlling for this measure lowers the coefficient on education in the smoking regression by 5 percent but increases the coefficient of education in the obesity regression.
The third set of measures of discounting are the answer to the questions “I spent a great deal of time on financial planning” and “I spent a great deal of time planning vacations”. More educated individuals are more likely to report that they agree than less educated individuals (Table 3) although the differences are small, especially for vacations. Adding the answers to these questions (a dummy for each possible answer: strongly agree, agree, agree somewhat, disagree somewhat, disagree, or disagree strongly or missing) has very little impact on our two measures of health.
Discounting may also take the form of impulsivity and lack of self-control, as suggested by Ross and Mirowsky (1999). More impulsive individuals may be less able to undertake actions with current costs but future gains, even if they know what is in their long-term interest. Individuals were asked a series of 14 questions, such as “I make hasty decisions”, “I do things on impulse that I later regret,” etc. Answers ranged from “disagree strongly” to “agree strongly”. We score the questions on a 1–5 scale and sum them, with an index that ranges from 14 (not impulsive) to 70 (greater impulsivity). High school dropouts are more impulsive than college graduates (Table 3). Adding the impulsivity index lowers the coefficient on education, but only by 3 percent for smoking and 6 percent for obesity.
It is possible that individuals discount health differently from money. A subset of the respondents was asked questions about time preferences for health: “20 extra days in perfect health this year would be just as good as ? extra days in perfect health X years from now”, where X was 1, 5, 10 and 20. As with financial discounting, the more educated are more patient, and the differences are greater for tradeoffs in the near future. Adding these questions to our regression lowers the coefficient on education by about 8 percent for smoking but increases the effect of education on obesity by 8 percent.
Even included together, the impact of these variables is not substantial. When all the discount measures are included, the coefficient on education falls by about 8 percent for smoking and 1 percent for obesity.
Neither MIDUS nor NHIS have measures of risk aversion. To investigate the role of risk aversion we use data from the Health and Retirement Survey (HRS). The HRS in 2002 asked hypothetical questions that allow for categorization of individuals into 4 risk aversion categories (Barsky et al., 1997). Respondents are first asked if they would risk taking a new job, given that family income is guaranteed now. The new job offers a chance to increase income but also carries the risk of loss of income. If the respondent says he/she would take the risk, the same scenario is presented, but with riskier odds. Risk aversion is scored on a 1 to 4 basis, from least to most risk averse (see the Appendix). Table 3 shows that education is not monotonically related to risk aversion; those with a high school degree are the most risk averse. This already suggests risk aversion is not a very promising factor in accounting for the education gradient.
More formal models are presented in Table 2. The addition of the risk aversion categories, shown in the last column of regressions, has virtually no impact on the education coefficient. The overall impact is within 1 percent. Indeed, the categories for risk aversion are not very consistently related to health behaviors. It may be that this measure of risk aversion is not ideal, but we do not have a way of testing this.29
All told, we attribute very little of the education gradient in health behaviors to utility function characteristics.
Even when people know what they want to do, translating intensions into actions may be easier for the better educated. We noted above the example of smoking: the better educated are more successful at quitting smoking than the less educated, not because they try to quit more frequently or use different methods, but because they are more successful when they do try.30 This parallels Rosenzweig and Schultz’s (1989) results on the success of contraceptive use. Many of these aspects of education were stressed by Grossman (1972); in his formulation, education allows inputs to be combined more productively.
One reason this might be the case is time constraints. The daily hassles of life (cooking, errands, children, etc.) may involve more intensive effort by the less educated, and hence leave them less time for health planning or the mental energy to devote to behavioral change. To test this theory, we looked at behaviors before and after retirement.31 If time constraints are a major issue, behavioral differences by education ought to decline after retirement, when leisure time increases. Results from the HRS (not shown) suggest this is not the case, however. The behavior of the more and the less educated does not change differentially after retirement, and in some cases the gradient increases.
Beyond time constraints, it may be that individuals differ in their psychological capacity to make behavioral changes. In many psychological theories, individuals need to be ‘ready’ to change, and feel able to do so. Depression or other psychological distress may hinder behavioral changes. Similarly, social integration and reinforcement may be helpful.
The NLSY asks a battery of questions about personality traits and sense of control. These include two self esteem scores (the Rosenberg self-esteem score, measured in 1980 and 1987), a score about one’s self-control (the Pearlin score, measured in 1992), a score about a sense of control over one’s life (the Rotter scale, measured in 1979), depression (the CES-D, administered in 1992 and 1994), and two indicators for whether the person is shy (one at age 6 and one in 1985). The Appendix discusses the questionnaires in more detail. Table 3 shows the mean of these variables by education. In general, there are differences in these measures across education groups, particularly in depression scales.
Table 5 shows the impact of adding the personality scales in the NLSY (in addition to economic resources). The impacts on exercise and regular doctor visits are among the largest effects (17 to 35 percent). But personality measures actually increase the gradient in illegal drug use measures and have minimal effects on smoking, drinking, and obesity. The average reduction in the education coefficient is 4 percent using the mortality weights (though a bit larger – as much as 13 percent – using other measures). This table suggests personality might matter for some outcomes. We explore this issue further with other data sets.
