Document 6waL67omE2D3NQ2Xywg45kGQ1

1 / Cm EXHIBIT A ATTACHMENT III -R RAYTHEON COMPANY aytheo)nN> c:__I cnv.bonmcntT: A oceanogwphic services P.O. BO* >60 POATSMOljtH, Rt 09071 T F. L 1 40*`l4f `IOOQ Tw* ?lO.fi}-eQ35 1 fc L t X 095-n7 EVALUATION OF STATISTICAL METHODOLOGY USED IN DERIVING EPA WATER QUALITY CRITERIA % (as stated in the May 18, 1978 Federal Register, Vol 43, No. 97, pp. 21506-21518) Prepared for: THE AMERICAN PETROLEUM INSTITUTE P/Ea.^1 A. [v\ C- L A. Priya J. Wickramaratne, Principal Investigator July 25, 1978 CMA 049450 Section A B C D TABLE OF CONTENTS Pap>0 Introduction ............................................................... 1 Summary of Results ................................................ 2 Discussion of Results ....................................... g References . . . . :................................................24 CMA 049451 A. INTRODUCTION This report documents an evaluation of the stat is tical methodology used in deriving EPA water quality criteria, as stated in the Hay 18, 1978 Federal Register, Vol. 43, No. 97, pp. 21506 - 21518. * 1 Oft O49452 4 B. SUMMARY of results i 3 A summary of significant conclusions reached in our evaluation are itemized below: 3 1. Definition of Correction Factor - Preliminary analysis of data shows that use of a 3 ratio as a description of the functional rela tionship between standardized LC^g and exist H ing LC^g values seems to be an adequate descrip i tion for the majority of cases. However, analy sis of data from Table 3, indicates that the re lationship may not be valid for the flow through/ a static correction factor for all compounds. We suggest that a more statistically rigorous method such as regression analysis be used to derive correction factors. 3 2. Definition of the mean ratio as the geometic mean of the individual ratios is valid under the jastir 3 fiable assumption that the concentrations of the compounds are log-normally distributed. 3. Estimation of the ratios in Table 1 as the "aver aged measured concentration divided by the average nominal concentration in each toxicity test" is not appropriate because it is generally assumed that measured concentration is log-normally distributed. n rT 2 CMA 049453 4. The use of one correction fnctor for all compounds is justified only if it can be correctly assumed that the correction factors are independently dis tributed random variables with a common mean and variance, i.e. the correction factors are not com pound dependent. If this is not the case, then the estimates derived from these data bases could be highly biased. Similarily the "pooling" of data from different tests, procedures, waters, species, etc. without any compensating weighting factors is only justifiable under the assumption that the correction facters are independent of these variables. Neither of these assumptions, to our knowledge, have been validated or even tested. We suggest that the standard statistical methods of analysis of variance and covariance be used to investigate these aspects of data vari ability. 5. The estimation of flow through/static correction factors in Table 3, using only 12 compounds, the majority of which are pesticides, will probably result in a biased estimate, i.e. not be represen tative of other, non-pesticide compounds. We recommend that a larger and more representative data base be used to derive these correction factors. 3 CMA 049454 6. Implicit in 'he use of one sensitivity factor to represent nil compounds for various species of fish, (Table 4) is the assumption that the variance of LC^q values over different species of fish is independent of compound.. This hypothesis should bo investigated using standard statistical methods for investigating homogeneity of variance before making this assumption. Since it is intended that the sensitivity factor be applied to standardized LC^q values, vc suggest that the values used in computing the sensi tivity factor be standardized LC^q values, i.e., LCdcrUv values observed at 96 hrs, measured conccntrations using flow through methods. 7. Comments in Section C above, also apply to Table 5, Table 8, and Table 9. 