Final Report: Relating Cardiovascular Disease Risk to Ambient Air Pollutants Using Geographic Information Systems Technology and Bayesian Neural Networks: The AHSMOG Study

EPA Grant Number: R830547
Title: Relating Cardiovascular Disease Risk to Ambient Air Pollutants Using Geographic Information Systems Technology and Bayesian Neural Networks: The AHSMOG Study
Investigators: Knutsen, Synnove F. , Beeson, Larry , Ghamsary, Mark , Soret, Samuel
Institution: Loma Linda University
EPA Project Officer: Chung, Serena
Project Period: February 1, 2003 through December 31, 2006 (Extended to January 31, 2009)
Project Amount: $964,436
RFA: Epidemiologic Research on Health Effects of Long-Term Exposure to Ambient Particulate Matter and Other Air Pollutants (2002) RFA Text |  Recipients Lists
Research Category: Health Effects , Particulate Matter , Air

Objective:

Specific Objectives:
1.   To assess the long-term effects of particulate and gaseous pollutants on risk of cardiovascular disease (CVD), including fatal and non-fatal coronary heart disease (CHD), during 23 years follow-up (1977-1999) using the unique data from the Adventist Health and Smog (AHSMOG) Study.
2.   To assess the long-term effects of ambient air pollutants on risk of fatal and non-fatal CHD among sensitive subgroups (e.g., prevalent CVD, hypertensives, diabetics, elderly).
3.   To assess the long-term effects of mixed pollutants on the endpoints in objectives 1 and 2.
4.   To investigate the effect of lag-times on the ambient air pollution-CVD association.
 
Approach:
1.   Utilize data from the existing AHSMOG Study, which has been updated through March 2000 through the current EPA STAR Grant (R827998).  These data include monthly indices of air pollutants to zip code centroids, monthly residence and work location histories, outcome assessment (CHD, fatal and non-fatal) and assessment of relevant confounders (smoking, environmental tobacco smoke, diet, exercise, etc.).
2.   Develop new indices of ambient air pollutants for the individual subjects in the AHSMOG Study using geographic information systems (GIS) technology and stochastic models that include error estimates of the indices.
3.   Develop non-linear statistical models using Bayesian neural networks to develop alternative analytical strategies for modeling the relationship between different ambient air pollutants and risk of CHD where several pollutants and latent (unobserved) and missing values can be incorporated.
4.   Compare new methods developed under approaches 2 and 3 to the classic or conventional methods previously used in the AHSMOG Study.

Summary/Accomplishments (Outputs/Outcomes):

The AHSMOG Study has been able to accomplish most of the objectives of this grant.  However, for various reasons we were not able to assess the association between ambient air pollution and non-fatal CHD. 
 
Monthly air pollution estimates for each subject were available since start of study in 1973 and for some pollutants (PM2.5) back to 1966 and through 2000. These were developed using a deterministic method with interpolation to the centroid of each zip code. The AHSMOG study had used information from all relevant monitoring stations in California to develop their ambient air pollution estimates.  Under this grant, we have developed individual mean estimates of the different air pollutants from 1977-2006 (or date of death) based on monthly residence history and monthly workplace zip code, which have been geocoded.  In collaboration with the research team at Environmental Systems Research Institute (ESRI), we have developed a software program that can combine the geocoded residence and workplace information with the EPA air pollution database to assess subject-specific ambient air pollution estimates using geostatistical data analyses.  Thus, we have subject- and residence-specific ambient air pollutant values for the entire AHSMOG cohort.
 
Unfortunately, we were not able to assess the association between ambient air pollution and incident CHD due to problems in obtaining medical records to verify incident cases of CHD.  However, we have completed analyses for risk assessment of fatal CHD and a paper was published in 2005 (Chen LH, Knutsen SF, Shavlik D, Beeson WL, Petersen F, Ghamsary M, Abbey D. The association between fatal coronary heart disease and ambient particulate air pollution:  are females at greater risk?  Environmental Health Perspectives 2005;113:1723-1729). A significantly elevated risk of fatal CHD was found for each 10 µg/m3 increase in PM2.5.  The effect was mainly present in females and was strengthened when ozone was included in the model.
 
Under this funding, we also have assessed the risk between ambient particulate matter and fatal CHD in sensitive subgroups using three different cohorts: the AHSMOG-1 (with PM10), the AHSMOG-2 (with PM2.5) and a population of renal transplant recipients (with PM2.5). The associations were of similar magnitude in the potentially sensitive subgroups and the total population cohorts. Thus, we cannot conclude that the sensitive subgroups tested are at higher risk than non-smokers in general.  However, our risk estimates tend to be higher than what others have found, possibly due to the fact that our analyses are not confounded by smoking.
 
Under this grant, we also have continued general analyses on the association between ambient air pollution and other disease outcomes (cancer and respiratory disease).  We find a moderate association with overall cancer mortality, lung cancer incidence and mortality and non-cancer respiratory mortality.
 
