Skip to main content
 

Joint mixed-effects models to address unmeasured confounding and selective attrition in vitamin D epidemiology research

Thursday, September 13, 2018 — Poster Session III

12:00 p.m. – 1:30 p.m.
FAES Terrace
NIA
EPIG-6

Authors

  • MD Shardell
  • L Ferrucci

Abstract

The body produces 25-hydroxyvitamin D [25(OH)D], a vitamin D biomarker, in response to sunlight, supplements, and dietary intake. These behaviors may influence health outcomes through multiple mechanisms; therefore, statistical adjustment is needed. However, information on these behaviors is not always collected or is measured with error in epidemiologic studies. As a result, vitamin D epidemiology research is vulnerable to unmeasured confounding. Furthermore, when interest is in longitudinal health outcomes, results are additionally vulnerable to selective attrition due to dropout or death. To tackle this problem, we propose a joint mixed-effects model for the study outcome, 25(OH)D, and missingness processes. This is a shared parameter model with random effects shared between the three processes. We present and illustrate this approach by examining the impact of 25(OH)D, measured by radioimmunoassay, on depressive symptoms, measured by Center for Epidemiology Studies Depression Scale (CES-D; range 0 to 60, higher indicates more severe symptoms), in participants aged ≥ 65 years enrolled in the Aging in Chianti study (n=1,203). Using data from three follow-up visits (occurring every three years), we found that participants with 25(OH)D ≥ 20 ng/ml had CES-D scores that were 1.80 points lower than scores for participants with 25(OH)D < 20 ng/ml (95% confidence interval 0.18 to 3.42). We also present results from a Monte Carlo simulation experiment showing that the proposed method performs well with respect to bias relative to conventional linear mixed-effects models. Lastly, we present SAS PROC NLMIXED software code to enhance the accessibility of the method for applied researchers.

Category: Epidemiology