We also checked whether there was an effect modification according to physical activity level (sedentary versus active) and BMI category at baseline (BMI <25 kg.m−2 and BMI ≥25 kg.m−2), by modelling interaction terms between physical activity/BMI category and dietary factors, and conducting stratified analyses.
We tested the proportional hazards assumption by modelling interaction terms of time and categories of alcohol intake and found no statistically significant interactions.
The proportional hazards assumption was verified by modelling interaction terms of age and our main exposures as well as other fixed covariates.
We evaluated the differences in the associations between miRNA and MN frequency in workers with different drinking status, smoking status, or age group by modeling interaction terms of (miRNA * stratum variables) in Poisson regression models.
Effect modification by sex, baseline BMI category (BMI <25, 25 to <30, and ≥30 kg/m), and smoking status (never, current, former smokers) was assessed by modeling interaction terms, in model 4, between these variables and total flavonoid intake, and conducting stratified analyses.
We evaluated effect modification by modeling interaction terms between each exposure (modeled as a continuous variable) and age (≤ 65 years, > 65 years), sex (males, females), ETS, CHD, intake of statins (yes, no), area (north, center, or south), city of residence (Mülheim, Essen, or Bochum), and wind direction (east vs. west, or north vs. south).
Effect modifications by sex, age-group (<50, 50 59, and ≥60 years), baseline BMI category (BMI <25, 25 to <30, and ≥30 kg/m), smoking status (former smokers, current smokers, and never-smokers), and history of diabetes in a first-degree relative were assessed by modeling interaction terms between these variables and rMED and conducting stratified analyses.
Because phthalate effects may be sex specific (Boberg et al. 2008; Feige et al. 2010; Hao et al. 2012), we evaluated effect heterogeneity by introducing in the models interaction terms between the exposure variable and sex and by stratifying models according to sex.
We furthermore repeated all analysis by sequentially including in models interaction terms for deprivation and gender, deprivation and age group, and deprivation and diagnosis period.
The proportional hazards assumption was tested by modeling multiplicative interaction terms between time and the primary exposure variable.
As a test of robustness, we estimated associations according to race/ethnicity by modeling product interaction terms for the exposure and potential modifier, in addition to lower-order terms and covariates, in whole-sample regression models controlling for all covariates.
