Cancer incidence and community exposure to air emissions from petroleum and chemical plants in Contra Costa County, California: a critical epidemiological assessment.

Contra Costa County covers an area of nearly 800 square miles in the northeastern portion of the San Francisco Bay area. It has a population of about 850,000. Its central and eastern sections are two of the fastest growing residential areas in the entire Bay area. In marked contrast, the "older" western and northern portions of the county are considerably more urban and industrial, with a total of five petroleum refineries and numerous chemical and other industrial plants concentrated along the rim of the bay.

Since the late 1970s, controversy has raged off and on as to whether emissions from these petrochemical operations are contributing to higher cancer rates -- particularly lung cancer -- in the so-called industrialized portions of the county. This was fueled initially by a National Cancer Institute (NCI) study in 1977 (1) which found Contra Costa, along with 38 other U.S. counties, with the heaviest concentrations of petroleum refining, to have significantly higher cancer mortality rates for a number of sites, including lung, nasal cavity and skin. However, a later study by researchers at the Kaiser-Permanente Medical Center in 1980 (2) did not find any association between cancer incidence and residence near petrochemical operations.

In 1984, two studies, by Kaldor et al. (3) and Austin et al. (4), both at the California Department of Health Services (DHS), came to different conclusions concerning the relationship between lung cancer and air pollution patterns throughout the county. Each employed different methods to measure and model specific air pollution contaminants for census tracts in the county, and to assess their correlation with cancer incidence. While Kaldor et al. included only industrial sources, Austin et al. also considered emissions from mobile sources, the highest single source of air toxics. A recent report published by the Bay Area Air Quality Management District (BAAQMD) (5) showed the central corridor along the Interstate Freeway 680 to have the highest concentration of benzene, and many other toxic compounds, in the county.

This paper will focus on the pitfalls of air pollution modeling from an epidemiologic perspective using these studies as examples. Vastly different conclusions can be reached, depending on how the data are derived and used. Furthermore, the models must be tested and validated with actual data. Finally, as with any epidemiological investigation, adequate control of confounding exposures is essential. For example, in studying the relationship between lung cancer and environmental exposures, smoking, occupational exposures and residential histories must be taken into consideration.

The Kaldor et al. (1984) study

Industrial air emissions -- In the Kaldor et al. (1984) study (3), the epidemiologists from the California DHS relied on a model developed by the BAAQMD to assess residential exposures to industrial air emissions. Since 1972, the BAAQMD has been estimating the amount of sulfur dioxide (S|O.sub.2~), hydrocarbons (HC) and nitrogen oxides (N|O.sub.x~) emitted from all major industries in Contra Costa County. These estimates were based on measurements taken at the point sources and the amount of chemicals used or produced by selected industrial sites. BAAQMD also estimates emissions by non-industrial sources such as automobiles, aircraft and small businesses.

In 1975, according to BAAQMD, 180 tons/day HC and 130 tons/day of S|O.sub.2~ were emitted by all sources in the county. Petroleum refineries and chemical plants accounted for 68 tons/day (38 percent) HC and 91 tons/day (70 percent) S|O.sub.2~. Power plants accounted for slightly less than 26 tons/day (20 percent) S|O.sub.2~, but less than 1 ton/day (1 percent) HC.

Air pollution model -- The BAAQMD developed an air pollution model based on the following information:

1) 1975 industrial emission estimates;

2) topographic data from U.S. geologic survey; and

3) 1973 meteorologic data.

Emissions produced by motor vehicles and other non-industrial sources were not TABULAR DATA OMITTED included in the model. The meteorological data were based on measurements taken at BAAQMD monitoring stations in 1973. Variables included spatial distribution of average, hourly wind speeds and annual distribution of inversion base heights. Average annual ambient concentrations of industrial emittants for each square kilometer area were estimated, based on the Hanna-Gilford method of air dispersion calculations. The model did not include air pollution episodes such as emergency releases or fires.

In Contra Costa County, there were 107 census tracts at the time of the Kaldor study (3). These 107 census tracts were merged into 21 groups that were geographically contiguous and demographically similar (family income, age distribution and race). Based on population density maps and the BAAQMD air pollution estimates for each square kilometer, a population-density-weighted average exposure to S|O.sub.2~, HC and N|O.sub.x~ for each of the 21 groups for 1975 was calculated.

