Eur J Epidemiol DOI 10.1007/s10654-012-9745-z
CARDIOVASCULAR DISEASE
Education and risk of coronary heart disease: assessment of mediation by behavioral risk factors using the additive hazards model Helene Nordahl • Naja Hulvej Rod • Birgitte Lidegaard Frederiksen Ingelise Andersen • Theis Lange • Finn Diderichsen • Eva Prescott • Kim Overvad • Merete Osler
•
Received: 16 April 2012 / Accepted: 5 November 2012 Ó Springer Science+Business Media Dordrecht 2012
B. L. Frederiksen M. Osler Research Center for Prevention and Health, Glostrup University Hospital, Glostrup, Denmark
proposed methods for mediation based on the additive hazards model, and compared with results from the Cox proportional hazards model. Short (vs. long) education was associated with 277 (95 % CI: 219, 336) additional cases of CHD per 100,000 person-years at risk among women, and 461 (95 % CI: 368, 555) additional cases among men. Of these additional cases 17 (95 % CI: 12, 22) for women and 37 (95 % CI: 28, 46) for men could be ascribed to the pathway through smoking. Further, 39 (95 % CI: 30, 49) cases for women and 94 (95 % CI: 79, 110) cases for men could be ascribed to the pathway through BMI. The effects of low physical activity were negligible. Using contemporary methods, the additive hazards model, for mediation we indicated the absolute numbers of CHD cases prevented when modifying smoking and BMI. This study confirms previous claims based on the Cox proportional hazards model that behavioral risk factors partially mediates the effect of education on CHD, and the results seems not to be particularly model dependent.
B. L. Frederiksen National Board of Health, Hospital Services and Emergency Management, Copenhagen, Denmark
Keywords Education Coronary heart disease Behavioral risk factors Mediation
Abstract Educational-related gradients in coronary heart disease (CHD) and mediation by behavioral risk factors are plausible given previous research; however this has not been comprehensively addressed in absolute measures. Questionnaire data on health behavior of 69,513 participants, 52 % women, from seven Danish cohort studies were linked to registry data on education and incidence of CHD. Mediation by smoking, low physical activity, and body mass index (BMI) on the association between education and CHD were estimated by applying newly H. Nordahl (&) N. H. Rod I. Andersen F. Diderichsen Section of Social Medicine, Department of Public Health, Faculty of Health Sciences, University of Copenhagen, Øster Farimagsgade 5, 1014 Copenhagen, Denmark e-mail:
[email protected]
T. Lange Section of Biostatistics, Department of Public Health, University of Copenhagen, Copenhagen, Denmark E. Prescott Department of Cardiology, Bispebjerg University Hospital, Copenhagen, Denmark E. Prescott The Copenhagen City Heart Study, Bispebjerg University Hospital, Copenhagen, Denmark K. Overvad Department of Epidemiology, School of Public Health, Aarhus University, Aarhus, Denmark
Introduction A number of studies have shown that lower socioeconomic position is associated with a higher risk of cardiovascular morbidity and mortality [1–11]. A better understanding of the mechanisms responsible for the gradients may enable the development of public health interventions, because such knowledge may help identify those modifiable risk factors that mediate or interact with the effects of socioeconomic position [12, 13]. Modifications of the fundamental social stratification may be desirable, but is often
123
H. Nordahl et al.
unfeasible and not directly within the realm of public health. However, one may be more optimistic about changing risk factors mediating the social inequality in cardiovascular diseases. Quantitative assessment of mediation is therefore crucial for prioritizing between risk factors to tackle such health inequalities. Although established behavioral risk factors for cardiovascular diseases such as smoking, physical inactivity, alcohol use, and diet are unequally distributed between different socioeconomic groups, the pathways through which socioeconomic position affects incidence of cardiovascular diseases are not fully understood and may vary in both time and place [2]. According to previous studies, most of the excess risk of cardiovascular mortality among men and women with short education is attributable to behavioral risk factors [7–9, 14, 15]. Other studies have illustrated that both behavioral and biological risk factors only partially explain the socioeconomic gradients of cardiovascular diseases [2, 6, 10, 16]. These previous evaluations of the mechanisms behind social inequalities in cardiovascular diseases have primarily been based on reductions in relative risk by the use of proportional hazards model. In the present study we applied an alternative absolute risk approach to understand social inequalities in coronary heart disease (CHD). Our main objective was to quantify how much of the effect of educational level on risk of CHD was mediated through three behavioral risk factors, i.e. smoking, low physical activity, and BMI. To separate the effects mediated (indirect effect) and not mediated (direct effect) through behavioral risk factors we used contemporary techniques that provides a simple measure of mediation and quantifies additional number of CHD cases per unit of time due to educational level, and decompose this number to additional cases attributable to the direct and indirect pathway [17, 18]. Further, we compared this alternative mediation method with the more common approach.
