 RESEARCH AND PRACTICE 
Variability and Vulnerability at the Ecological Level: 
Implications for Understanding 
the Social Determinants of Health
| Adam Karpati, MD, MPH, Sandro Galea, MD, MPH, Tamara Awerbuch, PhD, and Richard Levins, PhD
Recent research into the role of the social en-
Objectives. We examined variability in disease rates to gain understanding of the
vironment as a determinant of individual complex interactions between contextual socioeconomic factors and health.
health has reinvigorated inquiry into the rela- Methods. We compared mortality rates between New York and California counties in
tion between context and health.1–3 Questions the lowest and highest quartiles of socioeconomic status (SES), assessed rate variability
regarding mechanism follow naturally from between counties for various outcomes, and examined correlations between outcomes’
this work. A key aspect of many contextual sensitivity to SES and their variability.
variables is that they cannot be measured at Results. Outcomes with mortality rates that differed most by county SES were among
the individual level; they are essentially group, those whose variability across counties was high (e.g., AIDS, homicide, cirrhosis). Lower-
or ecological, characteristics. Contextual fac- SES counties manifested greater variability among outcome measures.
tors likely interact with the large number of Conclusions. Differences in health outcome variability reflect differences in SES im-
individual characteristics that determine pact on health. Health variability at the ecological level might reflect the impact of
stressors on vulnerable populations. (Am J Public Health. 2002;92:1768–1772)
health and illness, such as genetics, behavioral
choices, and access to medical care.
Analyses of community factors attempt to after myocardial infarction.12 For public states in the United States and their county
elucidate how context affects the health of in- health surveillance or for epidemiological mortality rates are based on relatively large
dividuals.4 Although multilevel analysis al- analysis, variability in population health or its denominators. We excluded from the analysis
lows statistical determination of the relative determinants may be a more informative any county in either state with a population
effect of individual and community factors,5 characteristic than the absolute level of par- of less than 15000 persons.
the manner in which these measures exert ticular components. Similarly, for policy or
their effects on public health is likely to be program evaluation, variability might be a Data
more complex than is suggested by general- useful measure of the relative effects of dif- We used New York State Department of
ized multilevel linear models.6,7 A more accu- ferent interventions. Health and California Department of Health
rate understanding of the interplay between We studied the relation between contextual Services data to obtain age-adjusted mortality
individuals and their environments requires effects and population health outcomes by ex- rates in each county for the various out-
construction of models that take into account amining mortality rates associated with sev- comes.13,14 Table 1 presents the mortality
our knowledge of interactions on various lev- eral conditions in counties in New York and rates for the outcomes studied. We selected
els, contextual and otherwise, and the fact California. We hypothesized that, first, certain outcomes on the basis of data availability,
that system components are interconnected diseases or health outcomes (e.g., traumatic range of clinical conditions, and consistency
and likely display feedback loops.8–10 events or communicable diseases) are more with previously published studies. Rates for
One approach to understanding complex sensitive to population socioeconomic factors New York were from 1997; rates for Califor-
systems is to examine variability among their than are others, reflecting the degree to which nia were either from 1997 alone or were an
components. Variability refers to the extent to those outcomes are avoidable or preventable. average of rates from 1995 to 1997.
