13
Social Science & Medicine 62 (2006) 779–791 Further examination of the cross-country association between income inequality and population health Rati Ram Economics Department, Illinois State University, Normal, IL 61790-4200, USA Available online 26 July 2005 Abstract Several scholars have put forward the view that the estimates by Rodgers [(1979). Income and inequality as determinants of mortality: An international cross-section analysis. Population Studies, 33 (2), 343–351], Flegg [(1982). Inequality of income, illiteracy and medical care as determinants of infant mortality in underdeveloped countries. Population Studies, 36 (3), 441–458] and Waldmann [(1992). Income distribution and infant mortality. Quarterly Journal of Economics, 107 (4), 1283–1302] showing a negative cross-country association between income inequality and population health, cannot be replicated from recent data. In view of the importance of this matter, the present study further examines the issue from the most recent, and probably more accurate, data for the largest cross-country sample used in this line of research. The main conclusion is that the negative cross-country association between income inequality and good health, reported by Rodgers, Flegg, and Waldmann, is replicated very well. The different findings indicated by some scholars may have been due to their samples or the models being unusual. Therefore, the recent skepticism about the existence of such a negative association needs to be reconsidered. Several additional points are also noted. First, income inequality shows significance even after an index of ethnic heterogeneity is included. Second, ethnic heterogeneity itself has a negative association with population health. Third, income inequality retains significance in the presence of a measure of social capital. Fourth, however, the association between the measure of social capital and population health appears weak. Fifth, a simple analysis does not support the view that the positive association between income inequality and infant mortality in less developed countries (LDCs) may just be a reflection of the role of poverty. Finally, there is some support for the proposition that while income may be relatively more important for health in LDCs, the role of income inequality may be stronger in developed economies. r 2005 Elsevier Ltd. All rights reserved. Keywords: Income inequality; Population health; Cross-country data; Poverty Introduction As many scholars, including Deaton (2003), Lynch et al. (2004), Subramanian and Kawachi (2004), and Wagstaff and Doorslaer (2000), have indicated, relation- ship between income inequality and population health has lately been subjected to immense scholarly research. In addition to the estimates reported for the United States and a few other countries, cross-country evidence on the relationship has been an important component of this research. As Ellison (2002, p. 561) noted, a compelling body of ecological research had found statistical association between various measures of inequality and average health status at the population level. In particular, Rodgers (1979), Flegg (1982), ARTICLE IN PRESS www.elsevier.com/locate/socscimed 0277-9536/$ - see front matter r 2005 Elsevier Ltd. All rights reserved. doi:10.1016/j.socscimed.2005.06.034 Tel.: +1 309 438 7101; fax: +1 309 438 5228. E-mail address: [email protected].

Further examination of the cross-country association between income inequality and population health

Embed Size (px)

Citation preview

ARTICLE IN PRESS

0277-9536/$ - se

doi:10.1016/j.so

�Tel.: +1 309

E-mail addr

Social Science & Medicine 62 (2006) 779–791

www.elsevier.com/locate/socscimed

Further examination of the cross-country association betweenincome inequality and population health

Rati Ram�

Economics Department, Illinois State University, Normal, IL 61790-4200, USA

Available online 26 July 2005

Abstract

Several scholars have put forward the view that the estimates by Rodgers [(1979). Income and inequality as

determinants of mortality: An international cross-section analysis. Population Studies, 33 (2), 343–351], Flegg [(1982).

Inequality of income, illiteracy and medical care as determinants of infant mortality in underdeveloped countries.

Population Studies, 36 (3), 441–458] and Waldmann [(1992). Income distribution and infant mortality. Quarterly

Journal of Economics, 107 (4), 1283–1302] showing a negative cross-country association between income inequality and

population health, cannot be replicated from recent data. In view of the importance of this matter, the present study

further examines the issue from the most recent, and probably more accurate, data for the largest cross-country sample

used in this line of research. The main conclusion is that the negative cross-country association between income

inequality and good health, reported by Rodgers, Flegg, and Waldmann, is replicated very well. The different findings

indicated by some scholars may have been due to their samples or the models being unusual. Therefore, the recent

skepticism about the existence of such a negative association needs to be reconsidered. Several additional points are also

noted. First, income inequality shows significance even after an index of ethnic heterogeneity is included. Second, ethnic

heterogeneity itself has a negative association with population health. Third, income inequality retains significance in

the presence of a measure of social capital. Fourth, however, the association between the measure of social capital and

population health appears weak. Fifth, a simple analysis does not support the view that the positive association between

income inequality and infant mortality in less developed countries (LDCs) may just be a reflection of the role of

poverty. Finally, there is some support for the proposition that while income may be relatively more important for

health in LDCs, the role of income inequality may be stronger in developed economies.

r 2005 Elsevier Ltd. All rights reserved.

Keywords: Income inequality; Population health; Cross-country data; Poverty

Introduction

As many scholars, including Deaton (2003), Lynch

et al. (2004), Subramanian and Kawachi (2004), and

Wagstaff and Doorslaer (2000), have indicated, relation-

ship between income inequality and population health

e front matter r 2005 Elsevier Ltd. All rights reserve

cscimed.2005.06.034

438 7101; fax: +1 309 438 5228.

ess: [email protected].

has lately been subjected to immense scholarly research.

In addition to the estimates reported for the United

States and a few other countries, cross-country evidence

on the relationship has been an important component of

this research. As Ellison (2002, p. 561) noted, a

compelling body of ecological research had found

statistical association between various measures of

inequality and average health status at the population

level. In particular, Rodgers (1979), Flegg (1982),

d.

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791780

Waldmann (1992), and Wilkinson (1992, 1996) found a

highly significant association between income inequality

and population health in several types of cross-country

data and models. Ellison (2002, p. 562) further observed

that the remarkable consensus between such a wide

variety of different studies led Wilkinson (1996, p. 105)

to conclude that ‘‘the income distribution relationship is

now firmly established’’, and to propose the ‘‘relative

income hypothesis’’, which says that an individual’s

health is affected by the distribution of income within

society.

Apart from the legitimate view that aggregate data do

not provide a good basis for an inference about the role

of income inequality in health at the individual level,

some recent research has questioned the very existence

of a negative association between income inequality and

population health in cross-country data. For example,

Baumbusch (1995) indicated that, contrary to Wald-

mann’s (1992) estimates, an increase in the income share

of the top 5% lowers infant mortality. Similarly, Judge,

Mulligan, and Benzeval (1998) considered the best

available data for 15 high-income countries and found

insignificant parameters for Gini in regression models of

life expectancy. More recently, based on 1990 data for

47 countries, Mellor and Milyo (2001, p. 499) indicated

that inclusion of average income and secondary-school

enrollment causes the coefficients of Gini to have a

perverse sign in models of life expectancy and infant

mortality. Similarly, Gravelle, Wildman, and Sutton

(2002) estimated Rodgers’s (1979) life-expectancy model

with more recent data and found (p. 580) that the

income-inequality parameter lacks statistical significance

at any meaningful level, and their extensive explorations

in terms of alternative specifications (p. 581), structural

changes with respect to time (p. 582), and variations

across high- and low-income levels did not indicate a

significantly negative parameter for income inequality.

