Predictors of 25-hydroxyvitamin D and its association with risk factors for prostate cancer: evidence from the prostate testing for cancer and treatment study.
Gilbert R., Martin RM., Fraser WD., Lewis S., Donovan J., Hamdy F., Neal DE., Lane JA., Metcalfe C.
PURPOSE: Circulating 25-hydroxyvitamin D (25(OH)D) may protect against prostate and other cancers. Few epidemiology studies have measured 25(OH)D on all participants, weakening the evidence-base through reduced statistical power and the potential for bias. We developed a score to predict individual 25(OH)D based on potential predictors, including sun exposure, nutrient intake, and vitamin D pathway genes, providing a method of substituting missing values. We assessed the usefulness of predicted 25(OH)D by comparison with multiple imputation of 25(OH)D levels. METHODS: Amongst 1,091 controls from a population-based case-control study (ProtecT), we quantified relationships of sun exposure, demographic, clinical, anthropologic, nutrient, and genetic data with circulating 25(OH)D and constructed several prediction scores from subsets of these measures. We investigated associations of three prostate cancer risk factors (PSA level, BMI, family history of prostate cancer) with 25(OH)D levels in sensitivity analyses based upon participants with measured 25(OH)D only and based upon the addition of all participants with missing 25(OH)D levels substituted by prediction score values or by multiple imputation. RESULTS: Our score accounted for 27.7% of the variation in measured 25(OH)D. Associations with risk factors of prostate cancer were consistent across the different estimates of 25(OH)D. However, standard deviations for the prediction score did not incorporate extra error from prediction. Multiple imputation of missing 25(OH)D values predicted a more realistic range of 25(OH)D. CONCLUSION: In epidemiological studies of cancer risk associated with vitamin D, multiple imputation of missing 25(OH)D is preferable to prediction scores, as a wider range of 25(OH)D levels are imputed and appropriate confidence intervals calculated.