Document 7zerb5L76xoK8Vjxwjda5rra

*76. R. N. Sawyer, Environ. Res. 13, 146 (1977). > 77. TLV's. Threshold Limit Values for Chemical * Substances In Workroom Air Adopted by ,\ACGIHfor I960 (American Conference of Gov* !' ervmental Industrial Hygienists, Cincinnati, r^Ohio. 1980). 71. Asbestos--Genera! Information (Asbestos In formation Association of North America, Wash* ' ' `ington. D.C., 1973). 79. M. Villeco. Archil. Forum 133, 30 (1970). 80.' A. N.' Rohl and R. N. Sawyer, paper presented at the International Symposium on Indoor Air Pollution. Health, and Energy Conservation, Amherst. Mass., 13 to 16 October 1981. 81. R. L. Murphy, B. W. Levine. F. J, Al Baiiai, J. J. Lynch, w. A. Burgess, Am. Rev. Respir. Dis. 104,376(1971). 82. Occupational Safety and Health Administration, "Standard for exposure to asbestos dust," Fed. Regiit. 31, 11318 41972). , 83. "Revised recommended asbestos standard," ' U.S. Dept. Health Educ. Welfare Fubi. (NIOSH) 77-169 (t977). 84. R. N. Sawyer and C. M. Spooner. U.S. Environ. Frol. Agency Publ. EPA-l50i2-78-0H (1978). 83. U.S- Environmental Protection Agency, "Amendments to asbestos standard." Fed. Re- ist. 43. 26372 (1978). Workshop on Indoor Air Quality Research Needs, Interagency Research Group on Indoor Air Quality, Leesburg, Va.. 3 to 5 December 1980. 87. "Position statement on indoor quality," J. Air Pollut. Control Assoc. 32. 348 (1982). 88. The American Society of Heating, Refrigera tion. and Air Conditioning Engineers, Inc., "Po sition statement on indoor air quality," p- proved by the ASHRAE board of directors on I July 1981. 89. J- L. Repace, Environ. Int. I. 2) (1982). 90. K. Sexton and R. Rcpctto, ibid., p. 5. 91. This work was supported in pan by N1EHS grant ES-OI108. EPRI grant RP-1001, and EPA grants 68-02-3201 and 2974. We thank the mem bers of the staff and faculty at Harvard Universi ty who commented on earlier versions of this manuscript: M- J. Ellenbecker. J. S- Evans, M. W. First. R. Leu. D. W. Moeller, P. B. Ryan, S. R. Thomas, R. D. Treitman, and W. A. Turner. We are also grateful to A. Maskell who typed the article and provided editorial comments. Cost-Effective Priorities for Cancer Prevention Milton C. Weinstein Environmental factors are responsible pie, but current understanding leaves us for 80 to 90 percent of cancer deaths in., far short of being able to prevent most the United States (/, 2). This conclusion, ^cancer in fact. The challenge of the comwhich once aroused considerable contro- ing decades will be to identify the specifversy, is now generally accepted, pro ic agents that cause or prevent cancer vided that the "environment** is broadly and, after identifying them, to develop y_ Summary. Faced with limited resources, the United States must set priorities for research to identify preventable causes of cancer. A quantitative approach to priority setting, based on principles of decision analysis and cost-effectiveness analysis, can offer guidance In this process. An illustrative application of such a model suggests that the National Institutes of Health-supported dbiical trial of dietary $-carotene offers a greater expected reduction in cancer mortality per research dollar than carcinogen bioassays of high-volume industrial chemicals such as p-dichlorobenzene. National research priorities should reflect the relative cost-effectiveness of such investments. defined to include not only industrial chemicals and pollution, but also diet, reproductive behavior, and other ele ments of life-style and culture, as well as such natural phenomena as infectious agents and nonionizing radiation. Doll and Peto have placed the contribution to U.S. cancer mortality of occupational and environmental exposures to industri al chemicals at less than 5 percent, in cluding 2 percent due to asbestos (2). Growing hope during the 1970's that cancer could be controlled in large part by detecting and eliminating carcinogens has been tempered during the 1980's by the sober realization that preventing can cer will not be simple. Epidemiologic data firmly support the proposition that most cancers are preventable in princi- 1 JULY 1983 - i and implement interventions to alter hu man exposure to them. The problem of identifying carcino gens in the environment seems formida ble enough when attention is focused on the 70,000 or so industrial chemicals in production. The cost of testing this in ventory of chemicals, let alone the thou sands of new chemicals entering produc tion each year, would be huge. Even if financial cost were not a constraint, the limited supply of toxicologists and laboratorics would constrain the volume of long-term bioassays. . Epidemiologic insights should, howev er, lead us to examine the priority-setting problem in a broader framework. If in dustrial chemicals other than asbestos account for 3 percent of cancer deaths. the prospect of saving 12,000 lives each year (out of 400,000 cancer deaths) should encourage us to discover the key industrial agents and control exposure to them. But if 35 percent of cancer deaths are related to diet (2), efforts to discover dietary factors in cancer might deserve an even greater claim on resources. Toxicologic studies ofindustrial chem icals and epidemiologic studies of dietary agents are, in general, funded from dif ferent budgets, and might seem not to be in competition for the same limited re sources. For the society as a whole, however, it is imperative to ask how best to spend resources in the general domain of cancer prevention. Priorities need to be set among alternative research strate gies for detecting carcinogenic and anticarcinogenic agents, and such priority setting should encompass the full range of environmental factors (broadly de fined) in cancer prevention. This article illustrates a quantitative approach to priority setting, based on principles of cost-effectiveness and deci sion analysis. It also shows how the approach may be used to compare the cost-effectiveness of toxicologic studies of industrial chemicals and prospective trials of dietary constituents. The indus trial chemical examined is p-dichlorobenzene, the active ingredient in moth balls. The cost-effectiveness of a ran domized prospective trial of dietary 13carotene, a close relative of vitamin A, is also assessed. This comparison and oth er considerations lead to policy implica tions regarding the optimal use of re sources in investigating the cancer-relat ed effects of environmental agents. Uncertainty is inherent in this kind of prospective analysis, and the attempt to quantitate this uncertainty may make some readers Uncomfortable. However-, policy decisions must and will be made in the face of uncertainty, and analysis The author is professor of Policy and Decision Sciences. Department of Bkwtausucs, Harvard School of Public Health. Boston. Massachusetts 02113. . 