Methods The SIC Cohort Study This study was based on the recently established Social Inequality in Cancer (SIC) Cohort Study, in which prospective data from the following seven Danish cohort studies were combined: the Copenhagen City Heart Study [19], the Diet, Cancer and Health Study [20], and five selected studies from the Cohorts at the Research Centre for Prevention and Health [21]: MONICA I, II, and III, the 1936 Cohort Study, and the Inter99 Study. Of the 83,000 participants enrolled in the SIC Cohort Study 7,483 was excluded due to diagnose of cardiovascular disease registered in the Danish National Patient Registry before
123
baseline. To obtain a robust measure of education, 491 participants age 29 or below were excluded (we assume that these people may not yet have reached their final level of education), as were 3,863 participants born before 1921 (the central registries did not have information on education from that period). All other participants with missing information on education were also excluded (n = 1,683). Thus 69,513 participants (36,969 women) were included in our analyses. Follow-up and end point Follow-up was achieved through linkage of the cohort data to national registers. Information on fatal and nonfatal cases of CHD was obtained from the Registry of Causes of Death and the Danish National Patient Registry, using the International Classification of Diseases: 8th revision code 410-4, and 10th revision code I20-I25. All participants were followed from baseline examination between 1981 and 2001 until the first incident case of CHD, death, emigration, or end of follow-up on December 31, 2009. The median follow-up time (5th and 95th centile range) was 13.7 years (5.5–27.2). Educational level Information on educational level was obtained by data linkage to the Integrated Database for Labour Market Research administered by Statistics Denmark since 1980. The highest attained educational level 1 year before baseline examination was categorized in three groups: short education (8–11 years of basic schooling), medium education (11–14 years; upper secondary or vocational education), and long education (C15 years of education). Behavioral risk factors All subjects in the SIC Cohort Study had participated in one of the seven population-based cohort studies, and information about behavioral factors had been collected by use of self administrated questionnaire and physical examinations. Detailed lists of what has been measured in each cohort are available elsewhere [19–23]. The establishment of the SIC Cohort Study was centered on a series of consensus meetings bringing together researchers who had great expertise in using data from the different cohort studies. Its construction was based upon flexible procedures [24] to retrospectively harmonize data by generating a selected list of core variables providing the basis for harmonization. Further, we used formal pairing rules to create each variable. Although the phrasing of questions differed in various cohorts it was possible to maintain sufficient consistency in the synthesized data set.
Education and risk of coronary heart disease
Information on smoking was based on questions covering current smoking status, and dichotomized into smoker/ ex-smoker versus never-smoker. Physical activity in leisure time was harmonized into a four-category variable about the participants’ usual physical activity in the past year, and dichotomized by collapsing sedentary and light activity into low physical activity indicating less than 4 h of light activity (housekeeping, walking, bicycling) per week, and collapsing moderate and high activity into moderate/high physical activity. The participants were also given a health examination with anthropometric measurements and body mass index (BMI) was calculated as weight (kg)/height (m2), based on data collected by trained nurses. A risk factor score was defined based on all three behavioral risk factors combined. The participants were defined as having a low risk score when all of the following applied at baseline: BMI \30 kg/m2, never-smokers, and moderate/high physical activity. Statistical method Distributions of baseline characteristics in the three educational levels were compared by one-way ANOVA F-test (continuous data) and Pearson v2-test (categorical data). Initially, we assessed the impact of risk factor adjustments on the association between educational level and CHD by means of the Cox proportional hazards model, with age as the underlying time scale. The proportional hazards assumption was evaluated for all variables by comparing estimated log– log survivor curves and by tests based on the generalization of Grambsch and Therneau [25]. The percent reduction in excess hazard ratio of CHD attributable to risk factors was calculated as previously reported [4, 7, 9, 11]. Drawing causal conclusions from a mediation analysis whether it is based on the Cox proportional hazards model, the additive hazards model or any other model, we must assume that there are no unmeasured confounders for; (1) the exposure–outcome, (2) the mediator–outcome, and (3) the exposure–mediator relationship, and we assume that there is no effect of the exposure that confounds the mediator–outcome relation (no exposure-dependent confounders) [18, 26]. We used crude mathematical derivations of Corollary 2 in [27] to estimate bias in the estimates of direct and indirect effect due to confounding (by unmeasured ‘country of birth’) of the exposure–mediator relationship and of the mediator-outcome relationship. Further, we examined potential exposure-dependent confounding by physical activity on the education–BMI–CHD relationship by means of auxiliary analyses in the Cox proportional hazards model. If the included confounders are indeed adequate to ensure the absence of unmeasured confounders the causal effect of exposure, e.g. educational level, can be separated into a
component mediated through behavioral risk factors and a direct effect component by applying an additive hazards model [17, 18]. In the additive hazards model absolute changes rates between groups (e.g. short vs. long education and medium vs. long education) are estimated. Thus the effect estimates are directly interpretable as additional CHD cases per person year compared to the reference group. The first step of this approach was to estimate the effects of educational level on each of the three behavioral risk factors adjusted for baseline confounders (age and cohort) by linear regression models. Diagnostic tests indicated that the models were well specified. The second step was to estimate the effect of both educational level and each of the behavioral risk factors on the onset of CHD by fitting the additive hazards model using age as the underlying time scale and adjusting for cohort. Standard techniques [28] found no indication of age-dependent effects of either educational level nor did behavioral risk factors, i.e. education level affect the rate of CHD onsets by equal amounts at all age levels. In addition, no interaction terms between educational level and behavioral risk factors were statistically significant. The indirect effect (IE) through each of the behavioral risk factors was given by the product of the parameter estimates for the regression of educational level on the mediator and the parameter estimate of the effect of the mediators on CHD from the additive hazards model. The direct causal effect (DE) of exposure was given directly from the additive hazards model. Finally, the total effect (TE) was found as the sum of the direct and indirect effects. For the direct effect, 95 % confidence limits were readily available from the additive hazards model, while limits for the indirect and total effects were computed by bootstrap (using 100,000 replications). All analysis was preformed separately for women and men, and fitted to adjust for potential confounding from baseline covariates. We used our prior knowledge and the methods of Directed Acyclic Graphs [29] to identify potential confounders which in addition to age and sex included cohort study.
