Are Particulate Matter Exposures Associated with Risk of Type 2 Diabetes?

Background Although studies have found that diabetes mellitus (DM) modifies the impact of exposures from air pollution on cardiovascular outcomes, information is limited regarding DM as an air pollution-associated outcome. Objectives Using two prospective cohorts, the Nurses’ Health Study (NHS) and the Health Professionals Follow-Up Study (HPFS), we investigated the relationship of incident type 2 DM with exposures to particulate matter (PM) <2.5 μm (PM2.5), PM <10 μm (PM10), and PM between 2.5 and 10 μm in aerodynamic diameter (PM10–2.5) in the previous 12 months and the distance to roadways. Methods Cases were reported and confirmed through biennial and supplemental questionnaires of diagnosis and treatment information. During follow-up from 1989 to 2002, questionnaires provided information on time-varying covariates and updated addresses. Addresses were geocoded and used to assign air pollution exposures from spatiotemporal statistical models. Results Among participants living in metropolitan areas of the northeastern and midwestern United States, there were 3,784 incident cases of DM in the NHS, and 688 cases in the HPFS. Pooled results from random effects meta-analysis of cohort-specific models adjusted for body mass index and other known risk factors produced hazard ratios (HRs) for incident DM with interquartile range (IQR) increases in average PM during the 12 months before diagnosis of 1.03 [95% confidence interval (CI), 0.96–1.10] for PM2.5, 1.04 (95% CI, 0.99–1.09) for PM10, and 1.04 (95% CI, 0.99–1.09) for PM10–2.5. Among women, the fully adjusted HR for living < 50 m versus ≥ 200 m from a roadway was 1.14 (95% CI, 1.03–1.27). Conclusions Overall, results did not provide strong evidence of an association between exposure to PM in the previous 12 months and incident DM; however, an association with distance to road (a proxy marker of exposure to traffic-related pollution) was shown among women.


Research
The association between cardiovascular out comes and exposure to ambient levels of air pollutants is well supported in the litera ture (Dockery et al. 1993;Laden et al. 2006;Miller et al. 2007;Pope et al. 2002;Puett et al. 2008Puett et al. , 2009. Studies have also reported that individuals with diabetes mel litus (DM) have increased susceptibility for adverse cardiovascular outcomes related to acute increases in exposures to air pollution (Bateson and Schwartz 2004;Dubowsky et al. 2006;O'Neill et al. 2005;Peel et al. 2007;Zanobetti and Schwartz 2001). What is not known is whether DM is in itself an adverse outcome of air pollution.
There are few studies exploring a potential causal role of air pollution in DM develop ment. Brook et al. (2008) studied the relation ship between DM and exposures to traffic pollution among more than 7,600 men and women in two Canadian cities using nitrogen dioxide (NO 2 ) measurements taken from 2002 to 2004. Metaanalytic models showed a statis tically significant increase in the odds of DM among women with each increase in 1ppb exposure to NO 2 . No association was found among male study participants. A case-control study reported that prediagnosis average par ticulate matter (PM) < 10 µm in aerodynamic diameter (PM 10 ) exposure was significantly higher for children diagnosed with DM com pared with controls (Hathout et al. 2002). Finally, diabetesrelated mortality has been associated with PM and other ambient air pol lutant exposures (Kan et al. 2004;Maynard et al. 2007;Ostro et al. 2006). However, based on studies (Bateson and Schwartz 2004;Dubowsky et al. 2006;O'Neill et al. 2005;Peel et al. 2007;Zanobetti and Schwartz 2001) that showed DM as an effect modifier of air pollution-associated DM outcomes, these findings may reflect susceptibility rather than incidence.
In this study, we used data from two estab lished prospective cohorts, the Nurses' Health Study (NHS) and the Health Professionals Followup Study (HPFS), to examine the role of chronic exposures to PM < 2.5 µm in aerodynamic diameter (PM 2.5 ), PM 10 , PM between 2.5 and 10 µm in aerodynamic diam eter (PM 10-2.5 ), and proximity to roads (as a proxy for trafficrelated pollutant exposures) as risk factors for type 2 DM. Biennial ques tionnaires from each of these cohorts and detailed monthly exposure modeling provided a unique opportunity to control for known individuallevel risk factors and to estimate ambient PM exposures specific to all reported mailing addresses for each participant during the 13year followup period.

