Grantee Research Project Results
Final Report: Probabilistic Modeling of Variability and Uncertainty in Urban Air Toxics Emissions
EPA Grant Number: R826790Title: Probabilistic Modeling of Variability and Uncertainty in Urban Air Toxics Emissions
Investigators: Frey, H. Christopher
Institution: North Carolina State University
EPA Project Officer: Chung, Serena
Project Period: October 1, 1998 through September 30, 2001 (Extended to March 31, 2004)
Project Amount: $329,425
RFA: Urban Air Toxics (1998) RFA Text | Recipients Lists
Research Category: Air
Objective:
The objectives of this research project were to: (1) develop and refine methods for quantification of variability and uncertainty in estimating emissions of urban air toxics (Tasks 1-4); (2) develop and refine methods for identifying key sources of variability and uncertainty in assessments of urban air toxics emissions and exposures (Task 5); (3) develop probabilistic process engineering models for making realistic estimates of emissions of urban air toxics (Task 6); (4) demonstrate the methods with a detailed case study of urban air toxics emissions and exposures (Task 7); and (5) characterize the benefits of the methods with respect to environmental and research management (Task 8).
A critical first step in any type of exposure or risk assessment is to clearly identify the assessment endpoint, including a clear definition of health effects. This then motivates the selection of: (1) an appropriate averaging time; (2) geographic area; (3) exposed population; and (4) exposure scenarios (e.g., direct inhalation, indirect pathways, long-range transport, etc.). The assessment endpoint significantly affects what type of data must be collected or produced for input into the assessment. For example, if the assessment is concerned with exposures to an acute toxin, short averaging times of perhaps 1 hour should be used. If the assessment is concerned with exposures to low concentrations of a carcinogenic compound, annual or lifetime average exposures may be of concern. The variability in annual emissions and exposures will typically be substantially less than the variability in hourly emissions or exposures. Similarly, there may be more uncertainty regarding the possible values of hourly emissions than for annual averages. Thus, the development of quantitative estimates of variability and uncertainty is highly dependent on the type of assessment. In many cases, the data available for averaging times or exposed populations are not what is needed for a given assessment.
The methodological focus of this research was how to quantify variability and uncertainty given that the assessment endpoint had been properly defined. Therefore, the methods we used in this research addressed how to: (1) quantify temporal and spatial variability in emissions and exposures using frequency distributions; (2) convert data from one averaging time to other averaging times relevant to the assessment endpoint (e.g., acute versus chronic health effects); (3) address both statistical sampling error (small data sets) and measurement error in characterizing frequency distributions for variability; (4) quantify uncertainties associated with statistical sampling error and measurement error; (5) quantify variability and uncertainty when some data are missing, such as for data sets containing nondetected values; (6) properly deal with correlation structures between variable and/or uncertain quantities; (7) propagate both variability and uncertainty through a model while retaining correlations among the inputs and maintaining sufficient precision to make predictions of outcomes for highly exposed individuals; and (8) identify the key sources of variability and uncertainty in model inputs that most significantly affect variability and uncertainty in model outputs. The development of methods to address the issues listed above was the focus of Tasks 1 through 5 of this research.
The new methods were applied to the development of several probabilistic emissions estimates in Task 6. The model inputs included process-specific information such as feedstock compositions, design parameters, and operating (activity) inputs. Thus, the emissions of hazardous air pollutants (HAPs) were predicted as a function of factors that led to temporal variability in emissions. The models were developed in a manner similar to previous efforts by Frey, et al. (1996) and Frey and Rhodes (1996). The main focus of this effort was mobile source emissions.
The methods for quantitative analysis of variability and uncertainty and the probabilistic emissions models were applied to a detailed case study regarding the development of a HAPs emissions inventory for an urban area in Task 8. Because this research program was primarily focused on the development of a modeling methodology, we devoted Task 8 to identifying and illustrating the benefits of the method with respect to environmental and research management.
Summary/Accomplishments (Outputs/Outcomes):
The motivating questions for this research were:
(1) What are the most significant sources of toxic pollutants of concern in urban areas?
(2) What would be the effect of control measures on the emissions?
(3) How can an estimation of emissions best be linked to estimate exposure and risk?
