Objectives. We examined income-related inequalities in self-reported health in the United States and Canada and the extent to which they are associated with individual-level risk factors and health care system characteristics.
Methods. We estimated income inequalities with concentration indexes and curves derived from comparable survey data from the 2002 to 2003 Joint Canada–US Survey of Health. Inequalities were then decomposed by regression and decomposition analysis to distinguish the contributions of various factors.
Results. The distribution of income accounted for close to half of income-related health inequalities in both the United States and Canada. Health care system factors (e.g., unmet needs and health insurance status) and risk factors (e.g., physical inactivity and obesity) contributed more to income-related health inequalities in the United States than to those in Canada.
Conclusions. Individual-level health risk factors and health care system characteristics have similar associations with health status in both countries, but they both are far more prevalent and much more concentrated among lower-income groups in the United States than in Canada.
Increasing evidence indicates that the roots of health inequalities lie in an array of social, economic, and political attributes of nation-states.1–4 Nations differ both in their average levels of population health and in the extent to which health is distributed unequally by socioeconomic status. Income-related inequalities in mortality, which are relatively stable in Canada,5 have been increasing steadily in the United States. In the early 1980s, the life expectancy gap in the United States between the poorest and most affluent decile was 2.8 years. By the late 1990s, this gap had increased to 4.5 years.6 The socioeconomic distribution of infant mortality in the two countries is also different, with declines across socioeconomic groups in Canada over recent decades,5 but widening gaps in the United States that are attributable to relatively higher declines for the affluent.7
What are the underlying causes of these disparities? Are determinants of population health distributed differently by income in these neighboring countries? Or is the association between these determinants and health stronger in the United States than in Canada? We aimed to identify the potential influence of health care and other policies on income-related inequalities in health by decomposing those inequalities in both the United States and Canada into relative contributions from a set of known determinants of health.
Our data came from the Joint Canada–US Survey of Health (JCUSH), a 1-time random-digit-dialed telephone survey conducted in both countries in 2002 to 2003.8 The JCUSH collected information on a wide range of health status and health system factors.
The JCUSH had a national sampling frame, with stratification by province in Canada and 4 geographic regions in the United States (Northeast, Midwest, West, and South). The target population included individuals 18 years or older residing in private dwellings with a landline telephone. Response rates were 50% and 66% for the United States and Canada, resulting in sample sizes of 5183 in the United States and 3505 in Canada. Poststratification adjustments were made and weights were applied to ensure that the sample reflected population estimates derived from the 2002 Current Population Survey in the United States and the 1996 Census of Population in Canada. The auxiliary variables used to create the poststratification adjustments were age, gender, and race/ethnicity for the United States and age, gender, and region for Canada.8 We used these weights in all analyses.
The outcome variable of interest was health-related quality of life as measured by the Health Utilities Index (HUI), version 3. The HUI is a multidimensional, preference-based cardinal measure of health that has been used both in clinical settings and in studies of population health.9–13 It has a theoretical range between −0.3 (living in a state worse than death) and 1 (perfect health) and is intended to capture an individual's overall health utility via responses to a series of questions covering 8 domains: vision, hearing, speech, mobility, dexterity, cognition, emotion, and pain. These domains are transformed into an overall score by a multiattribute utility function.14,15 A difference in HUI of 0.03 is considered practically significant, meaning that the difference would be meaningful to or discernible by individuals.16
The independent variables represented factors that are known to be associated with individual-level health and thus may be related to inequalities in health status at the population level.17–21 The domains included demographic characteristics (age, gender, marital status, race/ethnicity), socioeconomic status (education, income), individual-level (lifestyle) risk factors (body mass index, smoking status, physical activity), and health care system factors (access to a regular medical doctor, unmet needs for health care, insurance for hospital and physician services, insurance for drugs).1 Respondents were dropped if they had missing data for any of these variables or if their HUI was less than zero. The final samples used in the analyses included 3574 respondents in the United States and 2744 in Canada. The distribution of health across key variables was unaffected by the loss of respondents (data not shown).
