Upload
rati-ram
View
216
Download
4
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 61 (2005) 2568–2576
www.elsevier.com/locate/socscimed
Income inequality, poverty, and population health:Evidence from recent data for the United States
Rati Ram�
Economics Department, Illinois State University, Normal, IL 61790-4200, USA
Available online 15 June 2005
Abstract
In this study, state-level US data for the years 2000 and 1990 are used to provide additional evidence on the roles of
income inequality and poverty in population health. Five main points are noted. First, contrary to the suggestion made
in several recent studies, the income inequality parameter is observed to be quite robust and carries statistical
significance in mortality equations estimated from several observation sets and a fairly wide variety of specificational
choices. Second, the evidence does not indicate that significance of income inequality is lost when education variables
are included. Third, similarly, the income inequality parameter shows significance when a race variable is added, and
also when both race and urbanization terms are entered. Fourth, while poverty is seen to have some mortality-
increasing consequence, the role of income inequality appears stronger. Fifth, income inequality retains statistical
significance when a quadratic income term is added and also if the log–log version of a fairly inclusive model is
estimated. I therefore suggest that the recent skepticism articulated by several scholars in regard to the robustness of the
income inequality parameters in mortality equations estimated from the US data should be reconsidered.
r 2005 Elsevier Ltd. All rights reserved.
Keywords: Income inequality; Poverty; Population health; Mortality; USA
Introduction
The relationship between income inequality and population health has received much attention during the last several
years at the theoretical and the empirical levels. Deaton’s (2003) survey explains the various theoretical aspects and
provides an overview of the empirical evidence. Subramanian and Kawachi (2004) provided a fine review of the
multilevel studies.
In the empirical context, evidence for the US has been an important component of the debate, and the conclusions
reached by various scholars have been mixed and sometimes contradictory. For instance, Kaplan, Pamuk, Lynch,
Cohen, and Balfour (1996) showed that, across the US states, income inequality was significantly associated with
adverse health outcomes and mortality trends. Also using state-level data and several measures of income inequality,
Kawachi and Kennedy (1997) noted that the choice of the inequality indicator did not appear to alter the conclusion
that income inequality is linked to high mortality. Similarly, Lynch et al. (1998) concluded that, across metropolitan
statistical areas (MSAs) in the US, higher income inequality is associated with increased mortality at all per-capita
income levels. McLaughlin and Stokes (2002) showed that the relation between income inequality and mortality is
e front matter r 2005 Elsevier Ltd. All rights reserved.
cscimed.2005.04.038
438 7101; fax: +1 309 438 5228.
ess: [email protected].
ARTICLE IN PRESSR. Ram / Social Science & Medicine 61 (2005) 2568–2576 2569
robust for counties in the US. Wolfson, Kaplan, Lynch, Ross, and Backlund (1999) showed that the observed state-level
association between income inequality and mortality cannot be wholly explained as a statistical artifact of an
underlying relation between income and mortality at the individual level. Similarly, multilevel studies by Subramanian
and Kawachi (2003a, b) have indicated that the association between income inequality and poorer health is not
confounded by the effect of race or racial composition.
On the other hand, besides the suggestion that studies at the aggregate level are not useful, or that the aggregate
relation is a statistical ‘‘artifact’’ of the nonlinear income–health nexus at the individual level, soundness of the
conclusions relating income inequality with higher mortality has been econometrically questioned. The main point
made in these critiques is that, while there is a negative correlation between income inequality and aggregate health
indicators, what is being observed is not the effect of income inequality, but of some other variables (confounders) for
which income inequality serves as a proxy. Two variables that have been particularly emphasized in this context are
education and race or racial composition of the population. For example, Mellor and Milyo (2001, p. 508) stated that if
controls are included for educational attainment and race (and urbanization), income inequality is no longer associated
with death rates in their pooled sample of 48 continental states. While Kawachi and Blakely (2001), besides others,
critiqued that view, Deaton (2003, p. 143) cited the Mellor–Milyo study and observed that inclusion of controls for the
average level of education in each state eliminates the significance of the Gini coefficient, and that once controls are
added for the fraction of people in each state who are urbanized and who are black, ‘‘the Gini coefficient attracts a
negative sign.’’ Similarly, Deaton and Lubotsky (2003, p. 1139) stated that ‘‘Conditional on the fraction black, neither
city nor state mortality rates are correlated with income inequality,’’ and that view has been expressed by Deaton quite
emphatically and repeatedly. In fact, after saying that the ‘‘city and state relationship between inequality and mortality
in the USydisappears once we control for racial composition,’’ and ‘‘Across states in 1990, the relationship also
disappears once we control for education,’’ Deaton (2002, p. 548) stated that ‘‘As of the time of writing, little appears to
remain of the whole enterprise.’’ Some of the observations by Deaton (2002) and Deaton and Lubotsky (2003) have
been prominently highlighted by Lynch, Harper, and Smith (2003). Similar sentiments have also been articulated by
some other scholars. For instance, Mackenbach (2002) noted that ‘‘most of the correlation between income inequality
and mortality at the aggregate level in the United States can be explained away by differences in average levels of formal
education.’’