17 TT" :& URL 05407  s can help to organize information, expose the sources and magnitude of uncertain ty, and facilitate rational debate about policy options. Further, if mqjor changes in estimates do not change the main conclusions, confidence in those conclu- - sions is increased despite the uncertain ly. An Analytic Framework for Priority Setting 4 Health effectiveness. Let us assume that the most important objective of lab oratory or clinical studies of the carcino genic effects of specific agents is to pre vent cancer mortality, and.that the ulti mate value of such studies can be mea sured by the number of cancer deaths prevented or the number of years of life saved (J). Overall, the number of years of life saved may be approximated by multiplying the number of cancer deaths averted by 12 years, the mean loss in life expectancy per cancer death in the Unit ed States (4). An advantage of using life years saved as the measure of effec tiveness is that it is general enough to permit comparisons with uses of re sources in widely different health pro grams. ; At least three steps are needed to translate a carcinogen bioassay or epide miologic study into reduced cancer mor tality. The tes(^system must detect an effect, the study result must lead to inter vention to alter exposure to the sub4, stance in question, and the intervention It/ nftist lead to reduced cancer mortality. The last step implies that the observed effect must be real and not an artifact of the experimental or statistical methods used. These events are not guaranteed to occur, they are probabilistic in that we do not know in advance whether each of them will occur (for example, whether the test will uncover an effect). Thus, probabilistic reasoning is needed to pre dict the benefits of studies; a priori, we must settle for a statistical measure of the expected, or average, number of cancer deaths averted or life years gained. Both objective and subjective probabilities are involved. Decision analysis provides a frame work for calculating the expected values of alternative testing strategies (5) from estimates of the quantities listed in Table t.' The prior probability of effect is the ' best estimate, prior to the study, of how likely it is that the agent in question affects human cancer mortality. For ex ample, one might assess that there is a 10 Table 1. Some factors influencing the expect ed value, or effectiveness, of a cancer study. 1. Prior probability of effect oo cancer mor tality and its magnitude Previous tests (in vitro, in vivo) Epidemio}ogic evidence Biologic understanding or theory 2. Sensitivity of the test system Sample size Technical design features Relevance to human exposures Tolerance for false positives (positivity criterion) 3. Effect of research findings on behavior and exposure Regulation Self-regulation by industry Personal behavior 4. Cancers prevented, over time, given effect and exposure change Potency or relative risk Exposure change Latency .... Lag to implementation percent chance that chemical X is a human carcinogen on the basis of previ ous tests (in vitro and in vivo), epidemio logic evidence (including negative evi dence), and structure-activity consider ations. XThe sensitivity of the test system is the probability that the test will show a posit tive result if an effect is truly present. It depends on design elements such as sam ple size and dose of agent, as well as on such technical design features as alloca tion of treatments, quality control, and observer reliability. If one is using a nonhuman test system to detect effects in humans, another sourcc of insensitiv ity (false negatives) arises from interspe cies differences in response. Finally, the sensitivity depends on the methods of statistical analysis and on the threshold chosen for calling a result positive. The effect of a study result on behav-. ior and exposure may depend on the actions of many parties. Government may regulate; industry may self-regulate; individuals may alter personal behav iors. It is difficult to predict what actions would follow from specific test out comes, or even to assign probabilities to these actions. But to fail to assess jhe prospects - for effective intervention would implicitly assign equal value to study results that would probably lead to intervention and results that would only raise anxiety levels or elicit reactions of denial, 'u * *=. The number and timing of cancer deaths prevented may be calculated from estimates of five items: an exposureresponse relation, summarized perhaps by a measure such as carcinogenic po tency or relative risk; magnitude of ex posure change; latency of biological ef fect (w); lag between the decision to undertake the study and its findings (L[); and lag between study findings and al tered exposure (Lj)- The result will typi cally be a stream of benefits over time, beginning at the end of the lag-pluslatency periods and extending into the indefinite future (Fig. 1). A linear expo sure-response relation is assumed over the range of actual exposures, with po tency k defined as the number of cancer deaths prevented per year per unit of exposure change. The model can easily be generalized to allow for nonlinear exposure-response relations. The expected, or average, value of a study may be expressed mathematically as follows. Let us assume a discrete formulation with / possible values of potency, k, (0. k2, . . . , k, _ t), each assigned a prior probability, pt * proba bility [,-]. Thus po may refer to the prior probability pf no effect; p\ to the prior probability that the potency is cancers prevented per year per unit of exposure reduction, and so forth. Next, let rj rep resent each of J possible study results (/ 0, . . . , J - 1); we assess the prob ability of result rj given true effect k, as qy probability [rj I k(]. Then, let em represent each of M possible changes In exposure (0, elf . .. , em _ |), resulting from regulatory control or behavior change, and we assess the probability of exposure change em given result rj (and all prior information) as probabili ty [*m oJ. Then the annual expected benefit (or effectiveness, ) of the study is given by , I -1 j - i 1 " Z Z Z (,z^q\jPi)ktem 1-0 (l) ' Cost. Testing costs include costs of protocol design, implementation, data collection, and analysis. Strictly speak ing, however, it would be wrong to bal ance only these costs against any expect ed benefits. Interventions to alter expo sure are also costly, at least as perceived ex ante; otherwise, why not reduce all exposures to suspected carcinogens to zero? The main analysis excludes the cost of intervention from consideration. Because this would be inappropriate if resources for lifesaving were viewed as constrained in the