As part of this grant, we have explored different methods of assessing individual air pollutants, using inverse distance versus kriging.  This has been presented at scientific meetings and the findings also are included in a paper that is soon to be submitted for publication. We have further assessed alternate statistical methods including Bayesian Neural Networks and Bayesian Cox (using BUGS software program). The findings have not been published yet.
 
1. Specific Objective I
 
“To assess the long-term effects of particulate and gaseous pollutants on risk of cardiovascular disease (CVD), including fatal and non-fatal coronary heart disease (CHD), during 23 years follow-up (1977-1999) using the unique data from the AHSMOG Study.”
 
1.A. Assessment of Non-fatal (incident) CHD
 We have not been able to fulfill the objective of assessing risk of non-fatal CHD as associated with ambient air pollution.
 
From 1977-1982, we have information on and verification of incident myocardial infarctions (MI). However, the number of cases was too small for meaningful analyses. For the period 1983-1999, we have self-reported incidence of acute myocardial infarction with additional information on name and address of the hospital in which these were diagnosed.  Validity of this information was planned through obtainment of medical records from the individual hospitals. However, this has proved impossible as hospitals do not keep medical records in-house for more than 5 years. After that they are put in remote storage and if records are older than 10 years, they often are destroyed. Thus, we have not been able to obtain more than 56% of the records or a total of 318 of 568.
 
To compensate for a lack of data for incident CHD, we requested permission from EPA to modify this part of the study and instead assess incident CHD in a new cohort, the Adventist Health Study 2 cohort. This cohort of 97,000 subjects recently has been assembled through funding from NCI to study the effect of lifestyle, especially diet, on cancer outcomes. We proposed to use the information from bi-annual hospitalization forms to study the association between particulate air pollution and incident CHD in the 6 western U.S. states using this population and a nested case-control design. Although we did engage a graduate student to work on this, it turned out to be more time consuming than expected. Therefore, it was decided to request separate funding and assess this in the entire AHS-2 population covering all 50 states of the United States and the 6 provinces of Canada. This will be a separate, stand-alone study once the datafile has been linked to air pollution estimates.
 
1.B. Fatal CHD
This objective has been more than fulfilled in that mortality has been completed through 2006 (instead of 1999 as stated in the objectives). We also have assessed the association between ambient air pollution and fatal CHD in both the AHSMOG-1 and AHSMOG-2 cohorts.
 
1.A.a. Fatal CHD and results from the AHSMOG-1
A total of 373 (6% of total cohort or 8% of total deaths) subjects died from coronary heart disease (CHD) during follow up (1977-2006).The paper on associations between fine particles (PM2.5) and fatal CHD (through 1998) was published in December 2005 (Environment Health Perspectives 2005;113:1723-1729).
 
We found that women seemed to be at substantially higher risk for fatal CHD when living in areas with high PM2.5 levels compared to men.  This also was reflected in the observed associations with PM10 in the geographic areas around airports, but not in other areas. The strong finding for women recently has been replicated in the larger cohort of the Womens Health Initiative study (Miller, et al., New England Journal of Medicine 2007;356(5):447-458). 
 
We have tried, with the help of a graduate student, to assess whether our observed gender difference could be due to a greater degree of misclassification of exposure among males.
from the original AHSMOG study.  However, it has been very difficult to separate the air pollution at residence and work location due to the way the data were set up and thus we are not able to assess this as planned.
 
1.A.b. Fatal CHD and Results from AHSMOG-2 (the AHS-2 cohort of 97,000)
As with the AHSMOG-1 study, we find a consistent association between ambient levels of PM2.5 and fatal CHD.  However, as opposed to the original AHSMOG Study, we also find an association among males.  The magnitude of effect is relatively strong.  After adjusting for age, education, past smoking history (all are non-smokers), race and exercise, the RR was 1.63 (0.92-2.9) in females and 2.19 (1.25-3.85) in males (Table 1.A.a. below). In runs not shown, we controlled additionally for BMI, vegetarian status, alcohol use, red meat consumption, high blood pressure, angina, diabetes, statin use, aspirin use, high blood pressure medicine use, and hormone replacement therapy.  Controlling for these variables did not change the results.An abstract was presented at ISEE 2009 in Ireland.
 
