Income Inequality and Health: Evidence from Developed and Developing Countries

We assess the effect of income inequality on life expectancy by performing separate estimations for developed and developing countries. Our empirical analysis challenges the widely held view that inequality matters more for health in richer countries than for health in poorer countries. Employing panel cointegration and conventional panel regressions, we find that income inequality slightly increases life expectancy in developed countries. By contrast, the effect on life expectancy is significantly negative in developing countries. Even though the quantitative effects are small, the contrast between the two country groups proves to be robust to modifications in measurement, specification and methodological choices. JEL I14 C23


INTRODUCTION
Income disparities have widened in various developed and developing countries during the process of economic globalization. Critics have called for redistributive policy interventions, not only for reasons of fairness but also to avoid economic and social costs of wide income gaps. Impaired health could add considerably to such costs.
The earlier literature suggests that the case for health-related redistribution is particularly strong for developed countries. 1 According to Wilkinson (1996: 4), the distribution of income is "one of the most powerful influences on the health of whole populations in the developed world to have come to light." 2 Lynch et al. (1998Lynch et al. ( : 1074 reckoned that the loss of life from income inequality in the United States "is comparable to the combined loss of life from lung cancer, diabetes, motor vehicle crashes, human immunodeficiency virus (HIV) infection, suicide, and homicide in 1995." The so-called Whitehall studies on British civil servants showed "that, even among people who are not poor, there is a social gradient in mortality that runs from the bottom to the top of society" (Marmot, 2003: S10). 3 In contrast to developed countries, income inequality may play a secondary role in lowincome countries with pervasive absolute poverty. Under such conditions, it could be low income per se that matters most for health and mortality, rather than income relative to other peoples' incomes (Deaton, 2003). Importantly, this so-called absolute income or poverty hypothesis implies that previous findings on inequality and health in developed countries do not necessarily hold for developing countries. 4 We account for this possibility by performing separate estimations for two samples of developed and developing countries.
Empirical evidence on the health effects of income inequality continues to be scarce for developing countries, largely because of lacking data on income inequality for a sufficiently large sample and a sufficiently long period of time. Our analysis contributes to filling this important gap by drawing on a relatively new data set, the Standardized World Income Inequality Database (SWIID, 2013), developed by Solt (2009). This data set combines information from several sources, resulting in greater coverage and better comparability (for details, see Section 3.b). Furthermore, many previous studies share limitations that we attempt to overcome in the subsequent analysis.
Most importantly, the endogenous nature of income inequality is often acknowledged, but rarely addressed appropriately in the empirical analysis. The panel cointegration approach pursued in the following accounts for endogeneity concerns and is robust to omitted variables.
Before describing in more detail our methodological approach and the data employed (Section 3), we provide an overview of the analytical background in Section 2, focusing on the theoretical ambiguity of the relationship between income inequality and health outcomes. Section 4 presents the empirical results for our samples of developed and developing countries, and Section 5 concludes. We find that income inequality has a small, but robust and significantly positive impact on health outcomes in developed countries. In contrast, the effect on life expectancy is significantly negative in developing countries. While the quantitative effects are small, the striking contrast between the two country groups proves to be robust to modifications in measurement, specification and methodological choices.

ANALYTICAL BACKGROUND
Several lines of reasoning in the relevant literature suggest that a more equal distribution of income is associated with better average health outcomes such as longer life expectancy and lower mortality. Nevertheless, there is considerable theoretical ambiguity in various respects. Preston's (1975) finding of a non-linear relationship between life expectancy and average per-capita incomes across countries provided an important building bloc of the so-called absolute income hypothesis, which has also been coined the poverty hypothesis (e.g., Deaton, 2003). The most obvious explanation for this non-linearity is that it reflects diminishing returns to increases in income economics. 7 Wilkinson (1996;, its most prominent proponent, argues that the epidemiological transition from infectious diseases to chronic and degenerative diseases implies that the major reason for differences in mortality and health shifts from (absolute) material deprivation to (relative) social disadvantage. Social disadvantage is supposed to give rise to psychosocial stress and relative deprivation. Unequal societies are characterized, according to Wilkinson (2000: 4), by "much more stressful strategies of dominance, conflict and submission." At the same time, biologists have shown that chronic stress impairs health by permanently perturbing the physiologic balance (Sapolsky, 2004).
