Statistical Science

Analysis of Local Decisions Using Hierarchical Modeling, Applied to Home Radon Measurement and Remediation

Andrew Gelman, David H. Krantz, Chiayu Lin, and Phillip N. Price

Full-text: Open access

Abstract

This paper examines the decision problems associated with measurement and remediation of environmental hazards, using the example of indoor radon (a carcinogen) as a case study. Innovative methods developed here include (1) the use of results from a previous hierarchical statistical analysis to obtain probability distributions with local variation in both predictions and uncertainties, (2) graphical methods to display the aggregate consequences of decisions by individuals and (3) alternative parameterizations for individual variation in the dollar value of a given reduction in risk. We perform cost­benefit analyses for a variety of decision strategies, as a function of home types and geography, so that measurement and remediation can be recommended where it is most effective. We also briefly discuss the sensitivity of policy recommendations and outcomes to uncertainty in inputs. For the home radon example, we estimate that if the recommended decision rule were applied to all houses in the United States, it would be possible to save the same number of lives as with the current official recommendations for about 40% less cost.

Article information

Source
Statist. Sci., Volume 14, Number 3 (1999), 305-337.

Dates
First available in Project Euclid: 24 December 2001

Permanent link to this document
https://projecteuclid.org/euclid.ss/1009212411

Digital Object Identifier
doi:10.1214/ss/1009212411

Zentralblatt MATH identifier
1059.62745

Keywords
Bayesian decision analysis hierarchical models small area decision problems value of information

Citation

Lin, Chiayu; Gelman, Andrew; Price, Phillip N.; Krantz, David H. Analysis of Local Decisions Using Hierarchical Modeling, Applied to Home Radon Measurement and Remediation. Statist. Sci. 14 (1999), no. 3, 305--337. doi:10.1214/ss/1009212411. https://projecteuclid.org/euclid.ss/1009212411