Table 1.A.b. Risk of fatal CHD according to increments of PM2.5 and ozone.  Single- and two-pollutant models.
Model
Pollutant
(increment)
Combined
Males
Females
 
 
N=60,652
N (cases)=171
N=20,829
N (cases)=85
N=39,823
N (cases)=86
Single Pollutant Model
PM2.5
 (10 µg/m3­)
1.9 (1.27-2.84)
2.19 (1.25-3.85)
1.63 (0.92-2.9)
Ozone (10 ppb)
1.26 (0.92-1.73)
1.2 (0.77-1.88)
1.37 (0.88-2.13)
Two Pollutant Model
PM2.5
(10 µg/m3­)
1.9 (1.27-2.84)
2.18 (1.24-3.83)
1.65 (0.93-2.93)
Ozone (10 ppb)
1.27 (0.92-1.75)
1.2 (0.75-1.9)
1.38 (0.88-2.17)
 
2. Specific Objective II
 
To assess the long-term effects of ambient air pollutants on risk of fatal and non-fatal CHD among sensitive subgroups (e.g., prevalent CVD, hypertensives, diabetics, elderly).
 
2. A. Sensitive subgroups in the original AHSMOG cohort
             The following sensitive subgroups were identified:
·         Older age (>64 yrs and > 74 yrs)
·         Past smokers
·         Prevalent CHD or stroke
·         Prevalent diabetes
·         Prevalent chronic obstructive pulmonary disease (COPD)
 
Only non-smokers were enrolled in the AHSMOG study.  However, some of the subjects had been smokers before enrolling in the study and thus our study population was stratified on past smokers versus never smokers. For prevalent diabetes, information was only available at baseline (1977).  Many may since that time have developed diabetes.  Thus, any results could be biased towards the null. 
 
For the groups with COPD or cardiovascular disease, this was assessed at each questionnaire (1977, 1982, 1992 and 2000).  Therefore, we have used all this information and moved subjects into these sensitive subgroups as of the date they report having been diagnosed with the specific condition.  Likewise, the sensitive subgroup of elderly has been put into this group as they aged throughout the study.  Thus our final sensitive subgroups with the different outcomes are as follows:
 
Table 2.A. Number of total natural cause death, fatal CHD and cardio-pulmonary deaths during 30 year follow up of the AHSMOG-1 Study.  1977-2006.  N=6,338.
Subgroup
 
 
     N
All natural cause Mortality
N
Fatal CHD
 
N
Cardio-pulmonary death
N
> 64 yrs
6,050
3,219
856
2,006
Prevalent CHD, stroke
906
679
285
518
Prevalent Diabetes
371*
279
97
187
Prevalent COPD
1,698
777
204
493
Past smokers
1,441
775
203
453
 
The results have been drafted into a paper.  Briefly, the effect of PM10 or gases on risk of CHD in sensitive subgroups seems to be small and of similar magnitude as those found in the entire cohort.  However, among diabetic males, the risk of CHD, cardiopulmonary, and all natural cause mortality seems to be elevated.  We have too few numbers to assess the effect of PM2.5.
 
2.B. Sensitive Subgroups in the AHSMOG-2
The new AHSMOG-2 cohort of 97,000 non-smoking subjects gives much better power to assess the effect of ambient air pollution in sensitive subgroups.  Thus, we have identified the following sensitive subgroups in this new cohort:
 
Older age:  > 80 years
Past smokers
Prevalent diabetes (Type I and II)
Prevalent cardiovascular disease (CHD, hypertension, stroke)
Prevalent CHD (MI)
Prevalent COPD (asthma, bronchitis)
 
The analyses in the AHSMOG-2 are not yet completed as they are part of a doctoral dissertation. Analyses have so far been limited to the 60,652  who have not moved during follow-up.  Preliminary results indicate similar magnitude of effect of PM2.5 among different sensitive subgroups and they are similar to the overall effect in the full cohort. Thus, our preliminary conclusion is that the effect of various ambient air pollutants on fatal CHD in the potentially sensitive subgroups of this population, does not seem to be higher than among the population as a whole.  Further analyses are needed in this new cohort before final conclusions can be made.
 
2.C. Sensitive Subgroup of Renal transplant recipients
To further assess the effect of air pollution in potentially sensitive subgroups, one of our doctoral students, has studied the effect of ambient air pollution on risk of fatal CHD in another sensitive subgroup, e.g., non-smoking renal transplant recipients.  An abstract was accepted for presentation at the ISEE conference in Ireland in August 2009, and a paper is in the process of being submitted for publication.
 
3. Specific Objective III
 
“To assess the long-term effects of mixed pollutants on the endpoints in objectives 1 and 2.”
 
This criterion has been met and reported for each of the results reported under objectives 1 and 2 or in the respective appendices (which are not included with this report). In general, the effect of ozone is not affected to any significant degree by including either particulate or gaseous pollutants in a two-pollutant model one at a time.  However, the effect of PM (especially PM2.5) is clearly strengthened by adjusting for ozone in a two-pollutant model in the AHSMOG-1 and the renal transplant recipients.   However, ozone does not seem to modify the PM effect to any great degree in the AHSMOG-2.  This discrepancy warrants further investigation. The effect of other gaseous pollutants are smaller and do not modify the PM effect to any large extent.
 
4.  Specific Objective IV
 
“To investigate the effect of lag-times on the ambient air pollution-CVD association.”
 