Some authors have posited that stress, depression, and anxiety are the mediating factor in behavioral changes (Salovey, Rothman, and Rodin, 1998). Individuals suffering from these conditions may not think their future will be very good or may not be able mentally to make behavioral changes. We have already included some of these measures in the previous NLSY analysis. But we have additional measures in other data sets. The MIDUS survey has several measures of whether individuals are under stress and whether they worry a lot. Table 3 shows that the less educated are under more stress than the better educated, but that extreme stress (answering yes to all three questions about stress) is relatively constant across education groups. This survey also contains a depression scale, an anxiety scale, a scale for sense of control, a scale for positive affect, and a scale for negative affect (the appendix shows how these are constructed). Table 6 shows that controlling for all of these measures (personality and stress) has no significant effect on the education gradients (again with a few exceptions); the overall change is essentially zero.32
Beyond individual attributes, we consider measures of social integration. The MIDUS asks a variety of questions about social integration, including scales for social ties, social contributions, positive and negative relations with spouse, and positive and negative relations with friends (see the appendix). These social measures pick up a number of different traits. Some part reflects individual personality—some individuals are more social than others. These measures also represent resources. Family and friends can be sources of information or reinforcement about behaviors. They can provide help in times of need or alternatively be the source of one’s troubles. They might also pick up other aspects of the environment such as the ability to meet other people easily. The questions in the MIDUS survey attempt to capture the extent of an individual’s social connections and the quality of these connections, both of which might matter. Interestingly many of these variables do not show steep education gradients, except for the extent to which individuals feel they are socially integrated and that they contribute to society (Table 3).
The final column of table 6 shows the impact of social integration on education gradients in behaviors in the MIDUS. There is a modest impact of these social integration measures. The coefficient on current smoking falls by 9 percent when social integration measures are added, and the coefficient on obesity falls by 3 percent. The average effect, shown in the last rows of the table, is 7 to 22 percent.
Overall we find that the vast bulk of personality measures relating to sense of control, stress, and psychological impairment account for very little of the education gradient. On the other hand our measures of social integration do account for a part of the gradient, though it is not entirely clear why they matter.33
Our results to this point have focused on the United States. As noted earlier, education gradients are pervasive in the developed (and developing) world. Analyzing data from other countries can help determine if the results in the United States carry over in other settings.
Data from the National Child Development Study (NCDS) in the United Kingdom are available to address these issues. The NCDS is a study of everyone born in a given week in Great Britain in 1958. We use data from the 6th interview wave, conducted in 1999–2000, when the participants were 41–42 years old. Nearly 6,500 people are surveyed. Years of schooling is a less meaningful measure in the U.K. than it is in the U.S. We form a dichotomous variable for whether the person passed the A levels, roughly equivalent to a college degree in the U.S.
The NCDS contains a number of health measures, detailed in the first column of Table 8. The four biggest risk factors are all asked about: smoking, drinking, diet/exercise, and illegal drug use. On many measures, people in the U.K. are comparable to the U.S. Smoking rates are similar, though a bit higher in the U.K., while obesity rates are somewhat lower. Because of its longitudinal nature, the NCDS has a large set of income and background controls. These include height at age 15, birth weight, SES of father at birth, age 7, 11, and 16, marital status of mother at birth, mother’s and father’s birthplace, own birthplace, and mother’s and father’s education. Because these were collected during earlier waves, they are less likely to be misreported than in surveys such as the HRS, which asks respondents about these measures retrospectively.
The first set of regression results relates behaviors to demographic and background controls only. As in the US, more education is associated with better health behaviors in the U.K. (though our measures of education are not quite comparable). Passing the A levels is associated with a 12 percent lower probability of smoking and a 4 percent lower probability of being obese. As in the U.S. more educated individuals are more likely to drink (1 percent), but less likely to be heavy drinkers (3 percent). The next column shows the impact of adding economic controls. As in the U.S., these controls have a significant impact on the education gradient in behaviors. The impact of education on current smoking falls by 21 percent, but the impact of education on weight measures increases. The average reduction is between 17 and 24 percent, depending on the measure used. This degree of explanatory power is somewhat greater than in the U.S. but not much.
The NCDS has a number of tests of cognitive ability. Cognitive tests were administered at age 7 (math and drawing), age 11 (reading, math, verbal, non-verbal, and drawing), and age 16 (math and reading comprehension). The next column of the table includes the results of all these cognitive tests. As in the U.S., scores on cognitive tests predict a significant part of the education gradient. Controlling for cognitive ability reduces the impact of education on current smoking by 45 percent and the impact on obesity by 18 percent. The share of the education effect that is attributable to cognitive ability ranges between 15 and 44 percent.
The NCDS has measures of current and expected future life satisfaction (each is a scale from 1–10 where 10 is the highest; see the appendix), although there are no measures of discount rates. The next column shows that life satisfaction does not affect the education gradient. The average decline is 1 to 2 percent, roughly the same as in the U.S.