They concluded (p. 587) ‘‘we have found, using a wide of

variety of specifications, that the association (between

income distribution and life expectancy) in our data

setyis never statistically significant’’. Wildman, Grav-

elle and Sutton (2003) stated that they could not

replicate Waldmann’s (1992) results with more recent

data. The studies claiming lack of a significant associa-

tion between income inequality and population health in

cross-country data seem to have been highly influential

and, for example, led Mackenbach (2002, p. 1) to state

that ‘‘Evidence favouring a negative correlation between

income inequality and life expectancy has disappeared’’.

Even more tellingly, in a recent discussion on Rodgers’s

(1979) paper, Deaton (2002, p. 548) stated that the

‘‘cross-country relationships for wealthy countries esti-

mated by Rodgers and Wilkinson do not show up in the

LIS (Luxembourg Income Study) data, nor can Rodgers

original regressions be replicated on the international

data from the World Bank’’, and that ‘‘As of the time of

writing, little appears to remain of the whole enterprise’’.

In the same discussion, Lynch and Smith (2002, p. 549)

wondered whether there was the ‘‘end of the story’’ in

regard to the relation between income inequality and

health. In the same vein, Deaton (2003, p. 139) stated

that ‘‘Later research has cast considerable doubt on the

robustness and reliability of many of these (cross-

country) findings’’ (of a negative relation between

income inequality and good health).

The motivation for this study is provided by the

consideration that, as noted above, while Rodgers

(1979), Flegg (1982), and Waldmann (1992) reported a

significant negative cross-country association between

income inequality and population health, several scho-

lars have suggested that such a negative association is

not found in the studies that tried to replicate the earlier

results from more recent data. In view of the obvious

importance of the point, the primary objective of this

study is to re-examine the cross-country association

between income inequality and population health by

using the most recent and the best available data for the

largest possible cross-country sample. The dataset used

in the present paper has the following desirable

characteristics that make it appropriate for the purpose:

1.

It covers a cross-section of 108 countries and is thus

larger than that used in almost any previous work.

2.

The variables pertain to the year 2000 or the late

1990s, and the data are thus more recent by 10 to 25

years than those used in earlier studies.

3.

The information on income inequality (and income

shares) reported in World Development Report 2003

(World Bank, 2003) is very good, and should be a

considerable improvement over that used in pre-

vious studies. As noted by World Bank (2003,

pp. 246–247), besides the LIS database for high-

income countries, information from nationally re-

presentative household surveys has been used, and an

effort has been made to ensure that the data on

income distribution are as comparable as possible.

Moreover, since the Gini coefficients (and income

shares) are based on income for some countries and

consumption for others, the present work makes an

adjustment so as to enhance cross-country compar-

ability of the numbers.

4.

Data on per-person GDP are in PPP (international)

dollars which have much greater cross-country

comparability than those used by Rodgers (1979)

and Flegg (1982).

5.

World Bank’s current information on infant mortal-

ity is likely to be better than that available 10 to 25

years ago.

6.

A lag of about 5 years is allowed between the health

indicators, which are for the year 2000, and each of

the regressors which typically pertain to mid- to late

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791 781

1990s. This makes the regressors econometrically

predetermined, and attenuates the ‘‘simultaneity’’

problem mentioned in the literature.

Besides the focus on a re-estimation of the basic

models used in the studies by Rodgers (1979), Flegg

(1982), Waldmann (1992), and some variants of these,

several related points are also studied. These include (a)

judging the robustness of the income inequality para-

meter in the presence of an index of ethnic hetero-

geneity, which might be an ‘‘omitted variable’’, (b)

considering the role of social capital, which might be a

channel for the effect of income inequality on health, (c)

shedding some light on Deaton’s (2003, p. 151) thesis

that the association between income inequality and

infant mortality in less developed countries (LDCs)

might be a reflection of the effect of poverty, and (d)

conducting a preliminary exploration of the view

expressed by Wilkinson (1996) and Deaton (2003,

p. 121) that while income might be relatively more

important for population health in LDCs, income

inequality is likely to be relatively more important in

developed countries (DCs).

Despite the usefulness of a comprehensive and

updated investigation to see whether the negative

association between income inequality and population

health reported by Rodgers, Flegg, and Waldmann can

be replicated from the most recent data, a major

shortcoming of all such studies that work with aggregate

data should be noted. One important issue addressed by

studies in this line of research is whether income

inequality at the aggregate level has an independent

‘‘contextual’’ or ‘‘ecological’’ effect on an individual’s

health. A proper investigation of that question requires a

multilevel research design in which income inequality at

the aggregate or community level is entered along with

individual characteristics in a model of individual health.

Such multilevel studies have been conducted by many

scholars, including Subramanian and Kawachi

(2003a, b), and have been ably reviewed by Subramanian

and Kawachi (2004). Aggregate models of the kind used

in the present study, and those which it replicates, cannot

shed any direct light on the role of income inequality in

individual health. The main point addressed in the

present study is to judge afresh the presence or absence

of a negative association between income inequality and

population health in cross-country data.

Models and data

Models

The models for which final estimates were provided by

Rodgers (1979, Table 6, p. 349) are used in this

replication. These may be written as

Hi ¼ a1 þ b1ð1=PCYÞi þ c1ð1=PCY2Þi

þ d1ðGINIÞi þ ui1 ð1Þ

and

Hi ¼ a2 þ b2½1= lnðPCYÞ�i þ d2ðGINIÞi þ ui2, (2)

where Hi denotes life expectancy at birth (LIFE) or

infant mortality rate (IMR) for country i, PCY is GDP

per capita, PCY2 is the square of PCY, GINI is the Gini

coefficient, ln denotes natural logarithm of the variable,

and u’s are stochastic error terms. As indicated by

Rodgers (1979), he explored a wide variety of nonlinear

specifications for the income variable, and reported in

his Table 6 (p. 349) the results that were ‘‘the best or

most interesting’’.

Flegg’s (1982, p. 445) model of infant mortality

included logarithm of real GDP per capita, Gini

coefficient, female illiteracy rate, supply of nurses, and

supply of doctors. However, data on nurses are not

available in the sources used for this study, and Flegg

did not report estimates with the exclusion of nurse-

supply alone. Therefore, the following model, for which

he reported estimates (Flegg, 1982, p. 445, Table 1,

second line) and for which data are now available, is

estimated so as to facilitate comparisons:

lnðIMRÞi ¼ a3 þ b3 lnðPCYÞi þ c3 lnðGINIÞi

þ d3 lnðFILITÞi þ ui3, ð3Þ

where FILIT denotes female illiteracy rate. Since he

worked with a sample of LDCs, this study also focuses

on that group relative to the above model.

Waldmann’s (1992, pp. 1283, 1287) basic model has

the following form:

lnðIMRÞi ¼ a4 þ b4 lnðPCYPOORÞi þ c4 lnðPCYMIDÞ

þ d4ðSHARE5Þi þ ui4. ð4Þ

where PCYPOOR is per-person income of bottom 20%,

PCYMID denotes per capita income of the middle

group (between bottom 20% and top 5%), and

SHARE5 is the income-share of top 5%. Three

variations over Waldmann’s model are adopted in this

study. First, World Bank data on the share of bottom

20% pertain to income for some countries and

consumption for others, and it is not possible to make

a reasonable estimate of per capita income of bottom

20% from the information on their share. Also, the

poor-income variable lacked significance in his full

regressions at any meaningful level. Therefore, that

variable is not included here. Second, due to the shares

being a mix of consumption and income, it is also not

possible to get good estimates of per capita income of

the ‘‘middle’’ group lying between bottom 20% and top

5%. Therefore, per capita income of the entire popula-

tion (PCY) is used, which is fairly close to his variable.

