Objectives. We examined the associations between socioeconomic position, co-occurrence of behavior-related risk factors, and the effect of these factors on the relative and absolute socioeconomic gradients in coronary heart disease.
Methods. We obtained the socioeconomic position of 9337 men and 39 255 women who were local government employees aged 17–65 years from employers’ records (the Public Sector Study, Finland). A questionnaire survey in 2000–2002 was used to collect data about smoking, heavy alcohol consumption, physical inactivity, obesity, and prevalence of coronary heart disease (myocardial infarction or angina diagnosed by a doctor).
Results. The age-adjusted odds of coronary heart disease were 2.1–2.2 times higher for low-income groups than high-income groups for both men and women, and adjustment for risk factors attenuated these associations by 13%–29%. There was no further attenuation with additional adjustment for the number of co-occurring risk factors, although socioeconomic disadvantage was associated with the co-occurrence of multiple risk factors. The absolute difference in coronary heart disease risk between socioeconomic groups could not be attributed to the measured risk factors.
Conclusions. Interventions to reduce adult behavior-related risk factors may not completely remove socioeconomic differences in relative or absolute coronary heart disease risk, although they would lessen these effects.
Coronary heart disease (CHD), a leading cause of morbidity and mortality in all Western countries, is more prevalent among lower socioeconomic position (SEP) groups than among groups that have higher SEP.1–7 Although the evidence of such a socioeconomic gradient in CHD is robust, the extent to which this gradient is the result of different distributions of coronary risk factors between SEP groups remains controversial. Several epidemiological studies suggested that most (60%–95%) of the CHD burden can be attributed to established risk factors: smoking, hypertension, diabetes, unfavorable cholesterol profile, and physical inactivity; appropriately, public health interventions target these risk factors to reduce the CHD epidemic.8–13 However, several studies that compare the magnitude of the socioeconomic gradient before and after adjustment for these risk factors suggest that they explain only 15%–40% of the association between SEP and CHD.7–14 Thus, the contradiction: most of the population burden of CHD can be attributed to established risk factors, but these same risk factors only explain a small part of the association between SEP and CHD. We raise the possibility that multivariable adjustment for risk factors may not have correctly estimated the contribution of these risk factors to SEP differences in the occurrence of CHD.12,15,16
In the most commonly used approaches, such as logistic regression and proportional hazards regression, risk factors are entered into the model to examine how they change the effects of SEP indicators on CHD.14 Any change in the effect estimate of SEP’s effects on CHD after adjustment for these risk factors is then used as an indication of the extent to which the risk factors “explain” the relative socioeconomic gradient in CHD. Underestimation might result from the failure of such models to fully account for clustering of individual risk factors. The overall effect of the risk factors would be underestimated if they were clustered (i.e., there was a greater than expected number of persons with either no risk factors or many risk factors), and this clustering was substantially more common in low- than high-SEP groups.17 Underestimation can also occur if the risk factors have synergistic effects (i.e., the effect of combined risk factors exceeds the predicted effects from separate risk factors, which assumes independence within the particular multivariable model).18 In most studies, the effect of synergism is examined by including interaction terms, but then removing them from the final model if the associated P value is large (conventionally >.05). However, most studies have a limited ability to detect multiple statistical interactions, and very large data sets are required to examine this possibility.
Multivariable adjustment may underestimate the effect of established risk factors when explaining most of the “excess” cases among the lower SEP groups (i.e., the effect of established risk factors on the absolute risk difference between SEP groups). A recent study of 2682 Finnish men in the Kuopio Ischemic Heart Disease Risk Factor Study found that although adjustment for smoking, hypertension, dyslipidemia, and diabetes resulted in a modest (24%) attenuation of the relative socioeconomic gradient of CHD risk, these same risk factors accounted for most (72%) of the absolute socioeconomic gradient, that is, the excess risk among those from the lowest SEP compared with those in the highest.19
The role of behavior-related risk factors in CHD, particularly those that might be modified through health promotion, is likely to lead to important policy implications. We used a large employee sample of people who were participating in the Finnish Public Sector Study20,21 to examine the associations between SEP, co-occurrence of behavior-related risk factors (smoking, physical inactivity, obesity, and heavy alcohol consumption),8,11,22–25 and the effect of these factors on the relative and absolute socioeconomic gradients in CHD.
