Education, income inequality, and mortality: a multiple regression analysis
BMJ 2002; 324 doi: https://doi.org/10.1136/bmj.324.7328.23 (Published 05 January 2002) Cite this as: BMJ 2002;324:23
All rapid responses
INDIVIDUAL-LEVEL EDUCATION DOES NOT EXPLAIN THE ASSOCIATION OF STATE-LEVEL INCOME INEQUALITY WITH HE
INDIVIDUAL-LEVEL EDUCATION DOES NOT EXPLAIN THE ASSOCIATION OF STATE-
LEVEL INCOME INEQUALITY WITH HEALTH
Muller (2002) has shown in an ecological study that lack of high
school education accounts for the state-level association of income
inequality with mortality in the US.[1] The implicit inference is that
education at the individual-level is responsible for the income inequality
association. However, ecological studies are weak study designs to assess
the independent associations of ecological (income inequality) and
individual-level (education) variables with an individual-level outcome
(health status). In particular, aggregate data are prone to problems of
collinearity rendering it difficult to tease apart independent effects.
Multi-level study designs overcome this limitation by including data at
both the individual and ecological levels.[2 3]
We have previously reported results for the association of income
inequality at the state-level with self-rated health using Current
Population Survey (CPS) data.[4 5] However, we have not previously
reported the effect of controlling for education at the individual-level
among adults in the 1995 and 1997 CPS as shown in the Table below.
Controlling for education attenuated but did not completely explain the
relationship between levels of state income inequality and self-rated
health. Our results do not support the contention that education at the
individual-level fully confounds or mediates the association of income
inequality with health. The remaining portion of the income inequality
association may be due to contextual effects (e.g. average educational
attainment, historical and political features that vary by states in the
US) or other individual-level characteristics such as lifestyle.
Table: Odds ratios of fair/poor self-rated health by quintile of state-level income inequality, with and without adjustment for individual education Quintile of Model 1 Model 2 inequality White males (n=59,341) High 1.22 (1.00-1.50) 1.16 (0.97-1.39) Medium-high 1.44 (1.18-1.76) 1.38 (1.15-1.65) Medium 1.25 (1.02-1.54) 1.20 (0.99-1.45) Medium-low 0.99 (0.79-1.25) 0.98 (0.79-1.22) Low 1.00 1.00 White females (n=62,404) High 1.29 (1.01-1.64) 1.19 (0.95-1.48) Medium-high 1.41 (1.12-1.78) 1.34 (1.08-1.65) Medium 1.40 (1.09-1.78) 1.33 (1.07-1.66) Medium-low 1.11 (0.86-1.45) 1.11 (0.87-1.41) Low 1.00 1.00 Model 1 = Age and income as covariates at the individual-level, and quintile of average income as covariate at state-level Model 2 = Model 1 plus individual-level education
REFERENCES
1. Muller A. Education, income inequality, and mortality: a multiple
regression analysis. BMJ 2002;324:23-25.
2. Diez-Roux A. Bringing context back into epidemiology: variables and
fallacies in multilevel analysis. Am J Public Health 1998;88:216-222.
3. Blakely T, Woodward A. Ecological effects in multi-level studies. J
Epidemiol Community Health 2000;54:367-374.
4. Blakely T, Kennedy B, Kawachi I. Socio-economic inequality in voting
participation and self-rated health. Am J Public Health 2001;91:99-104.
5. Blakely T, Kennedy B, Glass R, Kawachi I. What is the lag time between
income inequality and heath status? J Epidemiol Community Health
2000;54:318-319.
Competing interests: Table: Odds ratios of fair/poor self-rated health by quintile of state-level income inequality, with and without adjustment for individual educationQuintile of Model 1 Model 2inequalityWhite males (n=59,341) High 1.22 (1.00-1.50) 1.16 (0.97-1.39)Medium-high 1.44 (1.18-1.76) 1.38 (1.15-1.65)Medium 1.25 (1.02-1.54) 1.20 (0.99-1.45)Medium-low 0.99 (0.79-1.25) 0.98 (0.79-1.22)Low 1.00 1.00 White females (n=62,404) High 1.29 (1.01-1.64) 1.19 (0.95-1.48)Medium-high 1.41 (1.12-1.78) 1.34 (1.08-1.65)Medium 1.40 (1.09-1.78) 1.33 (1.07-1.66)Medium-low 1.11 (0.86-1.45) 1.11 (0.87-1.41)Low 1.00 1.00Model 1 = Age and income as covariates at the individual-level, and quintile of average income as covariate at state-levelModel 2 = Model 1 plus individual-level education
Makenbach correctly points out that the causal order of income
inequality and educational attainment is debatable in my cross-sectional
analysis. He writes that educational attainment can be viewed as a
“confounder, or an intermediary between income inequality and mortality.”
