Contingent Valuation of Ill Health Caused by Pollution: Testing for Context and Ordering Effects
Richard Readya, Ståle Navrudb, Brett Dayc, Richard Dubourgd, Fernando Machadoe, Susana Mouratof, Frank Spanninksg and Maria Xosé Vázquez Rodriquezh a
Pennsylvania State University, USA Agricultural University of Norway, Norway c Centre for Social and Economics Research on the Global Environment (CSERGE), University College London and University of East Anglia, UK d Entec UK, London, UK e Catholic University of Portugal f Imperial College, London, UK g Amsterdam Free University, the Netherlands h University of Vigo, Spain b
This research was supported by the European Union’s Environment and Climate Research Programme: Theme 4 – Human Dimensions of Environmental Change (contract no. ENV4CT96-0234)
Corresponding Author: Richard Ready Department of Agricultural Economics and Rural Sociology 112-A Armsby Building Pennsylvania State University University Park, PA 16803 USA tlf: +1 814 863 5575 FAX +1 814 865 3746 e-mail: [email protected]
Contingent Valuation of Ill Health Caused by Pollution: Testing for Context and Ordering Effects
Contingent valuation is being increasingly used to value episodes of ill health caused by environmental pollution. In contrast to studies that have used contingent valuation to value other types of non-market goods, health episode valuation studies have tended to 1) value several ill health episodes or symptoms in the same survey, and 2) be vague in the survey instrument about the cause of the ill health, how it would be avoided, or how the improvement would be paid for. The resulting values are then combined with exposure-response functions to generate economic estimates of health damages from pollution. This study tests whether episode valuation responses are sensitive to two of these design features. In a five-country study using split sample treatments, neither episode ordering nor mention of the cause of the ill health influenced stated willingness to pay to avoid specific ill health episodes.
Contingent Valuation of Ill-health Caused by Pollution: Testing for Context and Ordering Effects
INTRODUCTION It has now become common to perform cost-benefit analyses on proposed actions aimed at reducing levels of air and water pollution. Often, the dominant category of benefits from a pollution-reducing action stems from the resulting reduction in adverse health impacts, both mortality and morbidity. The standard approach to estimating the benefits from these reductions is to first project the number of early deaths and ill-health episodes that will be avoided as a consequence of the pollution reduction, based on previously-estimated dose-response functions, and then multiply those numbers of cases by per-death and per-ill-health-episode values (see for example Rabl and Spadaro). The estimation of per-death mortality values is controversial, and an area of active research, but is not the focus of this study. Here, we focus on the estimation of the value of avoiding a non-lethal ill-health episode caused by exposure to environmental pollution. The costs to society of an episode of ill health include three categories of impacts. They are, 1) the productivity lost when a worker gets sick, 2) the health care costs generated by the episode, both to the patient and to the rest of the health care system, and 3) the pain, suffering, and inconvenience experienced by the person who experiences the episode of ill health. The value of the first of these impacts can be estimated from wage or productivity data. The value of the second can be estimated by directly observing health care expenditures for a variety of ill health episodes. Placing a value on the third impact is more problematic. In some cases, the value that a potential sufferer places on avoiding an episode of ill health can be estimated based on defensive expenditures - money paid for market goods that reduce the likelihood or severity of an ill health episode. An example would be purchase of an indoor air filter to improve respiratory health. However, in most cases, the link between the purchase of such a market good and the perceived
change in health is unobservable to the researcher, or is complicated by other positive or negative side effects from the use of the market good, so that the value attributable to the ill health episode can not be identified. A more common method for valuing the disamenity from ill health is the contingent valuation method (CVM). In that method, survey respondents are asked to imagine that they face a situation where it is possible to, in effect, trade money for better health. Respondents’ statements about how they would behave in that constructed situation reveal information about the value that respondents place on alternative health outcomes. Early studies that used CVM to value ill health episodes related to pollution include Dickie et al and Tolley et al. More recent studies include Alberini et al., Johnson et al, and Navrud. These studies are referred to here as episode valuation studies, to distinguish them from studies that value alternative courses of treatment for a specific disease. This latter group of studies are referred to here as health care program valuation studies. Diener reviewed studies that have used CVM to value health care programs. Many episode valuation studies (for example Dickie et al., Tolley et al., Navrud) differ from the typical contingent valuation study for an environmental good or a health care program in two important ways. First, it is common in episode valuation studies to have respondents value several different episodes of ill health in the same survey. This practice is less common when valuing other types of amenities or when valuing health care programs, out of fear that sequencing effects will “contaminate” values obtained for later-valued goods (Giraud et al). Second, the scenario within which these episodes are valued often includes no reference to the cause of the ill-health episode, how it would be avoided, or how money would be collected. The resulting values are then assumed to be applicable for policy analyses of any program that results in changes in numbers of that type of episode. In contrast, for environmental valuation, the value of an environmental amenity is usually thought to be highly context-specific. Indeed, the NOAA expert panel on contingent valuation (NOAA) concluded that respondents could not reliably
answer CVM questions about environmental goods unless the hypothetical program to provide the good is described in rich detail. Similarly, in health care program valuation, the cause of the ill health and the way that it would be treated are inseparable from the program benefit. This study examines these two unique characteristics of health episode valuation. In five different European countries, similar contingent valuation studies were conducted simultaneously. Split-sample experiments tested whether either the ordering of episodes within the surveys or mention of the cause of the episodes influenced the values generated. Tests based on parametric regressions showed little evidence that episode ordering or information about the cause influenced stated WTP values. Likewise, non-parametric tests showed no significant difference in WTP associated with either episode ordering or the level of information provided. Taken together, these tests show little evidence that WTP values were sensitive to the common practices of valuing multiple episodes in each survey, and of not mentioning the cause of the episodes.
