UNITED STATES ENVIRONMENTAL PROTECTION AGENCY * WASHINGTON D.C. 20460 OFFICE OF THE ADMINISTRATOR SCIENCE ADVISORY BOARD August 22, 2006 EPA-SAB-06-009 The Honorable Stephen L. Johnson Administrator U.S. Environmental Protection Agency 1200 Pennsylvania Avenue, N.W. Washington, D.C. 20460 Subject: Review of Agency Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models and Mode Is Knowledge Base by the Regulatory Environmental Modeling Guidance Review Panel of the EPA Science Advisory Board Dear Administrator Johnson: The EPA Regulatory Environmental Modeling (REM) Guidance Review Panel of the Science Advisory Board has completed its review of the Agency's Council on Regulatory Environmental Models (CREM) Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models, dated November, 2003 (also referred to as the Draft Guidance), and the Models Knowledge Base (MKB), an online database of environmental models. The Panel commends the Agency's REM initiative, which provides a much needed vision for modeling across all EPA programs and offices. The Draft Guidance in particular provides a comprehensive overview of modeling principles and best practices. The Panel notes that the Agency has been very responsive to previous SAB advice on environmental modeling, and recommends that special recognition be accorded to Agency CREM participants for their leadership. However, the Panel is concerned that the CREM activities have been funded through an Adv Hoc approach and the REM vision is not matched by a commensurate, and steady, allocation of resources on the part of the Agency. It is therefore recommended that the Agency provide a meaningful commitment of resources to the REM initiative. The Panel also commends the Agency for recognizing the need for and beginning development on the Models Knowledge Base (MKB). This type of resource has been needed for some time and even in its draft form, the MKB provides an easily accessible resource for the modeling community that, if maintained and used, will significantly improve the development and application of models both internal and external to the Agency. ------- The Panel's report emphasizes a number of ways in which the Draft Guidance andMKB can be improved, including: • Care in articulating the audience to which the Draft Guidance is directed; • The need to develop and apply models within the context of a specific problem; • Caution in the way that information on modeling uncertainty is evaluated and communicated, and the need for the Draft Guidance to more fully discuss uncertainty and sensitivity analysis methods; • More consistency in conforming the terminology used in the Draft Guidance and MKB to previous uses and meanings through the REM Glossary; and • The need to gather, and in many cases to develop, additional information to be included in the MKB, including the framework, evaluation, and limitations of models included, and to implement a mechanism within the MKB that allows the community of users to submit feedback on their experiences. In summary, the SAB finds that the Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models is an important document, and the Mode Is Knowledge Base an important tool that will guide the Agency and others in developing and using models for environmental purposes. In the Panel's judgment it is essential that these efforts be revised and updated regularly in order for their full value to the Agency to be realized. The Panel stands ready to provide additional advice and review on this effort as it continues to progress. Sincerely, /signed/ /signed/ Dr. M. Granger Morgan Dr. Thomas L. Theis Chair Chair Science Advisory Board REM Guidance Review Panel Science Advisory Board ------- NOTICE This report has been written as part of the activities of the EPA Science Advisory Board (SAB), a public advisory group providing extramural scientific information and advice to the Administrator and other officials of the Environmental Protection Agency. The SAB is structured to provide balanced, expert assessment of scientific matters related to problems facing the Agency. This report has not been reviewed for approval by the Agency and, hence, the contents of this report do not necessarily represent the views and policies of the Environmental Protection Agency, nor of other agencies in the Executive Branch of the Federal government, nor does mention of trade names of commercial products constitute a recommendation for use. Reports of the SAB are posted on the EPA website at http;//www\epa,,gQWgab, ------- U.S. Environmental Protection Agency Science Advisory Board Regulatory Environmental Modeling (REM) Guidance Review Panel CHAIR Dr. Thomas L. Theis, Professor, Civil and Materials Engineering, and Director, Institute for Environmental Science & Policy, University of Illinois at Chicago, IL MEMBERS Dr. Linfield C. Brown, Professor, Civil and Environmental Engineering Department, Tufts University, Medford, MA Dr. Joseph DePinto, Senior Scientist, Limno-Tech, Inc., Ann Arbor, MI Dr. Panos Georgopoulos, Professor, Environmental and Occupational Medicine, UMDNJ - Robt Wood Johnson Medical School, Piscataway, NJ Dr. Steven Heeringa, Research Scientist, Institute for Social Research (ISR), University of Michigan, Ann Arbor, MI Dr. Bruce K. Hope, Senior Environmental Toxicologist, Oregon Department of Environmental Quality (DEQ), Air Quality Division, Portland, OR Dr. Alan J. Krupnick, Senior Fellow & Director, Quality of the Environment Division, Resources for the Future, Inc., Washington, DC Dr. Randy L. Maddalena, Scientist, Environmental Energy Technologies Division, Indoor Environment Department, Lawrence Berkeley National Laboratory, Berkeley, CA Dr. June Fabryka-Martin, Staff Scientist, Los Alamos National Laboratory, Los Alamos, NM Mr. David Merrill, Principal, Gradient Corporation, Cambridge, MA Dr. Paulette Middleton, President, Panorama Pathways, Inc., Boulder, CO Dr. Mitchell J. Small, The H. John Heinz III Professor of Environmental Engineering, Department of Civil & Public Policy, Carnegie Mellon University, Pittsburgh, PA Dr. Douglas G. Smith, Principal Environmental Health Scientist, Risk Assessment Department, ENSR International, Inc., Westford, MA Dr. James H. Smith, Lead Technical Project Manager, State Photochemical Smog Group, Texas Commission on Environmental Quality, Austin, TX 11 ------- Dr. Richard L. Wetzel, Chair, Department of Biological Sciences and Professor of Marine Science, Virginia Institute of Marine Sciences (VIMS), College of William & Mary, Gloucester Pt, VA Dr. Peter Wilcoxen, Associate Professor of Economics and Public Administration, Syracuse University, Syracuse, NY SCIENCE ADVISORY BOARD STAFF Dr. K. Jack Kooyoomjian, Designated Federal Officer, US EPA Science Advisory Board, Washington, DC in ------- U.S. Environmental Protection Agency Science Advisory Board CHAIR Dr. M. Granger Morgan, Carnegie Mellon University, Pittsburgh, PA SAB MEMBERS Dr. Gregory Biddinger, ExxonMobil Biomedical Sciences, Inc, Houston, TX Dr. James Bus, The Dow Chemical Company, Midland, MI Dr. Trudy Ann Cameron, University of Oregon, Eugene, OR Dr. Deborah Cory-Slechta, University of Medicine and Dentistry of New Jersey and Rutgers State University, Piscataway, NJ Dr. Maureen L. Cropper, University of Maryland, College Park, MD Dr. Virginia Dale, Oak Ridge National Laboratory, Oak Ridge, TN Dr. Kenneth Dickson, University of North Texas, Denton, TX Dr. Baruch Fischhoff, Carnegie Mellon University, Pittsburgh, PA Dr. A. Myrick Freeman, Bowdoin College, Brunswick, ME Dr. James Galloway, University of Virginia, Charlottesville, VA Dr. Lawrence Goulder, Stanford University, Stanford, CA Dr. Rogene Henderson, Lovelace Respiratory Research Institute, Albuquerque, NM Dr. Philip Hopke, Clarkson University, Potsdam, NY Dr. James H. Johnson, Howard University, Washington, DC Dr. Meryl Karol, University of Pittsburgh, Pittsburgh, PA Dr. Catherine Kling, Iowa State University, Ames, IA Dr. George Lambert, Robert Wood Johnson Medical School/ University of Medicine and Dentistry of New Jersey, Piscataway, NJ Dr. Jill Lipoti, New Jersey Department of Environmental Protection, Trenton, NJ IV ------- Dr. Genevieve Matanoski, Johns Hopkins University, Baltimore, MD Dr. Michael J. McFarland, Utah State University, Logan, UT Dr. Jana Milford, University of Colorado, Boulder, CO Dr. Rebecca Parkin, The George Washington University, Washington, DC Mr. David Rejeski, Woodrow Wilson International Center for Scholars, Washington, DC Dr. Joan B. Rose, Michigan State University, E. Lansing, MI Dr. Kathleen Segerson, University of Connecticut, Storrs, CT Dr. Kristin Shrader-Frechette, University of Notre Dame, Notre Dame, IN Dr. Robert Stavins, Harvard University, Cambridge, MA Dr. Deborah Swackhamer, University of Minnesota, Minneapolis, MN Dr. Thomas L. Theis, University of Illinois at Chicago, Chicago, IL Dr. Valerie Thomas, Georgia Institute of Technology, Atlanta, GA Dr. Barton H. (Buzz) Thompson, Jr., Stanford University, Stanford, CA Dr. Robert Twiss, University of California-Berkeley, Ross, CA Dr. Terry F. Young, Environmental Defense, Oakland, CA Dr. Lauren Zeise, California Environmental Protection Agency, Oakland, CA SCIENCE ADVISORY BOARD STAFF Mr. Thomas Miller, Designated Federal Officer, US EPA Science Advisory Board, Washington, DC ------- Table of Contents EXECUTIVE SUMMARY 1 BACKGROUND 4 1. BEST PRACTICES 8 1.1. INTERPRETATION OF "BEST AVAILABLE AND PRACTICABLE SCIENCE" 8 1.2. GENERAL COMMENTS 8 1.3. PROBLEM SPECIFICATION 9 1.4. MODEL CALIBRATION AND SENSITIVITY ANALYSIS 10 1.5. MODEL POST-AUDIT 11 1.6 DOCUMENT ORGANIZATION 11 2. GOALS AND METHODS 14 2.1. INTRODUCTION 14 2.2. INTENDED AUDIENCE AND SCOPE OF USE 14 2.3. GLOSSARY 16 2.4. MODEL DOCUMENTATION, PROJECT DOCUMENTATION, AND USER MANUAL 17 3. GRADED APPROACH 18 3.1. DEFINITION OF "GRADED APPROACH" 18 3.2. MODELING COMPLEXITY AND ASSOCIATED EVALUATION NEEDS 18 3.3. EVALUATING MODEL RESPONSE 19 3.4. USE OF MULTIPLE AND LINKED MODELS 20 3.5. USE OF MODEL-DERIVED DATA 21 4. PRACTICAL ADVICE FOR DECISION-MAKERS 22 4.1. GENERAL COMMENTS ON UNCERTAINTY 22 4.2. SENSITIVITY ANALYSIS VIS-A-VIS UNCERTAINTY ANALYSIS 24 4.3. UNCERTAINTY ANALYSIS PRACTICES/METHODS (REM GUIDANCE SECTION C.6) 25 4.4. VALUE OF INFORMATION - IDENTIFYING "UNCERTAINTIES THAT MATTER" 27 4.5. COMMUNICATING UNCERTAINTY 28 5. IDENTIFICATION AND STRUCTURE OF OPTIMAL SET OF INFORMATION FOR ALL USERS 30 5.1. GENERAL COMMENTS AND SUGGESTIONS 30 5.2. SPECIFIC SUGGESTIONS BY THE PANEL 33 5.4. LISTING OF KEY PUBLICATIONS AND APPLICATIONS OF MODELS 35 5.5. CLARIFICATION OF MKB ENTRY SHEET ITEMS C1-C3 35 6. DATA DICTIONARY AND DATA STRUCTURE 37 6.1. GENERAL COMMENTS 37 6.2. MODEL PERFORMANCE INFORMATION 38 6.3. ADDITIONAL RECOMMENDATIONS 39 7. QUALITY OF INFORMATION PROVIDED ABOUT THE MODELS 42 7.1 GENERAL COMMENTS 42 7.2 VISION FOR THE KNOWLEDGE BASE 43 7.3 QUALITY ASSURANCE AND QUALITY CONTROL 44 7.4 LAYOUT AND NAVIGATION OF KNOWLEDGE BASE 46 7.5 UPDATING THE KNOWLEDGE BASE 46 7.6 THE ROLE OF THE KNOWLEDGE BASE AS A "MODEL SELECTION TOOL" 47 REFERENCES 48 VI ------- APPENDIX A - ENHANCEMENTS TO THE GLOSSARY 52 APPENDIX B - THE CREM MODELS KNOWLEDGE BASE DATA ENTRY SHEET 54 APPENDIX C - PANEL MEMBERS EXPERIENCES USING THE MKB 56 C-l CALPUFF (THE ILLUSTRATIVE Am MODEL) 56 C-2 THE INTEGRATED PLANNING MODEL (TPM - THE ILLUSTRATIVE ECONOMIC MODEL) 59 C-3 AQUATOX (THE ILLUSTRATIVE WATER QUALITY MODEL) 60 C-4 OTHER MODELS 61 APPENDIX D-ACRONYMS 63 vn ------- EXECUTIVE SUMMARY The Regulatory Environmental Modeling (REM) Panel of the SAB has reviewed the Agency's Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models, dated November, 2003 (referred hereafter as the Draft Guidance), and the Agency's Mode Is Knowledge Base (referred to as the MKB). Major points of consensus are summarized below.1 The Panel commends the Agency's REM initiative, which provides a much needed vision for modeling across all EPA offices. The Draft Guidance in particular provides a comprehensive overview of modeling principles and best practices. The Panel notes that the Agency has been very responsive to previous SAB advice on environmental modeling, and recommends that special recognition be accorded to Agency REM participants for their leadership. The Panel believes that the Regulatory Environmental Models (REM) program at EPA will provide leadership and guidance for improving the quality of model development, evaluation, and application in the use of environmental models for decision support. As a part of this program, the MKB will provide a web-based database of information on selected models, including key operational and scientific features, model downloads, guidance for use, and examples of model applications provided by model developers. Nevertheless, the Panel is concerned that the activities of EPA's Council on Regulatory Environmental Modeling (CREM) have been funded through an Ad Hoc approach and the REM vision is not matched by a commensurate, and steady, allocation of resources. It is therefore recommended that the Agency provide a meaningful commitment of resources to the REM initiative. The Draft Guidance is comprehensive, and will most likely be read and used by a wide variety of audiences including model developers, analysts, managers at various levels, decision- makers, and other stakeholders who come from Federal, State, and private sectors. Yet it is written, and most comprehensible, primarily to those who develop and/or those who "use" or run models to generate output. Accordingly the Panel recommends that the Agency clarify carefully the use of the Draft Guidance for a variety of audiences, describing or suggesting how it can be used beneficially by different participants in a modeling project. In the same vein, the Panel finds that the use of modeling terminology is sometimes inconsistent with Agency past uses, or usage common in the modeling community. It is recommended that these inconsistencies be recognized through developing and using a common reference, the Glossary, in which these and other terms are carefully defined. The current Glossary in the Draft Guidance should be expanded to make it as comprehensive as possible. 1 This report contains the consensus views of the REM Panel on the current state of the REM program within the Agency, as presented in the Draft Guidance and the MKB documents. The report is organized by responses of the Panel to charge questions posed by the Agency. Generally speaking, each set of responses consists of statements and explanatory materials that present the Panel's point of view on a given topic, which are followed by formal recommendations, or in some cases commendations. For ease in discerning the plain meanings and actions of the Panel, these recommendations and commendations are boldened. Less urgent, but still important observations, suggestions, and concerns are not boldened, and/or are contained in the appendices of the report. ------- In the Panel's view it is important that the specifics of the problem posed be explicitly stated and agreed upon by all stakeholders, and be used to guide model conceptual development, complexity, data needs, and interpretation of output. Toward this end, the Panel suggests an alternative version of Figure 1 (page 7) in the Draft Guidance in which Problem Specification is given greater emphasis (page 12 in this review). The Panel believes that this alternative figure better reflects the central role of stakeholders in the public policy process, and provides a more accurate representation of the modeling process and its iterative nature. As noted in the Draft Guidance the evaluation of uncertainty in the application of models is an important element in both understanding a system and in presenting results to decision- makers, a point with which the Panel concurs. Indeed the use of Quantitative Uncertainty Assessment (QUA) methods is a desirable, and often necessary component of modeling, but experience suggests that the use of increasingly complex QUA techniques without an equally sophisticated framework for decision-making and communication may only increase management challenges. Accordingly the Panel recommends that the Draft Guidance strongly advise modelers to select particular QUA methods only after becoming aware of how the decision-maker plans to use the information on uncertainty. This is an important component of the Problem Specification as well. The Panel finds that the Draft Guidance provides a generally adequate discussion of sensitivity analysis methods; however it is deficient in articulating a more tangible set of alternatives for assessing model uncertainty, and a clearer distinction between sensitivity and uncertainty analysis. While references cited provide an array of applicable methods to address model uncertainty, the Draft Guidance does not provide sufficient discussion, context, and recommendations necessary to provide a model user/decision-maker with "practicable" information relating to appropriate uncertainty analysis methods and how to convey the results of such analyses. In addition, recommendations for uncertainty analysis could identify focusing resources on those processes to which the model state variables are most sensitive and are less certain in terms of their formulation and/or parameterization. The topic of propagation of uncertainty in modeling frameworks relying upon linked models, is not addressed in the Draft Guidance, and warrants specific discussion. The Panel also recommends that both the Draft Guidance and the MKB provide more practicable information through inclusion of "case study" examples of where and how EPA is currently incorporating QUA in environmental models as an integral component of decision-making. The Panel commends the Agency for recognizing the need for and beginning development on the MKB. This type of resource has been needed for some time and even in its draft form, it provides an easily accessible resource for the modeling community that, if maintained and used, will significantly improve the development and application of models both internal and external to the Agency. In its review of the MKB, the Panel arrived at several suggestions for modifying the data entry sheet that are given in our response to Charge Question 5. Perhaps the most important recommendation is the need to clarify and in some cases gather additional information on models including their framework (which in the Panel's opinion needs to be redefined), evaluation, and limitations. The Model Evaluation section of the Model Science MKB information page considers many of the key issues needed to evaluate the scientific rigor behind the underlying model development and previous applications, and addresses many of the ------- elements of good modeling practice that are emphasized in the Draft Guidance. Indeed, the Panel views an important purpose of the MKB as providing an incentive for model developers and stakeholders to conduct and openly communicate their efforts in model evaluation. From this perspective, the Panel recommends some additional pieces of information that should be elicited and reported, including: 1) Documented examples of peer review for the model, including reviews conducted by the EPA, other agencies or panels, and papers presented in the peer reviewed literature. Key limitations and needs for improvement that were identified in these evaluations should be reported; 2) Benchmarking studies in which the model's predictions and/or accuracy were compared with other models; 3) Provision of a mechanism that actively solicits feedback from the user community regarding application experience and model performance, both inside and outside the agency, beyond voluntary e-mails to designated contacts for individual models; and 4) Information on revision tracking, which should be incorporated into the MKB. The Panel also recommends that the Agency follow its own standard QA/QC program procedures for ensuring quality of all of the underlying information in the MKB system. A meaningful commitment to QA/QC is necessary to ensure the quality of information in the MKB, without which it is doubtful the MKB will achieve its potential value and utility. This QA/QC function will require the allocation of an appropriate level of resources on the part of the Agency. Finally, this report contains specific experiences of Panel members (Appendix C) on the use of the MKB for three specific models that it contains. These experiences can help guide efforts by the Agency as they continue to modify the MKB in the future. ------- BACKGROUND2 The impetus for much of the Council for Regulatory Environmental Modeling's (CREM) current activities derives from the Data Quality Act and the Act's requirement that EPA and other executive agencies establish mechanisms to allow the public to raise questions about information they issue.3 Because environmental models and their analytical results were generally construed to fall within the Act's ambit, EPA's Administrator charged the CREM to establish guidelines to clarify the Agency's views on model quality.4 While the Data Quality Act was passed in 2000, the history of EPA's and the SAB's interest in the nexus between policy and environmental models actually date back a few decades, as described in the following paragraphs. In December 1984, the Chairman of the Executive Committee of the SAB first addressed the issue of best modeling practices in a letter to the EPA Administrator, recommending that a "systematic effort of model validation be initiated, including identification of the appropriate balance between monitoring and modeling." In 1989, the SAB's Environmental Engineering Committee, noting common problems among the models brought before the Committee for review, recommended that EPA establish "a central coordinating group within the EPA to assess the status of environmental models currently used or proposed for use in regulatory assessment, and to provide guidance in model selection and use by others in the Agency."5 In subsequent years, SAB addressed a variety of modeling issues, such as the need for generic models to account for site-specific circumstances,6 and the need to conduct sensitivity and uncertain analyses to better characterize modeling uncertainties.7 Among the efforts to respond to SAB suggestions, EPA established an ad hoc committee in 1992 to address challenges related to generating and using models. This committee, the Agency Task Force on Environmental Regulatory Modeling (ATFERM), produced guidance on 2 CREM Background Materials: A web version of the CREM related background information, with links to pertinent documents, is available at www .cpa. gov/crcm/sab. 3 U.S. Congress. 2001. Pub. L. No. 106-554. 2001. The Data Quality Act, Section 515 of the Treasury and General Government Appropriations Act for Fiscal Year 2001, Pub. L. No. 106-554. 