The Finnish Public Sector Study focused on all local government employees of 10 towns and all employees in 21 public hospitals that provided specialized health care in the districts where the towns are located.20,21 These employees cover a wide range of SEPs, from city mayors to semiskilled cleaners. The largest groups were nurses and teachers. A total of 48 592 (9337 men, 39 255 women), aged 17–65 years responded to a questionnaire survey in 2000–2002 (response 68%). Women were slightly overrepresented among the respondents (81% women) compared with eligible employees (n = 70 961, 76% women), but the differences in mean age (44.7 vs 44.0 years) and SEP (15% vs 17% performing manual labor) were small. The gender and age of respondents were also representative of Finnish public sector employees (77% women; mean age 44.6 years).26 However, the predominance of women did not correspond to the gender distribution of the Finnish general working population (48% female; mean age 45.5 years).26
Occupational status, income, and education were used as indicators of SEP. We obtained the participants’ occupational titles from the employers’ records (1931 different titles)27 and used the occupational classification by Statistics Finland27 to classify individuals into 3 categories on the basis of these titles: upper nonmanual workers, lower nonmanual workers, and manual laborers. Average monthly income figures for men and women were obtained by occupational title from Statistics Finland,27 and the distribution was divided into thirds separately for the men and women (referred to as high-, intermediate-, and low-income groups). Educational level was self-reported in the surveys and was categorized as primary or secondary versus tertiary.
We assessed 4 behavior-related risk factors by using standard questionnaire measurements in the surveys. We requested the participants’ smoking status and the habitual frequency and amount of beer, wine, and spirits intake. Responses to the alcohol questions were transformed into units of alcohol per week.28 Binge drinking was determined by requesting whether the participant had passed out as a result of alcohol consumption more than once during the past 12 months. Physical activity was measured by the metabolic equivalent task index29 and was expressed as the sum score of metabolic equivalent task-hours per day (h/d). Self-reported weight and height were used to measure body mass index (kg/m2).
CHD was measured by using a self-administered checklist of common chronic diseases.30 For each disease, the respondent was asked to indicate whether or not a physician had diagnosed him or her as having the disease. Prevalent CHD was determined by affirmative responses for myocardial infarction or angina. The agreement between these self-assessments and data from medical records has previously shown to be substantial for myocardial infarction and angina (κ>0.70).30
We coded all of the risk factors as binary variables (0 or 1). Risks were defined as ever a smoker (current or past smoking), heavy alcohol consumption (>21 units of alcohol per week or binge drinking),22,23 physical inactivity (<2 metabolic equivalent task h/d),29 and obesity (body mass index>30 kg/m2).25 The age-adjusted association between SEP and binary risk factors was estimated in a logistic regression analysis that used upper nonmanual labor employees, high-income, and tertiary education as the reference groups. We counted the number of risk factors on the basis of these binary variables. Thus, the participants with all 4 risk factors had a score of 4; those with any 3 risk factors scored 3, and those with no risk factors scored 0. We performed a multinomial logistic regression analysis to examine the association between SEP and co-occurrence of risk factors (this analysis can assess associations that have a categorical outcome variable, such as ours).31 The multinomial models were used to assess the likelihood of having 1 risk factor, 2 risk factors, and 3 or 4 risk factors versus having no risk factors (the reference). The corresponding age-adjusted odds ratios were calculated for the levels of SEP by using the same reference groups that were used in the logistic regression analyses for each of the risk factors examined individually as described above. Finally, we tested the extent of clustered risk factors in the whole population and within each SEP group. The expected frequencies were those predicted given the prevalence of the risk factors within each SEP group (i.e., assuming independence or no clustering). Clustering is indicated when individuals are more likely to have no or many risk factors and are less likely to have a single risk factor than would be expected if the risk factors were independent (i.e., the observed-to-expected ratio is >1 for no risk factors, <1 for a single risk factor, and >1 for 3 and 4 factors). We calculated χ2 statistics to test the difference in the distributions of observed and expected counts within each occupational group.
