All individuals in South Korea are covered by the single-payer Korea National Health Insurance Service (NHIS). The Korea National Health Insurance has been operated based on a fee-for-service payment system for both inpatient and outpatient services and thus all of health care utilization data including medication history are available. The Korea NHIS built up the national representative cohort database, which was comprised of 1,025,340 nationally representative random subjects in 2002, amounting to 2% of the entire population [10]. Proportionate stratified random sampling was used based on a total of 1,476 strata (two categories for sex, 18 categories for age group, and 41 categories for income). This cohort was followed-up over 12 years until 2013, and random samples of new infants were added to compensate for the loss of individuals due to death or emigration for each year between 2003 and 2013. The data consist of complete medical records, including socio-demographic variables, diagnostic codes, deaths, medical service use, and health examinations. Recently, the NHIS publicly released this National Health Insurance Service-National Sample Cohort 2002–2013 (NHIS-NSC 2002–2013). Detailed information on the NHIS-NSC profile was described in a previous study [10]. We used a subset database of the NHIS-NSC, which consisted of approximately 570,000 people who received a medical check-up without charge that provided comprehensive social and medical information about the population. The protocol for this study was approved by the Institutional Review Board of Asan Medical Center (S2015-0982) and informed consent was waived by the board.

We identified patients who were diagnosed with HZ, stroke, and MI by use of the relevant diagnostic codes of the International Classification of Diseases, Tenth Revision (ICD-10). The cases of HZ were comprised of VZV meningitis (B020), VZV encephalitis (B021), zoster in other divisions of the trigeminal nerve (B022), zoster ophthalmicus (B023), and other types of zoster (B027, B028, or B029). The cardiovascular disease (CVD) referred to all episodes of stroke and MI. Stroke consisted of ischemic brain infarction (I63), hemorrhage stroke (I60, I61, or I62), and unspecified stroke (I64). MI included myocardial infarction (I21, I22) and complications following myocardial infarction (I23).

In addition, clinical characteristics that could potentially become confounding factors for HZ, stroke, and MI were searched using ICD-10 code or medical records; these included age, gender, body mass index, overweight, cigarette smoking, alcohol consumption, physical activity, income, hypertension, diabetes (fasting plasma glucose), dyslipidemia (total cholesterol), angina pectoris, transient ischemic attack, heart failure, atrial fibrillation/flutter, valvular heart disease, chronic renal disease, peripheral vascular disease, chronic liver disease, rheumatoid disease, inflammatory bowel disease, malignancy, transplantation recipients of solid organs or hematopoietic stem cell, and human immunodeficiency virus infection. We considered the variables listed in Table 1 as potential confounders between statin use and CVD, including stroke and MI, between statin use and HZ, and between HZ and CVD, and we adjusted for these variables in the main analysis. These covariates were selected a priori as common risk factors of HZ and CVD (Supplementary Fig. 1). Chronic pulmonary disease is a major risk factor for HZ and CVD [11]. As such, we did not include chronic pulmonary disease as a variable in the main analysis to quantify the impact of unmeasured mediator-outcome confounding on mediation effect.

Initially, 569,999 individuals were enrolled in this cohort (Fig. 1). Of this population, 148 persons under 18 years old were excluded, and a total 569,851 persons ≥ 18 years old were followed up for 11 years from 2002 through 2013. To identify the first-ever episodes of HZ, stroke, and MI, we excluded individuals with evidence of these diseases during the first year of the observation period, which we designated as a wash-out period. As a result, we excluded 3,794 patients with a history of HZ, stroke, and MI. Furthermore, patients who lacked medical records prior to the occurrence of HZ, stroke, and MI were excluded (n = 45,762). All patients were required to have used statins for at least 6 months prior to the beginning of the study period; those who used statins for less than 6 months prior to the beginning of the study period were excluded (n = 12,499). Individuals lost to follow up were also excluded (n = 415). In addition, people ≤ 40 years old were excluded due to low prescription rate of statin. A total of 275,382 persons were finally observed over an 11-year period from 2003 to 2013 (Fig. 1). A detailed schematic diagram is shown in Fig. 2. Furthermore, we performed subgroup analyses according to subtypes of age groups, considering the different risks of stroke after HZ by age groups [2].

Our main interest was quantifying whether or not, and by how much, the effect of the statins on CVD is mediated by HZ. The total effect of statins on CVD is decomposed by natural direct and indirect effects (or mediation effect) operating through mediator HZ (Supplementary Fig. 1). The definitions of natural direct and indirect effects within a potential outcome framework have been published elsewhere [12].

Categorical variables are presented as numbers and percentages, and were compared using the Pearson’s chi-square test. Continuous variables are expressed as mean ± SD, and were compared using the Student’s t test.