The NCDS also has several personality measures. There are three measures of self-efficacy: whether the respondent gets what they want out of life, how much control they have over life, and whether they can run their life how they want. These variables are most related to the self-esteem and self-control measures in the NLSY. The survey also contains two scales that measure mental health and stress: the Malaise index and the General Health Questionnaire (GHQ12). The impact of adding these variables is shown in the next column of the table. Relative to economic and background controls only, personality controls have a negligible impact on the education gradient in behaviors. The overall effect is about 2 percent change in any of the average measures.
Finally, the NCDS has a variety of measures of social integration: whether the respondent’s parents are alive, whether the respondent sees their parents, and whether they frequently eat together as a family, visit relatives, go out as a family, spend holidays as a family, go out alone or with friends, and attend religious services. These differ in nature from those in the MIDUS: they capture frequency of interactions, but not their quality. The next column of the table presents the results from adding these measures. Again we find that social measures have an impact on the education gradient in behaviors, reducing the coefficient by about 15 percent (in comparison to the 7 percent in the U.S.).
The final column of the table shows the combined impact of cognitive ability, future valuation, personality factors, and social integration on the education gradient in behavior. The cumulative impact is 48 percent using the weighted measure and less with the unweighted ones. Along with the 24 percent of the education gradient that is attributable to economic and background factors, we can account for up to 72 percent of the education gradient in health behaviors. Overall these results from the U.K. are remarkably similar to those from the U.S. data.
One of our most interesting results is that a non-trivial share of the education gradient in health behaviors can be accounted for cognition measures. Previous literature has considered whether the relationship between education and health (rather than health behaviors) is mediated by cognition, and finds mixed results. Most notably, Auld and Sidhu (2005) find that including test scores has a large effect on the education gradient in self-reported health status, whereas Grossman (1975) finds that it does not.
Causality is a central issue in this debate. It may be that education leads to greater intelligence (by this we mean better decision making abilities), and that intelligence matters for outcomes—we term this the learning channel. An equally plausible hypothesis is that people who are more intelligent go on to more education, and education matters for outcomes. Alternatively, there may be some third factor that influences both education and cognitive ability and also determines health behaviors. Of course these mechanisms are not mutually exclusive. To trace out these pathways one would need to estimate causal effects of education and cognition on health (or health behaviors), as well as causal effects of cognition on education and vice-versa. The studies we know of cannot establish all of these, nor can we.34
In this section we focus instead on whether there is any evidence for the learning channel: the idea that education is causally related to health because of its impact on cognition. Some previous work supports this idea. For example, several studies point out that education seems to have a causal effect on health (as discussed in the introduction). In addition, other studies find causal evidence that schooling increases AFQT (or other measures of cognition). For example, Hansen, Heckman and Mullen (2004) find that that one year of schooling increases AFQT scores between 2 and 4 percentage points (see also Neal and Johnson, 1996, and Winship and Korenman, 1997).35 Finally cognition is associated with better health and health behaviors (Gottfredson and Deary, 2004), although we know of no causal evidence. Together this evidence points to a causal effect of education on health and health behaviors through increases in cognition.
We can present some additional, albeit imperfect, evidence that is consistent with the learning channel using our data sets. In particular, both the NLSY and the NCDS have test scores taken at different ages. A finding that cognitive ability at later ages is more important in mediating the education effect would suggest that education influences later life cognitive ability, which in turn explains differences in health behavior. If cognitive behavior at younger ages were more important, in contrast, it would suggest that early cognitive ability influences education and health behaviors.
Table 9 presents the results using a small subsample of the NLSY for which early test score measures are available.36 For most outcomes the effect of including late IQ is much larger than that of early IQ. Overall, late IQ (controlling for early IQ) reduces the effect of education by 8 to 32 percent, whereas controlling for early IQ alone has no effect on average.
We can repeat this exercise using the British data as well, which has test scores for all individuals in the sample at ages 7, 11 and 16. These data are better suited for this exercise because of the larger sample, the fact that all individuals were administered the same test and that the tests are available at 3 different ages rather than 2. Table 10 shows the results. The pattern of the cognitive test scores again suggests that education is causally related to behaviors, rather than the reverse. Adding cognitive test scores at age 7 often increases education gradients and on average has no effect. Conditional on test scores at age 7 and background measures, adding test scores at age 11 reduces the effect of schooling on average by 14–23 percent. But together test performance at age 11 and 16 reduce the coefficient on A levels by 22–45 percent relative to its size when income, background and test performance at age 7 are accounted for. To the extent that performance in these test reflects learning in school, the results suggests that what is learned from age 7 to 11, and then from age 11 to 16 accounts for a significant portion of the education gradient.
Finally we examine the types of cognitive abilities that appear to “explain” the effect of education on behaviors. Using the 1992 HRS we investigated how different commonly used measures of cognition among adults and the elderly affect the education gradient in behaviors.37 Table 11 shows the results. We find that indicators of higher level processing (such as scores on the WAIS test38 or self reports of one’s ability to read a map, follow instructions or use computers) reduce the education gradient, whereas memory measures (the ability to recall a list of words for example) do not appear to account for any of the education gradient.39
Similarly we also found that vocabulary and spelling test scores at age 16 in the British Cohort Survey (results available upon request) did not impact education gradients, while math scores did. In the NLSY, most components of the ASVAB test scores (math, science, verbal, speed, or vocational) account for about an equal reduction in the education gradient, but the effects are quite heterogenous depending on the outcome of interest (results available upon request). Overall it would appear that measures of abstract thinking, rather than memory-based or knowledge-based questions, are more important in explaining the education gradient.
