Objectives. We tested the association between the availability of primary care and income inequality on several categories of mortality in US counties.
Methods. We used cross-sectional analysis of data from counties (n=3081) in 1990, including analysis of variance and multivariate ordinary least squares regression. Independent variables included primary care resources, income inequality, and sociodemographics.
Results. Counties with higher availability of primary care resources experienced between 2% and 3% lower mortality than counties with less primary care. Counties with high income inequality experienced between 11% and 13% higher mortality than counties with less inequality.
Conclusions. Primary care resources may partially moderate the effects of income inequality on health outcomes at the county level.
There is strong theoretical and empirical evidence for the association between strong national primary care systems and improved health indicators.1–4 US ecological studies have demonstrated an association between the primary care physician-to-population ratio and various health outcomes. Better health outcomes were found in states with higher primary care physician-to-population ratios after sociodemographic measures (elderly populations, urban residents, minority populations, education, income, unemployment, pollution) and lifestyle factors (seatbelt usage, obesity, and smoking) were controlled for.5 Geographic areas with more family and general physicians had lower hospitalization rates for conditions preventable with good primary care.6 Individual-level and state-level measures of primary care resources were also significantly associated with lower heart disease and cancer mortality rates.7–11
Although many of the previous studies controlled for both environmental and individual-level determinants of health, a new appreciation of the role of contextual-level determinants of population health emerged during the past decade. In particular, there is an ongoing debate over the role of the distribution of income as a determinant of population health, with evidence both supporting and refuting what has become known as the relative income hypothesis—the proposition that the greater the gap in income between the rich and the poor in a given area, the worse the health status for the overall population of that area.12–15 Although international and cross-country studies of the relative income hypothesis have resulted in conflicting conclusions,16–18 there is considerable evidence that, at least within the United States, income inequality is associated with poorer population health.19–21
The pathways through which income inequality might affect health are still unknown, and hypotheses include psychosocial and material pathways.12,19 In this study, we use income inequality as a proxy for social inequalities regardless of their cause. We include the measure of unequal income distribution as a measure of underlying social inequalities in order to test whether primary care resources might be one strategy to help remedy the poorer-than-expected health profile seen in communities that suffer from income and other social inequalities.
Our previous studies suggested that the availability of primary care partially attenuates the adverse effects of other community-level risk factors for poorer health, such as income inequality. Shi and colleagues22 found that the supply of primary care services had an independent and positive impact on health indicators, and that in multivariate models when demographic, income, and health system covariates were controlled for, a higher supply of primary care services actually reduced the magnitude of the deleterious impact of income inequality on health outcomes. Using a multilevel model that included individual, community, and state-level variables, Shi and Starfield23 found that even when they controlled for all covariates, an increase of 1 primary care physician per 10000 population was associated with a 2% increase in the odds of reporting excellent/good health.
Although previous studies have examined the relation between income inequality, primary care, and health outcomes, questions remain about the level of aggregation in which contextual effects occur. This study tests the extent to which these effects are present at a smaller level of analysis: US counties. We chose counties because they represent distinct political and geographic units, are an important unit of local decisionmaking, and are used by health and social services organizations to plan and provide many (but not all) public health and other social programs.
We tested the effects of local-level primary care and social inequalities on a variety of health outcomes at the county level. The first outcome measure was total (all-cause) mortality, one of the most commonly used health status indicators.24,25 The second outcome measure was mortality from heart disease, one of the principal causes of death in the United States and amenable to prevention and control by primary care.26,27 Heart disease and its risk factors have also been associated with income and other social inequalities.28,29 The third outcome measure was mortality from cancers. Although cancer mortality depends on access to specialty services and treatments, primary care has a clear role in the prevention and early detection of some cancers, in the integration and coordination of care given during treatment and follow-up, and in the treatment of other comorbidities.30
All the mortality data were drawn from the Centers for Disease Control and Prevention compressed mortality files. We used data from 1990, the most recent period for which the complete set of other county-level variables was available. The Centers for Disease Control and Prevention WONDER/PC software31 was used to directly standardize data for age and gender to the 1980 US population. Standardized mortality rates for each county are expressed as the number of deaths per 100 000 population.