which a characteristic of a complex system Second, the rates of the outcomes that are We obtained the following measures of
(e.g., heart rate or stock prices) changes over most sensitive to socioeconomic factors also county socioeconomic status (SES) from the
time or space. Variability in a complex system vary the most among counties, reflecting the US Census Bureau: unemployment rate
might reflect the effect of external influences wide distribution of responses to the stressors (1997), percentage living in poverty (1995),
(“stressors”) through their interaction with the to which populations are exposed. percentage of children aged less than 18
system’s homeostatic mechanisms.11 years living in poverty (1995), median house-
Most evaluations of variability in complex METHODS hold income (1995), and high school gradua-
physiological systems have been done in the tion rate (1990).15 Percentage living in pov-
context of individual clinical characteristics. The unit of analysis for this study was the erty is defined as the percentage of
For example, a decrease in heart rate vari- county. We selected New York and California households under the federal poverty thresh-
ability has been shown to predict mortality because they are among the most populous old, adjusted for family size and composi-
1768 | Research and Practice | Peer Reviewed | Karpati et al. American Journal of Public Health | November 2002, Vol 92, No. 11
 RESEARCH AND PRACTICE 
TABLE 1—Mortality Rates from Various Causes in New York and California, 1997.a
New York California
Outcome Mean Range Range / Mean Mean Range Range / Mean
All-cause mortality (all ages) 829 274 0.3 453.8 228.5 0.5
All-cause mortality (≥ 75 y) 21 655 9070 0.4 19 284.2 8714 0.5
All-cause mortality (10–24 y)b 120 310 2.6 75.0 155.1 2.1
Infant mortality 6.7 12.5 1.9 7.0 12.3 1.8
AIDS 4.9 53.7 10.9 3.4 24.6 7.3
Pneumoniac 34.4 65.1 1.9 15.8 22.3 1.4
COPD 42.7 58.3 1.4 23.5 29.6 1.3
Cardiovascular disease 286 234 0.8 89.5 75.0 0.8
Stroke 47.9 64.0 1.6 26.4 22.6 0.9
All neoplastic disease 204 107 0.5 118.8 56.1 0.5
Lung cancer n/a n/a n/a 34.9 32.3 0.9
Female breast cancer n/a n/a n/a 19.1 19.1 1.0
Cirrhosis 8.0 15.6 2.0 10.1 24.7 2.4
Accidentsd 29.2 33.8 1.2 18.5 29.5 1.6
Homicide 2.6 16.7 6.3 6.7 17.7 2.6
Suicide n/a n/a n/a 12.1 18.3 1.5
Note. COPD = chronic obstructive pulmonary disease; n/a = not available.
aNew York data from 1997. California data from 1997 or aggregated from 1995–1997. New York rates are age-adjusted to New York State census population 1990. California rates are standardized
to US standard population 1940. (As such, these rates are not directly comparable.) All rates are per 100 000 population.
bCalifornia—ages 15–24 years.
cCalifornia—includes influenza.
dCalifornia—motor vehicle accidents only.
tion.16 These socioeconomic factors were cho- correlation coefficients between rate ratios Table 1 shows the mean, range, and range di-
sen on the basis of data availability and con- and variability measures and examined vari- vided by mean across counties in both states.
sistency with previously published studies.11,17 ability in outcomes across counties, stratified In New York, the largest variability in out-
by SES. comes was in AIDS mortality (range/mean=
Analysis We also calculated smoothed county- 10.9), followed by homicide (6.3), all-cause
We analyzed counties in California and specific rates, in which the observed rate in a mortality among persons aged 10–24 years
New York separately. We stratified counties county was “stabilized” by replacing it with (2.6), and mortality from cirrhosis (2.0). The
into quartiles by each of the SES measures, the weighted average of the county rate and smallest variability was observed in all-cause
calculated the average rate of each health all adjacent county rates; weights were pro- mortality across all ages (0.3) and among per-
outcome for the bottom and top quartiles, portionate to population size.18 We repeated sons aged more than 75 years (0.4), as well as
and obtained the rate ratio for a given health all of the analyses described here on the in mortality from neoplastic disease (0.5) and
outcome by comparing counties in the lowest smoothed rate estimates. Finally, we com- mortality from cardiovascular disease (0.8).
and highest socioeconomic quartiles. pared outcome rankings and correlations de- In California, variability was highest for
We also calculated the variability of each rived using the range-divided-by-mean mea- AIDS (range/mean=7.3), followed by homi-
health outcome across all counties in each sure with rankings obtained using 2 other cide (2.6), mortality from cirrhosis (2.4), and
state. Following Levins and Lopez, we used variability measures: interquartile range di- mortality among persons aged 15–24 years
the range of values divided by the mean vided by mean, and the coefficient of varia- (2.1). The lowest variability was in rates for
value as the measure of variability; this mea- tion (SD/mean×100%). all-cause mortality across all ages and for per-
sure provides a useful estimate for qualitative sons aged more than 75 years as well as mor-
analyses.11 The larger the range divided by RESULTS tality from neoplastic disease (0.5 for each)
the mean value, the higher the variability of and mortality from cardiovascular (0.8) or
a particular health outcome across counties. We included 61 (98%) of 62 counties in cerebrovascular (0.9) disease. The ordering of
No statistical inferences were based on this New York and 53 (91%) of 58 counties in diseases by their intercounty variability was
measure of variability. We calculated Pearson California in the analysis. For each outcome, similar between the 2 states.