This paper makes a modest contribution on the subject by working with state-level data for the year 2000, which no
scholar has apparently considered so far, along with that for 1990, so as to generate a recent 2-period panel of the kind
used by Mellor and Milyo (2001). The main focus is on considering whether, in models of (overall) mortality, statistical
significance of income inequality is lost when (a) education variables are added, (b) fraction of the population that is
black is included, (c) an urbanization term is introduced, and (d) various combinations of these variables are entered in
the model. An effort is also made to judge whether income inequality may simply reflect the effect of poverty.
Moreover, significance of the income inequality term is studied after adding a quadratic income term and in log–log
form of a fairly inclusive model.
Data and model
The study starts with state-level data for the year 2000 and supplements these with the information for 1990 so as to
generate a recent 2-period panel. To facilitate comparisons with earlier studies, a simple linear model of the kind used
by Mellor and Milyo (2001) is specified. The basic form of the model may be written as
DRit ¼ a0 þ a1GINIit þ a2INCOMEit þ a3HSit þ a4COLLEGEþ a5BLACKit þ a6URBANit þ uit, (1)
where DRit denotes overall death rate in state i and year t, GINI is household Gini coefficient, INCOME stands for
personal income per capita, HS and COLLEGE indicate the percentage of population (25+) that has at least a high
school diploma or college degree, BLACK denotes the proportion of population that is black, URBAN is the
percentage of urban population, and u is a standard stochastic term. In addition to the variables shown in Eq. (1),
percentages of population under 18 and over 65 are included as controls for age composition, and percentage of
population in age-group 18–65 serves as the ‘‘base’’. In pooled samples, an intercept dummy for the year 1990 is also
included.
A major shortcoming of such studies that work with aggregate data should be noted. One crucial issue sought to be
addressed in such research is whether income inequality at the aggregate level has an independent ‘‘contextual’’ effect on
health at the individual level. A proper investigation of that question requires a multilevel research design in which
income inequality at the aggregate (community) level enters along with individual characteristics in a model of
individual health. Such multilevel studies have been conducted by several scholars, including Subramanian and
ARTICLE IN PRESS
Table 1
Variable definitions, data sources, and descriptive statistics
Variable Mean SD Min Max N
Death rate (per 100,000) 863.23 134.88 400.00 1200.00 102
Gini (0–1 scale) 0.44 0.03 0.38 0.51 102
Percent below poverty line 12.07 3.85 5.20 25.70 102
Personal income per capita 24.05 4.48 15.36 37.85 102
(thousand 1996 dollars)
Percent with completed 79.05 5.74 64.30 88.30 102
high school or more
Percent with college 22.05 4.89 12.30 39.10 102
degree or more
Percent black 10.82 11.90 0.25 66.23 102
Percent urban 70.53 15.26 32.20 100.00 102
Percent under 18 25.90 2.32 19.50 36.50 102
Percent over 65 12.48 1.99 4.00 18.20 102
Data sources: Crude death rate (DR) 1990: Statistical abstract of the United States: 2002, p. 77 (US Census Bureau, 2002). 2000:
National Vital Statistics Report, March 14, 2003, p. 19 (US Department of Health and Human Services & National Center for Health
Statistics, 2003).
Percent below poverty line (POOR) 1990: State and metropolitan area data book, 1998, p. 25 (US Census Bureau, 1998). 2000: US
Bureau of Labor Statistics and Bureau of the Census (2001).