For the purpose of this study, primary care physicians refers to doctors of medicine per 10000 civilian population in active office-based patient care in family medicine, internal medicine, and pediatrics.32 This variable is called primary care throughout this article. Counties were classified as to whether they fell below or above the national 75th percentile in terms of primary care physicians per 10000 population. Physician data were obtained from the American Medical Association’s annual publication Physician Characteristics and Distribution in the US.33 Data were drawn from 1990.
Income distribution was measured by the Gini coefficient, wherein higher values indicate greater inequality in income distribution.34,35 Household income data used to calculate county-level income inequality came from the US Census Bureau. The Gini coefficient was calculated using software developed by E. Welniak (unpublished software, US Census Bureau, 1988). Because of evidence of a skewed distribution of Gini data, we analyzed income inequality by grouping counties into Gini quartiles, that is, each county was assigned to 1 of 4 groups based on whether its level of income inequality (Gini) was below the national 25th, between the 25th and 50th, between the 50th and 75th, or above the 75th percentile. The group with the lowest levels of income inequality (the under-25th-percentile group) was used as the reference group in analyses. This technique has been employed in other studies.20,36,37
Additional sociodemographic variables known to be associated with population health were included in the analyses as covariates. Because of problems of intercorrelation, each of these variables (except education) was analyzed as a series of categorical variables in order to avoid problems of multi-collinearity in multivariate analyses. The proportion of each county’s population aged 25 years and above with at least a high school education (percentage high school education) was analyzed as a continuous variable. Counties were categorized as either above or below the national median for the proportion of their working population currently unemployed (percentage unemployed), and the percentage of their population identified as Black or African American (percentage Black). Previous studies have shown that there is an important relation between the percentage of African Americans in a given geographic area and other social inequalities.38,39 Counties were also categorized into high (at or above the 75th percentile) or low (below the 75th percentile) for average household income (average income). These data came from the US Census Bureau.
This study was an unmixed ecological analysis. The unit of analysis was each US county, as defined by the 1990 census. Because of missing or incomplete data, not all counties were included. The effective sample size was 3075 counties instead of the US total of 3081. Despite missing values, this sample represents 99.9% of all US counties. This analysis used only county-level ecological variables. In order to avoid an ecological fallacy, we make no inferences about individuals based on the results of these analyses.40,41
In analyzing the data, we first tested the bivariate association between health outcomes and primary care, income inequality, and sociodemographic covariates. Because the independent variables were primarily expressed categorical variables, bivariate analyses were performed using 1-way analysis of variance. Scheffe multiple comparison tests were used in order to compare the differences among the 4 categories of income inequality.42 We employed ordinary least squares regression for multivariate analyses because all dependent variables were continuous with independent and identically distributed errors.42 A set of nested models was designed to examine the extent to which primary care and other covariates moderated the adverse effect of income inequality on mortality. In model 1, only income inequality was used as a predictor of mortality. Model 2 examined the same relation while adjusting for both primary care and the sociodemographic characteristics of the county. Regression coefficients and standard errors are presented along with tests of significance. We compared nested models (model 2 vs model 1) by calculating a delta sum of squares (F) test. This allowed us to determine whether adding the additional covariates in model 2 improved on model 1.43
Table 1 shows the means and standard deviations for the outcome and control variables used in this study. In 1990, mean all-cause mortality was 1011 per 100 000 and varied across counties from a low of about 132 to a high of 2176 that was almost twice the national average. Heart disease mortality represented, on average, nearly 45% of all-cause mortality, ranging from a low of 0 to a high of 1108 per 100 000. Cancer mortality accounted for approximately 22% of all mortality, ranging from 0 to about 567 per 100 000.
The mean county-level Gini coefficient was 0.42, with a range from 0.29 to 0.56. Twenty-six percent of counties were classified as high (above the 75th percentile) income inequality counties, whereas only 20% were classified as below the 25th percentile. Primary care physicians per 10000 population ranged from a low of 0 to a high of approximately 40. The average value was nearly 5 physicians per 10000 population.