Several standard approaches exist for measuring relative inequalities in health by income level. We used virtually identical methods to those proposed by van Doorslaer and Jones.20 These methods were developed in the health economics field20,22–25 and have gained increasing acceptance in public health and epidemiology.26
In brief, a health Lorenz curve represents the cumulative proportion of a population's health as a function of the cumulative proportion of the population, where the population is ranked from lowest to highest health (Figure 1a). The situation representing perfect equality—the case in which every person has the same health—is shown by the diagonal line. In general, a health Lorenz curve will lie below the diagonal, because the health distribution in any population will be unequal. A Gini coefficient for health inequality measures the deviation from the line of equality; it ranges from 0 to 1 and is defined as twice the shaded area shown in Figure 1.
Similar principles can be applied for measuring the degree of inequality in health by another ranking variable, such as income. In this case, individuals are ranked not by their health but by their income along the x-axis, and the cumulative distribution of health is represented by a concentration curve.27 Concentration indexes are calculated the same way as Gini coefficients but can range from −1 to 1, depending on whether the health (or ill health) variable is more concentrated among individuals with lower income (negative concentration index values) or higher income (positive concentration index values). Gini coefficients and concentration indexes can be computed conveniently by regression.22,23 We focused mainly on income-related inequalities because these can be considered unfair and are at least potentially amenable to policy.
Wagstaff et al. showed that a linear regression equation relating a dependent health variable to a set of explanatory variables can be used to decompose the health concentration index into a weighted sum of the concentration indexes of each of the independent variables, with the weights equal to health elasticities.24 Health elasticities are the product of the variable's (partial) regression coefficient and a ratio of the variable's mean over the mean of the dependent variable.
In other words, the overall inequality in a population's health can be decomposed into the inequality in each of its contributing factors, weighted by the influence of each of those determinants. For any factor to help explain income-related health inequalities, there are 2 requirements: it must be unequally distributed by income (i.e., have a nonzero concentration index), and it must show a partial association with health after control for other factors (i.e., have a nonzero regression coefficient).
The required steps to compare income-related health inequalities in the United States and Canada were (1) to measure overall inequality in health by income in each country with a concentration curve and index; (2) for each country, to explain individual health with a linear regression of health on a set of explanatory variables to obtain partial regression coefficients; and (3) to decompose the overall inequality in health by income into contributions that arose from the strength of association of each variable with health (i.e., partial regression coefficients), the prevalence of each variable, and the extent to which its distribution in the population was related to income. Of course, the resulting decomposition depended on the validity of the underlying explanatory regression model.
The reference category individual in the model was a married White man aged 18 to 44 years with a high income and a postsecondary degree who did not smoke, was physically active, had a regular doctor and no unmet health care needs, had insurance covering pharmaceuticals, and (in the United States) had private health insurance.
Mean health (e.g., utility) was slightly lower in the United States than in Canada (Table 1), although this difference was not statistically significant. Mean health decreased with age, as expected, and was significantly higher in Canada than in the United States among the 2 lowest education groups and the lowest income group.
Sample Characteristics and Average Health Utility Score: Joint Canada–United States Survey of Health, 2002–2003
|No.||HUI, Mean Score||No.||HUI, Mean Score||HUI Gapa||Pb|
|Overall health-related quality of life||3574||0.881||2744||0.889||0.008||.111|
|Less than high school||332||0.786||539||0.822||0.036||.049|
|Equivalized household income,c $|
|< 10 000||385||0.757||195||0.818||0.061||.008|
|10 000–20 000||630||0.837||518||0.823||−0.014||.467|
|20 000–30 000||711||0.884||583||0.879||−0.005||.658|
|30 000–50 000||927||0.911||849||0.910||−0.001||.682|
|> 50 000||921||0.928||599||0.933||0.004||.469|
|Gini Coefficientd (95% CI)||0.093 (0.087, 0.099)||0.085 (0.079, 0.091)||0.007||.044|
|Concentration indexe (95% CI)||0.031 (0.026, 0.036)||0.026 (0.021, 0.030)||0.006||.048|
Note. HUI = Health Utility Index; CI = confidence interval.