Results A total of 8,046 (11.6 %) participants were diagnosed with CHD during the follow-up time: 3,132 women and 4,917 men. Baseline levels of behavioral risk factors differed between the three educational levels (Table 1). Women and men with short education were more likely to be current smokers, have a sedentary leisure time, have a higher BMI, and have a less favorable risk factor score than those with long education. All three behavioral risk factors and the risk factor score were found to be associated with risk of CHD (data not
123
H. Nordahl et al.
shown). Being a smoker or ex-smoker was associated with a higher risk of CHD in women (hazard ratio (HR) = 1.39, 95 % confidence interval (CI): 1.29, 1.50) and men (HR = 1.43, 95 % CI: 1.33, 1.54). A 5-unit increment in BMI was also associated with a higher risk of CHD in women (HR = 1.23, 95 % CI: 1.19, 1.27) and men (HR = 1.35, 95 % CI: 1.30, 1.40). The risk factor score was linearly related to the risk of CHD and for one risk factor increment at baseline, the hazard ratio for women was 1.28 (95 % CI: 1.22, 1.34), and for men 1.27 (95 % CI: 1.23, 1.32). Being low physically active was associated with a higher risk of CHD among men (HR = 1.08, 95 % CI: 1.03, 1.15) but statistically insignificant among women (HR = 1.05, 95 % CI: 0.97, 1.13). Table 2 shows the results of the proportional hazards model, with the hazard ratios of CHD by educational level before and after adjustments for the behavioral risk factors (potential mediators). The educational gradient in CHD was skewed, since short education was associated with a higher risk of CHD. In the baseline model among women, those with short education (HR = 1.65, 95 % CI: 1.47, 1.85) and medium education (HR = 1.28, 95 % CI: 1.14, 1.43) were at higher risk of CHD compared to women with long education. Correspondingly, among men a higher risk of CHD was observed in those with short (HR = 1.55, 95 % CI: 1.42, 1.69) and medium (HR = 1.36, 95 % CI: 1.26, 1.47) education compared to long education. Adjustment for a 5-unit increment in BMI reduced the hazard ratio of both women (HR = 1.55, 95 % CI: 1.38, 1.73) and men (HR = 1.42, 95 % CI: 1.30, 1.55) with short education accounting for 15 % and 23 % of the excess risk of CHD, while adjustment for smoking reduced it by 6 % (HRWomen = 1.62, 95 % CI: 1.44, 1.81 and HRMen = 1.50, 95 % CI: 1.38, 1.64). Low physical activity had no noticeable influence on the relative difference in CHD. Adjusting for the risk factor score reduced the relative difference between short and long education in risk of CHD by 11 % among women, and 15 % in men. The results derived from the additive hazards model estimating direct, indirect, and total effects of educational level on CHD are presented in Table 3 (women) and Table 4 (men). The direct and indirect effects were expressed as a single number since no indication of time dependent effects for either the exposure or mediators was observed. Parallel to the results in Table 2, the incidence of CHD was lowest among women and men with long education. Being in the short educational level compared with long educational level was associated with 277 (95 % CI: 219, 336) additional cases of CHD per 100,000 personyears at risk among women. Of these additional cases, 17 (95 % CI: 12, 22) cases or 6 % (95 % CI: 4, 8) could be attributed to the pathway through smoking (Table 3). Furthermore, 39 (95 % CI: 30, 49) cases or 14 % (95 % CI:
123
11, 20) could be attributed to the pathway through a 5-unit increment in BMI. Among men, being in the short educational level compared with long educational level was associated with 461 (95 % CI: 368, 555) extra cases per 100,000 person-years at risk, of which 37 (95 % CI: 28, 46) cases per 100,000 person years or 8 % (95 % CI: 6, 11) and 94 (95 % CI: 79, 110) cases or 21 % (95 % CI: 16, 27) could be attributed to the pathway through smoking and BMI, respectively (Table 4). The effect of low physical activity was negligible both among women and men. A 1-unit increment in the risk factor score seemed to be attributable to 30 (95 % CI: 23, 37) and 61 (95 % CI: 49, 74) additional cases per 100,000 person years of the social gradient in CHD among women and men, respectively.