The associations of SEP, risk factors, and the number of co-occurring risk factors with CHD were studied first with age-adjusted logistic models. To estimate the contribution of the risk factors to the association between SEP and CHD, each risk factor, all their interactions on a multiplicative scale, and the number of co-occurring risk factors were added to the model as covariates. We wanted to examine whether greater attenuation was achieved if the risk factors were more finely categorized and linear associations were not assumed. To do this, we split body mass index, physical activity (metabolic equivalent task h/d), and alcohol consumption (units per week) into fifths of their distributions and entered these into the regression model as 3 variables together with 4 indicator variables that represented past smoking, current smoking, high alcohol consumption (> 21 units per week), and binge drinking. Finally, we estimated the absolute risk of CHD associated with SEP in the whole population and in a low-risk group (anyone who was free of all of the measured risk factors) to determine how many excess cases among the lowest (compared with the highest) SEP group would be removed if these risk factors were eliminated.
The analyses were performed separately for men and women and for each SEP indicator using SAS 8.2 (SAS Institute Inc, Cary, NC) software. The findings were consistent across all of the 3 SEP indicators, so we reported full results for income only (indicator with evenly distributed categories) and have summarized the main findings for occupational status and education in the text.
Information was missing on income for 2227 (5%) of the participants, on any risk factor for 4044 (8%) of the participants, and on CHD for 4700 (10%) of the participants. A total of 39 631 employees (82% of all of the respondents) had full data for all of these variables. They differed slightly from participants who had some missing data in terms of gender (81% vs 79% women), mean age (44.2 vs 46.8 years), and SEP (15% vs 18% manual laborers). In spite of this, participants who had complete data were representative of all of the respondents (81% women, mean age 44.7 years, 15% manual).
In men and women, 31%–45% had no risk factors, 53%–63% had 1 or 2 risk factors, and ≤1% had all 4 risk factors. There were graded associations between SEP and each risk factor (except for heavy alcohol consumption). For both genders, the highest risk occurred for individuals who had the lowest SEP. The risk factors were all positively associated with CHD risk among both genders (odds ratios between 1.1 and 2.4). The one exception was heavy alcohol consumption, which was not associated with an increased incidence of CHD. The odds of CHD were 2.4–3.3 times greater for the men and women who had 3 or 4 risk factors than for those who had no risk factors. Statistical tests for interaction terms across the risk factors resulted in no strong evidence of synergism between the risk factors (P for all interaction terms ≥ .14).
Table 1 presents the age-adjusted association between SEP and the co-occurrence of risk factors. Having a low SEP, compared with a high SEP, was associated with 1.5–1.7 times higher odds of having a single risk factor for CHD versus having no risk factors. However, SEP had a stronger association with having 3 or 4 risk factors (odds ratios 2.4–3.3). The findings were similar when occupational status or educational attainment (instead of income) were used as the explanatory SEP variables.
In Table 2, the expected and observed numbers of participants who had 0, 1, 2, and 3 or 4 risk factors show that the risk factors were clustered within all of the SEP groups and that the clustering pattern was similar in the groups. These findings were also replicated with other SEP indicators.
Table 3 shows the multivariable association between SEP and CHD. For both genders, a simple adjustment for each risk factor entered into the model simultaneously as single covariates resulted in a 13%–29% reduction in the relative SEP gradient of CHD (20%–23% reduction after the binary risk factors were replaced with more finely categorized risk variables). No further attenuation in the relative socioeconomic gradient was found when we used a model that included both individual risk factors and a score that represented the total number of risk factors as covariates. Moreover, we used a backward elimination approach that removed all of the 4-, 3-, and 2-way interaction terms with P > .05 from a saturated model that included all of the risk factors and their interaction terms. This approach led to a final model that contained no interaction terms, a further indication that the interactions between the risk factors did not explain the association between SEP and CHD.
To illustrate what might happen to CHD risk and absolute SEP gradient if risk factors were removed from the population, we formed a low-risk subgroup that consisted of all the employees who had none of the measured risk factors (Table 4). Comparing this subgroup to the entire population suggests that CHD risk would have been reduced in all SEP groups by 6%–48%. However, a marked socioeconomic gradient in absolute risk remained in the subgroup that was free of the measured risk factors. The analyses with other SEP indicators replicated these findings.