domain of public health interventions (6 7), costs of inter vention are introduced later in the dis cussion of the examples used to illustrate the model. `The cost-effectiveness ratio. Given es timates of the cost of testing (C) and of the expected number of cancer deaths prevented () for a range of studies, the 18 SCJENCE. VOL. 221 URL 05408 \* following question arises: Given that re sources do not permit undertaking all studies, y^ith positive expected health benefits ( > 0), how do we set prior ities? The answer, given the objective of maximizing the expected reduction in cancer mortality, centers on the cost- effectiveness ratio, that is, the cost per cancer death averted, for each study being contemplated. If studies are ranked in increasing order of this ratio, and undertaken in the implied order of priority, the total health-effectiveness of the overall testing program will be maxi mized (5). . . In forming the cost-effectiveness ratio, one must express costs and benefits in temporally comparable units. Recall that we have an expenditure at time t0, fol lowed by a delay bf 1| 11. + w years before healthy benefits accrue (Fig. i). Suppose that the number of cancer deaths averted beginning at time // /o + L\ + Li + *v is per year. In order to render costs and benefits temporally comparable, we will convert the cost, C, into an equivalent constant annual stream, C*, commencing at fu ture time where s /'C* - K7(l + rY'~ - "s`(2) and r is the real (that is, inflation-correct ed) long-term discount rate, taken to be 0.05 in the following analyses. C* is analogous to the annual payment on a long-term mortgage at rate r whose prin cipal value is C, but whose first payment is deferred for f/ - f0 years (9). The ratio C*/* is then the time-corrected costeffectiveness ratio (10). 0 4 Cost-Effectiveness of the Bioassay of p-Dlchlorobenzene ; The carcinogen bioassay in small ro dents (CBSR) is the mainstay of carcino genesis testing in the United States. The National Toxicology Program (NTP) of the federal government has published over 200 reports on chemicals tested under the CBSR, including 27 completed in 1982. Hundreds more chemicals have been tested by private groups, including industry. Separate from the government's own testing program, the Environmental Pro tection Agency (EPA), under authority of Section 4 of the Toxic Substances Control Act (TSCA), may require pri vate industry to test chemicals for carci nogenicity. Substantial effort has gone into setting priorities for testing (//). This article examines the cost-effective ness of testing a chemical that emerged at the top of the priority-setting process to to + L i . to + l + Lj to + Li+Lj + w ----- | ------------- 1-------------------- 1 ---------------1------------------------------------ Tima Study Study Action I Initiated completed Initia,te.d Cancera prevented Fig. I. Timeline for events following initiation of a study. The total delay between initiation of the study and the beginning of the benefit stream is Z., + L2 + w (where L, is the lag to end of study, Lj is the lag to initiation of intervention, and it' is the cancer latency period). that ied thc*EPA to identify p-dichloro- benzene as one of the first three chemi cals to be tested under TSCA. Approximately 33 million kilograms of p-dichiorobenzene were produced in the United States in 1978, of which an esti mated 24 million kilograms were re leased into the air. Of the industrial output, 55 percent is used in space de odorants, 35 percent in mothballs, and 10 percent in a variety of products including pesticides, dyes, door waxes and finish es. abrasives, and agricultural chemicals (12). Despite Us widespread use, this compound has not been adequately test ed for carcinogenicity. Prior probabilities and potency. No fully satisfactory basis exists for estimat ing the probability that a particular chemical is a human carcinogen. Histori cal experience with the CBSR is a misleading guide because selection of chemicals by the NTP often is based on scientific reasons related to molecular structure rather than public health con siderations such as extent of exposure and environmental persistence 03). In one random survey of compounds tested prior to 1974 it was estimated that 5 percent ofcompounds were carcinogenic 04). Information from short-term tests would raise or lower the probability for any particular chemical, depending upon the results (7). Apparently, p-dichlorobenzene is not mutagenic in the Ames test (15). All things considered, let us take 10 percent as the prior probability that p-dichlorobenzene is a carcinogen. This still leaves open the question of bow potent it might be. For simplicity, let us apply a point estimate of 0.005 lifetime cancers per milligram per kilo gram of body weight per day. This is based on the data compiled by Crouch and Wilson on carcinogenic potency of chemicals tested in animals and man (16). This potency is treated as if it applied to fatal cancera only. Sensitivity oftest system. Of 26 known human carcinogens, 18 have been sub jected to an adequate CBSR; of these, only two (arsenic and benzene) are not rodent carcinogens (77). Thus, one has a rough estimate that 16/18 (89 percent) of human carcinogens are rodent carcino gens. Assuming a statistical power of 0.9, the overall test sensitivity is (0.89)(0.9) 0.80. Exposure reduction. Occupational ex posures are estimated from EPA data to average 28 mg/kg-day for 5000 workers (IS), and environmental exposures from the air are estimated to average 0.7 jig/ kg-day for each of 230 million Americans (19). It is impossible to predict what the response of government, industry, or consumers would be if mothballs were found to cause cancer. Let us assume an expected 50 percent reduction in expo sure levels given such a finding (20). Timing of cancers prevented. The CBSR typically requires at least 6 years (Li) including planning, analysis, and reporting. Allowing at least another 4 years (l2) for public response, and a carcinogenic latency period of 20 years (h>), the total delay would be 30 years before cancer deaths were actually pre vented. Cost of the bioassay. Xu fiscal year-s 1981, the NTP let 16 private contracts for' CBSR's on 49 chemicals. The mean con tract budget per chemical tested by feed ing was $467,000. This calculation as sumes a cost of $500,000. Cost-effectiveness calculation. Total occupational and environmental expo sure to this compound (i| in Eq. 1) is calculated from the above data to be 3 x. 10* person-mg/kg-day. Multiplying by the assumed potency, k\ (0.005 life time cancer deaths per person-mg/kgday divided by 70 years per lifetime), yields 21 cancer deaths per year poten tially averted. The expected effectiveness of testing equals this potential benefit, times the prior probability