METHODS Survey Design The study was carried out in five European countries: the United Kingdom, the Netherlands, Norway, Portugal and Spain. The goods valued in the CV surveys were avoidance of six different ill health episodes. Episode descriptions were written by a medical doctor to reflect typical episodes that would be classified as a mild symptom day, a minor restricted activity day, a work-loss day, a bed day, an emergency room visit, and a hospital admission in epidemiological studies estimating dose-response relationships between ill health and air and water pollution. Table 1 presents brief synopses of the six episode descriptions.
Table 1: Ill-Health Episode Descriptions Episode Name EYES (E)
Epidemiological End Point 1 Mild Symptom Day
One Day with mildly red, watering, itchy eyes. A Runny nose with sneezing spells. Patient is not restricted in their normal activities. One day with persistent phlegmy cough, some tightness in the COUGH 1 Minor (Co) Restricted Activity chest, and some breathing difficulties. Patient cannot engage in strenuous activity, but can work and do ordinary daily Day activities One Day of persistent nausea and headache, with occasional STOMACH 1 Respiratory vomiting. Some stomach pain and cramp. Diarrhea at least (S) Work-Loss Day twice during the day. Patient is unable to go to work or leave the home, but domestic chores are possible. Three days with flu-like symptoms including persistent BED 3 Bed Day phlegmy cough with occasional coughing fits, fever, headache (B) and tiredness. Symptoms are serious enough that patient must stay home in bed for the three days A visit to a hospital casualty department, for oxygen and CASUALTY Emergency Room medicines to assist breathing problems caused by respiratory (Ca) Visit for COPD distress. Symptoms include a persistent phlegmy cough with and Asthma occasional coughing fits, gasping breathing even when at rest, fever, headache and tiredness. Patient spends 4 hours in casualty followed by 5 days at home in bed Admission to a hospital for treatment of respiratory distress. HOSPITAL Hospital Symptoms include persistent phlegmy cough, with occasional (H) Admission for coughing fits, gasping breath, fever, headache and tiredness. COPD, Patient stays in the hospital receiving treatment for three days, pneumonia, followed by 5 days home in bed respiratory disease and asthma Note: COPD = Chronic Obstructive Pulmonary Disease
All of these symptoms, with the exception of STOMACH, could be caused by air pollution, while EYES, BED, COUGH and STOMACH could be caused by bathing in water contaminated by poorly-treated sewage. Thus, three of the episodes included in this study could be caused by two different types of pollution, allowing a test of whether willingness to pay to avoid episodes of ill health depends on their cause. The baseline survey instrument used to value these episodes contained 5 sections. In the first section, respondents were asked questions about their own health, including whether they had been diagnosed as having asthma, bronchitis or respiratory allergies, how many days during
the previous month they had experienced upper and lower respiratory symptoms, and whether they had visited an emergency room or been admitted to the hospital for respiratory problems within the past year. In the second section, the respondent ranked the ill health episodes based on how bad they would be to experience. This ranking exercise was designed to force the respondent to think about all of the ill health episodes before any were valued. In the third section, the episodes were valued sequentially. The fourth section, which included attitude and behavior questions, varied among the countries participating in the study. The fifth section collected standard socio-demographic information. The valuation scenario in the baseline survey had no detail about the cause of the ill health episode, or how it would be avoided. Respondents were told to assume that they would, with certainty, experience the episode some time within the near future, but could, with certainty, avoid it by paying some amount of money. The amount respondents would willingly pay was elicited using a modified iterative bidding format accompanied by a payment card. This protocol was first used by Dubourg et al. (1994, 1997) and is similar to the multiple bounded response potential protocol developed by Welsh and Poe. For each ill health episode, respondents were shown a payment card with 35 money amounts ranging from 0 to the local equivalent of £3500. Money amounts were chosen such that the each amount was roughly a constant multiple of the preceding amount, and such that the highest value included was larger than almost all anticipated WTP responses, based on pretest interviews. Alongside the list of money amounts were listed everyday consumer goods that cost similar amounts of money in that country, to reinforce the idea that expenditures on health leave less money for other goods1. Respondents were instructed to look at the money amount at the top of the list, and asked whether they were almost certain (95% sure) that they would pay that amount. In this case, that money amount was very small (less than £0.5). If the respondent indicated that she was 95% sure
Many respondents commented that they liked having the consumer-good references points as they answered the valuation questions.