This criterion has been met. In our investigation of lag-times for the air pollution-CVD association, we tested out the effect of long lag-times versus more recent lag times.  The original AHSMOG study only had a maximum of 4 years air pollution prior to a CHD death.  Thus, we were limited in our ability to assess various lag-times.  However, we were able to assess the effect of using the 4 years immediately prior to fatal CHD as the exposure window (with the exclusion of the last month before death to avoid short-term effects) and compare this with the use of a fixed time period immediately before the AHSMOG Study began (1973-1977).  Of these two options, the moving 4 year average immediately prior to CHD death gave the strongest estimates with the most narrow confidence intervals (Table 4.A. below).  This suggests that the effect of particulate air pollution on risk of CHD death is related to more recent ambient air pollution.
 
Table 4.A.  The association between PM2.5 and fatal CHD using different exposure estimates
 
Increment
Fixed time period, 1973-77
4-year moving average
 
 
RR
95% CI
RR
95% CI
Single pollutant model
 
 
 
 
 
PM2.5
10 µg/m3
1.10
0.93-1.31
1.42
1.11-1.81
Ozone*
10 ppb
1.01
0.99-1.04
0.97
0.71-1.32
 
 
 
 
 
 
Two pollutant model
 
 
 
 
 
PM2.5
10 µg/m3
1.08
0.86-1.35
1.99
1.37-2.88
Ozone*
10 ppb
1.01
0.97-1.04
0.54
0.34-0.84
 
5. METHODS
In our proposal, we mentioned three methodological approaches to be used.  We have explored all of these.
 
  1. Develop new indices of ambient air pollutants for the individual subjects in the AHSMOG Study using geographic information systems (GIS) technology and stochastic models that include error estimates of the indices.
  2. Develop non-linear statistical models using Bayesian neural networks to develop alternative analytical strategies for modeling the relationship between different ambient air pollutants and risk of CHD where several pollutants and latent (unobserved) and missing values can be incorporated.
  3. Compare new methods developed under approaches 2 and 3 to the classic or conventional methods previously used in the AHSMOG Study.
 5.A. New Indices
The previous work done on the AHSMOG-1 Study used inverse distance interpolations based on their residence history through 2000. From 2001 and onwards, using ArcGIS,  we developed new inverse distance interpolations based on the 2000 residence through the remainder of the follow-up (through 2006) or to time of censoring.  In order to assure high reproducibility between the old indices and the new ones, we estimated ambient air pollution levels for the year 2000 and compared it with the old estimates. The correlations ranged from r=0.99 to r=0.89. 
 
For the new AHSMOG-2 Study we interpolated using both inverse distance and kriging while varying the parameters of these methods. Using ordinary kriging and setting a lag size of 50 km (with 4 sectors, maximum of 5 neighbors per sector, and minimum of 2) resulted in a correlation with inverse distance with a 100 km radius of r=0.60 (no sectors, maximum neighbors of 3, this was similar to AHSMOG 1 except this study used a 50 km radius).  When comparing this kriging model with an inverse distance model that accepts the defaults (1324 km, maximum 15 neighbors, minimum 10, power=2), the correlation increases to r=0.71.  When comparing the ordinary kriging model that accepts the default parameters to the previous mentioned inverse distance model set at a radius of 100 km, the correlation is r=0.89.  When comparing this default ordinary kriging model to the default inverse distance model, the correlation is r=0.96.
 
The development of GIS-based individual ambient air pollution estimates has been completed for PM10, PM2.5, NO2, SO2 and ozone. We have worked in close collaboration with ESRI and have developed an automated program that can estimate individual and time specific measures based on GIS kriging of the air pollutant and the subject’s residence and work location. One of the co-investigators, Dr. Samuel Soret, is continuing to work on two papers, one cross-validating air pollution estimates using GIS-based kriging versus a deterministic model.  In the second paper, he is estimating the errors introduced, if any, by using zip code centroids versus actual street address of residence to estimate individual measures of air pollution.
 
Using GIS, we also are pursuing two papers related to mobile sources of PM by:
1) relating risk of fatal CHD to time spent on freeways and high traffic roads and
2) relating risk of fatal CHD to distance of residence from freeways/highways.
 
One of our investigators and a doctoral student, David Shavlik, is pursuing further work evaluating spatial interpolation methods on PM10 and assessing and adjusting for bias in RRs using spatial interpolation methods.  He also is working closely with Dr. Soret on papers where he will adjust AHSMOG data for spatial auto correlation forming clusters based on location and pollution values using a sandwich estimator (implemented in SAS) and a random effects Cox model (implemented in S-Plus). 
 
Preliminary analysis indicates that varying the parameters significantly affects the results. Increasing the radius or lag size will increase the effect estimate, decreasing the standard errors of the estimates, but increasing the confidence interval width.  Choosing these parameters wisely is obviously very important to getting unbiased estimates.
 