of carcinogenicity (pi0.l), limes the test sensitivity (qti 0.8), times the expected exposure reduction (zM * 0.5), so that E* * 0.85 cancer death averted per year. From Eq. 2, we calculate the annua lized cost of the bioassay, at the point of benefit, as (0.05) ($500,000) (1.05)" $110,000. The cost-effectiveness ratio, then, is (C*/*) ($110,000 per year)/ (0.85 cancer per year) * SI30.000 per cancer prevented, or about SI),000 per year of life saved. This excludes the social and economic cost of actually re ducing exposure to p-dichlorobenzene. URL 05409 I JULY 1983 19 i-. / * . / Cost-Effectiveness of the Prospective Trial of 3-Carotene Let us now examine the cost-effectiveness of prospective studies of dietary h'Jb factors in human cancer. Specifically, let us consider the ongoing prospective trial of dietary p-carotene. Epidemiologic evidence suggests that retinoids and carotenoids (that is, vita- min A) may contribute to cancer preven tion. Petoe/ al. cite 20 dietary studies, of which ten found relative risks of 1.5 to 3.0 for lowversus high vitamin A diets, seven found relative risks of 1.3 to 1.5, and only three found no statistically sig nificant effect (21). A more recent study of the correlation between cancer inci dence and 3-carotene consumption (rather than total vitamin A or consump tion of vegetables) found a relative risk for lung cancer of 7.0 for all subjects and 8.1 for smokers only (22). In addition, several studies in mice have shown that 3-carotene can reduce or delay tumor incidence (2/). The epidemiologic results are not con clusive for several reasons. One reason is that recall of dietary data is imperfect; another is that the observed negative' association between p-carotene con sumption and cancer might be an artifact stemming from the carcinogenic effect of animal fat and an inverse association between 3-carotene and animal fat levels in the diet. Ityat least three epidemiolog ic studies (23, 24), however, no associa tion was found between fat or fiber con>4 sumption and the cancer under study V . `(larynx, lung, or breast), while a strong protective association with vitamin A consumption was found. Only a prospective study, in which pcarotene is administered independently of specific foods, can resolve this ques tion. A double-blind controlled trial of pcarotene (30 milligrams every 2 days) in American physicians was funded by the National Institutes of Health, after some deliberation, as an add-on to a trial ex amining the relation between aspirin and myocardial infarction, using a 2 by 2 factorial design. The protocol calls for a 5-year intervention and follow-up <2S). The fo{lowing data and assumptions were used in the cost-effectiveness anal ysis of the p-carotene trial. Prior probabilities and potency. Esti mates of association between P-carotene and cancers of the lung, bladder, larynx, esophagus, and breast were derived from epidemiologic studies. Two studies pro vided estimates of relative risk for lung cancer. Metttin et al. (24) found a 70 percent increased risk of lung cancer in that half of a population with a lower p- carotene consumption compared to the upper quartile of the population, in a similar study of p-carotene consump tion, Shekeile et al. (22) reported a sev enfold risk for males in the lowest quar tile compared to the highest quartile. in this.artide it assumed that an increase in intake to 15 mg/day would correspond to moving to the low-risk quartile, and the lower of the risk' estimates from the two studies is used. A total of 30,600 lung cancer deaths would therefore be avert ed per year, or 32 percent of all such deaths in the United States. Estimates for cancers of the bladder, larynx, esophagus, and breast were cal culated analogously from epidemiologic data (26) as 3300. 1500, 2400, and 6900 deaths averted per year, respectively, or 35, 50, 30, and 20 percent of cancer deaths at these sites. With a S percent reduction in cancer mortality being as sumed at all other sites combined, an additional 13,000 deaths would be avert ed, bringing the total to 58,000 cancer deaths. Therefore, the potential reduction in cancer mortality if the p-carotene hy pothesis were true and if the population altered its dietary habits, is estimated to Ws60,000 deaths per year, or a 15 per cent reduction. ` Finally, we need a subjective estimate of the probability that the hypothesis is, in fact, correct. Let us use a subjective probability estimate of 10 percent. Test system sensitivity. The statistical power of the study depends on at least four factors: the magnitude of any true effect, the sample size, the duration of the study, and the latency period prior to manifestation of the effect. Approxi mately 15 percent of all male U.S. physi cians aged 50 to 75 years have been enrolled in the study. The approximately 20,000 subjects are divided randomly be tween treated persons and placebo con trols, and follow-up will be for S years. Assuming a 2-year latency period, and a IS percent reduction in male cancer incidence during years 3 through 5, we would expect 3.17 percent cancer inci dence in the controls and 2.69 percent incidence in the treated group. With a significance level of .05, under these assumptions, the probability of the study detecting the effect, if present, would be 64 percent (27). Since there is no issue of interspecies correlation, 0.64 is used as the estimate of test system sensitivity. Exposure change. The public health impact of a positive finding would de pend on the responses of both public health officials and private individuals. Public health officials could declare such a finding cause for a mqjor public health campaign analogous to fiuoridation of public water supplies. Already, dairy products are fortified with vitamins; they could be fortified with higher doses of 3carotene, provided the public would ac cept foods with a slight orange tint. Fail ing such a mass intervention, individuals could be urged to increase their 3-caro tene intake by promotional campaigns; or subsidies for high 3-carotene foods could be increased. Finally, individuals may elect to take 3-carotene as a drug; several commercial preparations are available. What proportion of the population would modify their diet in response to a positive finding? What is thelflcelihood of a public health initiative to fortify foods? These are difficult questions, but it might be easier for parents to inculcate a tolerance for carrots than an abhor rence of smoking. Let us suppose, sub jectively and perhaps conservatively, that 10 percent of the potential benefit would be realized by some combination of public and private initiatives. Timing of cancers prevented. Let us assume a lag equal to the study duration (5 years), plus an additional 10 years for dissemination and