4 U.S. EPA. 2003a. Council for Regulatory Environmental Modeling, Administrator Memorandum, 2003. Available at http://www.epa.gov/osp/crem/library/whitman.PDF. 5 U.S. EPA. SAB. 1989. Resolution on the Use of Mathematical Models by EPA for Regulatory Assessment and Decision-Making, by the Modeling Resolution Subcommittee of the Environmental Engineering Committee, Science Advisory Board, EPA-SAB-EEC-89-012, January 13, 1989. Available at http://www.epa. gov/osp/creiii/librai3f/sab_89iEsolirtion_iiiodels.pdf. 6 U.S. EPA. SAB. 1990. Review of the CANSAZ Flow and Transport Model for Use in EPACMS, Report of the Saturated Zone Model Subcommittee of the Environmental Engineering Committee (EEC), Science Advisory Board 1990, EPA-SAB-EEC-90-009, March 27, 1990. Available at http://www.epa.gov/osp/creni/libraiy/sab_cansaz.pdf. 7 U.S. EPA. SAB. 1995. Commentary on Bio accumulation Mo deling Issues, Report from Bioaccumulation Subcommittee, Science Advisory Board, EPA-SAB-EPEC/DWC-COM-95-006, September 29, 1995. Available at tep://www.epa.gov/osp/crem/libraiy/sab_bioaccumulation.pdf. 4 ------- the peer review of models, suggested model acceptability criteria, and proposed a charter for a Council for Regulatory Environmental Modeling (CREM). In 1999, the SAB recommended that EPA establish policies and procedures for the development, validation and use of environmental regulatory models. SAB further suggested that EPA should collaborate with model users not just inside the Agency, but outside as well, seeking their feedback to continually improve model development and use. In February 2000, EPA's Administrator formally established the CREM to continue the initiatives toward building consensus and consistency in modeling efforts by the Agency.9 In February 2003, the Administrator stated her expectations for the CREM to lead EPA in, among other things: 1. providing "guidance for the development, assessment, and use of environmental models;" and 2. making "publicly accessible an inventory of EPA's most frequently used models, which will include information on a model's use, development, validation, and quality assessment." It is with regard to these two items that the CREM has now turned to the SAB's Regulatory Environmental Modeling Guidance Review Panel for advice. Specifically, the CREM has submitted the following charge questions to the Panel. Specific Charge Questions Charge Question 1: Has EPA sufficiently and appropriately identified the best practices, such that decisions based on models developed and used in accordance with these practices may be said to be based on the best available, practicable science? Charge Question 2: Has EPA sufficiently and appropriately described the goals and methods, and in adequate detail, such that the guidance serves as a practical, relevant, and useful tool for model developers and users? If not, what else would you recommend to achieve these ends? Charge Question 3: Has EPA sufficiently and appropriately proposed a graded approach, such that users of the guidance can determine the appropriate level of evaluation for a particular model use. If there are deficiencies in the proposed approach, what would you recommend to correct it, and why? U.S. EPA. 1994. Agency Guidance for Conducting External Peer Review of Environmental Regulatory Modeling, 1994. Available at http://cfpiib.epa.gov/crem/niodelpr.cfiii. 9 U.S. EPA. 2000. Framework for the Council on Regulatory Environmental Modeling, Available at http://www.epa.gov/osp/crem/libraiy/crem%20franiework.htm. ------- Charge Question 4: Has EPA sufficiently and appropriately provided practicable advice for decision-makers who must deal with the uncertainty inherent in environmental models and their application? What additional advice should EPA consider in dealing with uncertainty, and why? A number of researchers recommend a Bayesian approach to help decision-makers incorporate uncertainty into their decisions and to do so in a transparent fashion (see, e.g., Attachments B and C). Is the use of methods such as Bayesian networks an effective and practicable way for EPA decision-makers to incorporate uncertainty within their decisions and to communicate this uncertainty to stakeholders? If so, how? Are there alternative methods available? Models Knowledge Base: As noted above, the SAB recommended that the CREM coordinate EPA efforts to collaborate and seek input from model developers and users both inside and outside EPA. One mechanism to implement this collaboration is through a web-accessible knowledge base for environmental models. EPA has developed such a knowledge base to communicate more clearly the data, algorithms, assumptions, and uncertainties underlying each model; to facilitate the use of individual models or the combined use of multiple models; and to enable developers and analysts to more easily identify information needs. Charge Question 5: The Panel should consider that environmental models will be used by people whose technical sophistication will vary widely. EPA has therefore attempted to cull information about models that broadly serve the needs of all users, using a data template to collect this information (see Attachment D). Has EPA identified, structured and developed the optimal set of information to request from model developers and users, i.e., the amount of information that best minimizes the burden on information providers while maximizing the utility derived from the information? Charge Question 6: EPA has developed a data dictionary and database structure to organize the information it has collected on environmental models (see Attachments E and F). Has EPA provided the appropriate nomenclature needed to elicit specific information from model developers that will allow broad intercomparisons of model performance and application without bias toward a particular field or discipline? Charge Question 7: To facilitate review for this particular charge question, the Panel should focus on three models that represent the diversity of model information housed within the Models Knowledge Base. These models are: (1) Aquatox, a water quality model, with information found at hjtp;//cfj)^ (2) Integrated Planning Model, a model to estimate air emissions from electric utilities, with information found at httpj/Mpjjb.,.epa^^^ ancj NWPCAM, an economic model with information at http://cfpub.epa.gov/crem/crem report.cfm?deld=74918.10 10 The final model selections from the MKB for observation and examination by the Panel include CALPUFF (The Illustrative Air Model - see Appendix C-l in this Report); IPM (Integrated Planning Model - The Illustrative Economic Model - see this Appendix C-2 in this Report); and AQUATOX (The Illustrative Water Quality Model - see Appendix C-3 in this Report). Other models are discussed generally in Appendix C-4 of this Report. ------- Using these three models as examples and emphasizing that EPA is not seeking a review of the individual models, but rather the quality of the information provided about the models, EPA poses the following questions to the Panel. Through the development of this knowledge base, has EPA succeeded in providing: (7a) Easily accessible resource material for new model developers that will help to eliminate duplication in efforts among the offices/regions where there is overlap in the modeling efforts and sometimes communication is limited? (7b) Details of the temporal and spatial scales of data used to construct each model as well as endogenous assumptions made during model formulation such that users may evaluate their utility in combination with other models and so that propagation of error due to differences in data resolution can be addressed? (7c) Examples of "successful" models (e.g., widely applied, have been tested, peer reviewed etc.)? (7d) A forum for feedback on model uses outside Agency applications and external suggestion for updating/improving model structure? ------- 1. BEST PRACTICES Charge Question 1: Has EPA sufficiently and appropriately identified the best practices, such that decisions based on models developed and used in accordance with these practices may be said to be based on the best available, practicable science? 1.1. Interpretation of "Best Available and Practicable Science" In developing and applying a model for supporting a regulatory action or decision, it is important to meet the criterion stated in Charge Question 1--"based on the best available, practicable science'' To the Panel, this means that the model uses the best current science that is consistent with the model's intended use, whether that use is regulatory, management or scientific. The term "practicable" refers to consideration of problem specification and programmatic constraints (data quality and availability, and limitations of time and resources) in selection of model complexity (i.e., spatial, temporal, and process resolution). Thus in the context of Figure 2 (page 11) of the Draft Guidance document, the Panel suggests that the location of the minimum (both in the x- and y-directions) in the uncertainly versus model complexity curve will depend on the problem specification and programmatic constraints. The Panel believes that when a model complexity is most appropriate for the problem and available data and resources, it is obtaining the minimum possible uncertainty and, hence, using the best available, practicable science. The Panel interprets this question as asking whether the Guidance aids the modeler in finding that level of model complexity. 1.2. General Comments In general, the Panel finds the REM initiative provides a common and much needed vision for modeling across all of the offices within the Agency. The draft document in particular provides a comprehensive overview of modeling principles and best practices, in a concise manner. The Panel also finds that the Agency has been responsive to previous SAB advice on modeling practices and commends the REM participants for their leadership. In particular the Panel applauds the emphasis in the document on using the peer review process to insure that a Regulatory Environmental Model is using the best available, practicable science However, the Panel is concerned that the CREM activities have been funded through an Ad Hoc approach and the REM vision is not matched by a commensurate, and steady, allocation of resources on the part of the Agency. It is therefore recommended that the Agency provide a meaningful commitment of resources to the REM initiative. The Panel believes that successful implementation of this recommendation will require a commitment from the top of Agency management, will require institutional change in the Agency, will take significant time to implement, and will require the establishment of a formal institutional mechanism responsible for review, oversight and coordination of model use in EPA. ------- The Panel encourages the Agency to recommend that any regulatory modeling project include peer review as part of its Quality Assurance Project Plan (QAPP). Furthermore, the Panel suggests that the peer review plan implement ongoing peer review through all stages of the modeling process, not just after the model application. Such a proactive practice will assist in avoiding technical errors or omissions that are often difficult or impossible to rectify after the project is over. Also, the Panel favors an open modeling process for Regulatory Environmental Models, in which modeling decisions and results are shared with stakeholders through model development and application. This practice avoids a situation where the model fails to address the regulatory questions as conceived by the various stakeholders in the process. Consistent with the above discussion concerning ongoing peer review and interaction between modelers and stakeholders and to reflect the recommendations of the Panel presented in more detail below, the Panel suggests an Alternative Figure 1 to the EPA's Figure 1 shown in the Draft Guidance (U.S. EPA. 2003). The Alternative Figure 1 represents the same general logic and information flow provided in the EPA's original Figure 1, but it has been amended to enhance the detail of some of the particular steps. It has also been expanded to represent the Panel's perception of the interaction with stakeholders in both the identification and specification of the problem to be solved and in the ongoing review of the quality of the regulatory modeling tools. 1.3. Problem Specification The Panel appreciates the distinctions made in the Draft Guidance between model framework development and model application. Nevertheless, the Panel finds that this distinction is not consistently maintained throughout the document. For example, the terms "application tool" in Section 2 means problem-specific model implementation whereas "model application" in Section 4 means model-based decision making. The Panel recommends that the term application tool be replaced with "problem-specific implementation." The Panel believes that Problem Specification is a critical element of any modeling project. It guides the development of the conceptual model and it governs the model complexity. It must, therefore, include a clear and complete statement of policy, management, and/or scientific objectives, model spatial and temporal domain and resolution characteristics, as well as program constraints (e.g., legal, institutional, data, time and costs). This process must involve interactions among all stakeholders. The Panel recommends that Problem Specification be given greater emphasis in the Draft Guidance by elevating it to a separate, initial step in the modeling process, as shown in the Alternative Figure 1 offered below. In accordance with this observation, the Panel offers the following suggestions that should be included for completeness and clarity in the expanded problem specification portion of the Draft Guidance for each of the above aspects of problem specification: ------- 1) Regulatory or research objectives are statements of what questions a model has to answer. The statement of modeling objectives should include the state variables of concern, the stressors (model inputs) driving those state variables and their control options, appropriate temporal and spatial scales, user acceptance of the model, and very importantly, the degree of accuracy and precision of the model. The discussion of Data Quality Objectives (DQOs) in the document is good, but the relationship between total uncertainty, accuracy, and precision of the model needs to be further clarified. 2) An alternative description of model types as a component of problem specification should compare and contrast: empirical vs. mechanistic, static vs. dynamic, simulation vs. optimization, deterministic vs. stochastic, lumped vs. distributed. 3) Specifying the model domain characteristics includes: identification of the environmental domain being modeled; specification of transport and transformation processes within that domain that are relevant to the policy /management/research objectives; specification of important time and space scales inherent in transport and transformation processes within that domain in comparison with the time and space scales of the problem objectives; and any peculiar conditions of the domain that will affect model selection or new model construction. 4) Problem specification should include a discussion of the potential programmatic constraints. These address time and budget, available data or resources to acquire more data, legal and institutional considerations, computer resource constraints, and experience and expertise of the modeling staff. These factors, collectively, allow the modeler to determine the "complexity" of a model that is necessary and sufficient for the application under consideration (see recommended definition of model complexity in Charge Question 2 response). 1.4. Model Calibration and Sensitivity Analysis The Panel applauds the overall treatment of model quality assurance and evaluation in Appendices B and C of the Draft Guidance. However, the Panel recommends that the process of "model calibration" receive increased attention regarding guiding principles and best practices, both in the main text of the document and in the appendices. While calibration of air models may not always be feasible or justified, it is an integral part of water quality modeling and one of the more poorly understood steps in the modeling process. For example, the document could discuss how sensitivity analysis can be used during the calibration process. Most process-oriented environmental models are underdetermined; that is, they contain more uncertain parameters than state variables that can be used to perform a calibration. Therefore, good model calibration practice uses sensitivity analysis to determine key processes for a given problem-specific implementation and then recommends empirical determination of 10 ------- the rate of those key processes as part of the calibration process in addition to measuring the time and space profile of relevant state variables. This practice can help further constrain a model for which parameterization by calibration (i.e. ground-truthing with empirical data or statistical techniques with data to estimate unknown parameters) is difficult. An example of this practice would be to measure the rate of photosynthesis (process) in a lake in addition to the biomass of phytoplankton (state variable). 1.5. Model Post-Audit The practice of model post-auditing is defined as the ongoing observation of the response of the system to the actual implementation of a policy or management action relative to the model's forecast of how that system would respond, and is crucial to the ongoing improvement of environmental models. The Panel recommends that the Draft Guidance acknowledge the value of post-auditing of models and associated data collection. This practice deserves a section of its own in the model application area (note the addition of a reference to post- auditing in Alternative Figure 1). That section should also discuss the role of regulatory modeling in adaptive management of environmental systems. 1.6 Document Organization The Panel believes that there are best practices for the development of a generic model framework (for example, WASP, QUAL2E, and AQUATOX) however most users of the Draft Guidance will not be model developers. Therefore, the document should contain additional best practices that should be followed for a site-specific or problem-specific implementation of a model framework. In order to clarify the guiding principles that should be considered for each type of project, the Panel recommends that the Agency consider organizing the Draft Guidance according to the steps involved in carrying out a modeling project from inception to completion as indicated in Alternative Figure 1. The Panel identifies the steps in Alternative Figure 1 to be: Problem Specification; Model Identification/Selection (the document should recognize that a site-specific modeling project may be conducted by either new model construction or by selection of an existing model framework); Model Development (including problem- and site-specific model conceptualization, model formulation and configuration, and model calibration); Model Evaluation (through peer review, data quality assessment, model code verification, model confirmation/corroboration, sensitivity analysis, and uncertainty analysis); Model Application (including diagnostic analysis,11 problem 1: Diagnostic use of models has great value for both model evaluation and problem-specific application. For example, plotting the cumulative distribution of observations of a state variable on the same plot as the cumulative distribution of model computation of that state variable on the same spatial and temporal scale is valuable for identifying whether the model is biased at high or low concentrations. As another example, development of a model mass balance diagram of a given state variable over appropriately chosen space and time scales (e.g., whole lake water column over the course of a year) is useful for identifying significant mass flow pathways, for addressing specific management questions, and for helping to guide monitoring programs. 11 ------- solution, and application support for decision-making); and, after implementation of a regulatory action, Model Post-Audit. These activities should be covered in a QAPP for any given modeling project. Furthermore, the entire Modeling Process should be detailed in a report that includes documentation of all of the above steps in the process. 1 • for the t k. Environmental ;^ Controls | i t tf . I I » of » and »Courts and gcsfiinniifrt (i,|*, • Advocacy flroups (o.ij.j pubic, fttwiistwiirtilj industry, cw On the left side, a few additional elements of the "public policy process" are suggested to clarify the stages of modeling that occur after a decision is made including the use of monitoring programs and post audit reviews of the outcome of previous or new regulatory actions to support model improvement. Alternate Figure 1 also expands on the role of models in supporting regulatory decisions, identifying needed environmental controls, and implementing these controls through enforcement actions when necessary. In addition, the centralized interactive role of all types of stakeholders is emphasized. These stakeholders include source facility owners and other responsible parties, neighboring property owners and other directly affected members of the public, courts and interested government agencies or related entities, and advocacy groups representing various environmental, industry, or trade organizations. 12 ------- The expanded format for the right side of the diagram illustrating the Environmental Model Development and Application Process also maintains the same basic logic of the original EPA Figure 1; but the individual steps have been expanded somewhat including details for problem specification, model selection, model calibration and uncertainty analysis to represent the recommendations of the Panel. Finally, the added emphasis provided by the addition of the continuous "Model Review Process" emphasizes the strong support of the Panel for the processes already occurring in much of the REM development program. The Panel commends the continued and expanded application of this model review process to the further development of the Models Knowledge Base. 13 ------- 2. GOALS AND METHODS Charge Question 2: Has EPA sufficiently and appropriately described the goals and methods, and in adequate detail, such that the guidance serves as a practical, relevant, and useful tool for model developers and users? If not, what else would you recommend to achieve these ends? 2.1. Introduction The general goals of the Draft Guidance are clearly stated (page 6), i.e., to provide guidance on how to assess the quality of regulatory environmental modeling. The assessment is to be made on the basis of a number of "performance criteria" or "specifications" (page 3) that characterize the three major components of regulatory environmental modeling; namely (1) model development, (2) model evaluation, and (3) model application. The Draft Guidance provides specific (and alternative) methods by which the performance criteria for each of these three components may be assessed. The Panel agrees that the Draft Guidance is an excellent start to defining the process of and providing the measurement tools for quality assurance in regulatory environmental modeling. Furthermore, the Panel makes particular note of the critical importance of problem specification at the beginning of any modeling project Problem specification supplies the modeling objectives and constraints that thereafter guide implementation of the modeling steps described in the Draft Guidance. 