Ethical considerations.

Written informed consent was obtained from all study individuals. The MONICA Northern Sweden study was approved by the Regional Ethical Committee at Umeå University, Sweden. The FINRISK 1997 study was approved by the Ethical Committee of the National Public Health Institute, Finland. The SHHEC was approved by the then Privacy Advisory Committee and Chief Scientist Committee of the Scottish Home and Health Department and, subsequently, by approximately 30 individual local research ethics committees in Scotland. The MONICA/KORA study was approved by the local authorities in southern Germany and conducted in accordance with the data protection regulations. The KORA study was further approved by the Ethics Committee of the Bavarian Chamber of Physicians, Germany. The MONICA Brianza study was approved by the Comitato Etico Azienda Ospedaliera San Gerardo-Monza, Italy. The Moli-sani study was approved by the Rome Catholic University Ethical Committee, Italy. The MATISS study was approved by the Ethical Committee of the Istituto Superiore di Sanità-ISS, Italy. The MORGAM/MONICA-Catalonia study was approved by the Board of the former Institute of Health Studies, Department of Health and Social Security, Generalitat of Catalunya, Spain.

Additional information regarding the ethical, cultural, and scientific considerations specific to inclusivity in global research is included in the Supporting information (S2 Checklist).

Cross-sectional analyses.

A total of 79,570 participants were eligible for cross-sectional analyses (Fig 1). To study the cohort-specific associations between 25(OH)D status and covariates, we used quantile regression models adjusted for sex, age (continuous using four-knot restricted cubic splines at the fifth, 35th, 65th, and 95th percentile, years), and calendar season of sampling (winter, spring, summer, and fall).

Study populations and follow-up data in prospective analyses.

The number of eligible participants in time-to-event analyses varied by endpoint and ranged from 56,888 participants in the heart failure analysis to 79,570 participants in the CVD mortality analysis (Fig 1). Time-at-risk for an incident endpoint was censored at the date of that endpoint, the date of death, or the end of the follow-up period (with separate models for each endpoint). For composite endpoints, time-at-risk was censored at the date of the first endpoint (if multiple endpoints had occurred), the date of death, or the end of the shortest follow-up period (if endpoints had different follow-up periods, as they had in Moli-sani and MONICA/KORA). Time-at-risk for CVD mortality was censored similarly, ignoring non-fatal endpoints.

Seasonality of CVD outcomes and 25(OH)D status.

Poisson regression models were used to calculate the incidence and mortality rates of CVD by calendar month. A detailed description of the case and person-year calculation is given in S1 Text, but, in brief, we split the follow-up time for each participant on the last day of each calendar month throughout the entire follow-up period (creating up to 12 million observations, depending on the endpoint). To examine whether the seasonal variation in CVD corresponded with that of 25(OH)D, we then calculated sex-, attained age-, and cohort-adjusted rate ratios according to two-month calendar periods. The calendar periods were, in turn, ranked from highest to lowest according to baseline 25(OH)D concentrations using the sex- and age-adjusted median values: August, September, July, October, November, June, December, January, May, February, March, and April. As such, the 25(OH)D ranking of calendar months was based on cross-sectional rather than prospective data (see S1 Fig for unadjusted 25[OH]D data). In sensitivity analyses, the calendar periods were re-arranged according to three-, four-, and six-month groups. Subgroup analyses were performed on persons with “low” and “high” baseline 25(OH)D concentrations, using the sex-, cohort-, and calendar month-specific median value as cut-off value, with the purpose to test for a heterogeneous seasonal variation of CVD outcomes by individual 25(OH)D concentrations. Subgroup analyses were also performed by attained age (≤70/>70 years), sex (men/women), and cohort (eight levels) (S1 Text).

Association between 25(OH)D exposure and CVD outcomes.

Cox regression models, with attained age as the time scale, were used to calculate hazard ratios (HRs) of CVD endpoints by quarters of individual 25(OH)D concentrations and, in a sensitivity analysis, by tenths of individual 25(OH)D concentrations. To account for within- and between-cohort differences in the calendar month of blood sampling, as well as the very large variation in 25(OH)D by sunlight exposure, the categories of 25(OH)D were sex-, cohort-, and calendar month-specific. We chose this methodology since it is reasonable to assume that individuals who are (i) ranked high according to percentiles during the winter also are ranked high during the summer and (ii) of the same sex and from the same geographical area are likely to have a similar sunlight exposure. For the same reasons, 25(OH)D status was not analyzed as a continuous variable. The standard assumption of proportional hazards along the time scale was evaluated by modeling the possible interaction between the time scale and the categorical exposure variable (S1 Text) [33].