We performed the two-stage regression methods proposed by VanderWeele to evaluate the natural direct and indirect effects [12]. First, we fit a logistic regression for mediator HZ conditional on statin use and confounders. We then fit a Cox proportional hazards regression model for CVD risk on statin, HZ, statin-HZ interaction term, and confounders. Because the statin-HZ interaction effect was not statistically significant, we dropped the exposure-mediator interaction from the final model. The natural direct and indirect effects were estimated using the coefficients of the above regressions and expression provided in VanderWeele’s book [11]. Finally, we calculated the “percentage of proportion mediated” using tthe formula HRNDE (HRNIE -1)HHRNDE HRIDE -1, where HR ^NDE and HR ^NIE are the natural direct and indirect effects of the hazard ratio, respectively. This measure evaluates how much of the overall effect of statins on CVD was due to the effect of statins on the mediator. In the sensitivity analysis, we included the variable chronic pulmonary disease as a variable in order to quantify the impact of unmeasured mediator-outcome confounding on the mediation effect, because chronic pulmonary disease is a well-known major risk factor for HZ and CVD [11]. Also, we conducted additional sensitivity analyses to examine the robustness of our results. All reported p values are two-sided, and p values of < 0.05 were considered statistically significant. Data manipulation and statistical analyses were conducted using SAS version 9.2 (SAS Institute Inc., Cary, NC, USA) and a macro of % mediation [13].

Table 1 presents the baseline clinical characteristics in the cohort. Of a total of 275,382 study subjects, 13,849 individuals (4.9%) developed cardiovascular events during the 12-year follow-up while 15,439 (5.6%) subjects developed herpes zoster. Of total study subjects, 11,415 (4.1%) individuals were using statin.

The multivariate analysis on the risk factors for stroke, MI, CVD, and zoster are shown in Supplementary Table 1. Old age, high body mass index, obesity, diabetes, dyslipidemia, angina pectoris, peripheral vascular disease, and rheumatoid disease were positively associated with CVD and zoster. While, male gender was more common in patients with CVD, while female gender was more common in HZ patients. However, the HZ patients were more likely to be higher income groups and non-smokers, while CVD patients were less likely to be high income groups and smokers.

The occurrence of zoster was significantly associated with statin use (adjusted odds ratio [OR], 2.47; 95% confidence interval [CI], 2.32 to 2.64) (Supplementary Table 1). The total effects of the stroke, MI, and CVD in the statin group were lower than in the non-statin group (Table 2); the adjusted hazard ratios (HRs) were 0.90 (95% CI, 0.82 to 0.98), 0.88 (95% CI, 0.80 to 0.98), and 0.91 (95% CI, 0.79 to 1.07), respectively. However, the CVD was higher in patients who experienced HZ than in those who did not (Table 2); the adjusted HRs were 1.22 (95% CI, 1.12 to 1.34), 1.18 (95% CI, 1.07 to 1.31), and 1.27 (95% CI, 1.09 to 1.49), respectively. We observed significant indirect or mediation effects of statin use on the CVD; the adjusted HRs were 1.01 (95% CI, 1.01 to 1.02), 1.01 (95% CI, 1.00 to 1.02), and 1.02 (95% CI, 1.00 to 1.03), respectively. The direct effects of adjusted HRs for statin use, through pathways other than HZ, were 0.89 (95% CI, 0.81 to 0.96), 0.87 (95% CI, 0.79 to 0.96), and 0.90 (95% CI, 0.77 to 1.05) for the risk of the stroke, MI, and CVD, respectively. The direct effect of statin use was, through pathways other than HZ, a bit bigger than the total effect since statin use increased the incidence of HZ and thereby also increased CVD. When we calculated the mediating effect of cardiovascular events by statin use through HZ, 11.6% of the total beneficial effect of statin use on CVD was mitigated through the occurrence of statin-induced HZ. In addition, this mediating effect was higher in the younger age group (< 60 years) than in older age group (≥ 60 years); the adjusted HRs of the mediation effects were 1.03 (95% CI, 1.01 to 1.04) and 1.01 (95% CI, 1.00 to 1.01) for the risk of the CVD in the younger age group and older age group, respectively (Table 2).

First, we tried to validate our findings by using “chronic pulmonary disease,” which was not included in the multivariable analysis, as one of the unmeasured confounding variables for the sensitivity analysis. The number of patients with chronic pulmonary disease was 3,888 (8.4%) for the HZ group and 42,361 (16.3%) for the Non-HZ group. The adjusted OR for HZ was 1.52 (95% CI, 1.47 to 1.59), and the adjusted HRs were 1.12 (95% CI, 1.07 to 1.16), 1.07 (95% CI, 1.02 to 1.12), and 1.24 (95% CI, 1.15 to 1.34) for the risk of the stroke, MI, and CVD, respectively; the mediation effect through the occurrence of HZ caused by statin use to prevent cardiovascular events was 10.8%. Second, we included individuals between 18 and 40 years old who were excluded in the main analysis due to a low prescription rate of statin for sensitivity analysis. In this scenario, the mediation effect through the occurrence of HZ caused by statin use to prevent cardiovascular events was 9.5%. Third, we included people between the ages of 18 and 40 years who used statins for less than 6 months prior to the study initiation as part of the non-statin group. The mediation effect through the occurrence of HZ caused by statin use to prevent cardiovascular events in this scenario was 9.1%. The detailed results of these sensitivity analyses are presented in Supplementary Table 2.