To shed some light on this issue, I related age-adjusted mortality rates
(1989-90) to Gini coefficients for families in 1969. The lagged income
inequality measure accounts for 10.7% of the variation in age-adjusted
mortality rates. The addition of the educational attainment measure
increased the value to 54.7%. The 44-percentage point increase in
accounted variance suggests that education makes an independent
contribution and appears to be a strong confounder.
Parker raises the question whether educational inequality is
considered in the regression analysis. The answer is: yes in a simple
way. States with low high school completion rates also tend to have low
completion rates for advanced degrees. In an earlier draft of the paper I
included a second educational attainment measure, the percent of persons
18 years of age and older with associate and higher educational degrees.
The variable had the expected negative effect on mortality, however the
measure was statistically insignificant when the analysis was weighted by
population size. Dropping the second education variable did not alter the
main finding of the study.
Jerret cautions that the disappearance of the income inequality
effect may be due to spatial autocorrelation bias. I examined this
possibility by looking at contiguous and non-contiguous states (AL, AR,
IA, KS, KT, ME, MI, MD, NV, NM, NY, ND, SC, WA, WY). A regression model
allowing for different slopes showed no statistically significant
differences between state types for the income inequality, the educational
attainment, and the per capita income measure. This suggests that spatial
autocorrelation may not be a problem, although the issue deserves further
examination.
Howard suggests an unexpected interpretation of the strong education-
mortality effect reported in my paper. He argues that differences in
testosterone levels among racial groups may account for the effect.
Several findings do not agree with his interpretation. A substantial
education-mortality effect is present among females (b=7.05, se=1.42) who
produce substantially lower amounts of testosterone than men. A further
analysis by race indicates that the education-mortality effect is quite
similar for the Caucasian (b=6.47; se=1.11) and the African American
population (b=7.51; se=1.67). The similarity in size of the regression
coefficients suggests that the education-mortality effect is unlikely to
reflect racial differences in testosterone levels.
According to Wilkinson’s comments, the substantial education-
mortality effect represents contextual effects of social status and
dominance that may result in higher rates of violence, and perhaps other
forms of social dysfunction. This prediction may be correct, but will
need to be empirically tested. We should not ignore potentially simpler
explanations of the same phenomenon. Part of the education-mortality
effect might simply be a reflection of conditions typically associated
with low levels of education: unstable employment in low wage jobs,
employment in riskier jobs, inadequate access to health care, smoking,
unhealthy diet, living in areas lacking health promoting resources.
Exposure to these conditions over a lifetime will deplete the stock of
health and shorten life span. There is a need for more detailed analyses
that sort out the relative importance of what may turn out to be different
aspects of a broader explanation.
Competing interests: No competing interests
I suggest another hypothesis may explain "Education, income
inequality, and mortality: a multiple regression analysis," BMJ 2002; 324:
23. Excessive testosterone may be the cause. That is, failing to
graduate high school and shortened life span may be coincident
consequences of high testosterone.
It is my hypothesis that human evolution is driven by increases in
testosterone ("Androgens in Human Evolution. A New Explanation of Human
Evolution," Rivista di Biologia / Biology Forum 2001; 94: 345-362). I
suggest all hominid characteristics are directly influenced by increasing
testosterone. Periodically, however, where circumstances are very
favorable to reproduction, testosterone increases too rapidly. My work
suggests this causes problems in increased infection rates, lowered sperm
counts, etc. These circumstances may have produced the “bottleneck in
hominid populations” which I connect with “Mitochondrial Eve” and
excessive testosterone, which was overcome by mutations controlling sperm
production, which I connect with “Y Chromosome Adam.” (“Mitochondrial Eve,
Y Chromosome Adam, Testosterone and Human Evolution,” accepted for
publication by Rivista di Biologia / Biology Forum.) Populations of high
testosterone individuals sometimes experience very negative consequences.