2.2. Intended Audience and Scope of Use The Draft Guidance identifies the intended audience as being composed of two general categories: model developers and model users. Upon closer reading, however, other important modeling constituencies are explicitly or implicitly identified, each with distinctly different roles in the modeling process, leading the Panel to conclude that the term "model user" is overly broad and imprecise. For this reason, the Panel is concerned that the Draft Guidance elaborate on the distinction between the model users who generate model output (those who setup, parameterize, run, calibrate, etc, particularly with model framework software such as WASP or QUAL2E), and those who are managers and are principally users of model output. They are both users, but play different roles in regulatory environmental modeling, and as such are likely to use this Draft Guidance to assess different quality criteria. It would also help to clarify the intent of the Draft Guidance and its relationship to its different regulatory audiences (at least 2 groups): regulatory decision makers, and regional and state "assessors'Vadvisors for permit applicants. Panel discussions also suggested including other stakeholders in this audience, e.g., those to whom the results will apply or affect. For less experienced audiences, the Draft Guidance may be insufficiently explanatory. The Panel recommends that the Agency clarify the use of the 14 ------- Draft Guidance for the variety of intended audiences and suggests that the Agency identify which sections will be most useful to the various stakeholders in a modeling project. A general concern about the overall Draft Guidance is its scope of use. The Panel finds that it provides a valuable resource to modelers in a wide range of disciplines, but unlike typical EPA guidance documents, it does not lay out a step-by-step course of action. Instead, it identifies a set of key "best practices" which should be adhered to, along with supporting materials Because this Draft Guidance differs in scope and content from other "guidance," and because the term "guidance" has specific connotations in certain areas of model application, the Panel suggests that EPA consider using a term such as "guiding principles" instead of "guidance," both in the body of the Draft Guidance and in its title. A second general issue related to the scope of the Draft Guidance is that much of the introductory parts of the Draft Guidance refer exclusively to regulatory applications of models, yet it is clear that the intent of the REM process is to bring consistency to all environmental applications of models, (e.g., regulatory support, research, resource assessment, evaluating alternative management actions, economic evaluations, etc.). Therefore, the Panel recommends that the Draft Guidance, including its stated purpose, be revised to reflect these additional uses. A final issue regarding scope concerns the types of models to which the Draft Guidance is intended to apply. The executive summary states "this Guidance provides recommendations for environmental models drawn from Agency white papers, EPA's Science Advisory Board, and peer-reviewed literature." The Panel presumes that the intended application is to a broad range of models. However, this intention (if correct) is not clearly articulated in the "Scope of Guidance" in the Introduction to the Draft Guidance, nor are the classes of models (i.e., economic, behavioral, physical, engineering design, health, ecological, and fate/transport models) explicitly identified. This concern is particularly apparent in the Models Knowledge Base (see also CQ5), where much of the information elicited is highly focused on models for pollutant fate, transport, exposure, and effects. Models that address economic activity, behavior, and emissions are differentiated by other key criteria, including whether they predict at the level of the individual, household, firm, sector, region, or national or global economy; whether they are normative (predicting how people should behave under various assumptions of rationality and information) or descriptive (reporting how people actually do behave); and whether they address the costs or benefits of environmental regulations. Clearly the Draft Guidance is primarily intended to address regulatory environmental models, particularly those models used for policy analysis, national regulatory decision-making, and implementation applications. However, it should also be noted that it applies equally to a far broader category of models than its original targeted audience, and hence most of the Draft Guidance is expected to be useful for other modeling audiences as well. According to the EPA's CREM home page, "The Models Knowledge Base is intended to be a living demonstration of the recommendations from the Guidance for Environmental Models. 15 ------- In this way, these two products work in tandem to describe and document good modeling practices." In pursuit of this goal, the Panel recommends that the Draft Guidance clearly articulate the broad range of model types to which it is to apply earlier in the document, and ensure that the guiding principles for problem specification, model development, model evaluation, and model application reflect this diversity of types. 2.3. Glossary One of the keys to a workable Draft Guidance for quality assurance in environmental modeling is that the various modeling constituencies share a common language and definition of key ideas and terms. The Panel believes the Agency has made a commendable effort in attempting to establish a common vocabulary for the purpose of environmental modeling. The glossary is an excellent component of this Draft Guidance for providing the basis of that shared understanding. However, there is room for improvement and a need for consistency, not only in the glossary, but also in the text. For example, some of the terminology and definitions are subject to multiple interpretations, which is to be expected for a document that combines vocabularies from a variety of fields. The Panel notes that the Draft Guidance's use of certain terms, e.g. "guidance," as described in the preceding section, is at times at variance with past definitions, including some of the Agency's own previous modeling documents many of which are cited in the references. The Agency should clarify the Draft Guidance's use of terminology and definitions that may not always be used consistently. The current terminology used to describe the graded approach needs to be clarified. For example, "managerial controls" should be replaced with a more generic terms such as "level of effort" or "allocation of resources." Another problematic area is the potentially misleading or overly generalized use of common statistical terms such as "reliability" and "sampling errors." Where the Agency's use of terms is intentionally different from prior or accepted use, they should be noted as such, and a brief, appropriate rationale should be provided. The Panel suggests that the Glossary be expanded to include more terms to make it as comprehensive as possible. Some key terms that should be added are: "validation" (add a note: see model validation), "documentation," "user manual," "proprietary models," "secondary applications," "flow chart (code)," etc. Some panel members questioned whether the glossary definitions are the consensus of those in the Agency, or in the modeling community, or both? For example, "corroboration" is an interesting and appealing substitute for "validation," but one that is not yet widely used in practice. Appendix A contains specific suggestions for enhancing the utility of the Glossary. 16 ------- 2.4. Model Documentation, Project Documentation, and User Manual A variety of types and levels of "documentation" are required for a successful modeling project. The Draft Guidance discusses model documentation only in the model application component, i.e. a comprehensive project documentation to address "transparency" issues, (see box "Recommended Elements for Model Documentation," in Section 4: Model Application, on page 26 of the Draft Guidance). However there is a need for model documentation during development, especially for complex modeling frameworks. In addition, the Draft Guidance makes no mention of the need for an adequate user manual (or user guide) for the "analyst" group of model users. It is unclear whether this is assumed to be part of the overall modeling project documentation. Some Panel members believe it is separate and distinct from model project documentation, and is essential. In addition to the items already included in the box on page 26 of the Draft Guidance, the Panel believes it is important to note the need for documentation of choices made during model development, and for a model user manual. 17 ------- 3. GRADED APPROACH Charge Question 3: Has EPA sufficiently and appropriately proposed a graded approach, such that users of the guidance can determine the appropriate level of evaluation for a particular model use? If there are deficiencies in the proposed approach, what would you recommend to correct it, and why? 3.1. Definition of "Graded Approach" The concept of a "graded approach" is implicit throughout the Draft Guidance, as it should be. Usually "graded" is expressed implicitly through the use of the descriptor "appropriate." The term "graded approach" first appears under "Model Evaluation" (introduced on page 18). However, the graded concept applies to all phases of modeling—development, evaluation and application—not just evaluation. The Panel is concerned that the concept of a graded approach be introduced earlier in the document, before the discussion of model development, as an example of overarching concepts that are part of all of the modeling stages. Accordingly, the Panel recommends that the material on the graded approach be modified to reflect that model development, evaluation and application should always be conducted using a graded approach that is adequate and appropriate to the decision at hand, as determined by the Problem Specification process described in the Panel discussion of Charge Question #1. This introduction should then be followed by a brief discussion of how "graded" applies throughout the modeling process. For example, in the context of model development, "graded" refers to the extent to which existing models are modified to fit the problem specification or that screening models are used where appropriate, instead of more complex models. 3.2. Modeling Complexity and Associated Evaluation Needs The scope (i.e., spatial, temporal and process detail) of models that can be used for a particular application can range from the simplest models to the very complex depending on the problem specification and data availability, among other factors. In addition to providing some additional comment on where the model continuum starts (i.e., what is the simplest model to be considered in the Draft Guidance or in the MKB), the Draft Guidance needs to comment in more detail on the level of evaluation or "grade" of evaluation that might be appropriate for models of varying degrees of complexity. Currently, the discussion on page 18 dealing with the graded approach to evaluation is brief and the discussion of model complexity on page 11 only touches on evaluation complexity. In addition to the example of a "screening test" noted as a case where less rigorous model evaluation is required, examples of more complex situations should also be addressed in order to clarify the extended scope of evaluation that may be needed in different cases. 18 ------- The Draft Guidance also needs to alert the reader that external circumstances can affect the rigor required in model evaluation. For example, in cases where the likely result of the modeling will be costly control strategies, court actions, or alienation of some sectors of the population, detailed model evaluation may be necessary. In those cases, all aspects of the modeling will come under close scrutiny, and it is incumbent upon the modeler to probe deeply into the model's inner workings (sometimes called "process analysis") to support subsequent regulatory decisions. This level of deeper model evaluation also would be appropriate when modeling unique or extreme situations not previously encountered. The draft document should also note that gradation in evaluation can apply within complex model applications. For example, in modeling urban air quality, most areas use a regional modeling domain nested to provide higher resolution over the region of primary interest (e.g., Amar et a/., 2004). Clearly the most intensive performance evaluation should be directed towards the object of the modeling (the "fine grid"), but at least some level of evaluation should be applied to more distant areas (the "coarse grid"). The Panel finds that the Draft Guidance acknowledges the scope and complexity of the models being used, but recommends that it provide more examples of appropriate evaluation steps for different models and model systems (i.e., combinations of models linked to address a particular issue) and for their particular applications. The Panel recommends that the Draft Guidance broaden the discussion of the graded evaluation approach to discuss evaluation requirements for additional circumstances such as using models in potentially litigious applications or in unfamiliar or unique situations. Model evaluation in most every situation basically involves expert judgment, examination of model output under changes in key driving variables, intercomparison with other similar models, sensitivity and uncertainty analysis and comparison with observational data. The Draft Guidance needs to discuss the appropriateness of using the more qualitative evaluation steps such as expert judgment to "screen" the model performance and application appropriateness (i.e., how well does the numerical model agree with the conceptual model under current and scenario conditions) before launching into more formal and complex, or higher grade, intercomparisons with observations or sensitivity analyses. In addition, the Draft Guidance should offer examples of some particular practical methods, complementary to evaluation (e.g., use of relative reduction factors and ensemble modeling) that can be used to address uncertainty in the decision-making process. 3.3. Evaluating Model Response The Draft Guidance provides a comprehensive discussion of methods for evaluating a model's performance in terms of its ability to replicate historical situations. However, in regulatory applications the most important feature of a model usually is its response to changes in its input (e.g., response to growth and/or control of emissions). Aside from a discussion of post-audit, the guidance provides little direction for model users to evaluate whether a model will respond correctly to changes in critical inputs. Certainly a solid performance evaluation of how well the model replicates historical events, including analyses of the model's processes as well as 19 ------- its predictions, is an important component of evaluating its response. However, additional analyses focused on evaluating the performance of model response should also be conducted when the goal of the modeling is to predict a future state under expected or hypothesized changes to inputs. EPA provides a good discussion on evaluating model response in its recently-released draft final Guidance on the Use of Models and Other Analyses in Attainment Demonstrations for the 8-hour Ozone NAAQS [U.S. EPA, 2005]. Recommended techniques include retrospective analyses (similar to post-audit), use of various probing tools, comparison to observation-based models, and conducting sensitivity analyses for both the base and predictive cases using a variety of assumptions (a detailed discussion of these techniques is beyond the scope of this review). The Panel recommends that the guidance be expanded to specifically discuss evaluation of model response, and to include suggested techniques such as those provided in [U.S. EPA, 2005]. 3.4. Use of Multiple and Linked Models Many environmental problems require use of multiple models, with the models often linking together and interacting to varying degrees. For example, air quality modeling often links meteorological, emissions, and air chemistry/transport models. Integrated assessments that attempt to evaluate multiple, interdependent benefits and costs of a problem such as the overall value of the Clean Air Act as is done in EPA's studies on Section 812 of that act (U.S. EPA, 1997, 1999) and the work of the Grand Canyon Visibility Transport Commission (GCVTC, 1996) require linkage of a wide variety of atmospheric, environmental, economic and social models. In cases in which multiple models are linked together to address a particularly complex issue, each model needs to be evaluated individually to assure that the model is being used within its proper domain and that it is performing properly in the context of the integrated assessment. In addition, evaluation of the full modeling system needs to take place to make sure that the overall analysis is adequate and appropriate for the application. Just because individual modeling components are behaving properly does not necessarily mean that the fully linked system will provide authentic overall analyses. When using such a system of models, it is essential to beware of compensating errors, which can lead to "getting the right answer for the wrong reason." A classic example of compensating errors occurs in air quality modeling applications, where emission rates of pollutants are developed using an emissions model and meteorological parameters are generated with a meteorological model. Pollutant concentrations are then simulated using a dispersion model, using as inputs the emissions and meteorological model outputs. Modeled wind speeds that are too slow will lead to over-prediction of pollutant concentrations by the dispersion model, while modeled emission rates that are too low will lead to under-prediction of pollutant concentrations. These errors can be mutually offsetting, producing modeled pollutant concentrations that meet accepted performance standards. However, the fundamental flaws in a model's formulation will likely cause the modeling system 20 ------- to respond incorrectly to changes in the inputs (e.g., application of emission controls). The Panel recommends that the Draft Guidance acknowledge that many applications require the linkage of multiple models and that this linkage has implications for assessing uncertainty and applying the team of models. Each component model as well as the full system of integrated models needs to be evaluated for a given application. 3.5. Use of Model-Derived Data The Panel commends the Agency for recognizing that the definition of data includes data sets generated from modeling exercises as well as from the literature and existing databases. However, the guidance also needs to clearly discuss treatment of uncertainty associated with the application of these diverse model-generated data as well as data sets derived directly from observations. Data derived from modeling analysis that are then used for another modeling application also must be evaluated for uncertainties, caveats, and limitations in applicability. The evaluation then must be carried with the data throughout their future uses. One example of this need for propagation of data uncertainties and limitations is the use of emission inventories in regional air quality modeling. The emission inventories often are the result of complex data collection, analysis and emissions modeling. The inherent uncertainties in the emissions data and the emissions modeling need to be somehow quantified. Model users must recognize that the use of data as input for the next phase of modeling carries uncertainties, thereby impacting the next modeling steps. Sometimes, these uncertainties can be treated explicitly and quantitatively, but at other times the uncertainties can only be acknowledged qualitatively. Regardless, the uncertainties need to be noted and considered throughout the modeling system. This complex relationship between data and models needs to be discussed in the Draft Guidance. 21 ------- 4. PRACTICAL ADVICE FOR DECISION-MAKERS Charge Question 4: Has EPA sufficiently and appropriately provided practicable advice for decision-makers who must deal with the uncertainty inherent in environmental models and their application? What additional advice should EPA consider in dealing with uncertainty, and why? A number of researchers recommend a Bayesian approach to help decision-makers incorporate uncertainty into their decisions and to do so in a transparent fashion. Is the use of methods such as Bayesian networks an effective and practicable way for EPA decision-makers to incorporate uncertainty within their decisions and to communicate this uncertainty to stakeholders? If so, how? Are there alternative methods available? 4.1. General Comments on Uncertainty Experience suggests that shifts toward new, more informative, but potentially more complex, quantitative uncertainty assessment (QUA) methods inevitably present decision makers with challenges. A greater knowledge of uncertainty, absent an equally sophisticated framework for decision-making and communication, may only increase management challenges. More sophisticated QUA techniques do not automatically create more sophisticated regulatory decision-making. Thus the effective incorporation of uncertainty in decisions by decision makers, and the acceptance of these decisions by stakeholders, will not be accomplished with different or ever more elaborate QUA tools alone. Specific methods for performing sensitivity and uncertainty analysis are discussed in Section C.5 and Section C.6, respectively, of the Draft Guidance. The guidance appropriately recommends a sequential approach to evaluating the sensitivity of the model to its components and boundary values, to be followed by more in-depth investigation of components and potential interactions that prove to exert the greatest influence on the variability of model outcomes. This is a sound recommendation for developing an understanding of sensitivity in complex models with many factors and many possible interaction effects among those factors. In addition to the work by Saltelli et a/., 2000 cited in the report, other authors have proposed experimental test frameworks (Kleijnen, 2005) for formally examining sensitivity to individual effects and interactions in multi-parameter models. The matrix of statistical methods in Section C.5.7 provides a convenient comparison of the strengths and weaknesses of a progressively more complex set of approaches to sensitivity analysis. The merits of various methods for QUA have been discussed, debated, enthused over, and at times derided, including everything from simple bounding analyses through 1-D and 2-D Monte Carlo analyses, to Bayesian techniques. However, the REM Guidance should remind readers that incorporation of uncertainly into decisions is not just a function of finding the right mathematical or modeling QUA "tool." Because scientists and researchers are often more comfortable focusing on the "hard science" of models/tools than on the "soft science" that governs the decision making process, often too little attention is given to problem formulation (in 22 ------- its fullest meaning), risk communication, or the perspective of decision makers (Thompson and Bloom. 