she would pay that amount to avoid the illness episode, she was instructed to go down the money amount list and ask herself the same question for the next larger money amount. The instructions to the respondent were very clear that she should consider each amount separately. The respondent placed a tickmark next to each money amount she was 95% sure she would pay, and stopped when she reached a money amount she was either unsure she would pay, or felt she would not pay. Respondents were then asked to proceed further down the card, until they reached an amount they were 95% sure they would not pay, and place a cross next to that amount. In this paper, only results regarding the first measure, the largest amount respondents are 95% sure they would pay, are reported. This conservative level of respondent certainty was chosen to mitigate the potential for overstatement of WTP that can arise in CVM. For example, Champ et al. showed that CVM respondents tended to overstate their willingness to participate in a hypothetical program providing an environmental good, but that the proportion of respondents who were “very certain” that they would participate was similar to the proportion that actually participated in a parallel sample with real expenditures. This procedure was repeated for each of the illness episodes valued. The valuation section concluded with questions about why the respondent would or would not pay money to avoid episodes of ill health, in order to identify respondents who did not accept the valuation scenario. Any respondent who gave a maximum WTP of 0 for all episodes, and stated as a reason “I would like to avoid this illness episode, but I can not say how much it would be worth to me to do so” or “I think the whole idea of paying to avoid illness is unrealistic”, was viewed as a protest or scenario rejection response. Responses that were treated as valid zero responses included “I cannot afford to pay anything” and “I don’t think the episode is bad enough to pay to avoid it.” Second, a few respondents who stated that they were willing to pay the largest amount listed on the payment card for all of the episodes valued were excluded. A second survey version was identical to the first, except that it included an additional section after the ranking exercise and before the valuation section. In this new section,
respondents were told the cause of the ill health episode, which for Portugal was contamination by untreated sewage in bathing water, and for all other countries was air pollution. Respondents were asked their opinions about air (water) quality, and their own views about the relationship between air (water) quality and health. The ill health episodes were then valued in the same manner as in the baseline, non-context version. To test for ordering effects within the surveys, the episodes were valued in two different orders. Table 2 shows the orderings used in both the context and non-context survey versions, and the total sample sizes. Surveys were conducted in person, in home by professional survey firms. The specific cities in which surveying was conducted were Amsterdam, Oslo, Lisbon and Vigo, Spain. In the United Kingdom, a representative sample was drawn for all of England. In Portugal, contextual surveys were conducted on beaches located in or near Lisbon, in order to reach the target population of beach users. In England, contextual surveys were conducted as part of a larger study into the role of context on values (see Day). To shorten the interview and save on survey costs, not all episodes were valued in all survey versions. Table 2. Episode orders used in the study countries
Non-context Order 1 Order 2 Co,H,B,Ca,E,S B,H,Co,Ca,E,S 139 142
Context Order 1 *
Order 2 *
* results for contextual surveys in England reported in Day ** conducted on-site at beaches
E,S,Co ** 401
Statistical Analysis The experimental design shown in Table 2 resulted in 27 different country/episode combinations. Of those, 18 were valued using two different orderings, and 11 were valued using both contextual and non-contextual surveys. Thus we have 18 independent within-country tests for ordering effects, and 11 independent tests for context effects. Three different statistical approaches were employed to test whether ordering or context influenced the WTP responses. First, the responses were analyzed using parametric regression, with dummy variables for episode ordering and the presence of context. The WTP responses were treated as interval data, where the largest value the respondent is 95% sure she would pay serves as a lower bound on WTP and the smallest value the respondent is not 95% sure she would pay serves as an upper bound on WTP. In the parametric analysis, WTP for an individual with characteristics X was assumed to follow a distribution with cumulative distribution function F(P|X), so that the probability that WTP will lie in any interval [PL, PU) is given by F(PU|X) F(PL|X). The parameters of the distribution F can then be estimated using maximum likelihood estimation. This was done using the Lifereg procedure in SAS. The explanatory variables used in the parametric regressions are listed in Table 3. For each country/episode combination, several alternative distribution families were tried. Of these, the log-normal and Weibull distributions clearly provided the best fits, as measured by McFadden’s R-square. Of these two, the Weibull outperformed the lognormal in some cases, and the opposite occurring in others. We present the results for the Weibull distribution, because in those cases where there was a clear difference between the two, the Weibull outperformed the lognormal.2
Regression results using the lognormal distribution are broadly consistent with those presented here.