5.B. Non-linear Statistical Models Using Bayesian Neural Networks
The development of statistical models using neural networks and Bayesian neural networks is completed. The first paper on this work was presented at the Hawaii International Conference on Statistics in Honolulu in June 2004. Three MSPH students have worked on one thesis each:
1) developing a model using survival neural network analysis based on a method developed by Drs. Dipley of University of Oxford, UK;
2) developing a model using Bayesian Neural Network using an approach described by Dr. Rad Neill of the University of Ontario, Canada. 
3) developing a new model for dealing with missing values. 
 
We also have been able to use the BUGS statistical software to do Bayesian Cox analyses.  Dr. Ghamsary has worked closely with the developers of this software in order to make it run with large datasets such as ours. Comparing the risk estimates using BUGS with estimates using traditional Cox analysis, was reassuring as they gave virtually the same estimates.  A paper describing these analyses was submitted to Statistics in Medicine. However, the paper was rejected and we now are pursuing other publication options.
 
A modified paragraph from the thesis of one of our students describes some of the results:
70% of the total sample size in the training dataset (N=3,047), as well as 30% in the validation dataset (N=1,306) were conducted to fit the Cox regression neural network approach. In addition, the independent variables (smoke-pack-years in 1976, years worked with smokers, hours per week spent outdoors, PM10, ozone, nut consumption, water consumption) were standardized for the model. The traditional multivariable Cox regression could detect only 59.2% of the sensitivity for the training dataset and 59.1% for the validation dataset. By increasing the number of neurons to 8, the sensitivity increases to 62.8% in the training dataset and 61.3% in the validation dataset. The overall increase in the specificity was very slight in the training dataset, and the accuracy increased by 1.1% in both the training and the validation datasets. The number of neurons could not be increased beyond 8 in this particular dataset to avoid over fitting.  Although neural networks are very useful in predictive modeling, the method is limited to this application and cannot help with improving estimations of the effect of air pollution.
 
5.C. Compare New Methods to Classic Methods
The answer to the question of whether enhanced positional accuracy of subject locations results in reduced exposure misclassification is not only of scientific interest, but of great importance operationally because street-address geocoding is labor intensive and requires substantial expertise and ancillary databases. 
           
This issue is of particular interest to the AHSMOG Study because this study traditionally has relied on zip codes to model residential and work locations and to assign exposures derived through inverse distance weighting (IDW) interpolations.  Although easy to geocode, zip codes have problems.  They are set up for mail delivery, not to represent health processes. They change over time and they introduce positional error.  In urban areas where zip codes tend to be small and relatively similar in shape, they can be reasonable approximations in representing street-level subject locations.  In rural areas, however, zip codes tend to cover larger areas and vary more in shape and they may not accurately represent the true location of place of residence.
           
We have utilized cross-validation methodology to evaluate the agreement between ozone concentrations produced by using three different GIS-based exposure models and the actually measured values across a set of 159 monitoring stations in California.  The monitors themselves were simulated as study subject locations represented alternatively by their true locations and zip code locations.  We examined several cross-validation statistics, including error metrics and concordance correlation, with respect to proximity to the monitoring network. 
           
The results for the mean bias indicate that regardless of exposure model, street level seems to offer an advantage when observations are analyzed with respect to distance categories.  Subjects residing closer to air monitors will tend to be assigned exposures that overestimate the true ozone concentrations. In contrast, subjects residing further away from the monitoring network will tend to be assigned exposures that notably underestimate true ozone concentrations. To minimize mean bias, one would use the proximity-based (Prox) method for deriving exposure estimates at the street level for subjects residing within 10 miles of the monitoring network.  To minimize the MB corresponding to exposure estimates for subjects residing beyond 10 miles from existing air monitors, one would use IDW and the street level spatial resolution.  Mean prediction error (PE) yielded similar results to mean bias (MB).  Results for root-mean-square prediction error (RMSE) also seem to favor the use of street address data but replacing the Prox method with ordinary kriging (OK) within the A zone.
           
In general, accuracy was similarly high across methods and spatial resolutions. It seems to be influenced only moderately by the spatial resolution of the data.  Prox produced better accuracy when observations that are near the monitoring network are resolved at the street level.  This is reversed in the case of OK and IDW, which are associated with higher accuracy in ozone predictions when observations close to the monitoring network are resolved at the zip code level. This contrast likely is related to the nature of the methods themselves and how central location plays within the method. Hence, accuracy does not seem to be a deciding factor in choosing among exposure models or the level of spatial accuracy. 
           
The concordance correlation results were largely influenced by precision.  This is not surprising because, as we have mentioned above, differences in accuracy are small compared to differences in precision.  Thus, concordance correlation was highest for OK because it yielded the greatest level of precision of the entire data set.  The combination OK-street works best within 10 miles of the monitoring network, while the optimal combination beyond 10 miles was the OK-zip combination. 
           