2000). The Panel cautions that searching for the "right" modeling tool (or uncertainty analysis) may miss the point; namely that models for regulatory purposes are a service to decision makers, and are not intended as a substitute for the hard task of selecting the "right" answer. Before deciding on a QUA tool, it is incumbent on the modeler to seek input from decision makers and stakeholders as to how and to what extent they may accommodate uncertainty in their regulatory decisions. To a scientist, expressing and quantifying uncertainty is a good thing. But the single value has a long history of use in regulatory decision-making. Asking decision makers and stakeholders how they view scientific uncertainty, how they would like to see it expressed, and how they see it being used in the decision-making process is the necessary precursor to effective and transparent use of any QUA method. In short: a) How much discretion does the decision maker have in addressing uncertainty? During policy development or for an action not directly governed by statute or rule, they may have considerable leeway to do so. Once a statute or rule is in place, they may have much less or no such leeway. Procedural regulations seem particularly resistant to incorporation of uncertainty. Many regulations work with reference to a fixed point (a "brightline" standard) and, despite an awareness that uncertainty exists in where this "fixed" point is actually located, decisions are simply based on whether or not the outcome is above or below that value. b) How will stakeholders react to knowledge of uncertainty and how will this reaction shape the decision-making process? To a stakeholder, expressions of uncertainty can be interpreted that experts "don't know," or could also imply inadequate effort, incompetence, or otherwise a lack of credibility of the responsible party, which undercuts support for regulatory decisions. Knowledge of uncertainty also allows opposing interests in a regulatory decision to focus on the highest or lowest value, regardless of its probability. Because there are often significant costs associated with choosing one specific value over another, arguments can erupt over differences in values that are, because of "uncertainty," statistically indistinguishable. The definition of the term "uncertainty" has been a source of considerable confusion in EPA documents and discussions of models used in environmental risk assessment. The REM Draft Guidance attempts to clarify the use of the term by: 1) identifying types of uncertainty (model, data, application niche) in Section 3.1.3.1; 2) distinguishing uncertainty from natural variability in model inputs and parameters for different modeling applications; and 3) defining uncertainty analysis (parameters) as distinct from sensitivity analysis (model form and importance of model factors). The Panel recommends that the Agency more clearly identify and discuss the various sources of uncertainty in model application, including: 23 ------- a) Stochastic variability, over space, time, and/or from individual to individual. Uncertainty arises from incomplete or improper representation of stochastic variability and the associated uncertainty in future system outcomes (e.g., of weather); b) Model (structure) uncertainty, including errors due to missing or improperly formulated process equations, inadequate spatial or temporal resolution, and incorrect model use; c) Model input uncertainty, resulting from data measurement errors, inconsistencies between measured values and those used by the model (e.g., in their level of aggregation/averaging), and parameter value uncertainty; and d) Scenario uncertainty, resulting from incomplete knowledge of current or future economic, regulatory, or physical conditions for which the model is applied. In addition to identifying sources of uncertainty, the Guidance should also discuss the implications of propagating uncertainties within model frameworks where models use the output of one model as input to another, or where model frameworks are assemblages of individual models. The Guidance provides some useful but too brief advice (Guidance §4.1.2) on how this uncertainty might be effectively communicated to decision makers and stakeholders. Much more emphasis must be placed on performing a robust and iterative problem formulation with modelers, decision makers, and stakeholders and on correctly conveying model results using non-technical, non-quantitative, and non-condescending communication techniques. Any transparency of QUA methods is only possible if decision makers and stakeholders are engaged early on by inclusive, effective communication and outreach strategies. The Panel recommends that the REM Guidance strongly advise modelers to begin model development or use only after they have obtained an awareness of how a decision maker plans to use the information on uncertainty that they will be providing. This is an important component of the Problem Specification as well. 4.2. Sensitivity Analysis vis-a-vis Uncertainty Analysis Section C.5 would benefit from improved clarity in the distinction between sensitivity and uncertainty analysis. For example, in Section C.5.1, the REM guidance obscures the distinction between the goals of sensitivity analysis and uncertainty analysis, where it states ".. .the distinction between these two related disciplines may be irrelevant" (p. 50). While the Panel agrees that the two are interrelated and sometimes confused, the distinction should be clarified in the guidance. Sensitivity analysis is an examination of the overall model response to a perturbation of model inputs. The analysis thus can be used to inform model users, decision makers and 24 ------- stakeholders on where to focus the most resources in terms of developing a better understanding and characterization of the uncertainties for particular components of the model identified as "most sensitive" to perturbations of underlying model parameters. Rather than perpetuating any possible confusion between the focus or goal of these two analyses, the REM guidance should be more transparent in describing the purpose of each, their interrelationship, and the distinction between them. For example, the discussion in Section C.5.5 relating to Monte Carlo analysis currently reads more like a discussion of uncertainty analysis, rather than sensitivity analysis. As noted in Cullen and Small (2004), sensitivity analysis is an important adjunct of uncertainty analysis, determining the impact of particular model inputs and assumptions on the estimated risk. Sensitivity analysis is often conducted as a precursor to uncertainty analysis, helping to identify those model assumptions or inputs that are important. If the model outcome is not sensitive to a particular input or set of inputs, there is no need to examine these inputs as part of a more sophisticated uncertainty analysis. Sensitivity analysis is revisited in the subsequent phases of an uncertainty analysis to identify those inputs and assumptions that are significant contributors to the overall variance of the output and/or critical to pending decisions (for an example of the latter, see Merz et a/., 1992), thereby identifying the uncertainties that matter. In this manner, priorities can be established for further research and data collection efforts. Therefore, the Panel recommends that the guidelines articulate a more tangible set of alternatives for addressing model sensitivity/uncertainty. In particular, recommendations for uncertainty analysis should identify the need to focus resources on those processes to which the model state variables are most sensitive and, in addition, are less certain in terms of their formulation and/or parameterization. 4.3. Uncertainty Analysis Practices/Methods (REM Guidance Section C.6) Section C.6 of the Draft Guidance on uncertainty analysis is incomplete in relation to the coverage given to sensitivity analysis in Section C.5. Returning to the discussion of types of uncertainty in Section 3.1.3.1, this section tries to address the "niche uncertainty" under the label of model suitability and "data uncertainty" through a weakly defined discussion of frequentist and Bayesian interpretations of probability. Unlike the rather detailed discussion of methods for corroboration and model sensitivity analysis, there is little true guidance on how to evaluate uncertainty in model parameters and the effect of this uncertainty in decision-making based on model outcomes. The current Draft Guidance touches on the notion of a Bayesian framework and the use of prior knowledge and expert advice to reflect uncertainly in the model inputs (including parameter values). However, it does not distinguish carefully between Bayesian estimation of posterior distributions and associated inferences and decision theoretic approaches which incorporate explicit loss functions for certain errors in inferences. It would be very useful to have a "Box" example of an uncertainty analysis in which there is an established prior for an "uncertain" model parameter, a likelihood for the input data, and an updated posterior distributions and associated inferences and decision theoretic approaches which incorporate explicit loss functions for certain errors in inferences. Thus, the Panel recommends that the 25 ------- REM Guidance (and MKB ) provide more practicable information through inclusion of "case study" examples of where and how EPA is currently incorporating uncertainty analysis in environmental models as an integral component of decision-making. In addition, the Panel recommends that Section C.6 be enriched to a level comparable to that of Section C.5 on sensitivity. The Panel agrees that Bayesian approaches are one of several candidate methods suitable for quantifying data uncertainty in appropriate situations. Bayesian methods are certainly appropriate for treating uncertainty in environmental modeling and may be particularly effective in modeling applications where empirical data on the distribution of model parameters in real applications are sparse and expert judgment may provide the most realistic assessment of the prior distributions. A Bayesian treatment of a simple model application or a more complex model with a network of dependencies (conditional relationships) is a theoretically appealing approach to incorporate prior uncertainty into posterior distributions of model outcomes (e.g. exposures, concentrations, expenditures, morbidity, mortality, etc.). Current software and iterative estimation algorithms have removed many of the computational barriers that once stood in the way of Bayesian treatment of a model application. Yet the removal of computational barriers does not eliminate the need for a solid understanding of the scientific basis for the model and in fact may require a heightened understanding (subjective, expert knowledge) of the prior distributions of parameters. Furthermore, adoption of Bayesian uncertainty analysis methods does not reduce the importance of sensitivity analysis to establish the importance of the model components and their interactions. The effectiveness of the Bayesian approach will be greatest when information on the prior distributions is accurate and new data to support the model application are plentiful. If the prior information is weak or uninformative or the amount of new data available for model parameter estimation is large, the model results will be dominated by the new data. If the new data inputs to the model are weak, the posterior distributions for outputs will be dominated by the prior distribution assumptions. The Panel endorses the recognition that QUA should be an inherent consideration when using models to support regulatory decisions. Yet, given the enormous breadth of modeling paradigms (spatial and temporal scope and degree of complexity), the Panel remains cautious in its recommendations regarding specific methods of QUA (e.g., "frequentist" vs. Bayesian as suggested in the charge question). The nature and complexity of any particular model, its application within a particular regulatory program, availability of data and resources, etc. will all influence the choice of QUA that is appropriate. Thus, as with all other aspects of modeling, a graded approach is warranted for conducting uncertainty analyses. In some applications, simple sensitivity analyses may be all that is required. Regulatory decisions with far-reaching impacts should endeavor to use QUA tools to provide the public and stakeholder community with greater appreciation for the uncertainty range in the model output decision variables that ultimately define regulatory decision points. 26 ------- 4.4. Value of Information - Identifying "Uncertainties that Matter" After identifying model inputs and assumptions that contribute significantly to variance in the output, it is necessary to consider how to use this knowledge (Cullen and Small, 2005). Value of Information (VOI) techniques seek to identify situations in which the cost of reducing uncertainty is outweighed by the expected benefit of the reduction. In short, VOI is helpful in identifying model inputs that are significant because: a) they contribute significantly to variance in the output, and b) they change the relative desirability of the available alternatives in the decision under consideration. The Panel recommends that the REM Guidance acknowledge the potential utility of VOI techniques available to assess the importance of the variability and uncertainty contributed by individual inputs to the expected value (or conversely, the "loss") associated with a decision under uncertainty (Raiffa, 1968; Morgan and Henri on, 1990; Finkel and Evans, 1987; Massmann, et al., 1991; Dakins et al., 1996; Yokota and Thompson, 2004). While the Panel understands that the REM guidance is not intended to be prescriptive in its effort to provide an overview of QUA methods, it does not provide sufficient context currently for an end user (e.g., modeler within the regulatory community) to be able to determine the level of QUA that would be appropriate within a particular context or application. The REM Guidance might consider providing a more concrete decision framework to help guide the choice of appropriate/available QUA methods. As a starting point, the REM Guidance should include examples of, or references to, the nature and degree of QUA currently being implemented or adopted within various EPA programs. For example the Panel is aware of the extensive uncertainty analysis that is an integral component of the 3MRA modeling system. While it is clear that this one example should not be taken to endorse a particular QUA, the MKB would provide one means of assembling a "library" of such examples with the nature of the QUA, the data requirements, limitations, etc. This would provide at least some options by way of example that model users and decision makers could turn to as a resource beyond the cited statistical references. The appeal of QUA is that it can be used to provide quantitative estimates of the "degree of confidence" when using model results as a component of regulatory decisions. Nevertheless the results should be presented with some caution. It might be tempting to assign a high degree of confidence in the uncertainty analysis based on the adoption of a highly elaborate or complex analysis. Yet, the validity of the QUA is of course dependent on the quantity and quality of the information available for the analysis. The choice of an appropriate QUA method (frequentist, 1-D versus 2-D Monte Carlo, Bayesian, etc) can only be made if the intended audience of the REM Guidance understands the data requirements and associated level of effort to conduct the analysis of the various types of QUA. As compared to the REM Guidance describing best practices for model development/evaluation, the guidelines do not contain a similar set of "best practices" for evaluating, presenting, and incorporating model uncertainty into decision-making. While references cited in the REM Guidance provide an array of applicable methods to address model uncertainty, the draft guidelines do not provide sufficient discussion, 27 ------- context, and recommendations needed to provide a model user/decision-maker with "practicable" information relating to appropriate uncertainty analysis methods and how to convey the results of such analyses. The Draft Guidance should offer some practical methods that can be used to address uncertainty within the decision-making process. For example, one is the concept of Weight-of- Evidence (WoE), in which the model is only one (albeit an important) component in a suite of analyses feeding into the decision framework. A second possible approach is to use the model in a relative, rather than absolute, predictive mode. This approach uses "relative reduction factors" multiplied by observed (measured) conditions in place of absolute predictions. In theory, such an approach can avoid or cancel out systematic biases in the model formulation, hence reducing the uncertainty in the predictions used for decision-making. A third example approach to dealing with uncertainty is the use of ensemble modeling. This approach involves running several different models and using a composite of the results. While ensemble modeling can be very resource-intensive, it may be worth considering for applications or decisions involving extreme cost or risk. These example approaches could be included, among others, with the REM Guidance to provide decision-makers with practical examples of methods incorporating uncertainty in the decision framework. 4.5. Communicating Uncertainty Independent of the choice of particular QUA tools, the Panel recommends that the REM Guidance provide more discussion on the importance of the manner in which results of QUA are communicated to the decision-maker (and to public/stakeholders). Graphical methods often serve to convey complex statistical/probabilistic results in a more understandable manner, and the REM Guidance should consider including a range of examples in the document. Again, the MKB may be useful as a library of such examples. As the analyst/modeler and the decision-maker are usually not the same individual, it is important to accompany results with the key assumptions and caveats encompassed in the analysis. How can uncertainty or probabilistic results be interpreted to help identify the uncertainties that matter most, and to point the analyst to further study or data collection activities that can be most beneficial in reducing these critical uncertainties? As noted earlier, often only a relatively small subset of inputs is responsible for a majority of the variance in a model's output. Morgan and Henrion (1990), Cullen and Frey (1999) and others describe the use of summary statistics, visual methods, regression approaches and other sensitivity analysis tools to help find the most important input uncertainties. Broader approaches to risk communication and methods for testing the effectiveness of alternative presentations are discussed in Finkel (1990), Bostrom et al. (1992), Morgan etal. (1992), Fischhoff etal. (1998), Thompson and Bloom (2000), and Cullen and Small (2005). The preponderance of QUA methods focus on what the REM Guidance defines as "data uncertainty." Quantitative "model uncertainty" and "application niche uncertainty" present significant challenges that are rarely feasible to address. In addition, empirical or observational 28 ------- data are themselves subject to uncertainty depending on the quantity and quality of those data, and it is important to recognize these uncertainties in the context of evaluating the importance of model uncertainties. In the case of directly observed data, there are uncertainties associated with the measurement techniques and with the data analysis processes themselves. In the case of data that are generated by modeling, uncertainties arise as a result of modeling analyses that produced the data. A common example is the difficulty of comparing environmental data (collected at a particular point in time and space) to a model prediction based on averaged conditions for a grid cell with spatial parameters and time steps necessarily much different from the conditions under which the measurement was made. As discussed earlier, a clear description that discusses the main sources of uncertainty, including an indication of the types of uncertainty that are most readily addressed, would be helpful in communicating these concepts to the reader. Therefore the Panel recommends that the REM Guidance be clear on the types of model uncertainty that most QUA tools address. These data uncertainties mean that using data to evaluate models is very much an imperfect process. As a result, the discrepancy between observed data and model simulations does not mean that the model is wrong or not useful. It is particularly important to communicate this concept to decision-makers who may favor discounting modeling results if the comparisons between observations and models are less than perfect. In addition, when analysis of data is used in lieu of modeling results because the modeling results do not completely agree with observations, the potential errors and/or uncertainties in the data used for the analysis must be acknowledged. In some cases, these uncertainties may actually be more significant than the uncertainties determined for the modeling itself. The complex nature of data uncertainties and modeling uncertainties needs to be carefully communicated to decision-makers. To promote this discourse as part of the general practice of modeling, the Panel recommends that the Draft Guidance should stress the importance of communicating model sensitivity and uncertainty both in the context of model evaluation and when interpreting and applying model outcomes in the context of decision-making. 29 ------- 5. IDENTIFICATION AND STRUCTURE OF OPTIMAL SET OF INFORMATION FOR ALL USERS Charge Question 5: The Panel should consider that environmental models will be used by people whose technical sophistication will vary widely. EPA has therefore attempted to cull information about models that broadly serve the needs of all users, using a data template to collect this information (see Attachment D). Has EPA identified, structured and developed the optimal set of information to request from model developers and users, i.e., the amount of information that best minimizes the burden on information providers while maximizing the utility derived from the information? 