Katie O'Brien

MSPH, PHD

National Institute of Environmental Health Sciences

post-doctoral fellow

PLCO (Learn more about this study)

PLCO-219

Jul 20, 2016

Temporal changes in melatonin and vitamin D in relation to health outcomes

Errors in assessing exposures can often be subject to influences from systematic exposure measurement modifiers (SEMMs), with influences that can be modeled. For example, the correspondence between the long-term average exposure and the measured exposure could vary by time of specimen collection (time of day, time of menstrual cycle, time in pregnancy, time of year), batch of assay, time specimen spent in storage prior to assay, diluteness of urine, or a combination of such factors. A consequence of measurement error in general is that estimates of the exposure-outcome association will be subject to bias.

In this methods paper, we will show that bias due to SEMMs can be mitigated by statistically adjusting for them, even when they would not seem to be of concern based on causal diagrams. We will compare methods using a simulation study and then apply the methods to real-life examples from the Prostate, Lung, Colorectal and Ovarian Cancer Screening Trial (PLCO). We are requesting data specifically for 25(OH)D, a vitamin D biomarker that varies seasonally, and for melatonin, a biomarker that varies based on time of day. We will look at the association between these biomarkers and obesity, type II diabetes, another prevalent metabolic health outcome assessed in the baseline questionnaire, or incident cancer (likely prostate or breast).

1. We will use simulation studies to compare the accuracy and precision of methods for controlling for seasonal variation in vitamin D levels when examining the association between vitamin D and a health outcome. Methods will include no adjustment, covariate adjustment using monthly indicators, harmonic models, and a 2-stage local regression model (LOESS). Performance will be assessed based on percent bias, mean squared error, confidence interval coverage and Akaike Information Criteria (AIC). We hypothesize that the LOESS model will outperform the other approaches.

2. We will apply the above approaches (no adjustment, covariate adjustment using indicators, harmonic models, and LOESS) to an applied example from PLCO. We hypothesize that the method used to control for the SEMM (either vitamin D or melatonin), will influence the magnitude and variance of the effect estimate for the outcome of interest. Based on the results of our simulation studies, we will make recommendations of when and how to control for possible SEMMs.