Table 1 presents the average exposures for the 21 groups of census tracts used in the Kaldor study (3). The 21 grouped tracks were further consolidated into four areas based on degree of exposure. Area 4 had relatively high concentrations of S|O.sub.2~ and HC (|is greater than~ 1 mg/|m.sup.3~), and Area 3 had relatively low concentrations of S|O.sub.2~ and HC (0-1 mg/|m.sup.3~). Since Area 2 had no industrially emitted S|O.sub.2~ and almost no HC exposures, it was considered in the Kaldor study to be unexposed to emissions from the petroleum and chemical industries. Area I had no exposure to industrially produced S|O.sub.2~, HC or N|O.sub.x~. Therefore, it was considered to be unexposed to industrial air emissions.

Cancer incidence data -- The San Francisco Bay area counties were part of the National Cancer Institute's Surveillance Epidemiology and End Results Program (SEER). As part of the SEER program, cancer incidence data in Contra Costa County for 1969-1977 were reported based on death certificates and hospital records. Based on previous investigations, the cancer incidence data were believed to be 98 percent complete. The diagnoses were confirmed by pathology reports in 90 to 95 percent of cases.

Correlation analysis -- Based on cancer incidence data, age-adjusted rates were computed for 20 selected cancer sites and types (such as esophagus, stomach, colon, liver, pancreas, lung, leukemia, melanoma, prostate, breast, bladder, kidney, brain, lymphoma, etc.) for each sex for the four areas. Areas 2, 3, and 4 were judged to have similar socioeconomic status (SES), while Area 1 was of higher SES. Based on this argument, Area 1 was not included in the subsequent correlation analysis in the Kaldor study. The Bartholomew |X.sup.2~ procedure was used to test for any increasing or decreasing trends in rates among the three areas. It should be noted that, in applying the Bartholomew's test to the three data points (three areas), Kaldor combined the middle point (Area 3) with either the low (Area 2) or the high point (Area 4) in order to create an upward trend. Based on this highly subjective procedure, Kaldor et al. were able to create a number of statistically significant trends. Based on these trends. Kaldor et al. concluded the following:

"In both males and females, residential exposure to petroleum and chemical air emissions was associated with an increased incidence of cancer of the buccal cavity and pharynx. In males, age-adjusted incidence rates for cancers of the stomach, lung, prostate and kidney and urinary organs were also associated with petroleum and chemical plant air emission exposures."

The Austin et al. (1984) study

Air pollution monitoring -- In the Austin et al. (1984) study (4), air pollution pattern s in the county were determined by 15 high volume particulate samplers (five permanent air monitoring stations and 10 temporary stations) strategically placed at 13 locations in Contra Costa County and two in adjacent counties. Air particulate material was collected at these 15 stations every sixth day from November 1978 to October 1979. The collected samples were analyzed for total suspended particulates (TSP), benzene soluble organics (BSO), sulfate (S|O.sub.4~), nitrate (N|O.sub.3~), lead (Pb), selected polycyclic aromatic hydrocarbons (PAH), and mutagenic activity.

Air pollution model -- For the purpose of epidemiologic analysis, levels of air pollution by census tract were needed. Contour maps of distribution of air pollutants were constructed using a computer program called SYMAP. The location of the monitoring station and the corresponding levels were used to create a matrix of pollution levels for the entire county. The contour maps were derived from these matrix values.

Based on these contour maps, pollution levels for each population center from each of the census tracts in the county were estimated. To validate the extrapolation from monitoring stations to census tracts, correlation coefficients between pollutants computed for the census tracts were compared to those for the monitoring stations. The comparability of the two sets of correlation coefficients indicated that the relationships of air pollution constituents originated from the monitoring stations were preserved by the computer mapping and that it would be valid to use the census tract values.

Correlation analysis -- Correlation coefficients between various air pollution constituents and average 10-year age-adjusted lung cancer incidence rates by sex for each census tract were computed. Only one air pollutant, S|O.sub.4~, showed a significant correlation with lung cancer in white males (r = 0.46, p |is less than~ 0.0001), but not in white females. Since certain census tract socioeconomic variables were also correlated with lung cancer, partial correlation coefficients controlling for these variables were computed. It can be seen from the table that controlling for percent of Hispanics, percent of blue collar workers and education reduced the correlation to nonsignificant level. Thus, S|O.sub.4~ itself was not correlated with lung cancer rates.

Case-control study -- In addition to the correlation analysis, a case-control study consisting of 249 lung cancer cases and 373 controls (matched for race, sex and age) was conducted by Austin et al (1984) (4). Detailed employment, residential and TABULAR DATA OMITTED smoking histories were obtained through the use of questionnaires. Using multiple logistic regression analysis, all air pollutants were examined for their relationships with lung cancer. The potential impact of various occupations controlling for smoking, drinking, asbestos exposures and S|O.sub.4~ dose was examined. No industry (including metal, construction and petrochemical) was found to adversely affect lung cancer. In males, both smoking duration and smoking dose showed a significant adverse impact on lung cancer risk, while the consumption of green vegetables reduced lung cancer risk significantly. In females, only smoking dose increased lung cancer risk significantly.

