The Seoul Metabolic Syndrome cohort study led by Seoul Metropolitan Government has collected data of citizens suspected of metabolic syndrome from 2014 (Metabolic Syndrome Cohort in Korea, NCT02077530). A total of 1,370 participants were recruited from 25 public healthcare centres in Seoul between 2014 to 2015. The initial check-up and follow-up visit were performed in Korea University Anam Hospital. In the cohort, 870 participants took the annual follow-up visit and underwent coronary CT in 2016. The dataset includes self-administered questions about lifestyle, disease, and medication history as well as medical records including physical examination, laboratory measurement, and coronary CT. Finally, data of 847 participants (425 men and 422 women) after removing the subjects with missing values were analysed. The present study was approved by the institutional review board of Korea University Anam Hospital (IRB NO. ED13087) and performed in accordance with the principles of the Declaration of Helsinki. Written informed consent was also obtained.

Medication history including statins was confirmed by checking the prescriptions of the participants. Statin intensity was categorized in accordance with the 2013 ACC/AHA guideline [1]. Blood sample was obtained under at least an 8-hour fasting status: Serum glucose level was determined using a UV assay; Serum levels of total cholesterol, LDL cholesterol, HDL cholesterol, and triglycerides were measured using homogeneous enzymatic colorimetric assay; The hsCRP was gauged using immunoturbidimetry assay. Covariates (pre-treatment confounders) such as age, sex, SBP (systolic blood pressure), and smoking were controlled in the mediation analysis.

Coronary CT was performed using a second-generation dual-source CT scanner (Somatom Definition Flash; Siemens Healthcare, Forchheim, Germany) with a 2 × 128 × 0.6-mm³ section collimation and a 280 ms ration time. Initially, CT scanning was performed to evaluate the CAC score with 120-kV tube voltage, 80-mAs effective tube current, and 3-mm section thickness. The total estimated radiation dose for cardiac CT examinations ranged from 3 to 15 mSv. The CAC score was quantified on the basis of the Agatston scoring methods [21]. CAC as outcome variable was coded into binary 0 for zero CAC scores and 1 for positive scores.

Characteristics of the study population were described as mean and count for continuous and categorical variables, respectively. Associations (represented by a, b and c in Fig 1) between statin use and CAC with each of the potential mediators were examined using regressions with bootstrapping. The confidence intervals (CIs) were estimated via a bias-corrected bootstrapping with 1,000 replications, considering the skewed nature of the measurement such as CAC and hsCRP.

R Mediation package 4.4.6 [22] was adopted to perform the causal mediation analysis. The causal mechanism describes that treatment (statin) affects outcome (CAC) through an intermediate variable (e.g., HDL-cholesterol) decomposing total effect into direct effect (Statin → CAC) and indirect effect (Statin → Mediator → CAC) (Fig 1).

Causal mediation mechanism can be identified in the satisfaction of assumptions called “Sequential Ignorability” (SI) [23]:

This study followed the model-based causal mediation analysis assuming the SI assumptions [22, 23]; the patients who have the same values of the pre-treatment covariates are regarded as if randomized, since the covariates were adjusted in analysis [24]. The mediator was modeled with a linear regression including treatment and pre-treatment covariates (i.e., age, gender, systolic blood pressure (SBP), and smoking). The outcome model was a logistic regression including the mediator, the treatment, and the pre-treatment covariates. These models were adjusted with the pre-treatment covariates.

Sensitivity analysis examines how the study results are robust to the violation of the last assumption above. Sensitivity parameter measures the correlation ρ between the two error terms ϵ₂ and ϵ₃ in the mediator (M) and outcome (Y) regression models. If the assumption holds, then the correlation has a value of 0. Data were analysed using R 3.4.0 software.

Table 1 shows the demographic, laboratory, CAC, and statin treatment characteristics of the study sample (n = 847) comprising 50.2% men. Mean age was 56.69 (SD = 6.57) years old. A total of 202 subjects took statin: 86 men (23.8%) and 116 women (76.2%). Majority of them (89.1%) received moderate intensity statin treatment. Atorvastatin and rosuvastatin accounted for over 80% of the statins prescribed. The group taking statin (n = 202) had lower LDL cholesterol (p < .001), triglyceride (p = 0.006) levels and higher HDL-cholesterol (p = 0.001) and glucose (p < .001) levels than the group not taking statin (n = 645). Smoking, waist circumference, and BMI were similar in both groups. In addition, there was a high prevalence (37.6%; n = 76; p = 0.002) of CAC along with a high CAC score (80.27; SD = 237.91; p = 0.005) in the group who received statin therapy.

After adjustment of the covariates, regression analysis showed the association between statin, potential mediators, and CAC (Table 2).

