Vitamin D and Neurodegenerative Disease With Selected Topics Related to Correlated and Missing Outcome Data

Abstract

The following dissertation addresses of two themes: the role of vitamin D and neurodegenerative diseases and methodologic concepts related to correlated missing outcome data. In the first chapter, the relation between vitamin D (characterized by circulating levels of 25Zhydroxyvitamin D 25[OH]D) and multiple sclerosis (MS) activity and progression is assessed in a secondary analysis of a population of over 1400 relapsing remitting MS patients with multiple asynchronous assessments of 25(OH)D. Overall, using a combination of clinical and magnetic resonance imaging endpoints, higher levels of vitamin D were associated with lesser MS activity. Results were more equivocal for clinical and brain volume measures. In the second chapter, the relation between dietary vitamin D and risk of amyotrophic lateral sclerosis (ALS) is assessed in pooled analysis of five prospective cohort studies. The study included nearly 1500 cases of ALS that occurred in a population of over 1 million individuals. Using dietary intake derived from food-frequency questionnaires, no association between dietary intake of vitamin D and risk of ALS was observed in any model. Similarly null results were observed for vitamin D intake from food and supplemental vitamin D. The third chapter addresses a methodologic concern when disease outcomes can be classified into multiple subtypes. It was not well understood of how to properly address marker-specific effects of a particular risk when markers are correlated and some are missing. This was addressed using an analysis of breast cancer in the Nurses’ Health Study (NHS) considering 5 markers with varying levels of correlation and missingness. Correlation among outcome measures was addressed through the calculation of an adjusted hazard ratio and four approaches for missing data were evaluated: the complete case, inverse probability weighting, missing indicator and multiple imputation. In the NHS, 4380 cases (with pathology reports) of breast cancer occurred; however, only 1551 cases had information on all five markers. We considered a list of established breast cancer risk factors and calculate adjusted marker-specific effects addressing missing using each of the 4 approaches. Effect estimates were generally similar for each method but corresponding standard errors were smaller using the multiple imputations and missing indicator approaches. Subsequent simulation studies suggest the missing indicator approach to produce the least bias and increases in standard error compared with datasets with complete information on all markers.