Other covariates showed wide variation among counties. The percentage of the population in each county classified as Black or African American averaged almost 9%, ranging from 0% to 86%. The proportion of the active population currently unemployed averaged 6% and ranged from a low of 0% to a high of 36%. Household income ranged from$5559 to almost $39000 with a mean value of $15209. The county average for adults having completed high school was nearly 70% but ranged from 32% to 96%.
Table 2 shows the bivariate relations between each of the covariates and the health outcomes as reported from a 1-way analysis of variance. Figures reported in the first column under each outcome represent the average value of the outcome in counties within each category.
The Gini coefficient of income inequality showed a positive relation with health outcomes. Counties with higher levels of inequality were associated with higher levels of mortality. Counties in the lowest quartile (that is, those with the least inequality in household income) had, on average, statistically significantly (P< .05) lower all-cause, heart disease, and cancer mortality than those counties with higher income inequalities. Primary care showed the opposite effect. Counties in the highest primary care category (≥ 75th percentile) had significantly lower all-cause, heart disease, and cancer mortality than those counties with fewer primary care resources.
Counties with higher than the national median levels of African Americans, unemployed people, or those lacking a high school education had, on average, higher levels of all-cause, heart disease, and cancer mortality. Those counties with average household income above the 75th percentile had lower average rates of mortality from all causes, heart disease, and cancer.
Table 3 shows the results of multivariate ordinary least squares regressions that test the association of income inequality and primary care on health outcomes, while holding other health determinants constant. In the first column (all-cause mortality), model 1 shows that each category of increased income inequality was associated with higher average levels of mortality than the reference group. A similar pattern is present in model 2, although the magnitude of the association was lower for all categories than in model 1. In model 2, the highest magnitude of association occurred in the middle category (50th to 75th percentiles). When all other variables were held constant, counties in this quartile had about 12% higher mortality than those in the lowest quartile.
Primary care was significantly associated with lower all-cause mortality. Counties in the lowest 3 quartiles of primary care had approximately 2% higher mortality than those in the reference category.
Other covariates associated with higher mortality included a higher-than-median percentage of the population that was African American and higher-than-median levels of unemployment. The percentage of population with a high school education was negatively associated with mortality. Each 1% increase in the population completing high school was associated with about a 0.6% decrease in mortality rates. Average income was not statistically significantly associated with mortality in this model.
The second column in Table 3 examines the relation among mortality from heart disease and income inequality, primary care, and other covariates. The pattern is consistent with that observed with all-cause mortality. In model 1, each category of increased income inequality was associated with higher levels of mortality from heart disease. In model 2, the magnitude of this association was somewhat reduced but still significant. Counties in the middle Gini quartile had approximately 11% higher rates of heart disease mortality than those in the reference group. Other covariates associated with increased heart disease mortality include a higher-than-median proportion of African American population and higher-than-median levels of unemployment.
Primary care and education were associated with lower rates of heart disease mortality. Those counties with low levels of primary care (the lowest 75%) had, on average, 3% higher mortality than those counties with more primary care resources. Each 1% increase in the percentage of the population having completed high school was associated with a little less than a 0.7% decrease in mortality from heart disease. The effect of average household income was not statistically significant.
The third column of Table 3 presents results for cancer mortality. Model 1 shows that income inequality was associated with higher cancer mortality, a relation that was somewhat reduced in magnitude in model 2. When other health determinants were controlled for, counties in the middle Gini category experienced about 13% increased mortality over those in the reference category.
Higher primary care was associated with lower cancer mortality. Counties in the lowest primary care quartiles had about 2% higher cancer mortality than those in the reference group. Counties with a higher-than-median proportion of African Americans also experienced higher levels of cancer mortality. Each 1% increase in the percentage of the population having completed high school was associated with a less than 0.5% decrease in cancer mortality. Unemployment and average household income had no statistically significant effect.