This reasoning, though plausible, does not necessarily imply impaired health due to income inequality. First, inequality may be closely linked with relative deprivation, "but there is little that suggests it is income inequality" (Deaton, 2003: 152). Second, rank matters and (upward) comparisons of one's own well-being with higher ranked individuals in relevant reference groups may be stressful. 8 However, deprivation and adverse health effects would be contained if inequality within specific reference groups was low compared to economy-wide inequality. Indeed, inequality appears to be lower in groups of people who have much in common, such as co-workers, friends, relatives, and neighbors . Furthermore, people typically belong to various reference groups and tend to reduce stress by deriving self-esteem from the reference group where their ranking is highest. 9 Third, while social subordination often involves stress and an increased risk of stress-related diseases, Sapolsky (2004: 397 and 408) concludes from surveying the relevant literature that there are "numerous" and "dramatic" exceptions to this profile. In unstable hierarchies, stress centers on the higher ranks as dominant individuals constantly need to defend their position against emerging competitors. 10 Income inequality could also impair health conditions by eroding social trust and affecting the political process of delivering public goods. It is widely agreed that mutual trust and social capital are associated with better health (e.g., Kawachi et al., 1997;d'Hombres et al., 2010;Ronconi et al., 2012). 11 Trust and social capital could also help contain violent crime that may have minor direct effects on mortality and life expectancy, but could have considerable second-order effects by creating chronic stress among potential victims . It is less obvious that adverse health effects of a disrupted social fabric can be traced back to income inequality. On the one hand, Alesina and La Ferrara (2002) find that a higher degree of income disparity is one factor among others eroding mutual trust among individuals in US localities. According to Leigh (2006), a negative effect of income inequality on trust can also be observed across countries. On the other hand, reverse causality from trust and social capital to inequality cannot be ruled out. Indeed, Knack (2002: 71) shows that social capital is "progressive, in the sense that it helps the poorer classes more than it helps the richer classes." This seems to suggest that inequality represents the intermediating variable through which social capital affects health, rather than being the ultimate cause of impaired health. 12 Ambiguity also prevails on whether income inequality is associated with less or more public spending on health. 13 As noted by , the Meltzer-Richard theorem predicts that an increase in the mean income, relative to the income of the median voter, increases the size of government (Meltzer and Richard, 1981). A wider gap between the poor and the rich would thus encourage redistribution. Saint-Paul and Verdier (1993) model redistribution through public spending on education, showing that more inequality is associated with higher spending on education since the median voter (being poorer than the mean) prefers a higher rate of taxation. The reasoning of Saint-Paul and Verdier would also apply to health-related spending. It cannot be ruled out, however, that more inequality reduces the support for public spending on health or education.
This could happen if poorer population segments participate less in the electoral process than richer population segments. The preferences of the poor may also be underrepresented because of the political clout of the rich elite and the associated pressure for lower taxes. 14 Finally, it is debatable whether the above noted superior-goods character of health care would strengthen the economic justification of the relative income hypothesis. According to Waldmann (1992), infant mortality could be expected to increase if the rich receive a higher income share and health care is a superior good. Waldmann (1992Waldmann ( : 1291 argues that the relative cost of health care would increase when the rich demand more medical services, leaving fewer medical resources for the poor. This reasoning rests on fairly restrictive assumptions. In public health systems with universal access, the poor may even benefit from living where income inequality is relatively high and where the demand of the rich results in better medical facilities (Miller and Paxson, 2006). The cost of medical services supplied at the request of the rich must not necessarily rise if the plausibly high fixed costs of sophisticated facilities are distributed among larger numbers of (rich and poor) users. Miller and Paxson (2006) make a similar argument with respect to the provision of health-related public goods such as stricter environmental regulations. Health conditions could improve with income inequality if the demand for environmental quality stems mainly from people whose income exceeds a certain threshold.
In summary, there is considerable theoretical ambiguity so that the health effects of income inequality are essentially an empirical issue. The subsequent panel cointegration analysis provides a particularly useful empirical approach to help clarify the causal links between income inequality and health outcomes. Clearly, cross-section analyses face problems to establish the direction of causality (Kawachi et al., 1997(Kawachi et al., : 1497. 15 More surprisingly perhaps, even recent panel studies do not systematically address causality issues. 16 Reverse causality is possible, or even likely, as ill health may widen income gaps in several ways (Borghesi and Vercelli, 2004;Deaton, 2003;. Measures that equalize health conditions across the population, e.g., clean water supply in relatively poor countries, are also likely to narrow income gaps. Better health enhances people's earning capacity by reducing absenteeism from work and improving productivity at work.
Health conditions within poor families affect the level of education and, thus, the income potential of their children. Income differences across countries could be reduced if health conditions improved in poorer countries through faster diffusion of superior health technology and drugs.
Hence, it appears essential to employ empirical methods that account for bidirectional causality.

EMPIRICAL MODELS AND DATA
Our objective is to examine the effect of income inequality on health in developed and developing countries using panel cointegration techniques as well as conventional regression techniques. In this section, we present the empirical models and discuss some econometric issues (Subsection a). Then, we describe the data, report descriptive statistics, and present some preliminary evidence (Subsection b).

(a) Empirical models and econometric issues
Following common practice in (panel) cointegration studies (see, for example, Pedroni, 2007;Herzer, 2008;Moscone and Tosetti, 2010), we consider a parsimonious model which includes only the two variables of empirical interest: inequality and health. Thus, the basic model takes the is reasonable to assume that LE it and Gini it are non-stationary integrated processes. If this assumption is correct, the linear combination of the two variables must be stationary, or, in the terminology of Engle and Granger (1987), LE it must be cointegrated with Gini it . Otherwise, there is no long-run relationship between life expectancy at birth and income inequality; Eqn.
(1) would in this case represent a spurious regression in the sense of Granger and Newbold (1974). Entorf (1997) and Kao (1999) demonstrate that the tendency for spuriously indicating a relationship may even be stronger in panel data regressions than in pure time-series regressions. Thus, the necessary conditions for our model to be a correct description of the data are that LE it and Gini it are nonstationary or, more specifically, integrated of the same order, I(1), and cointegrated.
A regression consisting of (non-stationary) cointegrated variables has the property of superconsistency such that coefficient estimates converge to the true parameter values at a faster rate than they do in standard regressions with stationary variables, namely rate T rather than T (Stock, 1987). The important point in this context is that the estimated cointegration coefficients are superconsistent even in the presence of temporal and/or contemporaneous correlation between the stationary error term, it ε , and the regressor(s) (Stock, 1987), implying that cointegration estimates are not biased by omitted stationary variables (see, for instance, Bonham & Cohen, 2001).
