The Relationship Between Primary Care, Income Inequality, and Mortality in US States, 1980–1995

Data and Measures

Data for this study came from the Compressed Mortality Files, the US Department of Commerce and the Census Bureau,17 the National Center for Health Statistics,18 and the Centers for Disease Control and Prevention (CDC).19 Physician data were obtained from the American Medical Association Physician Master File.20 Data were drawn from 1980, 1985, 1990, and 1995.

The outcome measure used was total all-cause mortality. Mortality is among the most commonly used health status indicators, especially in studies on income inequality and health.3,19,21,22 It is significantly and inversely associated with life expectancy (r = −.90, P < .001, based on 1990 US state data). The CDC Wonder/PC software23 was used to obtain mortality data and directly standardize them for age to the 1980 US population. Standardized data for each state are expressed as the number of deaths per 100,000 population.

Income distribution was measured by the Gini coefficient, a commonly used indicator of income inequality. The Gini coefficient is based on the Lorenz curve, a cumulative frequency curve that compares the distribution of a specific variable (eg, income) with the uniform distribution that represents equality. This equality distribution is represented by a diagonal line; the greater the deviation of the Lorenz curve from this line, the greater the inequality. Thus, the Lorenz curve is a mechanism to graphically represent the cumulative share of the total income accruing to successive income intervals.24 The Gini coefficient ranges from 0 to 1; 0 represents perfect income equality and 1 total income inequality. The reduction of the Gini coefficient requires corresponding redistribution in income or wealth from the rich to the poor. Although the level of income inequalities is reflected in the value of the Gini coefficient itself (for example, a value very close to 0 will represent a low level of inequality), the interpretation of the coefficient is usually done in comparative terms, by contrasting the calculated value with that of other geographic units, population groups, etc. For example, a coefficient of 0.2 will represent a lower level of inequality than a coefficient of 0.4. For examples of Gini coefficient calculation, please refer to specific web sites (http://www.paho.org/English/SHA/be_v22n1-Gini.htm).

Data used to calculate the income inequality measure came from the US census population and housing summary tapes for the years 1980 to 1995. This file provides annual data on household income for 25 income intervals. Counts of the number of households that fall into each income interval along with the total aggregate income and the median household income were obtained for each state. The Gini coefficient was calculated using software developed by E. Welniak.25 We also used the Robin Hood index, another measure of income inequality that is less sensitive to highly skewed income distributions.26 Because both measures provided similar results, we generally only present results using the Gini coefficient.

Primary care physician specialties included family practice and general practice, general internal medicine, and general pediatrics.17,27 Family and general practice are often combined into a single group called family medicine. Thus, primary care physicians referred to doctors of medicine per 10,000 persons who were in active office-based patient care in family medicine, internal medicine, and pediatrics. For the sake of brevity, this variable is called primary care throughout this article. In the analysis, we compare the effect of this combined primary care measure with the independent effects of each primary care subspecialty.

We also evaluate the impact of specialty care on population health. Specialty care is defined as non-primary care doctors of medicine per 10,000 persons (excluding physicians in residency training or osteopathic physicians) who were in active office-based patient care.

Analyses excluded other potential health determinants such as education, health insurance, racial/ethnic minorities, and poverty for several reasons. First, previous studies using the same data sources and unit of analysis (US states) have already confirmed the relationship between income inequality and health while controlling for these covariates.16,28 Second, many covariates are highly correlated with the income inequality measure, so including them together in the same model could result in multicollinearity. For example, Blakely and Kawachi29 found that using median household income in multiple regression models, including income inequality, can ‘overcontrol’ the association of income inequality with health outcomes. This could result in a greater chance of a type II error, because any absence of effect of income inequity might be explainable by its relationship with these other measures instead of its lack of relationship with the health outcome.30

Design

The present study was an ecologic study of the unmixed type; ie, our analyses correlated ecologic variables with an ecologic outcome.31 The unit of analysis was the 50 US states. Because we avoided making inferences about individual persons from grouped data, no cross-level bias occurred.31,32

One advantage of such an approach is the lower likelihood of random fluctuations in both numerators and denominators of mortality rates through geographic aggregation at the state level. Using state-level aggregate data also had the advantage of attenuating the likely “crossover” effect encountered when smaller units of analysis are used for measuring availability of medical care and mortality.19,33 The “crossover” effect refers to the likelihood that those who require specialized care may seek care in areas where that care is more available. However, patients are less likely to move across states than across smaller geographic units, such as counties or cities, to seek specialized care.

Analysis

In analyzing the data, we first looked at the correlation between contemporary and time-lagged income inequality, primary care and primary care subspecialties, and specialty care variables with health outcomes using Pearson’s correlation coefficients at each 5-year period (1980, 1985, 1990, and 1995).

Multivariate models used weighted multiple regression. This procedure takes into account a weight (based on state population size) assigned to each observation that reflects the “relative amount of information” embodied in the observation.34 The multiple regression model was chosen because the variables included in the analysis were either interval or ratio measures and they seemed to be normally distributed. This technique allowed us to examine the association of any given independent variable on the dependent variable while holding all other independent variables constant.

Because the effects of the predictors on mortality are expected to materialize over time, we also conducted analyses with time lags. A total of 3 time lag models (each with a 5-year lag) were presented. The first model compares total mortality in 1985 as a function of the independent measures from 1980; the second compares mortality in 1990 with predictors from 1985, and the third compares mortality from 1995 with predictors from 1990.

