Article 3 2/21/03 4:20 PM Page 113

Journal of Economic Literature Vol. XLI March 2003) pp. 113–158

Health, Inequality, and Economic Development

ANGUS DEATON1

1. Introduction offer their citizens more public goods, more social support, and more social capital; and UPPOSE THAT INCOME causes good health. they satisfy humans’ evolved preference for People live longer and are healthier in S fairness. That equal societies are healthier is rich countries than in poor countries, people an argument particularly associated with live longer and are healthier than their Richard Wilkinson (1992, 1996, 2000), as grandparents and great-grandparents who well as the collection of papers edited by lived in poorer times and, within a country at Ichiro Kawachi, Bruce Kennedy, and a moment in time, rich people live longer Wilkinson (1999). One study (John Lynch et and are healthier than poor people. Suppose al. 1998) goes so far as to claim that, in the too that this relationship is concave, so that United States in 1990, the loss of life from in- income has a larger effect on health and come inequality “is comparable to the com- longevity among the poor than among the bined loss of life from lung cancer, diabetes, rich. Then income redistribution from rich to motor vehicle crashes, HIV infection, sui- poor, within countries, or between countries, cide, and homicide in 1995.” If even a frac- will improve population health (Samuel tion of this effect were real or if, more Preston 1975). Yet there may be more to it broadly, income distribution affects popula- than that. Income inequality, or other related tion health even indirectly, economic and fis- social inequalities with which it is correlated, cal policy has effects on well-being that are may be directly hazardous to individual typically ignored by economists or policy health. According to a recent body of litera- makers. And if economists are skeptical of ture, equal societies have more social cohe- such mechanisms, many policy makers are sion, more solidarity, and less stress; they not; shortly after being elected, British Prime Minister Tony Blair stated in Parliament that “there is no doubt that the published statis- 1 Research Program in Development Studies and Center for Health and Wellbeing, Princeton Univer- tics show a link between income, inequality, sity. Based on a paper prepared for Working Group 1 of and poor health.”2 the WHO Commission on Macroeconomics and This paper explores the theoretical basis Health. I am grateful to Anne Case for many helpful discussions during the preparation of this paper and to and empirical evidence for a connection be- George Alleyne, David Cutler, Sandy Jencks, Michael tween inequality and health, among poor as Marmot, Jim Smith, Duncan Thomas, Richard well as rich countries. The proposition that Wilkinson, and three referees for comments on earlier versions. I am grateful for financial support from the John D. and Catherine T. MacArthur Foundation and from the National Institute of Aging through grants to 2 Quoted in Michaela Benzeval, Ken Judge, and Sue Princeton and to the NBER. Shouls (2001). 113 Article 3 2/21/03 4:20 PM Page 114

114 Journal of Economic Literature, Vol. XLI (March 2003)

income inequality is an individual health risk preference. For much of this paper, I shall was first proposed for wealthy countries that take it as given that a causal link from in- have passed through the epidemiological come to health is at least part of the story, transition, where chronic diseases have re- and follow the consequences. Even so, I in- placed infectious diseases as the main cause clude in section 2 a brief discussion of the of mortality. But as we shall see, many of the topic, recognizing that an adequate discus- arguments that income inequality is a health sion of the relationship between income and risk are as plausible for poor as for rich health would require a survey much longer countries and, in some cases, more so. than the present one. I devote a substantial portion of the paper I discuss a number of other theoretical to theory, albeit interspersed with evidence. links between inequality and health, includ- With a few exceptions, the literature does ing possible effects of income on invest- not specify the precise mechanisms through ments in health and education, the two-way which income inequality is supposed to af- link between nutrition and earnings at low fect health. In consequence, there is little levels of income, and the arguable negative guidance on exactly what evidence we effects of inequality on the ability of the po- should be examining, or whether the propo- litical process to deliver public goods. I also sitions are refutable at all. Section 2 lays out consider arguments that our evolutionary some of the possible stories, starting with history predisposes us toward fairness, and the simple case in which health is affected sickens us when we live in unequal environ- by income, and there is no direct effect of ments. Such an account can be made consis- inequality. This is sometimes referred to us tent with a story in which relative depriva- the “absolute income hypothesis,” to empha- tion is a cause of ill health and according to size that it is income that matters for health, which, within groups, inequality has no ef- not income relative to other peoples’ in- fect on health conditional on income, while comes, nor income inequality. A name that across groups, both average income and in- would be at least as good is the “poverty” hy- come inequality determine health. Finally, I pothesis, that ill health is a consequence of consider the important case in which in- low income, in the sense that more income come inequality is in part a consequence of improves health by more among those with ill health, so that policies that reduce the low incomes than among those with high in- likelihood of sickness, shorten its duration, comes. It is important to start by establish- or ameliorate its effects on earnings can also ing the full range of implications for this narrow income inequalities. Some income simple case, and in particular what role is inequality is a consequence of the fact that played by income inequality. I also discuss earnings cannot be completely insured what happens when we make health de- against ill health, so that better health insur- pend, not on absolute income, but on rela- ance is likely to help reduce inequalities in tive income, and what this version of the rel- income. Better insurance will typically re- ative income hypothesis implies for the duce inequalities, both in health and in relationship between health and inequality. income. Much of the health-economics literature Section 3 turns to the evidence, most of does not accept the existence of any causal which comes from developed economies. I effect running from income to health, ex- review cross-country studies on adult mor- cept possibly through the purchase of health tality for rich countries and on child mortal- care, arguing that the correlation between ity for both rich and poor countries. A major them is driven in part by a causality running question is whether the international data on from health to income, and in part by third income distribution is of sufficiently high factors, such as education, or rates of time quality to support the inferences that are Article 3 2/21/03 4:20 PM Page 115

Deaton: Health, Inequality, and Ecomomic Development 115

being made. Better data, or, at least, more tions to this work, and in turn are likely to consistent data are available within coun- benefit from other social scientists’ some- tries, and some of the most interesting evi- times well-argued skepticism of economists’ dence on inequality and health comes from methodologies and prejudices. studies looking across areas and over time More broadly, ideas about income, in- within developed countries, such as Britain, come inequality, and health are important Canada, and the United States, and in a few for welfare economics. Health is a compo- poor countries. I also review the evidence nent of well-being, so that if health were af- from studies that link individual mortality fected by income inequality, tax and transfer and morbidity to the ambient level of in- policies that affect the distribution of in- come inequality, as well as those that study come would have effects that work, not only the links over time between income inequal- through the usual mechanisms—for exam- ity and patterns of mortality. ple, through an equity-preferring social- My conclusion is that there is no direct welfare function defined over levels of in- link to ill health from income inequality per come or consumption—but also through se; all else equal, individuals are no more individual levels of health. If income has a likely to be sick or to die if they live in places nonlinear effect on health, and even if there or in periods where income inequality is is no direct effect of income inequality on higher. The raw correlations that exist in health, redistribution of income toward the (some of the) data are most likely the result poor will improve their average health by of factors other than income inequality, more than the loss of health among the rich. some of which are intimately linked to People who are income-poor are also health- broader notions of inequality or unfairness. poor, so that seeing well-being as dependent The fact that income inequality itself is not a on both income and health reveals wider health risk does not mean that inequalities disparities between rich and poor than more generally are not important, let alone are recognized by standard, income-based that the social environment in which people approaches. live is irrelevant for their health. Indeed, I shall argue precisely the opposite. But we 2. Theoretical Accounts of Income must narrow and focus our search if we are Inequality and Health to make the leap from correlation to policy. The literature that is reviewed in this pa- 2.1 Individual, Group, and National Health per comes from a number of fields other than economics, particularly epidemiology, 2.1.1 Health, Income, and Poverty public health, sociology, psychology, and his- tory. Different fields have different styles of Figure 1 shows a recent version of the presenting theory and evidence, as well as Preston (1975) curve, the international rela- different standards for what counts as credi- tionship between life expectancy and na- ble evidence. Nevertheless, the ideas are of- tional income in current purchasing-power ten important and should not be dismissed parity dollars. Among the poorest countries, by economists, if only because they are increases in average income are strongly as- widely accepted by many policy makers and sociated with increases in life expectancy, by scholars in other disciplines, and econo- but as income per head rises, the relation- mists need to confront, not ignore, them. In ship flattens out, and is weaker or even ab- this spirit, it is worth presenting material sent among the richest countries. As Preston that is not fully worked out theoretically nor originally noted, if this relationship were convincingly demonstrated empirically. also to characterize the relationship between Economists can make important contribu- mortality and income within countries, Article 3 2/21/03 4:20 PM Page 116

116 Journal of Economic Literature, Vol. XLI (March 2003)

France Japan Spain Italy 80 Mexico

China UK Germany USA Korea 70 Brazil India Argentina Russia Indonesia 60 Pakistan Bangladesh

Gabon

life expectancy, 2000 life expectancy, 50 Equatorial Guinea Nigeria South Africa Namibia

40 Botswana

0 10,000 20,000 30,000 40,000 GDP per capita, 2000, current PPP $

Figure 1. The Preston Curve: Life Expectancy versus GDP Per Capita Source: World Development Indicators CD-ROM, World Bank (2002) Note: Circles are proportional to population and some of the largest (or most interesting) countries are labeled. The solid line is a plot of a population-weighted nonparametric regression. Luxembourg, with per capita GDP of $50,061 and life expectancy of 77.04 years, is excluded.

there would be a negative relationship the public release data) who were inter- across countries between income inequality viewed in a current population survey or in and life expectancy. Redistribution of in- a census-related sample around 1980, and come from rich to poor, whether within or whose deaths over a follow-up period of between countries, will increase the health 3,288 days were ascertained by matching to of the poor more than it hurts the health of the National Death Index (see Eugene the rich, and thus improve average national Rogot et al. 1992 for details). The survey or world health. Countries with particularly collects each person’s family income high income inequality, such as the United (within one of seven ranges) so that it is States, will have lower life expectancy than possible to link the probability of death their average income would warrant. during the follow-up to family income and As Preston surmised, there is indeed a other variables, the most important of relationship between income and mortality which are sex and age. For adults between within rich countries, even those on the flat the ages of 18 and 85, the log odds of mor- part of the international curve. Figure 2 tality are approximately linear in age, so a was calculated from the National convenient way to summarize the data is to Longitudinal Mortality Survey in the estimate a logit model for the probability of United States. This is a national follow-up death in which age is entered linearly, to- study of about 1.3 million people (with gether with a series of dummy variables for about three-quarters of a million people in the income categories. Article 3 2/21/03 4:20 PM Page 117

Deaton: Health, Inequality, and Ecomomic Development 117

0.15

white males 0.10

0.05

white females probability of death at age 50

0

0 20,000 40,000 60,000

family income in 1980, 1980 $

Figure 2. Age Adjusted Probability of Death and Family Income, National Longitudinal Mortality Study Source: Author's calculations from public use sample of the NLMS. Note: Probability of death is the probability that a 50-year-old man or woman at the time of interview (around 1980) died during the follow-up period of 3,288 days. Restricted to people aged 15 to 85 at the time of interview.

As it turns out, the log odds of mortality are ably more relevant. For both men and close to being linear in the logarithm of fam- women, the effect of income on mortality is ily income, so that the probability of death is greater among the poor than among the as illustrated in the figure, which shows, for rich; that the curve for women appears less men and women separately, the probabilities curved than the curve for men is an insignif- of a fifty-year old dying during the follow-up icant consequence of the choice of scale. If period. The protective effects of income are these curves come from a causal link be- substantial; Rogot et al. calculate that people tween income and health, income redistri- whose family income was more than $50,000 bution will improve population health. in 1980 have a life expectancy that is about It is often useful to think of the redistribu- 25 percent longer than people whose family tive story, not in terms of inequality, but in income was less than $5,000. This difference terms of poverty. If a country with a high av- is overestimated by the assumption that peo- erage income has a great deal of income in- ple stay in the same income class throughout equality, then there is a relatively large num- their lives, but underestimated by the poor ber of people with low income whose health measurement of income in much of the is poor. Although figure 2 shows no poverty NLMS, as well as by the fact that current in- line below which income matters and above come is likely a poor proxy for the longer- which it does not, it is at the bottom of the term measures of income that are presum- income distribution that the relationship Article 3 2/21/03 4:20 PM Page 118

118 Journal of Economic Literature, Vol. XLI (March 2003)

between income and health really matters. though large enough by postwar standards, And if a rich country has a lot of poor peo- are probably not large relative to earlier ple, it will have low average health relative to compressions, particularly those associated its per capita income. To illustrate, with the world wars (see Peter Lindert 2000, Wilkinson (1989) looks at mortality differ- and Claudia Goldin and Lawrence Katz ences by social class in Britain from 1921 to 2001). They are also characterized more by 1981, and argues that mortality fell most increases in the incomes of the rich than by rapidly at times when income differentials decreases in the incomes of the poor. were narrowed, particularly at times when The fact that a concave relationship be- incomes of the poor rose more rapidly than tween health and income implies that aver- those of the rich, such as during World War age health is negatively associated with in- II. Similarly, Amartya Sen (1999, figures 2.2 come inequality has come to be known in and 2.3) shows how life expectancy in the literature as a “statistical artefact” (Hugh England and Wales from 1901 to 1960 grew Gravelle 1998). The usage is designed to dis- most rapidly in the decades 1911–21 (by 6.5 tinguish it from mechanisms in which in- years) and 1940–51 (by 6.8 years), and more come inequality has a direct effect on indi- slowly at other times, 4 years in 1901–11, 2.4 vidual health, but it is unfortunate in years in 1921–31, 1.4 years in 1931–41, and suggesting that there is no real link between 2.8 years in 1951–60. Sen shows that the income inequality and health, and that redis- decadal rate of growth of GDP per capita is tributive policy cannot improve average strongly negatively correlated with decadal population health. This is far from the case; increases in life expectancy and, like if income causes health, and if there are di- Wilkinson, focuses on the degree of sharing minishing returns, redistribution from rich during the two wars, as well as on the direct to poor will improve average population nutritional and health interventions that health. took place during and immediately after the 2.1.2 Does Income Cause Health? second war. Both wars brought well-paying employment opportunities to many people The nonlinearity argument for income re- in Britain for the first time, including many distribution, in common with almost all of women. Richard Hammond (1950) dis- the literature in epidemiology and public cusses how wartime food policy in the 1940s health, takes it for granted that increases in not only protected essential food supplies, income cause improvements in health and but brought fresh milk and vitamins to work- reductions in mortality. Yet Preston’s original ing people to the extent that, unlike most of paper emphasized that income is only part the citizens of Europe, their nutritional sta- of the story, and he shows that the curves tus improved during the war. Reductions in have shifted upward over time, a movement income inequality during the wars, if indeed that he attributes to improved public health they took place, marked an improvement in measures, particularly in middle-income the conditions of the working people, among countries. And the 2000 Preston curve in whom better incomes and better nutrition figure 1 has developed a non-monotonicity would have had the largest effects on around $5,000, driven by the collapse of life mortality. expectancy in Russia, and by AIDS in Africa, In contrast to these arguments, I shall ar- particularly in South Africa; neither of these gue in section 3.5 that it is not possible to developments is income driven. Beyond link recent increases in income inequality in this, much of the economics literature has the United States and Britain to mortality been skeptical about any causal link from in- changes. It should be noted, however, that come to health, and instead tends to empha- recent increases in income inequality, size causality in the opposite direction, from Article 3 2/21/03 4:20 PM Page 119

Deaton: Health, Inequality, and Ecomomic Development 119

good health to higher earnings, and from against the argument that differences in poor health to the inability to work and, in mortality across population groups are pri- some countries, to running down assets to marily driven by differences in access to or pay for care. Although these arguments are quality of health care. Yet McKeown’s argu- obviously correct in some cases, and no one ments, even if true historically, need not be seriously challenges that ill health is often a true now, and the case against health care is reason for early retirement from the work- far from well-established. Indeed, many force, they have been strongly resisted by studies today take it for granted that differ- many noneconomists, particularly in Britain; ences in health care are the primary and ob- see Sally MacIntyre (1997) for an insightful vious reason for disparities in outcomes, for discussion of the political background of example “unequal treatment” of blacks and these debates. Economists have also made whites in the United States (Brian Smedley, the argument that the correlation reflects Adrienne Stith, and Alan Nelson 2002). the operation of some third factor, particu- For the purpose of the current discussion, larly education (Michael Grossman 1972, I need only that there be some effect run- 1975, 2000), or preference factors such as ning from income to health, a position that discount rates. Those who are impatient or has increasingly gained currency among have poor self-control are likely both to fail economists in recent years; see for example to make investments to protect their health the discussions in Jonathan Feinstein and to fail to acquire the education and skills (1993), Preston and Paul Taubman (1994), that generate higher earnings (Victor Fuchs and more recently, James Smith (1999). In 1993). Such arguments also focus on the role Deaton (2002) I have tried to summarize my of health-related behaviors, such as smok- own position. An earlier, usefully skeptical ing, binge-drinking, or lack of exercise, review is by Alan Garber (1989). A full dis- which are correlated both with low income cussion of the theory and evidence relating and poor health. income and health would occupy another re- In private or partially private systems, view, at least the length of this one. Even so, money can also pay for more expensive it is worth emphasizing a few key points. health care, but this mechanism is often dis- The literature outside of economics has counted, sometimes because of a general documented many hundreds of cases of a skepticism about the degree to which death positive correlation between health and can be postponed by medical care or, more some measure of socioeconomic status; the specifically, because of doubts about the rel- “modern” literature goes back to the eigh- ative effectiveness of expensive over basic teenth century, with precursors in ancient health care. It is also true that the relation- Rome and China; see Nancy Krieger (2001). ship between income and mortality, the One of the most influential of the current “gradient,” exists in countries with and with- studies is of civil servants in Britain, the out public provision of health care; see Whitehall study, which has documented Anton Kunst, Feikje Groenhof, and Johan substantial morbidity and mortality differ- Mackenbach (1998) for a review of ences by administrative rank; see Michael European countries. Thomas McKeown Marmot, Martin Shipley, and Geoffrey Rose (1976, 1979) famously demonstrated that, (1984), Marmot et al. (1991), and the more for disease after disease, the introduction of general discussions in Marmot (1994, 2002). effective prophylaxis was not the primary This literature tends to see income as a factor driving historical declines in mortality, “marker” for an underlying concept of so- and the persuasiveness of his arguments has cioeconomic status, which is treated as the left a presumption, at least among medical underlying cause of health discrepancies. sociologists and historical demographers, Because income is only one such marker, it Article 3 2/21/03 4:20 PM Page 120