Personal per capita income (INCOME)(1996 dollars): 1990 and 2000, Statistical abstract of the United States: 2002, p. 426 (US Census
Bureau, 2002).
Household Gini (GINI) 1990: US Census Bureau b. 2000: US Census Bureau c.
Percent (25 or over) with high school or more (HS) 1990: US Department of Education, National Center for Education Statistics
(2002, p. 20). 2000: Statistical Abstract: 2002, p. 865.
Percent (25 or over) with college degree or more (COLLEGE) 1990: US Department of Education, National Center for Education
Statistics (2002, p. 20). 2000: Statistical Abstract: 2002, p. 865.
Percent urban (URBAN) 1990: Census website, Urban and Rural Population 1900 to 1990 (US Census Bureau a). 2000: Statistical
Abstract: 2002, p. 861
Percent black (BLACK) 1990: Compiled from total and black population data on pp. 2, 6 of State and Metropolitan Area Data Book,
1998 (US Census Bureau, 1998). 2000: Statistical Abstract: 2002, p. 27.
Percent below 18 and over 65 (18�, 65+) 1990: State and Metropolitan Area Data Book, p. 4. 2000: Based on Statistical Abstract:
2002, p. 25.
R. Ram / Social Science & Medicine 61 (2005) 2568–25762570
Kawachi (2003a, b), and have been ably reviewed by Subramanian and Kawachi (2004). Despite the aforesaid limitation
of the present work, it may be useful in several ways. First, it uses more recent data to revisit the research done at the
aggregate level, which has been highly influential and has suggested that statistical significance of income inequality in
models of population health goes away when a race variable or an educational attainment term or both are included in
the model. Second, even if such studies may not accurately reveal the contextual effect of income inequality on
individual health, these can indicate the partial association between income inequality and population health (holding
average income, education, race, and urbanization constant), which should be of considerable relevance to income and
distributional policies.
Table 1 provides descriptive statistics for the sample, defines each variable, and lists the data sources.1 It is perhaps
useful to note that, in terms of coefficient of variation, income inequality has the smallest variance across states while
percent-black has the highest dispersion. The poverty ratio has the second highest dispersion, followed by urbanization
and college education. Therefore, from a simple econometric perspective, parameter estimates for GINI may be
expected to be less precise, and thus show lower significance, than those for percent-black, poverty ratio or
urbanization.
Table 2 shows simple correlations across the variables. While the patterns are of the expected kind, it may be
interesting to see that, apart from the expected high covariation between the age-composition variables and the sizable
1Note that household GINI for 1990 is computed by the Bureau of Census from census data (US Census Bureau b), and that for
2000 is computed by the Bureau from CPS. Personal income per capita is taken as a measure of average income because it is a better
indicator than household income of the potential for private and public health outlays in the state.
ARTICLE IN PRESS
Table 2
Simple correlations across the main variables
DR GINI POOR INCOME HS COLLEGE BLACK URBAN 18�
GINI 0.42
POOR 0.28 0.38
INCOME �0.11 0.30 �0.55
HS �0.45 �0.31 �0.70 0.49
COLLEGE �0.27 0.23 �0.43 0.83 0.58
BLACK 0.38 0.55 0.40 0.16 �0.49 0.13
URBAN �0.25 0.20 �0.23 0.61 0.20 0.54 0.18
18� �0.62 �0.36 0.15 �0.46 0.12 �0.28 �0.31 �0.14
65+ 0.80 0.15 0.01 �0.08 �0.18 �0.21 �0.07 �0.14 �0.56
See Table 1 for variable notations and definitions.
R. Ram / Social Science & Medicine 61 (2005) 2568–2576 2571
correlation between income and college, most correlations are relatively modest. As an aside, it might be noted that the
correlation between mean income and GINI is a moderate positive number, although Deaton (2003, p. 142) had
indicated that across the US states, income inequality is strongly negatively correlated with income.
Results
Table 3 shows the basic results from estimating Eq. (1) by the ordinary least-squares procedure (OLS). The estimates
are shown for pooled data for 1990 and 2000 for the 51 states (N ¼ 102), for each year separately, and pooled data for
the 48 contiguous states. The subsample of contiguous states is considered to facilitate comparisons with Mellor
and Milyo (2001, p. 508) and also because Deaton (2003, p. 142) indicated that the District of Columbia is an outlier.