Bivariate Analyses

Table 1 presents Pearson correlations between mean income inequality (Gini coefficient and Robin Hood Index), primary care, primary care specialty physicians, specialty physicians, and all-cause mortality for each state for the periods 1980, 1985, 1990, and 1995. Both the Gini coefficient and the Robin Hood Index were statistically significantly associated with state all-cause mortality during every period (P < .01 for all measures, except the Gini coefficient in 1995, where P < .05). The magnitude of the association does not show a clear trend. The highest correlations were seen in 1990, whereas the lowest correlations were observed in 1995. Five-year lagged income inequality measures were likewise significantly correlated with mortality but the magnitude of the association increased each period from 0.46 in 1985 to 0.55 in 1995 (P < .01 for each period).

Primary care was negatively associated with mortality in 1980 (P < .01), 1985 (not significant), 1990 (P < .1), and 1995 P < .01). Five-year lagged primary care measures showed a stronger negative association than contemporary primary care measures. The primary care subspecialty of general practice was likewise negatively associated with all-cause mortality for each period. The strength of the association decreased with each period, and statistical significance declined to 90% in 1995. Lagged general practice measures showed a similar pattern. Neither contemporary nor lagged measures of internal medicine, general pediatricians, or specialty care physicians were significantly associated with mortality at any period.

Multivariate Analyses

Table 2 presents the weighted regression coefficients of primary care, income inequality, and specialty care on all-cause mortality. Four columns are presented, 1 for each 5-year period. The results show that income inequality and specialty care were strongly associated with higher mortality, while primary care was associated with lower mortality in each period.

Each period shows essentially the same pattern in terms of the significance direction and magnitude of the association between the explanatory variables and mortality. In ordinary least-squares regression, the β-coefficient estimates the effect of a 1-unit increase in the independent variable on the dependent variable (mortality). For example, in 1995, a 1-unit increase in the Gini coefficient (ie, income redistribution favoring the rich) was associated with as many as 719 deaths per 100,000, whereas an increase of 1 primary care physician per 10,000 persons was associated with a reduction of 35 deaths per 100,000. An increase of 1 specialty physician per 10,000 population was associated with approximately 15 additional deaths per 100,000.

The square of the multiple correlation coefficient (R²) is a measure of the proportion of the variance in the dependent variable explained by all the independent variables in the model. All R² values reported here are adjusted for the number of covariates in each model.30 The adjusted R² values (ranging from 0.27 to 0.35) indicate that the models explain only about 30% of variance in age-adjusted total mortality.

Table 3 is similar to Table 2, except that primary care was divided into its 3 specialty components (family medicine, internal medicine, and general pediatrics). In weighted multiple regressions for the 4 time periods, the Gini coefficient continued to be positively associated with all-cause mortality (P < .05 for all periods, except in 1995, when the statistical significance fell to the P < .1 level). The magnitude of the regression coefficient was reduced, however. For example, in 1995, a 1-unit increase in the Gini coefficient was associated with an additional 657 deaths per 100,000, as opposed to 719 in the previous table.

Among the primary care specialties included in the model, only family medicine was statistically significant for all 4 periods (P < .01). An increase of 1 family medicine physician per 10,000 persons was associated with a decrease of 71 deaths per 100,000 in 1980 and 39 deaths per 100,000 in 1995. The internal medicine primary care specialty was statistically significant only in 1995 (P < .05): an increase of 1 general internist per 10,000 persons was associated with a decrease of 30 deaths per 100,000 in 1995. However, the number of general pediatricians was not significantly associated with mortality. As in the previous table, specialty physicians were not significantly associated with mortality during any period. The adjusted R² measures range from 0.30 to 0.43, indicating that the models are improved over those presented in Table 2. Each explains more than 30% of variance in age-adjusted total mortality.

Because the effect of income inequality and primary care on mortality is expected to manifest over time, Table 4 presents results of weighted regressions of time-lagged primary care, specialty care, and income inequality measures on mortality. We also tested the effects of 2 different income inequality measures: the Robin Hood index and the Gini coefficient.

Each column presents the regression of the 5-year time-lagged predictors on mortality rates. Separate models compare the effect of the Robin Hood index with that of the Gini coefficient.

In each of the periods tested, the relationship is similar: both time-lagged income inequality measures were positively associated with mortality (P < .01 for both measures at each period). The magnitude of the β-coefficient for the Robin Hood index was consistent at about 18 for all periods, whereas the magnitude of the Gini coefficient increased from 1340 in 1985 to 1448 in 1995.

The time-lagged primary care measure exhibited a consistent pattern as well. It was negatively associated with mortality at each period (P < .05 for all models). The magnitude of the primary care regression coefficient did not significantly differ based on the income inequality measure used in the model. Time-lagged primary care measures were associated with a decrease of about 40 deaths per 1000,000 in 1985 to less than 30 deaths per 100,000 for in 1995. The magnitude of this association was similar to that of the contemporary primary care measure.

We also conducted time-lagged analyses examining the influence of primary care specialties on mortality (results not shown but available on request). As with contemporaneous analysis, only family medicine was statistically significant for all 4 periods (P < .01). An increase of 1 family medicine physician per 10,000 persons was associated with a decrease of about 70 deaths per 100,000. The internal medicine and pediatrics primary care specialties were inversely but not statistically significantly associated with mortality (P > .05).

Time-lagged specialty care was positively associated with mortality in 1985 and 1990 (P < .05), but statistical significance declined to the 90% level in 1995. The magnitude of the regression coefficient was similar for each income inequality measure at each period and declined from approximately 17 in 1985 to approximately 10 in 1995.

The adjusted R² measures range from 0.34 to 0.39, indicating that the models explain more than 30% of the variance in age-adjusted total mortality. There was no clear difference between adjusted R² values for models using the Gini coefficient versus those using the Robin Hood index.