Discussion In this paper we applied a simple and intuitively understandable measure of the mediating effects of behavioral risk factors on educational-related inequality in CHD. The additive hazards model allowed an interpretation of absolute measures indicating that 277 CHD cases per 100,000 persons years among women and 461 CHD cases per 100,000 person years among men could be prevented, assuming the associations were causal and we were able to change the educational level from short to long for the total population. A modest number of the additional cases arising from these differences in educational level could be ascribed to the pathway through smoking and BMI. For example, if an intervention could encourage women and men with a short education to improve their BMI from being overweight to normal weight 39 (women) and 94 (men) CHD cases per 100,000 person-years at risk could be eliminated, assuming that there was a causal relation between education-BMI-CHD. Qualitatively, the results of the additive hazards models (Tables 3, 4) were in accordance with the results of the Cox proportional hazards models (Table 2). Thus, our findings suggested that behavioral risk factors partially explained the relative and the absolute educational gradients in CHD, and seem not to be model dependent. The Cox proportional hazards models is often employed to assess mediation, but the traditional use of this model only gives direct and indirect effects with a causal interpretation when the outcome is rare [26]. The advantage of the additive hazards model is that it not only gives effects on the absolute scale, but can also be used with a common outcome [18]. In the present study the outcome is relatively rare (116 cases of CHD for every 1,000 people) and the two approaches give similar results. Previous studies have established results similar to our findings [6, 10, 16]. Kivima¨ki et al. [16] found that the
8,046 (11.6) 5,333 (7.7)
CHD events, no (%)
Death, no (%)
18,166 (26.1) 27,805 (40.0) 137 (0.2)
Ex-smoker
Smoker Missing
40 (0.3)
23,133 (33.3) 5,919 (8.5) 181 (0.3)
Moderate activity
High activity
98 (0.1) 25.3
Obese
Missing
Median BMI kg/m2
29,115 (41.9) 27,792 (40.0) 3,586 (5.2) 277 (0.4)
1 risk factor
2 risk factors
3 risk factors
Missing 6,717 (9.7) 3,514 (5.1) 1,311 (1.9) 1,789 (2.6) 49,477 (71.2)
CCHS (1981–1983)
MONICA I (1982–1984)
MONICA II (1986–1987)
MONICA III (1991–1992)
DCH (1993–1997)
Cohort (examination years), No (%)
8,743 (12.6)
Low risk
(20.1–33.2)
9,482 (13.6)
Overweight
(5th–95th percentile) Risk factor score, no (%)
31,815 (45.8) 27,397 (39.4)
Normal weight
721 (1.0)
Underweight
Body mass index, no (%)
Missing
793 (6.0)
28,144 (40.5)
Light activity
8,831 (33.5)
349 (38.3)
273 (41.2)
778 (44.5)
2,065 (54.7)
65 (0.5)
847 (6.4)
5,911 (44.5)
5,077 (38.2)
1,390 (10.5)
(19.7–34.7)
25.1
25 (0.2)
2,280 (17.2)
4,505 (33.9)
6,249 (47.0)
231 (1.7)
3,929 (29.6)
5,748 (43.3)
12,136 (17.5)
2,780 (20.9)
6,031 (45.4) 32 (0.2)
2,674 (20.1)
4,553 (34.3)
1,276 (9.6)
1,480 (11.1)
54 (40–64)
Short
12,389 (47.0)
466 (51.1)
290 (43.7)
766 (43.9)
1,296 (34.3)
56 (0.3)
648 (3.8)
6,533 (38.5)
7,172 (42.3)
2,549 (15.0)
(19.6–33.2)
24.4
23 (0.1)
2,032 (12.0)
5,457 (32.2)
9,173 (54.1)
273 (1.6)
33 (0.2)
1,103 (6.5)
5,116 (30.2)
7,617 (44.9)
3,089 (18.2)
5,823 (34.3) 30 (0.2)
3,681 (21.7)
7,424 (43.8)
803 (4.7)
1,276 (7.5)
53 (36–63)
Medium
Women’s educational level
Sedentary
Physical activity, no (%)
23,405 (33.7)