Table 4 reports pooled-sample OLS estimates with several combinations of the variables. The combinations are not
exhaustive, but include the ones that seem most relevant to the issues addressed. The following points may be noted
from Tables 3 and 4.
1.
The explanatory power of the specification is good, and around 90% of the variance in mortality across states isexplained by the full model.
2.
Estimates for the years 1990 and 2000 in Table 3 are similar. It is, therefore, reasonable to work with pooled samples.3.
The overall pattern of estimates is plausible, and most parameters have the expected sign and significance.4.
The pattern of estimates for the 51-state pooled sample is similar to that for 48 contiguous states. Therefore, thediscussion focuses mainly on the 51-state data.
5.
The most important point indicated by the tables is that the income inequality parameter is robust across the variedsamples and specifications. It shows statistical significance at the conventional levels for 1990 as well as 2000, and in
the pooled sample of contiguous states as well as that for 51 states. The parameter is, of course, highly significant by
itself, and remains so when entered with (a) income alone, (b) income and education, (c) income and proportion-
black, (d) income, proportion-black, and urbanization, and (d) income, education, proportion-black, and
urbanization. It is, therefore, not true that statistical significance of income inequality is lost when education is
entered in the model or when race or racial composition is accounted for. Moreover, it may be useful to recall that
the robust significance of GINI is observed despite a low dispersion of the variable in the sample.
6.
While Table 3 shows robustness of GINI-parameter in full specification, Table 4 reveals sturdiness of incomeinequality in several different combinations of the variables, and it seems useful to state more explicitly some of the
details indicated by the latter table.
First, column (3) in Table 4 shows that when the high-school variable is added, both GINI and mean income remain
highly significant and high-school education also shows significance in the expected direction.
Second, column (4) in the table shows that, far from rendering income inequality insignificant, inclusion of the
education variables (HS and COLLEGE) is associated with a large GINI parameter that is highly significant, and it is
ARTICLE IN PRESS
Table 3
Income inequality and death rates: state-level data for the US
Regressor Dependent variable is number of deaths per 100,000 population
All 51 states 48 contiguous states:
pooled
Mellor and Milyo (2001, p.
508 col. 5) and estimates
Pooled 1990 2000
(1) (2) (3) (4) (5)
GINI 856.04* 1095.90* 806.66* 513.82* �66.60
(3.30) (2.22) (2.13) (2.26) (�0.30)
INCOME 0.54 1.14 2.11 1.95 0.0013
(0.24) (0.26) (0.69) (0.91) (1.03)
HS 2.33 2.34 1.07 �0.04 �1.92*
(1.34) (0.88) (0.35) (�0.03) (�2.21)
COLLEGE �7.16* �4.40 �8.61* �8.56* �3.78*
(�3.74) (�1.54) (�2.98) (�4.79) (�2.07)
BLACK 4.30* 4.28* 3.91* 2.68* 2.07*
(8.05) (5.19) (4.99) (4.94) (4.46)
URBAN �1.58* �1.90* �1.68* �1.39* �0.76*
(�4.17) (�3.02) (�2.68) (�4.10) (�2.42)
Adj. R2 0.90 0.91 0.89 0.90 0.83
N 102 51 51 96 240
Notes: Relevant t-statistics are in parentheses. All equations include controls for proportion of population below 18 and over 65.
Estimates from pooled observations include a control dummy for year 1990 also. Please see Table 1 for variable notations and
definitions.
* Indicates statistical significance at least at the 5% level.
Table 4
Income inequality and death rates, pooled 1990 and 2000 data for 51 states: estimates for several combinations of variables
Regressor Dependent variable is number of deaths per 100,000 population
(1) (2) (3) (4) (5) (6)
GINI 1552.42* 1526.94* 916.81* 1123.04* 574.90* 658.51*
(5.38) (6.05) (2.81) (3.24) (2.31) (2.86)
INCOME �10.63* �6.76* �4.32 �9.08* �3.90*
(�5.54) (�2.93) (�1.59) (�5.70) (�2.04)
HS �5.07* �3.40+
(�2.81) (�1.66)
COLLEGE �4.12+
(�1.66)
BLACK 3.73* 3.98*
(6.88) (7.91)
URBAN �1.70*
(�4.24)
Adj. R2 0.75 0.81 0.82 0.82 0.87 0.89
N 102 102 102 102 102 102
Notes: Relevant t-statistics are in parentheses. All equations include controls for proportion of population below 18 and over 65 and an
intercept dummy for year 1990. Please see Table 1 for variable notations and definitions.