Evidence from a large contemporary population suggests that SEP is associated with the co-occurrence of behavior-related risk factors such as smoking, heavy alcohol consumption, physical inactivity, and obesity. Men and women who have low SEP tend to have multiple risk factors more often than those who have higher SEP. Although these risk factors were clustered in the total sample, no evidence was found that clustering was more common in low-SEP groups or that the risk clusters had a synergistic effect on reported CHD. The effect of adjusting for risk factors on the relative socioeconomic gradient of CHD was to produce a modest (13%–29%) reduction, which is a similar magnitude to reduction found in several other studies.7,14,19 Furthermore, when we examined absolute risk, we found that by removing the behavior-related risk factors, important reductions in the prevalence of CHD in all SEP groups would result without removing the socioeconomic gradient. These findings were replicable across the 3 SEP indicators of income, occupational status, and education. Thus, according to this study, behavior-related risk factors do not explain a large part of either the relative or absolute socioeconomic gradient in prevalent CHD.
Few data have been published on the association between SEP and the clustering of risk factors, and those studies were on the basis of much smaller sample sizes than was used in our study. Consistent with our finding, an investigation of 2900 older women in the United Kingdom found evidence of risk factor clustering in all of the SEP groups and no evidence that the extent of the clustering was greater among those from the manual-labor SEP than among those from the non-manual ones.17
By contrast, a study of 3600 Australian adolescents found that both the co-occurrence and clustering of smoking, high levels of television watching, overweight, and high blood pressure were more common in families that had a low SEP than in other families.32 Similarly, a study of 480 young adults, aged 18–24 years, reported that clustering of behavior-related risk factors was more common among participants who had a history of unemployment and less common among students.33 Thus, clustering of risk factors may be more common among individuals who have lower SEP in adolescence and young adulthood, perhaps because, at these ages, peer pressure related to the initiation or noninitiation of some risk behaviors is very socially patterned. However, the socioeconomic gradient in clustering does not appear to be retained in later adulthood.
In this large study, we had the adequate ability to test for statistical interactions between behavior-related risk factors and their effect on CHD, yet we found none for either gender. Participants who had more behavior-related risk factors had greater risk of CHD, but this effect was consistent with the additive rather than synergistic effects of each risk factor. There was no evidence of differences in risk factor clustering by SEP and no evidence that the risk factors combined synergistically. It is, therefore, not surprising that we found that a more-complex adjustment (either including interaction terms or including both individual risk factors and a score for the number of risk factors) did not result in greater attenuation of the relative socioeconomic gradient in CHD than did simple adjustment for each individual covariate.
We examined absolute differences in the prevalence of CHD within a group that was free of all measured behavior-related risk factors in order to illustrate the expected effects of removing these risk factors from the population. Our findings confirmed the modest contribution of these risk factors to the absolute socioeconomic gradient. The low-risk population had a lower prevalence of CHD throughout the entire social hierarchy, but the absolute (and relative) socioeconomic gradient in risk remained. This finding is in contrast to that of the Kuopio Ischemic Heart Disease Risk Factor study, in which the absolute socioeconomic gradient largely disappeared in a subgroup that was free of measured risk factors.19 A potential reason for the discrepancy between these studies involves the selection of risk factors. Three risk factors: hypertension, dyslipidemia, and diabetes, which were included in the Kuopio study but not in our study, are physiological markers of the underlying pathophysiological processes that end in manifest CHD. Only 5% of the CHD events occurred among the low-risk population in the Kuopio study. Because the elimination of these major disease mediators appears to eliminate much of the absolute socioeconomic gradient in CHD, it is likely that any underlying factors, whether socioeconomic, psychosocial, psychological, early life or genetic, may have their major influence through these disease mediators. In our study, with the exception of obesity, all of the measured risk factors were purely exogenous, which reflects the lifestyle of the participants. Of the CHD cases, almost 30% had none of these behavior-related factors. Therefore, it seems that several etiological pathways to CHD that can be related to SEP remain uncovered by these more distal risk factors.