The basic model was adjusted for sex (men/women), cohort (eight levels), and calendar week of blood sampling (continuous). The multivariable model was further adjusted for BMI (continuous, kg/m²), smoking status (no/yes), systolic blood pressure (continuous, mmHg), antihypertensive medication (no/yes), total cholesterol concentration (continuous, mmol/L), crea-eGFR (continuous, ml/min/1.73 m²), history of diabetes (no/yes), and history of CVD (no/yes; only included in the mortality analysis). Continuous covariates were modeled using four-knot restricted cubic splines at the fifth, 35th, 65th, and 95th percentile. Sex, smoking status, antihypertensive medication, and history of diabetes and CVD were accounted for by stratification of the baseline hazard function, because they did not always meet the proportional hazards assumption.

Missing data on 25(OH)D and covariates in the primary endpoint analyses were handled using multiple imputation by chained equations. Endpoint data were not imputed. The percentages of missing data by endpoint and analytical model are shown in S3 Table, and a detailed description of the multiple imputation method is reported in S1 Text. In a sensitivity analysis, we repeated the analysis using complete case data. In a first extension of the complete case model, in which participants from MATISS and subcohort 4 of MONICA/KORA were excluded due to lack of education data (n = 6924 and 7559 participants for CVD incidence and mortality, respectively), we further adjusted the multivariable model for education (three levels). In further extensions of the complete case model, which aimed to account for exposure misclassification over time and for long baseline examination periods in some cohorts, we restricted the follow-up time to a maximum of five years and re-categorized the exposure variable according to calendar year of examination (S1 Text). To avoid repeated imputations—and since results based on imputed and complete case data were almost identical for CVD incidence and mortality—the Cox regression models of secondary endpoints were only based on complete case data.

To examine reverse causality bias, we constructed three separate models with different baseline exclusions in another sensitivity analysis (S1 Text). Separate models by sex (men/women) and cohort (eight levels) were also performed as subgroup analyses (S1 Text).

Study populations and follow-up data.

As can be seen in Fig 1, the number of included cohorts in time-to-event analyses varied by CVD endpoint. The total number of participants, person-years, and cases for each endpoint are shown in Table 2. The incidence rate per 10,000 person-years ranged from 54.7 cases for coronary heart disease to 24.6 cases for stroke; and the mortality rate was 30.0 cases per 10,000 person-years. (Nota bene: The cohort-specific incidence and mortality rates are given in S5 Table, conditional on age, sex, and decade; and the co-occurrence of endpoints is presented S6 Table.)

Download:

• PNG
  larger image
• TIFF
  original image

Table 2. Follow-up data in time-to-event analyses.

https://doi.org/10.1371/journal.pone.0319607.t002

Seasonality of CVD outcomes and 25(OH)D status.

There was a large difference in the median 25(OH)D concentration by calendar month, with the peak observed in August and the trough observed in March (+48 and -18%, respectively, compared to the full year estimate) (Fig 2). For CVD incidence, the observed rates were rather stable over the calendar year except for a trough in July and August (trough and peak difference of -14 and +7% compared to the full year estimate). A more variable pattern was observed for CVD mortality, with a trough that started in May and reached a minimum in September and with a peak in January (trough and peak difference of -19 and +21% compared to the full year estimate).

Download:

• PNG
  larger image
• TIFF
  original image

Fig 2. Seasonal variation of (i) cardiovascular disease (CVD) incidence and mortality during follow-up and (ii) 25-hydroxyvitamin D (25[OH]D) at baseline. Hollow circles and capped spikes represent point estimates and 95% CI for Poisson regression-derived rates (left y-axis). Bars and capped spikes represent point estimates and 95% CI for quantile regression-derived median concentrations (right y-axis). The dashed lines represent point estimates for the entire calendar year.

https://doi.org/10.1371/journal.pone.0319607.g002

Sex-, attained age-, and cohort-adjusted rate ratios of CVD incidence and mortality by two-month calendar periods, ordered by 25(OH)D status, are shown in Table 3 (and by three- to six-month calendar periods in S7 Table). The trough months of 25(OH)D in March and April had a similar incidence but a higher mortality of CVD than the peak months of 25(OH)D in August and September (p for comparison = 0.12 and < 0.001, respectively). There were no differences in CVD outcomes between the trough months and the non-peak months of 25(OH)D (despite median concentration differences up to 62%; p for comparison ≥ 0.077). (Nota bene: Incidence rate ratios for coronary heart disease and stroke according to the same categories are presented in S8 Table, for which there was a difference between the trough and peak months of 25[OH]D for stroke but not for coronary heart disease.)