Within modern populations, I suggest these periodic, rapid increases
in testosterone within populations produce the “secular trend,” which is
real and vigorous at this time in the U.S.A. (“Secular Trends in Height
Among Children During 2 Decades. The Bogalusa Heart Study,” Archives of
Pediatric and Adolescent Medicine. 2000; 154: 155-161). The secular trend
is the increase in size and earlier onset of puberty. The phenomena
Muller examined are current problems, i.e., byproducts of the secular
trend. Testosterone is connected with numerous conditions Muller
describes as “economic resource deprivation, risk of occupational injury,
and learnt risk behaviour.” For example, domestic violence is connected
to testosterone: “Testosterone levels were significantly associated with
levels of both verbal aggression and physical violence self-reported by
the men.” (Psychoneuroendocrinology 2000; 25:721-39), learning
disabilities: “The matched analysis further substantiated the association
between testosterone secretion and learning disabilities. Thus, it is
possible that some learning disabilities may be associated in part with
abnormal testosterone levels.” (Physiol Behav 1993; 53: 583-6), adolescent
smoking: “In this study, salivary testosterone was positively associated
with cigarette smoking among 201 subjects 12-14 years of age.” (J Behav
Med 1989; 12: 425-33), impulsive behavior: “…the results of this study
support a positive relationship between testosterone and impulsivity.”
(Physiol Behav 2001; 73: 217-21), and infection rates: “…male gender was
significantly associated with an increased incidence of pneumonia after
injury…” (J Trauma 2001; 50: 274-80). I suggest the presence of
testosterone also accounts for phenomena described by Dr. Muller.
In the U.S.A. blacks males disproportionately fail to graduate high
school and exhibit “economic resource deprivation, risk of occupational
injury, and learnt risk behaviour.” I suggest this is due to the fact
that black males produce significantly more testosterone than other groups
vis-à-vis females and white males. “Mean testosterone levels in blacks
[males] were 19% higher than in whites [males], and free testosterone
levels were 21% higher. Both these differences were statistically
significant.” (J Natl Cancer Inst 1986; 76: 45-8). Black boys were
affected most by the secular trend in the Bogalusa Heart Study.
I suggest increases in testosterone within certain groups may also
explain Dr. Muller’s findings.
Competing interests: No competing interests
Could the regression analysis be redone including education
inequality a variable?
Competing interests: No competing interests
WHY IS EDUCATION LESS IMPORTANT AT THE INDIVIDUAL LEVEL OF ANALYSIS THAN AT THE STATE LEVEL?
Blakely and Kawachi show in a new analysis of self-reported health
that education measured at the level of individuals does not fully account
for income inequality measured at the state level. Lack of high school
education completely accounted for the income inequality effect in my
study of U.S. states. The discrepancy in findings is not surprising since
our analyses are at different levels of aggregation, use different methods
of analysis, and measure different outcomes among other differences.
More specifically, my paper points out that the large education-
mortality effect in part reflects other factors than education. [1 p3] I
briefly mention the results of an expanded regression analysis that
controls for additional confounders: (1) percent of persons 15 years old
and older married, (2) percent of employed persons 16 years and older in
high injury risk occupations, (3) percent of persons 18 years and older
current smokers, (4) percent of persons without health insurance, (5)
percent African American, or Latino population, and (6) percent of
population foreign born. The additional control variables improve model
fit significantly (R2 adj. = .82) and reduce the direct effect of
education by 40%. However, the education effect remains significant (b =
.066; p = .01) while the income inequality effect continues to be spurious
(b = .54; p = .89). The more detailed analysis was submitted to BMJ, but
not published. The results of the expanded regression analysis are
available upon request.
Blakely and Kawachi’s new data provide only weak support for the
causal contribution of income inequality when education is added as
control. The odds ratios comparisons (model 2) indicate that five out of
eight comparisons are statistically insignificant. In addition, exposure
to medium levels of income inequality results in poorer self-rated health
than exposure to high levels of income inequality. This is an unexpected
result that suggests that some confounding variables may be at work. My
question is: will the income inequality effect disappear with a more
complete model specification? My guess is that it will be further reduced
and the education effect will stay intact. This result would also be
consistent with studies of individuals that found education to be an
important predictor of mortality. [2 3 4]
REFERENCES
1. Muller A. Education, income inequality, and mortality: a multiple
regression analysis. BMJ 2002;324:1-4.
2. Backlund E, Sorlie PD, Johnson NJ. A comparison of the
relationships of education and income with mortality: the national
longitudinal mortality study. Soc Sci Med 1999;49:1373-84.
3. Elo IT, Preston SH. Educational differentials in mortality: United
States, 1979-85. Soc Sci Med 1996;42:47-57.
4. Hoyert DL, Arias E, Smith BL, Murphy SL, Kochanek KD. Deaths:
Final data for 1999. National Vital Statistics Reports, vol. 49 no
8.Hyattsville, Maryland: National Center for Health Statistics. 2001.
Competing interests: No competing interests