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791782

Third, information on the share of top 10% (SHARE10)

is directly available from World Development Report

2003, and that is used in place of the share of top 5%.

However, since the share-information is based on

income for some countries and on consumption for

others, an intercept-dummy has been included to make

allowance for this characteristic of the data on shares.

The dummy variable (DCON) takes the value one for

countries where the distribution data are based on

consumption and zero for others. Thus the following

variant of Waldmann’s model is estimated:

lnðIMRÞi ¼ a04 þ b04 lnðPCYÞi þ c04ðSHARE10Þi

þ d 04ðDCONÞi þ u0i5. ð40Þ

This has a fair resemblance to Waldmann’s specification.

Also, since SHARE10 is a well-known measure of

income inequality, the modified model can be perceived

as a format that contains average income along with

income inequality, but inclusion of the share of top 10%

along with the logarithm of mean income may attenuate

the ‘‘aggregation problem’’.

In addition to estimation of models of the kind

specified by Rodgers (1979), Flegg (1982) and Wald-

mann (1992), an effort is made to estimate the following

cross-country IMR model reported by Mellor and Milyo

(2001, p. 499) whose estimates show a dramatic

difference from those of Rodgers (1979).

IMRi ¼ a5 þ b5ðGINIÞi þ c5ðPCYÞi

þ d5ðSCHOOLÞi þ ui5, ð5Þ

where SCHOOL stands for secondary school enrollment

rate. They reported the estimates for the full sample, and

that is done for the replications also.

Data

The information on income inequality is taken

from World Development Report 2003 (World Bank,

2003, pp. 236–237). Every country for which data on

Gini index and share of top 10% was available has

been included, leading to a cross-section of 108

countries. Most information is from mid- to late

1990s. The Report (pp. 234–235) also provides informa-

tion on life expectancy at birth for the year 2000, which

allows a lag of a few years between the dependent

variable (health indicator) and Gini index. Data on IMR

are taken from the 2002 edition of World Development

Indicators on CD-ROM (World Bank, 2002), and, like

life expectancy, are for the year 2000. Information on

income, female illiteracy rate, and secondary school

enrollment is also taken from the CD-ROM and is for

1995 (or a year closest to that), allowing a 5-year lag

relative to the dependent variables. GDP per person in

current-price PPP (international) dollars is the proxy for

income.

It is perhaps of some importance to note again that

the Gini index is based on income for some countries

and on consumption for others. Following Deininger

and Squire (1996, p. 582), Gini coefficients based on

consumption have been adjusted upward by 6.6 so as to

make this important variable more comparable across

countries. It may also be noted that both Rodgers (1979)

and Flegg (1982) used income measures in conventional

US dollars since PPP measures were not easily available

at that time.

The Appendix lists the 108 countries and also

identifies the OECD members. However, since avail-

ability of data for the variables differs, the sample size

varies across the models.

Main results

Tables 1 and 2 provide a preliminary view of the data.

Table 1 indicates descriptive statistics for the sample

and Table 2 shows simple correlations. Besides the

variables discussed above, the tables include three others

that are introduced later. One is an index of ethnic

heterogeneity (ETHNIC). Another is generalized trust

(TRUST), which is a measure of social capital. The third

is the poverty index (POOR), which is the percent of

population whose income is below the international

poverty line of one PPP dollar per day (per person). The

tables indicate several points. First, cross-country

dispersion in infant mortality, income, and female

illiteracy (and poverty) seems much greater than in life

expectancy, Gini coefficient, and share of the top 10%.

Second, life expectancy and IMR have fairly high

correlations of the expected kind with both income

and Gini, and the correlations with female illiteracy and

school enrollments are also high and have the expected

signs. Third, correlations between life expectancy and

IMR and between GINI and share of top 10% are

expectedly very high (�0.94 and 0.94). Moreover,

correlation between female illiteracy and school enroll-

ments is high, and so is that between GINI and school

enrollments. Fourth, correlations associated with the

three other variables (ETHNIC, TRUST and POOR)

are of the expected kind, and, as discussed later, seem

informative.

Table 3 contains the main results. Estimation is

done by the ordinary least-squares (OLS) procedure,

but, due to the diversity of the sample, the t-statistics

are based on White’s (1980) heteroscedasticity-consis-

tent standard errors. Besides the models specified in

Eqs. (1)–(5), a modified version of Mellor and Milyo

(2001) is also estimated so as to permit a nonlinear

relation between income and infant mortality. The

modification is that their linear IMR equation is

changed to the Flegg-type log-log form, but the

ARTICLE IN PRESS

Table 1

Descriptive statistics for the main variables

Mean Std. dev. Min. Max. N

LIFE (life expectancy at birth, years, for year 2000) 65.59 12.20 38.00 81.00 108

IMR (infant mortality rate, per thousand live births, 2000) 38.62 37.13 3.40 153.60 108

PCY (GDP per capita, PPP dollars, around 1995) 7106 7569 474 28284 108

GINI (adjusted Gini coefficient, percent, mostly late 1990s) 43.38 10.94 19.50 69.50 108

TOP10 (income/consumption share of top 10%, percent, mostly

1990s)

31.13 7.48 18.20 48.80 108

FILIT (female illiteracy rate, percent, around year 1995) 24.18 27.41 1.00 93.40 101

SCHOOL (secondary school enrollment rate, percent, around 1995) 56.45 30.92 4.90 98.60 66

POOR (percent of population that is below the international poverty

line of PPP $1 per day, mostly 1990s)

18.52 19.57 2.00 72.80 77

TRUST (proxy for social capital, 2000 or mid-1990s) 26.72 14.74 2.83 66.53 68

ETHNIC (index of ethnic fractionalization, around 1990s) 0.42 0.25 0.002 0.93 107

Note: Data on GINI, TOP10, POOR and LIFE are from World Development Report 2003 (World Bank, 2003, pp. 234–237).

Information on IMR, PCY, FILIT, and SCHOOL is from the 2002 edition of World Development Indicators on CD-ROM (World

Bank, 2002). For most OECD countries, the female illiteracy rate is zero or nearly zero. To enable a logarithmic transformation for

Flegg’s (1982) model, a value of 1 has been used in these cases. Information on ETHNIC is taken from Alesina, Devleeschauwer,

Easterly, Kurlat, and Wacziarg (2003, pp. 184–189) and most data are for the 1990s. TRUST is taken from Bjornskov (2006,

Appendix Table 1), and most numbers are based on 4th wave of World Values Survey and the reference year is around 2000.

Table 2

Simple correlations between the main variables

LIFE IMR PCY TOP10 FILIT SCHOOL POOR ETHNIC TRUST

IMR �0.94*

PCY 0.65* �0.62*

TOP10 �0.45* 0.42* �0.46*

FILIT �0.73* 0.83* �0.57* 0.32*

SCHOOL 0.77* �0.81* 0.78* �0.68* �0.72*

POOR �0.79* 0.82* �0.60* 0.49* 0.64* �0.77*

ETHNIC �0.61* 0.59* �0.48* 0.37* 0.48* �0.48* 0.54*

TRUST 0.40* �0.31* 0.57* �0.51* �0.18 0.51* 0.03 �0.34*

GINI �0.57* 0.55* �0.59* 0.94* 0.42* �0.75* 0.57* 0.46* �0.55*

Note: See Table 1 for variable definitions. An asterisk (*) indicates significance at least at the 5% level, based on the p-values generated

by SAS for Windows (version 8), which was used for all computations.