120 Journal of Economic Literature, Vol. XLI (March 2003)

is often treated interchangeably with wealth from eliminated, by controlling for behav- (or parts of wealth, such as vehicle owner- ioral risk factors (in the Whitehall studies, ship or housing), or with consumption, or which have full medical and behavioral earnings, or even with rank, power, social or data, only about a quarter of the association occupational class, education, or even race between rank and mortality is eliminated by and geography. Such a framework is not very controlling for risk factors), or for employ- useful for thinking about social policy as ment status. health policy—for example, in a simple tax The most difficult issue of all is sorting system, a higher marginal rate of income tax out the dual causality between income and will change disposable income, but will often health. In poor countries, where malnutri- have no effect on rank in the income distri- tion remains a major issue, there is wide bution—nor does it take us very far in think- agreement that income has a direct causal ing about whether redistribution of income, effect; for excellent empirical evidence on as opposed to wealth, or even power, is likely the effects of South Africa’s generous social to improve public health. pension, see Anne Case (2001, 2002). Many That said, there is an increasing number have argued the same about rich countries, of studies that try to separate out different emphasizing the negative effects of poverty influences and confront economists’ argu- and of poor housing, and the positive effects ments about confounding by third factors. of not spending one’s life worrying about There are studies that look at the separate money. Even malnutrition may still have ef- effects of both education and income. In fects in rich countries today; Gabriele the NLMS (and figure 2) the protective ef- Doblhammer and James Vaupel (2001) and fects of income are not much affected by Doblhammer (2002) have established a re- controlling for education (Irma Elo and lationship between month of birth and re- Preston 1996; Deaton and Christina Paxson maining life-expectancy at age fifty (Austria, 2001a; and E. Backlund et al. 1999), and Denmark, Australia) and mean age at death similar findings have come from other data (those who died between 1989 and 1997 in sets (Evelyn Kitigawa and Philip Hauser the United States). Those born in October 1973; Paula Lantz et al. 1998; Richard and November do better than those born in Rogers, Robert Hummer, and Charles Nam March and April, although the seasonal pat- 2000). Income and education are also sepa- tern is reversed in the southern hemisphere rately protective for self-reported health (except for those who died in the south but status (Jacob Nielsen 2002). Economists’ were born in the north.) These effects are skepticism about a direct role for income is plausibly attributed to the effects of in- supported by relationships between mortal- trauterine nutrition more than half a cen- ity, income, and education at the state or tury ago, particularly to the seasonal avail- metropolitan statistical area level, where av- ability of fresh fruit and vegetables, and are erage education drives out average income, indeed wearing off for more recently born or even turns it into a risk factor (see cohorts. Theoretical mechanisms for such Richard Austen, Irving Leveson, and effects are developed in the work of David Deborah Sarachek 1969; Joseph Newhouse Barker (1994, 1995). The fetal origins and Lindy Friedlander 1980), a finding that (“womb with a view”) hypothesis “states continues to hold in more recent data (see that fetal undernutrition in mid to late ges- Christopher Ruhm 2000; and Deaton and tation, which leads to disproportionate fetal Darren Lubotsky 2003). The conflict be- growth, programs later coronary heart dis- tween the individual and aggregate data re- ease” (Barker 1995, p.171). There is now mains unresolved. The relationship be- extensive empirical evidence supporting tween income and health is reduced, but far these views, including evidence on the very Article 3 2/21/03 4:20 PM Page 121

Deaton: Health, Inequality, and Ecomomic Development 121

long-lasting (or long-postponed) effects on effect on health. Before that, there will still health. be health-diminishing poverty even in rich The studies of birth cohorts, particularly countries, so that income inequality will still in Britain, have thrown up a wide range of matter as well as average income. In conse- dynamic effects between health and earn- quence, the absolute income or poverty hy- ings, often operating with long lags. Panel pothesis implies that, among the poorest data in Britain (British Household Panel countries, average income is what matters Survey/BHPS) and the United States for population health, and income inequal- (Health and Retirement Survey/HRS) are ity is relatively less important. Among rich currently being used to examine health and countries, average income is less important, income dynamics. For example, in BHPS, and income inequality relatively more im- Benzeval and Ken Judge (2001) find that portant. Eventually, neither will matter controlling for initial (self-reported) health much for population health but, under plau- status reduces but does not eliminate the ef- sible assumptions, the effect of income in- fect of current income or current health sta- equality relative to that of average income tus. A more sophisticated approach, closely continues to grow as countries become related to Granger causality testing, is em- richer. These implications of the absolute ployed by Peter Adams et al. (2003) using income hypothesis are important because it data from HRS; they find no causal link is often claimed that the differential effects from wealth to health. This analysis, in par- of average income and income inequality on ticular, and the ability of “post hoc, propter health change with economic development hoc” inference to capture causality in gen- as described, and that the observation helps eral, has been seriously challenged by establish the case for a direct effect of in- Fabrizia Mealli and Donald Rubin (2002). It equality on health in rich countries. While is clear that there are influences between in- the finding is certainly consistent with such come and health that run in both directions, an effect, it is also consistent with a simple and that, in some cases, the lags can be as model in which income has a larger effect long as a human lifetime. In such circum- on health among the poor. stances, establishing clear causal patterns is Suppose that individual i lives in country a matter of great difficulty. s, and that her health his is a quadratic func- tion of her family income y so that if, for 2.1.3 Poverty, Health and the Direct and is convenience, I write everything relative to Indirect Effects of Inequality worldwide means, signified by overbars, we The absolute income (poverty) hypothe- have sis tells us a good deal about how we can ex- ¯ his − h = α + β(yis − y¯) pect average income and income inequality 2 −γ(yis − y¯) − θvs. (1) to affect population health at different lev- els of development. At the most obvious, Income promotes health, but by less for the the effects of per capita income mirror rich, so that both β and γ are positive. those of individual income, and become less Income inequality, here represented by the important the richer is the country. variance vs of income around the country Eventually, we would expect income in- mean, ys , is allowed to have a direct effect on equality to lose its effect too, but it is not health; I shall be particularly concerned with enough that average income be high the case where θ is zero, so that the only ef- enough, we also need everyone’s income to fects of nonlinearity are the aggregation ef- be high enough. The bottom tail of the in- fects associated with the nonlinearity. come distribution has to be pulled up be- Equation (1) is assumed to hold true every- yond the point at which income has much where, for all individuals, whether they live Article 3 2/21/03 4:20 PM Page 122

122 Journal of Economic Literature, Vol. XLI (March 2003)

in poor or rich countries, and wherever they the relationship between income and mor- are in the epidemiological transition. tality at the individual level (as in figure 2), Averaging (1) over everyone within each and then show that the implied relationship country, we get the cross-country relation- between income inequality and mortality ship across the states accounts for only a fraction ¯ of the actual relationship. I shall return to hs − h = α + β(ys − y¯) 2 the state data in section 3.3 below, but the −γ(ys − y¯) − (γ + θ)vs (2) contradiction between these two sets of re- where variables subscripted by s are popula- sults most likely reflects a deeper difference tion means. Even if θ is zero, the variance in the relationship between income and appears in (2) through the aggregation. mortality when estimated from individual Equation (2) can be used to illustrate a and aggregate data. number of points. First, consider the effect The quadratic model is useful for illustra- of economic development in the form of tive purposes, but not very plausible, if only higher levels of ys on the derivatives of because it implies that increases in income health with respect to average income and will eventually reduce health. A more realis- income inequality. The derivative with re- tic model links individual health to a latent β − 2γ(y − y¯) γ spect to ys is s , which, since variable, itself a function of income, and as- is positive, decreases with average income. sumes that death takes place when this la- By contrast, the derivative with respect to vs tent variable falls below some critical value. is constant. As a result, as a country becomes Under suitable assumptions, such a formula- richer and average income rises, the effect tion yields explicit predictions for the proba- on population health of income inequality bility of death for individuals, as well as the becomes more important relative to the ef- fraction dying in communities, and the re- fect on population health of average income. sulting equations provide a better way of θ This is true whether or not , the direct ef- linking population health to population in- fect of inequality, is zero. come and income inequality. As before, I The second point illustrated by (2), due to write h for the health of individual i in pop- Douglas Miller (2000), is that the parame- is ulation s, and assume that health is linear in ters γ and θ are both identified in the equa- the logarithm of individual income, tion so that it is possible, at least in principle, to test for a direct effect of inequality using his = a + blnyln is + dωs + is (3) only aggregate data. Intuitively, this works where the random term is is assumed to be because the curvature in the individual rela- normally distributed with mean zero and tionship (1) is transmitted to the aggregate variance σ2, and ω is the within-country relationship (2), so that it is possible, using s variance of log income, so that, once again I only aggregate data, to eliminate the part of allow for a possible direct effect of inequal- the inequality effect that works through the ity on health. The individual dies when h aggregation. Miller applies this technique to is falls below the critical level c. The probabil- mortality differences across the U.S. states, ity that this happens, or the probability of and finds that the relationship between mor- death, is written p which is tality and income inequality is entirely ex- is plained by the curvature in the aggregate re- pis = p(death | yisωs)= lationship between average income and c − a − blnyln is − dωs mortality, leaving no room for a direct effect Φ (4) σ of inequality. This result, however, is quite contrary to that of Michael Wolfson et al. where Φ is the distribution function of the (1999), who use the NLMS data to estimate standard normal. Suppose that, within each Article 3 2/21/03 4:20 PM Page 123

Deaton: Health, Inequality, and Ecomomic Development 123

country s, the logarithm of income is nor- ers. Richard Easterlin (1975) long ago found µ ω mally distributed with mean s variance s. evidence that happiness is independent of We can then rewrite (3) as income in the long run, and health may fol- h = a + bµ + dω low the same pattern. That health depends is s s on income relative to average incomes of +b(lnyln is − µs)+is (5) one or more reference groups is what we might call a relative income hypothesis. This so that p , the fraction of people dying in the s could happen in a number of different ways. state, or the probability of death conditional One case is where relative income deter- on the state mean and variance of log in- mines access to material goods, for example come, can be written when the people who live in a town are the p = p(death | µ ,ω ) s s s market for local land for housing, with the richest getting the hilltop plots with fine c − a − bµs − dωs =Φ . views, and the poorest getting the plots 2 2 (6) σ 1+b ωs/σ downwind of the smokestacks. The local housing case is an example where it is not Equation (6) links population mortality rates money itself that is important, but rank, to the population mean of log income, µ , s here determined by money. More generally, and the variance of log income, ω . s rank at work is important in determining The main points of (2) carry through to the how much control people have over their more realistic (6). The parameter d is identi- working lives and, as demonstrated in the fied in (6) from the aggregate data; even Whitehall studies, the degree of control at though the log variance appears in both the work accounts for much of the relationship numerator and denominator, its coefficient between occupational status and health in the latter is the same as the coefficient of (Marmot et al. 1997). mean income in the numerator. Second, sup- The relative income theory is consistent pose that d is zero, and we differentiate, first with an effect of income inequality, although with respect to µ , and then with respect to s it does not imply it. The distinction is of ω , it is readily shown that (a) both deriva- s some importance, if only because a direct in- tives tend to zero as µ goes to infinity, and s fluence of inequality on health, as proposed (b) that the ratio of the inequality derivative by Wilkinson, is frequently labeled “the rela- to the mean log income derivative grows lin- tive income hypothesis.” The argument ap- early in the latter. As countries become suffi- pears to be that, if health is lower for those ciently rich, and in the absence of a direct ef- whose income is relatively low, then higher fect of income inequality, neither population inequality makes the poor even poorer in mean income nor income inequality have relative terms, and so worsens population any effect on population mortality rates. But health. In section 2.1.6 below, I shall de- the effect of income inequality relative to the velop a model in which just this happens, effect of income is larger among richer coun- but the effect is not an automatic conse- tries. Once again, this has nothing to do with quence of a relative income model. a direct effect of inequality on health. Suppose that, as before, individuals are la- beled i, but that we use s to index the rele- 2.1.4 Relative Income, Absolute Income, vant comparison or reference group. An in- and Inequality dividual’s health might then be above or If income causes health, it is possible that below the population mean depending on health is determined, not by absolute in- whether her income is above or below the come, but by income relative to some aspira- average income in the reference group. tion level, or relative to the incomes of oth- Hence, with β positive, we might write: Article 3 2/21/03 4:20 PM Page 124

124 Journal of Economic Literature, Vol. XLI (March 2003) ¯ his = h + β(yis − ys). (7) ates a monotone increasing nonlinear rela- tionship between income and health, and For the group as a whole, group average there will be associated inequality effects h,¯ health hs is just average health which is over aggregated groups. If rank is all that the same for all groups, so that neither group matters about income, higher income for income nor group income inequality has any everybody will have no effect on anyone’s effect on group health. This annihilation of health. This rank hypothesis is worth noting the effect of income on health as we move for two reasons. First, rank is a plausible de- from the individual to the group will always terminant of power, of social status, or of happen if the reference group is within the control over others, all of which have been unit over which we are aggregating. For ex- associated with health. Second, rank is not ample, if reference groups are geographically necessarily affected by the usual redistribu- local, average income will be unrelated to av- tive policies. For example, consider a redis- erage health at the region, province, or state tributive policy that raises the marginal tax level. Such an account offers one explanation rate and uses the proceeds to pay lump-sum for why income might be related to health benefits to everyone. Provided the elasticity within groups, but not between them. of labor supply is greater than –1, there will The relative income story is also consis- be no effect on rank; if A has more income tent with a role for income inequality at the than B before the change, she will have more group level, and the mechanism is the same after the change (Kevin Roberts 1981). as for the absolute income hypothesis, the nonlinearity of the relationship between in- 2.1.5 The Impossibility of Identifying come, now relative income, and health. Reference Groups Again, an obvious example is provided by An immediate problem with the imple- the quadratic version of equation (7), in mentation of any model of relative income is which individual health depends on income the identification of the relevant reference relative to group income and the square of group. In a few cases, such as the Whitehall income relative to group income: civil servants, the reference group (or at ¯ 2 least what is likely people’s most important his = h + β(yis − ys) − γ(yis − ys) (8) reference group) is a ready-made part of the When we average health over individuals design. More usually, reference groups are within each group, the first term vanishes, as not clearly defined and people will often before, while the second term becomes the have multiple such groups, comparing variance of income, themselves to their neighbors, to their co- workers, to those they meet in social and re- ¯ − hs = h γvs (9) ligious organizations, and to those they see This provides us with one possible account on television or read about in newspapers. of the oft-cited phenomenon that was previ- One way of dealing with this is to recognize ously discussed, that within states or coun- that reference group incomes cannot be ob- tries, individual health is related to individ- served, and to work out the effects of the ual income, while between them, average omission on the relationship between the health is only weakly dependent on average two things that we can observe, health and income but is negatively related to income income. As shown in Deaton (2001a), this inequality. procedure brings inequality back into the Yet another possibility is that health is de- story even when it has no direct role. termined by rank in an income distribution. Figure 3 illustrates the simple case where In the individual data, this hypothesis gener- there are two groups, labeled “economists” Article 3 2/21/03 4:20 PM Page 125

Deaton: Health, Inequality, and Ecomomic Development 125 health

Group 1 “Ecomonists”

Group 2 “Doctors”

income

Figure 3. Health and Income in Two Reference Groups: the Effects of Within- and Between-Group Inequality on the Gradient.

(on the left) and “doctors” on the right. flatter, broken line. By mixing the two Income is measured on the horizontal axis, groups, omitting the relevant information on health on the vertical axis. The two ellipses group, the relationship is flattened out or show where economists and doctors are lo- attenuated. cated in terms of their health and incomes; Inequality comes into this story because members of each group are scattered within the degree of flattening depends on the ratio the two elliptical areas. Doctors have higher of within-group income inequality to be- incomes than economists, and within each tween-group income inequality. If doctors group, individual health depends on individ- and economists are moved further apart, by ual income relative to other members of the moving the two ellipses horizontally away group. The two parallel steep lines show the from each other, the broken line will be- relationships between income and health for come more attenuated. If within-group in- each of the two groups. Although doctors equality is increased, holding the between- have higher incomes than economists, their group difference fixed, so that the ellipses individual health is no better on average be- are stretched out along their individual in- cause their absolute income does not matter, come to health lines, the broken line will be- only their income relative to other members come steeper, increasing the gradient be- of their group. Suppose that an epidemiolo- tween income and health. The steepness of gist analyzes the data on economists’ and the gradient depends positively on the ratio doctors’ health, but without knowing which of within- to between-group inequality. For is which. When the data are pooled, the rela- example, in Whitehall, if there is only one tionship between health and income is the reference group, that of British civil servants, Article 3 2/21/03 4:20 PM Page 126

126 Journal of Economic Literature, Vol. XLI (March 2003)

the gradient of health with respect to in- equalities were unknown for most of our his- come (or rank) is likely to be steeper than in tory, modern inequalities are likely to be a a study containing a mixture of indistin- hazard to our health. This argument is force- guishable reference groups. More generally, fully and eloquently put by Wilkinson if health depends on relative, not absolute, (2000). income, and there is an increase in income That foraging groups were egalitarian ap- inequality that increases inequality within pears to be widely agreed. Such arrange- groups more than it increases it between ments could perhaps have come from lack of groups, the slope of the gradient will in- a technology for storing food. When a kill crease. has been made, and the meat is too much to In Deaton (2001a), I show that if health be consumed at once, sharing and subse- depends on relative income as in equation quent reciprocity are the only mechanisms (7), so that that can turn meat today into meat tomor- ¯ row. Humans, unlike other predatory his = h + β(yis − ys)+is (10) species, lack a long-term ability to store glu- ε cose, and so cannot summon the energy to where is is a random term that ensures that there is a scatter around the line, then the hunt after long periods without food, which expectation of health conditional on individ- may be why hunter-gatherer bands evolved ual income, but unconditional on reference in the first place (Arthur Robson 2002). group income, takes the form: Members of groups that enforced strict 2 sharing would therefore have a survival ad- ¯ βσw E(his | yis)=h + ys (11) vantage over members of those who did not, σ2 + σ2 w b so that a preference for sharing, fairness, 2 2 where σw and σb are the within- and and reciprocation may be evolved attributes between-group variances of income. The (see also Samuel Bowles and Herbert Gintis formula is the same as the attenuation bias 1998). With the invention of settled agricul- formula for measurement error in regression ture, with the associated ability to fill and analysis. hold food in granaries, as well as to build herds of animals, egalitarian and reciprocal 2.1.6 Evolution, Equality, Deprivation, sharing was less efficient, and could give way and Fairness to hierarchies within which rich and power- It might easily be supposed that hierar- ful individuals could dominate others. chic, unequal societies are an inevitable part Although such systems and their industrial of the human condition. Yet for the vast ma- successors are vastly more productive than is jority of our evolutionary history, humans foraging, the benefits come at the price of a lived in hunter-gatherer groups that were nagging and health-compromising outrage not only not hierarchic, but aggressively over the loss of equality. And while humans egalitarian; see the survey of the literature will perhaps evolve to suit this new environ- by David Erdal and Andrew Whiten (1996). ment, we have only given up foraging for a As has long been argued, perhaps first by very short time, only 10,000 to 20,000 years McKeown (1976, 1979), human health is of our one- to two-million-year history. maximized when we live under the condi- Adaptation to the new environment has its tions under which we evolved, pursuing reg- benefits, in terms of production, longevity, ular exercise (walking ten to fifteen miles a health, and population size, but it has a lin- day, as foragers and hunters did), and eating gering cost that prevents our health from low-fat, low-salt, low-meat, low-sugar, high- reaching its full potential. fibre, largely vegetarian diets. By the same Wilkinson and others have begun to token, given that hierarchies and social in- weave together a plausible story of the Article 3 2/21/03 4:20 PM Page 127