A key shortcoming of current approaches to emissions estimation, exposure assessment, and risk assessment is the failure to properly quantify both variability and uncertainty. Uncertainty and variability in data and models have important implications for decision making. Uncertainty is a lack of knowledge about a quantity’s true value because of measurement error, systematic error, irreducible randomness, disagreement among experts, or the lack of an empirical basis for an estimate. Variability comprises the heterogeneity of values for different members of a population. Information about uncertainty can help determine where additional research or alternative measurement techniques are needed to reduce uncertainty. Knowledge of variability is important, for example, in identifying subgroups susceptible to specific health risks from exposures to a given chemical.
Variability and uncertainty in HAP emissions can be large (Rubin, et al., 1993; Frey and Rhodes, 1996). Furthermore, to develop emissions inventories that are relevant for human exposure and risk assessment, it is critically important to characterize both variability and uncertainty based on appropriate averaging times. For example, the averaging times used for estimating risks associated with acute health effects are typically much shorter than those used for chronic health-effect risk assessments. In this research, we rigorously developed practical methods for properly quantifying variability and uncertainty in urban air toxics emissions.
This report focuses on an integrated presentation of information that builds on the methods developed in tasks 1-5 and the emissions models of task 6. Most of the material in this report represents tasks 7 and 8, which include the case studies of probabilistic emission inventories and the demonstration of environmental management implications based on an exposure assessment case study.
The results of this project have been reported in several peer-reviewed journals as well as other publications. Listed below are the results with respect to each of the major tasks:
1. Developed and refined methods for quantification of variability and uncertainty in estimating emissions of urban air toxics (Tasks 1-4) (Frey and Bammi, 2002, 2003; Frey and Zhao, 2004; Frey and Zheng, 2002; Frey and Small, 2003; Zhao and Frey, 2004b, 2004c; Zheng and Frey, 2004).
2. Developed and refined methods for identifying key sources of variability and uncertainty in assessments of urban air toxic emissions and exposures (Task 5) (Frey and Zhao, 2004; Zhao and Frey, 2004a).
3. Developed probabilistic process engineering models for making realistic estimates of emissions of urban air toxics (Task 6) (Bammi, 2001; Frey and Bharvirkar, 2002; Frey and Zheng, 2002).
4. Demonstrated the methods with a detailed case study of urban air toxics emissions and exposures (Task 7) (Frey and Zhao, 2004; Zhao and Frey, 2004).
5. Characterized the benefits of the methods with respect to environmental and research management (Task 8).
In this project, rigorous statistical methods to develop probabilistic urban air toxics emission inventories were demonstrated through detailed case studies with a focus on emission factors and statistical analysis of empirical data. The key contributions of this work include a synthesis of methods into one framework and case study results for specific pollutants, for which uncertainty estimates were not previously available. Interunit variability and uncertainty in mean emission factors were quantified using parametric, mixture, and empirical distributions combined with bootstrap simulation. Thus, the methodology avoided imposing restrictive normality assumptions on the uncertainty estimates for mean emission factors. A rigorous, asymptotically unbiased method was applied to deal with multiple censored data. Statistical and graphical methods for evaluating goodness-of-fit were demonstrated, including a novel graphical method applied to censored data. A methodology was proposed for and applied to developing surrogate emission factor uncertainty estimates for source categories in which data were lacking. For some source categories, uncertainty estimates were developed based on simple models or weighted averages among processes or time periods. Probabilistic emission inventories were developed for six urban air toxics for a specific urban area. Sensitivity analysis was used to show that the results for overall uncertainty were typically influenced by a small number of source categories. Comparative and sensitivity analysis results demonstrated that surrogate uncertainty estimates did not substantially influence the case study results. This is partially because only a small portion of the total emissions for each of the six pollutants involved the use of surrogate uncertainty assumptions.
There are many possible uses of probabilistic emissions estimates, including air quality modeling and exposure assessment. Depending on the pollutant, typical uncertainty for an urban area is a factor of two or three for short-term (e.g., daily) averages of the emission factor data. Therefore, uncertainty would propagate through an air quality model when making predictions of ambient concentrations. This range of uncertainty is greater than that for many exposure factors (e.g., inhalation rate for some subpopulations) and implies that uncertainty in emission inventories could be an important contributor to uncertainty in exposures. The uncertainty analysis also helps answer many key questions posed by decision makers regarding how good the numbers are, the key sources of uncertainty, and where resources should be targeted to reduce uncertainty.