latency. Cost of the study. The budget for the combined study of 3-carotene and can cer and of aspirin and myocardial infarc tion is $4,000,000. Let us attribute the lull cost to the 3*c&rotene study, recog nizing that this tends to overestimate its true incremental cost. Cost-effectiveness calculation. Under our central assumptions (64 percent study power, 10 percent compliance, 15 percent cancer mortality reduction), the estimated reduction in annual cancer mortality if an effect is present would be 3840. Multiplying by the prior probabili ty of 0.1, the expected annual benefit from the study is 384 cancer deaths averted. From Eq. 2, with r .05 and tf - 15 years, we calculate the an nualized cost of the study to be $420,000 per year. The cost-effectiveness ratio is C*/* - ($420,000 per year)/(384 cancer deaths per year) - $1100 per cancer death prevented, or about $91 per year of life saved. This is approximately l per cent of the corresponding estimate for the bioassay of p-dichlorobcnzene. Comparative Cost-Effectiveness and Sensitivity Analysis* To summarize, the expected cost per year of life saved is expected to be SI 1.000 for a rodent bioassay ofp-dichiorobenzene and $91 for a prospective trial of 3-carotene. Both figures are for re search studies to establish harmful or UHL 05410 4 4` 1 beneficial effects, and exclude the costs from $11,000 per year of life saved to considerably at a minimal cost. In any of intervention. approximately $1,000. case some economic dislocations would The data and assumptions underlying On the other side, it is unlikely that%ve surely be felt. these calculations are soft. But is the have been overly optimistic in our esti The annual retail cost of taking 15 mg 100-fold difference in cost-effectiveness mates about p-carotene. If we assumed a of O-carotene daily is $36.50 (.10). The large enough to withstand even rather 10 percent (rather than 15 percent) re long-run cost, with genetically available large errors in the estimates? Let us duction in cancer mortality, with prior 3-carotene, might be in the $20 to $30 approach this question by means of sen probability 10 percent, the cost-effec range. Less costly would be fortification sitivity analysis. tiveness ratio would change from $91 to of foods such as milk or butter, or indi In the calculation of cost-effectiveness $240 per year of life saved, still extraor vidual diet modification; two 100-g serv for the bioassay of p-dichlorobenzene, dinary by most standards. Thus, the ings of carrots contain 13.4 mg of p- the six critical parameters were as fol highest plausible figure for the 3-caro carotene. The side effects of p-carotene lows: the exposure estimates for workers tene study is still well below the lowest consumption are benign (}/). Most and for the general public, the assumed plausible figure for a CBSR assay of p- prominent is a coloring of the skin which carcinogenic potency, the prior probabil dichlorobenzene. Plausible assumptions some people find appealing. ity of carcinogenicity, the sensitivity of in the opposite directions would have Suppose the cost of arPannual regimen the test system, the proportion reduction increased the divergence from two or of p-carotene were $30. If 200,000,000 in exposure, and the lag and latency ders of magnitude to three or even four Americans paid this price (excluding periods. orders of magnitude. small children), the annual bill would be The 5000 workers assumed to be ex One caveat in advocating studies such $6 billion. If this could prevent 60,000 posed to two-thirds of the maximum as the 3-carotene trial is that the possibil cancers a year, and if the ratio of (he allowable time-weighted average con ity of a false-negative finding may be true-positive to false-positive study re centration of the potential carcinogen seen as unacceptably high. A negative sults is as assumed previously, (he include those directly involved in the result may engender future public mis steady-state cost per cancer death avert manufacturing processes. We might trust of public health information and ed would be $170,000, or about $14,000 have assumed another 50,000 to be ex may make further research on diet and per year of life saved. The cost-effective posed to 2 mg/m3, the equivalent of a cancer more difficult to justify. It will be ness ratio would be tower if the program mothball-filled closet, but this would add difficult to explain that, even if there were targeted at older age groups. Even only 100,000 person-mg/m* to the origi were a 15 percent reduction in cancer, as an upper bound, this cost per year of nal occupational estimate of 1,500,000 the study had a 36 percent chance of life saved compares favorably to many person-mg/m3. Even if all 500,000 per missing it. One remedy would be to preventive medical interventions in com sons employed in the industry were ex increase the duration of the study. Under mon use, such as treatment of high blood posed to 2 mg/m3, this would add only 60 our previous assumptions, a 10-year pressure (32) and cancer screening (33). percent to the occupational exposure es study would increase the statistical pow It is one or two orders .of magnitude timate, or about 30 percent to the overall er from 0.64 to approximately 0.95, while lower than ratios estimated for occupa- 7 exposure estimate. perhaps doubling the cost. Thus, the tional and environmental health mea The assumed carcinogenic potency is cost-effectiveness ratio would increase sures aimed at cancer prevention (34). It already five times the human potency of somewhat (though still less than $200 per is also likely to be substantially lower benzene, and more than double the year of life saved), but the chances of a than the corresponding estimate for a mouse potency of ethylene dichloride falsely negative result would be reduced ban on p-dichlorobenzene, which rein (2$). Moreover, we are assuming all of considerably. forces the conclusion from the main the inhaled chemical to be absorbed and analysis in which the costs of interven are using a conservative linear dose- tion were excluded. response model. Although the estimate Considering the Cost of Intervention of potency may be conservatively high already, we double it to 0.01 cancer Let us now consider the cost of inter Policy Implications deaths per person-mg/kg-day in the sen vention to alter exposures to these sub sitivity analysis. The prior probability of stances. The ^-carotene trial appears to be an 10 percent is already as high as is con The social and economic cost of ban excellent investment in health resources. sistent with the negative evidence from ning major uses of p-dichlorobenzene The carcinogen bioassay of p-dichloro short-term tests. would be great. The gross annual pri benzene may also be a reasonable use of The