Deaton: Health, Inequality, and Ecomomic Development 127

processes that support such an account. ties has been proposed by Ernst Fehr and Psychosocial stress is the main pathway Klaus Schmidt (FS) (1999), who note that through which inequality affects health. people with such preferences, unlike Wilkinson draws a contrast between soci- purely selfish actors, are willing to give up eties in which relationships “are structured their own resources to affect other people’s by low-stress affiliative strategies which fos- outcomes. They show that the assumption ter social solidarity” on the one hand, and that a fraction of the population has such societies characterized by “much more preferences can account for a wide range of stressful strategies of dominance, conflict experimental and actual social behaviors. and submission. Which social strategy pre- The FS utility function is written dominates is mainly determined by how ∞ equal or unequal a society is” (Wilkinson u(y)=y − β1 (x − y)dF (x) y 2000, p. 4). Equality is seen as a precondi- y tion for the existence of stress-reducing + β2 (y − x)dF (x) (12) networks of friendships, while inequality 0 β > and relative deprivation are seen as com- where 1 0 captures the extent to which it promising individual dignity, and promot- is harmful to have people with incomes β ing shame and violence. At the same time above you, while 2, which can be positive, the biological mechanisms through which negative, or zero, but in all cases greater chronic stress compromises health are be- than –1, captures the benefits or costs that ginning to be understood; excellent surveys arise from the existence of people with in- β can be found in Robert Sapolsky (1998), comes less than yours. FS assume that 2 is Eric Brunner and Marmot (1999), and negative, so that people with (12) are averse Wilkinson (2000, ch. 2.) to people being either below or above them, The story outlined above is persuasive in caring about “fairness,” or more precisely, many respects. That the social environment “self-centered inequity aversion,” but they in which we live helps determine our health note that almost all of their results are robust is surely right, and the effects of psychoso- to allowing β to be positive. β is less than β 2 2 cial stress are now well-documented in 1 which guarantees that people prefer that both human (particularly Whitehall civil their own incomes be higher. servant) and animal populations, where If we note that the integral of y – x over controlled experiments are possible. Yet it the range of x is y – µ, for mean income µ, is less clear why income inequality is the we can rewrite the FS utility function as only villain, or even one of the villains, in u(y)=y + β (y − µ) − (β − β )× the piece. One attempt to explain is given 2 1 2 ∞ in Deaton (2001b), which is extended in (x − y)dF (x)=y + β2(y − µ) the following account. I treat income rela- y tive to other members of a reference group −(β1 − β2)µR(y) (13) as the key variable, and hypothesize that stress on each individual depends on the where Ry is the measure of relative depriva- differences between that individual’s in- tion, proposed by Shlomo Yitzhaki (1979). come and those of others in the group. The Relative deprivation is a term frequently existence of incomes above you poses a used by Wilkinson in his discussions of social threat, while incomes below you can be ei- stress, and whose implications for health, ther costly, if you value fairness, or benefi- also using the Yitzhaki measure, have been cial, if you derive pleasure from the status explored by Christine Eibner (2001). associated with having more income than Although the meaning here is more special- others. A utility function with these proper- ized (and more precise) than that intended Article 3 2/21/03 4:20 PM Page 128

128 Journal of Economic Literature, Vol. XLI (March 2003)

by Wilkinson, the formalization appears to 2.2 Other Mechanisms, Other Inequalities be well-chosen, in that it generates many of 2.2.1 Credit Constraints, Health, and the effects that he proposes. The derivative Income Inequality of health (utility) with respect to income is 1 + β + (β − β )(1 − F(y)), which is positive, Income inequality may affect investment, 2 1 2 − β − β while the second derivative is ( 1 2)f(y) including investments in health and educa- which is always negative. So if we think of tion. If poor people are more likely to be health as proportional to utility, FS utility credit constrained than rich people, if only implies a concave relationship between in- because they have less collateral, redistribu- come and health within groups. Because tion from those without credit constraints to health is entirely determined by income, in- those with them may increase investment, a equality has no effect on health conditional mechanism that has been explored in the re- on income. cent literature on economic growth; see for If we look at group health, and compute example Roland Benabou (1996); Phillipe the average by integrating (13) over y, we Aghion, Eve Caroli, and Cecilia García- obtain Peñalosa (1999); Pranab Bardhan, Bowles, ∞ u(y)= u(y)dF (y) and Gintis (2000). Relevant health invest- 0 ments may include taking a sick child to a hospital, paying for vaccinations, or even = µ[1 − (β1 − β2)g] (14) providing adequate nutrition; even if these where µ is mean income, and g is the gini have high rates of return over the long run, coefficient of income inequality. Across poor parents may lack the money to pay for groups, utility—or by supposition health— them. Education is also relevant, in its own depends positively on mean income but neg- right, as an investment in human capital and atively on income inequality. (Note also that higher earnings, but also because of its links utility evaluated at mean income µ has the with health. For example, Elo and Preston same form as (14) with the gini coefficient (1996) estimate that, around the world, a replaced by the relative mean deviation, year of extra education reduces mortality sometimes called the Pietra or Robin Hood rates by about 8 percent, half of which works index, the last frequently used in the epi- through the effects of additional earnings, demiological literature. Equation (14) con- and half directly. forms almost exactly to Wilkinson’s (1997) Figure 4 shows some relevant data on ed- statement that “income inequality summa- ucation for rural India. The 52nd round of rizes the health burden of individual relative the National Sample Survey collected data deprivation.” in 1995–96 on education enrolment and sta- In summary, this theory of mortality risk tus of respondents. The graph shows the has three important implications: (i) within percentage of boys and girls aged from groups, health is a concave increasing func- seven to twelve inclusive that are currently tion of income; (ii) conditional on an indi- enrolled in school (as reported by the vidual’s income within the group, inequality household in the survey) as a function of the does not matter for individual health; and logarithm of total household expenditure (iii) for groups, average health depends per head, a close proxy for income per head. positively on group income and negatively The graphs are calculated using nonpara- on group income inequality. The results of metric regression using 20,307 boys and testing this theory are postponed to section 17,321 girls aged seven to twelve. Boys are 3, where they can be presented in the con- more likely to be in school than girls, both text of other, related work on geographical are more likely to be in school when they mortality differences in the United States. live in better-off households, and the effects Article 3 2/21/03 4:20 PM Page 129

Deaton: Health, Inequality, and Ecomomic Development 129

100

boys 80

girls 60 percentage reporting enrolment

40

100 300 500 700

household per capita total expenditure, rupees per head

Figure 4. School Enrolment of Boys and Girls Aged 7 to 12 in Rural India, 1995–96. Source: Author's calculations (locally weighted regressions) based on National Sample Survey data.

of additional resources are larger for girls holds, so that this graph indicated no major than for boys. Although the slopes of the effect of redistributing income on raising curve vary somewhat with per capita expen- child vaccination rates. Nor is there any diture, they are flatter among better off difference between vaccination rates of boys households. In consequence, and with the and girls. usual caveats about omitted variables and These examples are only suggestive. The causality, redistribution from richer to curves in figures 4 and 5 make no attempt to poorer households will increase the total control for other factors—such as parental percentage of children in school, and will education, school quality, or health service increase girls’ enrolment more than boys’. provision—that are likely to be positively cor- Education, in turn, is likely to improve related with both household income and the health, not only of those receiving it, but of outcome, so that the effects of income are al- their children, particularly in the case of most certainly overstated. Nevertheless, the women. graphs illustrate that income redistribution Figure 5 shows some contrasting data on may (or may not) increase health and educa- child vaccinations from the same Indian sur- tional investments in children, and may even vey. As was the case for school enrolment, differentially favor girls. the fractions of children under five who have 2.2.2 Nutritional Wages, Health, and all three of the most important vaccinations Inequality of Land (BCG, DPT, and OPV) is lower for poor than for rich households. However, unlike The nutritional wage model provides an education, the slope is only slightly flatter account of how inequality affects both health among better-off than among poorer house- and earnings while explicitly recognizing that Article 3 2/21/03 4:20 PM Page 130

130 Journal of Economic Literature, Vol. XLI (March 2003)

70 girls

boys 60

boys

50 girls

40 percentage reporting full vaccinations 30

100 300 500 700 household per capita total expenditure, rupees per head

Figure 5. Vaccination Rates of Children Under 5 in Rural India, 1995–96. Source: Author's calculations (locally weighted regressions) based on National Sample Survey data.

health and earnings are simultaneously de- derbid them because their undernourish- termined. It pinpoints inequality in land ment and poor health lowers their marginal holdings, not inequality in income, as the product to make them unattractive hires determinant of poor health, and provides even at the lower wage. More formally, there the first of several examples of an important may be no combination of work and calorie theme of this survey, that inequality in di- consumption that is feasible for the worker mensions other than income may have an ef- and acceptable to the employer. Under such fect on health. The ideas go back to Harvey conditions, unemployment, and the chronic Leibenstein (1957), with the fundamental malnourishment and sickness associated work by James Mirrlees (1975) and Joseph with it, cannot be eliminated by decreasing Stiglitz (1976); see Christopher Bliss and the wage. Nicholas Stern (1978) for a survey. That nu- What is the role of inequality in this story? tritional wage models can account for per- The root of the problem is seen in the distri- sistent poverty and destitution in poor coun- bution of land. If the destitute had sufficient tries is eloquently argued in Partha land of their own, they would be able to Dasgupta and Debraj Ray (1986, 1987) and meet their basic nutritional needs by grow- Dasgupta (1993); an excellent textbook sum- ing food on small plots without having to mary is provided by Ray (1998, ch. 13). work for wages. Seen this way, the problem According to this story, unemployed and of unemployment and malnutrition is a destitute workers, even though they are will- problem of landlessness, and can be re- ing to work for less than the wage being cur- solved by redistributing land so as to give to rently paid to employed workers, cannot un- every family a plot sufficient to guarantee Article 3 2/21/03 4:20 PM Page 131

Deaton: Health, Inequality, and Ecomomic Development 131

their basic needs, and to enable them to par- 1983, 2,000 calories (in the form of standard ticipate gainfully in the labor market. coarse cereals) could be purchased for less Destitution is therefore seen as a problem than 5 percent of the day wage, a finding that can be dealt with by a more equitable that is consistent with the observation that distribution of land. poor agricultural workers in India typically The nutritional explanation of destitution eat their fill of cheap calories at the end of and chronic malnutrition in India by the work day; see also Anand Swamy (1997). Dasgupta and Ray is echoed by Robert With food so cheap, the nutritional wage Fogel’s (1994) account of the economic and trap seems too easy to escape. More broadly, health history of Europe. He argues that, the model does not always draw a suffi- even at the end of the eighteenth century in ciently clear distinction between nutrition, England and France, food production was which comes from the food that can be so low that, given its likely distribution over bought for money, and nutritional status, people, perhaps a fifth of the population was which depends on disease as well as on nu- capable of no more than a few hours of light tritional inputs. A plentiful supply of food work each day. These people were chroni- will not nourish someone whose drinking cally malnourished, were short in stature, water and food are contaminated, and and died young. Only the increase in agricul- chronic malnutrition typically needs to be tural productivity in the nineteenth century addressed through public health measures permitted an escape from this nutritional as well as by increasing the supply of food. trap, and meant the beginning of the transi- This criticism applies, not only to the nutri- tion to better health, lower mortality, and tional wage theory as an account of destitu- longer life. tion in poor countries, but also to McKeown This story of nutrition and wages has (1976) and Fogel’s (1994) arguments about much to commend it. It directly incorpo- the historical importance of nutrition; see rates the two-way causality between health particularly Preston (1996). and earnings and provides a general equilib- 2.2.3 Public Goods and Political Inequality rium explanation of unemployment and poor health that has an obvious relevance to While the nutritional wage story pinpoints poor countries now as well as to the histori- inequality in land holdings as a cause of poor cal record in now-rich countries. It also pin- health, another line of literature sees the po- points land reform as the appropriate redis- litical process as a route through which in- tributive policy prescription. Yet the come and other inequalities can affect predictions of the model have been chal- health. One story is that heterogeneity in lenged; see for example Hans Binswanger preferences makes it more difficult for peo- and Mark Rosenzweig (1984), Rosenzweig ple to agree on the provision of public (1988, pp. 720–28), and John Strauss and goods, such as health, water supply, waste Duncan Thomas (1998). Workers who are disposal, education and police. Such mecha- trapped by their low nutrition and inability nisms have long been recognized in the lit- to work would devote all their energies to erature on political economy, and a simple finding food, and would have no energy for but suggestive account has been provided by consuming anything other than food, for Alberto Alesina, Reza Baqir, and William saving, or even for procreation (Mark Easterly (1999). They point out that the av- Gersovitz 1983). And even in the poorest erage value of public good to members of a economies, food is typically too cheap rela- community diminishes with the heterogene- tive to the wage rate to make the trap plausi- ity of their preferences, a heterogeneity that ble. Shankar Subramanian and Deaton could be located in racial, income, or geo- (1996) calculate that in rural Maharashtra in graphical differences. Provision is therefore Article 3 2/21/03 4:20 PM Page 132

132 Journal of Economic Literature, Vol. XLI (March 2003)

likely to be lower where heterogeneity is the atmosphere and the environment. Yet greater. Although they note the potential ap- many cities, in Britain and in Europe, were plication to income inequality, Alesina, slow to address these problems, even when Baqir, and Easterly think of their model as the necessary policies were well understood. applying to racial divisions in the United Although the germ theory of disease was States, and to ethnic fractionalization more generally accepted only after 1870, earlier widely, a variable that often appears to have explanations, such as the “miasma” theory, negative consequences in cross-country also emphasized the importance of cleaning growth studies. In the context of the cities up the environment. And while money is al- and counties of the United States examined ways a factor limiting public construction, by Alesina et al., ethnic fractionalization is these were periods of relatively rapid eco- closely related to the fraction of the popula- nomic growth. Indeed, the coexistence of tion that is black, which is positively related rapid economic growth and mortality in- to total spending, but negatively related to crease (as well as a decrease in stature) dur- the shares in spending of “productive” pub- ing this period is regarded as something of a lic goods, such as health, roads, and educa- puzzle by economic historians—such as tion. For health, the total effect offsets the Michael Haines and Hallie Kintner (2000), share effect, so that the absolute amount of Roger Schofield and David Reher (1991), health spending is positively associated with and Fogel (1997)—who are typically so con- fractionalization. Income inequality is in- fident of the link between health and in- cluded in the models, although the results comes that they often use measures of the are not presented in the paper. However, I former—such as stature—as reliable indica- understand from one of the authors that the tors of the latter (Richard Steckel 1995). effects vary from model to model and that Szreter (1988) argues that the key to there are no robust negative effects of in- understanding the mortality transition in equality on either total spending or its distri- England lies in local politics. Although the bution. However, it might also be noted that industrializing cities were in fact well sup- Robert Putnam (1983), in his study of social plied with fresh water, it was used for com- capital in Italy, also sees equality as an im- mercial purposes, not supplied to homes. portant element in a well-functioning civic And the new entrepreneurial class, rich community. though it was, saw no point in spending each Simon Szreter (1988) has provided a per- others’ money for public sanitation which suasive account of politics and sanitation in had no obvious commercial benefit. It was Britain in the mid-nineteenth century; see only after political reform, and particularly also Easterlin (1996). The urbanization of the limited political emancipation of work- population associated with the industrial ing men, that new political coalitions could revolution led to a sharp reduction in public develop that made sanitation and public health, with mortality higher in cities than in health a priority. the countryside, and a decline in overall life This is a story of nineteenth-century expectancy. Urban populations often had no England, not of the world today, and it is access to clean water, and no facilities for about political, not income, inequality. Yet disposal of human and other wastes, which there is a marked parallel with the effects of were allowed to accumulate as a perpetual civil rights legislation in the American South hazard to health. Crowding aided the trans- in the late 1960s. Douglas Almond, Ken mission of infectious diseases, some of Chay, and Michael Greenstone (2001), using which can only be sustained in populations detailed data from Mississippi, document above a critical size. Pollution from smoke that, prior to 1965, hospitals were strictly and other factory discharges contaminated segregated by race and that, especially in Article 3 2/21/03 4:20 PM Page 133

Deaton: Health, Inequality, and Ecomomic Development 133

rural areas, there was limited access for construction provides employment opportu- black mothers and their infants. Federal ac- nities for women) though, perhaps surpris- tions, i.e., the 1963 court decision outlawing ingly, not in the limited amount of preschool “separate but equal” clauses in federal hos- education that is under panchayat control. pital construction, and the 1964 Civil Rights The work on villages in India, on nine- Act, made segregation illegal, but it was the teenth-century Britain, and on cities in the passage of Medicare in 1965, which prohib- United States all point to the possible im- ited Medicare funding to segregated hospi- portance for health of political arrange- tals, that effectively enforced desegregation. ments, and in particular to the consequences Almond, Chay, and Greenstone show that, of inequalities in political power. Once from 1965 to 1971, there was a large reduc- again, the story is not about income inequal- tion in black post-neo-natal infant mortality ity per se, though income inequality can rates, particularly in conditions amenable to hardly be unrelated to political inequality. treatment, such as diarrhea and pneumonia, But by taking a narrow focus on income in- and that this period saw the only sustained equality, we may be missing the inequalities narrowing of the gap in black-white infant that have the largest effect on health. mortality rates after 1950. 2.3 Income Inequality as a Consequence In the story of Britain in the nineteenth of Ill Health century, the central actors in the plot are the Reform Acts and their (limited) extension of Most of this section has started from the democracy which may or may not have (di- position that income causes health, develop- rectly) reduced income inequality, but cer- ing the consequences for the relationship tainly reduced political inequality. In the between income inequality and health. But story of the South, the civil rights legislation health also causes income, and policies that not only improved the health of African directly affect health are likely to have ef- Americans, it also markedly improved their fects on income inequality. incomes relative to those of whites (John When interactions between income and Donohue and James Heckman 1991; and health are important, the distribution of in- David Card and Alan Krueger 1993). come will depend on the level and distribu- Reducing political inequality—that some can tion of health. Any measure that reduces the vote and others cannot, or some can get ac- spread of health conditions across the popu- cess to schools, hospitals or jobs while others lation, or improves the health environment, cannot—is not only valuable in its own right, will narrow the distribution of income. it also helps reduce other inequalities, in Health shocks are important determinants health, incomes, and education, in accord of earnings and consumption in developing with the general thesis of Amartya Sen countries; see for example Paul Gertler and (1999). In a contemporary context, it is hard Jonathan Gruber (2002). In some countries not to see political action, or at least the lack health care expenses can be large relative to of it, as one of the reasons behind the low income or wealth (Adam Wagstaff 2002). In level of provision of schools and clinics in agricultural villages around the world, the Indian villages. Rhagobendra Chattopadhyay poorest people are often those who cannot and Esther Duflo’s (2001) work in India work as a result of some long-term disability shows how the mandated representation of or injury. Anything that helps people recover women as leaders of village councils (in a more rapidly from an illness will reduce the randomly selected third of all Indian villages) persistence of ill health, which reduces the has led to small but perceptible gains in pub- long-term variance of health across the pop- lic goods important to women and children, ulation. Better insurance arrangements, or particularly water, fuel, and roads (whose better and more widespread clinics are Article 3 2/21/03 4:20 PM Page 134