aThe HUI is a multidimensional, preference-based cardinal measure of health. It has a theoretical range between −0.3 (living in a state worse than death) and 1 (perfect health) and is intended to capture an individual's overall health utility through responses to a series of questions covering 8 domains: vision, hearing, speech, mobility, dexterity, cognition, emotion, and pain. The HUI gap was calculated by subtracting the average HUI in the United States from the average HUI in Canada.
bTwo-tailed tests for difference.
cHousehold income was adjusted for household size. Income was reported in local currency.
dThe Gini coefficient measures inequality in the distribution of a variable, in this case health, with a range from 0 to 1 (higher numbers indicate greater inequality).
eThe concentration index measures inequality by the distribution of 2 variables, such as income and health. It ranges from −1 to 1, depending on whether the variable of interest is more concentrated among individuals with lower income (negative concentration index values) or higher income (positive concentration index values).
The difference in Gini coefficients quantified the greater overall health inequality in the United States than in Canada (Table 1). Both overall health inequality and income-related inequalities in health were smaller in Canada, as shown by the concentration indexes. Although health was more concentrated among higher-income groups in both countries, this concentration was more pronounced in the United States. In the concentration curves for both countries (Figure 1b), the Canadian curve was always closer to the line of equality. This implies that the population's accumulated health was better in Canada along the full range of the income distribution.
Table 2 provides the results of the regression analyses of individual-level variables on health. The regression coefficients can be interpreted as the reduction in HUI associated with a change in the explanatory variable, holding all else constant.
Decomposition Analysis of Factors Contributing to Income-Related Inequalities in Health: Joint Canada–US Survey of Health, 2002–2003
|Regression Coefficient (95% CI)||Prevalence, %||Concentration Indexa||Contribution, %||Regression Coefficient (95% CI)||Prevalence, %||Concentration Indexa||Contribution, %|
|Aged 45–64 y||−0.063 (−0.085, −0.043)||16.7||0.201||−7.9||−0.023 (−0.043, −0.004)||16.5||0.222||−4.2|
|Aged ≥ 65 y||−0.053 (−0.086, −0.019)||7.3||−0.125||1.8||−0.069 (−0.098, −0.040)||9.1||−0.177||5.4|
|Aged 18–44 y||0.008 (−0.008, 0.024)||26.2||−0.068||−0.5||−0.005 (−0.023, 0.013)||26.2||−0.058||0.4|
|Aged 45–64 y||−0.044 (−0.063, −0.026)||18.0||0.077||−2.3||−0.032 (−0.052, −0.012)||16.1||0.046||−1.2|
|Aged ≥ 65 y||−0.055 (−0.090, −0.019)||10.4||−0.218||4.6||−0.084 (−0.123, −0.045)||9.3||−0.292||11.1|
|Widowed||−0.045 (−0.081, −0.009)||8.2||−0.199||2.8||−0.017 (−0.053, 0.021)||7.7||−0.242||1.5|
|Separated/divorced||−0.018 (−0.036, 0.002)||15.9||−0.092||1.0||−0.065 (−0.090, −0.039)||12.3||−0.112||4.4|
|Single/never married||−0.004 (−0.021, 0.011)||18.6||−0.090||0.3||−0.019 (−0.036, −0.002)||20.2||−0.034||0.6|
|Black||0.010 (−0.015, 0.036)||7.2||−0.229||−0.6||0.000||0.0|
|Other||0.008 (−0.010, 0.026)||14.7||−0.302||−1.3||−0.012 (−0.030, 0.006)||15.1||−0.109||0.9|
|Less than high school||−0.045 (−0.072, −0.010)||9.3||−0.606||9.5||−0.054 (−0.077, −0.030)||19.6||−0.301||15.5|
|High school||−0.023 (−0.035, −0.007)||35.4||−0.155||4.7||−0.016 (−0.031, −0.001)||28.8||−0.082||1.9|
|Some postsecondary||−0.017 (−0.035, 0.002)||14.1||−0.027||0.2||−0.012 (−0.038, −0.004)||22.0||0.015||−0.2|
|Log of income adjusted for household size||0.027 (0.018, 0.040)||10.273c||0.045||46.4||0.026 (0.017, 0.037)||10.298c||0.038||49.8|