Never
Smoking status, no (%)
53 (40–63)
Mean age, years (5th–95th percentile)
Total
Table 1 Descriptive statistics by sex and educational level
5,156 (19.6)
97 (10.6)
100 (15.1)
203 (11.6)
413 (10.9)
14 (0.2)
192 (2.9)
2,248 (33.4)
3,077 (45.8)
1,190 (17.7)
(19.5–32.0)
23.7
13 (0.2)
585 (8.7)
1,834 (27.3)
4,180 (62.2)
109 (1.6)
2 (0.0)
487 (7.2)
2,260 (33.6)
2,902 (43.2)
1,070 (15.9)
1,830 (27.2) 5 (0.1)
1,781 (26.5)
3,105 (46.2)
184 (2.7)
376 (5.6)
53 (37–63)
Long
1,265 (16.2)
\0.01
\0.01
\0.01
\0.01
\0.01
\0.01
4,837 (20.9)
505 (28.6)
199 (30.7)
251 (28.6)
1,350 (44.3)
62 (0.8)
663 (8.5)
3,562 (45.6)
2,978 (38.1)
550 (7.0)
(21.3–34.0)
26.5
14 (0.2)
1,457 (18.6)
3,756 (48.1)
2,554 (32.7)
34 (0.4)
49 (0.6)
694 (8.9)
2,736 (35.0)
2,944 (37.7)
1,392 (17.8)
4,237 (54.2) 23 (0.3)
2,118 (27.1)
1,437 (18.4)
1,479 (18.9)
\0.01
54 (40–63)
\0.01
Short
12,170 (52.7)
1,026 (58.1)
361 (55.7)
498 (56.8)
1,260 (42.8)
63 (0.4)
991 (5.8)
6,954 (40.4)
7,282 (42.3)
1,909 (11.1)
(21.3–32.8)
26.0
18 (0.1)
2,475 (14.4)
8,397 (48.8)
6,250 (36.3)
59 (0.3)
45 (0.3)
1,902 (11.1)
6,230 (36.2)
6,409 (37.3)
2,613 (15.2)
7,465 (43.4) 33 (0.2)
5,259 (30.6)
4,442 (25.8)
1,445 (8.4)
2,604 (15.1)
53 (35–63)
Medium
Men’s educational level
\0.01
P*
6,094 (26.4)
236 (13.4)
88 (13.6)
128 (14.6)
378 (12.8)
17 (0.2)
245 (3.3)
2,584 (34.3)
3,529 (46.9)
1,155 (15.3)
(21.0–31.3)
25.3
5 (0.1)
653 (8.7)
3,448 (45.8)
3,409 (45.3)
15 (0.2)
12 (0.2)
904 (12.5)
2,862 (38.0)
2,524 (33.5)
1,192 (15.8)
2,419 (32.1) 14 (0.2)
2,653 (35.2)
2,444 (32.5)
360 (4.8)
834 (11.1)
54 (40–63)
Long
\0.01
\0.01
\0.01
\0.01
\0.01
\0.01
\0.01
\0.01
\0.01
P*
Education and risk of coronary heart disease
123
H. Nordahl et al.
123
*P value for comparison of between educational levels obtained from one-way ANOVA F test (continuous data) and Pearson v -test (categorical data)
7,530 (23.1)
Strengths and limitations
CCHS Copenhagen City Heart Study DCH Diet Cancer and Health Study
17,199 (52.8) 7,815 (24.0) 6,721 (18.2) 16,958 (45.9) 13,290 (35.9) 69,513 Total
55 (10.9)
697 (23.3) 1,516 (50.7)
235 (46.6) 214 (42.5)
Inter99 (1999–2001)
780 (26.1)
939 (1.4)
5,766 (8.3)
1936-Cohort (1993–1997)
2
56 (12.9)
550 (19.8)
258 (59.3)
1,626 (58.6)
121 (27.8)
597 (21.5)
Medium Short Medium Short
Table 1 continued
Total
Women’s educational level
Long
P*
Men’s educational level
Long
P*
effect of adjusting for smoking, heavy alcohol consumption, physical inactivity and obesity on the relative educational gradient of CHD produced a modest (13–29 %) reduction. Lynch et al. [10] found, that adjustment for smoking, high cholesterol, hypertension, and diabetes explained a modest proportion of the relative educationalrelated inequality in CHD. Further, in their attempt to explain absolute rather than relative gradients Kivima¨ki et al. found that a marked educational gradient in absolute risk remained in the subgroup that was free of the measured behavioral risk factors. However, Lynch and colleagues found that the risk factors explain the majority of absolute inequality in CHD, because these risk factors explained the vast majority of CHD cases. The absolute measures applied in both Kivima¨ki et al. and Lynch et al. was based on simple risk differences and did not take confounders or follow-up time into account. The absolute risk approach used in the present study provided a useful flexible method to analyze direct and indirect effects for time-to-event data and confounding adjustments.