134 Journal of Economic Literature, Vol. XLI (March 2003)

obvious candidates to reduce such persist- The parameters in (16) can be thought of ence, and so will not only improve popula- in terms of the various policies that might af- tion health (if they work at all), but also im- fect them. The variance of the health shock, σ prove the distribution of income. Clean 11, captures the health environment, and water, whose lack affects the poor more than would be reduced by vector eradication the rich, will improve the incomes of the poor campaigns, cleaning up the environment, or relative to the rich, and reduce income in- more comprehensive vaccinations. The pa- β equality. Malaria eradication campaigns and rameter 1 which is non-negative, controls vaccination drives will have the same effect, the speed of the healing process once health not only improving population health, but has been compromised so that, for example, also narrowing the distribution of income. better health insurance and better health- β β This would be true even in a Lewis world in care would reduce 1. The parameter 2 which there is an unlimited supply of labor at which is positive, controls the effect of the subsistence wage. Better health may not health on income. Disability insurance, improve the wage rate, and cannot do so un- which limits the effects of health on income, der the Lewis assumption, but it can enable would decrease β , towards zero. If we take 2 β more people to work at that wage. income to be net of health care expenses, 2 Once again, a simple formal model illus- will also reflect the costs of health care, and trates these points. Suppose that health and better health insurance, like disability insur- income satisfy the following equations: ance, will decrease it towards zero. The pa- rameter γ controls the direct causal link h = α + β h − + γ y +  1 t 1 1 t 1 1 t 1t from income to health and, for the purposes of the argument, I shall assume it is zero so yt = α2 + β2ht + 2t (15) as to examine the effects of health on in- ε ε where 1t and 2t are stochastic shocks to come inequality in the absence of any link health, each with zero mean, variances σ from income to health. σ σ 11 and 22 and covariance 12. For the process We can use (15) to calculate the variances to be stationary, I assume that 0 ≤ β1 < 1. and covariance of income and health. In par- The conceptual experiment here of a group ticular 2 of identical people, who differ only in the β σ11 var(y)=σ + 2 22 − 2 (16) random shocks that they receive to their in- 1 β1 comes and health. These shocks are drawn from the same joint distribution, but inde- so that the variance of income is the variance of the income innovation plus a term that pendently across people. In such circum- depends on the health environment and the stances, the joint distribution of health and insurance system. A reduction in the vari- income across people is the same as the ance of health shocks—vector control, for long-run time-series distribution of health example—reduces income inequality. So do and income from (15). It would be straight- improvements in health care, in health in- forward to introduce heterogeneity by surance, and in disability insurance. Places making the parameters individual specific that have more of these policies in place will and, if the health and income shocks are have better health and a more equal distri- distributed independently of the individual bution of income. Even without any causal effects, the inequality from the dynamics link from income to health, the regression can be added to the inequality from the coefficient of health on income (the gradi- heterogeneity. While stationarity is not re- ent) is positive, and can be written alistic for finitely-lived people, it helps us cov(hy) β σ isolate the contribution of random shocks b = = 2 11 2 − 2 (17) to inequality. var(y) β2 σ11 + σ22(1 β1 ) Article 3 2/21/03 4:20 PM Page 135

Deaton: Health, Inequality, and Ecomomic Development 135

This gradient is larger the larger is the vari- sufficient to permit us to make unambiguous ance of the health shock relative to that of inequality rankings between any two distri- the income shock, and is smaller the more butions of income. Instead, the axioms in- effective is the healthcare and health insur- duce a partial ordering whereby we can ance system. sometimes rank one distribution as more un- Income differences between countries are equal while, in other cases, we can judge the also affected by their differences in popula- inequality of distributions only by choosing a tion health. In consequence, the speed at specific inequality measure, with different which new health technology is transmitted measures giving different results. In particu- from the industrialized to developing coun- lar, different inequality measures give a dif- tries affects their relative population healths ferent emphasis to different parts of the dis- as well as their relative incomes; see Jeffrey tribution. For example, the gini coefficient is Sachs (2001). Current debates about interna- more sensitive to inequality (or to measure- tional pricing of drugs and patent protection ment error) at the top of the income distri- are important in this context. Whether the bution, whereas measures that work with faster transmission of drugs and vaccines will the logarithms of income, such as the Theil improve the distribution of income within measures, or the variance of logarithms, are poor countries depends on the prevalence of quite sensitive to inequality at the bottom. disease within the income distribution. For Although neither the conceptual issues nor AIDS, prevalence in poor countries is (still) the choice of inequality indicator are the higher among the rich, while for malaria or most important issues for health, it should tuberculosis, prevalence is higher among the be noted that many of the indexes that are poor. In the past, health innovations have of- used in the public health literature do not ten widened health inequalities when first in- satisfy even the generally accepted axioms. troduced; examples are better sanitation a For example, the Robin Hood index (more century or more ago (Preston and Haines usually known as the relative mean devia- 1991); tobacco and lung cancer more re- tion; Kennedy, Kawachi, and Deborah cently; and child health improvements in Prothrow-Stith 1996a) is unaffected by contemporary Brazil (Cesar Victora et al. transfers between individuals on the same 2000). Faster transmission of best-practice side of the mean. If a transfer program were health care may therefore widen the income to transfer incomes from those just below distribution within the receiving countries, at the median to those near the bottom, the least in the first instance. Robin Hood index would not change, even though there would have been a real reduc- tion in inequality (and very likely a decrease 3. Empirical Evidence on Inequality in mortality risk too.) Perhaps most of the and Health public health work uses as its inequality measure the share of income accruing to the 3.1 Measuring Income Inequality bottom x (often 50) percent of the popula- The measurement of income inequality tion. Once again, transfers within the bot- involves conceptual and practical issues. tom x percent, or within the top 1 − x per- Comprehensive treatments of the theory of cent, will leave the measure unaffected, inequality measurement were developed in even though such transfers are capable of the 1970s by Anthony Atkinson (1970) and having a substantial effect on income in- Sen (1973), the latter updated in James equality more broadly. Foster and Sen (1999). Although there are a Some authors, including Lynch et al. number of axioms on the nature of inequal- (1998), Luiza Franzini, John Ribble, and ity that are broadly accepted, these are in- William Spears (2001), Paul Brodish, Mark Article 3 2/21/03 4:20 PM Page 136

136 Journal of Economic Literature, Vol. XLI (March 2003)

Massing, and Herman Tyroler (2000), and income be assigned to individuals, and that Diane McLaughlin and Shannon Stokes inequality be calculated over persons, not (2001) also use the 90:10 ratio defined, not households. These apparently technical de- as is usually the case, as the ratio of the 90th tails can sometimes have serious effects on to the 10th percentile of the income distri- the measurement of income inequality, and bution, but the ratio of the share of income their treatment in the public health litera- accruing to the bottom 90 percent to the ture has often been cavalier. There are also share accruing to the bottom 10 percent. obvious constraints when authors restrict This “inequality” measure has the unfortu- themselves to the inequality measures pre- nate property (for an inequality measure) calculated by the Census Bureau; these re- that it is reduced by transferring income strictions may sometimes be necessary in or- from the middle of the distribution to the der to avoid having to deal with other data top tail; not surprisingly, this measure is problems with the public-release micro sometimes strongly negatively correlated data. with more usual measures such as the gini; Conceptual problems are dwarfed by see for example Franzini, Ribble, and measurement problems. Income itself is Spears. Those authors who use it to demon- hard to measure, and the difficulties multi- strate that inequality affects mortality may ply when measuring income inequality. well have shown the opposite! Measurement error in income, even if it has There has also been a good deal of discus- little effect on the measurement of mean in- sion about the appropriate definition of come, will inflate the measured variance and household income, and in particular the measured income inequality. The measure- treatment of household size. The standard ment of income is sensitive to survey design, procedure in economics would be to “equiv- particularly to the choice of the reference alize” household income by dividing income periods for income (longer reference peri- by some measure of household size, either ods give lower measured inequality), and to household size itself, or the number of exactly how the income question is asked. equivalent adults, for example the number The degree of disaggregation of income cat- of adults plus half the number of children, or egories is important, as is whether incomes the square root of the total number of peo- are reported as a number, or in a set of pre- ple in the household. Such per-equivalent defined ranges, typically ending with an measures attempt to capture the resources open-ended top category. The choice of cut- available to each person in the household, off points for the ranges is important, partic- and recognize that, at the same level of in- ularly the top band which effectively limits come, members of a larger household are the highest income that can be reported. worse off than members of a smaller house- Some surveys permit people to report nega- hold. Indeed, public health statisticians, tive incomes (losses from business activities) studying pellagra, were among the first to and some do not. Some surveys collect data develop equivalent measures; see Edgar on income, and some on consumption; the Sydenstricker and Willford King (1921). latter is almost always less unequally distrib- (Even so, it should be noted that there is ev- uted than the formal. The response rate idence in Elo and Preston 1996 that, condi- from surveys varies over space and time, and tional on family income, larger family size richer households are typically less likely to may not increase mortality. Nor do equiva- agree to participate, in many cases because lized measures give any purchase on they live in communities where the enumer- whether income is allocated equally within ators cannot reach them; see Robert Groves the household) When income inequality is and Nick Couper (1998). In rich countries, calculated, it is also important that equivalized and in some not-so-rich countries, response Article 3 2/21/03 4:20 PM Page 137

Deaton: Health, Inequality, and Ecomomic Development 137

rates have been falling over time, perhaps in data are well-documented and have been response to the increasing competition from widely analyzed, so that their properties are market researchers. There are also differ- well understood. Some authors have taken ences across countries in the degree to income inequality data for industrialized which people are prepared to cooperate countries from other, non-LIS sources, such with government surveys. Response rates in as Malcolm Sawyer (1976); these are now the United States are typically much higher superceded by the LIS. than in Western Europe. Matters are a good deal more difficult for Because different countries use different income distribution data from the large survey instruments, and have different sur- numbers of developing countries that are vey protocols, useful cross-country compar- not covered by the LIS. For many years, isons of income inequality require detailed popular sources of income distribution data knowledge of the specific surveys. Similar is- were Shail Jain (1975) and Felix Paukert sues sometimes arise with comparisons over (1973), which are essentially compendia of time within one country, even where the sta- inequality estimates then available in the tistical service is of the highest quality. World Bank and International Labor Office, Specifically, the U.S. Census Bureau, in the respectively. In more recent years, research summer of 2000, decided that the large in- on international patterns of income inequal- crease in household income inequality be- ity has been transformed by the availability tween 1992 and 1993, which it previously on the World Bank web site of the inequality presented as real, was in some unknowable data assembled by Klaus Deininger and Lyn part due to changes in survey methodology, Squire (DS) (1996). These data, which have particularly changes in the highest level of seen widespread use, contain more than income permitted in the questionnaires, as 2,600 observations on gini coefficients (and well as the introduction of computer-aided many quintile shares) for more than 100 de- interview technology; see Arthur Jones and veloped and developing countries for dates Daniel Weinberg (2000). In consequence, between 1947 and 1994. To be included in the United States no longer has a consistent the DS data set, estimates have to come continuous time series of household income from an identifiable source, be national in inequality. That the United States is worse coverage, and be based on either consump- (as opposed to more transparent) than other tion or income. (Which comes from which is countries seems unlikely. identified.) A subset of the observations are International data on income inequality labeled “high-quality” and these have been come from a number of standard sources. widely (and mostly uncritically) used in a Perhaps the most reliable is the large number of papers, including papers on Luxembourg Income Study (LIS) which income inequality and health. Much of the contains information on the distributions of high-quality data comes from industrialized disposable income for 25 (wealthy) coun- countries, so that many researchers inter- tries over a period of twenty years, although ested in poor countries have used at least not all countries have data for all years; see some data not so labeled. Peter Gottschalk and Timothy Smeeding While DS’s data and documentation are a (2000, pp. 273–74). The LIS permits access great improvement over what was previously to the micro data from broadly comparable available, they do not support the uncritical income data for the covered countries. Note use that has been made of them, as shown in that the underlying surveys do not use the an important study by Atkinson and Andrea same questionnaire, so that the comparabil- Brandolini (2001). Atkinson and Brandolini ity is not perfect, nor are response rates the focus their attention on the subset of the DS same for all countries. Nevertheless, the data for the OECD countries, for which Article 3 2/21/03 4:20 PM Page 138

138 Journal of Economic Literature, Vol. XLI (March 2003)

there are good, well-documented surveys income by state using the Current (including the LIS) which can be used for Population Survey, which has a sample size comparison. DS’s “high-quality” estimates of around 50,000 households each year. do not do well in this comparison, either However, it publishes inequality measures across countries, or in some cases over time by state only for three-year moving averages. within countries. For example, DS shows Several developing countries also have regu- Sweden as one of the more unequal coun- lar, national household surveys that are large tries in the OECD, with more income in- enough to support considerable disaggrega- equality than the United Kingdom, whereas tion; India, Indonesia, and Pakistan are ex- in the LIS (as reported in Gottschalk and amples. The Indian National Sample Survey Smeeding 2000), Sweden has the lowest in- (NSS) collects detailed consumption data come inequality and the United Kingdom from more than 120,000 households every the highest apart from the United States. In five years or so, and these surveys are de- some cases, such as Germany, the DS time signed to be representative for more than series of inequality is quite different from seventy regions of the country. Yet even this that computed directly from the surveys. survey does not support the measurement of Although the DS data may be more reliable income inequality for districts, the level at for poor countries than for rich, it is unlikely, which the Indian census publishes much of especially since the poor countries contain a its data on child mortality. much larger fraction of the data that are not Most household surveys have a two-stage endorsed by DS themselves. stratified design in which, at the first stage, We are currently in the position of not primary sampling units (PSUs) are randomly having any consistently reliable set of data selected with probability proportional to on income inequality outside the countries population size. These PSUs are typically covered by the LIS. This is in spite of the ex- small geographical units, such as villages or istence of the fifty or so surveys that have census tracts. Within each PSU, the same been collected under the aegis of the World number of households are selected so that, Bank’s Living Standard Measurement over the two stages together, each household Survey (LSMS), which was set up in 1980 in the population has an equal chance of be- with the original purpose of generating com- ing selected into the survey. In some studies, parable data on income distribution for a investigators have calculated measures of lo- wide range of countries. While the LSMS cal income inequality based on the house- surveys are broadly comparable, the ques- holds in each PSU. This procedure is obvi- tionnaires are not identical across surveys, ously dangerous when there are only a few and some have differed a great deal. households in each PSU and, even when this Research on inequality and health has also is not the case, respondents within PSUs are used data on measures of inequality for sometimes not randomly selected, so that areas within countries. In principle, such the relationship between the sample esti- measures are less problematic, if only be- mate and its population counterpart cannot cause they are usually calculated from na- be assessed. Note too that PSUs are selected tional surveys using a uniform survey instru- for statistical convenience, not analytical ment so that, even if there are errors, the meaning, and frequently do not correspond patterns across areas may not be much af- to any sensible definition of a community. fected. One problem is sample size, espe- 3.2 Cross-Country Studies of Income cially for small areas. Inequality measures Inequality and Health are usually less precisely estimated than means so that, for example, the U.S. Bureau Cross-country studies have played an im- of the Census publishes estimates of mean portant part in the literature on income Article 3 2/21/03 4:20 PM Page 139

Deaton: Health, Inequality, and Ecomomic Development 139

inequality and health. Preston’s (1975) semi- changes over time. Countries, such as nal analysis looked at international patterns France and Greece, that narrowed their in- of GDP and life expectancy, and it was on come distributions by reducing relative the basis of his findings that Preston sug- poverty, increased their life-expectancies, gested that there should be a negative rela- while those, such as the United Kingdom tionship between income inequality and and Ireland, whose income distributions health. G. B. Rodgers (1979) and A. T. Flegg widened, fell behind (Wilkinson 1996, figure (1982) were early studies that followed 5.4). Wilkinson interprets these results as Preston’s lead, explicitly looking for (and showing that, as countries become wealthier finding) effects of income inequality on and move through the epidemiological tran- mortality. Rodgers used the Paukert (1973) sition, the leading cause of differences in data for 56 (unnamed) countries and, con- mortality moves from material deprivation trolling for income and other variables, to social disadvantage. Material deprivation found hazardous effects of inequality on life- provokes poverty and infectious disease, expectancy at birth, life-expectancy at age while social disadvantage provokes stress five, and on the rate of infant mortality, with and chronic disease. the last only significant in the developed Later research has cast considerable countries in the sample. Flegg (1982) looked doubt on the robustness and reliability of only at child mortality, and found significant many of these findings. As expected, one of effects of income inequality on child mortal- the main difficulties lies in the unreliability ity in developing countries using the Jain of the data on income inequality. For exam- (1975) data. These authors, like Preston, ple, using the Deininger and Squire data, thought of the nonlinearity in the relation- Gravelle, John Wildman, and Matthew ship between income and health as the basis Sutton (2002) fail to replicate Rodgers for their results, and did not propose any di- (1979) results for developed and developing rect effect of income inequality at the micro- countries. Jeffrey Mellor and Jennifer Milyo economic level. A direct effect was found by (2001) use a sample of 47 developing and Robert Waldmann (1992) who used U.N. developed countries in 1990 and find that and World Bank sources, supplemented by the positive correlation between the gini co- income inequality data from Jain (1975), to efficient and infant mortality vanishes once investigate infant mortality on a cross sec- secondary school enrolment is controlled tion of up to 57 developing and developed for, while the negative correlation between countries. As expected, he found that, condi- income inequality and life expectancy is tional on mean income, the share of income eliminated by controlling for income per going to the poorest 20 percent of the popu- head. (Commenting on Mellor and Milyo, lation decreased infant mortality, and more Kawachi and Tony Blakely 2001 argue that surprisingly, that the share of income going poor education may be a consequence of to the top 5 percent increased infant mortal- high income inequality, and it is therefore ity. This is a direct effect of inequality; the invalid to include it as a control.) Mellor and infant mortality rates among the poor in- Milyo also fail to replicate Wilkinson’s re- crease when the rich get richer, even when sults for developed economies. Although the their own incomes do not suffer. DS data have their own problems, the origi- Perhaps the single most cited finding in nal results are clearly not robust; see also the literature is Wilkinson’s (1992, 1994, Lynch et al. (2001) who use the LIS data to 1996) demonstration of a relationship be- demonstrate how the choice of “sample” tween income inequality and life expectancy points (i.e. countries) affects the link be- across a number of industrialized countries, tween income inequality and adult mortality. not only in levels but, more impressively, in The same is true of Waldmann’s findings; Article 3 2/21/03 4:20 PM Page 140

