USE-IMT is an ongoing individual participant data meta-analysis of which the methods have been described in detail elsewhere.[11] In short, general population cohorts were identified using literature search and expert suggestions. For inclusion in USE-IMT, cohorts were required to have available baseline data on age, sex, blood pressure, cholesterol fractions, smoking status, use of antihypertensive medication, diabetes mellitus and CIMT-measurements and follow-up information on occurrence of cardiovascular events. For the current analysis, we included the participating cohorts of which data on ethnicity on an individual level were available (n = 9). Cohorts that did not have information on ethnicity were included if it was reasonable to assume that >95% of the participants belonged to one race/ethnic group due to either selection of participants or race/ethnic homogeneity of the source population (n = 6). Ethnicity was recoded when applicable, to create uniform race/ethnic groups for analysis. Details concerning this recoding process and the original classification can be found in S1 Table. As our research is focused on the first-time events in asymptomatics, only individuals to whom the Framingham criteria [7] are applicable were included in the current analyses.

This study was approved by the institutional review committee of the University Medical Centre Utrecht. Each individual cohort obtained approval from a local Ethical Review Board and written informed consent from all participants. All authors exchanged a material transfer agreement. This study conforms to the declaration of Helsinki.

Out of the 17 cohorts participating in USE-IMT, we included 15 cohorts[12–26], consisting of 66,213 individuals. After excluding individuals of whom individual ethnicity was not known (n = 285) or who had already experienced a cardiovascular event (n = 5,717) 60,211 individuals were included for analyses. This group consisted of 46,788 Whites, 7,200 Blacks, 3,816 Asians and 2,407 Hispanics.

Incomplete data on mean common CIMT, cardiovascular risk factors, and (time to) CV events, approximately 12% of total values, were imputed, as described previously.[11] This imputation was done to increase power and to reduce omitted variable bias.[27]

Average mean common CIMT was calculated for each individual using the maximum set of carotid angles, near and/or far wall measurements, and left and/or right side measurements that were assessed within each cohort. Time to first fatal or non-fatal myocardial infarction or stroke (hemorrhagic or ischemic) was used as a primary endpoint in this analysis.

Baseline data are represented as means with standard deviations (SD) for continuous variables and as percentages for categorical variables. We describe our population both by their original cohort (Table 1) and by race/ethnic group (Table 2). All statistical analyses were performed in R (version 2.15.1).

To examine the influence of ethnicity on associations of risk factors with mean common CIMT and CV events, we used linear regression models and Cox regression models, respectively. For both questions, we first fitted a model containing the Framingham risk factors (age, sex, systolic blood pressure (SBP), HDL-cholesterol, total cholesterol, current smoking, presence of diabetes and antihypertensive drug use)[7] and ethnicity as independent variables. Mean common CIMT (linear regression-log transformed) and a combined endpoint of first-time stroke or myocardial infarction (Cox regression) were used as an outcome, respectively. To take between-study-variation into account, we took a random effects approach for each original cohort. We then extended this model by adding the interaction terms for ethnicity with each risk factor. Whites were the reference group in all analyses.

We first compared the interaction model to the first model using a likelihood ratio test. Subsequently, we tested the significance of the interaction terms. Finally, we calculated hazard ratios and betas (with 95% confidence intervals) for each risk factor in each race/ethnic group, and tested for significance of individual interaction terms using a Wald test. Due to the large number of tests performed, we took a conservative threshold for interaction at p_interaction = 0.05 (commonly accepted p-values for interactions are 0.1 or 0.2). In these analyses, Whites were the reference group. Interaction terms were only considered significant when significantly adding to the model without interaction terms and when significantly different from Whites.

No comparisons between ethnicities were performed, except for the comparisons to Whites. A two-sided significance level of 0.05 was used.

Mean common CIMT was 0.74 mm in Blacks, Whites and Hispanics and 0.78 mm in Asians.

The model with risk factors and interaction terms with ethnicity fitted the data better than the model with risk factors alone (likelihood ratio test p<0.001). The interaction terms of ethnicity with age, HDL-cholesterol, smoking and systolic blood pressure were statistically significant (p <0.05).

We thus found that for specific ethnicities risk factor associations were significantly different from Whites.

The magnitude of the association (i.e. betas of the linear regression model) of age with logCIMT was significantly smaller in Blacks and Hispanics as compared to Whites: 0.08/10 years (91% of the White beta) and 0.09/10 years (97% of the White beta), respectively. This means that every 10-year increase in age gives less increase in mean common CIMT in Blacks and Hispanics than it does in Whites. The beta for HDL-cholesterol per 1 mmol/L increase was -0.05 (144%) in Blacks (a stronger inverse association) as compared to the beta in Whites. The beta for smoking (yes vs. no) was 0.01 (30% of the beta in Whites). In Asians the beta for SBP was 0.02 (131% of the White beta). All betas are plotted in Fig 1 and are presented in Table 3.