To our knowledge the present observational study is the first to comprehensively address the educational-related gradients in CHD and the mediating effects of behavioral risk factors in absolute rather than relative measures, which are of great public health relevance, as they indicate the absolute numbers of cases prevented when modifying specific mediators. Further, the results for the ratio of indirect to total effects are given with confidence intervals. Additional strengths of the present study include the prospective design, the linkage to high quality data from central and administrative registries that gives access to complete follow-up, as well as individual information on highest attained educational level, and the establishment of a large cohort study by synthesizing information already collected by existing studies. The analyses were based on pooled data from seven different cohort studies, which provided enough statistical power to study the aetiology and the mechanisms behind the gradients in CHD. However, study-related factors such as the use of different self-administered questionnaires, where phrasing of questions and time periods of responses varied between the different studies, might have influenced the potential for synthesis and full characterization of all multiple versions of the behavioral risk factors, which might have resulted in imprecise measures. One particular problem with non-differential misclassification in this study is related to our dichotomization of smoking status and level of physical activity, which was done partly as a consequence of harmonization procedures and partly to fit the model. Since there are multiple levels of e.g. physical
Education and risk of coronary heart disease Table 2 Hazard ratios (HRs) and 95 % confidence intervals for CHD using cox proportional hazards model, by educational level Women’s educational level Short
Medium
Long
% Excess riska
Men’s educational level Short
Medium
Long
% Excess riska
HR
95 %CI
HR
95 %CI
(REF)
HR
95 %CI
HR
95 %CI
(REF)
1.65
1.47, 1.85
1.28
1.14, 1.43
1.00
1.55
1.42, 1.69
1.36
1.26, 1.47
1.00
Tobacco: Smokers and ex-smokers
1.62
1.44, 1.81
1.27
1.13, 1.42
1.00
-6
1.50
1.38, 1.64
1.34
1.24, 1.45
1.00
-6
Physical activity:
1.65
1.47, 1.85
1.27
1.13, 1.43
1.00
0
1.54
1.42, 1.68
1.36
1.26, 1.47
1.00
-2
1.55
1.38, 1.73
1.23
1.10, 1.39
1.00
-15
1.42
1.30, 1.55
1.29
1.20, 1.40
1.00
-23
1.58
1.41, 1.77
1.25
1.11, 1.40
1.00
-11
1.47
1.35, 1.60
1.33
1.23, 1.44
1.00
-15
Baseline model Plus adjustments for
Sedentary-light activity BMI: (5-unit increase) Risk factor score: (1-unit increase) All estimates are adjusted for cohort a
HR comparing short versus long educational level
Table 3 Direct effects (DE), indirect effects (IE), and total effects (TE) of educational level on CHD for each risk factor derived from linear regression parameter estimates and the additive hazards model (WOMEN) Educational level
DE 9 10-5
95 % CI
IE 9 10-5
Tobacco:
Long ? Short
261
204, 319
17
12, 22
6
4, 8
Smokers and ex-smokers
Long ? Medium
107
54, 158
3
1, 6
3
1, 7
Physical activity:
Long ? Short
276
218, 333
1
-1, 3
0
0, 1
Sedentary and light activity
Long ? Medium
108
55, 160
1
-1, 2
1
-1, 3
BMI:
Long ? Short
233
175, 292
39
30, 49
14
11, 20
(5-unit increase)
Long ? Medium
89
37, 141
19
14, 24
18
11, 35
Risk factor score:
Long ? Short
247
188, 305
30
23, 37
11
8, 15
(1-unit increase)
Long ? Medium
97
45, 150
11
8, 15
11
6, 21
Potential mediators
95 % CI
IE/TE %
95 % CI
All estimates are adjusted for cohort Total effect (95 % CI) 9 10-5: Long ? Short: 277 (95 % CI: 219, 336) and Long ? Medium: 109 (95 % CI: 56, 161)
Table 4 Direct effects (DE), indirect effects (IE), and total effects (TE) of educational level on CHD for each risk factor derived from linear regression parameter estimates and the additive hazards model (MEN) Educational level
DE 9 10-5
95 % CI
IE 9 10-5
Tobacco:
Long ? Short
427
334, 521
37
28, 46
8
6, 11
Smokers and ex-smokers
Long ? Medium
286
213, 360
20
14, 26
6
4, 9 0, 1
Potential mediators
95 % CI
IE/TE %
95 % CI
Physical activity:
Long ? Short
460
367, 553
2
0, 5
1
Sedentary and light activity
Long ? Medium
304
231, 378
1
-1, 2
0
0, 1
BMI:
Long ? Short
361
267, 455
94
79, 110
21
16, 27
(5-unit increase)
Long ? Medium
247
172, 320
60
49, 71
19
15, 27
Risk factors score:
Long ? Short
401
307, 494
61
49, 74
13
10, 18
(1-unit increase)
Long ? Medium
277
202, 350
31
24, 39
10
7, 14
All estimates are adjusted for cohort Total effect (95 % CI) 9 10-5: Long ? Short: 461 (95 % CI: 368, 555) and Long ? Medium: 306 (95 % CI: 232, 380)