140 Journal of Economic Literature, Vol. XLI (March 2003)

Andrew Baumbusch (1995) replicated nor fully accurate. The debate between Waldmann’s analysis using data of the same Wilkinson (1998) and Judge, Mulligan, and vintage, but found that income accruing to Benzeval (1998) has focused on differences the top 5 percent reduced infant mortality in response rates across the LIS surveys, and once the data were updated from the 1993 their possible effect on the results. It is also edition of the World Bank’s World Devel- possible that, as with the difference between opment Report. infant mortality and life-expectancy, there The single most convincing study of the will be links between specific causes at spe- LIS countries is by Judge, Jo-Ann Mulligan, cific ages and the plausibly associated mea- and Benzeval (1997), who emphasize the sures of inequality, see for example Sandra poor quality of the data in previous work, McIsaac and Wilkinson (1997). Yet, it is and use the LIS data in their own examina- surely time to agree that there is currently tion of life expectancy and infant mortality in no evidence that income inequality drives Australia, Belgium, Canada, Finland, life expectancy and all-cause adult mortality France, Germany, Ireland, Italy, Luxem- within the industrialized countries. It re- bourg, The Netherlands, Norway, Sweden, mains to be seen whether this means there is Switzerland, the United Kingdom, and the no relationship, or whether there is a rela- United States. In these data, which are the tionship that is being obscured by still inade- best international data currently available, quate data. Judgement on that depends a the correlation between the gini coefficient good deal on whether there exists a relation- and life expectancy is –0.17, insignificantly ship between income inequality and health different from zero, and neither the gini nor in other contexts, on which more below. other measures of income inequality are That, conditional on income, there should close to significance in any of the regressions be a cross-country relationship between in- explaining life expectancy. The situation is fant mortality and income inequality, at least somewhat different for infant mortality in poor countries and possibly in rich coun- rates, where there is a significant positive tries, is both theoretically plausible, and (i.e. harmful) effect of the ratio of the 90th rather better supported by the (admittedly to the 10th percentile. This measure of in- inadequate) data that are available. The equality exerts a significant effect in several plausibility comes from recent work from of the regressions, though it becomes in- the World Bank which, following the significant when controls are added for the methodology pioneered by Deon Filmer negative effects on mortality of female labor and Lant Pritchett (1999b), has used demo- force participation, see also Lynch et al. graphic and health surveys around the world (2001) for a similar finding. In these data, to construct a synthetic measure of wealth, the raw correlation between infant mortality which is then used to explain infant mortal- and inequality is driven largely by the ity rates; see Davidson Gwatkin (2000) for United States, which is very unequal and has an overview. The measure of wealth is an relatively high infant mortality. Inequality in index based on the ownership of various the United States may indeed have some- durable goods, and while the measure is thing to do with its high infant mortality undoubtedly correlated both with actual rate; however, as I shall argue below, any sat- wealth and income, we have no way of cali- isfactory discussion of income inequality and brating the transformation and thus of using mortality in the United States must take ex- the results to relate income to child health. plicit account of the effects of race. (The transformation is also different in dif- If these results are not entirely definitive, ferent countries, which effectively precludes it is because the LIS data, although better meaningful cross-country comparisons.) than any other, are neither fully comparable Nevertheless, the results show very strong Article 3 2/21/03 4:20 PM Page 141

Deaton: Health, Inequality, and Ecomomic Development 141

gradients in child health, with infant mortal- mation. An exception to this generalization ity rates heavily concentrated at the bottom comes from the countries of Eastern Europe of the distribution. Wagstaff (2000) uses and the former Soviet Union, where life ex- nine (mostly) LSMS surveys from develop- pectancy has been falling as income inequal- ing countries to calculate child mortality ity has increased (see Marmot and Martin rates by quintile of equivalent consumption, Bobak 2000, fig. 3, which shows a twelve- and shows that child and infant mortality country correlation coefficient of –0.63). rates typically decline most rapidly between Peter Walberg et al. (1998) also show that the bottom and second quintile. Whether life-expectancy fell fastest in the initially these results imply that infant mortality rates most unequal regions of Russia, a finding are convex in income depends on the degree that is not obviously related to the inequality of convexity of the relationship between in- hypothesis. Furthermore, as is widely recog- come and the asset index, and on the density nized, the Eastern European experience is function of equivalent consumption, but that difficult to interpret because so much else infant mortality is concentrated at the bot- has been going on, so that it is hard to isolate tom of the income distribution seems likely. the effect of income inequality. Yet there is also some evidence on the other Finally, there are a number of cross-coun- side. In particular, Mamta Murthi, Anne- try studies that link other health outcomes to Catherine Guio, and Jean Drèze (1995) find income inequality. Steckel (1995) finds a re- very little effect of poverty on child mortality lationship between human stature (a mea- across districts in India once they control for sure of cumulative nutritional status) and in- other factors, most importantly female liter- come inequality on a sample of developed acy and urbanization. and developing countries using the income To the extent that the DS data are ac- distribution from Jain (1975). Mead Over cepted, there is a good deal of empirical evi- (1998) looks across cities in the developing dence from developing countries linking in- world and finds that the U.S. Census fant and child mortality to the DS measures Bureau’s estimates of HIV infection rates of income inequality conditional on the level are positively related to the DS measures of of GDP per head and a range of other vari- countrywide income inequality. He inter- ables (for example in Pritchett and prets his findings in terms of upper income Lawrence Summers 1996; Filmer and men demanding the services supplied by Pritchett 1999a; and Simon Hales et al. lower income women, and these results per- 1999). Whether this evidence extends to haps come closest to providing substance to adult mortality and life expectancy is diffi- Paul Farmer’s (1999) contention that disease cult to know, not only because of the data occurs along the “fault lines” in the income difficulties with income inequality, but be- distribution. Marie Gaarder (2001) argues cause of the quality of the data on adult mor- that income inequality is likely to worsen the tality. Few poor countries have complete health consequences of pollution because registration systems for deaths, so that good the poor have lower baseline health and are evidence on adult mortality (or life ex- therefore more susceptible. She includes pectancy at age five, for example) is hard to the gini coefficient in a meta-analysis of pre- come by. In a few cases, such as India, there vious estimated effects of particulate con- are sample registration surveys, and some centration on mortality at various sites data on adult deaths can be gleaned from around the world and finds significant posi- the demographic and health surveys. But for tive effects. Pablo Fajnzylber, Daniel many countries, data on life expectancy are Lederman, and Norman Loayza (2000) find extrapolated from the data on infant mortal- a significant relationship between DS gini ity rates, and contain little additional infor- coefficients and both homicide and robbery Article 3 2/21/03 4:20 PM Page 142

142 Journal of Economic Literature, Vol. XLI (March 2003)

rates for a group of 45 (for homicides) and Social Survey. Such an account is very much 34 (for robberies) developed and developing in the spirit of stories of psychosocial stress countries. A good deal of this is driven by within unequal social structures. In support Latin American countries, where both crime of this explanation, Wolfson et al. (1999) es- and inequality are very high. Using data on timate the degree of nonlinearity in the in- seventeen countries in the Americas, and come to mortality curve using the NLMS gini coefficients from DS, Oscar Mujica et (as in figure 2) and show that the effects of al. (2000) confirm the positive correlation income inequality on state mortality rates (0.55) between homicide and income in- are too large to be explained by the nonlin- equality, but find a negative correlation earity argument alone, so that there must be (–0.78) between suicide rates and income some direct effect of income inequality on inequality. individual mortality. The implications of these results for economic policy have not 3.3 Within-Country Area Studies gone unnoticed. See for example Kaplan of Income Inequality and Health and Lynch’s (2001) editorial in the As skepticism has grown about the inter- American Journal of Public Health entitled national relationship between income in- “Is Economic Policy Health Policy?” equality and health, attention has switched These within-nation results do not suffer to studies within countries, particularly of from the same data problems as do the in- mortality and income inequality across the ternational comparisons. Income inequal- states of the United States. Two studies, by ity is usually measured from incomes col- George Kaplan et al (1996), and by lected in the census, which is administered Kennedy, Kawachi, and Prothrow-Stith in the same form to all households in all (1996a,b), both published in the British states. Nor is there any question about the Medical Journal, and inspired by existence of the correlation. Figure 6 Wilkinson’s (1992) cross-country work, shows a typical scatter plot between the log found a relationship across the states be- odds of age-adjusted mortality (the log of tween various measures of income inequal- the ratio of the fraction dying to the frac- ity and age-adjusted all-cause mortality, as tion not dying) on the vertical axis, and the well as a number of other measures, includ- gini coefficient of household income per ing infant mortality rates, deaths from can- equivalent, with equivalents defined as 1 cer, coronary heart disease, and homicide, for adults and 0.5 for children aged eigh- as well as disability, low birth weight, and teen and less. The District of Columbia is crime. Kawachi and Kennedy (1997) estab- included and, although it is an outlier in lished that the results were robust to the the sense of having higher income inequal- choice of inequality indicator, while Lynch ity and higher mortality than any state, it et al. (1998) extended the results to 282 lies along the regression line defined by metropolitan statistical areas (MSAs) in the other observations. Across the states of 1990, finding that the loss of life from in- the United States, income inequality is come inequality “is comparable to the com- strongly negatively correlated with income; bined loss of life from lung cancer, diabetes, poor states, many of which are in the motor vehicle crashes, HIV infection, sui- South, also have the weakest safety nets for cide, and homicide in 1995.” Kawachi, the poor. Even so, the correlation between Kennedy, and Prothrow-Stith (1997) argue income and mortality is much weaker than that income inequality works by reducing the correlation between income inequality social capital, in particular the degree of and mortality, and adding income to a re- trust between people, a state-level measure gression does not eliminate the effect of of which is constructed from the General income inequality. Article 3 2/21/03 4:20 PM Page 143

Deaton: Health, Inequality, and Ecomomic Development 143

12 DC

MS LA 10 SC GA AL MD DE NY ME TX NH VT IN OK CA 8 WI FL IA AZ NM MN age adjusted mortality per 1000 UT HI

6 .33 .35 .37 .39 .41 .43 .45 .47 .49 gini of per equivalent income

Figure 6. Income Inequality and Age-Adjusted Mortality, U.S. States, 1990 Source: CDC for state (all race, both sex) mortality data, author's calculations from 1990 census for income inequality. Note: Circles are proportional to population size.

Nevertheless, there are serious questions Similar reversals are found for the fraction about whether the correlation between in- of births that are low birth weight while, come inequality and mortality is robust over the five decades, there is no relation- through time, and whether it comes from ship across states between deaths from car- the effects of income inequality or some diovascular disease, from malignant neo- other factor that is correlated with it. Mellor plasms, or from liver disease. Indeed, for the and Milyo (2001) use data for the 48 conti- first two, income inequality has a negative nental states from five census years—1950, and significant relationship with deaths once 1960, 1970, 1980, and 1990—and reproduce controls are entered for income, education, the strong hazardous effect of the gini coef- race, and urbanization. Only for homicides ficient on all-cause mortality when only year and, to a lesser extent, infant mortality and dummies, the age composition of the state, deaths from accidents, is the gini coefficient and median income are included as controls. a risk factor conditional on the other con- The inclusion of controls for the average trols. Mellor and Milyo also subject the hy- level of education in each state eliminates pothesis to a much more stringent test, look- the significance of the gini coefficient and, ing at the relationship between ten- and once the authors include controls for the twenty-year changes in mortality and the fractions of people in each state who are ur- corresponding changes in income inequality. banized and who are black, the gini coeffi- This is almost certainly too severe a test be- cient attracts a negative sign, though one cause it places a great deal of weight on the that is not significantly different from zero. timing of the link between income inequality Article 3 2/21/03 4:20 PM Page 144

144 Journal of Economic Literature, Vol. XLI (March 2003)

and mortality. Even so, it is worth noting significantly different from zero. This result that, with one exception, none of the income means that the effect of inequality on whites inequality to mortality relationships survives comes, not from the inequality of white in- the test. The exception is homicide, where comes, but from the inequality between the relationship with income inequality is whites and blacks, raising the suspicion that well-determined and holds over time as well the effect has more to do with race than with as in the cross-section. income inequality. Such a suspicion is borne Another concern is the pooling of data out by controlling for the fraction of the across racial or ethnic groups with different population that is black in each state. It incomes and different mortality rates. In the turns out that a high fraction of blacks raises United States, African Americans have mortality rates among both males and fe- higher mortality rates and lower incomes males (note that these are whites) and that than whites, so that states with a high frac- conditional on race, income inequality has tion of blacks tend to have higher mortality no effect on mortality. It is unclear why the rates as well as higher income inequality. As fraction black should exert such a strong ef- can be seen from figure 6, such states tend fect on white mortality (black mortality is to be predominately in the South where also higher in states where there are rela- many other special factors are likely to oper- tively many blacks), though it might be ar- ate; by contrast, the states with low mortality gued that the fraction black is itself some are states where there are few or no African sort of marker for the inequality that charac- Americans. If data are pooled for 1980 and terizes race relations in the United States. 1990, the log odds of age-adjusted mortality Even so, the effect is not one that works responds to the gini coefficient with a coeffi- through income inequality; once the fraction cient of 1.7 for males, and 1.1 for females. In black is included in the regression, the gini the same regression, the coefficient on the coefficient has no effect. mean of the logarithm of equivalized income There is an obvious concern here that I is negative for men, but barely significant, have simply replaced one invalid variable— and not at all for women (Deaton 2001b). To income inequality—with another—racial illustrate the size of the effect of inequality, composition—and that both stand proxy for the 1990 gini coefficients for Louisiana and something else. This is particularly the case New Hampshire were 0.45 and 0.33 respec- with the state data, where there are at most tively, which would account for a 12-percent 51 observations (or 102 observations if we difference in mortality rates, more than half pool data from 1980 and 1990), and where it of the difference shown in figure 6. would be easy to confound racial composi- If we now confine the calculations to tion (or income inequality) with geographi- white mortality alone, so that we no longer cal factors, especially given the peculiar role have the mechanical effects described of the South. Nevertheless, Deaton and above, the coefficient on the gini drops to Lubotsky (2003) show that the results carry 1.1 for men and to 0.6 for women, about a through to the 287 MSAs that can be consis- third lower than for all-race mortality. tently identified between 1980 and 1990. Nevertheless, these effects remain signifi- These data can be used to replicate the find- cantly different from zero, and still show ings of Lynch et al. (1998), and to show that, that inequality is a health hazard for the once again, the inclusion of racial composi- white population alone. If we recalculate the tion eliminates (and sometimes even re- gini coefficients so as to measure only in- verses) the effect of income inequality. And equality among whites, the effects are fur- because there are more MSAs than states, it ther reduced, to 0.6 for men, and 0.4 for is possible to work within regions, and to women, and only the former is (marginally) show that whether we look at cities in the Article 3 2/21/03 4:20 PM Page 145

Deaton: Health, Inequality, and Ecomomic Development 145

South, or cities in other regions, and condi- ity and that, in those areas, both blacks and tioning on city average income, white mor- whites have higher mortality rates. In the tality is higher in cities where the fraction of event of an acute myocardial infarction blacks is higher. The results also holds across (heart attack), people are taken to the near- nearly all age and sex groups. In apparent est facility whatever their income, health in- contradiction to these results, McLaughlin surance status, or race, so that whites, like and Stokes (2002) find that, with county blacks, will receive the health care provided data, controlling for race does not eliminate in their community. As a result, both blacks the effects of income inequality on mortality. and whites suffer from the lower-quality Perhaps these findings come from the un- health care provided to blacks. Furthermore, fortunate choice of inequality measure (the in the many metropolitan areas that have ratio of the share of the bottom 90 percent multiple hospital referral regions, the effects to the bottom 10 percent—see above); of the fraction black on white mortality will Deaton and Lubotsky’s findings are repli- be attenuated, and on black mortality exag- cated on the county data. gerated by residential segregation. Segre- That the fraction black eliminates the ef- gation effectively transfers the burden of fects of income inequality or mortality is poor treatment from whites to blacks; see easy enough to understand mechanically; Douglas Massey (1991) for this argument in blacks are poorer and whites richer in cities a more general context. with a large fraction black, so that income If the apparent effects of inequality or inequality is higher where the fraction black mortality are nothing to do with income in- is higher, but it is the fraction black, not in- equality per se, but are instead tied to race come inequality, that matters for both white in America, then there should be no rela- and black mortality. But the finding tells us tionship between income inequality and nothing about what it is about largely black mortality across regions in countries where cities that makes them less healthy. One ar- race does not have the same salience. This gument is similar to the case against income seems to be the case. Nancy Ross et al. inequality, that racism generates stress and (2000) find that there is no relation between lack of social capital, sickening both the per- income inequality and mortality for the ten petrators and the victims. But there are provinces and 53 metropolitan areas of other more mundane stories. While there is Canada, and Ross has similar, unpublished considerable evidence that the quality of findings for Australia. For Britain, there ap- health care is lower for African Americans pears to be no area study on income inequal- than for whites (Brian Smedley, Adrienne ity and health, though Yoav Ben-Shlomo, Stith, and Alan Nelson 2002), there is much Ian White, and Marmot (1996) find that less evidence that treatment is different be- mortality in the 8,464 wards of England is tween blacks and whites within the same affected not only by an index of deprivation treatment area, technically within a hospital based on household characteristics, but also referral region (Amitabh Chandra and by the within-area dispersion of the depriva- Jonathan Skinner 2002). There is also evi- tion index. Again, this is what is to be ex- dence in Fuchs, Mark McClellan, and pected if the deprivation measure is more Skinner (2002) that the association between closely linked to mortality among high- the fraction black and white mortality deprivation people. Likewise, Debbie among the elderly is strongest for cardiovas- Stanistreet, Alex Scott-Samuel, and Mark cular disease. So one possibility is that, in ar- Bellis (1999) find a link between mortality eas where African Americans make up a and imputed income and income inequality large fraction of the population, hospitals are across 366 English local authorities. But less well-funded or otherwise of lower qual- their income measures are imputed from Article 3 2/21/03 4:20 PM Page 146

146 Journal of Economic Literature, Vol. XLI (March 2003)

the occupational structure of each area, as- not make the point, a link between homicide suming that wage rates are identical every- among men and the shortage of women in- where, so that they are measures of local oc- vites a socio-biological explanation in terms cupational patterns, not of income nor of mating behavior. Rodolfo Peña, Stig Wall, income inequality. and Lars-Åke Persson (2000) find that infant Tung-liang Chiang (1999) looks at mortal- mortality risks are higher among the poor in ity rates in the 21 counties and cities of Nicaragua, and higher still when the poor Taiwan in 1976, 1985, and 1995 using house- live in relatively wealthy neighborhoods, hold survey data to calculate measures of in- which is consistent with a negative role for come and income inequality. He finds strong inequality. protective effects of income in 1976 and That there is no interregional relationship 1985, and little effect of income inequality, between income inequality and mortality is but finds that the situation is reversed in somewhat surprising, given the nonlinear re- 1995, at which date income inequality is a lationships between income and health with hazard, and income has no effect. Chiang in- which I began. Perhaps the curvature is too terprets his findings as support for mild to generate any strong relationship— Wilkinson’s idea that income is important at and Wolfson et al. (1999) provide such evi- low levels of income, and income inequality dence—or perhaps income is not causal. I at high levels of income, which, as we have have already referred to the regional esti- seen, is also consistent with a nonlinear ef- mates for the United States where education fect of income and no direct influence of in- drives out education from aggregated mor- equality. Enrique Regidor et al. (1997) find tality relationships. Another place to look is no relationship between (a nonstandard) the time-series data and the patterns of in- measure of income inequality and the preva- come, income inequality, and mortality over lence of long-term disability across the sev- time, and I shall examine that evidence in enteen regions of Spain. section 3.5 below. It is widely believed that there is a link be- 3.4 Studies of Income Inequality and tween income inequality and crime (includ- Health Using Individual Data ing homicide) in the United States. I have al- ready noted Mellor and Milyo’s (2001) Studies using individual-level data face finding that homicide was the only negative different data problems from either the na- health outcome that was robustly linked to tional or the area studies. They have the ad- income inequality in their tests, and such vantage of being able to look for a direct ef- findings have consistently appeared in the fect of income inequality without having to literature since Isaac Ehrlich (1973); see allow for the effects of inequality that work C. C. Hsieh and M. D. Pugh (1992) for an through aggregation. But there are compen- oft-quoted review and meta-analysis. sating difficulties. Because mortality is a rare There appear to be few relevant studies event, large sample sizes (or long follow- from developing countries, even where it ups) are required to give enough deaths to would be possible to do so, for example in estimate mortality rates reliably. At the same the work on fertility and child mortality in time, those few health-related surveys that India by Murthi, Guio, and Drèze (1995) follow people from interview to death are and Drèze and Murthi (2001). However, typically very poorly endowed with eco- Drèze and Reetika Khera (2000) find that nomic information, including incomes. homicide rates across India are unrelated to Nevertheless, there are several surveys in measures of consumption inequality, but are the United States that have been used to positively associated with the fraction of look at the determinants of mortality at the “missing” women. Although the authors do individual level. The National Longitudinal Article 3 2/21/03 4:20 PM Page 147