* Indicates statistical significance at least at the 5% level, and a plus sign (+) indicates significance at the 10% level.
R. Ram / Social Science & Medicine 61 (2005) 2568–25762572
the education variables that show marginal significance at the 10% level. The income parameter, however, loses
significance when both education variables are entered.
Third, when proportion-black is included along with GINI and income, the income inequality variable remains
significant.
ARTICLE IN PRESSR. Ram / Social Science & Medicine 61 (2005) 2568–2576 2573
Fourth, GINI retains significance when, along with income, both proportion black and urbanization are included.
Table 3 already shows that GINI retains significance in all samples when income, education, racial composition, and
urbanization variables are entered together.
Therefore, it is reasonable to say that the preponderance of evidence from recent data shows that increased income
inequality is associated with a higher mortality rate. At any rate, evidence supporting the hypothesis of a significant
adverse association between income inequality and mortality and population health seems no less robust than evidence
for the proposition that there is no such association.
Discussion
Tables 3 and 4 indicate that income inequality has a robust association with increased mortality in recent state-level
data for the US even after mean income, education, race, and urbanization variables are included in the model. It seems
useful, however, to discuss several points that are related to these estimates.
First, Deaton (2003, pp. 114, 151) suggested that the observed correlation between income inequality and mortality
might be a reflection of the effect of poverty on health. Given the availability of reasonably comparable data on poverty
for the US states, it is possible to investigate that possibility which Deaton (2003, p. 114) termed as the poverty
hypothesis. Table 5 contains estimates that are relevant to Deaton’s hypothesis. Columns (1) and (4) in the table
replicate columns (1) and (4), respectively, of Table 3. Other columns in Table 5 show the estimates when poverty is
entered in place of or in addition to income inequality. It is seen that income inequality retains statistical significance in
all cases, both with and without poverty included in the model. Also, in the presence of income inequality, poverty does
not show significance at any sensible level in either sample. Moreover, even when income inequality is excluded, poverty
is seen to be significant at the 10% level in the 51-state sample, but is not significant at any meaningful level in the
sample of contiguous states. Reflecting the low significance of the poverty variable, the pattern of estimates for other
variables is the same whether poverty is included or not. Therefore, it is reasonable to say that while poverty does seem
to have some mortality-increasing consequence in the 51-state sample, the role of income inequality appears stronger,
and it is unlikely that the income inequality parameters reflect largely or substantially the effect of poverty.
Table 5
Comparing roles of income inequality and poverty in mortality: pooled state-level US data, 1990 and 2000
Regressor Dependent variable is number of deaths per 100,000 population
All 51 states pooled 48 contiguous states pooled
(1) (2) (3) (4) (5) (6)
GINI 856.04* — 787.48* 513.82* — 537.14*
(3.30) (2.68) (2.26) (2.05)
POOR — 3.65+ 1.05 — 1.59 �0.35
(1.91) (0.50) (0.93) (�0.18)
INCOME 0.54 3.30 1.20 1.95 3.00 1.75
(0.24) (1.26) (0.45) (0.91) (1.27) (0.73)
HS 2.33 0.69 2.52 �0.04 �1.19 �0.13
(1.34) (0.41) (1.41) (�0.03) �0.78) (�0.08)
COLLEGE �7.16* �6.38* �7.33* �8.56* �7.84* �8.52*
(�3.74) (�3.22) (�3.75) (�4.79) (�4.32) (�4.70)
BLACK 4.30* 4.34* 4.26* 2.68* 2.75* 2.68*
(8.05) (7.74) (7.84) (4.94) (4.95) (4.90)
URBAN �1.58* �1.62* �1.60* �1.39* �1.35* �1.39*
(�4.17) (�4.09) (�4.18) (�3.55) (�3.89) (�4.07)
Adj. R2 0.90 0.90 0.90 0.90 0.90 0.90
N 102 102 102 96 96 96
Notes: Relevant t-statistics are in parentheses. All equations include controls for proportion of population below 18 and over 65 and a
dummy for year 1990. Please see Table 1 for variable notations and definitions.
