The epidemiologic study suggesting a relation between serum 25-hydroxyvitamin D (25[OH]D) and C-reactive protein (CRP) by Amer and Qayyum is timely, given the prominence of 25(OH)D in recent scientific research and the recent increase in vitamin D recommended daily allowances. However, the study appears to suffer from deficiencies in internal validity (due to the analysis methods used) and external validity (because vitamin D supplementation by subjects was not included in the analysis). Their findings contradict evidence for a dose-dependent reduction of coronary heart disease mortality with higher serum 25(OH)D, and furthermore, CRP has questionable utility in guiding treatment of patients with coronary heart disease. Best practices in data analysis suggest several steps that should have been taken to avoid bias and uncertainty.

The scale of the y axis in Figure 1 should have been back-transformed from the logarithm of CRP to its raw (original) value. The investigators’ multivariate analysis results are counterintuitive, because in Table 1, we can read that subjects with 25(OH)D levels <21 ng/ml had higher mean CRP levels than subjects with 25(OH)D levels >21 ng/ml. An appropriately scaled graph would allow readers to visually grasp the relations being examined and assess the conclusions being drawn: is it graphically obvious that CRP increases at serum 25(OH)D levels >21 ng/ml?

When building a statistical model, one can introduce bias into the model or, even worse, mathematically create associations that are not there in real life (e.g., by including variables such as coronary heart disease, which may be on the causal pathway between 25[OH]D and CRP). The investigators provided considerable generic details of how variables were measured in the National Health and Nutrition Examination Survey (NHANES) but did not justify their choice of including those variables in multivariate analysis. A structured approach to statistical model building with an a priori rationale, guided by modern understanding of causality and model-fitting tools such as the Akaike criterion, can help avoid or reduce these biases. It is thus possible that on reaching steps 5 to 9 of their stepwise model-building process (Table 2), the investigators may have achieved a model that yielded significant β coefficients but was in fact a poorly fitted model.

We believe the data do not support the investigators’ conclusions in their abstract. They did not test the association of serum 25(OH)D with mortality or morbidity but rather stopped short at serum CRP. Because they did not consider vitamin D supplement use by NHANES subjects, there is no evidence to support concluding the abstract of their highly visible report (which attracted media attention) with the recommendation that we readers should question the use of vitamin D as a dietary supplement in patients with 25(OH)D levels >21 ng/ml. Our concerns about their conclusion are warranted given results of another 25(OH)D study using NHANES data, in which serum 25(OH)D was inversely associated with all-cause and cardiovascular disease-specific mortality in hypertensive United States adults (using a statistical model that included CRP as an adjustment variable); comparisons were made against a 25(OH)D reference level of 29 ng/ml.

Finally, the investigators acknowledged that there were available data on other inflammatory markers but chose not to use them. Because their study was of an exploratory nature and not designed as a causal analysis, these other markers could have been included to make the study a comparative analysis.