We determined all risk factors from self-reports. Although self-reported height and weight have been shown to be strongly correlated with direct measurement, obese individuals who self-report tend to underestimate their body mass index.33–36 If this systematic misreporting of weight is similar across SEP groups in our study, it would tend to dilute rather than exaggerate the magnitude of the associations we observed. Any variation in misreporting by social class could bias our results in either direction.
The cross-sectional design of this study is open to reverse causality, healthy-worker bias, and survivor bias. If individuals who are diagnosed with disease change to adopt a healthier lifestyle, the associations of risk factors with CHD may be underestimated, and the extent to which these risk factors explain socioeconomic gradients may also be underestimated. However, the magnitude of the associations that we found between SEP, behavior-related risk factors, and CHD are similar to those reported in prospective studies,37–43 which suggests that the cross-sectional nature of the study did not result in a major bias. The only exception was heavy alcohol consumption, which was not associated with CHD, even though it has predicted CHD events in several, though not all, prospective studies.37,40 Replication of our risk cluster analyses with a prospective investigation on incident CHD is important, but the challenge will be to achieve sufficient statistical power, which will require a very large cohort that is followed for many years.
Our findings may not be generalized to other populations. However, a socioeconomic gradient in CHD of a similar magnitude has been reported throughout several different European and US populations. In addition, the associations between behavior-related risk factors and CHD risk are similar across these different populations. Thus, it is likely that our findings could be generalized to most developed countries.
This study has shown that smoking, heavy alcohol consumption, physical inactivity, and obesity do not fully explain the socioeconomic gradient in CHD. However, our data, along with previous data, have demonstrated that these behavior-related risk factors have some explanatory power. Therefore, strategies aimed at reducing these risk factors would reduce CHD risk in the whole population and would also attenuate some of the socioeconomic gradient. Although the more proximal mechanisms through which CHD risk is generated are well understood, the determination of these factors (circulating lipid levels, blood pressure, and insulin resistance) is not fully understood. Further research is needed to determine additional ways to eliminate socioeconomic inequalities in CHD.
Note. CI = confidence interval. Risk factors are ex- or current smoking (prevalence for men and women, 47% and 33%), heavy alcohol consumption (defined as > 21 units of alcohol per week or binge drinking; 25% and 6%), physical inactivity (27% and 25%), and obesity (body mass index > 30 kg/m2; 13% and 11%). Participants with no missing values for any of the risk factors were included in these models. Note. P values were the same (P < .001) across all income levels for both genders. P values were on the basis of χ2 test with 3 degrees of freedom testing the null hypothesis of no differences in the observed and expected frequencies across all of the number of risk factor categories. Within all of the income groups, the risk factors were clustered with a greater-than-expected number of participants who had no risk factors, a lower-than-expected number who had 1 risk factor, and a greater-than-expected number who had 3 or 4 risk factors among the men and women. aGiven the prevalence of the risk factors within each group and on the basis of the assumption that all of the risk factors were independent of each other. Note. CI = confidence interval. aIncludes participants who had no missing values for any of the risk factors. bPercentage difference in the odds ratios for low income versus high income between the presented model and the age-adjusted model. aIncludes participants with no missing values for any of the risk factors. bAge adjusted. cPercentage difference in the excess risk for low income versus high income between the low-risk group and the total sample.