Note that columns (1) and (4) replicate the estimates in columns (1) and (4) of Table 3.
* Indicates statistical significance at least at the 5% level, and a plus sign (+) indicates significance at the 10% level.
ARTICLE IN PRESSR. Ram / Social Science & Medicine 61 (2005) 2568–25762574
Second, it might be thought that due to the nonlinear relation between income and health at the individual level, the
state-level model should have a quadratic income term in addition to income inequality. However, if an income-square
term is added to the model specified in Eq. (1), the pattern of estimates, particularly the income inequality parameter, is
almost identical with that without the quadratic-income term.2
Third, it may appear that a log–log model, with elasticities as the parameters, would be more appropriate than the
linear model of Eq. (1). However, the pattern of estimates is almost the same if logarithms of the main variables are used
in place of the levels.3
Fourth, besides the linear model in levels, Mellor and Milyo (2001, pp. 510–511) also worked with specifications
based on 10-year and 20-year changes in the variables. However, as Deaton (2003, pp. 143–144) observed, this is almost
certainly too severe a test because it places a great deal of weight on the timing of the link between income inequality
and mortality.
Fifth, mortality is represented by crude death rate which is the number of deaths per 100,000 population in the year.
Such rates are not strictly comparable across states because these depend to some extent on the age composition of the
population which varies across the states. It is possible to estimate age-adjusted death rates that take into consideration
age composition of the population in each state. National Center for Health Statistics of the US Department of Health
and Human Services has been reporting age-adjusted death rates also for the last several years (US Department of
Health and Human Services, National Center for Health Statistics, 2003). However, three considerations make it
desirable to use the unadjusted rates for this study. First, while age-adjusted death rates are available for the year 2000,
these do not seem to have been published for 1990. Second, Rosenbaum and Rubin (1984) have shown that it may be
problematic to use age-adjusted dependent variables when the regressors are not age-adjusted. They suggested that it
would be better to use unadjusted values of the dependent variable along with appropriate age controls. Third, use of
unadjusted death rates enhances the comparability of the present study with that of Mellor and Milyo (2001) who
worked with unadjusted death rates along with age-composition controls. At any rate, a sensitivity test was conducted
by comparing estimates for the sample for 2000 from models of crude and age-adjusted mortality rates, and it was
found that the two sets of estimates were almost identical. In particular, size and significance of the income inequality
parameters are very similar in the two sets.4
Last, it is important to consider why the pattern of income inequality parameters in this study differs so dramatically
from that reported by Mellor and Milyo (2001) from their 5-period (1950–1990) panel data for the 48 continental states.
To provide a direct comparison of estimates for the most complete models, column (5) in Table 3 reproduces the
estimates from column (5) of their Table 7 (p. 508), which are comparable with the present study’s estimates in column
(4) of Table 3. It may be seen that most estimates in the two sets are fairly similar, and the most glaring difference is in
the parameters for the income inequality variable which is the primary focus of both studies.5
Although models underlying columns (4) and (5) of Table 3 are identical, and both sets represent OLS estimates,
there are several methodological differences between the two studies, and it is useful to consider whether any of these
could account for the large divergence between the income inequality parameters in the two sets.
The GINI variable used by Mellor and Milyo (2001, p. 506) is for family income while that used in the present study
is for household income which seems more appropriate. However, it is unlikely that the difference in the GINI
parameters is due to the slight variation in the variables which are very close to each other. Similarly, the income
variable used by Mellor and Milyo is median family income (in 1992 dollars) while that used in this study is personal
income per capita (in thousands of 1996 dollars), which, again, seems more appropriate for a model of state-level
mortality. However, apart from the scale, the two variables should yield very similar estimates.