The case study results represent a first step toward development of a set of probabilistic emission factors for several important urban air toxics that could be used by others to estimate uncertainty in emission inventories for other urban areas. These uncertainty estimates include random sampling and measurement errors.
A key implication of the sensitivity analysis is that in the future it may be possible to develop good preliminary estimates of uncertainty in urban emission inventories of specific air toxics based on analysis of uncertainty in a relatively small number of source categories. Similarly, it may be possible to substantially reduce uncertainty in the inventory by focusing collection of more or better data on just a small number of key source categories.
Because data analysis is based on short-term averaging times, the results can be used in acute exposure analysis (e.g., to study acute interstitial changes caused by mercury). Currently, there is a lack of long-term emissions data. Longer term emissions data are needed in the future to better support development of exposure assessments pertaining to chronic health effect endpoints.
The probabilistic emission inventories could be improved with additional data or the incorporation of expert opinion. Although biases in the mean emission factors may exist, especially for fugitive emissions and as a result of process upsets, there were insufficient data to quantify such biases. Other possible sources of bias include lack of representative data (e.g., measurements may have been for load or operating conditions not typical of annual average in-use activity) and the use of surrogate data for source categories in which data were lacking or not readily available. Expert elicitation could be used to encode judgments of additional uncertainty associated with nonrepresentative or surrogate data. As new data become available, the assessment can be updated.
A key obstacle to a quantification of uncertainty based on statistical data analysis is the way in which data are obtained. Often, data are measured and reported by multiple organizations. In the long term, a protocol for archiving such data and making them available would facilitate probabilistic analysis. Emission factor databases should contain measured values, detection limits, averaging times, and information regarding the precision and accuracy of measurement methods.
In the long term, quantifiable uncertainty in activity factors should be incorporated when empirical data are available for statistical analysis or based on other inference methods. For the cases that lack sample data, expert judgment may be required as the basis for a subjective estimate of uncertainty. Judgment-based approaches also could be used to introduce an uncertainty factor associated with the use of surrogate data, because the uncertainty estimates may need to be widened when applying a normalized uncertainty distribution to a different category. Also, there is a long-term need for time series data or models so that uncertainty in emission factors can be adjusted to averaging times appropriate for a particular need (e.g., acute versus chronic exposure assessment).
In addition, case studies were developed for probabilistic exposure assessments based on urban air toxic exposure scenarios. The case studies demonstrated methods for characterizing variability and uncertainty in exposure assessment. Specifically, the case studies demonstrated how to deal with missing values, censored data, and zero values in exposure data sets and their associated uncertainty. (These issues often are ignored or improperly handled in most current exposure assessments.) Improper handling of censored data may bring bias to variability and uncertainty estimates. Ignoring zero values may lead to loss of useful information. For example, in the case studies, addressing the zero values instead of discarding them provided information on uncertainty in the portion of the population with zero exposure to pollutants of interest. The information reasonably reflected the real-world exposure scenarios. Such information could not be obtained if zero values in the exposure factor data sets are ignored.Similarly, by appropriately dealing with censored data and missing values, the analysis was based on the best use of available data despite their imperfections. Sensitivity analysis was used to identify key sources of uncertainty, of which the pollutant concentration was found to be the single most important source. The pollutant concentration was directly influenced by the emissions of urban air toxics.
The results of this work demonstrate that random sampling and measurement errors lead to substantial quantifiable uncertainty in the emission inventories of selected urban air toxics. The positively skewed ranges of uncertainty account for the fact that emissions must be non-negative. The substantial ranges of uncertainty estimated here should be taken into account when conducting air quality modeling and exposure assessment. Furthermore, the identification of key sources of uncertainty in the inventory helps to prioritize resources for additional data collection or research to reduce uncertainty.
References:
Frey HC, Kini MD, Ranjithan SR, Fu SY. Uncertainty, bias, and variability in emission factors for light duty gasoline vehicles. In: Proceedings of the 89th Annual Meeting of the Air and Waste Management Association, Nashville, TN, June 23-28, 1996, Paper No. 96-108B.03.