Materials and Methods
Study population. The NHS is a prospective cohort study that began in 1976 with 121,700 female registered nurses age 30-55 years who lived in one of 11 states in the United States (California, Texas, Florida, Massachusetts, Pennsylvania, Ohio, New York, New Jersey, Michigan, Connecticut, and Maryland) at enrollment. Biennial questionnaires on risk factors and health outcomes have been mailed to participants from 1976 until the present (NHS 2009). The NHS was approved by the institutional review board (IRB) of the Brigham and Women's Hospital.
The HPFS cohort originated in 1986 with 51,529 male dentists, podiatrists, pharmacists, veterinarians, osteopaths, and optometrists located throughout the United States who were 40-75 years of age at enrollment. In 1986, the participants returned a mailed question naire that collected data on lifestyle character istics and medical history. As with the NHS, HPFS participants also were mailed question naires every 2 years to date (HPFS 2009). The HPFS was approved by the Harvard School of Public Health IRB. For both of these cohorts, baseline IRB approval included returning the completed questionnaires, which constituted implied consent to use the data in ongoing health research. Loss to followup in both of these cohorts is < 10%.
This study was restricted to NHS and HPFS participants living in metropolitan sta tistical areas (MSAs) in 13 U.S. contiguous northeastern and midwestern states (Maine, Vermont, New Hampshire, Massachusetts, Rhode Island, Connecticut, New York, New Jersey, Delaware, Pennsylvania, Ohio, Michigan, and Maryland) from 1989 through 2002 to facilitate comparisons with previous studies (Eftim et al. 2008;Pope et al. 1995Pope et al. , 2002 and because air pollution monitors used to estimate exposures that are sparsely distributed outside the MSAs. In addition, we excluded persons with a history of dia betes at baseline. The final study population comprised 74,412 women and 15,048 men, with the vast majority of exclusions from the original cohort population attributable to addresses outside the Northeast and the Midwest at baseline, rather than addresses outside an MSA. Nurses and health pro fessionals were excluded for any period of followup during which they lived outside this region, rather than being censored at the time they moved outside the region. Outcome assessment. Study participants who reported a diagnosis of DM on a biennial questionnaire were sent an additional ques tionnaire to ascertain the month and year of diagnosis and information about diagnostic tests and treatment. To be considered a con firmed case of DM, at least one of the follow ing National Diabetes Data Group criteria (National Diabetes Data Group 1979) had to be met: elevated plasma glucose concen trations on at least two different occasions (as defined below), one or more DM symp toms (e.g., weight loss, thirst, polyuria) and a single elevated plasma glucose concentra tion, or treatment with hypoglycemic medica tion. An elevated plasma glucose concentration was defined as a fasting plasma glucose > 140 mg/dL for cases diagnosed before or during 1997 or > 126 mg/dL for cases diagnosed after 1997 (Expert Committee on the Diagnosis and Classification of Diabetes Mellitus 1997), a random plasma glucose concentration > 200 mg/dL, or a plasma glucose concentration > 200 mg/dL after > 2 hr of oral glucose toler ance testing. Comparisons between medical records and selfreported DM for subsamples of men from the HPFS cohort and women from the NHS resulted in confirmation of 97% and 98% of cases, respectively (Hu et al. 2001a;Manson et al. 1991).