2If income-square term is added to the model of column (1) of Table 3, the GINI parameter is 822.25 with a t-statistic of 3.17, which
is almost identical with the estimate of 856.04 (t ¼ 3:30) without the quadratic income term.3For example, estimates for 51-state pooled sample are given below. These correspond to column (1) of Table 3 and include the
constant term. Relevant t-statistics are in parentheses, and ln denotes natural logarithm of the variable.
lnðDRÞ ¼ 7:158 þ0:484ðlnGINIÞ �0:102 ðlnINCOMEÞ þ0:295 ðlnHSÞ
ð9:93Þ ð3:42Þ ð�1:31Þ ð1:66Þ
�0:098 ðlnCOLLEGEÞ þ0:040 ðlnBLACKÞ �0:175 ðlnURBANÞ
ð�1:92Þ ð5:78Þ ð�5:21Þ4For example, Gini parameter in regressions of age-adjusted death rate for the year 2000 is 672.55 with a t-statistic of 2.11, compared
with 806.66 (t ¼ 2:13) of column (3) in Table 3 where the dependent variable is crude death rate.5The difference in the income parameters probably reflects scaling, and the variations in the HS parameters may be due to the huge
changes in the educational attainment of the population between 1950 and 1990.
ARTICLE IN PRESSR. Ram / Social Science & Medicine 61 (2005) 2568–2576 2575
While the dependent variable in both studies is crude death rate, Mellor and Milyo used four age-composition
controls, while only two (18� and 65+) are included in the present study. However, the estimates seem almost identical
whether four age-composition controls are included or only two are used.6
Finally, although both studies obtained parameter estimates by the OLS procedure, the usual t-statistics are shown in
Tables 3, 4 and 5, but those reported by Mellor and Milyo were based on White’s (1980) heteroscedasticity-consistent
standard errors. These standard errors are consistent if the error term is not homoscedastic, but are less efficient than
the usual standard errors if the errors are homoscedastic. Although use of White’s standard errors might be more
appropriate in cross-country data where there is a huge dispersion in the mortality rates (and other health indicators),
the pattern of income inequality parameters seems almost identical whether the standard errors used are the usual ones
or White’s (1980).7
Based on the above, it seems that none of the methodological differences are likely to explain the strong contrast
between the income inequality parameters reported in the two studies, which are reflected in columns (4) and (5) of
Table 3. My best guess is that, aside from the possibility of computational errors, it is possible that pooling observations
across such distant years as 1950 and 1990 by Mellor and Milyo obscures important parametric changes and generates a
somewhat distorted indication of the relation between income inequality and population health in recent decades. As is
well known, income inequality patterns in the US have undergone dramatic changes between 1950 and 1990, and the
year-dummies included by Mellor and Milyo apparently cannot capture the parametric variations related to these
changes.
Concluding observations
This study did not seek to deal either with profound issues concerning the channels through which income inequality
might affect population health or with an empirical exploration of such channels. In particular, it does not consider the
recent themes that income inequality may erode social cohesion (or social capital) or that other types of inequalities
might be more important than income disparities. It had only a modest objective. It used the latest state-level
information to reexamine recent skepticism about the econometric robustness of income inequality parameters in
mortality models estimated from US data. It considers whether, as suggested by Mellor and Milyo (2001), Deaton
(2002, 2003), Deaton and Lubotsky (2003), and other scholars, the income inequality parameter loses significance when
education variables are entered or when race, racial composition, or urbanization terms are included in the model. The
investigation is done by using state-level panel data for the years 1990 and 2000. The main points indicated by the study
may be summarized in five statements. First, contrary to what has been suggested by the aforesaid scholars, the income
inequality parameter seems robust and shows high statistical significance in several different sets of observations and a
wide variety of specificational choices. Second, there is no indication that income inequality loses significance when
education variables are included; on the contrary, the size and the significance of the inequality parameter remain high
when high-school and college-education variables are added. Third, similarly, the evidence does not support the view
that inclusion of race or a racial-composition variable renders the income inequality term insignificant or that inclusion
of both racial-composition and urbanization terms causes income inequality to lose significance. Fourth, although
poverty does show some mortality-increasing consequence, the role of income inequality appears stronger, and it is
unlikely that the correlation between income inequality and mortality largely reflects the effect of poverty on health.
Fifth, significance of income inequality is robust to the addition of a quadratic income term and also to a log–log
specification of the Mellor–Milyo linear model. These results deserve at least as much consideration as those on which
the recent skepticism about the role of income inequality is based. The evidence supporting the hypothesis of a
significant adverse association of income inequality with mortality and population health appears no less robust than
that in favor of the view that there is no such association.