|Current smoker||−0.014 (−0.030, 0.002)||22.7||−0.133||1.6||−0.014 (−0.030, 0.002)||25.8||−0.116||2.0|
|Underweight||−0.043 (−0.091, 0.006)||2.2||−0.218||0.8||−0.004 (−0.040, 0.032)||2.7||−0.268||0.1|
|Overweight||−0.001 (−0.015, 0.012)||33.8||0.011||0.0||−0.016 (−.030, −.002)||34.3||0.063||−1.7|
|Obese||−0.036 (−0.054, −0.018)||21.6||−0.085||2.5||−0.039 (−0.060, −0.017)||15.6||−0.038||1.1|
|Moderately active||−0.018 (−0.032, −0.002)||22.3||0.110||−1.6||−0.007 (−0.022, 0.008)||27.4||0.057||−0.6|
|Inactive||−0.046 (−0.059, −0.031)||55.8||−0.095||9.0||−0.036 (−0.050, −0.022)||46.1||−0.084||6.8|
|Health care system factors|
|No regular doctor||0.026 (0.010, 0.043)||19.0||−0.169||−3.1||0.027 (0.012, 0.043)||14.9||−0.023||−0.5|
|Unmet needs||−0.104 (−0.130, −0.077)||12.5||−0.217||10.6||−0.130 (−0.161, −0.099)||11.2||−0.119||8.4|
|Health insurance status|
|Public insurance only (not Medicaid)||−0.076 (−0.116, −0.035)||5.1||−0.278||4.0||0.000||0.0|
|Public plus private/other insurance||−0.036 (−0.058, −0.014)||21.7||−0.042||1.2||0.000||0.0|
|Medicaid||−0.136 (−0.181, −0.091)||5.3||−0.570||15.6||0.000||0.0|
|No insurance||−0.035 (−0.065, −0.001)||10.6||−0.449||6.2||0.000||0.0|
|No pharmaceutical insurance||0.020 (−0.003, 0.043)||19.9||−0.353||−5.3||0.009 (−0.007, 0.025)||21.9||−0.185||−1.8|
Note. BMI = body mass index; CI = confidence interval.
aThe concentration index measures inequality by the distribution of 2 variables, such as income and health. It ranges from −1 to 1, depending on whether the variable of interest is more concentrated among individuals with lower income (negative concentration index values) or higher income (positive concentration index values).
bBMI based on WHO categories: underweight was < 18.5 kg/m2, normal weight was 18.5 kg/m2–24.9 kg/m2, overweight was 25 kg/m2–30 kg/m2, and obese was > 30 kg/m2.
cLog of income value.
As expected, health decreased with greater age in both the United States and Canada. The sole exception to a clear age–health gradient was for men aged 45 to 64 years in the United States, among whom we found a stronger negative association with health than for men 65 years and older. This phenomenon occurred when the health insurance variables were added to the model and appeared to reflect the relationship between age and insurance status in the United States. It may also have related, in part, to the fact that these analyses did not include information on labor force participation, which is associated with both health and type of health insurance in the United States.
We observed a positive relationship between education and health in both countries, and increasing income was associated with greater health. We found a negative association between health and being widowed (in the United States) and between health and either being separated or divorced or being single or never married (in Canada). Individuals who were obese or physically inactive also reported lower health (all else being equal). We found a strong association in both countries between reporting an unmet need for health care and health status and a positive association between reporting not having a regular doctor and health status. This latter group likely included both those who wanted and could not find a regular doctor and those who were in good health and were content not to have one.
Finally, the insurance variables in the United States were significantly and negatively associated with health status. It is worth noting that race/ethnicity did not attain significance in the final model. Although self-reported Black race/ethnicity in the United States was associated with poorer health, this relationship disappeared when the income variable was added.