Exposure assessment. To ascertain the exposure of each participant to air pollution at each geocoded address where question naires were mailed, we developed separate spatiotemporal models to estimate monthly PM 2.5 and PM 10-2.5 exposures. Mailing addresses were residential for the women; however, some work addresses were included as mailing addresses for the men. These mod els and their previous use in assessing chronic PM exposures among the NHS cohort are described in detail elsewhere Puett et al. 2008Puett et al. , 2009Yanosky et al. 2008Yanosky et al. , 2009 (Spengler et al. 1996;Suh et al. 1997). The model also included geographic information system (GIS)-derived covariates such as mete orology, land use and elevation, population density, road network, and point source emis sion data. Land use data were from the U.S. Geological Survey 1992 National Land Cover Data Set . The process of estimating PM 2.5 exposures was similar to the process used to estimate PM 10 exposures but involved developing separate models for PM 2.5, before and after 1999, because U.S. EPA AQS monitoring data for PM 2.5 were unavailable before 1999. The pre1999 PM 2.5 model used a simpler spatiotemporal structure to estimate the PM 2.5 to PM 10 ratio seasonally and included estimated extinction coefficients from airport visibility data. Finally, PM 10-2.5 exposures were estimated, for each month and location, by subtracting each of the modeled PM 2.5 estimates from each of the modeled PM 10 estimates. Monitor data were included from 922 PM 10 sites and 498 PM 2.5 sites ). The PM 10 model and post1999 and pre1999 PM 2.5 models were evaluated using a crossvalidation approach, where a subselection of monitors were held out to compare predicted values with observed values Yanosky et al. 2008Yanosky et al. , 2009. The models were shown to exhibit little bias and high precision. For the PM 10 model, crossvalidation (R 2 = 0.62) showed the model performed substantially better than other approaches (e.g., inverse distance weighting crossvalidation R 2 = 0.29) . Crossvalidation for the PM 2.5 model was R 2 = 0.77 and 0.69 for post 1999 and pre1999 PM 2.5 models, respec tively ). The predicted PM 10-2.5 levels showed little bias but were less precise compared with PM 2.5 ). Yanosky et al. (2009) noted that the PM 2.5 levels in the study area were more spa tially homogenous than were the PM 10 and PM 10-2.5 levels.
Distance from each biennially updated address to the nearest road was used as a proxy for trafficrelated air pollution expo sure. Specifically, the distance (in meters) from each address to the closest U.S. cen sus feature class code A1 (roads with limited access, typically interstates), A2 (major, non interstate roads), or A3 (secondary roads, typi cally with more than two lanes) road segment was determined using a GIS (ArcGIS, ver sion 9.2; ESRI, Redlands, CA) and 2000 U.S. Census Topologically Integrated Geographic Encoding and Referencing system (TIGER) files (U.S. Census 2000). Based on informa tion from previous studies and the exposure distributions in these cohorts, distance to the closest road was categorized as 0-49 m, 50-99 m, 100-199 m, or ≥ 200 m (Adar and Kaufman 2007;Hart et al. 2009;Lipfert and Wyzga 2008;Zhu et al. 2002).
Covariates. To assess potential confound ing and effect modification, timevarying data from the biennial questionnaires were used for the following covariates: hyper tension (yes or no regarding a diagnosis from a health professional), smoking status (never, former, or current), hypercholesterolemia (yes or no regarding a diagnosis from a health profes sional), alcohol consumption (0, 0.1-4.9, 5.0-14.9, ≥ 15 g/day), and smoking pack years. Having a lowrisk diet (yes, no) was assessed by semiquantitative food frequency questionnaires administered every 4 years and defined as a diet with a high ratio of polyun saturated to saturated fat that also was high in cereal fiber and low in trans fat and glycemic load (details described elsewhere) (Hu et al. 2001b). Baseline body mass index (BMI; < 25.0, < 30.0, or ≥ 30.0 kg/m 2 ) and physical activity [< 3, 3 to < 9, 9 to < 18, 18 to < 27, or ≥ 27 metabolic equivalent (MET) hr/week] were also included in models. We modeled baseline rather than timevarying values of BMI and physical activity to avoid adjust ing for factors that might be a consequence of the outcome, given that insulin resistance and DM could lead to reduced physical activ ity and thus increased BMI. For selection of covariates in final fully adjusted models, we evaluated changes (10% difference) in model estimates, but also opted to include some variables that might not have produced such a change in estimates but that are known risk factors for diabetes based on prior knowledge. Although some of the covariates may not have been identified previously as predictors of the exposure, it can be useful to adjust for these factors, as they reduce unexplained variation in the outcome and therefore increase power to detect the exposure. volume 119 | number 3 | March 2011 • Environmental Health Perspectives Statistical analysis. Timevarying Cox proportional hazards models were used to assess the relationship of DM with