Download:

• PNG
  larger image
• TIFF
  original image

Table 3. Rate ratios of CVD incidence and mortality by two-month calendar periods (ordered by 25[OH]D status).

https://doi.org/10.1371/journal.pone.0319607.t003

The shape of the association between seasonal changes in 25(OH)D and seasonal changes in CVD did not differ by low or high baseline 25(OH)D concentrations (p for interaction ≥ 0.10; S4 Fig). The comparison of trough and peak months of 25(OH)D was also similar in subgroup analyses by attained age, sex, and cohort (p for interaction ≥ 0.21; S5 Fig).

Association between 25(OH)D exposure and CVD outcomes.

Cox regression-derived HRs of CVD incidence and mortality by quarters and tenths of 25(OH)D status are shown in Table 4 and S6 Fig. In the age-, sex-, and season-adjusted model, the highest quarter of 25(OH)D had a strong inverse association with CVD incidence and mortality: the HR was reduced by 31 and 49%, respectively, compared to the lowest quarter (Table 4). In the multivariable-adjusted model, the inverse associations were largely attenuated but remained strong with HR reductions of 18% for CVD incidence and 36% for CVD mortality. HRs of secondary endpoints are presented in S9 Table. In brief, coronary heart disease, stroke, and heart failure had inverse associations with 25(OH)D status and atrial fibrillation had a null association with 25(OH)D status.

Download:

• PNG
  larger image
• TIFF
  original image

Table 4. HRs of CVD incidence and mortality, based on multiple imputed data and according to quarters of 25(OH)D status.

https://doi.org/10.1371/journal.pone.0319607.t004

The results did not markedly change in sensitivity analyses based on complete case data, irrespective of further adjustment for education, restriction of the follow-up time, or re-categorization of the exposure variable (S10 Table). There seemed to be a small degree of reverse causality in the association of 25(OH)D status with CVD mortality, especially when excluding prevalent cases of CVD, but not in the association with CVD incidence (relative differences in HR of 6.3 and 1.2%, respectively, in the models excluding the first three years of follow-up; S11 Table).

There was no evidence of an interaction between 25(OH)D status and attained age or sex in relation to CVD incidence (p for interaction ≥ 0.17; S7 and S8 Figs), with multivariable-adjusted HRs of 0.83 (95% CI: 0.75–0.92) and 0.80 (95% CI: 0.70–0.91) in men and women, respectively, for the comparison of extreme quarters. In contrast, there was evidence of an interaction by attained age for CVD mortality in the highest 25(OH)D category (p for interaction = 0.020 for quarter four compared to quarter one), with deviation from the time scale-fixed HR in participants above 80 years of age (S7 Fig). In the same quarter, CVD mortality also differed by sex (p for interaction = 0.019 compared to quarter one), with multivariable-adjusted HRs of 0.59 (95% CI: 0.51–0.68) in men and 0.76 (95% CI: 0.62–0.93) in women (S8 Fig). The exposure-endpoint associations by cohort are shown in S9 Fig, for which there was no clear evidence of between-cohort heterogeneity (p for interaction ≥ 0.052).

Seasonality of CVD outcomes and 25(OH)D status.

An often overlooked aspect of 25(OH)D in relation to health outcomes is the Earth’s axis deviation and angular rotation, which leads to a never-ending natural experiment with oscillating 25(OH)D concentrations on a season-to-season basis, at least in the Northern hemisphere. In 1981, it was hypothesized by Scragg that hypovitaminosis D was a major contributor to the observed winter excess in CVD outcomes, either via direct or indirect effects on thrombosis formation [26]. Since then, short-term reductions in 25(OH)D concentrations have in fact been associated with different measures of arterial stiffness and vascular dysfunction in observational data [34]. One feature that Scragg found especially attractive with his ecological hypothesis was that it was “testable”. When it finally was tested, some 34 years later in the SHHEC (n = 13,224), there was no evidence that seasonal variation in 25(OH)D had a large impact on seasonal variation in CVD outcomes [27]. Despite a 2:1 ratio of 25(OH)D concentrations between August and March, there was no peak or trough in CVD incidence, and the peak and trough in CVD mortality occurred near the solstices, preceding the extremes of 25(OH)D by at least two months. In addition, there was no increased outcome variability in persons with “low” 25(OH)D status, defined by below-median 25(OH)D concentrations; that is, persons who should experience more hypovitaminosis D during the winter. As such, Scragg’s hypothesis was rejected by the authors, given that any underlying mechanism explaining a seasonal variation in CVD outcomes must occur acutely (as acknowledged by Scragg himself [26]).