8. No attempt has been made by EPA to estimate the pre diction error associated with each correction factor, and the propogation of errors using several correc tion factors. Data'in Tables 1, 2 and 3 were used to estimate the prediction error generated by standardizing an LC^q value measured at 24 hours, using static methods at nominal concentration to an LC^q value measured at 96 hours using flow-through method methods and mea sured concentration. The. minimum average error in prediction, under these conditions, was found to be 907.. 4 CMA. 049455 Data in Tables 6 ;''id 7 were used to estimate the prediction error generated by standardizing an value measured at 24 hours using static methods, to an LC^q value measured at 96 hours using flow-through methods (for invertebrates). The minimum average error in prediction under these conditions was found to be 1617.. Data in Table 5 were reported in mg/1 and yg/1 without units. Calculations were run as if they were the same units which will result in incorrect values of thc.averagc of the logarithms of the variance. For example, MATC values for Endrin, Chlordane and Methyl mercury are given in milligrams. Vfliile the MATC values for other compounds are given in V grams. In the estimation of a final fish chronic value, the use of equation X /YZ_(Section II.C.c) to compute a calculated fish chronic value is not justified under the following conditions: Z = Final fish acute value from I.G.2., where the LC5q value obtained from I.G.2. is as sociated with a species whose MATC value is known. CMA. 049456 c. DISCUSSION OF RESUL'l S This section gives a more detailed discussion of the items listed in the summary. Items 7, 9 and 10 have not been in cluded in this section, since we feel that they arc selfexplanatory. 6 4* # ,rij- (;. r r'-r-.i-- > A v 4 JN'f* " CMA 049457 1. Definition of ccrrcctitn factor: Let y. be the i-- standardized concentration and let x. be the i-- existing concentration, where y. represents either (measured concentration) - 'x^ (normal concentration) (concentration measured under static conditions) yi (LC^q measured at 96 hrS) - x^ (LC^q measured at 24, 48, 72 hrs) If we assume [1J that the y^ are random variables with a log-normal distribution, then the In y^ are random variables with a normal distribution. Let In y^ = f(ln x^) and assume that a linear relationship exists between In y^ and In xi., i.e., In = Mn x. + o + where e^ arc independent, normally distributed random variables. The standard statistical method of estimating S and a is by regression analysis, using the method of least squares [2J. This method gives estimates of p and a defined by: 3 = I (In x^ - In x) (In - In y) E(ln i - In x) 7 and 7 QA& 049458 On the other hand, if the correction factor is defined as the ratio: where k is a constant k xi Then, taking logs: In yi = In x^^ + In k. Comparing this formulation with that above for the regression analysis, the two formulations coincide if 0-1. Hence, we can conclude that the definition of the correction factor a* as a ratio is appropriate if 3-1. To see if E-l for toxicants of interest, consider Table la which gives a list of toxicants ^ith their corresponding A4 regression equations. The coefficients (a.3) were derived from analyses done on data taken from available references in Table 1 of the EPA document. Table lb gives a list of toxicants with the corresponding regression equations derived from analysis done on data from available reference for Table 3 of the EPA document. It will be noted the values of B in Table lb arc different, indicating that the ratio relationship may not be valid for some of the compounds in Table 3. Table lc gives a comparison of the equations defining the relationship between measured and nominal concentrations: a) using the EPA definition of the correction factor as a ratio, and b) using the correction factor derived from regression analysis. The comparisons are made for the toxicants listed in Tabic la (y is defined as the measured concentration, and x as the nominal concentration). 