Odds Ratio (95% CI) Gender and Income No. Participants 1 vs 0 Risk Factors 2 vs 0 Risk Factors 3–4 vs 0 Risk Factors Men High 2742 1.00 1.00 1.00 Intermediate 2623 1.48 (1.29, 1.68) 1.67 (1.43, 1.94) 2.33 (1.86, 2.92) Low 2644 1.74 (1.52, 1.99) 2.32 (1.99, 2.70) 3.32 (2.66, 4.14) Women High 12 031 1.00 1.00 1.00 Intermediate 11 002 1.16 (1.10, 1.23) 1.28 (1.17, 1.39) 1.59 (1.34, 1.89) Low 11 482 1.47 (1.39, 1.55) 2.09 (1.93, 2.26) 2.39 (2.03, 2.82) High Income Intermediate Income Low Income Gender and No. of Risk Factors No. Exp No.a Ratio No. Exp No.a Ratio No. Exp No.a Ratio Men 0 1028 882 1.16 777 641 1.21 654 532 1.23 1 983 1197 0.82 997 1170 0.85 986 1142 0.86 2 566 554 1.02 615 658 0.94 721 757 0.95 3–4 165 109 1.52 234 154 1.52 283 283 1.33 Women 0 6048 5796 1.04 5015 4741 1.06 4367 4081 1.07 1 4428 4812 0.92 4272 4653 0.92 4695 5120 0.92 2 1318 1293 1.02 1400 1442 0.97 2008 2000 1.00 3–4 237 130 1.82 315 166 1.90 416 285 1.46 Odds Ratio (95% CI), Adjusted For Income and Gradient Change No.a Cases Age (Model A) Model A + Risk Factors (Model B) Model B + Interaction Terms Between Risk Factors Model B + No. Co-occurring Risk Factors (Model C) Men Income High 2719 54 1.00 1.00 1.00 1.00 Intermediate 2593 60 1.84 (1.25, 2.73) 1.59 (1.07, 2.36) 1.62 (1.09, 2.41) 1.60 (1.08, 2.38) Low 2583 78 2.24 (1.55, 3.24) 1.88 (1.29, 2.74) 1.93 (1.32, 2.82) 1.91 (1.31, 2.79) Change in gradientb 0% –29.0% –25.0% –26.6% Women Income High 11 886 70 1.00 1.00 1.00 1.00 Intermediate 10 903 102 1.57 (1.14, 2.16) 1.53 (1.11, 2.11) 1.54 (1.12, 2.12) 1.53 (1.11, 2.11) Low 11 218 158 2.12 (1.57, 2.84) 1.98 (1.47, 2.67) 1.97 (1.47, 2.68) 1.98 (1.47, 2.67) Change in gradientb 0% –12.5% –13.4% –12.5% Men Women Population and Incomea No.a Cases Riskb (per 10 000) Excess Riskb (per 10 000) No.a Cases Riskb (per 10 000) Excess Riskb (per 10 000) All participants Income High 2589 51 132.5 0 11 417 66 62.3 0 Intermediate 2472 57 266.5 134.0 10 334 91 88.0 25.7 Low 2419 73 334.1 201.6 10 400 140 129.7 67.4 Low risk group (no measured risk factors) Income High 973 11 62.5 0 5753 18 35.8 0 Intermediate 740 9 158.4 95.9 4719 35 73.6 37.8 Low 609 13 249.6 187.1 3966 52 125.3 89.5 Change in gradientc –7.2% +32.8%
Note. CI = confidence interval. Risk factors are ex- or current smoking (prevalence for men and women, 47% and 33%), heavy alcohol consumption (defined as > 21 units of alcohol per week or binge drinking; 25% and 6%), physical inactivity (27% and 25%), and obesity (body mass index > 30 kg/m2; 13% and 11%). Participants with no missing values for any of the risk factors were included in these models.
Note. P values were the same (P < .001) across all income levels for both genders. P values were on the basis of χ2 test with 3 degrees of freedom testing the null hypothesis of no differences in the observed and expected frequencies across all of the number of risk factor categories. Within all of the income groups, the risk factors were clustered with a greater-than-expected number of participants who had no risk factors, a lower-than-expected number who had 1 risk factor, and a greater-than-expected number who had 3 or 4 risk factors among the men and women.
aGiven the prevalence of the risk factors within each group and on the basis of the assumption that all of the risk factors were independent of each other.
Note. CI = confidence interval.
aIncludes participants who had no missing values for any of the risk factors.
bPercentage difference in the odds ratios for low income versus high income between the presented model and the age-adjusted model.
aIncludes participants with no missing values for any of the risk factors.
cPercentage difference in the excess risk for low income versus high income between the low-risk group and the total sample.
The work presented in this article was supported by grants from the Academy of Finland (projects 117604 and 105195) and the participating towns and hospitals. The work of Debbie A. Lawlor was supported by a United Kingdom Department of Health, Career Scientist Award.
Human Participation Protection This study was conducted according to the guidelines of the Helsinki declaration, and the study protocol was approved by the Ethics Committee of the Finnish Institute of Occupational Health.