sensitivity of the test system (as mary sales of this compound are approx resources, although not nearly as high a sumed to be 80 percent) exerts little imately $30 million. Using this figure as priority as the p-carotene trial. leverage mi the analysis, and is as high as an estimate of the economic benefits Can it be concluded more generally is reasonable given statistical and inter forgone if a ban were implemented, tak that, as an approach to cancer preven 't v species considerations. Similarly, the as ing our estimate of 21 cancer deaths tion, studies designed to test dietary sumed 50 percent reduction in exposure averted per year if the chemical is a hypotheses in humans are likely to be seems as high as is realistic, although carcinogen, and assuming that there is at more productive (in terms of health conceivably a virtual LOO percent reduc least a 5 percent chance of a falsely benefits per dollar spent) than carcino tion could be achieved if the chemical positive bioassay result, we obtain a gen bioassays jh small rodents? The an were banned. Finally, the latency period cost-effectiveness ratio of $2.2 million swer depends on the degree to which the could be shorter than 20 years; as an per life saved, or S 186,000 per year of life examples chosen for analysis are typical extreme case for sensitivity analysis, we saved (29). Perhaps, on the other hand, of their classes. p-Dichlorobenzene was take it to be zero. acceptable and economical substitute given the highest priority for testing by Under all of these extreme assump products could be found, or minor ad the EPA on the basis of human exposure tions, the cost-effectiveness ratio for the justments in work practice and prudent and other considerations, and may there bioassay ofp-dichlorobenzene would fall use in the home could reduce exposures fore be considered a best case. p-Caro- URL 05411 * tche may also be considered a best case, uents discover what foods are good or since it is one of relatively few dietary bad for them. factors that are now ready for prospec As a final caveat, this framework as tive study. But even if the prior esti sumes that the value of the information mates for other agents such as vitamin E, yielded by a study lies in its ability to vitamin C, and selenium are ten times Influence decision-makers--in industry, less favorable than that for (3-carotene, in government, and private individuals-- such studies appear to be well worth the to alter exposures to the agent in ques costs when the expected benefits are tion. To the degree that a study contrib compared to those that might be derived utes to scientific knowledge per se, how from animal bioassays of industrial ever, it may lead indirectly to future chemicals or to current uses of health improvements in public health, and such care resources in preventive and cura considerations ought to affect the esti tive medicine. mated value of a study. From this per As an immediate policy implication, it spective, the value of basic research on may be concluded that, apart from pure mechanisms in carcinogenesis should ly scientific considerations, industrial not be underestimated as a result of chemicals should be carefully screened myopic applications of policy models on the basis of exceptionally high expo such as the one proposed here. sure or strong prior evidence of carcino This article began with the premise genic potential prior to the initiation of that the criterion for health resource long-term studies. The chemical-by-chem- allocation ought to be health benefit, ical approach to'discovering carcinogens somehow defined. This led to the criteri appears to be less cost-effective than on of cost-effectiveness. But how can the other uses of the same resources. institutions of our government--and so Large-scale national (and perhaps in- ciety, more generally--be structured to temrtional) tests of dietary hypotheses make these tradeoffs in domains as di seem to be promising and cost-effective verse as unimal toxicology and human uses of health resources. This analysis epidemiology, or cancer prevention and underscores the value of basic and epide cancer treatment? The mission of the miologic research to identify new dietary^ National Cancer Institute would seem to hypotheses. Given the strong evidence 'require that it reexamine these priorities linking diet to cancer, it seems likely that and .. allocate resources accordingly. large-scale epidemiologic investigations Moreover, the debate about hospital cost will generate hypotheses as promising as containment should not be carried on in (3-carotene, and some may have major isolation from concerns for environmen implications for cancer mortality in the tal health and chemoprevention. The United States. common value of all of these programs to Several countervailing observations society is health. As resources for might seem to lessen the strength of health-related activities become increas these conclusions. First, the population ingly constrained, we must ask ourselves of potential subjects for studies like that anew the question of whether we are of p-carotene is limited. It would be spending our resources wisely. Failure difficult to mobilize enough subjects to to do so may result in lost opportunities conduct more than a few such trials at to control the most dread diseases of our one time. society. Second, changes in personal behavior may be difficult to effect even if dietary factors are found to be protective. The response to data on smoking and lung cancer has been smaller and slower than might have been hoped for. However, a large share of the decline in cardiovascu lar mortality may reasonably be attribut ed to changes in health behaviors such as blood pressure control, diet, as well as tobacco use (JJ). Psychological and political factors 'may also favor continued vigilance over industrial chemicals, despite unfavorable cost-effectiveness ratios, and such fac tors must be recognized as legitimate. Involuntary exposures to cancer risks may be feared more than voluntary ex posures, and public officials do not gain much politically by helping their consul- hftrwcn and Nmn E. L. Wynder end O, B- Gori, J, Nail. Cancer hut. St. 129 (1977); I. Higginson and C. S. Muir. ibid. O, 1291 (1979). 2. R. DoU and R. Peto. Ibid. M, 1192(1981). J. Other, non complicated measure* could incor porate coocoras for the advene effects ofcancer on the quaHty of life, or the economic burden of cancer. It Is unlikely, however, that such addi tional comidorations would affect cbe relative cost-effectiveness of alternative cancer research investments. 4. J. Cairns. Cancer: Science and Society (Free man, San Francisco. 1978). This figure may be adjusted for specific tumor sites or demographic groups, but such adjustments will rarely make much difference for cancers in the adult age range. 