Comparison with other studies

Risk of myocardial infarction and other thrombotic complications is typically higher during the winter months than during the summer [20], [21], and in line with our study, fibrinogen (but not CRP) has been reported to vary by season [22]. However, there is little information on the direct associations between 25(OH)D and circulating markers of hemostasis in apparently healthy adults. In line with our findings, an earlier smaller study found the association between 25(OH)D and CRP to be explained by adiposity, while an independent relation persisted between 25(OH)D and tPA [23]. The lack of evidence for an association between current vitamin D status and CRP agrees with recent observations in other populations [24], including intervention studies where vitamin D supplementation at varying dosages (700–3332 IU per day) have failed to achieve changes in CRP concentrations [25], [26], [27]. There is, however, one earlier intervention which showed a significant (22%) fall in CRP following vitamin D supplementation (50,000IU every three months over one year) in vitamin D deficient British Bangladeshi adults [28]. It is possible that associations between 25(OH)D and CRP might not be detected in general population studies if effects are confined to extreme vitamin D deficient groups.

Explanations

Increased concentrations of tPA and D-dimer are thought to serve as markers for aggravated fibrinolytic activity reflecting increased future burden of CVD [29]. Hence, the inverse associations of 25(OH)D with tPA and D-dimer observed in our study support the role of vitamin D metabolism in maintaining antithrombotic homeostasis. The direct influences of hormonal vitamin D axis on hemostasis are not well established, although recent gene-expression studies suggest that vitamin D analogues may suppress thrombogenicity and enhance fibrinolysis thereby reducing intimal plaque formation [30], [31]. Vitamin D analogues have been observed to suppress PAI-1 expression in human coronary artery smooth muscle cells [32]. Up regulation of PAI-1 has been associated with increased risk of CVD, and it has been suggested that suppression of PAI-1 expression may contribute to the observations on improved survival among patients with chronic kidney disease who are taking vitamin D analogues [7], [32].

Hypovitaminosis D is believed to have wide-ranging influences on vascular physiology, which include both direct (e.g. influences on endothelial cells) and indirect pathways (endocrine, immunomodulatory) [33]. Vitamin D toxicity has been associated with adverse effects on vascular calcification, but available data indicates that calcification is increased also in hypovitaminosis D [34], [35]. Associations with increased blood pressure are believed to be mediated through decreased renin production [36] and it has been suggested that (independently of blood pressure), this could also affect vascular stiffness [33]. Vitamin D might also exert anti-proliferative effects on vascular smooth muscle cells, thereby affecting myocardial cell hypertrophy and proliferation.

Discussion regarding optimal status for 25(OH)D concentration is ongoing, and there is some debate about whether a threshold exists [37], [38], [39], [40]. The curved association between 25(OH)D and D-dimer, together with the suggestive elevations in fibrinogen and CRP at the extreme of high concentrations observed in this study, could support a threshold effect with the optimal concentration being between 60 and 120nmol/l. These results are in line with an evaluation using data on multiple health outcomes (including bone mineral density, lower extremity function, risk of falls, fractures and colorectal cancer) [40], which corresponds to an earlier consensus statement on osteoporosis [37] suggesting 75nmol/l as the lower reference value. Earlier studies have provided tentative evidence for a trend towards lower bone mineral density and prolonged sit-to-stand time at a higher extreme of 25(OH)D concentrations (>140 nmol/l) [40].

Methodological considerations

The main strength of this study lies in the large sample of participants, which provided adequate power for detailed investigation of the associations between these inter-related health indicators. Moreover, as data collection covered the full seasonal range, we were able to obtain further support for the key findings from the independent evaluation of seasonal patterns in inflammatory and hemostatic markers. Given the exceptional information available from the 1958BC, we were able to adjust for multiple factors in our analyses thereby controlling for confounding introduced by demographic, lifestyle or social variations. Final models evaluating the independent effect of 25(OH)D on inflammatory and hemostatic outcomes were adjusted for quadratic terms in both BMI and waist circumference in order to control for adiposity as fully as possible. The full attenuation of the association between 25(OH)D with CRP and fibrinogen after adjustment for the available indicators suggests that these measures were sufficient for this purpose.