TABULAR DATA OMITTED

Based on their findings, Austin et al. concluded: "Analysis indicated that the major contribution to lung cancer in the county was due to cigarette smoking. No significant association between lung cancer risk and measured constituents of air pollution was found."

Discussion

Two air pollution epidemiologic studies were conducted around the same time in Contra Costa County by epidemiologists in the California Department of Health Services. The Kaldor et al. (1984) report (3) concluded that lung cancer risk in males in the county was significantly increased as a result of their exposure to industrial emissions. The Austin et al. (1984) report (4) concluded that none of the air pollutants was correlated with lung cancer risk, and that the variation in lung cancer risk within the county could be explained by cigarette smoking. Why do the two studies point to two opposite conclusions?

Comparison of epidemiologic design and analysis between the two studies

There were many important differences between the two studies. Each study had two major components: air pollution assessment, and epidemiologic design and analysis. Although it is not the purpose of this paper to discuss the differences in epidemiologic design and analysis between the two studies in detail, for completeness, the major contrasts will be listed below.

Table 3. Comparison of the Pearson Correlation Coefficients for
the mean annual measured values for 15 monitoring stations and
for computes values for 113 census tracts, Contra Costa County,
CA, Nov. '78-Oct. '79

Comparison          Stations    113 census tracts

BSO vs. PB             .78             .70
BSO vs. Mut (-S9)      .24             .28
BSO vs. Mut (-S9)      .28             .39
BSO vs. BAA            .87             .89
BAP                    .80             .84
BGP                    .64             .71
CHR                    .83             .84
BSO vs. S|O.sub.4~     .19             .17
TSP vs N|O.sub.3~      .64             .75

Source: Austin, et al. (4)

First, the Kaldor et al. (3) study utilized TABULAR DATA OMITTED the ecologic study design to investigate the relationship between exposure to air emissions produced by the petroleum and chemical industries and average annual cancer incidence among residents of Contra Costa County. An ecologic study is one in which the units studied are populations or groups (e.g., census tracts, groups of census tracts, counties, etc.) rather than individuals. Correlations or associations found in this type of study may not hold true for the individual members of these populations or groups. This major limitation of ecologic analysis, known as "ecological fallacy," is well known to epidemiologists. One way to minimize the ecological fallacy is to make the groups as homogeneous as possible by using smaller units of analysis (6).

In the Kaldor study, data on cancer incidence and air emissions were estimated for each individual census tract. An analysis could have been done based on the 107 census tracts, although some of the larger census tracts might not be internally homogeneous and the ecological fallacy would be a problem even at this level. The investigators of the study magnified the ecological fallacy by first grouping the 107 census tracts into 21 groups and further merging these groups into four exposure areas. Although the 21 groups might be geographically contiguous and demographically similar, the eventual four exposure areas were not geographically contiguous and not necessarily demographically similar. For example, within the eastern portion of Area 2, the northern part (Pittsburg and Antioch) is heavily industrialized, while the lower part is primarily agricultural. On the other hand, the Austin correlation analysis was based on individual census tracts. Furthermore, Austin et al. also conducted a case-control study based on individual residents, a study design free from the ecological fallacy.

Second, smoking is known to be a potent lung carcinogen, but the use of tobacco was not considered in the Kaldor study. In the Austin investigation, smoking information was available in the case-control study, and it was found that the males in the industrial areas of the county smoked considerably more than those in non-industrial areas, and that the difference in smoking practices in these two areas accounted for the observed difference in lung cancer risk in that study.

Third, there was an inverse relationship between socioeconomic status (SES) and industrial pollution level among the four areas in the Kaldor study. Even after Area 1 was dropped from the analysis, there were still differences among the remaining three areas. For example, Area 2 had a higher income, lower percent of blue collar workers, and higher percent of high school graduates than Area 4. Furthermore, the SES data presented in the report were based on the 1970 census, which might not be representative of the entire study period. In the Austin correlation analysis, the significant correlation between S|O.sub.4~ and lung cancer was eliminated by controlling for various SES variables. Failure to control for these SES differences between different areas contributed to the apparent association between air pollutants and lung cancer in the Kaldor study.

