The mediating role of the venules between smoking and ischemic stroke

This study is based on the Rotterdam Study (RS), a large prospective population-based cohort study in the Netherlands that investigates causes and consequences of chronic diseases in the general population [11]. All inhabitants of the Ommoord district in the city Rotterdam, aged 55 years or older, were invited to the study in 1990 (RS-I). In 2000 those who had become 55 years of age or moved into the study district were invited (RS-II). In 2006 a further extension of the cohort was initiated and inhabitants aged 45 years or older were invited (RS-III). Home interviews including assessment of smoking, and physical examinations take place every 3–4 years. Retinal vascular calibers were measured at the baseline visit of RS-I (RS-I-1), in a random sample of the second visit of RS-II (RS-II-2), and at the baseline visit of RS-III (RS-III-1), see Fig. 1. We considered the date at which smoking assessment was done in RS-I-1, RS-II-2 and RS-III-1 as our baseline. We excluded persons without data on smoking, without gradable retinal photographs, persons with a history of stroke at baseline, and persons who did not give permission to monitor for future disease event. The Rotterdam Study has been approved by the Medical Ethics Committee of the Erasmus MC and by the Ministry of Health, Welfare and Sport of the Netherlands, implementing the “Population Studies Act: Rotterdam Study” (Wet Bevolkingsonderzoek: ERGO). All participants provided written informed consent to participate in the study and to obtain information from their treating physicians.

Information on smoking behavior was obtained using a computerized questionnaire during the home visits. Participants were classified as current smokers, or non-smokers (including both former or never smokers). Current smokers were participants who answered yes to the question: “are you currently smoking?” Former smokers were participants who answered no to this question and answered positively to the question: “are you a former smoker?”

Participants underwent a full eye examination at each subcohort’s baseline (1990, 2004, and 2006) including simultaneous stereoscopic retinal color photography of the optic disc (20° field, Topcon Optical Company, Tokyo, Japan) of both eyes after pharmacological mydriasis. For each participant, the image of one eye with the best quality was analyzed with the Retinal Vessel Measurement System (Retinal Analysis, Optimate, WI; Department of Ophthalmology & Visual Science, University of Wisconsin-Madison). For each participant one summary value was calculated for the arteriolar diameters and one for the venular diameters (in μm) of the blood column after correction for differences in magnification due to refractive status of the eye, enabling us to use the separate arteriolar and venular diameter sum values [12, 13]. We verified in a random sub-sample of 100 participants in RS-I that individual measurements in the left and right eye were similar. Measurements were performed by total four trained raters, masked for participant characteristics. Pearson’s correlation coefficient for interrater agreement were for arteriolar diameters 0.67–0.87, and for venular diameters 0.91–0.94. For intrarater agreement the correlation coefficients were 0.65–0.88 for arteriolar diameters, and 0.82–0.95 for venular diameters.

History of stroke was assessed using home interviews and confirmed by reviewing medical records. Participants were continuously followed up for stroke through digital linkage of general practitioners’ files with the study database [14]. Furthermore, nursing home physicians’ files and files from general practitioners of participants who moved out of the district were checked on a regular basis. Hospital discharge letters and information from general practitioners was collected for all potential strokes. Research physicians reviewed the information and an experienced neurologist verified the strokes. Strokes were further classified into ischemic or hemorrhagic based on neuroimaging reports. Subarachnoid hemorrhages were excluded. Infarcts that turned hemorrhagic were classified as ischemic stroke. If neuroimaging was lacking, a stroke was classified as unspecified. Follow-up was completed until 1 January 2016. Participants suffering stroke at any point in follow-up were dichotomized as yes/no.

We considered the “smoking/ischemic stroke” relation, and “wider venules/ischemic stroke” relation to be confounded by the same set of measured persons characteristics. All confounders were assessed at the same visit period when smoking was assessed and when retinal photographs were obtained. Blood pressure was measured twice in sitting position at the right brachial artery with a random-zero sphygmomanometer, and the average of two readings was used for analysis. Body mass index was computed as weight (kg) divided by height squared (m²). Serum total and high-density lipoprotein cholesterol concentrations were determined by means of an automated enzymatic procedure. White blood cell count was assessed in citrate plasma using a Coulter counter T540 (Coulter electronics, Luton, England). Diabetes mellitus was considered present if participants reported use of antidiabetic medication or when non-fasting serum glucose level was greater than 11 mmol/l, or when fasting glucose level in serum was greater than 7 mmol/l. Alcohol consumption was calculated as amount of alcohol in g/day. Atherosclerotic plaques were assessed by ultrasound at the carotid artery bifurcation, common carotid artery, and internal carotid artery on both sides. Plaques were defined as focal thickening of the vessel wall of at least 1.5 times the average intima-media thickness relative to adjacent segments with or without calcified components. The carotid artery plaque score (range 0–6) reflected the number of these locations with plaques. Information on blood pressure lowering medication use and education level (low: primary education, intermediate: secondary general or vocational education, or high: higher vocational education or university) was obtained during the home interview by a questionnaire. Definitions of a history of myocardial infarction, coronary artery bypass graft, and percutaneous coronary intervention, has been described extensively previously [15].