Deaton: Health, Inequality, and Ecomomic Development 147

Mortality Study starts from data collected in they are asked easily and quickly. They ask the Current Population Survey, mostly respondents to rate their health on a five- around 1980, and then uses the National point scale from “poor,” to “fair,” to “good” Death Index to check whether members of to “very good” and “excellent.” Many investi- responding households are dead by each gators convert this to a binary indicator of follow-up date. Currently, around 1.3 mil- poor health, corresponding to the “poor” lion people have been tracked for up to a and “fair” categories; such an indicator has decade. In principle, the CPS provides ex- been validated as a powerful predictor of cellent and detailed economic information, subsequent morbidity and mortality, even but many of the rounds used for the NLMS conditional on a physician’s examination; were not the March surveys, when income (Ellen Idler and Yeal Benyamini 1997). data are collected, and so contain only rudi- Even so, there is no automatic link between mentary information on household incomes. morbidity and mortality; indeed there are There are two other U.S. health surveys reasons why they might move in opposite di- with later merges of death-certificate data: rections; see Christopher Murray (1996) for the National Health Interview Survey a good discussion. Self-reported general (NHIS), which interviews around 50,000 health questions are included in the NHIS, households every year, and the National in the Behavioral and Risk Factor Sur- Health and Nutritional Examination Survey veillance Study (BRFSS) and, from 1995 on- (NHANES) the first round of which sur- wards, in the Current Population Survey. veyed more than 14,000 people between The interpretation of the individual stud- 1971 and 1975, for whom information on ies, and of the extent to which they support a deaths has been merged up to 1987. A final link from inequality to health depends a source of mortality data is the Panel Study of great deal on who is doing the interpreting. Income Dynamics (PSID), which has fol- Nevertheless, there is general agreement lowed around 5,000 households (and their that the results from these studies are children and split-offs) since 1968. Because weaker and more ambiguous than the area this is a panel survey returning regularly to studies. For example, Kimberley Lochner et each household, deaths are reported by sur- al. (2001), using the NHIS and merged mor- viving family members. All four of these sur- tality data, find only a small effect of state in- veys have been used to look at the relation- come inequality on mortality (relative risk of ship between mortality, income, and income living in the top five most unequal states inequality. compared with the ten most equal states of Sweden also has nationally representative 1.12). This effect, which is estimated with data that works in the same way as the controls for family income and the state- NLMS in the United States, though with level poverty rate, is only statistically signifi- even more comprehensive data. Since 1975, cant for near-poor whites. Kevin Fiscella Statistics Sweden has interviewed around and Peter Franks (1997) find that, once they 7,000 individuals each year in its Survey of control for individual income, there is no ef- Living Conditions, and these people have fect of PSU-level inequality on the probabil- been linked, not only to the national death ity of dying in the NHANES follow-up, but statistics, but to income information from there are real questions about whether their the national income tax statistics; see Ulf measure of inequality—the share of income Gerdtham and Magnus Johannesson (2001). accruing to the bottom 50 percent of the A second line of work has used, not mor- population—can be adequately measured tality, but self-reported measures of overall from the PSU data in the NHANES itself, health status. The questions are included in within which respondents are not randomly a large number of surveys, if only because drawn. Mary Daly et al. (1998) find no effect Article 3 2/21/03 4:20 PM Page 148

148 Journal of Economic Literature, Vol. XLI (March 2003)

of state-level income inequality on individual and 1998. This conclusion is echoed for five-year mortality rates using data from the Japan, where the raw correlation between PSID. As I shall explain in more detail be- poor health and the prefecture gini coeffi- low, state-level income inequality also has no cient is eliminated once individual income is detectable effect on individual mortality in controlled for (Kenji Shibuya, Hideki the NLMS. Hashimoto, and Eiji Yano 2002). Several studies find a range of results, As noted, Wolfson et al. (1999) show with at least some studies finding an effect that the degree of curvature in the NLMS of income inequality on self-reported mor- (figure 2) is insufficient to explain the large bidity. See Mah-Jabeen Soobader and effects of income inequality on mortality at Felicia LeClere (1999, using the NHIS with the state level. In Deaton (2001b), I ad- merged census county-level inequality data), dress more directly the role of state in- Kennedy et al. (1998, using the BRFSS come inequality on mortality in the NLMS. merged with state-level inequality data), The NLMS distinguishes seven income Fiscella and Franks (2000, using the NHIS), groups, so that at the first stage of the Mellor and Milyo (2002) and Blakely, analysis, I use a logit model to estimate the Lochner, and Kawachi (2002) using the CPS log odds of dying during the ten-year fol- with inequality data from the CPS at various low up as a linear function of age including geographical levels; and Roland Sturm and dummy variables for each of the seven in- Carole Roan Gresenz (2002) and Robert come groups. These logits are estimated Kahn et al. (2000) using other health sur- for white males and females separately, us- veys. The estimated effects are typically ing data only for those aged 18 to 75 at the modest, and in several cases are eliminated time of first interview. (The log odds of when controls are added for individual in- mortality is approximately linear in age come. For example, Mellor and Milyo show over this range.) In order to conduct a that their effects are removed once controls state-level analysis, each of these models is are introduced for income and its square, as fitted to data for a single state, thus allow- well as for fixed state effects. (Note again ing inequality—or any other state-level ef- that this last is a severe test; Mellor and fect—an unrestricted effect on the rela- Milyo only have three years of CPS data, so tionship between mortality and income. they are effectively demanding a link be- The first stage produces numbers for each tween changes in morbidity and changes in state like the points shown as circles in fig- income distribution between 1995 and ure 1, so that, at the second stage, it is pos- 1997). But LeClere and Soobader (2001) sible to examine whether these points are also demonstrate considerable fragility in higher in states where income inequality is the results, showing that the effects seem to higher. Note that this two-stage procedure work only for whites aged 18–44 in high- is as general as a single-stage model in inequality counties, and middle-aged whites which individual mortality is linked to in very-high-inequality counties. There are state-level data on income inequality. no effects for other whites, nor for non- My original concern was to test the model Hispanic blacks. Blakely, Lochner, and of relative deprivation presented in section Kawachi (2002) is the most recent and one 2.2.5 above. This was done by comparing the of the most careful studies; it concludes that, effects of each income group in each state after controlling for income, and apart from with the predicted values from computing the possible effects of income inequality that relative deprivation. While states are by no work through aggregation, there is little as- stretch of the imagination plausible refer- sociation between income inequality and be- ence groups, it is at the state level that the ing in poor health in the CPS data from 1996 inequality effects were first discovered, so it Article 3 2/21/03 4:20 PM Page 149

Deaton: Health, Inequality, and Ecomomic Development 149

is tempting to try to account for the effect mortality was followed until the end of 1996. using the relative deprivation story. Within Mortality was assessed relative to individual states, the relative deprivation story does income, community income, and commu- well, outperforming a simple model in nity income inequality, with the latter two which income itself accounts for the differ- measured from the survey data itself, a pro- ences across income groups; Eibner (2000) cedure subject to the reservations raised finds similar results using a range of defini- above. As in all similar studies, individual in- tions for reference groups. However, the rel- come was strongly protective, even allowing ative deprivation model accounts for essen- for education and a host of other variables, tially none of the variation in mortality including initial health status, but neither in- across states which, given the theory, means equality nor mean community income ap- that the gini coefficient does not predict in- peared to have any effect. The last result is terstate mortality differences in the NLMS evidence against the relative income hypoth- data. This finding is supported (even with- esis so that, once again, we are led back to out controls for income) using the 1.3 mil- the original model in which health is an in- lion observations in the full NLMS. Table 13 creasing nonlinear function of absolute in- in Rogot et al. (1992) shows no correlation come. Finally, in another impressive Nordic across the states between age-adjusted mor- study, Merete Osler et al. (2002) use data on tality and income inequality, a finding that is nearly 26,000 people from Copenhagen who in direct contradiction with figure 6 and with were followed for an average of 12.8 years. the findings listed at the outset of section 3.4 These survey data, which include data on above. This contradiction is resolved (at behavioral risk patterns, were merged with least in part) by the demonstration that, in administrative data on housing, income, ed- these individual-level data as in the aggre- ucation, and occupation, as well as income gate state-level data, the fraction black in equality (share accruing to the bottom 50 each state is a powerful predictor of white percent) at the parish level, calculated from mortality. population, not survey, data. Without con- Once again, there is no direct effect of in- trols, income inequality is positively related come inequality. Because the racial compo- to mortality, but the effect is eliminated by sition of areas is such a strong predictor of adding the other variables; indeed, adding mortality in the aggregate state, MSA, and individual income is sufficient to eliminate county-level data, as well as in the individual the effect of inequality on mortality. data from the NLMS, it would be interesting 3.5 Inequality and Mortality Decline to discover whether the Lochner et al. in the United States and Britain (2001) findings on mortality in the NHIS follow-up can also be attributed to racial A final source of evidence comes from ex- composition; given their partition of states amining whether there is a link between into inequality groups, it seems likely. mortality and the increase in income in- Taking income and mortality together, the equality in the 1980s in both Britain and the Swedish data used by Gerdtham and United States. Wilkinson (1996) argues that Johanesson (2001) are probably of higher for Britain, mortality rates for infants and for quality than anything currently available young adults fell less rapidly after 1985 than in the United States. Gerdtham and would have been the case had income in- Johanesson used the 284 municipalities of equality remained constant. Wilkinson Sweden as their communities, and examine (1996, figure 5.10, p. 97) plots a time series individual mortality for 41,006 individuals of mortality, not only of infants, but also of aged between 20 and 84 who were inter- children and young adults, and shows that viewed between 1980 and 1986 and whose the sum of age-adjusted mortality rates fell Article 3 2/21/03 4:20 PM Page 150

150 Journal of Economic Literature, Vol. XLI (March 2003)

less rapidly after 1985 than it did in the the deaths of low birth-weight babies. These decade from 1975 to 1985. These findings, techniques diffused more quickly in the together with the corresponding evidence United States than in the United Kingdom, for the United States were recently exam- so that the rapid decline in mortality started ined in Deaton and Paxson (2001b). Their first in the United States, and its possibilities results are as follows. were more rapidly exhausted there. There There were large increases in income in- seems no reason other than coincidence to equality in both Britain and the United link the timing of this exhaustion to the rise States in the 1970s and 1980s. In both coun- in income inequality. Third, among young tries, inequality in family and household in- adults, much of the increase in mortality is come increased from the early or mid-1970s attributable to HIV/AIDS, for which the rise until around 1990, with (arguably) little in- in income inequality in the mid-1980s is not crease but certainly no decline since. By the the cause. Finally, if we look at mortality early to mid-1980s, inequality had risen to rates of adults aged 45 and above, there is a new postwar highs and continued to in- period of unusually rapid mortality decline crease, at least until 1990. As pointed out by (particularly although not exclusively for Wilkinson, the rate of decline of infant mor- men) that began around 1970 (again a little tality was particularly rapid in the decade earlier in the United States), and continues from 1975 to 1985, and less rapid thereafter. to the present. So if income inequality is The same is true in the United States, hazardous for the young, it is protective for though the period of rapid decline starts their elders! Once again, a more convincing somewhat earlier, in the late 1960s, and fin- explanation lies in the increased use, first in ishes earlier, around 1980. In both countries, the United States and later in Britain, of life- the rate of decline of mortality rates among saving technologies for dealing with cardio- young adults has slowed steadily, and by vascular disease, angioplasty, coronary artery 1985 mortality rates are either flat or actu- bypass grafts, and the use of clot-busting ally rising. For infants and young adults drugs and even aspirin. taken together, the rate of mortality decline The time-series results also cast doubt on has therefore been a good deal slower in the role of income in promoting health. For recent years than in the period before the most age groups, the rate of mortality decline increase in income inequality. was slow during the period of rapid growth in Even so, it is unlikely that income in- median incomes from 1950 until the early equality has much to do with these mortality 1970s, and much more rapid thereafter dur- trends. First, the episodes of rapid decline in ing the period of the productivity slowdown. the infant mortality rate are episodes, not In Britain, there is no slowdown in the rate of trends. Prior to the periods of rapid decline growth of real disposable personal incomes in each country, progress was slower, with a after the mid-1970s, so that the pattern of in- rate of decline comparable to that after the come growth was very different from that in end of the episode. Yet income inequality the United States. Yet the patterns of mortal- was not high prior to the onset of the rapid ity decline were very similar. Neither income decline, so that in neither country since nor income inequality help us explain British 1950 has there been a consistent relation- and American mortality patterns in the two ship between income inequality and the rate quarter centuries after 1950. of decline of infant mortality. Second, we know a good deal about the causes of the de- 4. Summary and Conclusions cline in infant mortality, much of which can be attributed to declines in perinatal mortal- The stories about income inequality af- ity through new techniques for preventing fecting health are stronger than the evidence. Article 3 2/21/03 4:20 PM Page 151

Deaton: Health, Inequality, and Ecomomic Development 151

Judging by the explosion of interest and of poverty so that, conditional on average in- citations, there is a strong appeal to the idea come, income inequality is important only that before the epidemiological transition, because it is effectively a measure of income determines mortality, while after it, poverty. But it is low incomes that are im- income inequality determines mortality; that portant, not inequality, and there is no evi- in poor countries, income protects against dence that making the rich richer, however poor sanitation, unhealthy working and liv- undesirable that may be on other grounds, is ing environments, poor nutrition, and a hazardous to the health of the poor or their plethora of infectious diseases; that in rich children, provided that their own incomes countries, where these evils are but distant are maintained. The only exception to these memories, income inequality is an indicator generalizations is perhaps the case of homi- of the quality of social arrangements, of cide, where income inequality itself appears stress, and of mortality. Yet, as we have seen, to play a genuine role. even if it is true that, at higher income levels, These conclusions are not different from income inequality becomes relatively more those of earlier commentators, particularly important as a cause of death, there is no Judge (1995); Judge, Mulligan, and need to assume that the relationship be- Benzeval (1997); Wagstaff and Eddy van tween income and mortality changes with Doorslaer (1999); and Judge and Iain economic development. If it is poverty, not Patterson (2001). Yet these results must not inequality, that drives mortality, so that in- be misinterpreted. While they do imply that come has a much bigger effect on health at income inequality is not important per se, low than high incomes, average income will other than its effects through poverty, they eventually cease to be associated with poor do not imply that the social environment is health, while the effects of inequality will not important for individual health, let alone endure for much longer because, even in that individual health is determined by indi- rich economies, there are some who are not vidual characteristics and the provision of so rich. Income inequality will continue to personal medical care. We know from affect mortality until everyone ceases to be Whitehall and from other studies that posi- poor, which happens long after average in- tions in hierarchies matter, perhaps through come has risen out of the range of poverty. an ability to control one’s life, but in any case But it is not true that income inequality it- through some mechanism that works self is a major determinant of population through relationships with other human be- health. There is no robust correlation be- ings. In the United States in 1999, black tween life expectancy and income inequality males had a life expectancy at birth 6.8 years among the rich countries, and the correla- shorter than white males; for females, the tion across the states and cities of the United discrepancy was 5.2 years (CDC 2002). States is almost certainly the result of some- There is a twenty-year gap in life expectancy thing that is correlated with income inequal- between white men in the healthiest coun- ity, but that is not income inequality itself. ties and black men in the unhealthiest coun- The rapid increases in income inequality in ties (Murray 1998 quoted by Marmot 2001). the 1980s have not been associated with any Hispanic Americans have longer life- slowdown in the rate of mortality decline. expectancies than whites. And there are dif- Studies of individual mortality and income ferences within racial and ethnic groups by inequality show no link, except for one sur- education and income. As I have argued vey where the estimated effects are small above, across cities, states, and counties of and are confined to one population group. the United States, places that have a larger Infant and child mortality in developing African-American population are unhealth- countries is primarily a consequence of ier, with both whites and African Americans Article 3 2/21/03 4:20 PM Page 152