Comparison of the exposure classification of the Austin and Kaldor studies

Of 107 census tracts in Contra Costa County, 67 were classified by Kaldor et al. as industrial tracts with at least some exposure to emissions (Areas 2, 3 or 4); yet only 29 out of those 67 were classified by Austin et al. as industrial areas. There was an additional discrepancy in exposure classification between the two studies: 13 census tracts in Area 4 (the most highly polluted) in the Kaldor study were classified by Austin as non-industrial (non-exposed) at all. The Austin model was more appropriate for the following reasons.

The Kaldor industrial emission estimates were derived from measurements made at the point of emission, weighted by population density. Emissions from automobiles and other non-industrial sources or industrial sources adjacent to the county were not considered. Since the estimates were weighted by population density in 1975, in order for these estimates to be applicable to the entire 20 to 30 years of latency for lung cancer, the relative population density must be constant over such a long period -- a very unlikely situation. For example, in the county as a whole, the population was 298,984 in 1950 and 657,252 in 1980--an increase of 120 percent. However, for the City of Concord, the population was 6,953 in 1950 and 103,000 in 1980--an increase of 1,381 percent. On TABULAR DATA OMITTED TABULAR DATA OMITTED the other hand, for the City of Richmond, the population was 99,545 in 1950 and 74,676 in 1980--a 25 percent decrease. It is clear that the county has experienced tremendous growth in the last 30 years, but the growth is not uniform across the country.

It also seems peculiar that in the Kaldor study some regions of Area 4 (most polluted) lie alongside some regions of Area 1 (no pollution) in certain parts of the City of Concord. Census tracts 328, 329 and 330 on one side of Clayton Road/Willow Pass Road in Concord were classified in Area 4, while the adjacent census tracts 331,335 and 336 on the other side of the same streets were classified in Area 1. This part of Concord is flat, with low-rise commercial buildings, apartment buildings and residential houses. Since the major components in a dispersion model from point sources are distance from the sources as well as wind direction and wind speed, one would expect the air pollution pattern to be more continuous, instead of dropping from the most polluted level in some relatively geographically small census tracts on one side of the street to no pollution at all in another group of geographically small census tracts on the other side of the same street. Thus, the air pollution model or exposure classification in the Kaldor study failed a "common sense" validation.

On the other hand, the exposure estimates in the Austin study were based on actual air monitoring measurements collected by high volume particulate samplers strategically located in Contra Costa County as well as in adjacent counties. The Austin pollution model was validated by comparing the correlation coefficients of various air pollutants derived from actual monitoring data to estimated census data. The comparability supported the validity of the model. Furthermore, in the case-control study in the Austin investigation, residential histories were collected and analyzed, whereas Kaldor analyzed residence at diagnosis or death only, thereby implicitly assuming that the study subjects had lived at the same residence for the last 20 to 30 years (to allow for exposure and latency), when the average length of residence in the county was six years or less, according to Kaldor et al. (1984) (3).

Comparison to occupational studies in the county

As an additional means of assessing the validity of these air pollution studies, a plausible hypothesis to test is whether any occupational studies by the largest industrial emitters show consistent results. During this same time period, three of the county's largest petrochemical companies (Chevron, Dow and Shell) completed studies of their employees to assess cancer risks (7, 8, 9, 10, 11, 12). Five of these studies were retrospective cohort mortality studies based on deaths as far back as 1950; the sixth was a cancer incidence study based on data from the Bay Area Tumor Registry. In all these studies, lung cancer rates among petrochemical employees were found to be lower, in some cases significantly, than the expected. While the majority of these studies used the general U.S. population for comparison, and therefore the typical "healthy worker effect" must be taken into account, the results nevertheless argue against plant exposures (emissions) posing an environmental lung cancer risk.

The important role of exposure assessment in epidemiologic studies

One of the most important criteria in determining causation in epidemiology is an exposure-response relationship. In a broad sense, exposure is simply a classification scheme. The most fundamental classification scheme is "exposed vs. non-exposed." A more desirable classification is one with various levels of exposure. Epidemiologic analysis is simply a comparison of the disease patterns between the exposed and the non-exposed, or among the various levels of exposure.

Exposure assessment is a vital component in any epidemiologic study. The primary objective of exposure assessment in an air pollution epidemiologic study is to avoid, or at least to minimize, the misclassification of study subjects by exposure status or exposure level.

In this paper we discussed two air pollution epidemiologic studies. One of the major differences between the two studies was the air pollution model. As demonstrated in this paper and in the Austin investigation, the apparent association between air pollutants and lung cancer risk reported in the Kaldor paper was due primarily to exposure misclassification as a result of an inappropriate air pollution model (or, at least, the misuse of the original air pollution data by the epidemiologists) and to inappropriate statistical analysis. Based on our review, the following points are vital in a successful linkage between air pollution data and epidemiologic analysis.