6For example, GINI parameter in the full regression for 2000, with four age controls of the Mellor–Milyo type, is 859.45 with a t-
statistics of 2.32 compared to 806.66 (t ¼ 2:13) of column (3) in Table 3 based on two age controls.7For instance, the following is a comparison of the reported t-statistics with those based on White’s consistent standard errors for the
GINI parameters reported in the six columns of Table 4. Like the reported t-statistics, each t-statistic based on White’s standard error
is significant at least at the 5% level.
(1) (2) (3) (4) (5) (6)
Usual t-statistics reported in Table 4 5.38 6.05 2.81 3.24 2.31 2.86
t-statistics based on White’s SEs 4.56 4.77 2.02 2.54 2.09 2.15
ARTICLE IN PRESSR. Ram / Social Science & Medicine 61 (2005) 2568–25762576
Acknowledgements
The Senior Editor (Professor Ichiro Kawachi) and two anonymous reviewers of this journal gave many useful
comments on an earlier version. V. Cristina Iliuta provided helpful research assistance. The usual disclaimer, however,
applies.
References
Deaton, A. (2002). Commentary: The convoluted story of international studies of inequality and health. International Journal of
Epidemiology, 31, 546–549.
Deaton, A. (2003). Health, inequality, and economic development. Journal of Economic Literature, 41, 113–158.
Deaton, A., & Lubotsky, D. (2003). Mortality, inequality and race in American cities and states. Social Science & Medicine, 56,
1139–1153.
Kaplan, G. A., Pamuk, E. R., Lynch, J. W., Cohen, R. D., & Balfour, J. L. (1996). Inequality in income and mortality in the United
States: Analysis of mortality and potential pathways. British Medical Journal, 312, 999–1003.
Kawachi, I., & Blakely, T. A. (2001). When economists and epidemiologists disagreey. Journal of Health Politics, Policy and Law,
26(3), 533–541.
Kawachi, I., & Kennedy, B. P. (1997). The relationship of income inequality to mortality: Does the choice of indicator matter? Social
Science & Medicine, 45(7), 1121–1127.
Lynch, J., Harper, S., & Smith, G. D. (2003). Commentary: Plugging leaks and repelling boarders—where to next for SS income
inequality? International Journal of Epidemiology, 32, 1029–1036.
Lynch, J. W., Kaplan, G. A., Pamuk, E. R., Cohen, R. D., Heck, K. E., Balfour, J. L., & Yen, I. H. (1998). Income inequality and
mortality in metropolitan areas of the United States. American Journal of Public Health, 88(7), 1074–1080.
Mackenbach, J. P. (2002). Income inequality and population health. British Medical Journal, 324, 1–2.
McLaughlin, D. K., & Stokes, C. S. (2002). Income inequality and mortality in US counties: Does minority racial concentration
matter? American Journal of Public Health, 92(1), 99–104.
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(3), 487–522.
Rosenbaum, P. R., & Rubin, D. B. (1984). Difficulties with regression analyses of age-adjusted rates. Biometrics, 40, 437–443.
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 defence 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.
US Bureau of Labor Statistics and Bureau of the Census (2001). Table S4: Detailed poverty tables: at url http://ferret.bls.census.gov/
macro/032000/pov/new25_001.htm.
US Census Bureau a. Urban and rural population 1900 to 1990: at url http://www.census.gov/population/censusdata/urpop0090/txt.
US Census Bureau (1998). State and Metropolitan Area Data Book 1997–98 (5th ed.). Washington, DC.
US Census Bureau (2002). Statistical Abstract of the United States: 2002 (122nd ed.). Washington, DC.
US Census Bureau b. Gini ratios by state: At the Bureau’s web site for historical census data.
US Census Bureau c. Income distribution data by state based on CPS: personal correspondence.
US Department of Education, National Center for Education Statistics (2002). Digest of Education Statistics, 2001. Washington, DC.
US Department of Health and Human Services, National Center for Health Statistics (2003). National Vital Statistics Reports, Vol.
51(5). Hyattsville, MD.
White, H. (1980). A heteroskedasticity-consistent covariance matrix estimator and a direct test for heteroskedasticity. Econometrica,
48(4), 817–838.
Wolfson, M., Kaplan, G., Lynch, J., Ross, N., & Backlund, E. (1999). Relation between income inequality and mortality: Empirical
demonstration. British Medical Journal, 319, 953–955.