123
H. Nordahl et al.
activity which corresponds to the category ‘‘medium/high level of physical activity’’ our chosen cut point separating low from medium/high level of physical activity could have induced measurement errors and ambiguity about the actual level of physical activity. In turn this could lead to downwards bias of the indirect effect and upwards bias of the direct effect [30]. However, our performed sensitivity analyses applying different cut-off points for the physical activity as well as modeling current tobacco smoking (cigarette, pipe, and cigar) in gram per day continuously, did not have any important impact on the results. In our preparation of this study we used prior knowledge and the methods of Directed Acyclic Graphs to identify potential confounders. One of the assumptions required for mediation analysis is that there is no effect of the exposure that confounds the mediator–outcome relation (no exposure-dependent confounders). It seems that this assumption likely does not hold with BMI. For example physical activity probably confounds the BMI–CHD relationship and physical activity is itself affected by the exposure (education). This situation cannot be neglected. As suggested in the descriptive analyses (Table 1), the physical activity level in the three educational groups does vary. The highest prevalence of sedentary participants was found in the short educated group. However, these variations were relatively small from 16 to 21 % sedentary participants in the short versus long educated group. We performed an auxiliary analysis of CHD conditioning on education, BMI, and in addition physical activity. The results (not shown) indicated no association between CHD and physical activity. Thus, while physical activity is a plausible exposure-dependent confounder it cannot possibly have severely affected our results. The no-unmeasured confounding assumptions required in mediation analyses are quite strong (see VanderWeele [27] for a discussion and techniques to assess potential impact of unmeasured confounders in a non-survival setting). While studying educational-related inequality in risk of CHD and the mediating effects of behavioral risk factors in general one should consider plausible unmeasured confounders like ‘country of birth’, which could bias the estimates of for example the education–BMI relationship and the BMI–CHD relationship. However in our set up using the SIC Cohort Study more than 98 % of the participants are native Danes, and for a considerable number of non-native Danes, particularly women, the registrybased educational data do not contain information on achieved educational level in country of birth. Thus, in a population comparable to the SIC Cohort Study unmeasured confounding due to non-native Danes is at most 2.9 9 10-5 (computed from crude mathematical derivations of Corollary 2 of [27]; a CVD rate difference between non-native vs. native Danes of 0.00588 [31] and a maximal
123
difference in non-native Danes share between education categories of 0.005 [32]). Qualitative evaluation of such bias in estimates (2.9 9 10-5) seems relatively small and is mainly necessary to consider for the indirect effects through low physical activity which is no more than 2 additional cases per 100,000 person years (as suggested by Tables 3 and 4). The predictive value for CHD of the behavioral risk factor measured at one point in time is likely to be influenced by duration of follow-up time [33]. In the present study the follow-up time varied from 6 of 27 years and to accommodate for potential different hazard levels due to difference in follow-up time, we fitted separate models for each cohort under the restriction that the coefficients are equal but the baseline hazard functions are not equal [28]. Further, fitting models where the follow-up time was stopped after 6 years did not affect the results.
Conclusion The results from the additive hazards model estimating the mediating effects of behavioral risk factors on educational differences in CHD were in accordance with the results of the Cox proportional hazards model: Indicating that smoking and BMI partially explained the educational gradient in CHD. It is of great public health relevance to estimate the absolute numbers of CHD cases prevented when modifying specific mediators such as smoking and BMI. However, if this quantitative assessment of mediation should be used for prioritizing between modifiable risk factors to tackle social inequalities in CHD, further research on pathways through other potential mediators such as psychosocial and biological risk factors is needed, and the impact of trajectories of these modifiable risk factors must be taken into consideration. Acknowledgments We thank the collaborators behind the SIC Cohort Study: the Copenhagen City Heart Study, the Diet, Cancer and Health Study, and the cohorts at the Research Center for Prevention and Health. This work was supported by the Commission of Social Inequality in Cancer [Grant Number SU08004]. Conflict of interest of interest.
The authors declare that they have no conflict
References 1. Rose G, Marmot MG. Social class and coronary heart disease. Br Heart J. 1981;45(1):13–9. 2. Harper S, Lynch J, Smith GD. Social determinants and the decline of cardiovascular diseases: understanding the links. Annu Rev Public Health. 2011;32:39–69.
Education and risk of coronary heart disease 3. Kaplan GA, Keil JE. Socioeconomic factors and cardiovascular disease: a review of the literature. Circulation. 1993;88(4 Pt 1): 1973–98. 4. Lynch JW, Kaplan GA, Cohen RD, Tuomilehto J, Salonen JT. Do cardiovascular risk factors explain the relation between socioeconomic status, risk of all-cause mortality, cardiovascular mortality, and acute myocardial infarction? Am J Epidemiol. 1996;144(10):934–42. 5. Emberson JR, Whincup PH, Morris RW, Walker M. Social class differences in coronary heart disease in middle-aged British men: implications for prevention. Int J Epidemiol. 2004;33(2):289–96. 6. Albert MA, Glynn RJ, Buring J, Ridker PM. Impact of traditional and novel risk factors on the relationship between socioeconomic status and incident cardiovascular events. Circulation. 2006; 114(24):2619–26. 7. Beauchamp A, Peeters A, Wolfe R, Turrell G, Harriss LR, Giles GG, et al. Inequalities in cardiovascular disease mortality: the role of behavioural, physiological and social risk factors. J Epidemiol Community Health. 2010;64(6):542–8. 8. Ernstsen L, Bjerkeset O, Krokstad S. Educational inequalities in ischaemic heart disease mortality in 44,000 Norwegian women and men: the influence of psychosocial and behavioural factors. The HUNT Study. Scand J Public Health. 2010;38(7):678–85. 9. Laaksonen M, Talala K, Martelin T, Rahkonen O, Roos E, Helakorpi S, et al. Health behaviours as explanations for educational level differences in cardiovascular and all-cause mortality: a follow-up of 60 000 men and women over 23 years. Eur J Public Health. 2008;18(1):38–43. 10. Lynch J, Davey SG, Harper S, Bainbridge K. Explaining the social gradient in coronary heart disease: comparing relative and absolute risk approaches. J Epidemiol Community Health. 2006; 60(5):436–41. 11. Strand BH, Tverdal A. Can cardiovascular risk factors and lifestyle explain the educational inequalities in mortality from ischaemic heart disease and from other heart diseases? 26 year follow up of 50,000 Norwegian men and women. J Epidemiol Community Health. 2004;58(8):705–9. 12. Hafeman DM. A sufficient cause based approach to the assessment of mediation. Eur J Epidemiol. 2008;23(11):711–21. 13. Hafeman DM, Schwartz S. Opening the black box: a motivation for the assessment of mediation. Int J Epidemiol. 2009;38(3): 838–45. 14. van Lenthe FJ, Gevers E, Joung IM, Bosma H, Mackenbach JP. Material and behavioral factors in the explanation of educational differences in incidence of acute myocardial infarction: the Globe study. Ann Epidemiol. 2002;12(8):535–42. 15. Kilander L, Berglund L, Boberg M, Vessby B, Lithell H. Education, lifestyle factors and mortality from cardiovascular disease and cancer. A 25-year follow-up of Swedish 50-year-old men. Int J Epidemiol. 2001;30(5):1119–26. 16. Kivimaki M, Lawlor DA, Davey SG, Kouvonen A, Virtanen M, Elovainio M, et al. Socioeconomic position, co-occurrence of behavior-related risk factors, and coronary heart disease: the Finnish Public Sector study. Am J Public Health. 2007;97(5): 874–9. 17. Aalen O. A model for non-parametric regression analysis of counting processes. In: Klonecki W, Kozek A, Rosinski J, editors.