Export citation

References

  • (1994). Areas of the United States with elevated screening levels of Rn222. Health Physics 66 50-54.
  • Bogen, K. T. (1997). Do U.S. county data disprove linear nothreshold predictions of lung cancer risk for residential radon? A preliminary assessment of biological plausibility. Human and Ecological Risk Assessment 3 157-186.
  • Boscardin, W. J., Price, P. N. and Gelman, A. (1996). A spatial random effects model for indoor radon concentrations in Washington State. Technical report, Energy and Environment Div., Lawrence Berkeley National Lab.
  • Carlin, B. P. and Louis, T. (1996). Bayes and Empirical Bayes Methods for Data Analysis. Chapman and Hall, London.
  • Cohen, B. L. (1995). Test of the linear-no threshold theory of radiation carcinogenesis for inhaled radon. Health Physics 68 157-174.
  • Cole, L. (1993). Element of Risk: The Politics of Radon. Oxford Univ. Press. Dakins, M. E., Toll, J. E., Small, M. J. and Brand, K. P.
  • (1996). Risk-based environmental remediation: Bayesian Monte Carlo analysis and the expected value of sample information. Risk Analysis 16 67-79.
  • Eddy, D. (1989). Screening for breast cancer. Annals of Internal Medicine 111 389-399.
  • Eddy, D. (1990). Screening for cervical cancer. Annals of Internal Medicine 113 214-226.
  • Englehardt, J. D. and Peng, C. (1996). A Bayesian benefit-risk model applied to the South Florida Building Code. Risk Analysis 16 81-91.
  • Evans, J. S., Hawkins, N. C. and Graham, J. D. (1988). The value of monitoring for radon in the home: a decision analysis. JAPCA 38 1380-1385. Ford, E. S., Kelly, A. E., Teutsch, S. M., Thacker, S. B.
  • and Garbe, P. L. (1998). Radon and lung cancer: a costeffectiveness analysis. American Journal of Public Health. To appear.
  • Geiger, C. and Barnes, K. B. (1994). Indoor radon hazard: a geographical assessment and case study. Applied Geology 14 350-371.
  • Gelman, A., Carlin, J. B., Stern, H. S. and Rubin, D. B. (1995). Bayesian Data Analysis. Chapman and Hall, London.
  • Henschel, D. B. (1993). Radon reduction techniques for existing detached houses. U.S. Environmental Protection Agency, EPA/625/R-93/011, October.
  • Henschel, D. B. and Scott, A. G. (1987). Testing of indoor radon reduction techniques in eastern Pennsylvania: an update. In Proceedings of the Second APCA Indoor Radon II, International Specialty Conference 146-159. Air Pollution Control Association, Pittsburgh, PA.
  • Hobbs, W. E. and Maeda, L. Y. (1996). Identification and assessment of a small, geologically localized radon hot spot. Environment International 22 S809-817.
  • Klotz, J. B., Schoenberg, J. B. and Wilcox, H. B. (1993). Relationship among shortand long-term radon measurements within dwellings: influence of radon concentrations. Health Physics 65 367-374. Lagarde, F., Pershagen, G., Akerblom, G., Axelson, O., Baverstam, U., Damber, L., Enflo, A., Svartengren, M. and
  • Swedjemark, G. A. (1997). Residential radon and lung cancer in Sweden: risk analysis accounting for random errors in the exposure assessment. Health Physics 72 269-276.
  • Lubin, J. H. and Boice, J. D. (1997). Lung cancer risk from residential radon: meta-analysis of eight epidemiologic studies. J. National Cancer Institute 89 49-57.
  • Lubin, J. H. and Steindorf, K. (1995). Cigarette use and the estimation of lung cancer attributable to radon in the United States. Radiation Research 141 79-85.
  • Marcinowski, F., Lucas, R. M. and Yeager, W. M. (1994). National and regional distributions of airborne radon concentrations in U.S. homes. Health Physics 66 699-706.
  • Miles, J. and Ball, K. (1996). Mapping radon-prone areas using house radon data and geological boundaries. Environment International 22 S779-S782.
  • Mose, D. G. and Mushrush, G. W. (1997). Variable spacial and seasonal hazards of airborne radon. Atmospheric Environment 21 3523-3530. National Research Council, Committee on the Biological
  • Effects of Ionizing Radiation (BEIR VI) (1998). Health Risks of Radon and Other Internally Deposited Alphaemitters. National Academy Press, Washington, D.C. National Research Council, Panel on Dosimetric Assumptions Affecting the Application of Radon Risk Esti
  • mates (1991). Comparative Dosimetry of Radon in Mines and Homes. National Academy Press, Washington, D.C.
  • Nazaroff, W. W. and Nero, A. V., eds. (1988). Radon and Its Decay Products in Indoor Air. Wiley, New York.
  • Nazaroff, W. W. and Teichman, K. (1990). Indoor radonexploring United States federal policy for controlling human exposures. Environmental Science and Technology 24 774- 782. Nero, A. V., Gadgil, A. J., Nazaroff, W. W. and Revzan, K.
  • L. (1990). Indoor radon and decay products: concentrations, causes, and control strategies. Technical Report LBL-27798, Lawrence Berkeley Lab. Nero, A. V., Schwehr, M. B., Nazaroff, W. W. and Revzan, K.
  • L. (1986). Distribution of airborne Radon-222 concentrations in U.S. homes. Science 234 992-997. Owens, D. K., Harris, R. A., Scott, P. M. and Nease, R. F.
  • (1995). Screening surgeons for HIV infection. Annals of Internal Medicine 122 641-652.
  • Pinel, J., Fearn, T., Darby, S. C. and Miles, J. C. H. (1995). Seasonal correction factors for indoor radon measurements in the United Kingdom. Radiation Protection Dosimetry 58 127-132.
  • Price, P. N. (1997). Predictions and maps of county mean indoor radon concentrations in the Mid-Atlantic States. Health Physics 72 893-906.
  • Price, P. N. and Nero, A. V. (1996). Joint analysis of longand short-term radon monitoring data from the Northern U.S. Environment International 22 S699-S714.
  • Price, P. N., Nero, A. V. and Gelman, A. (1996). Bayesian prediction of mean indoor radon concentrations for Minnesota counties. Health Physics 71 922-936.