R. Ram / Social Science & Medicine 62 (2006) 779–791 783

variables are kept unchanged. The following points are

indicated by Table 3:

1.

1Rodgers (1979, p. 349) compared the full sample estimates

with those for the LDCs. Full sample estimates are reported in

Table 3. LDC estimates, which are very similar, are available

from the author. For example, LDC parameters for GINI in

models of life expectancy and IMR have White-consistent t-

statistics that range from 2.96 to 3.98 in absolute value.

Rodgers’s (1979, p. 349) models are remarkably well

replicated despite (a) the much larger cross-country

sample, (b) the data being more recent by at least 25

years, and (c) the income variable being in PPP

dollars. In every case, estimated parameter for GINI

is highly significant in both LIFE and IMR

equations. In particular, the estimates for LIFE are

a stark contrast from the replications reported by

Gravelle et al. (2002, p. 580, column 6) in regard to

the significance of the parameter for GINI. Part (ii)

in section A of the table provides a quick comparison

of the GINI parameter estimated in this study with

those reported by Rodgers (1979, p. 349) and

Gravelle et al. (2002, p. 580). The comparison

indicates the estimate reported by Gravelle et al. to

be ‘‘unusual’’.

Some differences from Rodgers’s estimates might

also be of interest.1 Besides the greater explanatory

power (R2 being 0.76–0.81, as compared with

ARTICLE IN PRESS

Table 3

Relation between income inequality and population health: life expectancy (LIFE) and/or infant mortality rate (IMR) equations

A. Rodgers (1979, p. 349) models

C 1/PCY 1/PCY2 1/ln(PCY) GINI R2 N

(i) Full sample estimates from this study

LIFE 87.184* �30.104* 6.565* �0.254* 0.78 108

(33.85) (�6.06) (2.63) (�3.10)

141.201* �548.368* �0.190* 0.75 108

(40.76) (�11.29) (�2.14)

IMR �23.078* 98.205* �22.050* 0.634* 0.81 108

(�3.87) (6.06) (�2.34) (3.32)

�194.203* 1734.173* 0.456* 0.76 108

(�15.64) (12.50) (2.12)

(ii) Comparison of GINI coefficients in LIFE equations from Rodgers (1979, p. 349), Gravelle et al. (2002, p. 580) and the present study for the common model

Rodgers (1979, p. 349, row iv) �36.47 (t-statistic �3.76)

Gravelle et al.(2002, p. 580, col. 6) �5.39 (t-statistic �0.97)

Present study (first row above, GINI on 0–1 scale) �25.42 (t-statistic �3.10)

B. Flegg’s (1982, p. 445) model: Dependent variable is ln(IMR): LDC sample

ln(Y) ln(GINI) ln(FILIT) R2 N

Flegg’s estimates �0.299* 0.774* 0.278* 0.66 46

(1982, p. 445, 2nd row) (�2.93) (3.03) (3.50)

Present study estimates �0.569* 0.626* 0.250* 0.84 74

(�7.62) (2.57) (5.97)

C. Models like those of Waldmann (1992, p. 1287): Dep. variable is ln(IMR)

C l(Ypoor) l(Ymid) Rich-share Year-dummy R2 N

Waldmann’s estimates 11.92 0.07 �0.80* 2.48* 0.06 0.79 57

(1992, p. 1287, col. 1) (13.2) (0.45) (�5.57) (2.82) (0.44)

Wildman, Gravelle, 10.07 �0.27 �0.65** �0.51 �0.19 0.77 71

Sutton (2003, p. 1002, column 4) (14.48) (�0.71) (�1.79) (�0.32) (�1.62)

Present study C ln(PCY) TOP10 DCON R2 N

7.33* �0.67* 3.52* 0.45* 0.88 108

(11.18) (�12.45) (5.31) (3.92)

D. Mellor and Milyo (2001, p. 499) IMR models: Dep. variable is IMR

GINI PCY SCHOOL R2 N

Mellor and Milyo (2001, p. 499, col. 3) estimates �0.364 �0.001 �98.15 0.72 47

(�0.99) (�0.73) (�4.32)

Present study �0.467 �0.076 �105.450* 0.66 66

(�1.09) (�0.24) (3.90)

Adaptation of Mellor–Milyo IMR model to log-log form: Dep. variable is ln(IMR)

ln(GINI) ln(PCY) ln(SCHOOL) R2 N

1.048* �0.664* �0.175** 0.91 66

(3.70) (�8.60) (�1.76)

Note: The models are explained more fully in the text. The variables are defined in Table 1. ‘‘ln’’ denotes natural logarithm of the

variable. DCON in Waldmann’s models is an intercept dummy variable which is included to take account of the fact that TOP10

represents income-share for some countries and consumption-share for others.

Relevant t-statistics are in parentheses. For the present study, these are based on White’s (1980) heteroscedasticity-consistent standard

errors. All computations are done on SAS for Windows (version 8).

*Statistically significant at least at the 5% level.

**Significant at the 10% level

R. Ram / Social Science & Medicine 62 (2006) 779–791784

ARTICLE IN PRESS

2

rep

equ3

cou

Ro

and

its

the

pan

esti

sec

R. Ram / Social Science & Medicine 62 (2006) 779–791 785

Rodgers’s 0.56–0.77), every income term in Table 3

(section A) is statistically significant at least at the

5% level, which increases the plausibility of the

estimates. Also, significance of the GINI parameter is

much stronger in the present study for IMR models.

The difference may reflect larger sample size and the

better income variable used in the present study. The

differences in the magnitude of the estimates are

probably due to different units. In the present study,

income is in thousand (1995) PPP dollars, and GINI

is on a 0–100 scale instead of the traditional 0–1

scale.

2.

Flegg’s (1982) model is replicated very well despite a

25-year difference in the timing of the data, improved

variable values, and a considerably larger sample

size. The difference in the size of the income

parameters probably reflects the better (PPP) income

measure used in the present study.

3.

Section C in the table reflects faithfully the main

points of Waldmann’s (1992) basic argument. De-

spite the differences in the models, there is a

remarkable similarity between the two sets of

estimates. The parameter for per capita income of

the entire population [ln(PCY)] is close to Wald-

mann’s for per capita income of the middle 75%

group [l(Ymid)]. More importantly, the coefficient

for the share of top 10% is close to his number for

the top 5% (rich-share). It is clear that given average

income of the population, income inequality proxied

by the income share of top 10% is associated with a

substantial increase in IMR.2

A comparison of the present study’s and Wald-

mann’s estimates with those of Wildman et al. (2003,

p. 1002) is also provided in section C of Table 3. The

comparison suggests that their parameter for the

rich-share is likely to reflect some unusual character-

istics of the sample or the data.

4.

The basic cross-country estimates of Mellor and

Milyo’s (2001, p. 499) IMR model are replicated well

in section D of Table 3. As shown in their full

regressions (p. 499, column 3), neither income nor

Gini has a significant association with IMR; only

secondary-school enrollment matters.3

Mellor–Milyo models that enter income linearly in

Following Waldmann, estimates for the IMR model are

orted. The results from the corresponding life expectancy

ation are similar.