Table 3. Explanatory variables used in WTP regressions Variable Household Income Higher Education Female Age Number Children Asthma/Bronchitis Coughing Days Itchy Eyes Days Context Effect Order Effect
Description Household annual income, after taxes = 1 if respondent completed high school or college = 1 if respondent is female Respondent’s age in years Number of children in the household = 1 if respondent diagnosed as having asthma or bronchitis # of days in the past month respondent experienced persistent coughing # of days in the past month respondent experienced persistent itchy eyes = 1 if survey mentions the cause of the illness = 1 if episode order 2 is used
Second, likelihood ratio tests were used to test whether the structural parameters of the regressions differed across subsamples. In these tests, the log-likelihood statistics from regressions performed on each subsample were summed and compared to the log-likelihood statistic from the full data set for that country/episode combination. As these tests tend to be sensitive to the form of the distribution chosen, they were performed using both the Weibull and the log-normal distribution. Third, to avoid any biases introduced by using a possibly-inappropriate distributional form, non-parametric techniques were used to test whether mean WTP differed between subsamples. Kristrom’s method for calculating a non-parametric estimate of mean WTP was used. For our data, that method constructs a survival function for WTP, which shows the proportion of respondents willing to pay any given amount, by plotting the proportion willing to pay each of the 35 amounts listed on the payment card, and then connecting those points with straight lines. Estimated mean WTP is then given by the integral of that survival function, truncated at the highest price listed on the payment card. Because the WTP data were interval-censored, estimates of the variance of WTP among individuals and of the sampling error of the parametrically-estimated means are not available. Instead, for the parametric estimations, empirical distributions for mean WTP were constructed
using a bootstrap approach, where the original data was re-sampled with replacement. 20,000 bootstrapped data sets were constructed in this manner for each subsample, generating 20,000 estimates of mean WTP for each subsample. Equivalence of mean WTP across subsamples was then tested using Poe et al.’s method of convolutions.
RESULTS WTP was regressed on the explanatory variables listed in Table 3 for each of 27 country/episode combinations. The parametric regression results are concisely summarized in Table 4, where each column represents an independent regression. That table shows which regression coefficients were significantly different from 0 at the 10% and 5% confidence levels, as well as the signs of the estimated coefficients. Not all regressions included all variables. Health status indicators were included only where they were closely linked to the episode in question. Some country/episodes combinations were valued using only non-contextual surveys, or only one episode ordering. Gray cells in Table 4 indicate variables that were not included or tests that were not performed.