The 10-year event rate was 6.7% in Asians, 7.8% in Hispanics, 8.1% in Whites and 9.2% in Blacks.

The model including the risk factors and the interaction terms with ethnicity fitted the data better than the model without interaction terms (likelihood ratio test p <0.001). Both age and total cholesterol had a significant interaction with ethnicity. The HR for CV events (first-time stroke or MI) per 10-year increase in age was lower in Blacks than in Whites (1.52 (1.44–1.60), 80% of the HR in Whites). The HR for a 1 mmol/L increase of total cholesterol was higher in Blacks (1.20 (1.13–1.26), 140%) compared to Whites. The hazard ratios of the Framingham risk factors per ethnic group are plotted in Fig 2.

The prevalence of risk factors in our cohort was unevenly distributed among the race/ethnic groups (Table 2). Race/ethnic differences in the prevalence of CVD risk factors have been previously described. Similar to our study, smoking was least prevalent in Asians.[28] Also diabetes was more prevalent in Blacks and Hispanics than in Asians and Whites[9], and the higher prevalence of diabetes was accompanied with higher BMI in these groups.[29] Also in accordance with literature, TC and LDL-cholesterol levels were highest in Whites, HDL-cholesterol levels were lowest in Hispanics and triglyceride levels were lowest in Blacks.[30]

However, in contrast with these similarities between our data and published literature, blood pressure was highest in Asians, which is opposite to the findings in the Asia Pacific cohort.[31] Possibly, this incongruence between our study and Asia Pacific can be explained by differences in age, as the Asians in our cohort are older than the Whites (in the Asia Pacific cohort the group from Australia and New Zealand is older than the Asians).

The observations made above indicate that our cohort is fairly similar to other cohorts described in literature, thereby suggesting that our results might be generalizable.

For Blacks, an increase in age was related to both an increase in mean common CIMT and a higher risk of CV events, in our analyses. The magnitude of these associations, however, was significantly smaller than in Whites. Although the direction of the association of age is the same in Blacks as in Whites, an increasing age has less effect on disease in Blacks than in Whites. This indicates that age is one of the factors that should be should weighted when developing a prediction model specifically for Blacks.

HDL-cholesterol and smoking in Blacks, SBP in Asians and age in Hispanics showed a different magnitude of the association with mean common CIMT as compared to Whites. However, these differences in the magnitude with mean common CIMT were not detected in the analysis of CV events. This might be due to the difference in power between linear regression and Cox proportional hazards analysis (which is driven by the number of events), or a difference in biology of CIMT and the actual occurrence of CV events.

The association between TC and CV event was significantly higher for Blacks as compared to Whites. The coefficient for TC on mean common CIMT was also higher than the coefficient in Whites, although this difference did not reach statistical significance.

When it comes to geographical differences, some of these differences have been reported before. The Asia Pacific study reported lower HRs for triglycerides and SBP for respectively coronary heart disease and hemorrhagic stroke, when comparing Asian countries with Australia and New-Zealand.[32] The INTERHEART study[33] showed in an analysis of 52 countries across the globe that odds ratios for myocardial infarction were comparable, although (small) differences in effect sizes exist. These geographical differences are in agreement with our race/ethnicity-specific findings.

Race/ethnicity-specific risk prediction should be exercised with discretion. Unwanted medical “discrimination” that could be perceived by patients should be avoided at all costs. On the other hand optimal risk estimation is of paramount importance for medical care. Adding race/ethnicity to risk prediction equations thus should be implemented with care and reasons to do so should be clarified to all parties, as to avoid undesirable situations in health care and society.

Due to differences in ascertaining and recording ethnicity per cohort, misclassification might have occurred and may have diluted our results to some extent. While we are aware of the growing number of people from mixed backgrounds, people of mixed race/ethnicity were excluded from the current analyses. Also, our Asian and Hispanic race/ethnic groups were relatively small, which may have led to a lack of power. Therefore, our conclusions on these groups should be interpreted with caution.

We were unable to take immigration status or acculturation into account, although time since immigration has been shown to influence mean common CIMT, risk factors levels and risk of cardiovascular events.[34,35] Therefore we cannot determine in which way immigration status might influence our results, or whether the results are influenced at all.

In our CV events analysis we were unable to take differences in prevention and treatment strategies into account, these might differ among the ethnic groups and influence the occurrence of CV events.