November 2002, Vol 92, No. 11 | American Journal of Public Health Karpati et al. | Peer Reviewed | Research and Practice | 1769
 RESEARCH AND PRACTICE 
TABLE 2—Relative Mortality Rates Comparing Counties in the Lowest to Highest Quartiles 
of Economic Indicators, New York and California, 1997.
New York California
Under-18 % high Row Under-18 % high Row
Outcome Employment Poverty poverty Income school grad mean Employment Poverty poverty Income school grad mean
All-cause mortality (all ages) 1.05 1.07 1.05 1.05 1.03 1.05 1.16 1.19 1.22 1.21 1.13 1.18
All-cause mortality (≥ 75 y) 0.97 0.97 0.96 1.01 0.95 0.97 0.95 0.94 0.94 0.91 0.92 0.93
All-cause mortality (10–24 y) 1.00 1.41 1.22 1.21 1.47 1.26 1.32 1.16 1.27 1.32 1.46 1.31
Infant mortality 1.04 1.14 1.04 0.96 1.17 1.07 1.24 1.35 1.41 1.29 1.17 1.29
AIDS 3.47 3.13 4.00 1.64 2.59 2.96 0.40 0.68 0.88 0.57 0.68 0.64
Pneumonia 0.92 1.11 1.09 1.10 1.01 1.04 0.93 1.07 1.09 0.88 0.99 0.99
COPD 1.01 1.22 1.10 1.16 1.12 1.12 1.35 1.32 1.34 1.47 1.19 1.33
Cardiovascular disease 1.12 1.06 1.07 1.05 1.04 1.07 1.18 1.28 1.33 1.18 1.23 1.24
Stroke 0.85 1.16 1.13 1.02 0.96 1.02 1.11 1.11 1.17 1.09 1.15 1.12
All neoplastic disease 1.01 0.97 0.97 1.00 0.94 0.98 1.07 1.02 1.05 1.12 1.01 1.05
Lung cancer n/a n/a n/a n/a n/a n/a 1.15 1.11 1.12 1.23 0.99 1.12
Female breast cancer n/a n/a n/a n/a n/a n/a 1.00 0.86 0.85 0.94 0.88 0.90
Cirrhosis 1.42 1.40 1.37 1.20 1.00 1.28 1.13 1.44 1.38 1.43 0.98 1.27
Accidents 0.91 1.25 1.13 1.17 1.23 1.14 2.56 1.71 1.75 2.38 1.76 2.03
Homicide 1.40 1.94 2.23 1.13 1.20 1.58 1.40 2.02 2.61 1.39 1.86 1.85
Suicide n/a n/a n/a n/a n/a n/a 0.93 1.00 0.98 1.06 0.73 0.94
Column mean 1.23 1.36 1.39 1.13 1.20 1.18 1.20 1.27 1.22 1.13
Note. COPD = chronic obstructive pulmonary disease; n/a = not available.
Rate ratios comparing mean disease-specific (0.90 for female breast cancer and 1.12 for living in poverty). In New York, the Pearson
mortality rates between counties in the lower lung cancer). The highest mean ratios were correlation coefficient was 0.97; in California,
and upper quartiles of various socioeconomic for mortality rates from motor vehicle acci- it was 0.70 (0.01 when AIDS was included in
markers are shown in Table 2. Rate ratios dents (2.03) and for homicide rates (1.85). the calculation).