Journal Articles on this Report : 9 Displayed | Download in RIS Format
Other project views: | All 21 publications | 12 publications in selected types | All 9 journal articles |
---|
Type | Citation | ||
---|---|---|---|
|
Frey HC, Li S. Methods for quantifying variability and uncertainty in AP-42 emission factors: case studies for natural gas-fueled engines. Journal of the Air & Waste Management Association 2003;53(12):1436-1447. |
R826790 (Final) R826766 (Final) |
Exit Exit |
|
Frey HC, Bammi S. Probabilistic nonroad mobile source emission factors. Journal of Environmental Engineering 2003;129(2):162-168. |
R826790 (2000) R826790 (2001) R826790 (2002) R826790 (Final) R826766 (2001) R826766 (Final) |
Exit |
|
Frey HC, Small MJ. Integrated environmental assessment, Part I: Estimating emissions. Journal of Industrial Ecology 2003;7(1):9-11. |
R826790 (Final) |
Exit |
|
Frey HC, Zhao Y. Quantification of variability and uncertainty for air toxic emission inventories with censored emission factor data. Environmental Science & Technology 2004;38(22):6094-6100. |
R826790 (Final) |
Exit Exit Exit |
|
Zhao Y, Frey HC. Uncertainty for data with non-detects: air toxic emissions from combustion. ASCE Journal of Environmental Engineering (submitted, 2003). |
R826790 (Final) |
Exit Exit |
|
Zhao Y, Frey HC. Development of probabilistic emission inventories of air toxics for Jacksonville, Florida. Journal of the Air & Waste Management Association 2004;54(11):1405-1421. |
R826790 (Final) |
Exit Exit |
|
Zhao Y, Frey HC. Quantification of variability and uncertainty for censored data sets and application to air toxic emission factors. Risk Analysis 2004;24(4):1019-1034. |
R826790 (Final) R826766 (Final) |
Exit Exit |
|
Zheng J, Frey HC. Quantification of variability and uncertainty using mixture distributions: evaluation of sample size, mixing weights, and separation between components. Risk Analysis 2004;24(3):553-571. |
R826790 (Final) R826766 (Final) |
Exit |
|
Zheng J, Frey HC. Quantitative analysis of variability and uncertainty with known measurement error: methodology and case study. Risk Analysis 2005;25(3):663-675. |
R826790 (Final) |
Exit |
Supplemental Keywords:
urban air toxics, benzene, formaldehyde, chromium, mercury, arsenic, 1,3-butadiene, uncertainty, variability, bootstrap, Monte Carlo, exposure, area sources, particulates, volatile organic compounds, VOC, heavy metals, solvents, pollution prevention, mobile sources, particulate matter, air pollution, air quality, air quality criteria, air quality model, chemical composition, emission control strategies, emission inventories, emission-based modeling, engineering models, hazardous air pollutants, HAPs, heavy metals, probabilistic modeling, quantitative analysis, toxicity, urban air pollution,, RFA, Scientific Discipline, Air, Toxics, Ecosystem Protection/Environmental Exposure & Risk, particulate matter, air toxics, Environmental Chemistry, HAPS, VOCs, Monitoring/Modeling, mobile sources, Environmental Monitoring, 33/50, Environmental Engineering, emission control strategies, urban air toxics, air pollutants, chromium & chromium compounds, Polycylic Organic Matter4, air quality models, emission-based modeling, air quality criteria, benzene, chemical composition, modeling, variability, air pollution, air quality model, toxicity, probabilistic modeling, hazardous air pollutants (HAPs), pollutants, urban air pollution, PM, Volatile Organic Compounds (VOCs), 1, 3-Butadiene, Benzene (including benzene from gasoline), emission inventories, Formaldehyde, heavy metals, engineering models, quantitative analysisProgress and Final Reports:
Original AbstractThe perspectives, information and conclusions conveyed in research project abstracts, progress reports, final reports, journal abstracts and journal publications convey the viewpoints of the principal investigator and may not represent the views and policies of ORD and EPA. Conclusions drawn by the principal investigators have not been reviewed by the Agency.