Path c: Statin had a positive association with the CAC (odds ratio 2.006; p < .001; 95% CI: 1.383 to 2.951), suggesting that the subjects with statin treatment had higher odds of positive CAC scores than those without statin treatment.

Path a: Statin was associated with the lower LDL-cholesterol (-37, 906; p < .001; 95% CI: -43.072 to -32.734) and triglyceride (-21.392; p < .030; 95% CI: -35.796 to -6.302) levels, and was associated with the higher HDL-cholesterol (2.993; p = 0.002; 95% CI: 1.168 to 4.993) and glucose (7.399; p < .001; 95% CI: 3.194 to 12.447) levels. On the other hand, no significant association was found between statin use and the hsCRP level.

Path b: HDL-cholesterol level was significantly associated with a lower prevalence of CAC (odd ratio 0.981; p = 0.008; 95% CI: 0.968 to 0.995). In the other hand, no significant relationships were observed between the remaining mediators and CAC.

Causal mediation analysis with each of the five biochemical factors as a mediator (Fig 1) was performed with the options of a nonparametric 1,000 bootstrapping replicates and 95% CI, while adjusting for the covariates. Table 3 shows the point estimates and 95% CIs of total, direct, and mediation effects from each mediation model. Direct and mediation effects are called average direct effect (ADE) and average causal mediation effect (ACME, also refer to as indirect effect). Statistically, total effect is the sum of ACME and ADE. The ACME states whether statin use (treatment) positively (the probability of increasing CAC) or negatively (the one of decreasing CAC) affects outcomes (CAC) through a mediator (e.g., HDL-cholesterol). Thus, if the upper limit of bootstrap CI is less than zero, then our hypothesis of negative causal mediation effect (i.e., H₀ = 0, H₁: ACME < 0) is supported at the one-tailed level of α = 0.025. On the other hand, if the lower limit of bootstrap CI is greater than zero, then our hypothesis of positive causal mediation effect, H₁: ACME > 0, cannot be rejected at the one-tailed level of α = 0.025.

Total effect and ADE (Table 3) show that statin is likely to increase the probability of CAC > 0 in all five mediation models. This result is consistent with previous studies [14, 25]. Importantly, ACME proposed only HDL-cholesterol as a significant mediator mitigating statin-induced increased probability of CAC > 0. It implies that statin increases HDL-cholesterol level, which in turn makes the subject less likely to have CAC > 0 (note that CAC was analysed as a binary variable of CAC score = 0 and CAC score > 0). Any other factors failed to show mediation effect between statin and CAC.

One measure of interest is the proportion of mediation defined as the proportion of total effect explained by the mediator. Average proportion of the HDL-cholesterol driven mediation effect becomes 8.3% (calculated as the percentage of ACME/total effect [0.011/0.132 X 100]). The total effect implies that statin increased the percentage chance of CAC > 0 by 13.2% in HDL-cholesterol causal mediation model. Other figures can be interpreted in the same context.

The robustness of the causal mediation effect of HDL-cholesterol was explored by (1) the additional analyses of three different datasets according to statin intensity and (2) sensitivity analysis.

The study population had 202 subjects treated with statin, and they can be divided into the 3 different subgroups according to statin intensity: low-intensity statin treatment group (10 patients), moderate-intensity statin treatment group (180 patients), and high-intensity statin treatment group (12 patients). Additional causal mediation analyses were performed with the 3 different datasets formed as follows:

The ACME with HDL-cholesterol mediator was estimated by applying the nonparametric bootstrapping method: ACME = -0.011 with 95% CI [-0.025, -0.003] for dataset 1; ACME = -0.011 with 95% CI [-0.027, -0.003] for dataset 2; and ACME = -0.010 with 95% CI [-0.026, -0.002] for dataset 3. These results were consistent with that of total population, and suggested that the mediation effect of HDL-cholesterol between statin and CAC might be independent to statin intensity.

This study performed sensitivity analysis to investigate the robustness of HDL-cholesterol mediation model when an unobserved confounder between HDL-cholesterol and CAC was considered. Fig 2 shows the ACMEs of HDL-cholesterol against varying sensitivity parameter ρ (the correlation between two error terms in ϵ2 and ϵ3) and 95% CIs represented by the shaded region. When the sensitivity parameter has a value in −0.1 ≤ ρ ≤ 1, the ACME of HDL-cholesterol could be identified. On the other hand, the sensitivity parameter ρ is less than -0.1, the ACME of HDL-cholesterol would have a positive value contrary to our study results (negative point estimate of ACME, -0.011 in Table 3). The vertical dashed line represent the point of ρ when the ACME becomes zero.