The fact that a regression consisting of cointegrated variables has a stationary error term also implies that no relevant non-stationary variables are omitted. Any omitted non-stationary variable that is part of the cointegrating relationship would become part of the error term, thereby producing non-stationary residuals, and thus leading to a failure to detect cointegration (see also Everaert, 2011).
If there is cointegration between a set of variables, then this stationary relationship also exists in extended variable space. In other words, the cointegration property is invariant to model extensions (see also Lütkepohl, 2007), which is in stark contrast to regression analysis where one new variable can alter the existing estimates dramatically (Juselius, 2006, p. 11). The important implication of finding cointegration is thus that no additional variables are required to account for the classical omitted variables problem. More specifically, the result for the long-run relationship between life expectancy and inequality would also hold if additional variables were included in the model (see also Juselius, 1996).
Of course, there are several other factors that may affect population health and/or income inequality. Therefore, adding further non-stationary variables to the model may, on the one hand, result in further cointegrating relationships. If, however, there is more than one cointegrating relationship, identifying restrictions are required to separate the cointegrating relationships.
Otherwise, multicollinearity problems may arise. On the other hand, adding further non-stationary variables to the regression model may result in spurious associations. More specifically, if a nonstationary variable that is not cointegrated with the other variables is added to the cointegrating regression, the error term will no longer be stationary. As a result, the coefficient of the added variable will not converge to zero, as one would expect of an irrelevant variable in a standard regression (Davidson, 1998).
These considerations justify a parsimonious model such as Eqn.
(1) (if cointegrated). All the same, we check the robustness of the results to the inclusion of additional control variables. More specifically, we follow Leigh and Jencks (2007) and include GDP per capita and GDP per capita squared.
The superconsistency of the cointegration estimation also implies that the potential endogeneity of the regressors should not affect the estimated long-run coefficients; the estimated long-run coefficients from reverse regressions should be approximately the inverse of each other due to the superconsistency (Engle & Granger, 1987). However, although the standard least-squares dummy variable (LSDV) estimator is superconsistent under panel cointegration, it suffers from a second-order asymptotic bias arising from serial correlation and endogeneity in finite samples. As a consequence, its t-ratio is not asymptotically standard normal. To deal with this problem, one has to employ an asymptotically efficient (cointegration) estimator. Examples of such estimators include panel versions of the dynamic OLS (DOLS) and fully modified ordinary least squares (FMOLS) methods. As shown by Wagner and Hlouskova (2010), the panel DOLS estimator of Mark and Sul (2003) outperforms other asymptotically efficient panel cointegration estimators in obtaining reliable long-run coefficients. Therefore, this DOLS estimator is our preferred estimator, but in the robustness section we also present results based on alternative estimation procedures.
The idea behind the DOLS estimator is to account for possible serial correlation and endogeneity of the regressors by augmenting the cointegrating regression (given by Eqn. (1)) with lead, lag, and current values of the first differences of the I(1) regressor(s). Accordingly, in our case, the DOLS regression is given by: where Δ is the difference operator (such that ) and k is the number of leads and lags. We use one lead and lag in the DOLS estimations to preserve degrees of freedom, as is common practice in the literature (see, for instance, Spilimbergo & Vamvakidis, 2003;Thorbecke & Smith, 2010;Herzer et al., 2012).
Another empirical issue is the likely cross-sectional dependence among the variables. Crosssectional dependence may be the result of a common business cycle and other common factors such as health shocks. Examples of such shocks that affect health in multiple countries at the same time might include major influenza epidemics, the spread of HIV/AIDS, the introduction of new vaccines, and the diffusion of antibiotics . To control for potential crosssectional dependence, we estimate the long-run effect of income inequality on health status using both the raw data and demeaned data; that is, in place of LE it and Gini it , we also use which is equivalent to including time dummies. Moreover, we use a battery of panel unit root and cointegration tests, including so-called second-generation panel unit root and cointegration methods that explicitly allow for cross-sectional dependence.
A potential disadvantage of panel cointegration methods is that they typically require balanced panel data over a sufficiently long time period. Continuous time series data on some alternative measures of inequality and health are not available for many countries over long periods of time. To check the robustness of our results to alternative measures of inequality and health, we are thus forced to use conventional panel methods. More specifically, we use data on the tuberculosis incidence rate and the income share of the bottom quintile to estimate a standard where Health it and Inequality it stand for the measures of health and inequality, a i are countryspecific fixed effects (as before), and λ t represents time dummies. As control variables, we include GDP per capita and GDP per capita squared following Leigh and Jencks (2007), as discussed above. The number of lags is set to k = 2 when the time period is sufficiently long (about 30 years); otherwise (when T is about 20 years) we use one lag, k = 1. The long-run effect of a change in inequality on health is given by As is well known, the dynamic fixed effects model may suffer from the so-called Nickell (1981) bias; that is, the correlation between the lagged dependent variable and the fixed effects may bias the coefficient on the lagged dependent variable toward zero. However, the bias becomes small when T is about 20 or more. Judson and Owen (1999) compare the performance of different estimators in terms of Nickel bias and recommend the LSDV estimator in unbalanced panels with T = 30. Bun and Kiviet (2006) examine the performance of several dynamic panel estimators in samples where both T and N are moderate or small and conclude that none of these estimators (including GMM and LSDV) dominates the others in terms of bias or mean squared error. We use the LSDV estimator given the relatively long time dimension of our data.

(b) Data and preliminary evidence
We estimate both Eqn.