greater than 1.0 imply that counties with The mean ratio for cirrhosis was 1.27. All- Figure 1 shows the relation between the
lower SES have higher disease-specific mor- cause mortality for persons aged more than homicide rates in New York counties and
tality than do those with higher SES. 75 years, suicide, pneumonia and influenza their socioeconomic status. Homicide rates for
The range of all rate ratios in New York mortality, and female breast cancer mortality each county were plotted against tertiles of
was 0.85–4.00. The mean ratios across eco- each had rate ratios of less than 1.10 for all poverty (percentage of persons aged <18
nomic marker categories ranged from 0.98 economic markers. In addition, ratios for years living in poverty). Counties of lower
for neoplastic disease to 2.96 for AIDS. Mean AIDS mortality ranged from 0.40 to 0.88. economic status displayed greater variability
homicide and cirrhosis mortality ratios were Mortality from all causes in persons aged in their rates of homicide than do counties
1.58 and 1.28, respectively. For all economic 10–24 years had a mean rate ratio of 1.31, with high economic status. The general trend
markers, all-cause mortality, all-cause mortal- whereas for persons aged more than 75 years of the relation was linear, with a positive
ity for persons aged more than 75 years, and the mean ratio was 0.93. The mean ratios for slope; however, among counties with the low-
mortality from neoplasms had ratios less than all outcomes across economic measures est economic status, there were both low and
1.10. Mortality from all causes in persons ranged from 1.13 (high school graduation high rates.
aged 10–24 years had a mean rate ratio of rate) to 1.27 (percentage of persons aged Smoothed rates showed less variability than
1.26. The mean ratios for all outcomes across <18 years in poverty). did observed rates (range/mean for all out-
economic measures ranged from 1.13 (me- In both states, the variability (range/mean) comes varied from 0.2 to 6.4, compared with
dian household income) to 1.39 (percentage of health outcomes across counties was 0.3 to 11.0 in the original analysis); nonethe-
of persons aged <18 years in poverty). strongly correlated with the mean ratio of less, the trends and correlations present in the
In California, ratios ranged from 0.40 to rates between counties in the lowest and original analysis were preserved. In both New
2.61. The mean ratio across economic indica- highest quartiles of economic status (mea- York and California, AIDS rates and homicide
tors for neoplastic disease rates was 1.05 sured by percentage of children <18 years rates continued to display the most variability
1770 | Research and Practice | Peer Reviewed | Karpati et al. American Journal of Public Health | November 2002, Vol 92, No. 11
 RESEARCH AND PRACTICE 
comes that are sensitive to SES.1920 For exam-
ple, in California, lung cancer mortality
(largely a consequence of smoking) had a
higher mean ratio than did female breast can-
15 cer mortality. In New York, cirrhosis mortality
(largely a consequence of alcohol misuse) had
a higher rate ratio than did neoplasms. A
comparison of age-specific mortality further
supports this hypothesis. In New York and
10
California, county rates between economic
strata exhibit low ratios for mortality among
older persons and high ratios for youth mor-
tality. Youth mortality has more potentially
5 modifiable behavioral and social causes at its
root than does mortality among older per-
sons.
The economic sensitivity of AIDS was re-
0 versed in New York and California. In New
Low poverty High poverty York State, counties in the lowest economic
quartiles had an average of 2.96 times the
Tertiles of under-18 poverty AIDS rates of counties in the highest quar-
FIGURE 1—Intercounty variability in homicide rates stratified by tertiles of child (under 18 y) tiles, and AIDS is particularly prevalent in
poverty, New York State, 1997. poor communities of New York City. By con-
trast, in California, the populations with the
lowest SES are both urban and rural, and
AIDS incidence is more widely distributed in
populations of varying SES.
(New York AIDS range/mean=6.4, homicide variability of those rates across counties. We Our principal measures of interest were
range/mean=3.4; California AIDS range/ hypothesized that the outcomes most sensi- variabilities in outcomes rather than absolute
mean=2.0, homicide range/mean=1.4), and tive to SES would also exhibit the most vari- rates. Variability in mortality rates across
mortality among persons aged more than 75 ability. Outcomes in both New York and Cali- counties was highest for the outcomes with
years and neoplastic disease mortality contin- fornia that displayed high sensitivities to SES larger SES rate ratios. AIDS mortality and
ued to display the least (range/mean= were AIDS, homicide, cirrhosis, and acci- homicide were the outcomes with the largest
0.2–0.3). Similarly, as expected, the magni- dents. Outcomes in both states were most variability in the 2 states. Rates for all-cause
tudes of variability for each outcome were at- sensitive to the percentage of children living mortality across ages and for mortality in
tenuated when the coeffecient of variation or in poverty as a single indicator of SES. older persons, as well as rates for neoplastic
the interquartile range divided by mean were All-cause mortality and certain neoplastic disease mortality, exhibited small variability
used (ranges 12%.–104% and 0.1–0.9, re- disease mortality rates did not differ greatly and were generally not sensitive to socioeco-
spectively, in New York, and 7%–200% and between poorer and wealthier counties. The nomic conditions.