Comparison between the effect of adjustment for 25(OH)D concentrations in the observed seasonal patterns in the inflammatory/hemostatic factors, and the direct associations between 25(OH)D and these outcomes, demonstrates the limitations of cross-sectional analysis of data and the problem of possible over/under adjustment. Given the strong influence of obesity on 25(OH)D concentrations, the latter would be expected to be associated with any factor that is strongly related to obesity (given a tolerable degree of measurement error and sufficient sample size). This argues for the need to adjust for obesity fully to reduce the likelihood of a false positive association due to confounding. However, it could also be argued that adjustment for adiposity may lead to an underestimation of associations between 25(OH)D and inflammatory/hemostatic markers given that adiposity is a key determinant for 25(OH)D [19]. Possible over-adjustment could explain why we observed some evidence for effect mediation by 25(OH)D on our seasonal modeling of fibrinogen, while the inverse relation seen in the unadjusted cross-sectional analyses between these two factors was fully attenuated by the adjustment for indicators of adiposity and lifestyle/social class.

Some further limitations need to be considered in relation to these findings. Given the observational design, we cannot prove causality or fully discount residual confounding by unmeasured variations. Residual confounding may also affect our seasonal mediation models; however, relevant confounders are likely to differ given that potential factors presumed important for the direct associations (such as adiposity) would not necessarily have seasonal patterns. Although 25(OH)D is the best indicator for vitamin D status in humans [2] information on serum parathyroid levels or 1,25(OH)₂D is not available and this precludes more detailed investigation of vitamin D metabolism. Furthermore, information was available only for the most commonly used CVD risk markers, while other, perhaps more relevant, indicators were not measured. Finally, although the 1958BC has been reported to remain generally representative of the current UK population, there is an underestimation of some minority groups [41]. These analyses were restricted to individuals of white Caucasian origin, which will reduce population stratification, however, generalization of these findings to other ethnic groups should be done with caution.

Conclusions

Current vitamin D status was associated with circulating concentrations of tPA and D-dimer, which may suggest a role for vitamin D in maintaining antithrombotic homeostasis. Further studies, including randomised controlled trials, are needed to demonstrate the role of vitamin D metabolism in cardiovascular health, and whether vitamin D supplementation or improved vitamin D status could have beneficial effects.

Laboratory analyses

Venous blood samples were obtained without prior fasting and posted to collaborating laboratories. Fibrinogen was determined by the Clauss method and CRP assayed by nephelometry (Dade Behring) on citrated plasma samples after one thaw cycle. vWF antigen was measured by Decollates enzyme-linked immunosorbent assay (elisa) and tPA antigen by Biopool elisa. 25(OH)D was measured using automated application of the IDS OCTEIA elisa on the Dade-Behring BEP2000 analyser (sensitivity of 5.0 nmol/l, linearity ≤155 nmol/l, intra-assay variation CV 5.3–7.4% and inter-assay variation CV 7.7–8.5%) [41], [43]. Heterogeneity of 25(OH)D concentrations measured by different assays is well-known, therefore levels were standardized according to the mean of the Vitamin D External Quality Assessment Scheme (DEQAS) [43].

Demographic, lifestyle and social factors have been described in detail previously [19], [42], [44], [45], [46]. In brief, weight, height and waist circumference were measured at 45y. Socio-economic position at birth (1958) and at age 42y was assessed using the Registrar General's occupational classification categorised as I &II (professional/managerial), III non-manual, III manual, IV & V (unskilled manual) [44]. Recreation Metabolic Equivalent of Task (MET) hours per week at 45y was derived from reported frequencies and usual durations for up to 37 activities, and published MET scores. We divided recreational activity into gender-specific quartiles. An additional category was created for implausibly high values (participants with weekly recreation hours of 3 standard deviations (SD) above the gender mean). Participants engaging in vigorous activity were those who recorded an activity with a MET score greater than six. Time spent watching a television or using a computer was reported at age 45y [19]. Information on smoking was based on smoking history recorded at ages 23y, 33y and 42y [46] and alcohol consumption on report from age 45y.

