We studied participants in the Normative Aging Study (NAS), a longitudinal study established by the VA in 1963 (Bell et al. 1972
). In brief, the NAS enrolled 2,280 men from the Greater Boston area who were initially free of known chronic medical conditions. All participants provided written informed consent, and the study was approved by the institutional review boards of all participating institutions. Participants visited the Boston VA Hospital study center every 3 years to undergo physical examinations. At each of these visits, blood samples and extensive physical examination, laboratory, anthropometric, and questionnaire data were collected. Information about cigarette smoking, medical history, and medication use were obtained by self-administered questionnaire. Each subject was interviewed to confirm the identity and purpose of medications used, and all new disease diagnoses were noted.
Diabetes was defined as a physician diagnosis of diabetes, and obesity was defined as a body mass index (BMI) of at least 30 kg/m2. Self-reported data on diabetes status and statin use were updated at each study visit. In addition, BMI and obesity were updated based on height and weight measurements at each visit. Thus, NAS data reflect changes in disease status and medication use over time.
Measurements of sICAM-1 and sVCAM-1 began in 1999. For the present study, we included the 642 NAS participants with at least one measurement of sICAM-1 and sVCAM-1 and whose home address was in the Greater Boston area (1,423 total person-visits). Subjects who moved out of the area were excluded.
Measurements of sICAM-1 and sVCAM-1
Blood samples routinely collected during medical exam visits from 1999 through 2008 were analyzed for sICAM-1 and sVCAM-1 in N. Rafai’s laboratory at Children’s Hospital Boston (Boston, MA). Plasma sICAM-1 and sVCAM-1 concentrations were measured in duplicate using the enzyme-linked immunosorbent assay (ELISA) method (R&D Systems, Minneapolis, MN), with a sensitivity of 0.35 ng/mL for sICAM-1 and 2.0 ng/mL for sVCAM-1 (Lim et al. 1999
BC exposure prediction
The BC exposure model and the stationary air monitors used to develop the model have been described in detail previously (Gryparis et al. 2007
). Briefly, 82 sites were used; most sites measured BC continuously using aethalometers, and other sites collected particles on a filter over 24 hr and measured elemental carbon (EC) using reflectance analysis. The monitoring data used to develop our model included 6,031 observations from 2,079 unique exposure days.
Using a spatiotemporal model that we developed and validated previously (Gryparis et al. 2007
), we estimated the 24-hr average BC concentration at each geocoded participant address. Predicted daily concentrations showed a > 3-fold range of variation in exposure across measurement sites (adjusted R2
= 0.83). A validation sample at 30 additional monitoring sites showed an average correlation of 0.59 between predicted and observed daily BC levels. We averaged the 24-hr predictions to form estimates for the 4, 8, and 12 weeks before each participant visit. We also averaged the 24-hr predictions to form estimates for the 4-week average during the 5–8 weeks before the study visit and the 4-week average during the 9–12 weeks before the study visit, which are components of the 8- and 12-week averages, to use as a sensitivity analysis.
Covariates in the BC prediction model included measures of land use for each address (cumulative traffic density within 100 m, population density, distance to nearest major roadway, and percent urbanization), geographic information system (GIS) location (latitude, longitude), daily meteorological factors (apparent temperature, wind speed, and height of the planetary boundary layer), and other characteristics (day of week, day of season). The Boston central-site monitor was also included as a predictor to reflect average pollutant concentrations over the entire region on each day.
Separate models were fit for warm and cold seasons. Interaction terms between the temporal meteorological predictors and source-based geographic variables allowed for space–time interactions. Regression splines allowed main effect terms to nonlinearly predict exposure levels, and thin-plate splines modeled the residual spatial variability not explained by the spatial predictors. A latent variable framework was used to integrate BC and EC exposure data, where BC and EC measurements were treated as surrogates of some true, unobservable traffic exposure variable; see Gryparis et al. (2007)
for further details.
Estimation of health effects
We log-transformed sICAM-1 and sVCAM-1 levels to increase normality and stabilize variance in the residuals. Model covariates were selected a priori, and all models included age, BMI, diabetes, smoking status, pack-years, and season. We used mixed models to account for correlation among measurements on the same subject across different medical visits. Mixed models have the form
where Yij is log(sICAM-1) or log(sVCAM-1) in subject i on day j, and ui represents a subject-specific intercept that reflects unexplained heterogeneity in the outcome. BC averages and model covariates are modeled as fixed linear effects, and ui is modeled as a random effect. We assume that the ui are generated from a normal distribution with common variance, yielding the simple compound symmetry variance structure. This model requires estimation of two variance components, which represent between- and within-subject variation. Models with unbalanced data (i.e., varying numbers of repeated measurements on each subject) typically yield accurate estimates of within-subject variation, provided a sufficient number of repeated measurements contribute to the estimate.
Models used to examine effect modification by obesity, diabetes, and statin use included interaction terms that allowed associations between BC and the outcomes to vary among subgroups. Diabetes, obesity, and statin use were treated as time-varying covariates, where the status was updated at each visit to reflect changes since the last visit. The percentages of subjects whose status changed for these factors over the study period were 3% for diabetes, 9% for obesity, and 21% for statin use.
We also performed a sensitivity analysis to investigate whether the interaction with statin use was heavily influenced by the people who began taking statins after the study period began. Specifically, we limited the interaction analysis to the participants who never used statins during the study period (n = 284) and those who used statins throughout the entire study period (n = 223).
Because the outcomes were log-transformed, effect estimates are reported as percent changes in sICAM-1 and sVCAM-1 concentrations associated with a 0.30-μg/m3 increase in BC, which corresponds to the interquartile range (IQR) for average BC exposures over all three time intervals (4, 8, and 12 weeks). A level of α = 0.05 was used to determine statistical significance.