Statistical A nalysis

We used generalized linear mixed effects models to evaluate the relationship of DNAm-age with 1-year PM_2.5 and 1-year BC exposure levels, singularly and in two-particle models. To account for within participant correlation between the repeated measurements, the mixed effects models included a random intercept for each participant. DNAm-age, 1-year PM_2.5, and 1-year BC were all considered as continuous variables in all analyses.

The aforementioned models were adjusted for known confounders and covariates with a priori biological/clinical relevance using a tiered approach. Given that results from previous DNA methylation studies have been confounded by blood cell heterogeneity, we obtained cell type estimates for six blood cell types (i.e. plasma, CD4T, CD8T, NK, monocytes and granulocytes) using Houseman and Horvath methods [20, 65]. We first constructed chronological age and blood cell type adjusted mixed effects models for the relationships of PM_2.5 and BC with DNAm-age (Model 1). Next, we built models (Model 2) accounting for environmental/lifestyle factors by adjusting for average 1-year temperature (continuous), cumulative cigarette pack years (continuous), smoking status (current, former, or never), and season of visit (Spring [March–May], Summer [June–August], Fall [September–November], and Winter [December–February]), body mass index (BMI) (lean [<25], overweight [25–30], obese [>30]), alcohol intake (yes/no ≥ 2 drinks daily), and maximum years of education (continuous) in addition to the Model 1 covariates. We constructed a third (Model 3) and fourth set of models (Model 4) which accounted for aging-related diseases and disease-related medications respectively. Model 3 adjusted for cancer (yes/no history of lifetime cancer diagnosis), coronary heart disease (yes/no based on electrocardiogram, validated medical records, or physical exam), diabetes (physician diagnosis or a fasting blood glucose > 126 mg/dl), and hypertension (yes/no antihypertensive medication use or systolic blood pressure ≥140 mmHg or diastolic blood pressure ≥90 mmHg) in addition to the Model 2 covariates. Model 4 adjusted for subjects taking statins and/or any diabetes and hypertension medications in addition to the Model 2 covariates. Last, we constructed two-particle mixed effects models with both PM_2.5 and BC as predictors of DNAm-age using the covariates from the Model 1–4 framework.

To exclude sensitivity of our models to outliers, we repeated all analyses using robust regression. By iteratively reweighting data points such that points far from model predictions in the previous iteration are given smaller weights, robust regression is able to minimize the sensitivity of a model to outlying values. Iterations continue until the values of coefficient estimates meet a specified tolerance and weighted least squares regression is then used to compute model coefficients. We performed a set of additional sensitivity analyses: (i) we added a random intercept for 450k plate to account for potential batch effects, (ii) we explored our particle DNAm-age associations in participants with only one NAS visit to see how our results compared with the primary analysis on the full study sample and (iii) we stratified our study sample by season of NAS visit to further explore the contribution of season to the relationship between particle exposures and DNAm-age. We also looked at the Pearson correlation between change in particle exposure and change in DNAm-age between study visits using participants with at least two NAS visits.

Additionally, we evaluated the relationships of DNA methylation values at each of the 353 DNAm-age CpG probes with 1-year PM_2.5 and 1-year BC exposure levels using the aforementioned Model 2 covariates and technical covariates (450k plate, chip, row, and column). FDR correction was performed to account for multiple hypotheses testing for all CpG methylation analyses. Gene ontology analyses were performed on significant CpG results using the publically available DAVID bioinformatics platform [66, 67].

As a means of comparison with the DNAm-age results, we explored the relationships of a standard marker of aging, TL, with PM_2.5 and BC exposure levels. We constructed mixed effects multivariable linear regression models adjusting for chronological age, blood cell type, average 1-year temperature, cumulative cigarette pack years, smoking status, season of visit, telomere batch (categorical with four batches), BMI, alcohol intake, and maximum years of education. Similar to our DNAm-age analyses, we constructed two additional sets of models adjusting for age-related diseases and disease-related medications respectively. There was one relative TL observation of 12.7, while the remaining 856 TL observations were < 4. We kept the outlying observation in the TL mixed effects models, but re-ran the models using robust regression and without the outlying value as sensitivity analyses.

We performed all statistical analyses using R Version 3.1.1 (R Core Team, Vienna, Austria) and considered a P-value < 0.05 to be statistically significant.