(2) and Eqn. (4) for developed and developing countries separately. conduct a meaningful panel data analysis. Therefore, we use (unadjusted) life expectancy as our main indicator of population health. As an alternative summary measure of population health, we use the infant mortality rate (per 1,000 live births), and as a specific measure of health status, we use the tuberculosis incidence rate (per 100,000 population). These two variables are also from the WDI 2013 online database.
The second limitation is that average life expectancy does not reveal the variation of health conditions within countries. The health conditions of poorer population segments tend to be worse than those of richer population segments-for economic reasons such as spending on health care and/or for reasons of social or psychic deprivation. There is evidence to this effect from selected case studies, typically for high-income countries, including the well-known Whitehall studies in the United Kingdom (e.g., Marmot, 2003;Anderson & Marmot, 2012) and for the United States (Singh & Siahpush, 2006). Comparable data do not exist for a panel analysis. However, it appears that the results achieved for average life expectancy ought to hold for life expectancy of poor population segments, if such data were available. This can be concluded at least tentatively when considering the prevalence of malnutrition among children under five as a marker of subsequent poor health and low life expectancy. The correlation between malnutrition, which can reasonably be assumed to be prevalent among the poor and be absent among the rich, and average life expectancy is strongly negative. 19 As discussed above, we include GDP per capita (in constant 2005 US dollars) and GDP per capita squared as additional explanatory variables. These data are also taken from the WDI.
As far as data on income inequality are concerned, several studies have used the Gini coefficient data set constructed by Deininger and Squire (1996). At least since the work of Atkinson and Brandolini (2001) it is well known, however, that the Deininger-Squire data suffer from deficiencies such as sparse coverage, problematic measurements, and the combination of diverse data types into a single data set, thus limiting the comparability, not only across countries but also over time. Many studies therefore rely on Gini data from the Luxembourg Income Study (LIS) database or the World Income Inequality Database (WIID). The major deficiency of all these sources is the lack of continuous and consistent inequality data over time. More generally, it should be noted that the Gini coefficient, though widely available and often used in empirical studies, is an imperfect measure on inequality. Most importantly, the Gini coefficient is not consistent with the welfare principle. 20 In this study, we utilize a data source that combines the strengths of the LIS and WIID data-the Standardized World Income Inequality Database (SWIID, 2013) developed by Solt (2009). 21 The SWIID combines information from the LIS and WIID data to create an improved data set with greater coverage than the LIS data and greater comparability than the WIID data. The logic behind the methodology underlying the SWIID can be summarized as follows (see also Morgan & Kelly, 2013). The synchronization process for the SWIID starts by utilizing inequality data from both the LIS and the WIID. The WIID data contain several country-years not available from the LIS and often includes inequality statistics based on multiple income concepts (with some including and others excluding various cash and/or in-kind transfers) for the same country-year. The SWIID synchronization process treats inequality as a latent variable, with data from the LIS and the WIID acting as imperfect indicators of the underlying concept. With knowledge from country-years in which the two data sets overlap, the SWIID uses inequality estimates from the strongly comparable LIS data set along with inequality estimates and information about the income concept represented in the WIID to adjust the WIID data such that it mimics the comparability of the LIS data. This yields data with greater comparability and more coverage than any other available data set.
Although many of the more recent income inequality studies use the (gross) SWIID Gini coefficient (see, for example, Desbordes & Verardi, 2012;Cole, 2013;Morgan & Kelly, 2013), this index has the limitation that it is estimated, and estimates may be biased (for several reasons).
Therefore, we check the sensitivity of the cointegration estimates to the measure of inequality by using the top-decile income share data provided by Leigh (2007). 22 Leigh adjusts top incomes series from different studies to produce a comparable data set. However, these data are available only for a small number of high-income countries. For developing countries, we use the income share of the bottom 20 percent of the population from the WDI. 23 As noted above, these data, as well as the data on the tuberculosis incidence rates, are unbalanced, and cannot be employed in cointegration analysis. Therefore, we estimate the long-run effects from the ARDL model given by Eqn. (4).
In our main analysis, we focus on the cointegrating relationship between LE it and Gini it . In order to apply panel cointegration techniques, we need a balanced panel data set. The construction of such a data set involves a trade-off between the time span and number of countries in the sample.
For the sample of developed countries, we select all high-income countries for which complete time-series data are available over the period 1976-2010-the longest time period with complete data for a reasonably large number of high-income countries according to World Bank (1995) classification.  In Tables 1 and 2 we also report the sample means of the variables used in the analysis, along with the minimum and maximum values of the data. As expected, life expectancy in developed countries is, on average, significantly higher than in developed countries, while income inequality is lower in developed countries compared to developing countries.  Rogers, 1979;Wilkinson, 1992;Waldman, 1992). The results with country and time fixed effects are also in line with previous studies (see, for instance, Mellor & Milyo, 2001;Beckfield 2004, Leigh & Jencks, 2007: The coefficient on the inequality variable turns out to be insignificant in columns 4-6.

Finally, in
[ Table 3] Columns 3 and 6 show that when we add GDP per capita and GDP per capita squared to the basic specifications in columns 1 and 4, the coefficient on GDP per capita is positive and significant (as in column 2) and the coefficient on GDP per capita squared is negative and significant.
Accordingly, increases in GDP per capita are associated with increases in life expectancy, but the effects diminish as GDP per capita rises, which is consistent with the results of Preston (1975), Deaton (2003), and Leigh and Jencks (2007).