Using a variety of data sets in two countries, we examine the relation between education and health behaviors. Education gradients in health behaviors are large; controlling for age, gender, and parental background, better educated people are less likely to smoke, less likely to be obese, less likely to be heavy drinkers, more likely to drive safely and live in a safe house, and more likely to use preventive care. Given the similarity across so many different behaviors, we focus on broad explanations for health behaviors, rather than explanations specific to any particular behavior.
With a number of different theories, we are able to account for a good share of the education gradient. Table 12 summarizes our quantitative results. Resources are an important first factor. Income, health insurance, and other economic indicators account for 11 to 32 percent of the education gradient in behavior; a consensus estimate is perhaps 20 percent.
Our most surprising result is that education seems to influence cognitive ability, and cognitive ability in turn leads to healthier behaviors. As best we can tell, the impact of cognitive ability is not so much what one knows, but how one processes information. Everyone ‘knows’ that smoking is bad and seat belts are useful, but the better educated may understand it better. We estimate that cognitive ability is about as important as resources in accounting for health behaviors; a guess is about 30 percent. Specific knowledge by contrast accounts for about 12 percent of the gradient.
Many economic theories stress the role of tastes in accounting for behavioral differences: better educated people will have lower discount rates or risk aversion than the less educated. Our proxies for these taste parameters are possibly measured with error, though we attempted to obtain the best measures available. Nevertheless none of our proxies for discounting, risk aversion, or the value of future explain any of the education gradient in health behaviors.
The theory that is most difficult to test is the translation theory: more and less educated people each want to improve health behaviors, but carrying out these intentions is difficult. Our data do not support the hypothesis that self-esteem, sense of control, stress, depression, or anxiety are important mediating factors in the education gradient. But the social environment does appear to be somewhat healthier for the better educated. In both the U.S. and U.K., the degree of social integration accounts for about 11 percent of the education gradient in behavior.
All told, our different theories account for 60 to 80 percent of the education gradient. This is a very high share, given the magnitude of these effects and the persistent inability of previous research to make sense of these gradients. The explanation for the remaining one-quarter to one-third of the education gradient is a topic for future research. Our results suggest several possible candidates. First, measurement error in the various proxies we use may explain why in some data sets some mechanisms matter more than in others—in the data sets where income and background are better measured, they account for a larger share of the gradient, and the same is true for cognition. However, regardless of how many different proxies for personality or discounting we had, we did not find these mattered.
Another possibility is that there are important peer effects. The existence of peer effects cannot explain why educated groups adopt better behaviors than uneducated groups to begin with, but peer effects can magnify the effects of education. Finally we did not explore the possibility of interactions between our different mechanisms. It is possible that cognition matters only when individuals have knowledge, or that income matters less (or more) for those who are well-integrated in society.
We are grateful to Lisa Vura-Weiss, Tom Vogl and Rebecca Lowry for excellent research assistance, to Frank Sloan for generously sharing the Survey on Smoking data with us, to Chris Winship and Alan Block for help with data and programs, and the National Institutes on Aging for research support. This work greatly benefitted from comments from Michael Grossman, Joe Newhouse and the seminar participants at Princeton, the NBER, Duke and Harvard.
|National Health Interview|
|1990, 1991, 1992,|
|Whites aged 25 and older.|
Survey of Youth 1979
|1998, 2002||All white respondents with non-|
missing dependent and independent
|National Survey of Midlife|
Development in the
United States (MIDUS)
|1995–1996||All white respondents with non-|
missing dependent and independent
|Health and Retirement|
|1996||All white respondents aged 55–64 in|
wave 3 who did not have a proxy
interviews, and with no missing
values for risk questions.
|Survey on Smoking (SOS)||2005||Adults ages 50 to 70 who were|
current, former or never smokers
|Wave 6, year 2000||All respondents|
Brim, Orville G., Paul B. Baltes, Larry L. Bumpass, Paul D. Cleary, David L. Featherman, William R. Hazzard, Ronald C. Kessler, Margie E. Lachman, Hazel Rose Markus, Michael G. Marmot, Alice S. Rossi, Carol D. Ryff, and Richard A. Shweder. NATIONAL SURVEY OF MIDLIFE DEVELOPMENT IN THE UNITED STATES (MIDUS), 1995–1996 [Computer file]. 2nd ICPSR version. Ann Arbor, MI: DataStat, Inc./Boston, MA: Harvard Medical School, Dept. of Health Care Policy [producers], 1996. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2003.
Economic and Social Data Service, “SN 5560 -National Child Development Study Response Dataset, 1958–2005”, http://www.esds.ac.uk/findingData/snDescription.asp?sn=5560, accessed August 20, 2007.
Leacock, Cindi, “Getting Started with the Health and Retirement Study”, Ann Arbor: University of Michigan, 1996.