predicted PM 2.5, PM 10 , and PM 10-2.5 exposures in the 12 months before diagnosis and traffic related exposures, with each cohort analyzed individually. All survival models were based on a monthly time scale and were used to estimate hazard ratios (HRs) and 95% con fidence intervals (CIs). We assessed PM expo sure averaged over the 12 months before DM diagnosis because longer windows of exposure were highly correlated, and previous research has shown that time period to be the most relevant exposure for air pollutionrelated deaths . Personmonths of followup time were calculated from base line (30 June 1989 for NHS and 30 January 1989 for HPFS) until DM diagnosis, the end of followup (30 June 2002 for NHS and 30 January 2002 for HPFS), censoring (mov ing outside the geographic region of interest or loss to followup), or death. Separate mod els assessed each particulate fraction alone, and a combined model examined PM 2.5 and PM 10-2.5 simultaneously. Cox models were stratified by age in months and adjusted for year (linear term) and state of residence. Because our PM analyses focused on expo sures averaged over 12 months, we also strati fied by season to adjust for seasonal changes in PM. Persontime spent living outside the geo graphic region of interest was excluded, rather than censoring the participant at the time of the move, as were nurses and health profes sionals with DM reported before baseline. Confounders were included in multivariable models individually, and HR modification was evaluated using pvalues (< 0.05) from multiplicative interaction terms. In sensitiv ity analyses, we conducted analyses excluding women and men who reported myocardial infarctions (MIs) and cancer (except non melanoma skin cancer) prior to baseline. In addition, we performed sensitivity analyses restricted to symptomatic cases, defined as cases who had at least one of the following symptoms at diagnosis: ketoacidosis, unusual urinary frequency, coma, unusual hunger, unintended weight loss, unusual thirst, and visual changes for health professionals and nurses and, additionally, pruritis of the vulva/ vagina for nurses. To explore potential changes in the relationship of DM to air pollution exposures over time, we additionally estimated associations between DM and air pollution exposures during the first 2 years of followup and during the previous 2 years of followup, as well as for the average exposure across the study period. Random effects models were used to conduct a pooled metaanalysis to increase precision of the risk estimates and CIs, and heterogeneity was evaluated with the Q test (DerSimonian and Laird 1986). We performed all the statistical analyses using SAS (version 9.1; SAS Institute Inc., Cary, NC).

Results
There were 3,784 incident cases of DM among 74,412 eligible participants in the NHS (448 per 100,000 personyears) and 688 cases among 15,048 eligible participants in the HPFS (402 per 100,000 personyears) ( Table  1). At baseline in 1989, the mean age was approximately 57 years for HPFS and 55 years for NHS participants. Most were never (NHS: 43%, HPFS: 45%) or former smokers (NHS: 36%, HPFS 46%). NHS participants were more likely than HPFS participants to have baseline BMI < 25 or > 30, although the preva lence of hypertension at baseline was similar for both groups. Men in the HPFS were more physically active and consumed more alcohol than did women in the NHS. The means and SDs of baseline particulate exposures for the HPFS and NHS were similar for PM 2.5 [18.3 (3.1) and 17.5 (2.7) µg/m 3 ], PM 10 [28.5 (5.5) and 26.9 (4.8) µg/m 3 ], and PM 10-2.5 [10.3 (3.3) and 9.4 (2.9) µg/m 3 ]. Baseline addresses for HPFS participants were more likely than NHS addresses to be 0-49 m from the nearest road (22.3% vs. 9.8%) and less likely to be ≥ 200 m from the nearest road (63.7% vs. 77.5%).
Among the NHS, an interquartile range (IQR) increase of 4 µg/m 3 in estimated PM 10-2.5 averaged over the 12 months prior to diagnosis was associated with inci dent DM (HR = 1.07; 95% CI, 1.01-1.13) based on a single pollutant model stratified by age and adjusted for state of residence, year, and season (Table 2), but the HR was attenuated after additional adjustment for cigarette smoking, hypertension, BMI, alco hol intake, physical activity and diet (HR = 1.04; 95% CI, 0.98-1.10). Associations with IQR increases in PM 2.5 (IQR: 4 µg/m 3 ) and PM 10 (IQR: 6 µg/m 3 ) were similar to those estimated for PM 10-2.5 in fully adjusted mod els (PM 2.5 HR = 1.02; 95% CI, 0.94-1.09; PM 10 HR = 1.03; 95% CI, 0.98-1.09), and associations with PM 2.5 and PM 10-2.5 were similar when both were included in the same model. Estimates did not change appreciably when timevarying rather than baseline BMI was included or when family history of diabe tes, census tract median household value, or census tract median household income were added to models (data not shown). NHS par ticipants living < 50 m from the nearest road were more likely to be diagnosed with DM than those living ≥ 200 m away (fully adjusted HR = 1.14; 95% CI, 1.03-1.27) ( Table 3). A nonsignificant (p > 0.05) association was also evident for women with residences located 50-100 m from the nearest road.
