Fatty Acids In The De Novo Lipogenesis Pathway And Incidence Of ...

Statistical analysis in individual studies

Individual-participant data analyses were prespecified and documented in the protocol, with the primary exposure variables being 16:0, 16:1n7, 18:0, and 18:1n9. We examined Pearson correlation coefficients between these fatty acids within each lipid fraction. To assess associations of interest, Cox proportional hazard regression was modeled to time-to-event data, with sampling weights applied in EPIC-InterAct with a case-cohort design [17]. Each cohort calculated follow-up time from time of fatty acid measurement to either date of incident T2D, death from any cause, or loss to follow-up, or censoring at end of follow-up, whichever available and occurred first. In the 2 cohorts (AOC and MCCS) without individuals’ person-time data [14,22], logistic regression was used as the most efficient approach to obtaining estimates of interest from the 2 cohorts. The fatty acid variables were evaluated as a continuous linear variable in a unit of the study-specific interquintile range (the difference between the midpoints of the top and bottom quintiles) and, in a separate model, as categorical indicator variables (quintile categories, with the lowest quintile as the reference). We used an interquintile range and quintile categories in continuous and categorical approaches, respectively, because two approaches allowed estimation of the associations over the same exposure range and improvement of comparability between the two approaches.

Covariates for statistical adjustment were prespecified, including their categorization (e.g., continuous, quintiles, etc.). Each participating study prespecified the use of some study-specific covariates (e.g., the number of categories for education status), depending on availability. The primary model included field site, age, sex, race/ethnicity, occupation, education, smoking status, physical activity, alcohol consumption, prevalent hypertension (self-reported or treated), prevalent dyslipidemia (self-reported or treated), prevalent heart disease, and self-reported health status. The second model further adjusted for adiposity measures (body mass index [BMI] and waist circumference). For the mechanistic investigation, the third model further adjusted for circulating 16:0 (for analysis of 16:1n7, 18:0, and 18:1n9) and triglycerides to assess whether associations of 16:1n7, 18:0, and 18:1n9 with incident T2D would be independent of 16:0 and triglycerides and for analysis of 16:0, of triglycerides.

We assessed study-specific measures of interaction by age, sex, BMI, and race/ethnicity using the second model that adjusted for potential confounders including the adiposity measures. Each fatty acid, these prespecified potential effect modifiers, and their relevant cross-product terms and variance-covariance measures were analyzed to evaluate the potential interaction within each cohort.