5.1. General Comments and Suggestions As indicated in Attachment D of the MKB (included in this report as Appendix B), the major categories of information collected for the models in the REM Models Knowledge Base include: A. General Information, regarding the model name, contact information, overview, and web link; B. User Information, concerning technical requirements and basic instructions for obtaining and using the model; C. Model Science, including the conceptual basis for the model and discussion of evaluation steps that have been undertaken and documented for the model (code verification, corroboration with observed data, sensitivity and uncertainty analysis); and D. Model Criteria, summarizing applicable regulations and the problem domain(s) addressed by the model, including types of pollutants, sources, environmental media, and key fate and transport and exposure and effects processes. The information solicited in the current data entry sheet addresses most of the critical elements needed by potential users to assess the overall relevance and utility of a model in the MKB, and does so in an effective and efficient manner. However, some additional general subcategories of information should be added to the data entry sheet. A. General Information The general information entries for the MKB data sheet include: 1. Model Name (and acronym), 2. Model Overview/Abstract, 30 ------- 3. Contact Information, and 4. Model's Home Page. This information is appropriately informative and concise, and the examples we considered in the current MKB provide useful introductions to the models. B. User Information The user information entries include: 1. Technical Requirements a. Computer Hardware, b. Operating Systems, c. Programming Languages, and d. Other Requirements and Features. 2. Download Info (with URL) 3. Using the Model a. Basic Model Inputs, b. Basic Model Outputs, c. User's Guide, and d. Other User Documents. The information requested is useful and appropriate. Most users will not need to know the programming language used by the model, since they will access, download, and use an executable version of the model. Nonetheless, this information could be useful for some users and provides a useful context for system requirements. The MKB should indicate whether the underlying programming language(s) must be obtained or licensed for use of the model. Under the "Using the Model" section of data entry, the Panel believes that it would be useful to indicate the level of expertise, both environmental and computer, needed to understand and use the model, and the level of user support provided for the model by its developers, the Agency, or other sources. This information is provided for a number of the models currently in the MKB as part of the User's Guide or Other User Documents fields. Still, it would be useful to explicitly ask for this information as part of the data entry sheet. C. Model Science The model science categories include: 1. Conceptual Basis of the Model, 2. Scientific Detail for the Model, 3. Model Framework (equations and/or algorithms), and 4. Model Evaluation (verification (code), corroboration (model), sensitivity analysis, uncertainty analysis). ------- The requested information addresses many of the key elements needed to document and assess the scientific basis for a model. However, the Panel does recommend some modifications and additions to the list above. First, defining the Model Framework as the 'equations and/or algorithms' for the model (as is also done in the Model Glossary) appears counter to the usual use of the word "framework." This term is usually associated with the broader conceptual basis for the model or (by some, see the U.S. EPA, 2003 and in particular, EPA's Modeling QAPP Draft Guidance, page 54) as "the model and its supporting hardware and operating system." A clearer request for the underlying model equations and/or algorithms would be provided using the descriptor "Model Structure and Calculation Methods." Second, the mention of corroboration (model) under Model Evaluation should explicitly mention the model's ability to predict observed monitoring data. The Model Evaluation section of the Model Science entry considers many of the key issues needed to evaluate the scientific rigor behind the underlying model development and previous applications, and addresses many of the elements of good modeling practice that are emphasized in the Draft Guidance. Indeed, the Panel views an important purpose of the MKB as providing an incentive for model developers and purveyors to conduct and openly communicate their efforts in model evaluation. From this perspective, the Panel recommends some additional pieces of information that should be elicited and reported, including: 1) Documented examples of peer review for the model, including reviews conducted by the EPA, other agencies or panels, and papers presented in the peer reviewed literature. Key limitations and needs for improvement that were identified during these evaluations should be reported, and 2) Benchmarking studies in which the model's predictions and/or its accuracy are compared with those of other models. The Panel also recommends the inclusion of a section, following Model Evaluation, for the model developer to summarize key limitations of the model and plans or needs for modifications and improvements. This type of self-critique would be both informative to users and motivating to the ongoing improvement of the models in the MKB. D. Model Criteria The model criteria elicited and reported include the major categories of: 1. Regulations, 2. Releases to the Environment, 3. Ambient Conditions, 4. Exposure or Uptake, and 5. Changes in Human Health or Ecology. 32 ------- The Panel notes that the criteria elicited are highly focused on models for pollutant fate, transport, exposure, and effects. Much of this information is not appropriate for models that address economic activity, behavior, and emissions. These models are differentiated by other key criteria, including whether they predict at the level of the individual, household, firm, sector, region, or national or global economy; whether they are normative (predicting how people should behave under various assumptions of rationality and information) or descriptive (reporting how people actually do behave); and whether they address the costs or benefits of environmental regulations. As such, the Criteria should first note the genre of the model, whether economic/behavioral vs. physical or engineering science models (though some models, e.g., for predicting emissions, could combine elements of both), and include different subsets of information for these. 5.2. Specific Suggestions by the Panel 1) Under Regulations, those entering information into the MKB should be given the opportunity to identify "Other Regulatory or Decision Support Applications." These could include US regulations, such as NEPA or Natural Resource Damage Assessments under CERCLA, or international agreements or treaties, such as those for ocean disposal or controls on persistent organic pollutants (POPs). It could also include non-regulatory decision support applications, such as for risk communication efforts by state environmental or public health agencies, or life-cycle assessment in support of green design decisions by firms. 2) Under the Releases to the Environment Section, a differentiation should be made between models for natural systems (emphasized in the current list) and engineered environments, such as buildings, treatment plants, and water distribution systems. (Models for the latter, such as EPANET, have received increased attention in recent years due to concerns regarding drinking water quality at the tap from accidental or purposeful (i.e. terrorist actions) contamination, and should be sought for inclusion in the MKB.) Also, under Source Type, area source models should be explicitly noted to include larger scale sources, e.g. for non-point source runoff in watersheds, biogenic emissions in regional air quality models, or distributed natural or anthropogenic sources to groundwater. 3) Under Ambient Conditions, the Panel feels that the terms included under Processes (transport, transformation, accumulation, and biogeochemical), while useful information for many fate-and-transport models, is specific enough that it need not be included in these general model criteria. The Panel suggests that this information be replaced with the following, more-general criteria: a) Time scales addressed in the model and whether the model predicts for dynamic or static conditions, 33 ------- b) Spatial scales or economic units addressed in the model and whether it provides a primarily distributed vs. lumped representation of the modeled system, and c) Whether the model is deterministic, predicting single values for model outputs, or stochastic, predicting a range or distribution of values to characterize variability and/or uncertainty. 4) Under Changes in Human Health or Ecology, the options should be expanded to include natural resource or materials damage, to consider effects, e.g., on visibility, historic buildings, or property value. 5) Model Applications: In addition, the Panel recommends that an additional major category of information be elicited and reported (in addition to the major items A- D). The additional category would be listed as "E. Model Applications," and should direct users to specific examples of regulatory or non-regulatory applications of the model (distinguishing between the two) in the public record and the peer-reviewed scientific literature. 5.3 Track Versions of Models The Panel recommends that revision tracking be incorporated into the MKB. Such a feature would have several benefits. First, it better reflects the realities of modeling than the current framework in which models are implicitly treated as unchanging. Second, it facilitates a tighter connection between policy analysis and modeling: the documentation for an analysis would specify a particular model version whose characteristics could be retrieved from the database. Third, it would provide valuable insight into the evolution of models over time. It would be possible to observe the extent to which changes in a model are driven by: developments in the underlying science; the availability of new data; the availability of new software or algorithms; the demand for new features; and the correction of programming bugs. Revision tracking could be implemented as follows: 1) A version field and a date field would be added to the data entry form. The contents of the version field would be a character string supplied by the model developer. The string should contain enough information that the developer (or a subsequent maintainer) could reconstruct and rerun that same version of the model at a later time. The date field would be the date at which that version of the model was released or placed in service, and 2) Each time a new version of the model is added to the database, there should be one or more fields describing the significant changes in the model from its previous version. In addition, all other fields associated with the model should default to their settings from the previous version. However, it should be possible to provide an updated 34 ------- version of any field without losing the corresponding field from the previous version of the model. The documentation burden imposed on model developers would be small. In particular, models whose development has been sponsored, at least in part, by EPA will already have significant changes spelled out in grant proposals or cooperative agreements. Ideally, the MKB would also include information on bugs fixed between versions. With revision tracking in place, the main page for each model would have a link to "Previous Versions," which would take users to a page showing the dates and revision numbers of all previous vintages of the model in the MKB. Each previous version should be a clickable link showing the list of changes embodied in that version (from above) and include links to other information specific to that version of the model. 5.4. Listing of Key Publications and Applications of Models The Panel believes that it would be useful to include a list of key references for each model: publications and reports where the model is described or documented, and important applications. Model developers will be able to provide this information easily and it will allow potential users to: (a) find out more about a model; and (b) avoid duplicating previous research; and (c) see example applications. This information would also address the concern raised in charge question 7c by showing how widely used and thoroughly peer-reviewed each model is. 5.5. Clarification of MKB Entry Sheet Items C1-C3 The distinction among items Cl, C2, and C3 in the MKB Data Entry Sheet should be made clearer, and the information requested by these items should correspond more closely to the parallel sections of the Draft Guidance that discuss this information. Question Cl and C3 are intended to match Section 2.2 and 2.3 of the Draft Guidance but most model builders and users will probably regard those sections as overlapping considerably. Section 2.2 (Conceptual Model Development) in the Draft Guidance, for example, requests a clear statement and description of each element of the conceptual model, plus documentation of the science behind the model, including: its mathematical form, key assumptions, the model's scale, feedback mechanisms, etc. It seems, in short, to be asking for essentially complete documentation for the model. Because of such great breadth of coverage, the types of information covered by Section 2.2 are solicited by items Cl, C2, and C3 on the Data Entry Sheet. Subsequently the Draft Guidance, Section 2.3 (Model Framework Construction), begins with a discussion of some of the same information: a formal mathematical specification of the concepts and procedures of the model. Assuming information provided under C3 is intended to parallel that discussed in Section 2.3, it is not clear how the mathematical formulation requested here differs from that requested under Cl. It appears that the intent of C1-C3 is the following. The answer to Cl would be a broad conceptual overview of the model that would be relatively free of technical detail (no equations) and would be accessible to readers from a wide range of backgrounds. It would usually include a 35 ------- diagram showing the relationship between major components of the model. The answer to C2 would provide the technical detail missing from Cl (namely, the model's key equations) and would have specialists as its intended audience. It would provide the theoretical basis for the model. The answer to C3 would describe the model's numerical implementation (data, algorithms, computer programming). This approach would be useful but needs to be spelled out more clearly in instructions accompanying the form. It would also integrate well with version tracking: the answer to C3 will usually change with each revision of the model; the answer to C2 will change periodically; and the answer to Cl - which defines the essence of the model - will generally be stable. 36 ------- 6. DATA DICTIONARY AND DATA STRUCTURE Charge Question 6: EPA has developed a data dictionary and database structure to organize the information it has collected on environmental models (see Appendices E andF of the Draft Guidance). Has EPA provided the appropriate nomenclature needed to elicit specific information from model developers that will allow broad inter-comparisons of model performance and application without bias toward a particular field or discipline? 6.1. General Comments The discussion of the elements of this question is based primarily upon relatively terse, and sometimes vague, information provided by the REM Data Dictionary and the REM Entity Relationship Diagram. The Panel's review of the Data Entry Sheet (CQ5) and related documentation of several individual models appearing in the REM Models Knowledge Base (MKB) were also considered in this question. This has led the Panel to recommend that the technical issues concerning the specific design of the MKB be addressed by either (a) a separate knowledge base topical report, or (b) an additional appendix to the current Draft Guidance, to allow the main report to concentrate on the Agency's overall plan for the use of this important tool, without ignoring the details of its functional design. The Panel's expectation is that the developers of the MKB database structure would also perform the necessary QA review of their Data Dictionary and entity relationships to assure that they are properly drawn and functioning. This aspect is virtually impossible for the Panel to evaluate thoroughly on the basis of the limited details provided on the database structure in the two documents provided. It is similarly difficult for panel members (who are not information technology specialists) to provide much useful advice without a better understanding of the strategy and implementation of the design. Perhaps the separate topic report or MKB Appendix could include all of this definition information and an outline of the database design strategy. Panel members were not sure this would be helpful. As noted below, review of the individual model documentation in the MKB provided the Panel with the most insight on the effective results of the application of these tools within its system. Although the Glossary presented in Appendix A of the Draft Guidance is an undisputed "plus" for the documentation effort, very few of the terms in the Data Dictionary are repeated there, as may be expected and appropriate, given the specialized nature of database terminology that is usually unique to the particular database software program for which it was specified. For a database, its functional terminology use has to be clear and internally consistent, regardless of its conformance to the "outside world." It has been noted elsewhere that several of the Glossary terms have varying definitions, as used in different sections of the Draft Guidance and MKB references—even though they are intended to conform to the Guidance definitions put forth in the Glossary. Although it initially appeared that ongoing efforts may have to include variant definitions (with footnotes to indicate model association); the use of "special guidance-specific" 37 ------- definitions for some terms may be satisfactory if the authors of the guidance carefully review their use of terminology for consistency of use, and alter the text accordingly. As suggested above, however, the MKB Data Dictionary can function independently and quite satisfactorily, as long as the translation of the Data Entry Sheet terminology to database definitions is precisely specified. The Panel therefore recommends that the Agency follow its own standard QA/QC program procedures for ensuring quality of the all of the underlying information in the MKB system. From evidence presented to the Panel, it appears that this has already been substantially completed for the functions currently defined. As new functions are added to support new features, including those recommended elsewhere in this report, it will of course be necessary to expand and update this Data Dictionary and repeat many of the QC checks to verify functionality. The Panel has varying opinions on whether the overall Glossary should include all of the Data Dictionary terminology to assure that referencing is clear to all users. For the reasons outlined above, it appears as though this would potentially add more opportunity for confusion than enlightenment Therefore, the recommended approach that would isolate the Data Dictionary in its own self-standing report would seem most advantageous at the current time. Regardless of the location of this documentation, the Panel re-iterates its encouragement to extend the QA/QC procedures followed to establish the initial quality of the MKB into the larger QA program. This is needed to maintain the information, as well as the hardware and software systems needed to implement it. 6.2. Model Performance Information This charge asks about including database information that is "unbiased." However, as indicated by the presentations made by Region 5 and 10 representatives before the Panel, there is also a need for a place in the database for additional "classification" information, which may go beyond that requested from the developer, and which may appear to support "apparent advocacy" or be otherwise "biased," if it includes "recommendation" information. This would be a subsection of the database specifically devoted to information that helps Agency regulatory- model application staff and "outside applicants" to identify the "most appropriate" candidate models. (A new "model selection program" that is under development by ORD was demonstrated at the Panel's review meeting. It appeared to be a potentially valuable tool, but several Panel members cautioned that it should produce an output file that includes a matrix of candidate models, rather than a single "recommendation," so that the user of the tool can more fully consider which of several candidates best fits the problem application at hand). Much final model-selection decision making is presently achieved by regional or state agency discussions that come to agreement on the most appropriate site-specific model choice for major projects at a particular decision point. However, as noted further below, the MKB would be more valuable, if cumulative EPA problem application experience could be more consistently represented in the database, along with the present basic model description information. The Panel is in concurrence on the importance of eliciting and including information on historical model performance and particular application experience from various model users 38 ------- (both other modelers and decision makers), as well as model developers. This was not especially motivated by any desire to minimize "biases" in reporting. There was some concern that developers of a model may not be in a position to fully (or objectively) judge its behavior in various contexts. Avoiding or minimizing bias would seem to require gathering reviews from as broad a user base as possible. It now appears that the current approach, which utilizes only information volunteered by the model developers, would tend to ensure that individual "biases" are included, without any real opportunity to neutralize them. This situation may be the unintentional result of using a more open narrative format for developers to explain features of the model. It may be noted that the Panel review of the current Data Entry Sheet, the Data Dictionary, and the Entity Relationship Diagrams did not suggest that there were any particular features that would "bias" the selection or representation of models. Instead, as noted both above and below, the reviewers were interested in seeing more information, as this could include application experience with "competing" models. In fact, the inclusion of additional information on the history of performance suggested by several Panel members would be more likely to include "opinions" as to the quality of performance, hopefully supported by comparison with appropriate measurement data sets. This extra information was viewed as important to prospective model users, even though it would be likely to also include some "biased" information. As long as instances of "preconceptual bias" can be identified and flagged or filtered, the availability of previous application experience (especially successes) would be a valuable component of the MKB information set. (Given the wide variety of models included, this "openness" may be helpful to both agency and "outside" users; but perhaps some form of warning of the risk of potential bias should be included with any new "performance history " element, so that the new users are fully aware of this limitation). The Panel recommends that the Agency clarify the intended roles of the "inside" and "outside" users of the MKB system and how that affects the priorities for the user interacting with the system (including supplemental, even if "biased," application history information). 