152 Journal of Economic Literature, Vol. XLI (March 2003)

dying younger than would otherwise be the equality, as the main correlate of the degree case. We do not know why this should be the to which people experience disease-induc- case, whether the nature of race relations in ing insults. Some of the institutions where America is so difficult that it is actually a rank is likely most protective—prisons or the health hazard, or whether what we see is the military—are not characterized by wide in- result of lower-quality health care in places come inequality, nor even by important dif- where there are large black populations. If ferences in income. Inequality may be im- the latter, residential segregation is likely to portant, but there is little that suggests it is improve the health of whites at the expense income inequality. of blacks, and there is some evidence in the My conclusions carry a number of implica- literature that, indeed, segregation harms tions for the direction of future research. the health of African Americans (Thomas The most obvious is that attention should be Laveist 1993 and Ingrid Gould Ellen 2000). directed away from income inequality per se. All of these cases involve inequality of one Instead, the urgent need is to investigate the kind or another, residential or racial, but not role that income plays in promoting health— income inequality per se. whether the effects of income come from in- I have also emphasized several other cases come itself, or from correlates such as educa- where reductions in deprivation in one di- tion, wealth, control, or rank. We need to mension—whether it be land ownership (the know why income is so important in the indi- nutritional wage story), democratic rights vidual level studies, and so apparently unim- (England in the 1870s or the American portant at the aggregate level over space or South in the 1960s), women’s agency (health over time. If income is indeed directly pro- and fertility in India), or income (through tective, we need to know whether the effect malnutrition or freedom from stress)—will is really nonlinear and by how much, because bring benefits not only in and of themselves, it is this, and not any direct effect of income but also to the relief from other deprivations, inequality on health, that determines in this case particularly the deprivation of ill whether and by how much income redistri- health. This is of course Sen’s (1999) theme bution can improve population health. in Development and Freedom, that relief

from any one of a number of interlinked dep- REFERENCES rivations, each of which is an important “un- Adams, Peter; Michael D. Hurd, Daniel McFadden, freedom” in its own right, helps promote re- Angela Merrill and Tiago Ribeiro. 2003. “Healthy, lief from the others. This is quite different Wealthy, and Wise?” J. Econometrics, 112, pp. 3–56. from a story in which income inequality is Aghion, Philippe; Eve Caroli and Cecilia García- Peñalosa. 1999. “Inequality and Economic Growth: the principal actor and main villain. The Perspective of the New Growth Theories,” J. These results are entirely consistent with Econ. Lit. 37:4, pp. 1615–60. what is known of about the biology of stress Alesina, Alberto; Reza Baqir and William Easterly. 1999. “Public Goods and Ethnic Divisions,” Quart. and disease. That rank can be protective of J. Econ. 114, pp. 1243–84. health, including enhancing the resistance to Almond, Douglas V.; Kenneth Y. Chay and Michael infection, has been demonstrated in both Greenstone. 2001. “Civil Rights, the War on Poverty, and Black-White Convergence in Infant Mortality in animal and human experiments (Sheldon Mississippi,” econ. dept., U. Cal. Berkeley. Cohen 1997, 1999). That repeated stress as- Atkinson, Anthony B. 1970. “On the Measurement of sociated with the insults and lack of control Inequality,” J. Econ. Theory 2, pp. 244–63. Atkinson, Anthony B. and Andrea Brandolini. 2001. that come from low rank has a well-devel- “Promise and Pitfalls in the Use of ‘Secondary’ Data oped biochemical basis (see for example Sets: Income Inequality in OECD Countries,” J. Robert Sapolsky 1993, and Bruce McEwen Econ. Lit. 39, pp. 771–99. Auster, Richard; Irving Leveson and Deborah 1998). Yet there is nothing that necessarily Sarachek. 1969. “The Production of Health: An Ex- implicates income, let alone income in- ploratory Study,” J. Human Res. 4, pp. 411–36. Article 3 2/21/03 4:20 PM Page 153

Deaton: Health, Inequality, and Ecomomic Development 153

Backlund, E.; Paul D. Sorlie and Norman J. Johnson. Center for Disease Control. 2002. National Vital Sta- 1999. “A Comparison of the Relationship of Educa- tistics Report, 50:6, March. tion and Income with Mortality: the National Longi- Chandra, Amitabh and Jonathan Skinner. 2002. “Geog- tudinal Mortality Study,” Soc. Sci. Med.. 49, pp. raphy and Racial Health Disparities,” econ. dept., 1373–84. Dartmouth College, processed. Bardhan, Pranab; Samuel Bowles and Herbert Gintis. Chattopadhyay, Raghabendra and Esther Duflo. 2001. 2000. “Wealth Inequality, Wealth Constraints, and “Women’s Leadership and Policy Decisions: Evi- Economic Performance,” in Handbook of Income dence from a Nationwide Randomized Experiment Distribution. Anthony B. Atkinson and François in India,” I.I.M. Calcutta and econ. dept. MIT. Bourguignon, eds. Vol. 1, Amsterdam: North-Hol- Chiang, Tung-liang. 1999. “Economic Transition and land, pp. 541–603. Changing Relation between Income Inequality and Barker, David J. P. 1994. Mothers, Babies, and Disease Mortality in Taiwan: Regression Analysis,” Brit. in Later Life. London: BMJ Publishing. Med. J. 319, pp. 1162–65. ———. 1995. “Fetal Origins of Coronary Heart Dis- Cohen, Sheldon; S. Line, S. B. Manuck, B. S. Rabin, ease,” Brit. Med. J. 311, pp. 117–4. E. R. Heise and J. R. Kaplan. “Chronic Social Stress, Baumbusch, Andrew P. 1995. “Income Inequality and Social Status, and Susceptibility to Upper Respira- Infant Mortality,” Princeton U., junior independent tory Infections in Nonhuman Primates,” Psychosom. work. Med. 59:3, pp. 213–21. Benabou, Roland. 1996. “Inequality and Growth,” in Cohen, Sheldon. 1999. “Social Status and Upper Res- NBER Macroeconomics Annual. Ben Bernanke and piratory Infection,” in Socioeconomic Status and Julio Rotemberg, eds. Cambridge: MIT Press. Health in Industrialized Nations: Social, Psychologi- Ben-Shlomo, Yoav; Ian R. White and Michael Marmot. cal, and Biological Pathways. Nancy Adler et al., 1996. “Does the Variation in Socio-Economic Char- eds. Annals NY Academy Science 896, pp. 246–53. acteristics of an Area Affect Mortality?” Brit. Med. J. Daly, Mary C.; Greg J. Duncan, George A. Kaplan and 312, pp. 1013–14. John W. Lynch. 1998. “Macro to Micro Links in the Benzeval, Michaela and Ken Judge. 2001. “Income and Relation Between Income Inequality and Mortality,” Health: The Time Dimension,” Soc. Sci. Med. 52, Milbank Quart. 76, pp. 315–39. pp. 1371–90. Dasgupta, Partha. 1993. An Inquiry into Well-being Benzeval, Michaela; Ken Judge and Sue Shouls. 2001. and Destitution. Oxford: Clarendon Press. “Understanding the Relationship Between Income Dasgupta, Partha and Debraj Ray. 1986. “Inequality as and Health: How Much Can Be Gleaned from a Determinant of Malnutrition and Unemployment: Cross-Sectional Data?” Soc. Policy Admin. 35, pp. theory,” Econ. J. 96, pp. 1011–34. 376–96. ———. 1987. “Inequality as a Determinant of Malnutrition Binswanger, Hans and Mark Rosenzweig. 1984. Con- and Unemployment: Policy,” Econ. J. 97, pp. 177–88. tractual Arrangements, Employment, and Wages in Deaton, Angus. 2001a. “Inequalities in Income and In- Rural South Asia. New Haven: Yale U. Press. equalities in Health,” in The Causes and Conse- Bliss, Christopher J. and Nicholas H. Stern. 1978. quences of Increasing Inequality. Finis Welch, ed. “Productivity, Wages, and Nutrition,” J. Devel. Econ. Chicago: U. Chicago Press, pp. 285–313. 5, pp. 331–98. ———. 2001b. “Relative Deprivation, Inequality, and Blakely, Tony A.; Kimberley Lochner and Ichiro Mortality,” Princeton U. Research Program Devel- Kawachi. 2002. “Metropolitan Area Income In- opment Studies and Center for Health and Wellbe- equality and Self-Rated Health—A Multilevel ing (NBER work. paper 8099). Study,” Soc. Sci. Med. 54, pp. 65–77. ———. 2002. “Policy Implications of the Gradient of Bowles, Samuel and Herbert Gintis. 1998. “Is Equality Health and Wealth,” Health Affairs 21, pp. 13–30. Passé?” Boston Rev. 23, pp. 4–35. Deaton, Angus and Christina Paxson. 2001a. “Mortal- Brodish, Paul; Mark Massing and Herman Tyroler. ity, Education, Income, and Inequality among 2000. “Income Inequality and All-Cause Mortality in American Cohorts,” in Themes in the Economics of the 100 Counties of North Carolina,” South. Med. J. Aging. David Wise, ed. Chicago U. Press for NBER 93, pp. 386–91. (NBER work. paper 7140) Brunner, Eric and Michael Marmot. 1999. “Social Or- ———. 2001b. “Mortality, Income, and Income In- ganization, Stress, and Health,” in Social Determi- equality among British and American Cohorts,” nants of Health. Michael Marmot and Richard G. Center for Health and Wellbeing, Princeton U., Wilkinson, eds. Oxford: Oxford U. Press. processed. Card, David and Alan B. Krueger. 1993. “Trends in Deaton, Angus and Darren Lubotsky. 2003. “Mortality, Relative Black-White Earnings Revisited,” Amer. Inequality and Race in American Cities and States,” Econ. Rev., Papers and Proceedings 83, pp. Soc. Sci. Med., forthcoming. 85–91. Deininger, Klaus and Lyn Squire. 1996. “A New Data Case, Anne. 2001. “Does Money Protect Health Sta- Set Measuring Income Inequality,” World Bank tus? Evidence from South African Pensions,” NBER Econ. Rev. 10, pp. 565–91. work. paper 8496, Cambridge, MA. Doblhammer, Gabriele. 2002. “Differences in Life- ———. 2002. “Health, Income, and Economic Devel- span by Month of Birth for the United States: The opment,” Proceedings of the World Bank Conference Impact of Early Life Events and Conditions on Late on Development Economics, 2001/2002, World Life Mortality,” Max Planck Institute for Demo- Bank, pp. 221–41. graphic Research, work. paper 2002–19. Article 3 2/21/03 4:20 PM Page 154

154 Journal of Economic Literature, Vol. XLI (March 2003)

Doblhammer, Gabriele and James W. Vaupel. 2001. Flegg, A.T. 1982. “Inequality of Income, Illiteracy and “Lifespan Depends on Month of Birth,” Proceedings Medical Care as Determinants of Infant Mortality in Nat. Academy Sci. 98, pp. 2934–9. Developing Countries,” Population Stud. 36, pp. Donohue, John J. and James J. Heckman. 1991. “Con- 441–58. tinuous Versus Episodic Change: The Impact of Af- Fogel, Robert W. 1994 “Economic Growth, Population firmative Action and Civil Rights Policy on the Eco- Theory and Physiology: the bearing of long-term nomic Status of Blacks,” J. Econ. Lit. 29, pp. processes on the making of economic policy,” Amer. 1603–44. Econ. Rev. 84, pp. 369–95. Drèze, Jean and Reetika Khera. 2000. “Crime, Gender, ———. 1997. “New Findings on Secular Trends in and Society in India: Insights from Homicide Data,” Nutrition and Mortality: some implications for pop- Population Devel. Rev. 26, pp. 335–52. ulation theory,” in Handbook of Population and Drèze, Jean and Mamta Murthi. 2001. “Fertility, Edu- Family Economics. Mark Rosenzweig and Oded cation, and Development: Evidence from India,” Stark, eds. Vol 1A, Amsterdam: Elsevier. Population Devel. Rev. 27, pp. 33–63. Foster, James and Amartya K. Sen. 1997. “On Eco- Easterlin, Richard A. 1974. “Does Economic Growth nomic Inequality After a Quarter of a Century,” an- Improve the Human Lot? Some Empirical Evi- nex to Sen, Amartya K., On Economic Inequality. dence,” in Nations and Households in Economic Oxford: Clarendon Press. Growth. P. A. David and W. R. Levin, eds. Stanford: Fuchs, Victor. 1993. “Poverty and Health: asking the Stanford U. Press, pp. 98–125. Right Questions,” in Medical Care and the Health of ———. 1996. Growth Triumphant: the Twenty-first the Poor. David E. Rogers and Eli Ginzberg, eds. Century in Historical Perspective. Ann Arbor: U. Boulder: Westview Press. Michigan Press. Fuchs, Victor; Mark McClellan and Jonathan Skinner. Ehrlich, Isaac. 1973. “Participation in Illegitimate Ac- 2001. “Area Differences in Utilization of Medical tivities: A Theoretical and Empirical Investigation,” Care and Mortality among U.S. Elderly,” NBER J. Polit. Econ. 81, pp. 521–65. work. paper 8628, Cambridge, MA. Eibner, Christine. 2001. “Relative Deprivation, Poor Franzini, Luiza; John Ribble and William Spears. 2001. Health Habits, and Mortality,” Princeton, Center for “The Effects of Income Inequality and Income Health and Wellbeing work. Paper, May (http:// Level on Mortality Vary by Population Size in Texas www.wws.princeton.edu/~chw/papers.html) Counties,” J. Health Soc. Behav. 42, pp. 373–87. Ellen, Ingrid Gould. 2000. “Is Segregation Bad for Gaarder, Marie M. 2001. “Can Population Characteris- Your Health? The Case of Low Birth Weight,” tics Account for the Variation in Health Impacts of Brookings-Wharton Pap. Urban Affairs pp. 203–38. Air Pollution? A Meta Analysis of PM10–Mortality Elo, Irma and Samuel H. Preston. 1996. “Educational Studies,” University College, London, processed. Differentials in Mortality: United States, 1979–85,” Garber, Alan M. 1989. “Pursuing the Links between Soc. Sci. Med.42, pp. 47–57. Socioeconomic Status and Health: critique, policy Erdal, David, and Andrew Whiten. 1996. “Egalitarian- implications, and directions for future research,” in ism and Machiavellian Intelligence in Human Evo- Pathways to Health: The Role of Social Factors. John lution,” in Modelling the Early Human Mind. Paul P. Bunker, Deanna S. Gombey and Barbara Kehrer, Mellars and Kathleen Gibson, eds. McDonald Insti- eds. Menlo Park, CA. Kaiser Family Foundation, pp. tute Monographs, pp. 139–50. 271–315. Fajnzylber, Pablo; Daniel Lederman and Norman Galor, Oded and Omer Moav. 2000. “From Physical to Loayza. 2000. “Crime and Victimization: An Eco- Human Capital Accumulation: inequality in the nomic Perspective,” Economía 1, pp. 219–78. process of development,” econ. dept. Brown U. and Farmer, Paul. 1999. Infections and Inequalities: the MIT, processed. modern plagues. Berkeley: U. California Press. Gerdtham, Ulf-G. and Magnus Johannesson. 2001. Fehr, Ernst, and Klaus M. Schmidt. 1999. “A Theory of “Absolute Income, Relative Income, Income In- Fairness, Competition, and Cooperation,” Quart. J. equality and Mortality,” Stockholm School Econ., Econ. 114, pp. 817–68. processed. Feinstein, Jonathan S. 1993. “The Relationship Be- Gertler, Paul and Jonathan Gruber. 2002. “Insuring tween Socioeconomic Status and Health: a review of Consumption against Illness,” Amer. Econ. Rev. 92, the literature,” Milbank Quart.71, pp. 279–322 pp. 51–70. Filmer, Deon, and Lant Pritchett. 1999a. “The Impact Gersovitz, Mark. 1983. “Savings and Nutrition at Low of Public Spending on health: does money matter?” Incomes,” J. Polit. Econ. 91, pp. 841–55. Soc. Sci. Med. 49, pp. 1309–23. Goldin, Claudia and Lawrence F. Katz. 2001. “De- ———. 1999b. “The Effect of Household Wealth on creasing (and Then Increasing) Inequality in Amer- Educational Attainment: evidence from 35 coun- ica: A Tale of Two Half-Centuries,” in The Causes tries,” Population Devel. Rev. 25, pp. 85–120. and Consequences of Increasing Inequality Finis Fiscella, Kevin, and Peter Franks. 1997. “Poverty or Welch, ed. Chicago: U. Chicago Press, pp. 37–82. Income Inequality as a Predictor of Mortality: lon- Gottshcalk, Peter and Timothy Smeeding. 2000. “Em- gitudinal cohort study,” Brit. Med. J. 314, pp. pirical Evidence on Income Inequality in Industrial 1724–28. Countries,” in Handbook of Income Distribution. ———. 2000. “Individual Income, Income Inequality, Anthony B. Atkinson and François Bourguignon, Health and Mortality: what are the relationships?” eds. Vol. 1, Amsterdam: North-Holland, pp. Health Services Res. 35, pp. 307–18. 261–307. Article 3 2/21/03 4:20 PM Page 155

Deaton: Health, Inequality, and Ecomomic Development 155

Gravelle, Hugh. 1998. “How Much of the Relation be- ical Health: Cross Sectional National Survey,” Brit. tween Population Mortality and Unequal Distribu- Med. J. 321, pp. 1311–15. tion of Income Is a Statistical Artefact?” Brit. Med. J. Kaplan, George A.; Elsie R. Pamuk, John W. Lynch, 316, pp. 382–85. Richard. D. Cohen and Jennifer L. Balfour. 1996. Gravelle, Hugh; John Wildman and Matthew Sutton. “Inequality in Income and Mortality in the United 2002. “Income, Income Inequality and Health: States: Analysis of Mortality and Potential Path- What Can We Learn from the Aggregate Data?” ways,” Brit. Med. J. 312, pp. 999–1003. Soc. Sci. Med. 54, pp. 577–89. Kaplan, George A. and John W. Lynch. 2001. “Editor- Grossman, Michael. 1972. “On the Concept of Health ial: Is Economic Policy Health Policy?” Amer. J. Capital and the Demand for Health,” J. Polit. Econ. Public Health 91, pp. 351. 80, pp. 223–55. Kawachi, Ichiro, and Tony A Blakely. 2001. “When ———. 1975. “The Correlation between Health and Economists and Epidemiologists Disagree,” J. Schooling,” in Household Production and Consump- Health. Politics, Policy, Law 26, pp. 533–41. tion. Nestor Terleckyj, ed. NY: Columbia U. Press Kawachi, Ichiro and Bruce P. Kennedy. 1997. “The Re- for NBER, pp. 147–211. lationship of Income Inequality to Mortality: Does ———. 2000. “The Human Capital Model,” in Hand- the Choice of Indicator Matter?” Soc. Sci. Med. 45, book of Health Economics. Anthony J. Culyer and pp. 1121–27. Joseph P. Newhouse, eds. Vol 1A, Amsterdam: Else- Kawachi, Ichiro; Bruce P. Kennedy and Deborah Pro- vier, pp. 347–408. throw-Stith. 1997. “Social Capital, Income Inequal- Groves, Robert M. and Mick P. Couper. 1998. Nonre- ity, and Mortality,” Amer. J. Public Health 87, pp. sponse in Household Interview Surveys. NY: Wiley. 1491–98. Gwatkin, Davidson R. 2000. “Health Inequalities and Kawachi, Ichiro; Bruce P. Kennedy and Richard G. the Health of the Poor: What Do We Know? What Wilkinson. 1999. The Society and Population Health Can We Do?” Bull. World Health Org. 78, pp. 3–18. Reader: Volume 1: Income Inequality and Health. Haines, Michael R. and Hallie J. Kintner. 2000. “The NY: New Press. Mortality Transition in Germany, 1860–1935,” Hist. Kennedy, Bruce P.; Ichiro Kawachi and Deborah Pro- Methods 33, pp. 83–104. throw-Stith. 1996a. “Income Distribution and Mor- Hales, Simon; Philippa Howden-Chapman, Clare tality: Cross-Sectional Ecological Study of the Robin Salmond, Alistair Woodward and Johan Macken- Hood Index in the United States,” Brit. Med. J. 312, bach. 1999. “National Infant Mortality Rates in Re- pp. 1004–07. lation to Gross National Product and Distribution of Kennedy, Bruce P.; Ichiro Kawachi, Roberta Glass and Income,” Lancet 354, pp. 2047. Deborah Prothrow-Stith. 1998. “Income Distribu- Hammond, Richard J. 1951. Food. History of the Sec- tion, Socioeconomic Status and Self-Rated Health in ond World War, United Kingdom Civil Series, Lon- the United States,” Brit. Med. J. 317, pp. 917–21. don, H. M. Stationery Office. Kennedy, Bruce P.; Ichiro Kawachi and Deborah Pro- Hsieh, C. C. and M. D. Pugh. 1993. “Poverty, Inequal- throw-Smith. 1996b. “Important Correction,” Brit. ity, and Violent Crime: A Meta-Analysis of Recent Med. J. 312, pp. 1194. Aggregate Data Studies,” Criminal Justice Rev. 18, Kitigawa, Evelyn and Philip Hauser. 1973. Differen- pp. 182–202. tial Mortality in the United States: A Study in So- Idler, Ellen L. and Yael Benyamini. 1997. “Self-rated cioeconomic Epidemiology. Cambridge: Harvard Health and Mortality: A Review of Twenty-Seven U. Press. Community Studies,” J. Health Soc. Behav. 38, Pp. Krieger, Nancy. 2001. “Theories for Social Epidemiol- 21–37. ogy in the 21st Century: An Ecosocial Perspective,” Jain, Shail B. 1975. Size Distribution of Income: A Int. J. Epidemiology 30, pp. 668–77. Compilation of Data. Washington, DC: World Bank. Kunst, Anton; Feikje Groenhof, Johan P Mackenbach Jones, Arthur F. and Daniel H. Weinberg. 2000. “The and EU Working Group on Socioeconomic Inequal- Changing Shape of the Nation’s Income Distribu- ities in Health. 1998. “Occupational Class and Cause tion,” P60–204, Washington, DC, Bureau of Census. Specific Mortality in Middle Aged Men in 11 Euro- http://www.census.gov/ftp/pub/hhes/www/p60204.ht pean Countries: Comparison of Population Based ml Studies,” Brit. Med. J. 316, pp. 1636–42. Judge, Ken. 1995. “Income Distribution and Life Ex- Lantz, Paula; James House, J. Lepowski, David pectancy: A Critical Appraisal,” Brit. Med. J. 311, pp. Williams, R. Mero and J. Chen. 1998. “Socioeco- 1282–85. nomic Factors, Health Behaviors, and Mortality,” J. Judge, Ken and Iain Patterson. 2001. “Poverty, Income Amer. Med. Assoc. 279, pp. 1703–708. Inequality, and Health,” work. paper No. 01/29, Laveist, Thomas A. 1993. “Segregation, Poverty, and New Zealand Treasury. Empowerment: Health Consequences for African Judge, Ken; Jo-Ann Mulligan and Michaela Benzeval. Americans,” Milbank Quart. 71, pp. 41–64. 1997. “Income Inequality and Population Health,” Lynch, John; George Davey-Smith, Marianne Hille- Soc. Sci. Med. 46, pp. 567–79. meier, Mary Shaw, Trivellore Raghunathan and ———. 1998. “Reply to Richard Wilkinson,” Soc. Sci. George Kaplan. 2001. “Income Inequality, Psy- Med. 47, pp. 983–85. chosocial Environment and Health: Comparisons Kahn, Robert S.; Paul H. Wise, Bruce P. Kennedy and across Wealthy Nations,” Lancet 358, pp. 194–200. Ichiro Kawachi. 2000. “State Income Inequality, LeClere, Felicia B. and Mah-Jabeen Soobader. 2000. Household Income, and Maternal Mental and Phys- “The Effect of Income Inequality on the Health of Article 3 2/21/03 4:20 PM Page 156