Two points about the Mellor and Milyo (2001) cross-

ntry estimates may be noted. First, perhaps following

dgers (1979), they reported estimates for both life expectancy

IMR. Table 3 reports the IMR equation so as to facilitate

reformulation into the Flegg-type logarithmic form. Second,

y also reported (2001, Table 3, p. 501) some results from

el cross-country data. However, their Table 2 (p. 499)

mates are more relevant to the present study that uses cross-

tion data.

health regressions are unusual in the literature, and

a flavor of the consequence of entering income

nonlinearly in their specification is provided in the

last part of section D of Table 3. It is done by

estimating a Flegg-type log-log model of infant

mortality with the same variables as were used

by Mellor and Milyo, which really amounts to

replacing Flegg’s female illiteracy rate by secon-

dary school enrollment. The estimates show that

the change in going from Mellor–Milyo linear

form to Flegg-type log-log form is rather dramatic.

The explanatory power of the model rises to 0.91,

income shows a highly significant negative sign,

and GINI has a highly significant positive associa-

tion with infant mortality. The schooling variable

is also marginally significant and has the expected

sign. These estimates seem much more plausible than

those of Mellor and Milyo and the corresponding

replication in the present study. It would be

appropriate to note here that there are obviously

many ways in which the models can be made

nonlinear in income, and the example given here is

somewhat arbitrary in choosing Flegg’s specification.

However, unreported estimates from an alternative

model (of logarithm of IMR), in which only PCY is

entered logarithmically while GINI and SCHOOL

enter linearly, also show GINI to be significant at

better than the five percent level with a White-

consistent t-statistic of 2.03.

5.

The overall picture suggested by Table 3 is that the

estimates reported by Rodgers (1979), Flegg (1982)

and Waldmann (1992), which all show income

inequality to be significantly negatively associated

with good health, are replicated well despite the

substantially larger sample size, much more recent

data, and considerably better variables. In particular,

the comparison in section A of Table 3 of the life-

expectancy models shows the stark contrast between

Rodgers’s (1979) and this study’s results from those

reported by Gravelle et al. (2002, p. 580). Similarly,

replication of the modified Waldmann (1992) model

in section C of Table 3 is a contrast from Wildman

et al. (2003). Moreover, Mellor and Milyo (2001,

p. 499, columns 3 and 6) estimates, which indicate

neither income nor inequality to have any signifi-

cant effect on life expectancy or infant mortality,

appear implausible in suggesting school-enrollment

to be the only driver of cross-country variations

in life expectancy and infant mortality, and it is

shown that the estimates undergo a dramatic change

when a Flegg-type nonlinear model of infant

mortality is used. It is, therefore, fair to conclude

that predominance of the cross-country evidence

supports the view that income inequality has a

significant negative association with the health status

of the population.

ARTICLE IN PRESS

4Samples for Rodgers’s models in Tables 3 and 4 differ

slightly due to missing data on ETHNIC for Yemen.

R. Ram / Social Science & Medicine 62 (2006) 779–791786

Discussion

The estimates in Table 3 can be interpreted as belying

the skepticism articulated by Deaton (2002, 2003),

Gravelle et al. (2002), Lynch and Smith (2002),

Mackenbach (2002), Mellor and Milyo (2001), Wildman

et al. (2003), and other scholars, about the replicability

of the studies by Rodgers (1979), Flegg (1982) and

Waldmann (1992) and the existence of a negative

association between income inequality and good health

in cross-country data. However, it seems useful to

consider some questions that have a bearing on the

topic, although these apply as much to the earlier studies

as to the present one. The point about the difficulty of

using aggregate data to shed light on the relation

between income inequality and individual health has

already been noted in the introductory section. The

difficulty of judging the concavity of the individual

income–health relation from aggregate data is also

evident.

An additional concern about using income inequality

in aggregate health models is that some relevant

variables may have been omitted and the effect of these

variables may be reflected in income inequality para-

meters. One such variable is ethnic or social hetero-

geneity. Table 2 indeed shows that the index of ethnic

heterogeneity (ETHNIC) has a sizable positive correla-

tion with GINI and a large negative correlation with

measures of good health. Therefore, absence of ethnic

heterogeneity in health models may bias the coefficient

of income inequality. Such a possibility is explored by

including index of ethnic heterogeneity along with GINI

in Rodgers’s (1979) specification and the modified

Waldmann model, for which the sample size is the

largest. The index of ethnic fractionalization is taken

from Alesina et al. (2003, pp. 184–189), and, following

the standard practice, has been computed by them as

one minus the Herfindahl index of ethnic groups’

population shares. Although the reference years differ,

most numbers are for the 1990s and are temporally

congruent with other variables in this study. Section A

in Table 4 shows the estimates with inequality (GINI

or TOP10) alone, with both inequality and ETHNIC,

and with ETHNIC replacing inequality. Six points

may be noted here. First, parameters of income

inequality (GINI or TOP10) are very similar whether

or not ETHNIC is added to the model. There is thus

no indication of any bias in the income inequality

parameters due to the omission of ETHNIC. Second,

significance of the income inequality variables is

hardly altered after ETHNIC is added to the models.

Third, ethnic heterogeneity has a negative associa-

tion with good population health even in the presence

of an inequality term. Fourth, as would be expected,

ETHNIC has a sizable negative association with

population health when income inequality is not

included and the association is stronger than when

income inequality is included. Fifth, explanatory

power of the models is somewhat weaker when

ETHNIC replaces income inequality, suggesting that

the latter may be more important. Last, in all cases, the

estimates bear a remarkable similarity with the corre-

sponding rows in Table 3, indicating robustness of the

estimates.4

Another important point concerns the mechanisms

through which income inequality may affect population

health adversely. Several well-known hypotheses have

been proposed to explain the association between

income inequality and health. As Wagstaff and Door-

slaer (2000) have shown, it is difficult to judge from

aggregate data relative merits of these hypotheses.

However, omission of a major mechanism from

aggregate health models could lead to a bias in the

income inequality parameter in much the same way as

the omission of another relevant variable, and it is

desirable to make an effort to shed some light on the role

of a possible mechanism. One important hypothesis

about a possible mechanism suggested by Wilkinson

(1996) and other scholars is that income inequality

lowers social cohesion or social capital and thus

adversely affects individual and population health.

Although social capital seems to have several dimen-

sions, ‘‘generalized trust’’ is considered to be a major

component, and it is possible to do an exploratory

analysis to see how a measure of generalized trust

correlates with income inequality and whether addition

of such a measure erodes the income inequality

parameter. As Bjornskov (2006) and other scholars

have noted, World Values Survey (WVS) provides a

good source for country-level information on general-

ized trust (TRUST). The numbers used in the present

study are taken from Bjornskov (2006, Appendix Table

1). He has explained that the data are based largely on

the fourth wave of WVS and most are for the year 2000.

The number for each country is the percentage of

population that answered ‘‘yes’’ to the question, ‘‘In

general, do you think that most people can be trusted, or

can’t you be too careful?’’

In terms of simple correlations, Table 2 shows that

TRUST has large negative correlations with income

inequality (GINI and TOP10) that are highly significant.