Table 4. Regression results Episode Country
Hospital Casualty Bed Eyes Cough Stomach HNPSUHNPSUHNPSUHNPSUHNPSU P U
Context Effect Order Effect
household income higher education female age number children asthma/bronchitis coughing itchy eyes
+++++++ ++++ + + ++ + + + + + + + - - - - - ++ - + + + +++++ + +++ - + + + + + + + + + -+ ++ +
+ - +++ ++ + +
+/- significant at 5% +/-
significant at 10%
Countries are: H - Netherlands (Holland) N - Norway P - Portugal S - Spain U - United Kingdom
The parametric regressions showed that WTP responses varied with the explanatory variables in systematic ways. Looking first at the characteristics of the respondents, we see that WTP increased with increasing income, as would be expected by theory. In 18 of 27 regressions, household income was significantly related to WTP, and in all of these cases the coefficient was positive, as predicted by theory. In 7 of 27 cases, number of children was significantly related to WTP, and in all of these cases, the coefficient was positive. Families with children face added difficulty and expense when a parent is sick. Higher levels of education tended to be associated with higher WTP, while females tended to be willing to pay less than males and older respondents tended to be willing to pay more than younger respondents. With regards to health status, respondents with poorer respiratory health tend to be willing to pay more to avoid an
additional episode than respondents with better health. While the regression results clearly indicate that the WTP responses vary in ways that are reasonable and consistent with economic theory, Table 4 may overstate that evidence somewhat. Because respondents provided more than one WTP value, regressions for the different episodes may not be entirely independent within each country. Turning next to the characteristics of the survey instrument, the two dummy variable intercept shifters provide simple tests for context and order effects in the WTP responses. As shown in Table 4, of 11 regressions where a test was possible, only 1 showed significant context effects at the 10% level, and none showed significant effects at the 5% level. From random sampling error we would expect 1 significant result at the 10% level. When testing ordering effects, of 18 tests, 3 showed significance at the 10% level, of which 1 showed significance at the 5% level. Here, random errors would predict 2 significant results at the 10% level and 1 significant result at the 5% level. These tests give little evidence that context or ordering effects exist. The dummy variable tests are somewhat restrictive, however, because they assume that the context and/or ordering effects affect only the intercept parameter in the WTP regressions. Likelihood ratio tests test whether all regression parameters are equivalent between subsamples. Here, it was found that more tests resulted in rejection of equivalence, but that the tests were very dependent on the choice of the distribution. Table 5 shows that using the Weibull distribution resulted in more significant ordering effects, but fewer significant context effects, than using the lognormal distribution. Further, relatively few episode/country tests were significant using both distributions; 2 of 11 context tests and 2 of 18 ordering effect tests were significant with both distributions. To conclude, the likelihood ratio tests show more evidence of context and ordering effects than the dummy shifter tests, but the positive test results are not very robust.
Table 5. Likelihood Ratio Tests (at 10% significance level) Total Tests Context Effects Ordering Effects
Significant Results Weibull
Because the parametric likelihood ratio tests are so dependent on the choice of the distribution, we tested whether the WTP values themselves differed between subsamples, using non-parametric methods to estimate and compare mean WTP. Of 11 tests for context effects, and 18 tests for ordering effects, none showed significant differences in mean WTP at the 10% level. In over half of the cases, mean WTP for the two subsamples were within 20% of each other. In only one of 29 tests was one subsample’s value more than double the other’s (the ordering effect test for England for STOMACH). Thus, even if there are significant ordering or context effects (and the evidence from the likelihood ratio tests that there are is very weak), they do not generate significant differences in mean WTP, and in most cases do not even generate meaningful differences.
CONCLUSIONS To conclude, this study found little evidence that the mention of the cause influenced respondents’ stated WTP to avoid an illness episode. Further, WTP values were not influenced by the order in which episodes were valued in the survey.3 Our results are reassuring for applied policy analysis. They support the cost-saving practices of considering the cause of the illness separately from its value and of estimating the values of several different ill health episodes in one interview.
This result is consistent with that of Kartman et al., who found no question order effects in a health care program valuation study.
We note three caveats, however. First, our “contextual” survey mentioned the cause of the illness, but did not mention how the episode would be avoided. Thus, even our contextual survey lacks some potentially-relevant context. Halvorsen (1996) found that in a study where specific clean-up programs are specified, their values are less robust to survey design issues such as episode ordering. There, however, the issue of ordering was confounded by potential partwhole sequencing effects. Rozan and Willinger found that the value of a reduction in respiratory ill health depended on whether that reduction came from a program aimed at reducing pollutant levels or reducing levels of naturally-occurring dust and pollen. In both the Halvorsen and the Rozan and Willinger studies, the values elicited included other benefits from the cleanup program, such as reduced soiling and improved visibility, which will tend to be very program and site specific. Second, these results do not technically prove that, for example, the value of a symptom day with itchy eyes caused by poor water quality is the same as the value of that same day caused by poor air. Instead, the results show that the value was the same whether the cause was mentioned or not. If respondents in Lisbon tended to believe that itchy eyes are most often caused by poor water, while respondents in the other countries believed that they are most often caused by poor air, then the information provided may not have been new to the respondents. However, epidemiological studies would suggest that, in all study sites where context effects were tested, more cases of itchy-eyes days are due to poor air than to poor water. Third, our study did not attempt to value avoidance by one individual of more than one ill health episode. Previous studies (Navrud, Dickie et al.) have found that WTP per-episodeavoided is lower when a respondent values avoidance of several episodes together, instead of just one, suggesting that the marginal value of avoiding an ill health episode decreases as the number of episodes experienced decreases toward zero. That result is consistent with the finding in this study that marginal value to avoid an episode is higher for respondents who currently have poorer health. These results suggest that, for individuals who will avoid more than one episode of ill
health, it is inappropriate to apply a single per-episode value to all of those avoided episodes. Unfortunately, available exposure-response relationships do not characterize the distribution of health impacts among individuals in the affected populations, that is, they do not tell us whether the predicted health impact will come as five episodes experienced by one individual or one episode experienced by each of five individuals.