6.3. Additional Recommendations To address details issues of CQ 6 more specifically, the Panel reviewers observed that the dictionary and database do capture much of the information necessary to assess model performance; but there were some noted exceptions: 1) CONCEPT: This results from problem formulation, but may or may not convey to the user useful information about the problem or set of problems (Draft Guidance §2.1) for which the model was developed. Another field should be added, "Problem Specification" (as noted in Section 1.2 of this review, to concisely capture descriptive information about the original application problem. 2) DECISIONDOCS: As written, this field seems to focus on how to use (run?) the model, how to produce output, and what experience there has been with running the model. This (or a new) field should include information or links to examples of when, 39 ------- how, and where the model was used to support an actual decision or decisions. Qualitative opinion on how the model performed would be acceptable/desirable. What benefits and problems did decision makers and stakeholders experience when using the model? This element should include a date entry so potential users can better judge the currency of the model. 3) DOWNLOADINFO: This should include information on the size of the model (zipped and unzipped), whether it is one file or a collection of files, and whether its setup will require changes in system files. 4) DIR ENTRY STATUS and REVISION_DATE: It is not clear what is meant by "last reviewed" — whether the date given would be for when the model itself was reviewed or when its entry into the dictionary was last updated? There should be information on when the model itself was last reviewed by its developer, as well as documentation (or links to such) of any and all changes, including errata and enhancements. It would also be useful to have documentation of problems encountered or improvements suggested by actual users of the model. All of this may be considered in MODELCONTACTINFO, but the database appears to be placing any "institutional memory" of the model's behavior in a person, who may or may not be available. The reviewers thought that there should also be fields consistently indicating whether model documentation is available online, who is responsible for preparing and maintaining this documentation, and the date it was last reviewed and/or updated. 5) EVALUATION includes four questions, but without performance information, the first three seemed less useful (recognizing that they might represent the only information available for newer models). 6) MODEL_CATALOG Table information given in Data Dictionary is too cryptic to tell whether any model performance information would fall into the descriptions provided there, and 7) PROG_LANGUAGE: This should also indicate whether any other software (particularly proprietary, e.g., ArcINFO) is required to operate the model. Panel reviewers considered their observations in reviewing the Aquatox (See Appendix C-3), CalPuff (See Appendix C-l), IPM (See Appendix C-2), and other models (See Appendix C-4) in reaching their conclusions about the performance of the identified database elements. Overall, the construction of the system appeared to be generally well-designed, but with opportunity remaining for expanding its focus to include more consistent information on model use experience and performance in a format that would make it more uniformly easy for users to compare models of interest for a particular candidate application. There are several key features that the Panel would like to see improved or expanded so that the MKB can be most effectively used by the EPA and its stakeholders. The existing Data Dictionary and Database Structure 40 ------- appear to be adequate to address existing features of the current MKB. However, as this tool is expanded to include new features recommended by either this Panel or the Agency's developers, it will be necessary to add new structural elements and data elements; and this will require an ongoing additional QA/QC effort. Therefore, the Panel recommends that the following issues should receive further consideration and attention: 1) A consistent QA review of the current content of the information contained in the MKB [some model feature/description errors (at the user interface level) were noted by Panel members, see Appendix C of this report], 2) Follow-up requests to developers who supplied original information to supply missing data for the minimum set of descriptors that the Agency decides are essential to proper model selection, 3) Entries into the data dictionary be clearly defined and made as consistent as reasonably possible, with the text in the Draft Guidance and data entry forms, and 4) Provision of a mechanism that actively solicits feedback from the user community regarding application experience and model performance, both inside and outside the Agency, beyond voluntary e-mails to designated contacts for individual models. 41 ------- 7. QUALITY OF INFORMATION PROVIDED ABOUT THE MODELS Charge Question 7: To facilitate review for this particular charge question, the Panel should focus on three models that represent the diversity of model information housed within the Models Knowledge Base. These models are: (1) Aquatox, a water quality model; (2) Integrated Planning Model (IPM), a model to estimate air emissions from electric utilities; and NWPCAM, an economic model. 12 Using these three models as examples and emphasizing that EPA is not seeking a review of the individual models, but rather the quality of the information provided about the models, EPA poses the following questions to the Panel. Through the development of this knowledge base, has EPA succeeded in providing: (7a) easily accessible resource material for new model developers that will help to eliminate duplication in efforts among the offices/regions where there is overlap in the modeling efforts and sometimes communication is limited? (7b) details of the temporal and spatial scales of data used to construct each model as well as endogenous assumptions made during model formulation such that users may evaluate their utility in combination with other models and so that propagation of error due to differences in data resolution can be addressed? (7c) examples of "successful" models (e.g., widely applied, have been tested, peer reviewed etc.)? (7d) a forum for feedback on model uses outside Agency applications and external suggestion for updating/improving model structure? 7.1 General Comments The Panel commends the Agency for developing the Models Knowledge Base (MKB) and strongly supports its continued improvement. This type of resource has been needed for some time and even in its draft form, the MKB provides an easily accessible resource for the modeling community that, if maintained and used, will significantly improve the development and application of models both internal and external to the Agency. In answering questions 7b-7d, the Panel focused primarily on two suggested models (i.e., AQUATOX and IPM) along with a third model selected by the Panel (CALPUFF). (The choice of models was governed by the past experiences of Panel members.) However, it was necessary The final model selections from the MKB for observation and examination by the Panel include CALPUFF (The Illustrative Air Model - see Appendix C-l in this Report); IPM (Integrated Planning Model - The Illustrative Economic Model - see this Appendix C-2 in this Report); and AQUATOX (The Illustrative Water Quality Model - see Appendix C-3 in this Report). Other models are discussed generally in Appendix C-4 of this Report. 42 ------- to go beyond these models to address Charge Question 7a. The Panel interprets this as being asked in the context of a model developer who might use the MKB to screen existing Agency models for use in a specific application or for existing model technology to include in a new model to support a specific decision. In this case the Panel found it necessary to identify a number of similar models (i.e., atmospheric dispersion models or water quality models) and assess first the number of models available to choose from and, second, the consistency, transparency and comparability of the data for these similar models. In answering CQ 7a, the Panel finds that the MKB has the potential to provide readily accessible information about models; however the amount and quality of information can be improved. For CQ 7b, the Panel recognizes that the information provided in the MKB is not highly detailed. As a result, sufficient level of detail about scales of data used and assumptions made during the formulation of any specific model in the MKB cannot be obtained from this tool alone. However, the MKB does allow for the initial identification of candidate models with links and references for obtaining further information. For CQ 7c, the Panel agreed that the three models considered in this review were all good examples of successful models both in their regulatory role and in the way they are presented in the Knowledge Base. For the final Charge Question, the Panel was not satisfied with the current form of feedback mechanism for the Knowledge Base. More detailed observations, suggestions and recommendations follow. 7.2 Vision for the Knowledge Base The issues surrounding which models to include in the MKB are not trivial; the Panel recognizes that this choice can have significant implications for the application of this tool in support of decision makers. The Panel is concerned that without a clear vision, the MKB may increase the burden on Regional and State offices by implying that a particular model is "endorsed" by the Agency. The disclaimer on the main page of the MKB makes it clear that models in the Knowledge Base are not endorsed by the Agency but the Panel suggests that this disclaimer be clearly presented at the top of each "Model Report" page as well. Part of the Vision for the MKB should specify the role of this resource in the development or life cycle of models. More specifically, there needs to be a clear statement about what models are included in the Knowledge Base and what models or types of models are excluded (if any). This will require that the Agency provide a clear definition of "Regulatory Model," or else that it move away from this restrictive terminology towards a more inclusive title. The Panel recognizes that in addition to providing a repository or library of mature models that are actively used by the Agency, the MKB can also play an important role in the development of new models and the improvement of existing models. For this reason, the Panel recommends that the Agency include models at all stages of their life cycle with a process for identifying to users those models that are new, actively being developed, currently used for decision-making and nearing retirement. 43 ------- An important aspect of any model repository from the perspective of a model developer or new model user is that it be as comprehensive as is feasible. In other words, users must be confident that when they use the MKB to identify an appropriate model for a task, it is likely that all relevant models have been considered. The draft MKB provides a good start but needs to continue to incorporate additional models used by the Agency. Many of the Agency's Offices, Programs, and Regions have developed their own clearinghouse for models; the Agency should make an effort to bring these existing data bases under the umbrella of the Knowledge Base. The Panel recommends that the Agency identify these parallel Agency supported databases (e.g., the Support Center for Regulatory Air Models (SCRAM), the Center for Exposure Assessment Modeling (CEAM), etc.) and develop a plan to incorporate them into the MKB. If it is not feasible to incorporate these existing databases at this time, then the Panel suggests providing a current list of- and links to - these additional databases on the main page and the search page of the MKB. In addition, there are ongoing efforts outside the Agency that are focused on developing common model documentation protocol (Benz and Knorrenschild 1997) and a searchable web-based registry for existing models (Benz et al. 2001) that may provide useful insight during the continued development of the MKB. The process of identifying and including existing models is clearly an important step to insure that the Knowledge Base is comprehensive. It is also important to continue to populate this MKB with new models as they emerge. To accomplish this, the Panel recommends that the Agency incorporate new models into the Knowledge Base as part of their initial application within the Agency. The information in the MKB for a given model is, or should be, part of the model development process so submitting this information as part of a model's initial application should not be an added burden to the model developers. Nevertheless, the Panel recognizes that it may be necessary for the Agency to provide additional incentive (or penalties) as part of their plan to encourage what is currently a voluntary effort by modelers to put their models in the MKB. To insure the continued improvement of what appears to be an extraordinarily valuable model information system, the Panel recommends that the Agency consider appointment of a Knowledge Base "System Librarian." This appointment might come from within the Agency or from an appropriately qualified contractor. The position would emphasize consistency in data collection and input of new information as well as system QA to improve information reliability with time, making the MKB a national resource for quality comparative information on both new and established models used in the regulation of the environment. 7.3 Quality Assurance and Quality Control In addition to its role as an institutional memory, the MKB, in its current form, is clearly a tool designed and developed to support regulatory decisions by delivering useful information about prospective models for specific applications. The database itself is not unlike other "models" developed to support regulatory decisions. As noted in CQ6, the development of the MKB and the information provided in it should be subject to the same level of quality control and quality assurance that any Agency modeling effort is expected to include. Therefore, in 44 ------- addition to the Vision Statement discussed earlier, the Panel recommends that the Agency provide a link on the main page of the Knowledge Base that takes the user to the Agency's plan for insuring the quality (integrity, utility and objectivity) of information provided. At a minimum, this should contain the following elements: 1) Problem specification that identifies the drivers for setting up the MKB (i.e. reduce duplication of effort, improve networking, facilitate model development, satisfy training needs, etc.); 2) Clear identification of the user community or "clients" for the MKB. There was some ambiguity among the Regional representatives at the face-to-face meeting about whether the Knowledge Base satisfied their specific modeling needs and as a result there appeared to be a lack of "buy in" from the Regions; 3) Identify specific performance criteria for the MKB information along with selection criteria for models in the database and identify who will be responsible for insuring that these criteria are met; and 4) If non-Agency models are eventually included in the MKB (see previous bullet on selection criteria) then the QA/QC plan should identify how these models will be treated or presented and who will absorb the burden of oversight for these models. The level of detail provided by each model should also be balanced. In the draft MKB, the details provided for models differ widely. An example of a model where information is very sparse is TRACI. Scientific detail is often just a statement of units used in the model (e.g., the SWTMODEL includes only the following statement under Scientific Detail "The model uses fixed units (S.I.)." and is missing Conceptual Basis all together). In other cases, it is not apparent that the sections include comparable information. For example, it is often difficult to distinguish between the Conceptual Basis, Scientific Detail and the Model Framework sections. The Panel recommends that improved guidance be provided as part of the data entry sheet to insure that the correct type of information is input into each field. This will also facilitate search functions by making sure those submitting the information realize what fields are searched. It may be necessary to request a keyword list from the model developer. As an example of this last point, the Panel found that the CALPUFF model was not identified in the keyword search using the phrase "air dispersion." Although "air" and "dispersion" are in the title or abstract, the phrase "air dispersion" is missing and as a result the model is not identified when the search is based on this common phrase. In another case, a search for "vapor intrusion" models (currently a timely topic) found no matches in the MKB. A search for "indoor air" models produced three matches, but none that appeared usable for the vapor-intrusion set of problems. This illustrates that there is still some significant work ahead to verify that the priority regulatory problems being addressed in Regional offices of EPA today are adequately considered in selecting candidate models to be included in the MKB. 45 ------- 7.4 Layout and Navigation of Knowledge Base In addition to the recommendations already provided in Section 5, the Panel identified several pieces of information that should be elicited when a model is introduced into the Knowledge Base. In this section, the Panel provides observations about the current layout of the MKB and provides suggestions for where new information should be presented. The current layout of the MKB is logical and generally easy to maneuver (with some exceptions noted later). The Panel found that much of the summary level material was readily accessible on the three main Report pages. The more detailed information is generally available through appropriate links. However, the Panel notes that in several cases, including the CALPUFF model, information is not provided for specific fields and rather than leave these fields blank, they are apparently removed from the Report. For example, the "Model Framework" and the "Model Evaluation" fields are often missing. The Panel recognizes that the Agency attempted to "cull information about models that broadly serve the needs of all users. but once this minimum information is identified, it should be provided for all models. The Panel recommends that if information is not provided for specific fields, those fields should be left blank rather than be removed from the Report. A blank field provides clear information about a model while a missing field is ambiguous. Overall, it was possible to use the MKB to obtain general information about the existence and availability of frequently used models and more detailed information about a specific model. However, a real understanding about how a given model works and what its specific strengths and weaknesses are would appear to require either going into the detailed documentation or contacting an actual user. Navigating the Knowledge Base was somewhat cumbersome, in that apparently different links go the same destination, links to critical information (e.g., model change bulletins) are obscure and the return link from the exit disclaimer page forwards the user to the keyword search page. In addition, several different pages (10 in the case of CALPUFF) needed to be accessed to gain a sense of model operation and capabilities. Perhaps accommodating the somewhat bewildering array of models and their varying characteristics is the cause of these navigational inefficiencies. Nevertheless, the Panel recommends that the MKB system be reconfigured so as to streamline access to model information. 7.5 Updating the Knowledge Base The Panel recognizes that the MKB is a "living demonstration of the recommendations from the Guidance for Environmental Models." This suggests that the Knowledge Base will evolve and adapt to the specific needs of the user community. The comments above also support the premise that this will be an ongoing process of optimization. Optimizing the MKB will ultimately require an understanding of the user community and an active and transparent feedback mechanism. To facilitate this, the Panel recommends that voluntary user profile 46 ------- and registration information be requested so that use profiles can be developed. This information can also provide a mechanism for announcements to be distributed when necessary. Improving the MKB and the models contained in it will ultimately depend on the quality of feedback from "external users" and the ability of new users to access this information. The MKB is currently limited to a single contact and does not provide any suggested format for comments, nor does it provide for open dialogue and discussion of users' modeling experiences. This seriously limits the Agency's ability to adapt the MKB and improve its utility. This lack of an open forum also limits the model developers from gaining experience from model users and it limits the ability of new modelers to learn about specific experience and application of a particular model The Panel recognizes the challenges associated with hosting an open forum on an Agency web site but recommends that the Agency reconsider including a transparent user feedback mechanism that will facilitate an open dialogue for the models in the MKB. 7.6 The Role of the Knowledge Base as a "Model Selection Tool" The Panel is not entirely convinced of the utility of a model selection tool or expert system that accesses the MKB to facilitate model selection. However, the Panel suggests that if such a tool is developed for application at the EPA Regions, labs and states, then the effort should be considered "model development" and as such should clearly follow the principles presented in the Draft Guidance. If such a model selection tool is developed, it will likely be used early in the life of a project, thus identifying and evaluating specific needs in a way that would facilitate a ranking of models that would otherwise be difficult to achieve. Therefore the Panel recommends that any tool developed by the Agency to facilitate model selection based on the Knowledge Base should simply present the models in a comparative matrix in the form of a side-by-side comparison table like one would see in the car sales industry. Appendix C provides more detailed information about Panel members' experiences in accessing and using specific models. 