Pooled metaanalysis models adjusted for age, year, season, and state suggested a slight increase in diabetes incidence associated with an IQR increase in estimated PM 10-2.5 and PM 10 averaged over the 12 months prior to diagnosis (for both exposures: HR = 1.06; 95% CI, 1.01-1.12), although associations were attenuated after full adjustment (for both exposures: HR = 1.04; 95% CI, 0.99-1.09) ( Table 2). Associations with PM 2.5 were some what weaker. Tests of heterogeneity between the NHS and HPFS were not significant for any of the pooled metaanalysis models, but a multiplicative interaction term between sex and PM 2.5 was statistically significant in a fully adjusted model (p = 0.04) (data not shown). The pooled metaanalysis of DM incidence and proximity to roadways indicated a statisti cally significant association with an address 0-49 m versus ≥ 200 m from the nearest road (HR = 1.11; 95% CI, 1.01-1.23) ( Table 3). Estimates from models adjusted for age, year, season, and state of residence, plus indi vidual covariates were generally similar to esti mates from fully adjusted models (data not shown). However, relative to basic models, adjusting for baseline BMI slightly attenu ated associations with all size fractions of PM in the NHS (e.g., HR for PM 2.5 in basic model adjusting for BMI = 1.02; 95% CI, 0.94-1.09) but slightly increased the HR for PM 2.5 in the HPFS (HR = 1.08; 95% CI, 0.93-1.25). No evidence of effect modifica tion was found for any of the covariates exam ined (data not shown).
Estimates from models that excluded par ticipants with a baseline history of MI (leaving 3,672 NHS and 604 HPFS cases) or prior cancer (except nonmelanoma skin cancer) (leaving 3,311NHS and 522 HPFS cases) were similar to the reported estimates (data not shown). Compared with estimates from our original HPFS study population, analyses restricted to symptomatic DM cases (366) suggested a slightly lower, although still null, risk of incident DM with an IQR increase in PM exposures (fully adjusted PM 2.5 HR = 0.96; 95% CI, 0.78-1.17; PM 10 HR = 0.97; 95% CI, 0.82-1.15; PM 10-2.5 HR = 0.98; 95% CI, 0.85-1.14). Results for symptomatic NHS participants (3,379) were comparable with those for all cases (data not shown).
In sensitivity analyses restricted to the first 2 years of followup, relative risks (RRs) for NHS women with IQR increases in all three particle fraction exposures were lower than for the full time period of followup, but higher when analyses were restricted to the final 2 years of followup (Table 4). Among men in the HPFS, associations were stronger with PM 2.5 during the first 2 years and with all size fractions in the last 2 years than for the full period of followup, whereas the asso ciation with PM 10-2.5 during the first 2 years was weaker.
When associations were estimated with IQR increases in PM exposures averaged over the entire followup period (instead of during the 12 months prior to diagnosis), there was no evidence of associations between incident DM and any PM size fraction among NHS women, and associations with PM 2.5 and PM 10 were attenuated for HPFS men (data not shown). In pooled metaanalyses, HRs for incident DM associated with average PM exposures over the entire followup period were 0.99 (95% CI, 0.89-1.10) for PM 2.5 ; 0.98 (95% CI, 0.91-1.06) for PM 10 ; and 1.02 (95% CI, 0.93-1.12) for PM 10-2.5 .