5. The measure of effectiveness used here is con ceptually identical to the "expected value of sample informalion" of statistical decision the ory [H. Raiffa. Decision Analysis: Introductory Lectures on Choices Under Uncertainly (Addlson-Wesley, Reading. Mass., 1968)]. 6. One remedy might add the cost of intervention (weighted by the probability that the study result will lead to intervention) to the cost of perform ing the study to give the total expected cost to society. Alternatively, one might weight the cost 22 of intervention differently from the testing costs, where the weights reflect the relative lifesaving potential (that is, opportunity costs) of re sources diverted from testing or intervention, respectively (7). The problems with ignoring costs of intervention are brought to the fore by the observation that the costs of false posi tives--that is, agents for whieh intervention it inappropriately implemented aa a consequence of a positive study result despite the absence of a.true effect--will be excluded from the formal analysis. 7. M. C. Weinstein, Public Policy Ti. 333 (1979). 8. The rationale and mechanics of the cost-effec tiveness approach to health program evaluation are discussed in M- C. Weinstein and W. B. Stason, N. Engl. J. Med. 294.716 (1977); M. S. Thompson, Benefit-Cost Analysis for Program Evaluation (Sage- Beverly Hills. Calif.. 1980); and K. E. Warner and B. K. Luce. Cost-Benefit and Cost-Effectiveness Analysis in Health Care: Principles. Practice, and Potential (Health Ad ministration Press. Ann Arbor. Mich.. 1982). Strictly speaking, the optimality of the costeffectiveness ratio rule depenSg-pn the assump tion that investments (for example, studies) are mutually independent and divisible; in practice, deviations from these assumptions rarely alter the priorities implied by the ranking by costeffectiveness ratios, and arc unlikely to affect the examples considered here. 9. The amortization formula presumes an infinite horizon. If the benefits of altered exposure are expected to be of finite duration, then the appro priate C* would be sli^itly greater. 10. An alternative to this method of temporal adjust ment would be to calculate the present value of all costs and benefit* at the discount rate r. This procedure, while mathematically equivalent to that chosen here, he* two drawbacks. First, it is more convenient, and intuitively appealing, in this context to work with constant streams than with lump sums. Second, the procedure of tak ing the ftuure amortized values of economic costs may appear to be ethically more reason able than discounting ftiture lives saved. The two procedures are equivalent, however, in that they increase the relative burden of early coats to their proper value at the time at which bene fits are realized. 11. For a review of such priority-setting schemes, see National Academy of Sciences. Straiefies to Determine Needs and Priorities for Toxicity Testing (National Academy Press. Washington, D.C.. 1981). vol. I. 12. U.S. Environmental Protection Agency, An Ex posure and Risk Assessment for Dichlorobensenes. Final Draft Report (Environmental Pro tection Agency. Washington. D.C.. 1981). 13. Although111 of the first 227 chemicals tested by NTP were carcinogenic in at least one species, one cannot conclude that 90 percent of chemi cals are rodent, or human, carcinogens. In fact, since 1979, the proportion testing positive has been falling as the selection of chemicals has become baaed more on public health concerns. 14. R. L. Dehn aad C. T. Helmes. An Automatic Procedure for Assessing Possible Carcinogenic Activity of Chemicals Prior to Testing [Report prepared for the National Cancer Institute (Stanford Research Institute. Menlo Park. Cal if.. 1974)]. The National Academy of Sciences ()/) is implementing an empirical study in which existing data on carcinogenicity of randomly selected chemicals will be examined, thus pro viding an updated basis for estimating a prior probability for chemicals in general. 19. T. Lawk*. S. R. Haworth. pTVoytek, Environ. Mutagen. 1. 143 (1979). 16. . Crouch and R. Wilson \J. Toxicol. Environ. Health S, 1079 (1979)1 fitted data from NTP bioauays in the B6C3FI mouse and the Fischer rat to a single-hit dose response model with correction for backmouad tumors. Excluding the chemicals with the two most extreme esti mated values of b for both species, and averag ing the rest of the estimated potencies, a mean potency of 0.002 cancer (lifetime) per milligram per kilogram per day ingested was calculated for B6C3FI mouse and O.008 for the Fischer rat. In accordance with the result of Crouch and Wil son that potencies in rodents aod men correlate well on average, with a regression coefficient of I. a point estimate of 0.009 lifetime cancers per milligram per kilogram per day is applied. A model in which potency is measured relative to fractions of maximum tolerated dose, rather than absolute quantity ingested, might lead to more stable ana more valid potency estimates, but the data have not yet been analyzed in that form. 17. L. Tomatis, Anna.Rev. Pharmacol. Toxicol. 19, 911 (1979). Benzene is added to the 17 human SCIENCE, VOL. 221 Jy ^ y,: / j T URL 05412 i ' , carcinogens considered by Tomatis iq have been ' adequately tested. ' According to the EPA (.12), estimates of the number or workers exposed range from 500.000 (estimate from the National Occupational Hazi ard Survey) to fewer than 5000 (an industry . estimate prepared by Hull and Co. for the Syn* thetic Organic Chemical Manufacturing Associ ation). The latler is probably more appropriate : at levels of exposure close to the permissible la bour time-weighted average (TWA) ambient lev els of 450 rag/m1, or to the empirically observed TWA levels of 300 mg/m1. Thus it is assumed that 5000 workers are each exposed to 300 mg/ nr-for 8 hours per day. 250 days per year. Assuming each worker inhales 1.2 mVhour and absorbs all of the o-dichlorobcnzene, annual exposure would be (250 days per year)(8 hours Sir dayX1.2 mVhourXJOO mg/m1) 720 g/year. onvertlng to an average daaily dose in miiligrams per kilogram per day, one obtains (720 g per yearV(70 kg)(365 days per year) *> 28mg/kgday for 5000 workers. 19. This estimate is based on (he EPA's estimated frequency distribution of persons exposed to various ambient levels from 100 pg/m1 to s 0.1 lig/or, and a mean inhalation rate of 1.2 m'/hour while awake for 16 hours a day. and 0.4 mVhour while asleep for 8 hours a day. Environmental exposures from water were calculated on the basis of EPA estimates, and were found to be negligible compared to the air (less than 0.1 percent of the amount inhaled). 20. A report prepared by the NTP for the Senate Appropriations Committee found that of 98 chemicals testing positive in at least one species prior to September 1979. 