47 ------- REFERENCES Amar, P.; R. Bornstein; H. Feldman; H. Jeffries; D. Steyn; R. Ramartino; and Y. Zhang (2004). Review of CMAQ Model, December 17-18, 2003. Submitted March 1, 2004. Available at: www. gov/cair/pdfs/PeerRevie w_of_CM AQ. pdf Benz, J. and M. Knorrenschild, (1997) "Call for a common model documentation etiquette". Ecological Modelling, 97, 141-143. Benz, J. and R. Hoch and T. Legovic (2001) "ECOBAS — modelling and documentation" Ecological Modelling 138 3-15. Bostrom, A., B. Fischhoff and M.G. Morgan. 1992. "Characterizing Mental Models of Hazardous Processes: A Methodology With an Application to Radon," J. Social Issues, 48(4), 85-100. Cullen, A.C., and Frey, H.C., 1999, Probabilistic Techniques in Exposure Assessment: A Handbook for Dealing With Variability and Uncertainty in Models and Inputs, ISBN 0-306- 45956-6, Plenum Press, NY. Cullen, A.C. and M.S. Small. 2005. "Uncertain Risk: The Role and Limits of Quantitative Assessment," In Risk Analysis and Society: An Interdisciplinary Characterization of the Field. Edited by T. McDaniels and M. Small, Cambridge University Press, Cambridge, UK. Dakins, M.E., Toll, I.E., Small, M.J., Brand, K.P. 1996. Risk-based environmental remediation: Bayesian Monte Carlo analysis and the expected value of sample information, Risk Analysis, 16: 67-79. Finkel, A.M. and Evans, J.S., 1987, "Evaluating the benefits of uncertainty reduction in environmental health risk management," J Air Pollut Control Assoc, 37:1164-1171. Finkel, A.M. 1990. Confronting uncertainty in risk management, Washington, D.C., Center for Risk Management, Resources for the Future, Washington, D.C. Fischhoff, B., D. Riley, D.C. Kovacs and M. Small. 1998. "What information belongs in a warning?" Psychology and Marketing, 15(7): 663-686. Grand Canyon Visibility Transport Commission (GCVTC) (1996) Recommendations for Improving Western Vistas: Report of the Grand Canyon Visibility Transport Commission to the United States Environmental Protection Agency. Dated June 10, 1996. Available at: http://wrapair. org/WRA P/Reports/GCVTCFinal. PDF Kleijnen, J.C. 2005. "An overview of the design and analysis of simulation experiments for sensitivity analysis". European Journal of OperationalResearch,Vo\. 164, No. 2, pp.287-300. 48 ------- March, J.G. and Simon, H.A., 1958, Organizations, John Wiley and Sons, New York. Massmann, J., R.A Freeze, L. Smith, T. Sperling, B. James. 1991. Hydrogeological decision analysis: 2. Applications to ground-water contamination. Ground Water, 29(4): 536-548. Merz, J., M.J. Small and P. Fischbeck. 1992. "Measuring decision sensitivity: A combined Monte Carlo-logistic regression approach," Medical Decision Making, 12: 189-196. Morgan, M.G., Henrion, M., and Morris, S.C. 1980. Expert Judgment for Policy Analysis, Brookhaven National Laboratory, BNL 51358. Morgan, M.G. and Henrion, M. 1990. Uncertainty: A Guide for Dealing With Uncertainty in Quantitative Risk and Policy Analysis, Cambridge University Press, Cambridge, UK. Morgan, M.G., B. Fischhoff, A. Bostrom, L. Lave, C. Atman. 1992. "Communicating Risk to the Public," ES&T, 26(1 1), 2048-2056. Raiffa, H., 1968, Decision Analysis: Introductory Lectures on Choices Under Uncertainty, Addison-Wesley Publishing, Reading, MA. Ramaswami, A., J.A. Milford and M.J. Small. 2005. Integrated Environmental Modeling: Pollutant Transport, Fate and Risk in the Environment. John Wiley & Sons, New York. Saltelli, A., K. Chan and M. Scott, eds., 2000. Sensitivity Analysis, John Wiley and Sons, New York Thompson, K.M., D.L. Bloom. 2000. "Communication of risk assessment information to risk rs" Journal of Risk Research, 3(4): 333-352. U.S. Congress. 2001. Pub L. No. 106-554.2001. The Data Quality Act, Section 515 of the Treasury and General Government Appropriations Act for Fiscal Year 2001, Pub L. No. 106-554 U.S. EPA. (no date). CREM Background Materials: A web version of the CREM related background information with links to pertinent documents, is available at: U.S. EPA. (no date). On-Line Models Knowledge Base (MKB, or KBase) Link is available at: http ://cfpub . epa. gov/crem/knowl ...... base/knowbase.cftn U.S. EPA. 1994. Agency Guidance for Conducting External Review of Environmental Regulatory Modeling, 1994. Available at: 49 ------- U.S. EPA. 1997. Final Report to Congress on Benefits and Costs of the Clean Air Act, 1970 to 1990. Report EPA 410-R-97-002. Available at: http ://www.epa. eov/air/sect8 1 27 U.S. EPA. 1999. Final Report to Congress on Benefits and Costs of the Clean Air Act, 1990 to 2010. Report EPA 4 10-R-99-001. Available at: http://www.epa.gov/air/sect8 12/ U.S. EPA. 2000. Framework for the Council on Regulatory Environmental Modeling. Available at: http ://www. epa. gov/osp/crem/1 i brary/crem%20framework. htm U.S. EPA. 2001. Memorandum from Dr. Gary Foley to Dr. Donald G. Barnes entitled "Request for Science Advisory Board Review of a draft outline of a proposed document entitled 'Guidance on Recommended Practices in Environmental Modeling, '" October 4, 2001 U.S. EPA. 2002. Memorandum from Christine Todd Whitman entitled "Strengthening Science at the Environmental Protection Agency, " May 24, 2002. Available at: http://epa.gov/osa/pdfs/saduties.pdf U.S. EPA. 2003a. Memorandum from Administrator Christine Todd Whitman entitled "Council for Regulatory Environmental Modeling, " February 7, 2003 [designating Dr Paul Gilman as the EPA Science Advisor and asking him to revitalize the CREM and accelerate its efforts.] Available at: httpj//ctj3ub,^ U.S. EPA. 2003. Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models, Prepared by The Council for Regulatory Environmental Modeling (CREM), November 2003, 60 pages. Available at: http ://www. epa. gov/ord/crem/library/CREM%20Guidance%20Draft%20 1 2 03 .pdf U.S. EPA. 2005. Guidance on the Use of Models and Other Analyses in Attainment Demonstrations for the 8-hour Ozone NAAQS (Draft Final), February 17, 2005, EPA Support Center for Regulatory Air Models. Available at U.S. EPA SAB. 1989. Resolution on the Use of Mathematical Models by EPA for Regulatory Assessment and Decision-Making, Modeling Resolution Subcommittee of the Environmental Engineering Committee, Science Advisory Board, EPA-SAB-EEC-89-012, January 13, 1989. Available at: htt|3i/Avww.j;|)a^ U.S. EPA. SAB. 1990. Review of the CANSAZ Flow and Transport Model for Use inEPACMS, Report of the Saturated Zone Model Subcommittee of the Environmental Engineering Committee, Science Advisory Board, EPA-SAB-EEC-90-009, March 27, 1990. Available at: U.S. EPA. SAB. 1995. Commentary on Bioaccumulation Modeling Issues, Report from Bioaccumulation Subcommittee, Science Advisory Board, EPA-SAB-EPEC/DWC-COM-95- 50 ------- 006, September 29, 1995. Available at: http://www.epa..gov/osp/crem/library/sab_bioaccumulation.pdf U.S. EPA. SAB. 2002. Panel Formation Process: Immediate Steps to Improve Policies and Procedures: An SAB Commentary, EPA Science Advisory Board, EPA-SAB-EC-COM-02-003, May 17, 2002 Yokota, F. and K.M. Thompson. 2004. "Value of information analysis in environmental health risk management decisions: Past, present, and future," Risk Analysis, 24(3): 635-650. HOTLINKS FOR SELECT SOURCES ARE AS FOLLOWS: AQUATOX: Available at: httg;//_cfjub^ CALPUFF: Can access this through the MKB (KBase) link at: http ://cfpub . epa. gov/crem/knowledge_base/knowbase.cfm EPA Center for Exposure Modeling (CEAM) Web Site at: Integrated Planning Model (IPM): Available at: http ://cfpub .epa. gov/crem/crem ....... report, cfm? dei d=749 1 9 Models Knowledge Base (MKB, or KBase) Link is available at: National Water Pollution Control Assessment Model (NWPCAM): Available at: http://cfpub.epa.gov/crem/crem_report.cfm?deid=749 1 8 U.S. EPA. 2003. Draft Guidance on the Development, Evaluation, and Application of Regulatory Environmental Models, Prepared by The Council for Regulatory Environmental Modeling (CREM), November 2003, 60 pages. Available at: http ://www. epa. gov/ord/crem/librarv/CREM%20Guidance%20Draft%20 1 2 03 .pdf 51 ------- Appendix A - Enhancements to the Glossary Consensus on a common nomenclature is a key requirement for implementing a consistent Agency-wide approach for environmental model development, use and quality assurance. The Glossary in the draft document is a preliminary step towards this goal. However, several aspects of the Glossary would benefit from additional technical and editorial attention: 1) The reader is likely to be frustrated when looking up underlined terms from the text when the terms are not listed in the Glossary in the same form that they appear in the text, e.g.: Spatial and temporal domain (p. 9; listed under Domain in glossary), code verification (p. 12), model evaluation (p. 16), model validation (pp. 16 and 43; also appears on p. 30 in the definition for corroboration), integrity (p. 16), proprietary models (p. 23). 2) Several terms are defined in the Glossary slightly differently from their definitions in the text; it is suggested that the definition be the same in both locations. Module (Box 2 on p. 37); Terms from Box 3: Applicability and Utility, Clarity and Completeness, Evaluation and Review, Objectivity, Uncertainty and Variability. Application Niche Uncertainty (p. 21). 3) Several terms are not in alphabetical order in the Glossary: Expert Elicitation, False Negatives, Forms (models), Model, Parameter Uncertainty, Quality,Variability. 4) Several additional terms should be added to the Glossary (and underlined in the text) and either defined at that location, or else cross-referenced to another existing term in the Glossary for the definition (as has been done for "Parameter Uncertainty"): Acceptance Criteria (Box 3), Bayesian view (p. 56), Beta test, bootstrap sampling (p. 48), Bug (computer), Configuration tests, Data, Data Acceptance Criteria (p. 43), Empirical data (p. 21 and 45), Errors, hyperplane (p. 51), Integration Tests (App B), Monte Carlo analysis (p. 53), Normal Distribution (p. 45), Paradigm (App C), Parameterize, Peer Review, Platform, Post-processing (model output), Qualifiers (for analytical data) (Box 5 on p. 43), Quality Assurance, Regimes (p. 48), Representativeness (p. 20; Box 5 on p. 43), structural error (p. 21), Type I error (p. 45), Type II error (p.45), User interface (p. 33, used in definition for Object-Oriented Platform). 5) Cross-references to more specific terms in the Glossary should be added to the definitions for generic terms, e.g.: a) Decision errors: See also False Negatives, False Positives, b) Errors: See also Accuracy, Bias, Data Uncertainty, Confounding Errors, Data Uncertainty, False Negatives, False Positives, Measurement Errors, Model Framework Uncertainty, Noise, Uncertainty, Uncertainty Analysis, Variability, and c) Model: See also Conceptual Model, Deterministic Model, Empirical Model, Mechanistic Model, Screening Model, Simulation Model, Statistical Model, Stochastic Model. 52 ------- 6) The definition of the term "model complexity" should be expanded to emphasize theoretical and process issues first (e.g. basis of the model; spatial and temporal resolution). The mathematical, numerical, and computational aspects of complexity should assume a secondary posture. 7) In addition to the Glossary, Since the environmental modeling is generally multidisciplinary, the Agency should consider adding a List of Acronyms to the Draft Guidance. Appendix E of this report provides an initial list. 53 ------- Appendix B - The CREM Models Knowledge Base Data Entry Sheet Instructions 1. Please complete this data entry sheet for each model that you want to be included in the CREM Models Knowledge Base. You may use as much space as necessary. 2. You are encouraged to include URLs to other sources of information, graphics, and other pertinent documents (in PDF or other formats). 3. The data entry sheet for the IPM model is provided as an example. 4. Any questions? Need assistance? Please contact Neil Stiber (202-564-1573). (A) General Information 1. Model Name: 2. Model Overview / Abstract: 3. Contact Information (name, affiliation, e-mail, phone #): 4. Model's Home Page: (B) User Information 1. Technical Requirements: a) Computer Hardware: b) Operating Systems: c) Programming Languages: d) Other Req'ts and Features: 2. Download Info (with URL): 3. Using the Model: a) Basic Model Inputs: b) Basic Model Outputs: c) User's Guide: d) Other User Documents: (C) Model Science 1. Conceptual Basis of the Model: 2. Scientific Detail for the Model: 3. Model Framework (equations and/or algorithms): 4. Model Evaluation (verification (code), corroboration (model), sensitivity analysis, uncertainty analysis): 54 ------- (D) Model Criteria Please use the shaded boxes on the left to select all criteria that are relevant to the model. Criteria should be selected based on an appropriate level of generality / specificity. Please note that selection of specific criteria (e.g., "Pollutant Type"); necessarily includes the more general (e.g., "Releases to the Environment") but, not the more specific (e.g., "Physical"). Regulations Clean Air Act (CAA) Clean Water Act (CWA) Safe Drinking Water Act (SDWA) Resource Conservation and Recovery Act (RCRA) Comprehensive Environmental Response, Compensation, & Liability Act (CERCLA) Federal Insecticide, Fungicide, and Rodenticide Act (FIFRA) Toxic Substances Control Act (TSCA) Releases to the Environment Pollutant Type Physical (e.g., radiation, heat, particles, fibers, noise) Chemical (e.g., organic, inorganic, toxics) Biological (e.g., microbial) Source Type Point source (e.g., tank, spill, stack, discharge pipe) Area source (e.g., spray, fertilizer, lagoon, holding area) Mobile source (e.g., automobiles, trains, ships, airplanes) Ambient Conditions Media Ground (e.g., soil, sediment) vjujinm. \^.g.; owii; pi/u-inii/iity Water (e.g., surface water, ground water) Air Ecosystem Processes Transport (e.g., advection, bulk, dispersion, diffusion) Transformation (e.g., chemical reaction, partitioning, biodegradation) Accumulation (e.g., deposition, sedimentation) Biogeochemical (e.g., cycling, growth, consumer-resource) Exposure or Uptake Exposure Characterization Location Frequency and Duration Pathway (e.g., inhalation, digestion, dermal, injection) Body Burden - Dose (e.g., phamacokinetics, retention, transformation) Changes in Human Health or Ecology Human Health Indicators Mortality Chronic and Acute Diseases Ecological Indicators Population Changes Acute & Chronic Disease Occurrence Land Use Change 55 ------- Appendix C - Panel Members Experiences Using the MKB This appendix summarizes comments related to the form and function of the Models Knowledge Base with specific emphasis on models selected to facilitate the review and response for Charge Question 7. Because the following narrative is meant to convey the individual reviewers experience with the MKB during the review, the narrative has not been heavily edited. C-l CALPUFF (The Illustrative Air Model) The CALPUFF example evaluation starts from the "Models Knowledge Base" page, and then goes to the listing of available models, and from that to the CALPUFF model report. With respect to CQ 7(a), if the user wasn't going to a specific model, it would be hard to decide, using this list alone, how to choose from among the several seemingly air-related models listed (however, the keyword search capability is helpful for this). A model overview on the "general information" page provides information that addresses, in part, CQ 7(b). Going to the "user information" page gives information on downloading and the availability of user's guides. Here the heading "Using the Model" is slightly misleading in that it implies information on how the model is used to make decisions but is actually about how a modeler would run the model. This section also provides no citations or links as to application of model results in actual decision making. Although the "Recommendations for Regulatory Use" section is informative, it also provides no citations or links as to how model results have fared in actual decision-making. The "Model Evaluation" section is clearly about evaluation of the model as a model and not as a decision support tool. The MKB does provide sufficient information to accomplish goal 7a for the CALPUFF model in that it allows users of the data base to locate candidate models which might serve their purpose. However, it should not be considered as providing a substitute (e.g., in summary report form) of the detailed information that has to be retrieved from the open literature in order to compare potentially relevant models for an application. It would be impractical for the MKB to provide the level of information necessary for users to determine which models are suitable for every application, but it can certainly help eliminate duplication by providing a limited number of candidates to consider. Evaluating these candidate models requires consistency in the presentation of information. The MKB cannot reasonably be expected to provide sufficient detail to fully address a model users/developer's questions about CALPUFF. However, it can and should answer basic questions such as "at what temporal and spatial scales has the model been shown to operate successfully?" and (for air models in the GAQM) "at what scales are these models considered to be preferred or acceptable alternatives?" This information should be sufficient to guide users of the MKB to ask the right questions, but probably cannot provide complete answers, since understanding the "endogenous assumptions made during model formulation" will require detailed understanding of the model algorithms beyond its scope. The models presently in the MKB differ widely in terms of ranges, attributes, objectives etc. The completeness/focus of the "model report" information also varies widely 56 ------- relative to the amount of information provided. For example, under User Information, essentially all that is provided for CALPUFF is links to the SCRAM and to the developer's web site, but for some other vendor-supplied models, summary information is provided in the MKB itself (plus appropriate links). Because vendors may provide information on models as they see fit, it would be beneficial to have at least a summary of basic information about each model in the MKB. As indicated in the Panel's Report, this information should include computational requirements (including operating systems supported and requirements for other software), descriptions of input data requirements, and descriptions of model output. Additional useful information could include some examples where the model was successfully applied, along with references and contact information to facilitate further research into the suitability of models for specific applications. As another example of the need for consistency, the CALPUFF site under the "user information" section, the link to "Technical Requirements" is missing. To facilitate identification of all candidate models for a specific task, each model should have the same major sections. Similarly, the Framework section on the Model Science page is missing for CALPUFF (as well as for AquaTox). Even if sections are left blank, they should be included for every model to facilitate use of the MKB. The main page of the CALPUFF model developer's website provides little information about the science of the model but does nicely summarize model updates, provides links to its regulatory status, a download, and training opportunities. The "regulatory status" page provides information similar to that found on the EPA "model science" page but goes further by offering links to notices and reports on regulatory use. This also highlights the need for some support by the Agency to synthesize information provided by the model developer in order to provide a consistent format and level of detail. Navigating the CALPUFF pages was somewhat awkward. The Environmental Indicators search feature was the least useful since it presupposes knowledge of how the Agency defines and uses such indicators. One of the download links from the "user information" page leads to EPA's SCRAM website, as does a similar link for "model homepage" on the "general information" page. The SCRAM website is apparently the only point at which it is possible to access the critical "Model Change Bulletin" and "Model Status" records, which are somewhat obscurely included only as "Notes" in smaller font. There appears to be considerable overlap in these two sets of information and the question arises why they couldn't be combined in one more accessible location (e.g., on the "user information" page). The link to the NTIS site is probably necessary but models without online documentation would appear to be at a disadvantage. Getting to CALPUFF on the SCRAM website from either the "general information" or "user information" pages provides one with a link to the model developer's website, who is a contractor and not the EPA. A link directly to this website is also on the "user information" page. Thus, there are three apparently different links on two different pages all leading to the same destination, a non-EPA website. This seems unnecessarily convoluted. It is not entirely clear until this point that genuinely useful information on the model resides with a contractor and not with the Agency. Something seemed to be wrong in the keyword search feature on the MKB primary panel, since entering "air dispersion" produced only three results, all related to the RAIMI. 