This study shows that in US counties, greater primary care resources are consistently associated with lower rates of all-cause, heart disease, and cancer mortalities—even in the presence of income inequality and other health determinants. In multivariate regressions, counties with the lowest levels of primary care resources experienced 2% to 3% higher mortality than counties with higher levels of primary care resources. Income inequality was also found to be a significant predictor of county-level variation in mortality. Counties in higher-income inequality quartiles experienced 11% to 13% higher mortality than counties in the lowest quartile. These findings are consistent with that found in prior studies performed at the state and national levels.11,44 Because this is an ecological study, it cannot show that the same individuals who suffer most from ill health because of income or other social inequalities would gain the most from improved availability of primary care. However, we believe that a plausible causal chain leads from income inequality to mortality. There is evidence that income inequality has negative effects on the psychosocial environment that can contribute to increased stress, poorer social relations, and reduced support at both the individual and population levels.45–47 These conditions may lead to a higher prevalence of certain diseases48 as well as risk factors (e.g., smoking, hypertension, and obesity) for heart disease and some cancers. Many of these risk factors have also been shown to be more prevalent in areas with higher social inequalities.29,49 The alternative neomaterial explanation for the observed relation between income inequality and health is that poor individuals living in unequal areas suffer from worse living conditions and have poorer access to quality education, health, and other social services.18,19 These material deprivations are thought to lead to more limited life chances and a poorer health profile.
There are several reasons why more county-level supply of primary care resources might be associated with better health outcomes. First, the effect of primary care resources may be partially because physicians trained in primary care are more likely to provide good quality primary care services than those who are not.50,51 A system of good-quality primary care services may be able to ameliorate some of the ultimate consequences of social inequalities at the population level by contributing to reduced levels of disease transmission, lowering aggregate levels of risk factors (such as hypertension, smoking, weight gain), improving countywide screening and early diagnosis activities, and developing systems to coordinate care.52,53 Exercise of these functions should contribute to improved functioning of the health system at large because strong primary care not only means more prevention but, ideally, can also lead to more efficient referral, coordination, and continuity of care.54,55
Moreover, the recommendation that expansion of primary care resources might be one strategy to improve population health in areas with high levels of income inequality does necessarily depend on the mechanism linking such inequalities with poorer health. Improving primary care resources in counties with high levels of social inequalities should contribute to incremental improvements in both material conditions and better management of conditions and behaviors associated with increased psychosocial stress.
In terms of the effect of income inequality on health outcomes, prior studies also found that the largest independent income inequality effects occur at the county rather than the census tract level.56 This finding is important because counties are often the local decision-making unit for issues affecting organization of local health, education, and other social services.57 One difficulty posed by county-level analyses is the likelihood of random fluctuations in both numerators and denominators of mortality and other rates through small area geographic aggregation. The presence of low-population counties that did not experience heart or cancer mortalities in 1990 may have led to some imprecision in multivariate models. Moreover, using state-level aggregate data would also have had the advantage of attenuating the likely “crossover” effect encountered when smaller units of analysis are used for measuring the availability of medical care and mortality.58 However, our results were broadly consistent with those in other studies that compensated for this potential limitation. McLaughlin and Stokes38 used a 5-year average (centered on 1990) for all-cause mortality and also found that US counties with higher income inequality had higher mortality rates, even when they adjusted for average income and percentage African American. For this reason, we do not believe that using the county as a unit of analysis has significantly biased our findings.
As a final sensitivity test, results were analyzed for geographic trends within the major US regions as defined in by the Census Bureau (i.e., Northeast, South, Midwest, and West). For all outcomes, regional variables were found to be statistically significant although their inclusion in the model did not affect other covariates.
In interpreting the results of this study, several limitations require consideration. First, the finding of a relation between an increased number of primary care physicians and lower population mortality does not necessarily imply that the mere presence of more primary care physicians ensures either that more individuals in the population will be exposed to primary care or that the delivery of primary care will produce better health outcomes at the individual level. Only studies performed at the individual level would test these hypotheses.
The coefficient of determination (R2) values for the multivariate models were smaller in magnitude than values obtained from similar analyses performed at the state level. Although this is not necessarily a study limitation, it does indicate that several factors affecting the health of populations could not be modeled here because of a lack of county-level data. One difficulty in constructing more comprehensive models of health determinants is that many of these factors are collinear. In this study, average county income was not found to be a statistically significant predictor of mortality, although it was in other studies.38 This discrepancy is likely to be the result of the strong correlation between average country income, education levels, and percentage African American. Each of these factors is interrelated and also correlated with income inequality.