Discussion
Overall findings for these two cohorts of nurses and health professionals living in the Northeast and Midwest indicate weak non statistically significant increased risks of inci dent DM associated with IQR increases in predicted PM 2.5 , PM 10 , and PM 10-2.5 aver aged over the 12 months prior to diagnosis. RR estimates were very similar for each size PM fraction, with overlapping CIs. Among NHS participants, the strongest fully adjusted association was with PM 10-2.5 , whereas    1989-1991. d 2000-2002. volume 119 | number 3 | March 2011 • Environmental Health Perspectives the weakest was with PM 2.5. This pattern was reversed in the HPFS, with the weak est association observed for PM 10-2.5 and the strongest with PM 2.5 . Although the addition of a multiplicative interaction term in the fully adjusted model for PM 2.5 was signifi cant, showing a greater risk among men, the pooled metaanalysis did not show a differ ence, possibly because of differences in how each analytic method deals with adjustment by covariates. Multipollutant models, analyses restricted to symptomatic DM, and analyses restricted to the first 2 years of followup were similar. In general, results were weaker for PM exposures averaged over the entire followup period. RRs were stronger for DM during the final 2 years of followup, although CIs were wider because of the reduction in sample size. Finally, fully adjusted estimates showed a sta tistically significant association with DM for NHS women living < 50 m versus > 200 m from the nearest road. However, distance to the nearest road was not associated with DM among the HPFS . The body of research on associations between diabetes and air pollution exposures is currently very limited. Although direct com parisons with our study results are not pos sible because of differences in the pollutants, diabetes type, and age groups studied, other studies have also shown weak evidence of an association between diabetes and PM expo sures and stronger evidence for trafficrelated pollutants (represented by distance from the nearest road in our study). Brook et al. (2008) reported a 4% increase in the adjusted odds (95% CI, 1.00-1.08, p = 0.03) of DM diag nosis with each ppb increase in NO 2 expo sure among 4,182 women in Hamilton and Toronto, Canada, but no association among the 3,452 men in the study [odds ratio (OR) = 0.99; 95% CI, 0.95-1.03]. In a study of 402 children in Southern California, type 1 diabetes was not significantly associated with a 10µg/m 3 increase in average PM 10 from birth to diagnosis (OR = 1.08; 95% CI, 0.87-1.34) or a 10ppb increase in NO 2 (OR = 1.03; 95% CI, 0.71-1.50) but was significantly associated with a 10ppb increase in O 3 (OR = 2.92; 95% CI, 1.86-4.58) and SO 4 (OR = 1.65; 95% CI, 1.20-2.28) (Hathout et al. 2006). In a smaller study of 100 children, designed as two casecontrol studies-one among younger children and another among older children, Hathout et al. (2002) reported a significant association between type 1 diabetes diagnosis and average prediagnosis PM 10 among children under age six years, but not among older children. In both age groups, type 1 diabetes was associated with average birth to diagnosis exposure to O 3 , but not to NO 2 or SO 4 .
Evidence of associations between air pol lution exposures and diabetes mortality is also somewhat inconsistent. In a time series study of 434 diabetes deaths in Shanghai, Kan (2004) found a weak association with a 10µg/m 3 increase in 1day lagged PM 10 (RR = 1.01; 95% CI, 1.00-1.01) and a 10µg/m 3 increase in 1day lagged NO 2 (RR = 1.01; 95% CI, 1.00-1.03). A study of mortality among Montreal, Canada, residents registered with the universal Quebec Health Insurance Plan included 3,677 deaths from diabetes (Goldberg et al. 2006). Some associations were reported between daily diabetes mortal ity and the 3day mean during the warm or cold season for certain air pollutants generated from combustion sources (e.g., sulfate, PM 2.5 , NO 2 ) in individuals who also had cardio vascular disease, cancer, or respiratory disease. No associations were reported for individuals without these conditions. In a casecrossover study of 100,000 deaths from 1995 to 2002, an interquartile increase (0.203 g/m black carbon) in traffic particle exposure the day before death was associated with a 5.7% (95% CI, -1.7 to 13.7) increase in deaths due to DM (Maynard et al. 2007).