57 ------- This search should produce several hits including CALPUFF. The Panel recognizes that the search is only performed on the title and abstract so if the word or phrase is missing from this field it will not be found. In CALPUFF, the abstract does not include the word "air" so it is not picked up by searching for "air dispersion." The "browse for models by selecting for environmental indicators" seems to have no search criterion which locates CALPUFF either. Also, after inadvertently selecting "Exit Disclaimer" on the CALPUFF User Information page, the "Return to Previous Page" takes the user to the "Browse to Knowledge Base" page rather than the previous page. On the CALPUFF model developer's website, a reference is made to the Guidelines on Air Quality Models (GAQM), while in the MKB, there is a reference to "Appendix W." In fact, both refer to the same document. The MKB should be clear that Appendix W to 40 CFR Part 51 and the GAQM are the same. Both the Model Knowledge Base and the model developer's web sites should provide links to the GAQM. The MKB includes many highly successful models (including CALPUFF), but it is not clear how users will be able to determine for themselves which ones are "successful." Clearly models "preferred" in the GAQM qualify, but a similar gold standard may not exist for other media. Other GAQM models may be assumed to have achieved some measure of "success." A list of the applications of a model could be useful in providing a measure of its success. To allow one to judge the level of success of a particular model, the summary report should provide a very simple summary of the "applicability range" of the model. For example the summary report states that "CALPUFF" is intended for use on scales from tens of meters from a source to hundreds of kilometers" but does not mention the fact that the minimum temporal resolution of the model (hourly averages) restricts its applicability to a simulation range that does not include important short-term phenomena (e.g., emergency events such as accidental spills), dispersion of heavy gases, etc. As indicated in the Panel's report, especially important information that should be included in the MKB are i) all input/output formats, ii) all software tools (public domain and proprietary - as well as potential substitutes) that are needed in order to fully utilize the model's capabilities, iii) available databases of inputs (potentially outputs from other models), and iv) past evaluations (especially cross-evaluations) studies involving the model(s) of concern. The MKB provides the opportunity to turn abstract discussions in the Guidance into specific examples; however, in order to achieve this, more detailed and consistent information needs to be included in the MKB. The role of the EPA as the "model contact" is not clear for the feedback forum. The appropriate or desired role of the model contact as either an Agency or external interface for the model should be made clear at this stage of the development of the MKB. It would also seem that a more direct link to the actual developer and maintainer of the model would be helpful. The MKB appears to have no formal feedback mechanism other than contacting Mr. Pasky Pascual. Feedback from model users could be extremely valuable to others who have specific modeling needs. The information would help users answer the charge questions posed in 7a-c. The MKB could solicit comments from users of the models, and post these comments on a bulletin board. Postings should allow for anonymity, as some model users might not want 58 ------- to be identified personally as users of the models - for example it is not unusual for busy modelers to get phone calls from graduate students wanting help running complex environmental models for thesis projects. C-2 The Integrated Planning Model (IPM - The Illustrative Economic Model) The write up on IPM in the MKB is very thorough. It is clear, concise and helpful as a first description of what this model contains and what it is used for. It turns out that almost all of the write-up is a verbatim cut and paste from the IPM Model Documentation. This is sufficient as long as the appropriate items are covered at sufficient depth. However, in examining the IPM Model Documentation, page 2-5 begins a section on Key Methodological Features (e.g., details of how the load duration curve is specified and information on how the dispatch order is determined) that could be simplified and incorporated into the MKB to bring the reader one level further down in detail. Thus, to maintain consistency in the level of detail presented in the MKB, it may be necessary for existing documentation to be re-written with a consistent format across all models. It is recognized that this would likely require a scientific editor/webmaster dedicated to the task of working with the model developer to prepare the documentation for upload onto the MKB. The Panel recognizes that the MKB alone is unlikely to provide sufficient information for new model developers who may require a detailed understanding of potentially competing models. This type of information can only be obtained, if at all, from model documentation. The IPM site, which can be accessed from the MKB, does contain links to such detailed documentation. In this sense new modelers may benefit. On the other hand an internet search or a search of the EPA's website would immediately bring up such documentation without the need for the MKB. New developers would be particularly keen on knowing the IPM's limitations and assumptions, none of which seems to be available. IPM in particular is extremely well entrenched in the Air Office and would be, therefore, unlikely to attract "new model developers." The level of detail on "endogenous assumptions" for a given model is dependent on the information provided by the model developer, so at some level this may be out of the realm or control of the developers of the MKB. Evaluating the utility in contrast to other models requires first that competing models be identified through the MKB, and second that the MKB provide enough information at a comparable level of detail so that appropriate choices on which model to use can be made. A high spatially resolved model is expected to be more accurate than one of lower resolution, but choices about resolution always involve tradeoffs, such as in model complexity, data availability, model flexibility, and the types of questions a model is designed to answer. The charge question does not encourage this kind of thinking (although earlier questions may) and the database is silent on providing information to aid in this type of thinking as well. For IPM, spatial resolution is clearly given - all 48 states plus DC are covered along with a number of coal producing regions that are identified. The forecasting horizon of the 59 ------- model is clear, however the temporal discretization is not explicit stated. Exogenous assumptions are not fully provided directly on the MKB model page, but model documentation accessed through links would surely provide this information for this model. There is a list of key assumptions (e.g., perfect foresight, pure competition) in the IPM Model Documentation document; this information should be provided in the MKB. Again, as noted earlier, modelers should be asked to provide a write-up for the MKB of significant limitations of their models in terms of simplifications, strong assumptions, and factors that have been ignored and/or are outside the scope. The Panel agrees that the IPM model is a good example of a "successful" model. A forum for feedback on model uses outside Agency applications and a means of collecting external suggestions for updating/improving model structure are currently inadequate. C-3 Aquatox (The Illustrative Water Quality Model) A new model developer would find the documentation and descriptive material on the technical and theoretical aspects of AQUATOX very helpful in eliminating duplication of effort. Processes in the model are well documented in the MKB and the associated model documentation provided on the AQUATOX web site. The technical documentation of Release 2.0 is reasonably thorough with regard to process documentation and assumptions inherent in the model. However, the format of the report does not follow the recommended elements for model documentation given in the Draft Guidance. The Panel would prefer to see a separate "Model Development" chapter that includes a conceptual model, a complete disclosure of all model assumptions and resulting caveats, and data used to convert the conceptual model to a mathematical model. Release 2.0 does specify that it can only be used in a non-dimensional or one-dimensional mode and does discuss the temporal scales of use. There are certainly limitations to use of the model imposed by these assumptions; the document does discuss these. This model has not had a long history of application in its current form, although it does have a long history of application of previous incarnations of the model (e.g., as CLEAN or CLEANER or PEST). The user manual presents several examples of applications of the model; however, only one of them (Onondaga Lake) shows system data that allows the user to assess the success of these applications. On the web site, model "validation" examples are offered in an EPA report published in 2000 that includes Onondaga Lake, PCBs in Lake Ontario, and agricultural runoff in the Coralville Reservoir. It does appear that these evaluation exercises compare AQUATOX with data and previous models for these systems, which is good. There is no discussion of regulatory use of the model. The documentation does make the point that this is a multi-stressor, multi-response model. Finally, the model web site provides an opportunity to become a registered user; however, it is not clear that this is the portal to provide feedback to the Agency on outside application experience or suggestions. 60 ------- C-4 Other Models As noted in the Panel's report, it was necessary to evaluate other models in the MKB in order to assess the level and consistency of detail and ease of use. The following comments are general observations from this survey. The Panel found that figures and diagrams were particularly helpful in the section describing the model's conceptual basis as used in the IPM. The information provided for a number of the models is not necessarily in line with the definition of "Conceptual basis" as described in the Guidance. The descriptions range in detail from providing a statement of what the model does to what inputs are required but not always clear on what the conceptual basis is (i.e., is it mechanistic, or empirical, or something in between). The BLP model has only two of the four sections in the model use section. There also appears to be some confusion between "Scientific Basis" and "Model Framework," which is illustrated by the similar level of information provided in the Scientific Basis section for CALPUFF and the Model Framework section of the IPM. With the IPM it appears that the text was pasted into the sections on conceptual basis, and that the framework section was used to capture overflow text. This reconstruction suggests confusion in populating the MKB system, either on the part of the person who filled out the original Data Entry Sheet or the person who uploaded this information from the data sheet into the MKB system. It would be useful if the web page on "User Information" provided an indication of the level of user expertise required to apply the model. For example, the IPM states that "The model's core LP code is run by ICF Consulting..." while at the other extreme, the THERdbASE states that "User needs only moderate level of technical education and/or modeling experience."This type of information is valuable for users planning to actually apply the models beyond just learning what is available. The Panel found that the level of detail provided in the MKB is very different across models. An example of a model that is very sparse is TRACI. Scientific detail is often just a statement of units used in the model (e.g., the SWEVIODEL includes only the following statement under Scientific Detail "The model uses fixed units (S.I.)," and is missing Conceptual Basis all together). The NWPCAM report is missing the model evaluation section. This speaks to the issue of quality control across the MKB. If the Agency is going to take responsibility for the quality of information provided on these pages, then there will need to be some oversight provided to the various people inputting data in order to get an acceptable level of consistency for the information provided. Or, as indicated earlier, there may be a need for a dedicated scientific editor. The Panel has recommended that the MKB include more detail on model version. A good example of a version tracking matrix or table is given on the PRIZM version index page that is found by following the links to the model web site that goes through the EPA Center for Exposure Assessment Modeling (CEAM) site (http://www. gov/ceampubl/products.htm) by selecting the model from the menu. 61 ------- It is important that the information in the MKB be kept current. It would be helpful for keeping the information up to date if an annual automated message was sent to individuals listed as the model contacts requesting updates or reviews of the material on the MKB. As an incentive, this could be accompanied with a report on the number of accesses that were made to the specific model. The user community for the MKB may provide a very effective policing mechanism to maintain model quality, especially when money is at stake. This provides a clear opportunity and incentive for improving the models it contains. However, this requires a more transparent feedback mechanism, which is currently lacking. Once this resource is developed, the Panel recognizes that the MKB will be a good candidate for technology transfer over the long-term. 62 ------- APPENDIX D-ACRONYMS AA-ship ADV ANSI AQCS AQUATOX ArcINFO ASQC ASTM BLP BMPs CAA CALPUFF CEAM CERCLA CFR CLEAN CLEANER Assistant Administrator-ship (within the U.S. EPA) Advisory American National Standards Institute Analytical Quality Control Services It is a tool in performing ecological risk assessments for aquatic ecosystems. It is a personal computer (PC)-based multi-stressor and multi-response ecosystem model that simulates the transfer of biomass and chemicals from one compartment of the ecosystem to another. It does this by simultaneously computing important chemical and biological processes over time. It predicts the fate of various pollutants, such as nutrients, organic toxicants and various chemicals in aquatic ecosystems, as well as the direct and indirect effects on the resident organisms and their effects on the ecosystem, including fish, invertebrates, and aquatic plants. It has the potential to help establish the cause and effect relationships between chemical water quality and the physical environment and aquatic life. A Geographic Information Modeling System American Society for Quality Control American Society for Testing and Materials Buoyant Line and Point source Gaussian plume dispersion model designed to handle unique modeling problems associated with air dispersion phenomena Best Management Practices Clean Air Act A multi-layer, multi-species non-steady state puff air dispersion model that simulates the effects of time- and space-varying meteorological and air quality conditions on pollution transport, transformation, and removal for assessing long range transport of pollutants and their impacts. Center for Exposure Assessment Modeling (U.S. EPA/ORD) Comprehensive Environmental Response Compensation and Liability Act Code of Federal Regulations Crops, Livestock and Emissions from Agriculture in the Netherlands: A Modeling Tool to Evaluate Policy Options for Reduction of Mineral Surplus, Ammonia Emissions to Air and Nitrogen and Phosphate Emissions to Soil Collaborative Large-Scale Engineering Analysis Network for Environmental Research. A networked infrastructure of environmental field facilities that enables formulation and development of engineering and policy options for the restoration and protection of environmental resources. 63 ------- CMAQ CONCEPT CQ CREM CWA D DECISIONDOCS DFO DOWNLOADINFO DQO EEC EPA EPANET FACA FACT FIFRA FR GAQM GCVTC HTTP IQG Community Multi-Scale Air Quality Model designed to simulate and model a wide range of physical and chemical processes relating to air quality at particular scales in the lower atmosphere. World Health Organization Concept Model of Children's Environmental Health Indicators which emphasizes the complex relationships between environmental exposures and children's health Charge Question Council for Regulatory Environmental Modeling Clean Water Act Dimension (e.g., as 1-D, 2-D, etc.) A Central Database and Clearing House Information System for Communication, Outreach, Terminology, Environmental Data for Monitoring, TMDLs, Water Quality, Ground Water Monitoring, etc. Designated Federal Officer A Listing of US EPA Environmental Models to Provide Information on Dispersion Models Supporting Regulatory Programs Required by U.S.Law Data Quality Objectives Environmental Engineering Committee (U.S. EPA/SAB/EEC) Environmental Protection Agency (U.S. EPA) Environmental Protection Agency Network simulation model. A windows program that performs extended period water network modeling simulation of hydraulic and water-quality behavior within pressurized pipe networks. It tracks the flow of water in each pipe, the pressure at each node, the height of water in each tank, and the concentration of a chemical species throughout the network using a simulation period comprised of multiple time steps. In addition to chemical species, water age and source tracing can also be simulated. Federal Advisory Committee Act (Public Law 92-463) Flow and Contaminant Transfer Model Federal Insecticide, Fungicide, and Rodenticide Act Federal Register General Air Quality Model (Also Guideline on Air Quality Models) Grand Canyon Visibility Transport Commission Hypertext Transfer Protocol (world wide web protocol) Information Quality Guidelines 64 ------- IPM KBase (also MKB) LP MKB (also KBase) NAAQS NAS NCSU NEPA NERL NIST NMSE NRC NSF NTIS NWPCAM OAT OECA OECM OMB ORD PCBs PDF PEST POPs PRIZM Integrated Planning Model. This model is used by the U.S. EPA to analyze the projected input of environmental policies on the electric power sector in the 48 contiguous states and the District of Columbia. It is a multi-regional, dynamic, deterministic linear programming model of the U.S. electric power sector. It provides forecasts of least-cost capacity expansion, electricity dispatch, and emission control strategies for meeting energy demand and environmental transmission, dispatch and reliability constraints. Models Knowledge Base An atmosphere-ocean model code for accumulation and printing of diagnostics for ocean dynamics. Models Knowledge Base National Ambient Air Quality Standards National Academy of Sciences North Carolina State University National Environmental Protection Act National Exposure Research Laboratory (U.S. EPA/ORD/NERL) National Institute of Standards and Technology Normalized Mean Skjuare Error National Research Council [of the National Science Foundation (NSF)] National Sicience Foundation National Technical Information Service National Water Pollution Control Assessment Model. It combines water quality modeling with economic analyses to translate concentration estimates to measures of "beneficial use attainment" used to characterize water quality for policy purposes. This is a national-level water quality modeling system that can simulate water quality changes and economic benefits that result from pollution control policies. It can develop place- specific water quality estimates for most of the nation's inland region. Office of Air Toxics (of the U.S. EPA) Office of Enforcement and Compliance Assurance (U.SA. EPA/OECA) Office of Enforcement and Compliance Monitoring Office of Management and Budget (U.S. OMB) Office of Research and Development (U.S. EPA/ORD) Polychlorinated Bi-Phenyls Portable Document Format (Also Probability Distribution Function - depends on context) Non-linear rjarameter estimation software for any numerical model Persistent Organic Pollutants A risk assessment model for pesticides to estimate environmental concentrations in surface waters (e.g., PRIZM/EXAMS). 65 ------- QA QA/QC QAPP QC QUA QUAL2E QUAL2EU RAIMI RCRA REM REM Panel REV SAB SCRAM SDWA SI Quality Assurance Quality Assurance/Quality Control Quality Assurance Project Plans Quality Control Quantitative Uncertainty Assessment An enhanced stream water quality model which is applicable to well- mixed dendritic streams. It simulates the major reactions of nutrient cycles, algal productions, benthic and carbonaceous demand, atmospheric reaeration and their effects on predicting temperature fluctuations on the dissolved oxygen balance. It is intended as a water quality planning tool for developing total maximum daily loads (TMDLs) and can also be used in conjunction with field sampling for identifying the magnitude and aquatic characteristics of non-point sources. This is an enhancement to QUAL2E which allows users to perform three types of uncertainty analyses, namely sensitivity analysis, first order error analysis, and Monte Carlo simulation. Regional Air Impact Modeling Initiative. A regional air impact modeling tool which is a set of software tools developed by U.S. EPA Region 6 to integrate emissions inventories, air dispersion models, risk models, and population models. EPA and state and local agencies can use this risk-based tool to evaluate the cumulative health impact on local communities of virtually an unlimited number of emissions sources. It has the ability to both predict potential risk to individual neighborhoods and differentiate from hundreds of pollution sources to a few where attention will yield the greatest health benefit. Results are generated in a fully transparent fashion such that risk levels are traceable to each source, each exposure pathway (e.g., inhalation, ingestion),and each contaminant, allowing for prioritization of remedial action based on the potential impact of a contaminant or source on human health. Resource Conservation and Recovery Act Regulatory Environmental Modeling Regulatory Environmental Modeling Panel (U.S. EPA/SAB/REM Guidance Review Panel; also referred to as "the Panel") Review Science Advisory Board (U.S. EPA/SAB) Support Center for Regulatory Air Models Safe Drinking Water Act International System of Units (from NIST) 66 ------- SWIMODEL THERdbASE TMDL TRACT TRIM FATE TSCA UK URLs VOI WASP WoE A (Also referred to as SWMM) dynamic rainfall-runoff storm water management simulation model, primarily but not exclusively for urban areas, for single event or long-term (continuous) simulation. Flow routing is performed for surface and sub-surface conveyance and groundwater systems, including the option of fully-dynamic hydraulic routing. Non-point source runoff quality and routing may also be simulated, as well as storage, treatment and other best management practices (BMPs). Total Human Exposure Risk Database and Advanced Stimulation Environment model. An integrated database and analytical modeling software system for use in exposure assessment calculations and studies. Total Maximum Daily Loading Tool for the Reduction and Assessment of Chemical and Other Environmental Impacts. This tool assists in impact assessment for sustainability metrics, life cycle assessment, industrial ecology, process design and pollution prevention. Total Risk Integrated Methodology Model FATE Module [It is an overall modeling framework intended to provide a flexible method for integrating the release(s) of pollutants from single or multiple sources to their multimedia, multipathway movement in order to predict exposure to pollutants and to estimate human health and ecological risks.] Toxic Siub stances Control Act United Kingdom Uniform Resource Locators Value of Information Water Quality Analysis Stimulation Program. This is a generalized framework for modeling contaminant fate and transport in surface waters and is used in TMDL water quality modeling applications. It is based on the flexible compartment modeling approach, and can be applied in one, two, or three dimensions. It is designed to permit easy substitution of user-written routines into program structure. Problems typically studied include biochemical oxygen demand and dissolved oxygen dynamics, nutrients and eutrophication, bacterial contamination and organic chemical and heavy metal contamination. Wei ght-of-Evi dence End of Document 67 ------- |