The MB and PE metrics seem to favor the use of street address data and the simple and straightforward Prox method for observations in proximity to the monitoring network and IDW for subject locations away from monitoring stations.  In contrast, the RMSE statistics and the concordance correlation analysis would favor the use of OK, because of its greater precision, and a street-level resolution within the A zone.  For subjects located beyond 10 miles from air monitors, OK would still be the method of choice but the optimal spatial resolution would be the zip code.
           
With regards to the assessmentvia the reliability coefficientof the potential direction of the bias in estimating the health effect, OK would produce the largest overestimation of the regression coefficient in an air pollution model.  IDW would have a similar effect. Only the Prox method would bias the effect toward the null for observations within 10 miles of the monitoring network.  The results also suggest that the spatial resolution of the data would have little influence with respect to the direction of the bias. 
           
Kriging traditionally has been considered the best method, using a true statistical approach and should return the least bias estimates. Our findings fit only in part with this traditional understanding. Looking at the data and comparing all other interpolation methods, OK may not be the method of choice across the board. Although the findings from the concordance analysis confirmed OK as the best method, the error metrics, MB and PE, did not. In addition, OK also has the potential for inducing the largest bias in the estimation of the health effect associated with ozone exposures.  Prox and IDW, on the other hand, are associated with lower levels of error but were outperformed in the concordance analysis. 
           
Overall, our results indicated that greater spatial accuracy improves exposure estimates only moderately.  The choice of exposure model, and above all, the distance (i.e., the density of) from subject locations to the monitoring network are likely to exert a greater influence on exposure error.  Based on our findingsobviously influenced by our methodological approach and data utilizedit is difficult to prescribe a straightforward, single solution to conduct the exposure assessment component of an air pollution study.  Different combinations of interpolation methods and spatial accuracies must be considered when assigning air pollution exposures to subjects residing at varying distances from existing monitoring stations.
 
6. OTHER FINDINGS
 
6.A. Total Mortality
Thanks to the no-cost extensions that have been approved for this grant, we have been able to extend mortality follow-up from 22 to 30 years.  Total mortality has been assessed in both single and two-pollutant models.  Our findings are in broad agreement with those reported by other studies.  In two-pollutant models with ozone, we find a 5% increase, although not statistically significant, risk of all natural cause mortality associated with a 10 μg/m3 mean increment in ambient levels of PM10 in both genders. For cardiopulmonary and coronary heart disease mortality as well as non-malignant respiratory mortality (poster presented at ISEE 2007), we find that the effect is somewhat stronger in females than in males.
 
Similar findings are found for cancer with PM2.5, with stronger effect among females than males (abstract presented at ISEA 2007).
 
A draft paper soon will be ready for submission. 
 
6.B. Cause-Specific Mortality
A total of 3,321 died during follow-up (1977-2006). Of these, 3230 (97%) were natural cause deaths (ICD-9<800 or ICD-10 A00-R99).
    
Through 1998, all death certificates were coded by a certified nosologist. Since 1998, we have linked with the National Death Index (NDI) and thus obtained cause-specific mortality. Age- and gender-specific total and selected cardiovascular deaths are given in the tables below for 30 years of follow-up both for the entire population (Table 6.B.a) and for those without known cardiovascular disease or cancer at baseline (Table 6.B.b).
 
TABLE 6.B.a. Mortality distributiona by cause of death, 1977-2006.
                         The AHSMOG cohort (N=6,338)
Cause of Death       
Males
(2,278)
Females
(4,060)
Totals
(n= 6,338)
All natural cause  (ICD9: 001-799)
1,218 (53.47)
2,012 (49.6)
3,230 (51.0)
Cardiopulmonary (ICD9: 401-440, 460-519)
743 (32.6)
1,267 (31.2)
2,010 (31.7)
CHD (ICD9: 410-414)
320 (14.1)
537 (13.2)
857 (13.5)
Total cancer (ICD9:140-172, 174-209)
242 (10.6)
338 (8.3)
580 (9.2)
Nonmalignant respiratory disease
(ICD9:460-519)
127 (5.6)
166 (4.1)
293 (4.6)
aNumber and percentage of popuation by causes.
 
TABLE 6.B.b. Mortality distributiona by cause of death, 1977-2006.  Ostensibly healthy AHSMOG subjects (without coronary, stroke, diabetes, cancer or COPD at baseline in 1977) (N=4,830).
Cause of Death       
Males
(1,750)
Females
(3,080)
Totals
(n= 4,830)
All natural cause  (ICD9: 001-799)
830 (47.4)
1329 (43.2)
2159 (44.7)
Cardiopulmonary (ICD9: 401-440, 460-519)
484 (27.7)
828 (26.9)
1312 (27.2)
CHD (ICD9: 410-414)
197 (11.3)
339 (11.0)
536 (11.1)
Total cancer (ICD9:140-172, 174-209)
176 (10.1)
228 (7.4)
404 (8.4)
Nonmalignant respiratory disease
(ICD9:460-519)
92 (5.3)
113 (3.67)
205 (4.2)
aNumber and percentage of population by causes.
 