Thus, there seems some merit in Wilkinson’s hypothesis

about income inequality eroding social capital. More-

over, TRUST has sizable positive correlations with

measures of good health. Therefore, it is possible that

exclusion of TRUST from aggregate health regressions

might bias the income inequality parameters. Part B in

Table 4 reports estimates of Rodgers’s and modified

Waldmann models with income inequality alone, with

ARTICLE IN PRESS

Table 4

Selected tests of sensitivity of income inequality parameters

A. Sensitivity to inclusion of ethnic heterogeneity index (ETHNIC)

Rodgers (1979, p. 349) models

C 1/PCY 1/PCY2 GINI ETHNIC R2 N

LIFE

With GINI but 86.792* �30.542* 6.473* �0.242* 0.79 107

Without ETHNIC (33.47) (�6.00) (2.52) (�2.92)

With both GINI 87.130* �27.725* 5.647* �0.208* �6.389* 0.80 107

And ETHNIC (35.08) (�5.27) (2.20) (�2.60) (�2.17)

With ETHNIC but 80.357* �32.739* 7.888* �8.697* 0.78 107

Without GINI (104.80) (�6.41) (3.06) (�2.63)

IMR

With GINI but �22.074* 99.323* �21.814* 0.603* 0.82 107

Without ETHNIC (�3.70) (5.98) (�2.22) (3.16)

With both GINI �22.815* 93.156* �20.006* 0.528* 13.987** 0.82 107

And ETHNIC (�3.86) (5.50) (�2.02) (2.81) (1.74)

With ETHNIC but �5.616* 105.887* �25.697* 19.846* 0.81 107

Without GINI (�2.47) (6.79) (�2.74) (2.37)

Modified Waldmann (1992, p. 1287) model: Dep. variable is ln(IMR)

C ln(PCY) TOP10 ETHNIC DCON R2 N

With TOP10 but 7.324* �0.671* 3.519* 0.445* 0.88 107

Without ETHNIC (10.74) (�11.98) (5.16) (3.90)

With both TOP10 6.819* �0.623* 3.285* 0.449* 0.435* 0.89 107

And ETHNIC (9.56) (�10.38) (5.10) (2.33) (4.03)

With ETHNIC but 8.695* �0.727* 0.631* 0.325* 0.85 107

Without TOP10 (12.60) (�10.72) (2.82) (2.32)

B. Sensitivity to the inclusion of social capital (TRUST)

Rodgers (1979, p. 349) models:

C 1/PCY 1/PCY2 GINI TRUST R2 N

LIFE

With GINI but 87.044* �25.795* 5.054** �0.274* 0.67 68

Without TRUST (24.08) (�3.90) (1.76) (�2.39)

With both 84.783* �25.763* 5.132** �0.244** 0.039 0.67 68

GINI and TRUST (18.28) (�3.95) (1.79) (�1.94) (1.03)

With TRUST but 73.905* �30.884* 7.398* 0.121* 0.63 68

Without GINI (37.94) (�5.37) (2.77) (2.62)

IMR

With GINI but �20.276* 78.172* �17.961* 0.661* 0.74 68

Without TRUST (�3.32) (4.45) (�2.22) (3.21)

With both GINI �26.813* 78.263* �17.734* 0.748* 0.114 0.75 68

And TRUST (�2.88) (4.45) (�2.21) (3.24) (1.20)

With TRUST but 6.523 93.956* �24.677* �0.136 0.68 68

Without GINI (1.35) (5.80) (�3.27) (�1.31)

Modified Waldmann (1992, p. 1287) model: Dep. variable is ln(IMR)

C ln(PCY) TOP10 TRUST DCON R2 N

With TOP10 but 6.114* �0.574* 4.645* 0.526* 0.86 68

Without TRUST (8.47) (�9.41) (5.73) (4.17)

With both TOP10 6.103* �0.581* 4.753* 0.001 0.524* 0.86 68

And TRUST (8.43) (�8.80) (5.28) (0.41) (4.16)

With TRUST but 8.613* �0.676* �0.008* 0.439* 0.78 68

Without TOP10 (10.81) (�7.47) (�2.01) (2.38)

Note. See the text for meanings and sources of ETHNIC and TRUST. Notes in other tables apply here also; in particular, t-statistics

are based on White’s consistent standard errors.

R. Ram / Social Science & Medicine 62 (2006) 779–791 787

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791788

both TRUST and income inequality, and with TRUST

replacing income inequality. These estimates also

suggest six points. First, parameters for income inequal-

ity variables (GINI and TOP10) are almost identical

whether or not TRUST is added to the models. In fact,

the income inequality parameters become larger with the

addition of TRUST in Waldmann-type and Rodgers’s

IMR models. Second, income inequality retains high

statistical significance in almost all cases when TRUST

is added. Third, despite the sizable correlations of

TRUST with IMR and LIFE in Table 2, when income

inequality and TRUST are included in the models,

coefficients of TRUST lack statistical significance at any

meaningful level and have perverse signs in both IMR

equations. Fourth, when TRUST replaces income

inequality, it has the expected sign and shows signifi-

cance in two of the three cases. Fifth, the explanatory

power of the models is considerably weaker when

TRUST replaces the income inequality variables. There-

fore, while erosion of trust (social capital) might

certainly be one channel through which income inequal-

ity affects health, the role of income inequality seems

more important and diverse. Last, despite the very

different and much smaller sample, estimates in part B

of Table 4 are similar to the corresponding rows in Table

3 and part A of Table 4, once again reaffirming

robustness of the estimates.

As the earlier discussion explains, while this study’s

estimates in Table 3 are remarkably similar to those of

Rodgers (1979), Flegg (1982) and Waldmann (1992) in

showing a negative association between income inequal-

ity and good health, there are important differences

from the patterns reported by several influential studies,

and it is useful to consider further possible reasons for

these differences.

The difference between the estimates reported by

Rodgers (1979), that are well replicated in this study,

and the results shown by Mellor and Milyo (2001) may

be due to the very different specification used by Mellor

and Milyo. Besides the addition of the schooling

variable, they entered income linearly although, follow-

ing Preston (1975), Rodgers (1979) explicitly and care-

fully explored several nonlinear forms of the income

variable. The work by Mellor and Milyo (2001) seems to

be the only instance in the literature where income is

entered linearly in a model of aggregate health. As

already observed, section D in Table 3 shows that if

Mellor–Milyo linear specification is changed to (Flegg-

type) log-log form, income inequality shows a highly

significant positive coefficient in the IMR equation,

which is a sharp contrast from Mellor–Milyo estimates

in the linear form. Estimates from the log-log form seem

much more plausible than those of Mellor and Milyo

which indicate school enrollment to be the only

significant driver of cross-country variations in health

and suggest that even average income does not matter.

Considerable effort was made to see why Gravelle

et al. (2002) were unable to replicate Rodgers’s (1979)

finding of a significant negative coefficient on income

inequality in models of life expectancy. It was noted that

their study differed from Rodgers’s in several ways.

Besides the more recent periods, a different (and

apparently smaller) set of countries, and pooling of

observations for two periods, they used male life

expectancy instead of overall life expectancy. A quick

ancillary project was undertaken to try to replicate the

Gravelle et al. results with the kind of data used by

them. Following their procedure, the years 1981 and

1989 (or close to these) were taken, income was obtained

from an update (PWT 6.2) of Summers and Heston

(1991), Gini was derived from the high-quality compo-

nent of Deininger and Squire (1996) dataset, and male

life expectancy was obtained from United Nations

(2001, Table A.30), which is an authentic and original

source. The sample size in that project was 110 for the

cases in which data for 1981 or 1989 (or both) were

available. However, coefficient estimates for Gini

corresponding to their columns 5 and 6 (p. 580) were

found to be significant at the one percent level, and the

White-consistent t-statistics were �3.43 and �2.85,

respectively. These estimates portray a picture that is

very different from theirs, but is very similar to that in

Table 3. It is possible that they just happened to get

atypical observations in their relatively limited cross-

country sample.