  • Prill, R. J., Fisk, W. J. and Turk, B. H. (1990). Evaluation of
  • 328 LIN, GELMAN, PRICE AND KRANTZ
  • and Schumann, R. R. (1998). Bayesian analysis of the relationship between indoor radon concentrations and predictive variables in U.S. houses. Technical Report, Energy and Environment Division, Lawrence Berkeley National Lab.
  • Rhoads, S. E. (1980). Valuing Life: Public Policy Dilemmas. Westview Press, Boulder, CO.
  • Stidley, C. A. and Samet, J. M. (1993). A review of ecologic studies of lung cancer and indoor radon. Health Physics 65 234-251.
  • Turk, B. H., Harrison, J. and Sextro, R. G. (1991). Performance of radon control systems. Energy and Buildings 17 157-175.
  • Upfal, M., Divine, G. and Siemiatychi, J. (1995). Design issues in studies of radon and lung cancer: implications of the joint effect of smoking and radon. Environ. Health Perspect. 103 58-63. U.S. Congress, Office of Technology Assessment. Researching Health Risks. U.S. Government Printing Office, OTABBS-570, Washington, D.C.
  • Viscusi, W. K. (1992). Fatal Tradeoffs: Public and Private Responsibilities for Risk. Oxford Univ. Press.
  • Warner, K. E., Courant, P. N. and Mendez, D. (1995). Effects of residential mobility on individual versus population risk of radon-related lung cancer. Environmental Health Perspectives 103 1144-1149.
  • Warner, K. E., Mendez, D. and Courant, P. N. (1996). Toward a more realistic appraisal of the lung cancer risk from radon: the effects of residential mobility. American Journal of Public Health 86 1222-1227.
  • Watson, S. R. and Buede, D. M. (1987). Decision Synthesis: The Principles and Practice of Decision Analysis. Cambridge Univ. Press. White, S. B., Clayton, G. A., Alexander, B. V. and Clifford,
  • M. A. (1990). A statistical analysis: predicting annual Rn222 concentration from 2-day screening tests. In Proceedings of the 1990 Symposium on Radon and Radon Reduction Technology 1 3-117-3-118. U.S. EPA, Washington, D.C. White, S. B., Bergsten, J. W., Alexander, B. V., Rodman, N. F.
  • and Phillips, J. L. (1992). Indoor Rn222 concentrations in a probability sample of 43000 houses across 30 states. Health Physics 62 41-50.
  • ance" (Environmental Protection Agency, 1997). But to Bayesians, this approach is tantamount to "sampling from the prior"; the Monte Carlo method is being used only to simulate values from assumed
  • e.g., Rai and Krewski, 1998). Still, with a few notable exceptions (Taylor, Evans and McKone, 1993; Brand and Small, 1995; Dakins, Toll, Small and Brand, 1996), the risk assessment literature seems in need of more formal Bayesian thinking, for which the present work (and the earlier work of Wolpert, Steinberg and Reckhow, 1993) may well serve as a blueprint. One way in which the risk assessment literature is ahead of that in statistics is in its willingness to discuss the value of human life on a dollar (or some other meaningful quantitative) scale. In clinical trials, for instance, real advances in decision theoretic solutions to the interim monitoring and final analysis problems have been stymied by the unwillingness of most statisticians, epidemiologists and clinicians to even contemplate such a mapping (though a few brave first attempts have been made
  • by Berry and Ho, 1988; Stangl, 1995). An important feature of the present paper is the authors' description of how already established government guidelines for what constitutes a radon exposure level worthy of remediation implicitly determines dollar values per microlife (Section 4.2). Clearly such a linear scale is not appropriate when we move far from the origin (no reasonable person would surrender one million of his own microlives for any dollar amount), but discussions of this sort may well have beneficial impact in risk assessment strategies far beyond environmental settings, if in no other way but informing decision-makers as to what implicit values their recommendations are placing on fractions of lives. Turning then to specific comments on the authors' approach, given the power of modern MCMC techniques I was surprised that the model components considered in Section 3 were essentially confined to normal distributions. The model apparently treats the variance parameters 2 and 2 (as well as a variety of tuning parameters in Section 4) as constants, instead of more plausibly assuming distributions for them. Indeed, some of the modeling is not even being shown: (2) is written as a prior (or a "predictive" in the authors' nomenclature), but in fact it must be the result of a preliminary prior-to-posterior calculation, combining some prior on the regression parameters with some preliminary data y on typical radon concentrations in U.S. homes. What is this preliminary data and model? Do its residuals suggest any evidence of lingering spatial correlation? Also, the two-stage implementation of the preliminary (Section 3.1) and house-specific (Section 3.2) models is odd, since it forfeits the usual Bayesian advantage of a single unifying model that enables all sources of variability and uncertainty to correctly propagate throughout its levels. As the authors mention, the paper's main focus is on the decision analysis in Section 4. Here there are any number of assumptions with which one could quibble (the flat $2000 to remediate any home regardless of location, the 70-year life expectancy for every occupant, etc.); one could either place distributions on these quantities as well, or simply undertake a variety of sensitivity analyses (as the authors describe in some detail in Section 6). While I don't wish to nitpick further here, I did find the approach for "discounting the value of a life," described