First, if the unit of analysis in epidemiology is something other than individuals (e.g., census tracts, groups of census tracts, counties), the unit should be as small as possible. Homogeneity of exposure within each unit should be examined. In many situations, it may be infeasible or impractical to test the entire data set for homogeneity. In such a case, the homogeneity test should be done on a sampling basis.

Second, like any other model, an air pollution model should be validated with actual data. This requirement usually presents some difficulties, since a model is developed because there is no or little real data in the first place. Nevertheless, a valid model has to make sense in the real world. For example, certainly the question of validity should be raised if the model predicts highest exposure on one side of a street and no exposure on the other. Since a model necessarily relies on numerous assumptions, it becomes prudent to check the validity of these assumptions.

Third, it is important to include emissions from all known sources in the model, both industrial and non-industrial. Even if the objective of the study is to assess the effects of industrial emissions only, it makes little sense not to include non-industrial emissions which would otherwise become strong confounders.

Fourth, air pollution is only part of the total exposure, or one of the many potential risk factors. In most air pollution studies, communities with different air pollution levels are compared. Communities are defined according to various geographical or political boundaries. In addition to air pollution level, usually these geographic or political units also differ in many socioeconomic aspects, some of which are associated with various disease risks. These SES differences must be investigated and addressed in the design and/or analysis of the study.

Conclusion

Air pollution modeling is a vital component in any epidemiologic study. An incorrect or inappropriate air pollution model will produce misclassification in the epidemiologic analysis, which in turn may create spurious associations. Air pollution models should be validated before they are incorporated into any epidemiologic analysis.

References

1. Blot, W.J., L.A. Brinton, J.F. Fraumeni and B.J. Stone (1977), Cancer mortality in U.S. countries with petroleum industries, Science 198:51-53.

2. Hearey, C.D., H. Ury, A. Siegelaub, M.K.P. Ho, H. Salomon and R.L. Cella (1980), Lack of association between cancer incidence and residence near petrochemical industry in San Francisco Bay area, J. Nail. Cancer Inst. 64:1295-1299.

3. Kaldor, J., J.A. Harris, E. Glazer, S. Galser, R. Neutra, R. Mayberry, V. Nelson, L. Robinson and D. Reed (1984), Statistical association between cancer incidence and major-cause mortality, and estimated residential exposure to air emissions from petroleum and chemical plants, Env. Health Persp. 54:319-332.

4. Austin, D.F., V. Nelson, B. Swain, L. Johnson, S. Lum and P. Flessel (1984), Epidemiological Study of the Incidence of Cancer as Related to Industrial Emissions in Contra Costa County, California, Report to the Environmental Protection Agency, Pub. No. EPA-600-S1-84-008.

5. Bateman, B. and T. Umeda (1988), Contra Costa County Air Toxics Exposure Study, Bay Area Air Quality Management District Planning Division, Berkeley, CA.

6. Morgenstern, H. (1982), Uses of ecologic analysis in epidemiologic research, Am. J. Public Health 72:1336-1344.

7. Ott, G.M., G.L. Carlo, S. Steinberg and G.G. Bond (1985), Mortality among employees engaged in chemical manufacturing and related activities, Am. J. Epid. 122(2):311-322.

8. Wong, O., R.W. Morgan, W. Bailey, R.E. Swencicki, K. Claxton and L. Kheifets, An epidemiologic study of petroleum refinery employees, Brit. J. Indus. Med. 43:6-17.

9. ... (1983), Cause-Specific Mortality Among Employees of the Chevron Chemical Facility in Richmond, Environmental Health Associates, Berkeley, CA.

10. ... (1983), Cause-Specific' Mortality Among Employees of the Chevron Research Company, Environmental Health Associates, Berkeley, CA.

11. Tsai, S.P., E.L. Gilstrap, S.R. Cowles, P.J. Snyder and C.E. Ross (1993), A Cohort Mortality Study of Two California Refinery and Petrochemical Plants, J. Occup. Med. 35(4):415-416.

12. Bond, G.G., D.F. Austin, M.R. Gondek, M. Chiany and R.R. Cook (1988), Use of a population based tumor registry to estimate cancer incidence among a cohort of chemical workers, J. Occup. Med. 30(5):443-448.

Otto Wong, Sc.D., F.A.C.E., Chief Epidemiologist, Applied Health Sciences, 181 Second Ave., Suite 628, P.O. Box 2078, San Mateo, CA 94401.