It may also be noted that while subjecting the life-

expectancy models to an extensive examination, Grav-

elle et al. (2002) did not report any estimates for

Rodgers’s infant-mortality equations. It is possible that

Gini coefficients were significant for infant-mortality

models even in their sample.

It is difficult to say why Wildman et al. (2003) were

unable to replicate from their data Waldmann’s (1992)

result that share of top 5% is associated with increased

infant mortality. A much larger sample and more recent

and better data in the present study show a pattern that

is strikingly similar to Waldmann’s despite several

differences in the models. It is possible that there was

something unusual or special about the Wildman et al.

(2003) sample or data.

One aspect related to the main exploration undertaken

in the present study is Deaton’s (2003, p. 151) suggestion

that infant (and child) mortality in less developed

countries is primarily a consequence of poverty and

that, conditional on average income, income inequality is

important only because it is effectively a measure of

poverty. While there are well-known difficulties in

comparing poverty incidence across countries, a simple

test was conducted by estimating for the LDCs the IMR

equations reported in Table 3 with the addition of a

poverty variable (POOR), which is the percent of

population below the international income-poverty line

ARTICLE IN PRESS

Table 5

Comparing roles of income inequality (GINI) and poverty (POOR) relative to infant mortality (IMR) in less-developed countries

(LDCs)

Rodgers’s model: Dep. variable is IMR

C 1/PCY 1/PCY2 GINI POOR R2 N

�17.976 85.156* �20.050** 0.450** 0.475* 0.82 69

(�1.52) (3.48) (�1.81) (1.65) (2.09)

Flegg’s (1982) model: Dep. variable is ln(IMR)

C ln(PCY) ln(GINI) ln(POOR) ln(FILIT) R2 N

4.384* �0.448* 0.474** 0.091** 0.249* 0.87 64

(3.76) (�5.60) (1.88) (1.90) (5.89)

Modified Waldmann model: Dep variable is ln(IMR)

C ln(PCY) TOP10 POOR DCON R2 N

6.774* �0.588* 2.801* 0.007 0.474* 0.79 69

(8.27) (�6.24) (2.91) (1.44) (3.61)

Mellor–Milyo (2001) model in log-log form: Dep variable is ln(IMR)

C ln(PCY) ln(GINI) ln(POOR) ln(SCHOOL) R2 N

5.431* �0.751* 1.016* 0.077 �0.007 0.81 37

(2.79) (�6.19) (2.44) (0.73) (�0.05)

Note: These are some of the IMR models of sections A, B, C and D of Table 3 with the addition of the poverty variable, and are

estimated for the LDC sample, which is obtained by excluding OECD countries from the full sample. Notes for that table are relevant

here also. In particular, t-statistics in parentheses are based on White’s (1980) heteroscedasticity-consistent standard errors, and an

asterisk indicates statistical significance at least at the 5% level, while ** indicates significance at the 10% level.

5Additional information is available from the author.

R. Ram / Social Science & Medicine 62 (2006) 779–791 789

of one PPP dollar per day per person. Despite its

weaknesses, the measure has some usefulness for cross-

country comparisons because, as World Bank (2003,

p. 246) noted, it holds ‘‘the real value of poverty line

constant across countries’’. Based on World Bank’s

(2003, pp. 236–237) data on poverty, Table 5 contains

estimates of the augmented IMR models. In two of the

four equations, POOR lacks significance at the 10% level

while income inequality variables remain significant at

better than the 5% level. In Flegg’s model, both income

inequality and poverty are significant at the 10% level. In

Rodgers’s model, while poverty is significant at 5%,

income inequality is significant at the 10% level. Based

on this preliminary evidence, it is difficult to say that the

association of income inequality with infant mortality in

LDCs is just a reflection of the effect of poverty. While

poverty does seem important, income inequality appears

to have a significance of its own.

Another related aspect is the proposition articulated

by Wilkinson (1996) and Deaton (2003, p. 121) that,

relative to income inequality, average income should

have a stronger association with population health in

less-developed countries, but income inequality should

be relatively more important in developed economies. A

preliminary exploration of that proposition was done by

comparing the implied marginal ‘‘effects’’ (parameters)

of GINI and per capita income on IMR from Flegg’s

model for the full sample and the LDCs. The

comparison is indirect since LDC estimates are com-

pared with those from the full sample and not for the

DCs which is a relatively small group, defined to consist

of the OECD countries. Based on sample means, the

implied marginal effects of GINI and per capita income

(in thousand PPP dollars) are of the following order:

Full

sample

LDCs

Ratio

of

LDCs

to full

sample

GINI

0.56 0.65 1.16

Income per capita (thousand

PPP dollars)

�3.08

�8.12 2.64

It is thus seen that while both income and GINI have

larger effects in the LDCs, the relative effect of GINI is

stronger in the full sample (and therefore in the DCs),

while that of income is stronger in the LDCs.5 Estimated

marginal parameters at the sample means from a simple

version of Rodgers’s model (that includes reciprocal of

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791790

income and GINI) also indicate a similar pattern. These

estimates should, however, be interpreted with caution

since the aggregate relation between health and income

may not faithfully reflect the concavity of the income–-

health relation at the individual level.

Concluding remarks

Careful research by Rodgers (1979), Flegg (1982) and

Waldmann (1992) provided solid evidence of a negative

association between income inequality and population

health in broad cross-country data. However, recent

studies by some scholars, including Mellor and Milyo

(2001), Gravelle et al. (2002), and Wildman et al. (2003),

suggest that the patterns reported by Rodgers, Flegg,

and Waldmann cannot be replicated from recent data

which indicate no significant association between

income inequality and population health in multi-

country contexts. In view of the importance of the issue,

the main purpose of the present study is to undertake a

fresh examination of the cross-country relation between

income inequality and population health by using the

largest possible sample and the most recent information

with a high level of data accuracy. In particular, it

considers whether it is true that the estimates reported

by Rodgers, Flegg, and Waldmann, indicating a

negative association between income inequality and

population health, can or cannot be replicated. The

main conclusion is that, using the largest cross-country

sample and the most recent and accurate data, this study

is able to replicate remarkably well estimates of the kind

reported by the three scholars, and there is consistent

evidence of a negative cross-country association between

income inequality and population health. Possible

reasons for deviations from the patterns reported by

Mellor and Milyo (2001), Gravelle et al. (2002) and

Wildman et al. (2003) were considered at some length.

The most likely reasons seem to be (a) unusual nature of

the Mellor–Milyo model in which income was entered

linearly, (since a Flegg-type log-log model with Mel-

lor–Milyo variables yields a very different scenario), and

(b) samples and data used by Gravelle et al. (2002) and

Wildman et al. (2003) being probably unusual or

atypical. Therefore, the widespread recent skepticism

about the presence of a negative association between

income inequality and population health, articulated by

many scholars, including Lynch and Smith (2002),

Mackenbach (2002), and Deaton (2002, 2003), needs

to be reconsidered. Besides the observation that such

aggregate studies are not well suited to shed light on the

effect of income inequality on health at the individual

level, and that concavity of the income–health relation

at the individual level may not be faithfully reflected in

aggregate data, this study notes seven additional points.