near the beginning of Section 4.1, to be somewhat confusing. At first blush, if Dd is the amount we are willing to pay to save one microlife now, then since lives saved 20 years in the future are worth less, it seems the revised Dd should be decreased (not increased) by a factor of 1 0520. However, recall that the paper does not really specify Dd from first principles, but rather "backs it out" by viewing the $2000 remediation cost as fixed. Thus if the value of the lives saved decreases, our cost per life saved must go up. Yet even here, it seems that the appropriate increased cost must be backed out from (5) and (6) as well, discounting each future year's risk separately in the thirty-year decision period rather than applying a single inflation factor to Dd. Of course, the actual dollar amount any given person would spend per microlife saved is probably more a function of their own financial resources and aversion to risk than any governmentally recommended remediation levels. I am personally acquainted with a suburban couple with three children who, after reading an early report on the alleged dangers of living near high-voltage electrical lines like the ones near their home, immediately sold the place and moved. Because they did this at just the time when popular concern over this potential risk was at its zenith, their total financial loss in the transaction (including moving expenses and remodeling their new home) was in the neighborhood of $80-100,000: in the light of more recent data on the subject, a colossal amount spent per expected microlife saved. In the language of (6), for this couple Dd (hence Dr) was essentially infinity for this perceived risk, and thus Raction Rremed.
  • 330 LIN, GELMAN, PRICE AND KRANTZ
  • (e.g., Stokey and Zeckhauser, 1978). But the critical importance of quantifying underlying probability distributions and sources of uncertainty-indeed, the understanding that decision-makers can even use such information-has been overlooked by other health researchers. This has created the sad situation where policy analysts are waiting for probability data for use in decision analysis, while health researchers are providing only the point estimates that they apparently assume to be preferable and useful. An issue like radon testing provides a useful way to introduce the social value of data gathering into public health without the medical arena's limits of allowable practice. Medical testing is functionally equivalent to other areas of public health, such as radon testing or finding out how safe a highway is by building it and watching what happens. But the culture of medicine and the legal climate complicate things. While allowing the gathering of patient data for the social good, the current culture makes it difficult to perform tests with expected social benefits but a net expected cost to the particular patient, or to withhold tests that have negative expected value but have any chance of improving a diagnosis. Indeed, it is sometimes difficult to even discuss making more efficient decisions in the medical arena. An environmental health issue like radon allows socially optimal recommendations because most people are amenable to persuasion about the right choice, given their underlying lack of knowledge, the relatively low individual costs and risks and the noninvasiveness of most actions. At the same time, the social costs and risks are fairly high, and it is worth the effort to try to minimize them. One major policy advantage of the situation described by Lin et al. (one which should probably be given more attention in the research literature and policy process) is that it allows individuals with different tastes for risk to take different actions. Unlike public health decisions that must be made by a central authority for everyone (cases ranging from effluent regulation to airplane safety features), the decision about radon parallels the decisions about medical care and consumption. If someone is more willing to risk disease or less willing to spend money to avoid it (or does not believe that the risk is actually real), then he has the option of making a different decision than the official recommendation. This would be particularly reasonable if, as suggested by Lin et al., EPA made recommendations that ignored household composition. Single nonsmoking assistant professors could rationally choose to ignore the radon risk in their homes. An extension of these principles in a different direction (and into more controversial areas) is to assess what population-level research would be most useful given our current data and priors (Phillips
  • and Maldonado, 1999). Epidemiologic studies, along with most quantitative health research, tend to conclude by calling for more research, but very seldom assess exactly what the further research should do. Further research can be used to eliminate some of the measurement error, simply assess the level of measurement error, eliminate confounders, measure confounders or just increase the sample size. Within all these choices, there are continuous ranges of choices along multiple dimensions. Yet the decision analysis principles are still the same as those in Lin et al. By fully assessing what we know and what more we would be likely to know following future research, we can better determine when to act, when to walk away or what more we want to know.
  • 332 LIN, GELMAN, PRICE AND KRANTZ
  • 334 LIN, GELMAN, PRICE AND KRANTZ