EMPIRICAL ANALYSIS
In this section, we estimate the long-run effect of income inequality on population health for developed and developing countries separately. We first analyze the effect for developed countries (Subsection a), as most previous studies have done (see, for instance, Wilkinson, 1992;Wennemo, 1993;Judge et al., 1998;Leigh & Jencks, 2007), and check the robustness of our results (Subsection b). Subsequently, we provide estimates of the long-run effect of inequality on health in developing countries (Subsection c).

(a) The long-run effect of inequality on health in developed countries
The pre-tests for unit roots and cointegration, reported in Appendix B, suggest that LE it and Gini it are non-stationary and cointegrated. This implies that there is a (non-spurious) long-run relationship between life expectancy at birth and the SWIID Gini coefficient. To estimate this relationship, we use the panel DOLS estimator suggested by Mark and Sul (2003). As discussed above, the DOLS estimator is superconsistent, asymptotically unbiased, and normally distributed, even in the presence of endogenous regressors. Table 4 presents the results of this estimation procedure both for the raw data (column 1) as well as for the data that have been demeaned over the cross-sectional dimension (column 2). The estimated coefficient based on the raw data is positive but significant only at the 10% level. When the demeaned data are used (to account for the problem of cross sectional dependence induced by common, unobservable factors), the coefficient becomes significant at the one percent level. This suggests that an increase in inequality is associated with an increase in life expectancy in highincome countries, which is in contrast to most previous studies.
[ Table 4] Given that the finding of a significantly positive relationship between inequality and health in high-income countries contradicts previous findings, we perform several robustness checks. First, we examine whether the positive relationship between inequality and health in developed countries is robust to alternative estimation techniques. A potential problem with the pooled results (in columns 1 and 2 of Table 4) could be that they are based on the implicit assumption of homogeneity of the long-run effects. While efficiency gains from the pooling of observations over the crosssectional units can be achieved when the individual slope coefficients are the same, pooled estimators may yield inconsistent and potentially misleading estimates of the sample mean of the individual coefficients when the true slope coefficients are heterogeneous. A comparative study by Baltagi and Griffin (1997) concludes that "the efficiency gains from pooling appear to more than offset the biases due to intercountry heterogeneities" (p. 317). Nonetheless, we allow the long-run coefficients to vary across countries by using the group-mean panel DOLS estimator suggested by Pedroni (2001). This estimator involves estimating separate DOLS regressions for each country and averaging the long-run coefficients, 1ˆ. The corresponding t-statistic is computed as the sum of the individual t-statistics (calculated using heteroskedasticity and autocorrelation-consistent standard errors) divided by the root of the number of cross-sectional units, . In addition, we use the pooled FMOLS estimator suggested by Phillips and Moon (1999). Like the time series FMOLS estimator, the panel FMOLS estimator incorporates a semi-parametric correction to the OLS estimator, which eliminates the second order bias induced by the endogeneity of the regressors. We report the results of these estimation methods in columns 3 and 4 of Table 4.
The results show a positive and significant effect of inequality on life expectancy.
Interestingly, the panel and group-mean DOLS estimators produce almost identical coefficients, suggesting that slope heterogeneity is not a serious problem in this sample. The FMOLS coefficient estimate in column 4 is somewhat smaller than the DOLS estimate in column 2, but still positive and significant at the one percent level. As shown by Wagner and Hlouskova (2010), these two estimators perform worse than the pooled DOLS estimator of Mark and Sul (2003). Therefore, we continue our robustness analysis using the pooled DOLS estimator (with the demeaned data).
To verify that the positive effect of inequality on health is not due to individual outliers, the DOLS regression is re-estimated excluding one country at a time from the sample. The sequentially estimated coefficients and their t-statistics are presented in Figure 1. They fluctuate between 0.033 (due to the exclusion of Ireland) and 0.043 (due to the exclusion of Australia) and are always significant at the one percent level, suggesting that the positive effect of inequality on health is not the result of individual outliers.
[ Figure 1] It is common practice in conventional panel studies to use time-averaged data to eliminate business cycle effects. However, as pointed out by Attanasio et al. (2000), annual data provide information that is lost when time-averaged observations are used. Moreover, is not obvious that averaging over fixed time intervals will effectively eliminate business cycle effects; the length of the interval over which averages are computed is arbitrary, and there is no guarantee that business cycles are cut in the right way, as their length varies over time and across countries. In addition, the use of time-averaged data decreases the number of observations, and hence statistical power.
Despite these concerns, we re-estimate the DOLS regression using five-year averages. The results of this estimation are reported in column 1 of Table 5. As can be seen, the estimate using five-year averages is close to the corresponding estimate based on annual data reported in Table 4, column 2. This is consistent with several studies showing that cointegration estimates are remarkably stable across frequencies (see, for instance, Chambers, 2001;Click & Plummer, 2005;Herzer, 2013).
In column 2 of Table 5, we estimate the coefficient on Gini it using a larger sample of 21 high-income countries from 1981-2005 (which is the period used in the next section to analyze the effect of inequality on health in developing countries). Once again, the estimated coefficient is positive and highly statistically significant.