A final potential limitation is that primary care physician availability is likely to be an inadequate proxy for the receipt of good primary care. Ultimately, we would like better information on the structural characteristics and practice features of primary care. Unfortunately, there are no data that make it possible to adequately characterize receipt of good primary care (as distinguished from receipt of ambulatory care services, which also include specialty care) at the state or county levels. Nonphysician providers of primary care were not captured in this study. For these reasons, it is likely that our analysis underestimates the overall contribution of primary care to improved population health.
a The mean and standard deviation are provided for continuous variables; the frequency and percentage are provided for categorical variables. bBecause of missing values, the total sample size may not reach the maximum of 3081. aOne-way analysis of variance was performed to find the overall difference among 4 groups for Gini. The Scheffe multiple comparison procedure was used to further compare paired differences between each of the 4 groups after rejection of the overall hypothesis that at least 1 group is different from the others. The results from Scheffe comparison procedure indicate that each mortality outcome for Gini ≤ 25th percentile group was significantly lower than that of the other Gini groups (P < .05). T-tests were used to compare variables with only 2 groups. * P < .05; **P < .01; ***P < .001; ****P < .0001. aThe coefficient for categorical variables is the difference between the reference group and the compared group, and the coefficient for continuous variables (e.g., percent high school education) is the slope, the magnitude of change of the outcome per 1 unit change in predictor. Values in parentheses represent standard deviations. * P < .05; **P < .01; ***P < .001; ****P < .0001.
n (total)b Percentage in Each Category Mean (SD) Range (Low–High) All-cause mortality per 100 000 population 3081 100.00 1011.49 (262.39) 131.78–2176.47 Heart disease mortality per 100 000 population 3081 100.00 458.91 (147.86) 0.00–1107.69 Cancer mortality per 100 000 population 3081 100.00 222.54 (60.57) 0.00–566.67 Gini 3075 100.00 0.42 (0.04) 0.29–0.56 > 75th percentile 808 26.28 . . . . . . 50th–75th percentile 922 29.98 . . . . . . 25th–50th percentile 727 23.64 . . . . . . ≤ 25th percentile 618 20.10 . . . . . . Primary care physicians per 10 000 population 3076 100.00 4.96 (3.24) 0.00–39.53 < 75th percentile 2307 75.00 . . . . . . ≥ 75th percentile 769 25.00 . . . . . . Percent Black 3076 100.00 8.56 (14.33) 0.00–86.20 ≥ Median 1531 49.77 . . . . . . < Median 1545 50.23 . . . . . . Percent unemployed 3076 100.00 6.20 (2.78) 0.00–36.10 ≥ Median 1520 49.41 . . . . . . < Median 1556 50.59 . . . . . . Average income (in US $) 3076 100.00 15 209 (3518) 5559–38 794 < 75th percentile 2305 74.93 . . . . . . ≥ 75th percentile 771 25.07 . . . . . . Percentage of population with high school education 3076 100.00 69.56 (10.34) 31.60–95.50 All-Cause Mortality per 100 000 Population Heart Disease Mortality per 100 000 Population Cancer Mortality per 100 000 Population Predictora Mean SD Mean SD Mean SD Gini > 75th percentile 1084.60**** 263.59 491.19**** 152.83 228.86**** 61.78 50th–75th percentile 1057.39**** 238.53 479.5**** 135.58 234.04**** 56.38 25th–50th percentile 1003.67**** 249.29 459.99**** 145.92 224.63**** 57.93 ≤ 25th percentile 856.24**** 244.44 384.76**** 135.67 194.59**** 59.66 Primary Care < 75th percentile 1030.90**** 262.53 471.37**** 148.85 225.51**** 