  • ing, rebuilt and new buildings (Cole, 1993). In the United States, EPA and state health department officials have told us that when people ask them for radon advice, they don't want to have to think about a lot of different issues; they just want to know what a "safe" radon level is. Whether or not the policymakers are right about the need for simplicity, it is clear that official radon recommendations will in fact be based on quite simple monitoring and remediation criteria. So when it came to deciding what policies to analyze, we decided to restrict ourselves to fairly simple variations on the EPA's current recommendations. We make no claim that the resulting policies are the best of all possible ones; we only claim that they would be improvements to the current recommendations.
  • 336 LIN, GELMAN, PRICE AND KRANTZ
  • treatment (Dehejia, 1998). In response to Carlin's comments about the underlying statistical model: yes, we previously fit a fully Bayesian model to a large set of shortand long-term radon measurements, along with other information on houses in the dataset and counties in the United States. We used the posterior distribution of that analysis as the prior distribution for the analysis in this paper. For each county and house type, we used the posterior simulation draws from our previous analysis to compute a prior mean and standard deviation for the mean log radon level for those counties and house types. We then assumed that the standard deviations of the measurements within that county and house type were estimated to a high precision (and could thus be summarized by posterior point estimates?), which is not too bad an approximation given the large datasets used to construct that posterior distribution. If we were less confident that the posterior distribution was close to normal, then we would have worked with the simulation draws themselves, but in this case, we wanted the convenience of the normal approximation, which allowed some of the steps of the decision analysis to be performed analytically. Another approach would be to use the normal approximation, but then check it (or correct for it) at the end of the analysis, using importance sampling. Finally, we believe it is important to link the concerns of statistical modeling to those of decision analysis. Sensitivity analysis is already recognized as a crucial step in any practical decision analysis. In addition, the iterative steps of modeling, fitting and model-checking are as relevant for decision analysis as for inference. In particular, in a decision problem, it makes sense to check that the decision recommendations for the model applied to the data are consistent with what would be expected under the model; that is, decision recommendations can be used as test variables in predictive checks
  • as in Gelman, Meng and Stern (1996). In the radon example, other natural predictive checks arise from concerns expressed with the model; for example, are there pockets of high-radon homes in otherwise lowradon counties, beyond that predicted by the model? More generally, the decision analysis should guide the model-checking as well as the inference and the modeling itself.
  • Berry, D. A. and Ho, C.-H. (1988). One-sided sequential stopping boundaries for clinical trials: a decision-theoretic approach. Biometrics 44 219-227.
  • Brand, K. P. and Small, M. J. (1995). Updating uncertainty in an integrated risk assessment: conceptual framework and methods. Risk Analysis 15 719-731.
  • Carlin, B. P., Kadane, J. B. and Gelfand, A. E. (1998). Approaches for optimal sequential decision analysis in clinical trials. Biometrics 54 964-975. Dakins, M. E., Toll, J. E., Small, M. J. and Brand, K.
  • P. (1996). Risk-based environmental remediation: Bayesian Monte Carlo analysis and the expected value of sample information. Risk Analysis 16 67-79.
  • DeGroot, M. H. (1970). Optimal Statistical Decisions. McGrawHill, New York.
  • Dehejia, R. (1998). Program evaluation as a decision problem. Technical report, Dept. Economics, Columbia Univ.
  • Environmental Protection Agency (1997). Guiding Principles for Monte Carlo Analysis. Technical Report EPA/630/R97/001, Risk Assessment Forum, U.S. Environmental Protection Agency, Washington, D.C.
  • Evdokimoff, V. and Ozonoff, D. (1992). Compliance with EPA guidelines for follow-up testing and mitigation after radon screening measurements. Health Physics 63 215-217.
  • Gelman, A., Meng X. L. and Stern, H. S. (1996). Posterior predictive assessment of model fitness via realized discrepancies (with discussion). Statist. Sinica 6 733-807.
  • Israeli, M. and Nelson, C. B. (1992). Distribution and expected time of residence for U.S. households. Risk Analysis 12 65-72.
  • Lin, C., Gelman, A., Price, P. and Krantz, D. (1999). Analysis of local decisions using hierarchical modeling, applied to home radon measurement and remediation. Statist. Sci. 14 305- 337.
  • Phillips, C. V. and Maldonado, G. (1999). Is More Research Really Needed? Unpublished manuscript, Univ. Minnesota School of Public Health, Division of Environmental and Occupational Health.
  • Rai, S. N. and Krewski, D. (1998). Uncertainty and variability analysis in multiplicative risk models. Risk Analysis 18 37-45.
  • Stangl, D. K. (1995). Prediction and decision making using Bayesian hierarchical models. Statistics in Medicine 14 2173-2190.
  • Stokey, E. and Zeckhauser, R. (1978). A Primer for Policy Analysis. Norton, New York.
  • Taylor, A. C., Evans, J. S. and McKone, T. E. (1993). The value of animal test information in environmental control decisions. Risk Analysis 13 403-412.
  • Vose, D. (1996). Quantitative Risk Analysis: A Guide to Monte Carlo Simulation Modelling. Wiley, Chichester.
  • Wolpert, R. L., Steinberg, L. J. and Reckhow, K. H. (1993). Bayesian decision support using environmental transportand-fate models (with discussion). Case Studies in Bayesian Statistics. Lecture Notes in Statist. 83 241-293. Springer, New York.