As discussed in Section 3, the finding of cointegration implies that there are no missing trending variables and that therefore no additional variables are needed to produce unbiased estimates. Nevertheless, we check the robustness of our results to the inclusion of GDP per capita and GDP per capita squared. A potential problem with this strategy is that it can introduce collinearity among the stochastic regressors or between the additional variables and the individual time trends. To account for this problem, we present estimates with and without individual time trends. As can be seen from columns 3 and 4 of Table 5, the impact of inequality on health remains positive and statistically significant when we include GDP per capita and GDP per capita squared, regardless of whether or not individual time trends are used in the analysis (to represent technological change). While in column 3 the coefficients on GDP per capita and GDP per capita squared have unexpected signs, the signs of these coefficients in column 4 are as we expect: the coefficient on GDP per capita is positive and significant and the coefficient on GDP per capita squared is negative and significant (in the regression without deterministic time trends).
[ Table 5] Next, we examine whether the results are robust to alternative measures of inequality and mortality. Leigh and Jencks (2007) Table 6 presents the results of the DOLS regressions using these two different measures, labeled TopDecile and IMR, both separately and jointly. All estimates suggest that income inequality increases health in high-income countries.
[ Table 6] A potential problem with these estimates is that life expectancy and infant mortality are measures of mortality rather than morbidity. While the two are highly correlated (see Section 3.b), it is morbidity rather than mortality which should be affected most by inequality. In other words, morbidity is, by definition, a better measure of the health response to income inequality than mortality (Soobader & LeClere, 1999). Regrettably, summary measures of morbidity which account for all diseases (whether physical, psychosomatic or psychiatric) are not available to conduct a meaningful panel analysis. A specific measure of morbidity is the tuberculosis incidence rate, which is available for a sufficient number of countries over the period 1990-2011. Because the data are unbalanced, we do not apply the DOLS estimator, but estimate the long-run effect of inequality on the incidence of tuberculosis using the ARDL model. Table 7 presents the results with and without control variables. The estimated long-run coefficient on Gini is always negative and highly significant, suggesting that inequality decreases the incidence of tuberculosis. This corroborates the finding that inequality has a positive effect on population health in developed countries. 24 [Table 7] In summary, we find that income inequality has a positive and robust effect on population health in developed countries. In the next subsection, we investigate whether this result also holds for developing countries.

(c) The long-run effect of inequality on health in developing countries
In Table 8, we present DOLS estimates of the long-run effect of inequality on health in developing countries using life expectancy and infant mortality as measures of health outcomes; the measure of inequality is the Gini index. The estimated coefficient in column 1 is negative and highly significant, suggesting that inequality is negatively related to life expectancy. Column 2 reports a positive and significant coefficient, suggesting that inequality is positively associated with infant mortality. Thus, in contrast to the results for developed countries, the results for developing countries show that inequality harms health. This is consistent with the results of the few previous studies that have examined the effect of income inequality on mortality in a (sub)sample of developing countries (see, for instance, Rogers, 1997;Flegg, 1982;Waldman, 1992). To quantify the effect, we multiply the coefficient of Gini in column 1 with the average change in the Gini index. The result implies that the average loss of life expectancy due to inequality is about one day per year (-0.003×365 = -1.095).
[ Table 8 about here] We also report results based on the income share of the bottom quintile, labeled BottomQuintile. Balanced panel data on this measure of inequality are not available, which prevents us from using the DOLS procedure. Instead, we employ the ARDL model. The results in Table 9 show that the long-run coefficient on BottomQuintile is always positive and significant, indicating that inequality has negative consequences for health in developing countries. Finally, we used the ARDL model to perform estimations with the incidence of tuberculosis as a specific measure of morbidity as in Table 7 above for the sample of developed countries. 25 The Gini index enters with a positive coefficient in the estimations for developing countries with the incidence of tuberculosis as the dependent variable, instead of life expectancy. This is in contrast to the corresponding result for developed countries, even though the coefficient on the Gini index fails to reach statistical significance at conventional levels. In other words, we again find that the results for developed countries do not carry over to developing countries.
The finding that inequality has a positive effect on health in developed countries and a negative effect in developing countries has an interesting implication once it is taken into account that the income distribution in developing countries is more unequal than in developed economies.
Taken together, our findings suggest that the health-impairing effects of inequality are stronger in more unequal countries than in more equal countries. This is consistent with the so-called threshold hypothesis (Kondo et al., 2009), which posits that income inequality harms health only if income gaps are sufficiently wide.

CONCLUSIONS
The widely held belief that more unequal societies are less healthy is politically highly relevant. Calls for redistributive policy interventions in order to improve health and ensure longer life expectancy would be justified, particularly if wide income gaps represent a major aspect of inequality within countries. This provided the motivation to re-assess Wilkinson's (1996) verdict that the distribution of income is one of the most powerful determinants of the health of whole populations. While recent studies have increasingly doubted this verdict for developed countries, the scant evidence available so far for developing countries posed the important question of whether the experience of developed countries would also hold for lower-income countries. Our empirical analysis addressed these unresolved issues. In addition, we attempted to overcome several limitations of previous research by employing panel cointegration techniques, which allowed us to account for endogeneity concerns.
We found that income inequality has a significantly positive impact on population health in developed countries. Even though wider income gaps increase life expectancy only by a quantitatively small margin, this result proved to be robust to modifications in measurement, specification and methodological choices. It was surprising to find wider income gaps to cause slightly better health outcomes in developed countries, although some recent studies pointed into the same direction, notably Mellor and Milyo (2001) as well as Leigh and Jencks (2007). Concepts and insights from different disciplines such as psychology, political science and economics offer some tentative explanations. For instance, certain stress-related health risks may center on higher-income ranks if hierarchies are unstable and dominant individuals constantly need to defend their position. More inequality may be associated with higher government spending on health care, as the Meltzer-Richard theorem would predict. Health conditions, including for poorer population segments, may also improve in line with Miller and Paxson (2006) if the superior-goods character of medical and environmental services induces stronger demand for such services by richer people beyond a certain income threshold.