Table 2 also provides the results of the decomposition analyses. Each variable's contribution to inequality can be read as follows: in the United States, with all else equal, men aged 45 to 64 years reported health that was on average 0.063 units lower (on the 0–1 HUI scale) than that of men aged 20 to 44 years (the reference group). The former group represented 16.7% of the sample and was more concentrated in higher-income groups (the positive concentration index value). As a result, men aged 45 to 64 years contributed −8% to the measured degree of income-related health inequalities. That is, measured income-related health inequalities would have been 8% higher if each income category contained equal proportions of men aged 45 to 64 years or if there were no relationship between this age–gender group and health (with all else equal).
The same age group in Canada contributed only −4% to health disparity, mainly because middle-aged men reported being only somewhat less healthy than the younger age group (relative to younger men, a health utility only 0.023 units lower), and they were only slightly more concentrated in the higher-income groups than were their peers in the United States (concentration index of 0.22 for Canada vs 0.20 for the United States). A similar interpretation and comparison pertained for each of the other explanatory variables.
By far the single most important variable associated with income-related health inequalities was income itself, accounting for close to 50% of the total inequalities in both countries. This importance was a result of both an unequal distribution of income and a strong positive association between income and health that remained even after we controlled for all other factors. Despite the much greater concentration of individuals with less than a high school education in lower-income groups in the United States, the group with the lowest educational status contributed more to overall income-related health inequalities in Canada. This contribution can be explained by this group's stronger negative association with health and by its larger size—nearly 1 in 5 Canadian respondents had less than a high school education, compared with only approximately 1 in 10 US respondents.
The measured risk factors did not contribute equally to inequality. Current smoking status contributed approximately 2% in both countries. Obesity contributed approximately 2.5% to income-related inequalities in the United States, compared with only approximately 1% in Canada. The United States had a higher prevalence of obesity, which was more concentrated among lower-income groups. Physical inactivity also contributed more to inequalities in the United States (9.0% compared with 6.8% in Canada) because of its stronger relationship with health, its higher prevalence, and its greater concentration in lower-income groups.
Universal insurance for hospital and physician services in Canada implies that coverage, by definition, could not contribute to income-related inequalities in health. Conversely, there was an almost 16% contribution from Medicaid coverage in the United States, because of its strong relationship with health and its concentration in lower-income groups. This does not imply, of course, that Medicaid coverage leads to poorer health or that removing Medicaid would reduce health inequalities. Instead, it reflects the fact that people who qualify for Medicaid tend to be both poor and sick. These effects persisted after statistical adjustment, meaning that Medicaid recipients tended to have far worse health even after we controlled for other factors, including income. This result may have been because income was measured only in quintiles, and Medicaid recipients were likely to be at the lower incomes even within the lowest-income quintile. Lack of insurance contributed 6.2% to income-related inequalities, and having public insurance contributed another 4%. The public insurance group was also very much concentrated at the lower end of the income spectrum, and, at 5% of the population, may have represented older Americans who lacked insurance coverage supplementary to Medicare.
The larger relative contribution to income-related inequalities in health of reported unmet needs in the United States (10.6%) than in Canada (8.4%) was driven mostly by the fact that unmet needs were more concentrated at lower incomes in the United States.
Figure 2 aggregates the percentage contributions to income-related inequalities in health of individual explanatory variables into broader categories. Socioeconomic factors explained more than 60% of income-related health inequalities in the United States and more than two thirds of those in Canada. The next most influential set of variables for the United States was health care system factors (29%), in particular insurance status, contributing approximately 22% of overall income-related health inequalities. In Canada the second most important factor was demographic characteristics, at approximately 20%. Lifestyle-related risk factors also contributed more to income-related health inequalities in the United States than they did in Canada.