61.11 ≥ 75th percentile 953.25**** 253.35 421.53**** 138.37 213.63**** 58.08 Percent Black ≥ Median 1030.91**** 277.79 471.26**** 161.02 226.75**** 66.87 < Median 992.24**** 244.76 446.67**** 132.47 218.37**** 53.30 Percent unemployed ≥ Median 1037.62**** 223.00 470.85**** 128.94 227.58**** 55.19 < Median 984.74**** 295.05 446.68**** 164.14 217.39**** 65.25 Average income < 75th percentile 1040.37**** 246.05 475.19**** 140.77 226.02**** 57.21 ≥ 75th percentile 925.15**** 289.54 410.24**** 157.64 212.14**** 68.66 Percent high school education ≥ Median 936.40**** 269.73 421.45**** 150.37 212.14**** 62.23 < Median 1087.07**** 231.45 496.61**** 135.23 233.01**** 57.00 All-Cause Mortality Heart Mortality Cancer Mortality Predictorsa Model 1 Model 2 Model 1 Model 2 Model 1 Model 2 Intercept 856.24 (10.02)**** 1367.52 (49.61)**** 384.76 (5.74)**** 642.57 (28.52)**** 194.59 (2.37)**** 271.18 (12.13) **** Gini > 75th percentile 228.36 (13.31)**** 147.42 (15.14)**** 106.43 (7.63)**** 63.17 (8.71)**** 34.27 (3.15)**** 24.16 (3.70) **** 50th–75th percentile 201.15 (12.95)**** 160.64 (13.14)**** 94.74 (7.42)**** 73.16 (7.55)**** 39.45 (3.06)**** 34.76 (3.21) **** 25th–50th percentile 147.43 (13.63)**** 131.59 (13.10)**** 75.24 (7.81)**** 66.92 (7.53)**** 30.04 (3.22)**** 28.51 (3.20) **** ≤ 25th percentile 0.00 0.00 0.00 0.00 0.00 0.00 Primary care < 75th percentile . . . 23.80 (10.76)* . . . 19.96 (6.19)** . . . 5.29 (2.63)* ≥ 75th percentile . . . 0.00 . . . 0.00 . . . 0.00 Percent Black ≥ Median . . . 101.33 (9.06)**** . . . 53.87 (5.21)**** . . . 17.36 (2.21) **** < Median . . . 0.00 . . . 0.00 . . . 0.00 Percent unemployed ≥ Median . . . 29.91 (9.42)** . . . 20.23 (5.42)*** . . . 0.53 (2.30) < Median . . . 0.00 . . . 0.00 . . . 0.00 Average income < 75th percentile . . . 8.52 (11.38) . . . 11.43 (6.54) . . . –4.36 (2.78) ≥ 75th percentile . . . 0.00 . . . 0.00 . . . 0.00 Percentage high school education . . . –8.10 (0.59)**** . . . –4.29 (0.34)**** . . . –1.18 (0.14) **** Adjusted R2 0.100 0.195 0.068 0.163 0.056 0.097 Delta sum of squares test (F-statistic) . . . 74.18*** . . . 70.48*** . . . 29.08***
a The mean and standard deviation are provided for continuous variables; the frequency and percentage are provided for categorical variables. bBecause of missing values, the total sample size may not reach the maximum of 3081.
aOne-way analysis of variance was performed to find the overall difference among 4 groups for Gini. The Scheffe multiple comparison procedure was used to further compare paired differences between each of the 4 groups after rejection of the overall hypothesis that at least 1 group is different from the others. The results from Scheffe comparison procedure indicate that each mortality outcome for Gini ≤ 25th percentile group was significantly lower than that of the other Gini groups (P < .05). T-tests were used to compare variables with only 2 groups.
* P < .05; **P < .01; ***P < .001; ****P < .0001.
aThe coefficient for categorical variables is the difference between the reference group and the compared group, and the coefficient for continuous variables (e.g., percent high school education) is the slope, the magnitude of change of the outcome per 1 unit change in predictor. Values in parentheses represent standard deviations.
* P < .05; **P < .01; ***P < .001; ****P < .0001.
This study was funded by the Bureau of Primary Health Care, Johns Hopkins Primary Care Policy Center.
Human Participant Protection This study was approved as exempt by Johns Hopkins University institutional review board because no human participants were involved in this study.