Biological mechanisms proposed to explain associations between air pollution and cardiovascular disease, specifically, inflamma tion, oxidative stress, and endothelial dysfunc tion, are also plausible mechanisms linking air pollution with the development or exac erbation of diabetic conditions. However, the majority of literature to date suggests that DM is an effect modifier of the relationship between air pollution exposures and cardio vascular outcomes, as opposed to being a direct consequence of air pollution expo sures. Schwartz (2001, 2002) reported that the percent increase in cardio vascular hospitalizations associated with a 10µg/m 3 increase in mean PM 10 exposure on the day of and day before admission was almost doubled among persons with diabetes compared with the percent increase among those without diabetes. Comparable findings of stronger associations between air pollution exposures and cardiovascularassociated hos pitalizations and emergency department visits among persons with diabetes compared with those without diabetes have been reported by other researchers (Peel et al. 2007;Pereira Filho et al. 2008). Research has suggested that inflammatory mechanisms are involved in strengthening the impact of air pollution among persons with diabetes. Increased levels of inflammatory biomarkers, Creactive pro tein and IL6 were more strongly associated with PM 2.5 exposures among persons with diabetes compared with persons without the disease (Dubowsky et al. 2006). Decreases in flowmediated vascular reactivity and nitro glycerinmediated reactivity were associated with increases in 6day moving average black carbon and PM 2.5 exposures, respectively, among a population of diabetics, whereas similar decreases were not evident among those at risk for diabetes (O'Neill et al. 2005).
Although this study helps address the gap of information regarding direct links between air pollution and diabetes, limitations and strengths must be considered. We included only confirmed and probable DM cases based on National Diabetes Group and American Diabetes Association criteria (National Diabetes Data Group 1979); however some persons with diabetes may have been mis classified because we relied initially on self reported diagnosis rather than on glucose testing for all cohort participants. Although we had the advantage of using data from large prospective cohorts, our power to detect an effect among the HPFS men was somewhat limited by the sample size, particularly in sensitivity analyses restricted to symptom atic DM cases. In addition, the metaanalyses and combined analyses with sexinteraction terms were dominated by the NHS because of a smaller number of HPFS participants. In addition, although results were consistent for the two time periods of average PM exposure, we evaluated (12 months before diagnosis and exposures averaged over the entire followup period), additional time windows of expo sure should be explored, including more acute exposures and exposures during childhood.
As in our previously published studies of the NHS , the GIS based temporal spatial smoothing of our PM predictive models reduced variability relative to measured concentrations. This Berkson measurement error should not cause sub stantial bias toward the null (Gryparis et al. 2009). However, it is possible that differences between study populations with respect to trafficrelated exposures (distance to the near est road) and PM 2.5 associations, may be due in part to differences in the type of address available for each study-residential only in the NHS versus occupational or residential in the HPFS-rather than differences in sus ceptibility between men and women. Thus, error in the predicted exposure estimates for the HPFS would differ from the NHS. For the HPFS, however, both address types are likely to represent PM exposures during a large part of a 24hr period, albeit different parts. Finally, unlike many previous stud ies, the availability of geocoded biennially updated addresses enabled us to adjust for a number of biennially updated individuallevel covariates and to assign predicted particu late exposure levels to each address for each participant throughout the followup period. Therefore, exposure estimates should be more accurate over time than estimates based on baseline addresses only. In addition, missing exposure data were minimized because of the use of a recently developed and highly specific GISbased spatial smoothing model.
Because of the mixed nature of the HPFS addresses, we did not believe it was appropri ate to assess arealevel SES as a confounder in this study. However, estimates from models adjusted for neighborhoodlevel SES, such as median house value and average household income at the censustract level (U.S. Census 2000) were similar to those presented for both the HPFS and NHS. An additional potential limitation of our analysis was that our two populations of health professionals represent a narrow range of SES, possibly limiting the generalizability of our results.

Conclusion
In summary, we did not find strong evidence for an association between exposure to PM 2.5 , PM 10 , or PM 10-2.5 in the 12 months before diagnosis and type 2 DM incidence among female nurses and male health professionals living in the northeastern and midwestern United States. However, our findings did sug gest an association between residential prox imity to roadways and incident DM among female nurses. Although questions remain regarding which pollutants play the most prominent role in incident DM and which subpopulations are most susceptible, our study results add to other findings reported in the current literature suggesting that DM may be an outcome of air pollution exposures in addition to being an effect modifier for air pollution-associated cardiovascular disease.