U.S. Dept. of Health and Human Services, National Center for Health Statistics. NATIONAL HEALTH INTERVIEW SURVEY, 1987: CANCER CONTROL STUDY [Computer file]. Washington, DC: U.S. Dept. of Commerce, Bureau of the Census, and Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producers], 1989. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 1990.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1990: HEALTH PROMOTION AND DISEASE PREVENTION (HPDP) SAMPLE PERSON SUPPLEMENT [Computer file]. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1990. Ann Arbor, MI: Interuniversity Consortium for Political and Social Research [distributor], 1993.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1991: AIDS KNOWLEDGE AND ATTITUDES SUPPLEMENT [Computer file]. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1992. Ann Arbor, MI: Interuniversity Consortium for Political and Social Research [distributor], 1993.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1991: HEALTH PROMOTION AND DISEASE PREVENTION SUPPLEMENT [Computer file]. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1992. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 1993.
---------. NATIONALHEALTH INTERVIEW SURVEY, 1992: CANCER CONTROL SUPPLEMENT [Computer file]. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1993. Ann Arbor, MI: Interuniversity Consortium for Political and Social Research [distributor], 1994.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1994 : YEAR 2000 OBJECTIVES SUPPLEMENT [Computer file]. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1996. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 1997.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1998 [Computer file]. ICPSR version. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1999. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2001.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 2000 [Computer file].2nd ICPSR version. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 2000. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2002.
---------. NATIONAL HEALTH INTERVIEW SURVEY, 1991: DRUG AND ALCOHOL USE SUPPLEMENT [Computer file]. ICPSR version. Hyattsville, MD: U.S. Dept. of Health and Human Services, National Center for Health Statistics [producer], 1991. Ann Arbor, MI: Inter-university Consortium for Political and Social Research [distributor], 2002.
U.S. Department of Labor, Bureau of Labor Statistics, “The NLSY79”, http://www.bls.gov/nls/handbook/2005/nlshc3.pdf, accessed August 20, 2007.
National Health Interview Survey (NHIS)
Knowledge of the harms of smoking and drinking
The 1987 NHIS asks three types of questions about smoking knowledge.
The correct answer to each question is yes.
In addition we use the answer to the following question: “I am going to read a list of things which may or may not affect a person’s chances of getting heart disease. After each one, tell me if you think it definitively increases, probably increases, probably does not, or definitively does not increase a person’s chances of getting heart disease—First cigarette smoking?”
Individuals who answered definitively or probably were coded as answered correctly.
The percent of questions answered correctly is calculated using all 12 questions for people aged 44 and lower, and for the first 7 questions for people aged 45 and older.
Two questions are asked about the harms of drinking:
The correct answer to each question is yes. The percent of answers that are correct is calculated for all 7 drinking questions for those age 44 and younger, and for the first three questions for paged aged 45 and older.
Depression and anxiety scales
ANXIETY: Anxiety scale is based on the following questions:
Answers were coded as follows: ALL of the time=4, MOST of the time=3 SOME of the time=2, A LITTLE of the time=1, NONE of the time=0
DEPRESSION: the depression scale is based on the following questions:
Answers were coded as follows: ALL of the time=4, MOST of the time=3 SOME of the time=2, A LITTLE of the time=1, NONE of the time=0
National Longitudinal Survey of Youth (NLSY)
Rosenberg Self-Esteem Scale (measured in 1980 and 1987)
Now I’m going to read a list of opinions people have about themselves. After I read each one I want you to tell me how much you agree or disagree with these opinions.
Pearlin Mastery Scale (1992)
“The next series of sentences describes the way some people feel about how much control they have over their lives. After each statement, please tell me whether you strongly disagree, disagree, agree or strongly agree….
The scale is constructed by adding together the responses from each item; thus, a range of 7 to 28 is possible.
Rotter Scale (1979)
We would like to find out whether people’s outlook on life has any effect on the kind of jobs they have, the way they look for work, how much they work, and matters of that kind. On each of these cards is a pair of statements numbered 1 and 2. (hand respondent card booklet 2.) For each pair, please select one statement which is closer to your opinion. In addition, tell me whether the statement you select is much closer to your opinion or slightly closer.
Respondents were asked to select one of each of the paired statements and decide if the selected statement was much closer or slightly closer to their opinion of themselves. The following shows how the scale is constructed:
|Internal Control Item||External Control Item|
|Much closer||Slightly closer||Slightly closer||Much closer|
Each of the four paired items is constructed in the same manner as the above example. The values for each item are then summed. The maximum possible score is 16, indicating high external control, while the minimum possible score is four, indicating high internal control.
Depression scale CESD
“Now I am going to read a list of the ways you might have felt or behaved recently.
After each statement, please tell me how often you felt this way during the past week…
1994: only asked questions 2, 5, 6, 7, 11, 18, 20.
We compute depression using the identical set of questions in both years.
The answers (after positives are reversed) are coded as
The scale is constructed by summing over answers. For the full scale 60 is the maximum possible score.
Age 6: “Thinking about when you were 6 years old, would you describe yourself as: 1 extremely shy 2 somewhat shy 3 somewhat outgoing 4 extremely outgoing”
Adult: “Thinking about yourself as an adult, would you describe yourself as: 1 extremely shy 2 somewhat shy 3 somewhat outgoing 4 extremely outgoing”
National Survey of Midlife Development in the United States (MIDUS)
The MIDUS survey asks several questions about the value of the present and future and discount rates:
Value of the future: “Looking ahead ten years into the future, what do you expect your life overall will be like at that time?” Coded from 0 (the worst possible) to 10 (the best possible).