0.2–0.7 in California); the ordering of out- underlying mechanisms for these health out- Although we hypothesized that variability
comes by variability, however, remained comes might account for the findings. Dis- reflects system-specific conditions (i.e., the
largely unchanged in comparison with the or- eases with an incidence or course that is in- balance among vulnerability, stressors, and
dering obtained with the range-divided-by- fluenced by behavioral or environmental protectors), variability may also be the prod-
mean statistic (e.g., AIDS and homicide rates factors would be expected to exhibit sensitiv- uct of random events, especially when the
remained most variable and all-cause mortal- ity to SES, whereas diseases with genetic or outcome of interest is rare or the population
ity and mortality from neoplasms remained other nonmodifiable causes would not. Our within which it occurs is small. Moreover,
least variable), as did the strong association analysis is consistent with earlier findings and there may be confounding of the SES–
between variability and SES sensitivity. builds on previous small-area analyses in sensitivity/variability relationship if rare
Kansas, Saskatchewan, and Cuba.11 events are also more sensitive to SES. To ad-
DISCUSSION A general process underlying these obser- dress these possibilities, we repeated all
vations has been articulated by Link and Phe- analyses using smoothed county-specific rates,
This analysis investigated the effects of lan, who postulated that access to protections which reduce intercounty population variabil-
counties’ SES on their mortality rates and the and avoidance of harms underlie health out- ity, as well as robust measures of variability,
November 2002, Vol 92, No. 11 | American Journal of Public Health Karpati et al. | Peer Reviewed | Research and Practice | 1771
Homicides per 100 000
 RESEARCH AND PRACTICE 
which have low sensitivity to outliers. The ob- Health, Harvard School of Public Health, Boston, Mass. increased mortality after acute myocardial infarction.
served rate variability was attenuated under Adam Karpati is also with the Epidemiology Program Of- Am J Cardiol. 1987;59:256–262.
fice, Centers for Disease Control and Prevention, Atlanta,
these circumstances, but trends in health out- 13. New York State Department of Health. AvailableGa. Sandro Galea is also with the Center for Urban Epide-
come variability and associations with county- at: http://www.health.state.ny.us/nysdoh/research/re-miologic Studies, New York Academy of Medicine, New
search.htm. Accessed April 2, 2002.
level SES were preserved. York. Tamara Awerbuch and Richard Levins are with the
Department of Population and International Health, Har- 14. California Department of Health Services. Avail-
In their examination of ecological factors vard School of Public Health. able at: http://www.dhs.ca.gov/hisp. Accessed April 2,
contributing to adverse health effects, Levins Requests for reprints should be sent to Adam Karpati, 2002.
and Lopez suggested that the relation be- MD, MPH, Bureau of Community HealthWorks, New York 15. U.S. Bureau of the Census. Available at http://
City Department of Health, 40 Worth St, Room 1607,
tween economic deprivation and variability in www.census.gov/hhes/www/saipe/estimatetoc.html.New York, NY 10013 (e-mail: aek3@cdc.gov).
health status might be mediated by the vul- Accessed April 2, 2002.This article was accepted January 10, 2002.
nerability of populations.11 They cited an ob- 16. Dalaker J. Poverty in the United States: 1998.