55 hud proposed or Anal regulations on the record. However, many of these regulations were guidelines and not mandatory standards, and, with the exception of tris-BP (the flame retardant once used in chil dren's sleepwear), it is impossible to document any change in exposure. For example, no new standards were promulgated after it was shown by means of a CBSR that ethylene dichloride is a carcinogen, but there may have been intensified voluntary efforts to reduce exposure to'this high-volume industrial chemical. 21. R. Peto, R. Doll. J. D. Buckley, M. B. Sporn, Nature (London) 290, 20l (1981). 22. R. B. Shekelie, S. Liu. W. J. Raynor, Jr., M. Leppcr, C. Maliza. A. H. Rossof, Lancet 1981U. 1185 (I9SI). 23. S. Graham. C. Mcttlin, J. Marshall, R, priore, T. Rzepka. D. Shedd. Am. J. Epidemiol. 113, 675 (1982). 24. C. Mettlin, S. Graham, M. Swanson, J, Nail. Cancer Inst. 62, 1435 (1979); S. Graham, J. Marshall, C. Mettlin. T. Rzepka, T. Nemoto, T. Bycra, ibid. 116.66(1982). 25. A 5-year trial may seem short in view of the usual latency periods in carcinogenesis. Howev er, Peto et al. (2/) interpret the evidence from animal studies as suggesting that ^-carotene inhibits a late stage in carcinogenesis; if so. its effects would be seen in just a few years. 26. C. Mettlin and S. Graham, Am. J. Epidemiol. 110, 255 (1979); S. Graham et al., ibid. 113,675 (1981); C. Mettlin, S. Graham, R. Priore, J. Marshall, M. Swanson, Nutr. Cancer 2, 143 (1981); S. Graham et ai.,Am. J. Epidemiol. 116, 68 (1982). The percent reduction in breast can cer was based on a 3; I weighting of the relative risks for persons over 55 and under 55 years of age, reflecting the age distribution or breast cancer mortality in the United States. 27. The power estimate, g. Is based on a one-tailed test and is given by e - * ( l(/, (1 Nl,: (P, p.m*-P,) + PojF" l,w] where N sample size per group. P,, - placebo mortality rate. Pia treated mortality rate, and - cumulative Gaussian distribution function. With a two-tailed test, the power estimate would be .52. 28. E. Crouch and R. Wilson, J. Toxicol. Environ. Health 5. 1095 (1979). 29. With a prior probability of 0.1, a true-positive rate of 0.8, and a false-positive rate of 0.05, the probability of carcinogenicity after a positive study is (0. IX0.8VK0.1X0.8)-f (0.9X0 05)10.64. Hence, the cost-effectiveness ratio is (530 X IO*y(21X0.64J, or S2.2 x 10* per cancer death averted. 30. Based on an average retail price of 520 per hundred at discount pharmacies in the Boston area. 31. g-Carotene may be taken safely in large doses, up to 180 rng/day (U.S. Food ana Drug Adminis tration, Evaluation of the Health Aspects of Carotene!Beta-Carotene) as a FoodIngredient, prepared by the Federation of American Societ ies for Experimental Biology (PB80-119837. Na tional Technical Information Service, Springfield, Va., 1979)]. . 32. M. C. Weinstein and W.'Bvfflason, Hyperten sion: A Policy Perspective (Harvard Univ. Press, Cambridge, Mass., 1976). 33. D. M. Eddy. Screening foe Cancer: Theory, Analysis and Design (Prentice-HaU, Englewood Cliffs, NJ.. 1980). 34. J. D. Graham and J. W. Vtupel, Risk Analysis 1. 89(1981). 35. R. Levy, Anna. Rev. Pubi. Health 2, 49 0981). 36. Supported by grants from the Alfred P. Sloan Foundation and the Mobil Foundation. 1 thank D. Atkins for research assistance and J. C. Bailor in, P. Braun, J. Cairns, M. Thompson, and two anonymous reviewers for sugges tions. URL 05413 -R--E--S--E--A--R--C---H---A--R--T--I-C--L--E-- \ 'N mental pathways. Individual recessive mutations within the complex give less extreme segmental transformations than those resulting from deletions of the whole complex. These mutations trans-v Molecular Genetics of the Bithorax form pan of a segment or segments into * tissue appropriate to a more anterior Complex in Drosophila melanogaster segment, toward the ground state. There are also dominant mutations, which Welcome Bender, Michael Akam, Francois Karch transforma segment or pan of a segment into more posterior structures, away Philip A. Beachy, Mark Peifer, Pierce Spierer E. B. Lewis, David S. Hogness from the ground state (2). These domi nant mutations seem to upset the regula tion of genes within the complex and turn on functions in an inappropriate segment. A genetic map of the complex is The bodies of insects are divided into a plex is deleted, the animal dies late in shown in Fig. I. Most of the recessive series of segments. The segments are embryonic development and shows mutants and several dominant mutants formed very early in the development of striking changes in the segmental pattern show no cytologically visible rearrange the embryo, and cells from one segment of the embryonic cuticle. The third seg ments in the salivary gland polytene do not, in general, mix with cells from ment of the thorax and all eight abdomi chromosomes, and they can be recom other segments throughout the rest of nal segments resemble the normal sec bined with each other. The recombina development </). In the fruit fly Dro ond thoracic segment (2). Thus the sec tion distances between some pairs are i. sophila melanogaster, there are muta ond thoracic segment, which gives rise shown. The recessive mutations bx and i tions that transform parts of segments or to the pair of wings and the second pair pbx affect development of the anterior entire segments into the form of other of legs in the adult fly, can be considered and posterior halves, respectively, of segments. These homeotic mutations de the developmental ground state, and the the third thoracic segment. In the abdo fine genes that direct cells into different bithorax complex directs the more poste men, bxd, iab-2, iab-5, and tab-8 affect developmental pathways in different seg rior segments to specialized develop the first, second, fifth, and eighth ab- ments. The bithorax complex in Dro sophila is one of the best studied clusters of such genes (2); these genes determine the developmental fate of many of the thoracic and abdominal segments of the animal. When the whole bithorax com- W. Bender. F- Karch. and M. Peifer are investigator? in the Department of Biological Chemistry at Harvard Medical School. Boston. Massachusetts 021)5. M. Akam Is an MKC Senior Fellow in the Department of Genetics. Cambridge University. Cambridge CB2 3EH, England. P. A. Beachy is s graduate student and D. S. Hogness is a professor in the Department of Biochemistry at Stanford University School of Medicine, Stanford, California 94305. ?. Spierer is an investigator in the Department of Molecular Biology, University of Geneva. 1211 Geneva 4, Switzerland. E. B. Lewis is a professor in the Division of Biology, California Institute of Technology, Pasadena 91125. I JULY 1983 23