In contrast to developed countries, we found that people living in developing countries with wider income inequality have a significantly lower life expectancy than people living in developing countries with a more equal distribution of income. The health-impairing effect of income inequality in developing countries is quantitatively small, but fairly robust. The case for redistributive policy interventions to improve health and increase life expectancy thus appears to be considerably stronger for developing countries than for developed countries. This is in some conflict with predictions from the absolute income or poverty hypothesis, according to which health primarily depends on the incidence of poverty in low-income countries. Unfortunately, our findings suggest that progressive income taxation might be advisable for health reasons exactly where wide income gaps tend to be most difficult to redress via taxation -due to insufficient administrative capacity and political resistance of local elites. Consequently, the preferred policy response may still consist of targeted pro-poor interventions with regard to the provision of health services.
Improving the education of poor population segments could provide another indirect handle to tackle the health-impairing effects of income inequality.
Generally speaking, health policies in both developing and developed countries should take into account that income disparity is just one manifestation of inequality. Paraphrasing Deaton (2003: 152), inequality may be important for health, even though the quantitative impact of income inequality on health is rather small and working in opposite directions. As indicated above, interdisciplinary research could provide further insights into the links between different aspects of inequality and health conditions in developed and developing countries. In addition, deeper insights may be gained once persistent data constraints are relaxed. Continued efforts to collect information on life expectancy adjusted for morbidity and time spent in poor health are of particular importance in this regard. Furthermore, the measurement of health conditions should be refined in order to reveal differences within countries, notably between particularly poor and richer population segments in developing countries.  , Bangladesh, Belgium, Brazil, Bulgaria, Canada, Chile, China, Colombia, Costa Rica, Cote d`Ivoire, Denmark, Egypt, El Salvador, Estonia, Ethiopia, Fiji, Finland, France, Georgia, Germany, Greece, Guatemala, Hong Kong, Hungary, India, Indonesia, Ireland, Israel, Italy, Japan, Jordan, Kazakhstan, Kenya, Korea, Kyrgyz Republic, Latvia, Lithuania, Madagascar, Malawi, Malaysia, Mauritius, Mexico, Morocco, Nepal, Netherlands, New Zealand, Nigeria, Norway, Pakistan, Panama, Peru, Philippines, Poland, Portugal, Puerto Rico, Russian Federation, Sierra Leone, Singapore, South Africa, Spain, Sri Lanka, Sweden, Switzerland, Tajikistan, Tanzania, Thailand, Trinidad and Tobago, Tunisia, Turkey, Turkmenistan, Ukraine, United Kingdom, United States, Uruguay, Uzbekistan, Venezuela, and Zambia. The countries from left to right are: Argentina, Australia, Azerbaijan, Bangladesh, Belgium, Brazil, Bulgaria, Canada, Chile, China, Colombia, Costa Rica, Cote d`Ivoire, Denmark, Egypt, El Salvador, Estonia, Ethiopia, Fiji, Finland, France, Georgia, Germany, Greece, Guatemala, Hong Kong, Hungary, India, Indonesia, Ireland, Israel, Italy, Japan, Jordan, Kazakhstan, Kenya, Korea, Kyrgyz Republic, Latvia, Lithuania, Madagascar, Malawi, Malaysia, Mauritius, Mexico, Morocco, Nepal, Netherlands, New Zealand, Nigeria, Norway, Pakistan, Panama, Peru, Philippines, Poland, Portugal, Puerto Rico, Russian Federation, Sierra Leone, Singapore, South Africa, Spain, Sri Lanka, Sweden, Switzerland, Tajikistan, Tanzania, Thailand, Trinidad and Tobago, Tunisia, Turkey, Turkmenistan, Ukraine, United Kingdom, United States, Uruguay, Uzbekistan, Venezuela, and Zambia.

APPENDIX B. PANEL UNIT ROOT AND COINTEGRATION TESTS (a) Panel unit root tests
One of the most commonly employed tests for unit roots in panels is that of Im, Pesaran and Shin (2003), the IPS test. It tests the null hypothesis that all of the individuals of the panel have a unit root against the alternative that some fractions are (trend) stationary using the augmented Dickey-Fuller (ADF) regression for the ith cross-section unit here k i is the lag order, z it represents deterministic terms, such as fixed effects or fixed effects combined with individual time trends, and Δ is the first-difference operator. To test the unit root null hypothesis, 0 : , 2, …, N, against the alternative of (trend) stationarity, 0 : where NT t is the average of the N (=19) cross-sectional ADF t-statistics, and μ and ν are, respectively, the mean and variance of the average of the individual t-statistics, tabulated by Im et al. (2003).
However, the IPS test procedure assumes cross-sectional independence and can thus lead to spurious inferences if the errors, ε it , are not independent across i (for instance, due to common shocks or spillovers between countries). Therefore, we also employ the cross-sectionally augmented IPS test proposed by Pesaran (2007). This test is designed to filter out the cross-section dependency by augmenting the ADF regression with the cross-section averages of lagged levels and first differences of the individual series. Accordingly, the cross-sectionally augmented ADF (CADF) regression is given by where t x is the cross-section mean of The results of the two tests for the variables in levels and in first differences are reported in  First differences ΔLE Constant -6.00*** -2.41*** ΔGini Constant -7.30*** -3.01*** Note: Three lags were selected to adjust for autocorrelation. The IPS statistic is distributed as N(0, 1). The relevant five-(one-) percent critical value for the CIPS statistics is -2.71 (-2.85) with an intercept and a linear trend, and -2.20 (-2.36) with an intercept. *** denote significance at the one percent level.