Alberini, A, M Cropper, T-T Fu, A Krupnick, J-T Liu, D Shaw and W Harrington (1997) “Valuing health effects of air pollution in developing countries: The case of Taiwan” Journal of Environmental Economics and Management 34(2):107-126.
Champ, P, R Bishop, T Brown, and D McCollum (1997) “Using donation mechanisms to value nonuse benefits from public goods” Journal of Environmental Economics and Management 33(2):151-162.
Dickie, M, S Gerking, D Brookshire, D Coursey, W Schulze, A Coulson, and D Tashkin (1987) “Reconciling averting behavior and contingent valuation benefit estimates of reducing symptoms of ozone exposure” Washington DC: Environmental Protection Agency.
Day, B (1999) “Do scenario context and question order influence WTP?” in Benefits Transfer and the Economic Valuation of Environmental Damage in the European Union – Report to DG XII.
Diener, A., B. Obrien and A. Gafni (1998) “Health care contingent valuation studies: A review and classification of the literature.” Health Economics 7(4):313-326.
Dubourg, W.R., M. W. Jones-Lee, and G. Loomes, (1994) “Imprecise preferences and the WTPWTA disparity.” Journal of Risk and Uncertainty 9:115-33.
Dubourg, W.R., M. W. Jones-Lee, and G. Loomes, (1997) “Imprecise preferences and survey design in contingent valuation.” Economica 64:681-702.
Halvorsen, B (1996) “Ordering effects in contingent valuation surveys: Willingness to pay for reduced health damage from air pollution.” Environmental and Resource Economics 8(4):485-499.
Giraud, K.L., J.B. Loomis and R.L. Johnson (1999) “Two valuation questions in one survey: is it a recipe for sequencing and instrument context effects?” Applied Economics 31(8):957-964.
Johnson, FR, M Ruby, W Desvousges, A Gable, P De Cevita, and D Stieb (Forthcoming) “Valuing reduced air pollution-related morbidity using stated preferences and health-state classification.” Environmental Research .
Kartman, B, NO Stålmammar, and M Johannesson (1996) “Valuation of health Changes with the Contingent Valuation Method: A Test of Scope and Question Order Effects.” Health Economics 5:531-541-
Kristrom, B (1990) “A nonparametric approach to the estimation of welfare measures in discrete response valuation studies.” Land Economics 66(2):135-139.
Navrud, S. (1998) “Valuing health impacts from air pollution in Europe: New empirical evidence on morbidity.” Discussion Paper 98-04, Department of Economics and Social Sciences, Agricultural University of Norway.
NOAA (1993) “Report of the NOAA Panel on Contingent Valuation” Federal Register 58(10):4601-4614.
Rozan, A. and M. Willinger (1999) “Does the knowledge of the origin of the health damage matter for WTP estimates?” Working paper, Department of Economics, B.E.T.A., Université Louis Pasteur, Strasbourg, France.
Poe, G L, E K Severance-Lossin, and M P Welsh (1994) “Measuring the difference (x-y) of simulated distributions - convolutions approach” American Journal of Agricultural Economics 76(4):904-915.
Rabl, A, and JV Spadaro (2000) “Public health impact of air pollution and implications for the energy system.” Annual Review of Energy and Environment 25:601-627
Tolley, G, D Kenkel, and R Fabian (eds) (1994) Valuing health for policy: An economic approach Chicago: The University of Chicago Press.
Welsh, M, and G Poe (1998) “Elicitation effects in contingent valuation: Comparisons to a multiple bounded discrete choice approach.” Journal of Environmental Economics and Management 36, 170-185.