Value of the present: “How would you rate your life overall these days?” Coded from 0 (the worst possible) to 10 (the best possible).
Discount rates: “I live one day at a time and don't really think about the future.”
1. Depression scale: Depression [DEPCON]: (continuous variable based on [DEP] and [DEP2]) : ranging from 0 to 7 : = 0, if a respondent was diagnosed as negative for both depressed affect and anhedonia: i.e., DEPX=0 and DEPZ=0.
[DEP]: (continuous variable based on 7 items)
(During two weeks in past 12 months, when you felt sad, blue, or depressed, did you)
Scaling: [DEP] was constructed by taking the number of “Yes” responses to the items.
[DEPX]: (dummy variable based on QA58, QA59, and [DEP]) = 1 if : The feeling of being sad, blue, or depressed lasted “All day long” or “Most of the day” (A58), AND You feel this way “Everyday” or “Almost every day” (A59), AND [DEP] is greater than or equal to “4.” = 0, otherwise.
[DEP2]: (continuous variable based on 7 items)
(During two weeks in past 12 months, when you lost interest in most things, did you)
Scaling: [DEP2] was constructed by taking the number of “Yes” responses to the items. [DEPZ]: (dummy variable based on QA70, QA71, and [DEP2]) = 1 if: The loss of interest in most things lasted “All day long” or “Most of the day” (A70), AND You feel this way “Everyday” or “Almost every day” (A71), AND [DEP2] is greater than or equal to “4.” = 0 otherwise.
2. Anxiety Disorder: [GADCON]: (continuous variable based on 10 items) How often - over the past 12 months-, you)
- A respondent answered s/he worries “A lot more” than most people (A80a), AND worried “Every day, Just about every day, or Most days” (A81), AND worries about “More than one thing” (A82), OR has different worries “At the same time” (A82a) Coding: 1 most days; 2 about half the days; 3 less than half the days; 4 never Scaling: [GADCON] was constructed by taking the number of “Most days” responses to the items.
3. Positive Affect [GOODMOOD]:
(During the past 30 days, how much of the time did you feel…)
Coding: 1 All of the time; 2 Most of the time; 3 Some of the time; 4 A little of the time; 5 none of the time.
4. Negative Affect [BADMOOD]:
(During the past 30 days, how much of the time did you feel…)
Coding: 1 All of the time; 2 Most of the time; 3 Some of the time; 4 A little of the time; 5 none of the time.
5. Perceived Control (combined scale):
Items: 12-item scale combining the 4 “personal mastery” items and the 8 “perceived constraints” items.
Scaling: Scale score can be constructed by calculating the mean of the 12 items. Items from “personal mastery” need to be reverse-coded so that higher scores represent higher levels of the overall perceived control.
Personal Mastery [MASTERY]:
Perceived Constraints [CONSTRNT]:
Coding: 1 Strongly agree; 2 Somewhat agree; 3 A little agree; 4 Don’t know; 5 A little disagree; 6 Somewhat disagree; 7 Strongly disagree.
Scaling: [MASTERY] and [CONSTRNT] were constructed by calculating the mean across each set of items. Items were recoded so that high scores reflect higher standing in each dimension. The scales can be constructed by calculating the sum of the reverse-coded values of the items.
6. Social Integration [SWBSI]:
7. Social Contribution [SWBSC]:
8. Family Support [KINPOS]:
Coding: 1 A lot; 2 Some; 3 A little; 4 Not at all.
Scaling: [KINPOS] was constructed by calculating the mean of four items. Items were recoded so that higher scores reflect higher support. The scale can be constructed by calculating the sum of reverse-coded values of the items.
9. Family Strain [KINNEG]:
Coding: 1 Often; 2 Sometimes; 3 Rarely; 4 Never.
Scaling: [KINNEG] was constructed by calculating the mean of four items. Items were recoded so that higher scores reflect higher strain. The scale can be constructed by calculating the sum of reverse-coded values of the items.
10. Friend Support [FDSPOS]:
Coding: 1 A lot; 2 Some; 3 A little; 4 Not at all.
Scaling: [FDSPOS] was constructed by calculating the mean of four items. Items were recoded so that higher scores reflect higher support. The scale can be constructed by calculating the sum of reverse-coded values of the items.
11. Friend Strain [FDSNEG]:
Coding: 1 Often; 2 Sometimes; 3 Rarely; 4 Never.
Scaling: [KDSNEG] was constructed by calculating the mean of four items. Items were recoded so that higher scores reflect higher strain. The scale can be constructed by calculating the sum of reverse-coded values of the items.
12. Spouse/ Partner Support (Marital Empathy Scale) [SPEMP]:
Coding: 1 A lot; 2 Some; 3 A little; 4; not at all.
Scaling: [SPEMP] was constructed by calculating the mean of six items. Items were recoded so that higher scores reflect higher support. The scale can be constructed by calculating the sum of reverse-coded values of the items.
13. Spouse/Partner Strain [SPCRIT]:
Coding: 1 Often; 2 Sometimes; 3 Rarely; 4 Never.