18. 19.
20.
21.
22.
23.
24.
25.
26. 27. 28. 29. 30. 31.
32.
33.
Lecture notes in statistics, mathematical statistics, and probability theory. New York: Springer; 1980. p. 1–25. Lange T, Hansen JV. Direct and indirect effects in a survival context. Epidemiology. 2011;22(4):575–81. Appleyard M, Hansen AT, Schnohr P, Jensen G, Nyboe J. The Copenhagen city heart study. A book of tables with data from the first examination (1976–1978) and a 5-year follow-up (1981–83). Scand J Soc Med Suppl. 1989;170(41):1–160. Tjonneland A, Olsen A, Boll K, Stripp C, Christensen J, Engholm G, et al. Study design, exposure variables, and socioeconomic determinants of participation in diet, cancer and health: a population-based prospective cohort study of 57,053 men and women in Denmark. Scand J Public Health. 2007;35(4):432–41. Osler M, Linneberg A, Glu¨mer C, Jørgensen T. The cohorts at the Research Centre for Prevention and Health, formerly The Glostrup Population Studies. Int J Epidemiol. 2011;40(3):602–10. Jorgensen T, Borch-Johnsen K, Thomsen TF, Ibsen H, Glumer C, Pisinger C. A randomized non-pharmacological intervention study for prevention of ischaemic heart disease: baseline results Inter99. Eur J Cardiovasc Prev Rehabil. 2003;10(5):377–86. Gerdes LU, Bronnum-Hansen H, Madsen M, Borch-Johnsen K, Jorgensen T, Sjol A, et al. Trends in selected biological risk factors for cardiovascular diseases in the Danish MONICA population, 1982–1992. J Clin Epidemiol. 2000;53(4):427–34. Fortier I, Doiron D, Burton P, Raina P. Invited commentary: consolidating data harmonization—how to obtain quality and applicability? Am J Epidemiol. 2011;174(3):261–4. Grambsch PM, Therneau TM. Proportional hazards tests and diagnostics based on weighted residuals. Biometrika. 1994;81(3): 515–26. VanderWeele TJ. Causal mediation analysis with survival data. Epidemiology. 2011;22(4):582–5. VanderWeele TJ. Bias formulas for sensitivity analysis for direct and indirect effects. Epidemiology. 2010;21(4):540–51. Martinussen T, Scheike TH. Dynamic regression models for survival data. New York: Springer; 2006. Greenland S, Pearl J, Robins JM. Causal diagrams for epidemiologic research. Epidemiology. 1999;10(1):37–48. VanderWeele TJ. Mediation analysis with multiple versions of the mediator. Epidemiology. 2012;23(3):454–63. Hempler NF, Larsen FB, Nielsen SS, Diderichsen F, Andreasen AH, Jorgensen T. A registry-based follow-up study, comparing the incidence of cardiovascular disease in native Danes and immigrants born in Turkey, Pakistan and the former Yugoslavia: do social inequalities play a role? BMC Public Health. 2011; 11:662. Statistical overview of integration: population, education, and employment. Copenhagen: The Ministry of Refugee Immigration and Integration Affairs; 2010. http://www.nyidanmark.dk/NR/ rdonlyres/373D7D82-3F91-4873-A3F0-1731A37DD18D/0/statis tical_overview_of_integration_2010.pdf. Accessed 24 Aug 2012. Suadicani P, Hein HO, Gyntelberg F. Socioeconomic status and ischaemic heart disease mortality in middle-aged men: importance of the duration of follow-up. The Copenhagen Male Study. Int J Epidemiol. 2001;30(2):248–55.
123