156 Journal of Economic Literature, Vol. XLI (March 2003)

Selected US Demographic Groups,” Amer. J. Public Mealli, Fabrizia and Donald Rubin. 2003. “Commen- Health 90, pp. 1892–97. tary: Assumptions Allowing the Estimation of Direct Leibenstein, Harvey. 1957. Economic Backwardness Causal Effects,” J. Econometrics, 112, pp. 79–87. and Economic Growth. NY: Wiley. Mellor, Jennifer M. and Jeffrey Milyo. 2001. “Re- Lindert, Peter H. 2000. “Three Centuries of Inequality examining the Evidence of an Ecological Association in Britain and America,” in Handbook of Income between Income Inequality and Health,” J. Health, Distribution. Anthony B. Atkinson and François Politics, Policy, Law 26, pp. 487–522. Bourguignon, eds. Vol. 1. Amsterdam: North-Holland, ———. 2002. “Income Inequality and Individual pp. 167–216. Health: Evidence from the Current Population Sur- Lynch, John W.; George A. Kaplan, Elsie R. Pamuk, vey,” J. Human Res., 37, pp. 510–39. Richard D. Cohen, Katherine E. Heck, Jennifer L. Miller, Douglas. 2000. “Income Inequality and Mortal- Balfour and Irene H. Yen. 1998. “Income Inequal- ity in the US: Aggregate Data and Micro Relation- ity and Mortality in Metropolitan Areas of the ships,” Princeton U. econ. dept., Ph.D. thesis, ch. 1. United States,” Amer. J. Public Health 88, pp. Mirrlees, James. 1975. “A Pure Theory of Underdevel- 1074–80. oped Economies,” in Agriculture in Development Lochner, Kim; Elsie R. Pamuk, Diane Makuc, Bruce P. Theory. Lloyd Reynolds, ed. New Haven, CT: Yale Kennedy and Ichiro Kawachi. 2001. “State-level In- U. Press. come Inequality and Individual Mortality Risk: A Mujica, Oscar J.; Carlos Castillo-Salgado, Jorge Ba- Prospective Multilevel Study,” Amer. J. Public callao, Enrique Loyola-Elizondo, Maria Christina Health, 91:3, pp. 385–91. Schneider, Manuel Vidaurre and George A. O. Al- Macintyre, Sally. 1997. “The Black Report and Beyond: leyne. 2000. “Income Inequality and Suicide Risk in what are the issues?” Soc. Sci. Med. 44, pp. 723–45. the Americas,” Pan-American Health Org. Marmot, Michael G. 1994. “Social Differences in Murray, Christopher J.L. 1996. “Epidemiologic and Health Within and Between Populations,” Daedalus Morbidity Transitions in India,” in Health, Poverty, 123, pp. 197–216. and Development in India. Monica das Gupta, Lin- ———. 2001. “Editorial: inequalities in health,” New coln C. Chen and T. N. Krishnan, eds. Delhi: Oxford England J. Med. 345, pp. 134–6. U. Press. ———. 2002. “The Influence of Income on Health: Murray, Christopher J.L.; M.C. Michaud, M. Views of an Epidemiologist,” Health Affairs 21, pp. McKenna and J. Marks. 1998. “US Patterns of Mor- 31–46. tality by County and Race, 1965–94,” Cambridge, Marmot, Michael and Martin Bobak. 2000. “Interna- MA. Harvard Center Population Development tional Comparators and Poverty and Health in Eu- Studies. rope,” Brit. Med. J. 321, pp. 1124–28. Murthi, Mamta; Anne-Catherine Guio and Jean Marmot, Michael G.; H. Bosma, H. Hemingway, Eric Drèze. 1995. “Mortality, Fertility, and Gender Bias Brunner and S. Stansfeld. 1997. “Contribution of in India: A District Level Analysis,” Population De- Job Control and Other Risk Factors to Social Varia- vel. Rev. 21, pp. 745–82. tions in Coronory Heart Disease Incidence,” Lancet Newhouse, Joseph P. and Lindy Friedlander. 1980. 350:9073, pp. 235–9. “The Relationship between Medical Resources and Marmot, Michael G.; Martin Shipley and Geoffrey Measures of Health: Some Additional Evidence,” J. Rose. 1984. “Inequalities in Death—Specific Expla- Human Res. 15, pp. 200–18. nations of a General Pattern,” Lancet 1:8384, pp. Nielsen, Jacob Arendt. 2002. “The Relationship Be- 1003–1006. tween Health, Education, and Income: Empirical Marmot, Michael G.; George Davey-Smith, S. Stans- Evidence,” U. Copenhagen dept. econ. feld, C. Patel, F. North, J. Head, I White, Eric Brun- Osler, Merete; Eva Prescott, Morten Grønbœk, Ulla ner and A. Feeney. 1991. “Health Inequalities Christensen, Pernille Due and Gerda Engholm. among British Civil Servants: The Whitehall II 2002. “Income Inequality, Individual Income, and study,” Lancet 337:8754, pp. 1387–93. Mortality in Danish Adults: Analysis of Pooled Data Massey, Douglas M. 1990. “American Apartheid: Seg- from Two Cohort Studies,” Brit. Med. J. 324, pp. regation and the Making of the Underclass,” Amer. J. 1–4. Sociology 96, pp. 329–57. Over, Mead. 1998. “The Effects of Societal Variables McEwen, Bruce S. 1998. “Protective and Damaging on Urban Rates of HIV Infection in Developing Effects of Stress Mediators: Allostatis and Allostatic Countries: An Exploratory Analysis,” in Confronting load,” New England J. Med. 338, pp. 171–79. AIDS: Evidence from the Developing World. Euro- McIsaac, Sandra J. and Richard G. Wilkinson. 1997. pean Commission, ch. 2. http://europa.eu.int/comm/ “Income Distribution and Cause-specific Mortality,” development/aids/limelette/index.htm Eur. J. Public Health 7, pp. 45–53. Pampel, Fred C. and Vijayan K. Pillai. 1986. “Pat- McKeown, Thomas. 1976. The Modern Rise of Popula- terns and Determinants of Infant Mortality in De- tion. NY: Academic Press. veloped Nations 1950–75,” Demography 23, pp. ———. 1979. The Role of Medicine: Dream, Mirage, 525–41. or Nemesis. Princeton: Princeton U. Press. Paukert, Felix. 1973. “Income Distribution at Differ- McLaughlin, Diane K. and C. Shannon Stokes. 2002. ent Levels of Development: A Survey of Evidence,” “Income Inequality and Mortality in US Counties: Int. Lab. Rev. 108, pp. 97–125. Does Minority Racial Concentration Matter?” Amer. Peña, Rodolfo; Stig Wall and Lars- ke Persson. 2000. J. Public Health 92, pp. 99–104. “The Effect of Poverty, Social Inequity, and Maternal Article 3 2/21/03 4:20 PM Page 157

Deaton: Health, Inequality, and Ecomomic Development 157

Education on Infant Mortality in Nicaragua, ———. 1998. Why Don’t Zebras Get Ulcers? An Up- 1988–1993,” Amer. J. Public Health 90, pp. 64–69. dated Guide to Stress, Stress-Related Diseases, and Preston, Samuel H. 1975. “The Changing Relation be- Coping. NY: Freeman. tween Mortality and Level of Economic Develop- Sawyer, Malcolm. 1976. “Income Distribution in ment,” Population Stud. 29, pp. 231–48. OECD Countries,” OECD Economic Outlook, Oc- ———. 1996 “American Longevity: Past, Present, and casional Studies. Paris, OECD. Future,” Syracuse U., Maxwell School, Center Pol- Schofield, Roger and David Reher. 1991. “The Decline icy Research, policy brief 7/1996. of Mortality in Europe,” in The decline of mortality Preston, Samuel H. and Michael R. Haines. 1991. Fa- in Europe. Roger Schofield, David Reher and Alain tal Years: Child Mortality in Late Nineteenth- Bideau, eds. Oxford: Clarendon Press, pp. 1–17. Century America. Princeton: Princeton U. Press. Sen, Amartya K. 1973. On Economic Inequality. Ox- Preston, Samuel H. and Paul Taubman. 1994. “Socio- ford: Clarendon Press. economic Differences in Adult Mortality and Health ———. 1999. Development as Freedom. NY: Knopf. Status,” in Demography of Aging. Linda G. Martin Shibuya, Kenji; Hideki Hashimoto and Eiji Yano. 2002. and Samuel H. Preston, eds. Washington, DC: Na- “Individual Income, Income Distribution, and Self- tional Academies Press, pp. 279–318. rated Health in Japan: Cross-Sectional Analysis of Pritchett, Lant and Lawrence H. Summers. 1996. Nationally Representative Sample,” Brit. Med. J. “Wealthier Is Healthier,” J. Human Res. 31, pp. 324, pp. 1–5. 841–68. Smedley, Brian D.; Adrienne Y. Stith and Alan R. Nel- Putnam, Robert D., with Robert Leonardi and Raf- son, eds. 2002. Unequal Treatment: Confronting faella Nanetti. 1993. Making Democracy Work: Racial and Ethnic Disparities in Health Care. Insti- Civic Traditions in Modern Italy. Princeton: Prince- tute of Medicine, Washington, DC: National Acad- ton U. Press. emy Press. Ray, Debraj. 1998. Development Economics. Prince- Smith, James P. 1999. “Healthy Bodies and Thick Wal- ton: Princeton U. Press. lets: The Dual Relation between Health and Eco- Regidor, Enrique; Pedro Navarro, Vicente Dominguez nomic Status,” J. Econ. Perspect. 13, pp. 145–66. and Carmen Rodriguez. 1997. “Inequalities in In- Soobader, Mah-Jabeen and Felicia LeClere. 1999. come and Long-term Disability in Spain: Analysis of “Aggregation and Measurement of Income Inequal- Recent Hypotheses Using Cross-Sectional Study ity: Effects on Morbidity,” Soc. Sci. Med. 48, pp. Based on Individual Data,” Brit. Med. J. 319, pp. 733–44. 898–902. Stanistreet, Debbie; Alex Scott-Samuel and Mark A. Roberts, Kevin W. S. 1977. “Voting Over Income Tax Bellis. 1999. “Income Inequality and Mortality in Schemes,” J. Public Econ. 8, pp. 329–40. England,” J. Public Health Med. 21, pp. 205–207. Robson, Arthur. 2002. “Evolution and Human Na- Steckel, Richard. 1995. “Stature and Standard of Liv- ture,” J. Econ. Perspect. 16, pp. 89–106. ing,” J. Econ. Lit. 33, pp. 1903–40. Rodgers, G. B. 1979 “Income and Inequality as Deter- Stiglitz, Joseph E. 1976. “The Efficiency Wage Hy- minants of Mortality: An International Cross-section pothesis, Surplus Labor, and the Distribution of In- Analysis,” Population Stud. 33, pp. 343–51. come in L.D.C.s,” Oxford Econ. Papers 28, pp. Rogers, Richard G.; Robert A. Hummer and Charles 185–207. B. Nam. 2000. Living and Dying in the USA: Behav- Strauss, John and Duncan Thomas. 1998. “Health, Nu- ioral, Health, and Social Differentials in Adult Mor- trition, and Economic Development,” J. Econ. Lit. tality. NY: Academic Press. 36:2, pp. 766–817. Rogot, Eugebe; Paul D. Sorlie, Norman J. Johnson and Sturm, Roland and Carole Roan Gresenz. 2002. “Rela- Catherine Schmitt, eds. 1992. A Mortality Study of tions of Income Inequality and Family Income to 1.3 Million Persons by Demographic, Social, and Chronic Medical Conditions and Mental Health Economic Factors: 1979–1985 Follow-up. Bethesda: Disorders: National Survey,” Brit. Med. J. 324, pp. NIH. 1–5. Ross, Nancy A.; Michael C. Wolfson, James R. Dunn, Subramanian, Shankar and Angus Deaton. 1996. “The Jean-Marie Berthelot, George A. Kaplan and John Demand for Food and Calories,” J. Polit. Econ. 104, W. Lynch. 2000. “Relationship between Income In- pp. 133–62. equality and Mortality in Canada and in the United Swamy, Anand V. 1997. “A Simple Test of the Nutri- States: Cross Sectional Assessment Using Census tion-based Efficiency Wage Model,” J. Devel. Econ. Data and Vital Statistics,” Brit. Med. J. 319, pp. 989. 53, pp. 85–98. Rosenzweig, Mark. 1988. “Labor Markets in Low In- Sydenstricker, Edgar and Willford I. King. 1921. “The come Countries,” in Handbook of Development Eco- Measurement of the Relative Economic Status of nomics. Hollis Chenery and T. N. Srinivasan, eds. Families,” Quart. Pub. Amer. Statist. Assoc. 17, pp. Vol 1. Amsterdam: North-Holland, pp. 713–62. 842–57. Ruhm, Christopher. 2000. “Are Recessions Good for Szreter, Simon. 1988. “The Importance of Social Inter- Your Health?” Quart. J. Econ. 115, pp. 617–50. vention in Britain’s Mortality Decline c. 1850–1914: Sachs, Jeffrey D. 2001. “Tropical Underdevelopment,” A Reinterpretation of the Role of Public Health,” Cambridge, MA. NBER work. paper 8119. Soc. Hist. Med., 1, pp. 1–37. Sapolsky, Robert M. 1993. “Endocrinology Alfresco: Victora, Cesar G.; J. Patrick Vaughan, Fernando C. Psychoendocrine Studies of Wild Baboons,” Recent Barros, Anamaria C. Silva and Elaine Tomasi. 2000. Progress in Hormone Research 48, pp. 437–68. “Explaining Trends in Inequities, Evidence from Article 3 2/21/03 4:20 PM Page 158

158 Journal of Economic Literature, Vol. XLI (March 2003)

Brazilian Child Health Studies,” Lancet 356, pp. Wilkinson, Richard G. 1989. “Class Mortality Differen- 1093–98. tials, Income Distribution and Trends in Poverty Wagstaff, Adam. 2000. “Socioeconomic Inequalities in 1921–1981,” J. Soc. Policy 18, pp. 307–35. Child Mortality: Comparisons across Nine Develop- ———. 1992. “Income Distribution and Life- ing Countries,” Bull. World Health Org. 78, pp. Expectancy,” Brit. Med. J. 304, pp. 165–68. 19–29. ———. 1994. “The Epidemiological Transition: From ———. 2002. “Poverty and Health,” Bull. World Material Scarcity to Social Disadvantage?” Daedalus Health Org., 80, pp. 97–105. 123, pp. 61–77. Wagstaff, Adam and Eddy van Doorslaer. 2000. “In- ———. 1996. Unhealthy Societies: The Affliction of In- come Inequality and Health: What Does the Litera- equality. London: Routledge. ture Tell Us?” Annual Rev. Public Health 21, pp. ———. 1997. “Commentary: Income Inequality Sum- 543–67. marizes the Health Burden of Individual Relative Walberg, Peter; Martin McKee, Vladimir Shkolnikov, Deprivation,” Brit. Med. J. 314, pp. 1727–28. Laurent Chenet and David A. Leon. 1998. “Eco- ———. 1998. “Letter to the Editor,” Soc. Sci. Med. 47, nomic Change, Crime, and Mortality Crisis in Rus- pp. 411–12. sia: Regional Analysis,” Brit. Med. J. 317, pp. ———. 2000. Mind the Gap: Hierarchies, Health, and 312–8. Human Evolution. London: Weidenfeld and Nicol- Waldmann, Robert J. 1992. “Income Distribution and son. Infant Mortality,” Quart. J. Econ. 107, pp. Wolfson, Michael C.; George A. Kaplan, John W. Lynch, 1283–302. Nancy A. Ross and Eric Backlund. 1999. “Relation- Wennemo, Irene. 1993. “Infant Mortality, Public Pol- ship between Income Inequality and Mortality: Em- icy and Inequality—a comparison of 18 industrial- pirical Demonstration,” Brit. Med. J. 319, pp. 953–55. ized countries, 1950–1985,” Sociology Health Illness Yitzhaki, Shlomo. 1979. “Relative Deprivation and the 15, pp. 429–46. Gini Coefficient,” Quart. J. Econ. 93, pp. 321–4.