Our study was an extension of the original SHHEC paper, in which Scragg’s hypothesis was examined once again but with data from several geographical regions and with a larger number of participants, person-years, and CVD cases. As expected, there was a large difference in the median 25(OH)D concentration by calendar month, with the peak observed in August and the trough observed in March (1.79:1 ratio). The peak-to-trough ratio was also diluted by the Swedish cohort, which had the highest median 25(OH)D concentrations despite a non-full year (winter-to-spring) examination period (1.93:1 ratio, if that cohort had been excluded). Nonetheless, the trough months of 25(OH)D in March and April had a similar CVD incidence as the peak months of 25(OH)D in August and September. CVD mortality was slightly higher in trough months compared to peak months but not compared to other non-peak months of 25(OH)D, leading to a rather nonsensical trend across categories with increasing differences in median 25(OH)D concentrations. Importantly, the peak in CVD mortality in January preceded the trough in 25(OH)D by two to three months, not adhering to the cause-must-precede-outcome criterion of the Bradford Hill criteria [35]. The trough in CVD mortality also started in May, in adjacent to the trough in 25(OH)D. Finally, similar to the SHHEC study, there was no evidence of an increased susceptibility of CVD outcomes during the trough months of 25(OH)D in persons with “low” baseline 25(OH)D status, defined by a 25(OH)D concentration below the sex- and calendar month-specific median value in each cohort.

All in all, and based on the findings from our study and the previous SHHEC study, we feel comfortable to reject Scragg’s hypothesis that short-term reductions in 25(OH)D concentrations are a major contributor to seasonal increases in CVD outcomes. While a winter excess of CVD cannot be denied, more so for mortality than for incidence in our data, the underlying cause is likely multifactorial and not due to one single factor (e.g., temperature, air pollution, and respiratory infections) [24,25,36]. In analogy, we do believe that a lack of sunlight exposure during the winter plays a role, but to reduce that role to a mere 25(OH)D pathway is an oversimplification.

Association between 25(OH)D exposure and CVD outcomes.

The existing literature of observational studies on the association between 25(OH)D status and CVD outcomes is extremely large, to the point where it is almost hard to get an overview. Previous studies have reported an increased mortality of CVD [1–3], as well as an increased incidence of coronary heart disease [4], stroke [5], heart failure [6], and atrial fibrillation [7], in persons with low 25(OH)D concentrations. Similar findings have also been observed in persons with comorbid conditions like diabetes [37,38]. Our study does not add novelty to the overall body of observational data but confirms most of the previous findings, with rather strong but surprisingly similar associations of 25(OH)D status with mortality of CVD and with incidence of coronary heart disease, stroke, and heart failure. Except for an interaction by age and sex on the highest 25(OH)D quarter in the CVD mortality analysis, with a weaker association in old and female participants, the exposure-outcome associations were also consistent across multiple sensitivity and subgroup analyses. In contrast, we did not observe an association between 25(OH)D status and incidence of atrial fibrillation. In a meta-analysis by Liu et al, which included data from six studies (n = 5503 cases; N = 66,139 participants), vitamin D deficiency (< 50 nmol/L) was associated with an increased incidence of atrial fibrillation [7]. However, as discussed by Gaksch et al [3], most meta-analyses on 25(OH)D status and CVD outcomes have been hampered by between-study variations in the assay method of 25(OH)D, an issue to which the meta-analysis by Liu et al is no exception.

Whether the observational associations between 25(OH)D concentrations and CVD outcomes are causal have been debated for many years. The HRs in our study were largely attenuated by multivariable adjustment, notwithstanding that we could not adjust for a number of potential confounders (e.g., diet, alcohol, and physical activity). A causal role of 25(OH)D has not been supported by randomized clinical trials, in which vitamin D supplementation has neither reduced incidence and mortality of CVD [16–20,38] nor improved different biomarkers of glycaemia, inflammation, and lipids [39]. It should, however, be noted that vitamin D supplementation has been associated with an improved cardiac function in patients with heart failure (by reducing the left ventricular end-diastolic diameter and increasing the left ventricular ejection fraction) [40].

Up until recently, there was no evidence from Mendelian randomization studies of an association between genetically-predicted 25(OH)D concentrations and incidence or mortality of CVD [8–11]. However, three Mendelian randomization studies published in 2021 and 2022 (all based on the UK Biobank and utilizing a so-called non-linear randomization) did observe an inverse association with CVD incidence [12,13] and CVD mortality [12,14], especially in subgroup analyses of persons with low 25(OH)D concentrations. In addition, another Mendelian randomization study found an inverse association with heart failure incidence [15]. Since then, the methods and model assumptions of non-linear Mendelian randomization studies have been under great scrutiny [41]. Subsequently, one of the aforementioned studies was recently retracted and republished with an altered conclusion, no longer supporting a causal relationship between genetically-predicted 25(OH)D concentrations and CVD outcomes [12,42].

Despite methodological issues, the above-mentioned Mendelian randomization studies have re-ignited the debate on vitamin D supplementation and to whom such supplementation should be focused on. However, given that more than 40 years have passed since the publication of Scragg’s hypothesis, the scientific and medical community must at some point decide on the importance of 25(OH)D in the primary and secondary prevention of CVD.