Only 2 and 4%, respectively, of females and males died of non-natural causes. The majority (26 and 24%, respectively, in females and males) of deaths were due to ischemic heart disease and occurred after the age of 70 as shown in Table 6.B.c. below.   

TABLE 6.B.c. All cause death, cardiovascular death, and other deaths in the AHSMOG Study, 1977-2006 according to age at time of death.

 

FEMALES

Age at time of death
Underlying cause of death
25-49
50-59
60-69
70-79
80+
Total
All Natural Causes (ICD:  001 – 799)
10
42
170
497
1293
2012
 
 
 
 
 
 
 
Ischemic Heart Disease (ICD:  410 - 414)
0
4
27
122
384
537
Nonmalignant respiratory (ICD: 460-519)
0
2
10
31
123
166
Total Cancer deaths (ICD: 140-171, 174-208)
5
17
77
104
135
338
 
 
 
 
 
 
 
All Causes (ICD:  001 – 999)
12
45
176
510
1305
2048
 

MALES

Age at time of death
Underlying cause of death
25-49
50-59
60-69
70-79
80+
Total
All Natural Causes (ICD:  001 – 799)
7
35
152
374
650
1218
 
 
 
 
 
 
 
Ischemic Heart Disease (ICD:  410 - 414)
0
11
48
96
165
320
Nonmalignant respiratory (ICD: 460-519)
0
3
7
28
89
127
Total Cancer deaths (ICD: 140-171, 174-208)
2
7
51
85
97
242
 
 
 
 
 
 
 
All Causes (ICD:  001 – 999)
11
42
163
387
670
1273
 
6.C.Cancer Assessment
Cancer incidence is available only through 1998 based on linkage with the California Tumor registries and registries in other states where the subjects have indicated they were diagnosed with cancer. The table below gives the incidence of cancer, excluding skin cancer, by age-groups during this follow-up.
 
6.C.a. Cancer incidence
 

FEMALES

Age at time of diagnosis
 
< 60
60-79
80+
Total
   All Cancer (ICD:  140-171, 174-208)
95
231
97
423
 
 
 
 
 
   Respiratory cancer (ICD:  161-163)
2
9
9
20
   Lung cancer (ICD:  162)
2
9
9
20
   Breast cancer (ICD:  174)
41
81
23
145
   Non-Hodgkin lymphoma (ICD:196)             
1
9
10
20

 

MALES

Age at time of diagnosis
 
< 60
60-79
80+
Total
   All Cancer (ICD:  140-172, 174-209)
24
187
69
280
 
 
 
 
 
   Respiratory cancer (ICD:  161-163)
0
16
5
21
   Lung cancer (ICD:  162)
0
12
4
16
   Prostate cancer (ICD:  185)
4
95
36
135
   Non-Hodgkin lymphoma (ICD:196)
7
16
4
27
 
 
6.C.b.Cancer mortality has been assessed through 2006 based on diagnoses from the death certificates or linkages with NDI. Total cancer death in the AHSMOG cohort is given in Tables 6.B.a, 6.B.b. and 6.B.c.
 
6.D. Lung Cancer
During 22 years of follow-up (1977-1998) there were 34 cases of incident primary lung cancer among person without cancer at baseline. 
 
During 30 years of follow-up (1977-2006), there were 37 fatal cases of primary lung cancer.
 
A consistent positive association was found between ambient particulate (PM10) air pollution and risk of lung cancer.  This association was strengthened in two-pollutant models with ozone. Unfortunately, we were not been able to assess the association with PM2.5 due to small number of cases. 
 
6.D.a. Fatal lung cancer
Compared to subjects without lung cancer, those who died of lung cancer were older and had a higher number of smoke-pack years (21.7 vs. 3.1 years).  A higher proportion were past smokers (54% vs. 20%), had lower education and male gender.  A larger proportion also reported having worked in dusty environments (16.2 vs. 5.6%). We found a positive association between PM10 and risk of fatal lung cancer and there was a dose-response relationship with increasing exceedance frequencies.  However, the associations did not reach statistical significance, most likely because of low power.
 
6.D.b. Incident lung cancer
Compared to subjects without lung cancer, those with incident lung cancer were older, had lower education and reported a higher number of smoke-pack years (13.6 vs. 3.2 years).  A higher proportion were former smokers (41.2 vs. 20.1%). There was a positive association between mean levels of PM10 and risk of incident lung cancer [RR=1.24 (95% CI:0.90-1.70)] and there was a clear dose-response relationship for each 30 hours/month exceedence of increasing concentrations of PM10 with the two highest levels (80 μg/m3 and 100 μg/m3) reaching statistical significance with RR = 1.34 and 1.63, respectively. 