Scaling: [SPCRIT] was constructed by calculating the mean of six items. Items were recoded so that higher scores reflect higher strain. The scale can be constructed by calculating the sum of reverse-coded values of the items.
Health and Retirement Study (HRS)
The HRS has a measure of risk aversion based on job choice. R is asked if he/she would risk taking a new job, given that family income is guaranteed now. The new job offers a chance to increase income but also carries the risk of loss of income. If R says he/she would take the risk, the same scenario but with riskier odds is presented. If R says he/she would not take the risk, the same scenario with less risky odds is asked.
The question wording is: Now I have another kind of question. Suppose that you are the only income earner in the family, and you have a good job guaranteed to give you your current (family) income every year for life. You are given the opportunity to take a new and equally good job, with a 50-50 chance it will double your (family) income and a 50-50 chance that it will cut your (family) income by a third. Would you take the new job?
If yes, then: Suppose the chances were 50-50 that it would double your (family) income, and 50-50 that it would cut it in half. Would you still take the new job?
If no, then: Suppose the chances were 50-50 that it would double your (family) income and 50-50 that it would cut it by 20 percent. Would you then take the new job?
This variable is set using the following:
Survey on Smoking (SOS)
It is constructed using the answer to the following questions:
For each statement, please indicate whether you disagree strongly, disagree, neither agree nor disagree, agree, or agree strongly.
National Child Development Study (NCDS)
GHQ12--General Health Questionnaire
The General Health Questionnaire (GHQ) is a self-administered screening test, designed to identify short-term changes in mental health (depression, anxiety, social dysfunction and somatic symptoms). It is a pure state measure, responding to how much a subject feels that their present state "over the past few weeks" is unlike their usual state. The 12 questions are
The final score is equal to the sum of the answers for each question (thus the score goes from 0 to 12). Questions are coded as 1 if individuals answered that their present state is unlike their usual state.
The Malaise Inventory provides a measure of for assessing psychiatric morbidity, developed by the Rutter and others at the Institute of Psychiatry from the Cornell Medical Index (Rutter M, Tizard J, and Whitemore K (1970) Education, Health and Behaviour. London). The 24 variables included on the dataset are identified below:
Malaise Index is calculated as the sum of the positive responses to the 24 items
The value of the present and future
Lifesat1: “Here is a scale from 0–10 where '0' means that you are completely dissatisfied and '10' means that you are completely satisfied. Please enter the number which corresponds with how satisfied or dissatisfied you are about the way you life has turned out so far.”
lifesat2: “Please use the scale once more to show how you expect to be in ten years time.”
The last few questions are to do with how you feel about your life so far. Please enter the number next to the statement which is most true for you.
Publisher's Disclaimer: This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final citable form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
1See Cutler and Lleras-Muney (2007) for additional references.
2Observed health behaviors however do not explain all of the differences in health status by education or other SES measures. We do not focus on this issue in this paper.
3Formal explanations for this phenomenon date from Grossman (1972) although there was less formal discussion earlier.
4The only exception would be BMI which is measured in the NHANES and which we do not use here because it contains no proxies to test our theories.
5There is no straightforward way to compute years of schooling using the information that is asked of respondents in Britain. Although using a dichotomous variable makes it difficult to compare the results to those for the U.S., we preferred this measure.
6For example we control for parental education, under the assumption that parental education is mostly determined prior to children’s education and that mothers and fathers do not make education decisions taking into account the possibility that their own education will determine their children’s education as well.
7Since the regression is a logit, the impact of changes in the X variables is non-linear. We evaluate the derivative around the average 10 year mortality rate in the population, 10.7 percent. We hold this rate constant in all data sets, even when age and other demographics differs.
8This methodology estimates a common education effect across outcomes, after standardizing the variables to have mean=0 and standard deviation=1. In each case, outcomes are redefined so that a higher outcome constitutes an improvement. Only outcomes that are defined for the entire population are included (so for example mammogram exam is excluded since it pertains to women only). The average effect of education is then computed as the unweighted average of the coefficient on education on each of the standardized outcomes.
9Later analyses use other years as well, specifically 1987 and 1992.
10Observed and self-reported obesity are not entirely similar. Measured obesity rates are generally 3 to 4 percent higher than self-reported rates (Cawley, 2004; Cawley and Burkhauser, 2006). Still, the two are highly correlated.
11Different health variables are available in different NHIS surveys, not all of which have information on health insurance. We note in the table which regressions do not have controls for health insurance.
12Note that since these outcomes come from different surveys we cannot compute the third overall measure of the effect of education which we report in subsequent tables.
13A variety of surveys show this response, including the 1987 NHIS Cancer Control Supplement.
14Gruber and Koszegi (2004) estimate elasticities of −1 for people without a high school degree, −0.9 for high school grads, −0.1 for people with some college, and −0.4 for college grads. Chaloupka (1991) estimates elasticities of −0.6 for people with a high school degree or less and −0.15 for people with more than high school.
15Obesity might be an exception. Food prices have fallen over time, especially for processed foods. Still, Cutler, Glaeser, and Shapiro (2003) argue that falling time prices are more important than monetary costs in explaining increased obesity.