Statistical Analysis

To describe the distribution of 25(OH)D concentration we used dichotomous indicators for levels below 25 nmol/l and above 125 nmol/l and a categorised factor into 25 nmol/l divisions with minimum <25 nmol/l and maximum ≥125 nmol/l tails. The natural logarithmic transformation was used to calculate geometric means to adjust the skewed distribution.

Inflammatory and hemostatic markers were transformed to gender-specific standard deviation scores (SDS) to compare variation across models. The SDS values were used as response variables in linear mixed effects regression models. Initial analyses included validation and graphical examination of data, statistical evaluation of linear and quadratic terms for 25(OH)D, adiposity measures (BMI and waist circumference), and single and joint effects of these measures on the inflammatory/hemostatic outcomes. Continuous measures were used in testing for interactions between the adiposity measures and 25(OH)D on the outcomes. For D-dimer three outlying observations were identified from graphical examination and model diagnostics (leverage and/or influence >2SD), and excluded from further analysis. We fitted linear regression models in three stages, starting with simple associations between 25(OH)D (categorized into 10 nmol/l groups, minimum <25 nmol/l and maximum ≥125 nmol/l tails) and the SDS inflammatory/hemostatic outcomes (model 1), next adjusting in addition for demographic, lifestyle and social factors (model 2), and finally adjusting for adiposity in addition to lifestyle and social factors (model 3). Models (1–3) included covariates gender and month of measurement and models with fibrinogen SDS as the outcome included laboratory assay batch. We also created an additional 25(OH)D category variable of <25, 25–74.9 and ≥75 nmol/l to summarize the effect size and repeated analyses for model 3. Missing information on the lifestyle factors (n = 514 with missing information on one or more factors) was imputed using the multiple imputation chained equations [47]. The models (1–3) were run on 10 imputed datasets and repeated for the sample restricted to participants with complete data on all confounders. The results were similar with both approaches; hence, results are only presented from the imputed models.

Seasonal variations were modeled using sine and cosine functions [48] with laboratory assay batch as a random effect on the intercept where appropriate (Likelihood Ratio Test (LRT) on between-batch variation p<0.05). The model equation was:Where is the response variable, i denotes the individual within the assay batch j , T is the time period (T = 365 days), n_c is the number of seasonal patterns with sequence () or combination of them, p_k is the period for the season () where , X_i3 is a matrix of linear predictors for the vector of parameters β₃ (inclusive of the mediator variable and seasonal confounders). The random effect in the intercept, , is defined by the laboratory assay batch as the dates of the blood samples were associated with the assay batches and we did not wish to estimate the effect of batch on the response; we assume ∼. Finally, the observational error terms were assumed to be normally distributed with mean 0 and variance .

Without the hierarchical structure implied by the random effects terms we may have under-estimated the standard error on the intercept. The n_c seasonal patterns investigated were yearly, half yearly, quarterly, and all their combinations of were included if appropriate. Models were adjusted for the potential seasonal confounders of respiratory infections, alcohol consumption, PC/TV time, physical activity and social class at birth and adulthood. These resulting models were further adjusted for 25(OH)D, its dominant yearly pattern and then re-tested for subsidiary cycles.

To quantify the seasonal effect of 25(OH)D on the outcomes we used the concept of mediation analysis [47], [49] where season, as modeled with the sine/cosine transformation, was the independent variable and 25(OH)D acted as its mediator to the outcomes. The product of coefficients test [50] used in mediation analysis was extended to allow for the amplitude (the seasonal variation around the mean) as derived from the sine/cosine transformation. The final seasonal models predict the mean levels of the outcomes from the partial regression coefficients of the seasonal functions. All analyses were carried out using STATA, version 10.0 (StataCorp LP, College Station, TX).