 

Conclusions:

To date, one paper has been published, and 13 presentations have been given (including the two at ISEE in August 2009) at scientific meetings.  Several papers are available in draft form or are being worked on by gradate students. Four doctoral students (Lie Hong Chen, David Shavlik, Rhonda Spencer Hwang and Shiva Metghalchi) are completing their dissertations addressing some of the aims of this study.

As mentioned earlier, the aim of assessing the relationship between ambient air pollution and incident CHD could not be fulfilled.  Therefore, we are moving forward with plans to assess this in the newly established NCI-funded cohort of Adventists, the Adventist Health Study 2 cohort. The air pollution part of this study is being called the AHSMOG-2 study. We will use a nested case-control study design with four controls per case and obtain medical records for all cases after obtaining their approval using a HIPAA appropriate consent form. This proposed study has not yet been funded, but we are actively pursuing funding opportunities.


Journal Articles on this Report : 4 Displayed | Download in RIS Format

Other project views: All 25 publications 4 publications in selected types All 4 journal articles
Type Citation Project Document Sources
Journal Article Chen LH, Knutsen SF, Shavlik D, Beeson WL, Petersen F, Ghamsary M, Abbey D. The association between fatal coronary heart disease and ambient particulate air pollution:are females at greater risk? Environmental Health Perspectives 2005;113(12):1723-1729. R830547 (2005)
R830547 (2006)
R830547 (2007)
R830547 (Final)
R827998 (Final)
  • Full-text from PubMed
  • Abstract from PubMed
  • Associated PubMed link
  • Full-text: EHP-Full Text HTML
  • Other: EHP-Full Text PDF
  • Journal Article Gharibvand L, Lawrence Beeson W, Shavlik D, Knutsen R, Ghamsary M, Soret S, Knutsen SF. The association between ambient fine particulate matter and incident adenocarcinoma subtype of lung cancer. Environmental Health 2017;16(1):71 (9 pp.). R830547 (Final)
  • Full-text from PubMed
  • Abstract from PubMed
  • Associated PubMed link
  • Full-text: BMC-Full Text HTML
    Exit
  • Other: BMC-Full Text PDF
    Exit
  • Journal Article Gharibvand L, Shavlik D, Ghamsary M, Beeson WL, Soret S, Knutsen R, Knutsen SF. The association between ambient fine particulate air pollution and lung cancer incidence: results from the AHSMOG-2 study. Environmental Health Perspectives 2017;125(3):378-384. R830547 (Final)
  • Full-text from PubMed
  • Abstract from PubMed
  • Associated PubMed link
  • Full-text: EHP-Full Text HTML
  • Other: EHP-Full Text PDF
  • Journal Article Spencer-Hwang R, Knutsen SF, Soret S, Ghamsary M, Beeson WL, Oda K, Shavlik D, Jaipaul N. Ambient air pollutants and risk of fatal coronary heart disease among kidney transplant recipients. American Journal of Kidney Diseases 2011;58(4):608-616. R830547 (Final)
  • Abstract from PubMed
  • Abstract: AJKD-Abstract
    Exit
  • Supplemental Keywords:

    Ambient air, ozone, particulate matter, exposure, risk, risk assessment, health effects, human health, sensitive populations, population, elderly, cumulative effects, susceptibility, epidemiology, modeling, monitoring, analytical, Bayesian neural networks, GIS, southwest, California, CA
     
    , RFA, Economic, Social, & Behavioral Science Research Program, Health, Scientific Discipline, Air, ENVIRONMENTAL MANAGEMENT, particulate matter, Health Risk Assessment, Risk Assessments, Susceptibility/Sensitive Population/Genetic Susceptibility, Disease & Cumulative Effects, Biochemistry, Environmental Statistics, genetic susceptability, Biology, Risk Assessment, ambient air quality, elderly adults, health effects, sensitive populations, health risk analysis, air pollutants, long term exposure, acute lung injury, Bayesian approach, cardiovascular vulnerability, exposure, Bayesian method, Bayesian neural networks, air pollution, chronic health effects, human exposure, statistical models, susceptibility, particulate exposure, sensitive subjects, Acute health effects, elderly, GIS, sensitive subgroups, cardiotoxicity, mortality, tobacco smoke, age dependent response, cardiovascular disease, cumulative effects, exposure assessment, human health risk, respiratory, genetic susceptibility, cardiopulmonery responses, toxics

    Relevant Websites:

    Loma Linda University

    Progress and Final Reports:

    Original Abstract
  • 2003 Progress Report
  • 2004 Progress Report
  • 2005 Progress Report
  • 2006 Progress Report
  • 2007 Progress Report