Washington, DC: US Census Bureau, Current Popula-
servation by a geneticist, I. I. Schmalhausen, Contributors tion Reports; 1999. Series P60-207.
that “a system at the boundary of its toler- A. Karpati and S. Galea collected the data, performed
17. Cubbin C, Pickle LW, Fingerhut L. Social context
ance along any dimension of its existence is the analysis, and wrote the article. T. Awerbuch con-
tributed to the study design, oversaw the analysis, and and geographic patterns of homicide among US black
more vulnerable to small differences in cir- contributed to the editing of the article. R. Levins con- and white males. Am J Public Health. 2000;90:
cumstance along any dimension.”20(p.276) Pop- tributed to the study formulation and design, oversaw 579–587.
ulations enduring social or economic depriva- the analysis, and contributed to the editing of the article. 18. Devine O, Parrish RG. Monitoring the health of a
population. In: Stroup DF, Teutsch SM, eds. Statistics in
tion will be more vulnerable to potential Public Health: Quantitative Approaches to Public Health
stressors than will populations of higher sta- Acknowledgments Problems. New York, NY: Oxford University Press;
Some of the ideas in this article were developed with
tus. Thus, acute outbreaks of infectious dis- 1998:58–91.the support of the Robert Wood Johnson Foundation
eases, environmental risks, or transient gaps Investigator Award and the Kansas Health Foundation. 19. Link BG, Phelan JC. Evaluating the fundamental
in public health services will likely affect an cause explanation for social disparities in health. In:
Bird CE, Conrad P, Fremont AM, eds. Handbook of
economically deprived population to a greater References Medical Sociology. 5th ed. Upper Saddle River, NJ:
degree than they would a less marginalized 1. Kawachi I, Kennedy BP, Lochner K, Prothrow- Prentice Hall; 2000:33–46.
population. Stith D. Social capital, income inequality, and mortality. 20. Schmalhausen II. Factors of Evolution. Philadel-
Am J Public Health. 1997;87:1491–1498.
We note, however, that these external phia, Pa: Blakiston Company; 1949.
2. O’Campo P, Xue X, Wang MC, Caughy M. Neigh-
stressors are not uniformly distributed across borhood risk factors for low birthweight in Baltimore: a
all disadvantaged communities, and therein multilevel analysis. Am J Public Health. 1997;87:
might lie the source of the observed variabil- 1113–1118.
ity. Although vulnerability might result from 3. Yen IH, Kaplan GA. Neighborhood social environ-
ment and risk of death: multilevel evidence from the
chronic economic deprivation, the range of Alameda County Study. Am J Epidemiol. 1999;149:
adverse health outcomes will depend on the 898–907.
degree to which each community experiences 4. Susser M. The logic in ecological, I: the logic of
stressors and the distribution within commu- analysis. Am J Public Health. 1994;84:825–829.
nities of counteracting protective factors. 5. Diez-Roux AV. Bringing context back into epide-
Variability in biological systems is increas- miology: variables and fallacies in multilevel analysis.
Am J Public Health. 1998;88: 216–222.
ingly seen as a marker for stresses to systems
6. Philippe P, Mansi O. Nonlinearity in the epidemi-
in homeostasis. Applying this insight to com- ology of complex health and disease processes. Theor
munities and health, we postulate that eco- Med Bioeth. 1998;19:591–607.
nomic deprivation produces vulnerability to 7. Goldenfeld N, Kadanoff LP. Simple lessons from
stressors whose nonuniform distribution complexity. Science. 1999;284:87–89.
across populations manifests as variability in 8. Sandberg S, Awerbuch TE, Gonin R. Simplicity vs
health outcomes. One possible implication of complexity in deterministic models: an application to
AIDS data. J Biol Systems. 1994;4:61–81.
this model is that interventions to improve
9. Koopman JS, Longini IM Jr. The ecological effects
public health might exert the greatest effect of individual exposures and nonlinear disease dynam-
not by targeting particular stressors, but ics in populations. Am J Public Health. 1994;84:
rather by focusing on improving general so- 836–842.
cial and economic well-being, thus reducing 10. Seely AJ, Christou NV. Multiple organ dysfunction
populations’ overall vulnerability. syndrome: exploring the paradigm of complex nonlin-
ear systems. Crit Care Med. 2000;28:2193–2200.
11. Levins R, Lopez C. Toward an ecosocial view of
About the Authors health. Int J Health Serv. 1999;29:261–293.
At the time of the study, Adam Karpati and Sandro Galea 12. Kleiger RE, Miller JP, Bigger JT Jr, Moss AJ. De-
were with the Department of Population and International creased heart rate variability and its association with
1772 | Research and Practice | Peer Reviewed | Karpati et al. American Journal of Public Health | November 2002, Vol 92, No. 11