Our analyses add to the existing evidence with new data and new methods. First, our analyses were based on a unique data set derived from a jointly developed and administered survey, which allowed for a direct comparison of the United States and Canada. The survey included a set of variables that captured comparable information for each country about both the health care system and individual-level demographic characteristics and risk factors. Second, our measure of individual health was the HUI, a generic measure of health-related quality of life. This allowed us to move beyond common comparisons of income-related mortality to an examination of inequalities in general morbidity as well. Third, we employed a decomposition method that enabled the identification of the relative degree to which various factors contributed to income-related inequalities in health.25
In general, we found average self-reported health (e.g., utility) to be lower and relative inequality in health to be greater in the United States than in Canada. In the context of our decomposition method, differences between countries in contributions to income-related health inequalities could only occur for 1 or a combination of the following reasons: (1) the relationship between health and its determinants was different, (2) the prevalence of health determinants was different, or (3) health-determining factors were distributed differently across income groups (as reflected in their concentration indexes).
The decomposition analysis revealed that income itself accounted for the single largest share of the measured degree of income-related health inequalities in both countries. This was consistent with findings for European countries,25 and the possibility of several causal pathways in both directions remained. Cross-country differences in income protection for people with poor health or disabilities are an example of possible reverse causation at work here. Our analysis was not able to examine such causation.
Health care system factors contributed far more to income-related health inequalities in the United States than in Canada. This was driven largely by the unequal distribution of health insurance, which accounted for more than one fifth of income-related inequalities in health in the United States. Self-reported unmet needs also made a larger contribution to income-related health inequalities in the United States than in Canada, mainly because they were far more concentrated among lower-income groups in the United States.
Physical inactivity and obesity were both more prevalent in the United States than in Canada and were also (especially obesity) more concentrated in lower-income groups. This explains why—despite the similar relationship between health and risk factors in the 2 countries—risk factors contributed more to income-related health inequalities in the United States.
The data and methods used in our analysis had some important limitations. First, the sample size of the JCUSH was relatively small and suffered from relatively high nonresponse rates. Although the low response rates may have created selection bias, comparisons of JCUSH results with other nationally representative surveys suggest that the data were still representative after applying sample weighting.8
In addition, the results of the decomposition analysis depended on the validity of the underlying regression model. These models cannot be given a causal interpretation. They are only helpful in unraveling the partial associations between the dependent and independent variables that were included. Thus, they do not allow for an assumption that, for example, if income inequality was reduced by X%, or obesity by Y%, health inequality would go down by Z%.
Previous analyses of the JCUSH data identified a health advantage for Canadians compared with Americans at the lower end of the income and education distributions28,29 and a strong positive relationship between having health insurance and reporting access to health care.30 Comparisons of mortality between the United States and Canada showed that lower-income groups had lower mortality rates in Canada than in the United States for specific causes such as cancer.31–33 In addition, mortality from causes amenable to health care have declined at a greater rate in Canada than in the United States since 1980, which was after the establishment of universal health insurance in Canada.34 Thus, there is increasing evidence that differences in health care are part of the explanation of higher average health status in Canada than in the United States3,35
We found that the relationships between health and its determinants were not too dissimilar in Canada and the United States, but that both risk factors associated with lifestyle and health care system factors associated with poorer health care access were more prevalent in the United States than in Canada and much more concentrated among lower-income groups. Because these differences mainly affected the low end of the income distribution, the Canadian health concentration curve dominated the US curve across the entire spectrum, implying that at every percentile point in the cumulative income distribution, Canadians had accumulated more health than had Americans.
Social and economic inequalities have been theorized to contribute to income-related differences in health,30,36,37 and our analyses provide more detail about potential pathways for those relationships. Behavioral risk factors are correlated with income everywhere, but it appears that this effect is stronger in the United States than it is in Canada. Further research is needed to identify the factors that put people at risk of inequality and the reasons for their differential effects across countries.
This study was supported by Statistics Canada and Erasmus University.
We are grateful to Jillian Oderkirk and Steven Lewis for thoughtful comments on earlier drafts of this article and to Dawn Mooney for creation of the figures.
Human Participant Protection
No protocol approval was required for this study, which used previously collected survey data. Permission to use the survey data was granted through institutional review and approval by Statistics Canada.