First, when ethnic heterogeneity, which may be

perceived as an omitted variable in such models of

population health, is added, there is very little change in

the income inequality parameters which retain high

statistical significance in the presence of the ethnic-

heterogeneity term. Second, ethnic heterogeneity also

has an adverse association with population health.

Third, when an index of social capital (TRUST), which

may be regarded as a channel through which income

inequality affects health, is included in the models,

income inequality parameters remain almost unchanged,

and retain high statistical significance in almost all

equations. Fourth, despite the sizable simple correla-

tions, the association of TRUST with population health

in the regression models appears weak. Fifth, the

regression patterns observed after the addition of ethnic

heterogeneity or TRUST indicate the estimates to be

quite robust. Sixth, a simple test does not support the

view that the (positive) association between income

inequality and infant mortality in the LDCs is simply a

reflection of the effect of poverty. Last, although there is

an obvious difficulty in capturing in aggregate data

concavity of the individual income–health relation, a

simple indirect test indicates that while both average

income and inequality have stronger marginal para-

meters in health models for the LDCs, the marginal

parameter for income is relatively stronger in the LDCs

while that for income inequality is likely to be relatively

stronger in the DCs.

Acknowledgements

Extremely useful comments on an earlier version from

two anonymous referees are gratefully acknowledged. V.

Cristina Iliuta provided helpful research assistance. The

usual disclaimer applies.

Appendix. List of sample countries (an asterisk indicates

OECD member, and all others are treated as LDCs)

Algeria, Armenia, Australia*, Austria*, Azerbaijan,

Bangladesh, Belarus, Belgium*, Bolivia, Brazil, Bulgar-

ia, Burkina Faso, Burundi, Cambodia, Cameroon,

Canada*, Central African Republic, Chile, China,

Colombia, Costa Rica, Cote d’Ivoire, Croatia, Czech

Republic*, Denmark*, Dominican Republic, Ecuador,

Egypt, El Salvador, Estonia, Ethiopia, Finland*,

France*, Georgia, Germany*, Ghana, Greece*, Guate-

mala, Guinea, Honduras, Hungary*, India, Indonesia,

Ireland*, Israel, Italy*, Jamaica, Japan*, Jordan,

Kazakhstan, Kenya, Korea (Rep.)*, Kyrgyz Republic,

Latvia, Lesotho, Lithuania, Madagascar, Malaysia,

Mali, Mauritania, Mexico*, Moldova, Mongolia, Mor-

occo, Mozambique, Nepal, Netherlands*, Nicaragua,

Niger, Nigeria, Norway*, Pakistan, Panama, Papua

ARTICLE IN PRESSR. Ram / Social Science & Medicine 62 (2006) 779–791 791

New Guinea, Paraguay, Peru, Philippines, Poland*,

Portugal*, Romania, Russian Federation, Rwanda,

Senegal, Sierra Leone, Slovak Republic*, Slovenia,

South Africa, Spain*, Sri Lanka, Sweden*, Switzer-

land*, Tajikistan, Tanzania, Thailand, Tunisia, Tur-

key*, Turkmenistan, Uganda, Ukraine, United

Kingdom*, United States*, Uruguay, Uzbekistan, Ve-

nezuela, Vietnam, Yemen (Rep.), Zambia, Zimbabwe.

References

Alesina, A., Devleeschauwer, A., Easterly, W., Kurlat, S., &

Wacziarg, R. (2003). Fractionalization. Journal of Economic

Growth, 8, 155–194.

Bjornskov, C. (2006). The multiple facets of social capital.

European Journal of Political Economy, forthcoming.

Baumbusch, A. P. (1995). Income inequality and infant

mortality. Junior independent work, Princeton University,

Princeton, NJ.

Deaton, A. (2002). Commentary: The convoluted story of

international studies of inequalities and health. International

Journal of Epidemiology, 31, 546–549.

Deaton, A. (2003). Health, inequality, and economic develop-

ment. Journal of Economic Literature, 41, 113–158.

Deininger, K., & Squire, L. (1996). A new data set measuring

income inequality. World Bank Economic Review, 10,

565–591.

Ellison, G. T. H. (2002). Letting the Gini out of the bottle?

Challenges facing the relative income hypothesis. Social

Science & Medicine, 54, 561–576.

Flegg, A. T. (1982). Inequality of income, illiteracy and medical

care as determinants of infant mortality in underdeveloped

countries. Population Studies, 36(3), 441–458.

Gravelle, H., Wildman, J., & Sutton, M. (2002). Income,

income inequality and health: What can we learn from

aggregate data? Social Science & Medicine, 54, 577–589.

Judge, K., Mulligan, J., & Benzeval, M. (1998). Income

inequality and population health. Social Science & Medi-

cine, 46, 567–579.

Lynch, J., & Smith, G. D. (2002). Commentary: Income

inequality and health: the end of the story? International

Journal of Epidemiology, 31, 549–551.

Lynch, J., Smith, G. D., Harper, S., Hillemeir, M., Ross, N.,

Kaplan, G. A., et al. (2004). Is income inequality a

determinant of population health? Part 1. A systematic

review. The Milbank Quarterly, 82, 5–99.

Mackenbach, J. P. (2002). Income inequality and population

health. British Medical Journal, 324, 1–2.

Mellor, J. M., & Milyo, J. (2001). Reexamining the evidence of

an ecological association between income inequality and

health. Journal of Health Politics, Policy and Law, 26,

487–522.

Preston, S. H. (1975). The changing relation between mortality

and level of economic development. Population Studies,

29(2), 231–248.

Rodgers, G. B. (1979). Income and inequality as determinants

of mortality: An international cross-section analysis. Popu-

lation Studies, 33(2), 343–351.

Subramanian, S. V., & Kawachi, I. (2003a). The association

between state income inequality and worse health is not

confounded by race. International Journal of Epidemiology,

32, 1022–1028.

Subramanian, S. V., & Kawachi, I. (2003b). Response: In

defense of the income inequality hypothesis. International

Journal of Epidemiology, 32, 1037–1040.

Subramanian, S. V., & Kawachi, I. (2004). Income inequality

and health: What have we learned so far? Epidemiologic

Reviews, 26, 78–91.

Summers, R., & Heston, A. (1991). The Penn World Table

(Mark 5): An expanded set of international compar-

isons, 1950–1988. Quarterly Journal of Economics, 106,

327–368.

United Nations. 2001. World population prospects: The 2000

revision (Vol. 1). New York.

Wagstaff, A., & Doorslaer, E. (2000). Income inequality and

health: What does the literature tell us? Annual Review of

Public Health, 21, 543–567.

Waldmann, R. J. (1992). Income distribution and infant

mortality. Quarterly Journal of Economics, 107(4),

1283–1302.

White, H. (1980). A heteroskedasticity-consistent covariance

matrix estimator and a direct test for heteroskedasticity.

Econometrica, 48(4), 817–838.

Wildman, J., Gravelle, H., & Sutton, M. (2003). Health and

income inequality: Attempting to avoid the aggregation

problem. Applied Economics, 35, 999–1004.

Wilkinson, R. G. (1992). Income distribution and life ex-

pectancy. British Medical Journal, 304, 165–168.

Wilkinson, R. G. (1996). Unhealthy societies: The afflictions of

inequality. London: Routledge.

World Bank (2002). World development indicators on CD-ROM.

Washington, DC.

World Bank (2003). World development report 2003. Washing-

ton, DC.