16These data are from the 1987 NHIS Cancer Control Supplement. The question was open ended; people were allowed to give multiple answers. We report the share of people volunteering the indicated response.
17Kenkel instrumented for health knowledge with variation including receipt of physician advice about lifestyle-related topics, industry and occupation dummies, and a dummy for employment in a health-related field. For smoking, years of schooling after 1964 are also included as an instrumental variable.
18In the General Social Survey, for example, about 15 percent of people with less than a high school degree had a “clear understanding” of scientific study, compared to nearly 50 percent of college graduates. Similarly, fewer than 10 percent of people with less than a high school degree can describe the use of a control group in a drug trial, compared to nearly one-third of college graduates. About one-third of the less educated reported “a great deal” of confidence in science, compared to over 50 percent of those with a college degree.
19We are grateful to Alan Block of the National Highway Traffic Safety Administration for making these data available to us.
20Scientifically, it is true that it is better not to be wearing a seat belt in some accidents, but it is more helpful to wear one on the whole.
21There is debate in the literature about whether these tests are IQ tests or not. For our purposes, this is not relevant. We term them measures of cognition as a general descriptor.
22The specifics of the AFQT have changed over time. Currently, it is a combination of word knowledge, paragraph comprehension, arithmetic reasoning, and mathematical knowledge.
23We have explored this in other data sets, as we are able. The British Cohort Study (BCS) is similar to the National Child Development Study; it surveys everyone born in England, Scotland, and Wales in one week in 1970. Measures of test scores in the BCS do not exacerbate the education gradient in illegal drug use.
24MIDUS was conducted in 1995–96 as part of a MacArthur Foundation Aging Network. Within the 25–74 year-old population, it is representative of the population as a whole, although the survey was on paper and was very long. Hence, response rates at the top and bottom of the income spectrum were relatively low (MIDMAC, 1999). There are about 3,000 observations in MIDUS, although for certain outcomes the sample is considerably smaller.
25Individuals were also asked to evaluate what various aspects of their lives might be like in the future, in several dimensions (health, willingness to learn, energy, caring, wisdom, knowledge, work, finances, relationship with others, marriage, sex and relationship with children). We investigated whether results differed when using these more detailed questions, but found essentially no difference, in terms of the education gradient. Similarly, there are other possible proxies for how future oriented individuals are. The results are not affected by the choice of proxy.
26These results could be explained if, relative to those who attended but did not complete college, high school graduates are better decision makers. Means from other data sets for example for AFQT do not suggest that this is the case, however.
27We estimated different versions of these regressions, using dummy variables for each category and making use of more detailed questions about current and future satisfaction that were asked in the survey (respondents ranked their overall life satisfaction but also their satisfaction with their health, finances, relationships, etc). The results from these alternative estimations were nearly identical to the ones presented here.
28We are grateful to Frank Sloan for providing us these data. See Khwaja et al (2007) for a description.
29We also estimated models where we included seat belt use as an explanatory variable as a proxy for discount rates or risk aversion. The results are very similar to those reported here.
30These results are from tabulations of the 2000 NHIS.
31One could alternatively consider time diaries, but the reporting of these is notoriously incomplete.
32The NHIS also contains information about depression and anxiety in 2000. We examined how these variables affect the education gradient for behaviors measured that year. Results from these regressions are in Appendix Table 2. The addition of these controls has a small effect of the education coefficient. The average across all outcomes is a reduction of 1 percent, and the mortality weighted average is 4 percent.
33Our regressions control for income, which may be endogenous, but the qualitative results are unaffected by this choice. Appendix Table 4 reports the NLSY results without income controls. The results are very similar to those in Table 5.
34Some papers have also explored interactions between education and IQ, see for example Elias (2004) or Auld and Sidhu (2005). A structural approach to the production of education and health, that includes the possibility that education and IQ are produced jointly, could be used to make some progress on the relationship between education, IQ and health. But these models depend on functional forms and are difficult to estimate.
35Similarly, Behrman et al. (2008) estimate that schooling as well as pre- and post-schooling experiences influence adult cognition.
36We follow Winship and Korenman (1997) and control for the type of test and the year that the test was taken when early IQ measures are included. We omit results for ever tried illegal drug use, since the education gradients increase when IQ is included in these regressions.
37We use a different HRS sample because it has a large set of cognitive measures for a large sample. Thus slightly different controls and dependent variables are used.
38The WAIS test score assesses higher level abstract reasoning. Each respondent is given seven pairs of words and asked to describe the way in which the items are alike.
39Other studies report similar results among diabetics in the HRS. Sloan and Ayyagari (2008) find that cognition mediates some of the effect of education on self-reported health status among diabetics. Goldman and Smith (2002) report that all of the effect of education on the probability that diabetics adhere to their treatment can be accounted for by controlling for the WAIS score, the same measure of higher level reasoning we use here. The memory test did not affect the education gradient.
David M. Cutler, Department of Economics, Harvard University, 1875 Cambridge Street, Cambridge, MA 02138, and NBER, Email: ude.dravrah@reltucd..
Adriana Lleras-Muney, Department of Economics, UCLA, 9373 Bunche Hall, Los Angeles, CA 90025, and NBER, Email: ude.alcu.noce@sarella..