In causal mediation analysis, the total causal effect of the exposure on the outcome can be decomposed into four components in the presence of an exposure-mediator interaction [10]: the controlled direct effect due to neither mediation nor interaction (in the current study: the effect of smoking on ischemic stroke which does not go through wider venules), reference interaction effect due to interaction alone (in the current study: the interaction of smoking with the observed level of wider venules), mediated interaction effect due to both mediation and interaction (in the current study: the interaction of smoking with wider venules that has been explicitly caused by smoking), and the pure indirect effect due to mediation alone (in the current study: the effect of smoking on ischemic stroke which purely goes through wider venules). Of note, for mediated interaction it is defined by the mediator having an effect on the outcome only in the presence of the exposure, but not in the absence of exposure, i.e. the exposure is necessarily present for the mediator to affect the outcome. Pure mediation is defined by the mediator having an effect on the outcome even in the absence of the exposure. Moreover, the controlled direct effect and reference interaction effect can be combined into the natural direct effect. Similarly, the mediated interaction effect and pure indirect effect can be combined into the natural indirect effect [16‐18]. Figure 2 shows a conceptual drawing of the four-way decomposition, and Fig. 3 shows a causal diagram representing the association of smoking with risk of ischemic stroke mediated by wider venules. Several conditions need to be met in order to identify each of the four components, including the following assumptions concerning confounding: (1) the effect of smoking on ischemic stroke should be unconfounded conditional on the baseline covariates; (2) the effect of wider venules on incident ischemic stroke should be unconfounded conditional on smoking and baseline covariates; (3) the effect of smoking on wider venules should be unconfounded conditional on baseline covariates; (4) the mediator-outcome confounders should not be affected by smoking. Of note, for the total effect only the first of these assumptions is required; the controlled direct effect only requires the first and second assumptions [10].

To obtain estimates of the four components from the data, a regression-based approach was used. First, we used a logistic regression model to regress incident ischemic stroke on smoking, retinal venular caliber, a product term denoting the interaction between smoking and the retinal venular caliber, and the covariates. Next, we used a linear regression model to regress retinal venular caliber on smoking and the covariates. We adjusted both regressions for the following covariates: age, sex, subcohort, education, systolic blood pressure, diastolic blood pressure, blood pressure lowering medication use, body mass index, total cholesterol, high-density lipoprotein cholesterol, white blood cell count, diabetes mellitus, alcohol intake, carotid plaque, history of cardiovascular disease, and the retinal arteriolar caliber. We obtained estimates of the four components by combining parameters from these two models according to the analytic expressions provided by VanderWeele [10]. Confidence intervals for the effect estimates were obtained via case resampling bootstrapping with 1000 iterations. We defined the proportion of the effect that was attributable to each component by dividing the estimate of a component by the excess relative risk. The overall proportion mediated was defined as the pure indirect effect plus mediated interaction divided by the excess relative risk; the overall proportion attributable to interaction was defined as the reference interaction plus mediated interaction divided by the excess relative risk; the overall proportion eliminated was defined as the excess relative risk minus the controlled direct effect divided by the excess relative risk.

In additional analyses, we aimed to address whether it is the smoking at the moment of the venular caliber assessment or the history of smoking that mattered most. Therefore, we redid the analyses by comparing ever smoking (i.e. past smoking and current smoking) to never smoking.

Finally, we performed sensitivity analyses to determine how much the estimates of the natural direct and indirect effect would change under different degrees of confounding by a hypothetical unmeasured binary confounder. Although we have adjusted for several potential confounders, the estimates will of course be biased if there remains important unmeasured or residual confounding of the relation between wider venules and ischemic stroke; here, for example, we may be concerned there is residual confounding by unhealthy lifestyle (e.g. unhealthy diets or physical inactivity). For this analysis, we used bias formulas for odds ratios for natural direct and indirect effects described in VanderWeele [19]. Briefly, bias-corrected-effect estimates for the natural direct and indirect effect were calculated by specifying a range of plausible values for the effect of the unmeasured confounder on ischemic stroke, and specifying a range of plausible values for the prevalence of unmeasured confounder among non-smokers and smokers. Next, these bias-corrected-effect estimates were subtracted from the crude natural direct and indirect effect (i.e. adjusted for age, sex, and subcohort) to address change in effect estimate if we had been able to adjust for this unmeasured confounder. Hence, a positive difference could be interpreted as an underestimation of the observed natural direct or indirect effect estimate failing to adjust for the unmeasured confounder, whereas a negative difference could be interpreted as an overestimation. All analyses were performed in R version 3.2.3.