(b) Panel cointegration tests
We use several panel cointegration tests to examine whether there is a long-run relationship between live expectancy at birth and income inequality. The first is the two-step residual-based procedure suggested by Pedroni (1999Pedroni ( , 2004, which can be intuitively described as follows. In the first step, the hypothesized cointegrating relationship where j is the respective panel or group statistic, and µ and ν are the expected mean and variance of the corresponding statistic, tabulated by Pedroni (1999).
However, standard panel cointegration tests such as those of Pedroni (1999Pedroni ( , 2004) assume cross-sectional independence and can have size distortions when this assumption is violated. To test for cointegration in the presence of possible cross-sectional dependence, we use a two-step residual-based procedure in the style of Holly et al. (2010). In the first step, we apply the common correlated effects (CCE) estimator of Pesaran (2006) , including an intercept. In doing so, we account for unobserved common factors that could be correlated with the observed regressors in both steps.
A drawback of residual-based (panel) cointegration tests is that they are generally not invariant to the normalization of the cointegrating regression. Therefore, we also use the Larsson et al. (2001) procedure, which is based on Johansen's (1988)  and then standardizing it as follows: overestimating the cointegrating rank, we additionally compute the standardized panel trace statistics based on small-sample corrected country-specific trace statistics. Specifically, we use the small-sample correction factor suggested by Reinsel and Ahn (1992) to adjust the individual trace statistics as follows: where k i is the lag length of the models used in the test.
The results of these tests are presented in Table A  4.39*** -0.78 Note: *** (**) indicate a rejection of the null hypothesis of no cointegration/no cointegrating vector at the one (five) percent level. The relevant one-percent critical value for the CIPS statistic is -2.36. All other test statistics are asymptotically normally distributed. The right tail of the normal distribution is used to reject the null hypothesis in the panel ν-statistic and the panel trace statistic, while the left tail is used for the other statistics. One lag was used to form the panel trace and the CIPS statistics. For all other statistics, the number of lags was determined by the Schwarz criterion with a maximum of seven lags. NOTES 1 Lynch et al. (2004) review almost 100 studies addressing the question of whether more unequal societies are less healthy. Informative reviews of the relevant literature are also presented by Judge et al. (1998), Wagstaff and van Doorslaer (2000), Deaton (2003), and Subramanian and Kawachi (2004). 2 In addition to Wilkinson's own extensive work, several prominent studies supported this view, including Rodgers (1979) and Waldmann (1992). 3 More recently, Anderson and Marmot (2012) used data from the Whitehall II study to show that promotions reduce the probability of developing heart disease. However, the pattern across OECD countries shown by Leigh et al. (2009: Figure 3) indicates that the increase in life expectancy and the decline in infant mortality were more pronounced where inequality widened. Leigh and Jencks (2007) present long-run evidence from a panel of 12 advanced countries; they do not rule out "the possibility that inequality raises life expectancy by a substantively significant amount" (page 19).
seem not even to be generalizable across primate species, let alone generally applicable to the health effects of hierarchical human social organization." 11 Putnam (1995: 67) defines social capital as "features of social organization such as networks, norms, and social trust that facilitate coordination and cooperation for mutual benefit." 12 Note that d 'Hombres et al. (2010) consider income inequality as an instrumental variable for social capital. 13 Recall from Deaton (2003) that is also disputed that higher spending on medical care would necessarily result in better health outcomes. 14 Kawachi and Kennedy (1999: 221) quote Paul Krugman to this effect.  also note that public spending on health could decline with more heterogeneous preferences of voters. This argument is based on Alesina et al. (1999) who show that the average value of public goods to members of a community diminishes with more pronounced heterogeneity. However, heterogeneity in Alesina et al.'s analysis is mainly linked to ethnic fractionalization, while they do not find robust negative effects of income inequality on public spending (see Deaton, 2003: 131-2). 23 Data availability does not allow us to conduct a panel data analysis of the effect of the share of the bottom 20 percent on health in developed countries. 24 Again, the coefficients of GDP per capita and GDP squared have the "wrong" sign. The positive association between GDP per capita and tuberculosis may be due to endogeneity bias if higher morbidity increases future per-capita income by reducing population size. 25 For the sake of brevity, the results are not shown in detail. They are available from the authors on request.  Bank (1995) classification) are marked with an asterisk ( * ).   Countries  30  30  30  Observations  547  537 537 Note: Standard errors (calculated by the Delta method) are in parentheses. *** indicate significance at the one percent level. The effects were estimated from an autoregressive distributed lag model with one lag on the endogenous variable and one lag on the exogenous variable(s). 1232 Note: t-statistics are in parenthesis. *** indicate significance at the one percent level. The DOLS results are based on a one lead/lag model. For Azerbaijan, Georgia, and Russia, complete time series data on infant mortality are not available for the period 1976-2010. Therefore, these countries are not included in the sample used to estimate the coefficient in column 2. The dependent variable is LE. Standard errors (calculated by the Delta method) are in parentheses. ** (*) indicate significance at the five (ten) percent level. The long-run effects were estimated from an autoregressive distributed lag model with two lags on the endogenous variable and two lags on the exogenous variable(s), given the relatively long period covered in the data.