8 i CMA. 049459 1 Tabic l.a. Regression Equations for Selected Toxicants From Table 1 of EPA Document I----------------------------------------------Toxicant Malathion llexa chlorobenzene Chlordane Cadmium Kepone Regression Equation In y = 1.01343 In x - 0.27785 In y = 0.97739 In x - 0.65373 In y = 0.88925 In x - 0.25437 In y*= 0.97519 In x - 0.14637 In y = 1.07580 In x - 0.48S56 Tabic l.b. Regression E quations for Selected Toxicants From Table 3 of EPA Document Toxicant Pyrethrum RU - 11679 SB? - 13842 TI'K , ............... Regre.s s ion Eq ua t ion In y = 1.25106 In x - 0.90202 In y = 0.67105-' In y - 0.48466 In y - 0.99301 In x - 0.81792 In x - 0.91404 In x + 0.18491 -- - 9 CMA 04 94 60 Table l.c. Comparison of correction factor derived by EPA with correction factor derived using regression equations for toxicants listed in Table l.a. Toxicant Malathion A IHexachlorobenzene Chlordane Cadmium iCepone i Correction Correction Factor Factor Derived Derived Using by EPA Regression Equation y " 0.7 9x y = 0.61x . y = 0.56x . . y = 0 *9 Ox y ~ 0.83x y = 0.76 x 1,01 y - 0.52 x 0,98 y - 0.78 x -89 y = 1.16 x -98 ' y = 0.61 x 1`08 10 CMA 049461 2. Definition of Mean Ratio as Geometric Mean Let the relationship between and x^ be defined as In y' i. = In xi. + In k i= 1', . . . N where y^ is the standardized concentration,' is the existing concentration and k is the ratio yJ i./x.i If it is assumed that the individual concentration of the toxicants are log-normally distributed, then the minimum A variance unbiased estimate of the mean ratio, In k, over the total number of N observations is given by [2] : In k = N \i=l N In V. - 1 In X. 1 i=l r` 1 f l ( In y. - In x. V " N \i=l ' 1 1J In (y x. 1 , N y. ii ii = N i=l xi ** In -n/r y_i V xi i~l - ln (^`W) This implies that the minimum variance, unbiased estimate at k is WN k V i~l (vl./x.i)` = the geometric mean of the individual ratios 11 CMA. 049462 It is noted that if the individual concentrations are not A log-normally distributed, then k = the geometric mean of the individual ratios is not necessarily the best estimate of k. Thus, implicit in the use of the geometric mean as an estimate of k, is the assumption that the individual con centration are log-normally distributed (See Section 3). 12 CMA 049463 \ 3. Estimation of Ratios in Table 1 Estimation of an average 'ralue for the ratio in each toxicity test in Table 1 (of the EPA document) as the "averaged measured concentration divided by the average nominal concentration in each toxicity test" is not appropriate under the condition that each measured concentra tion is log-normally distributed. The appropriate estimate of the mean ratio for each toxicity test is the geometric mean of the individual ratios as derived in section 2 above. 1 It should be noted that the "averaged measured concentration 1 divided by the average nominal concentration in each toxicity test" does not have a log-normal distribution, 1 under the condition that each individual measured concentra tion is log-normally distributed. ! I 1 1 3 0 0 a 3 b 13 CMA 049464 4. Use of One Correction Factor for All Toxicants The use of one correction factor for all toxicants is justified 'only if it can be correctly assumed that the correction factors are independently distributed random variables with a common mean and variance, i.e., the correction factors are not toxicant dependent. Similarly, the "pooling" of the data from different tests, procedures, water, species, etc., without compensating weighting factors is only justifiable under the assumption that the correction factors are independent of these variables. The objective of this analysis as we see it is not to estimate a "grand" mean correction factor that is repre sentative of all toxicants under a wide variety of con ditions. This, however, appears to be what EPA has taken as its objective. Rather, the objective should be to estimate a correction factor, or series of factors, which, will predict the desired concentration with a minimum of error. Let us assume that the correction factors corresponding to the individual toxicants arc random variables, ylt y2, ..........v Let In yi C N (Vi 1 ,02) In y? C N (u?,a2) In y s C N C U 3 , O 2 ) In >n C N CVoO 14 CMA 049465 i.e., In yi are normally distributed, with the same variance but different means. It is desired to find estimates of y1. y2.*-, yR. if y 1 = ^2 - = y,, = vj0 then the best estimate of y0 is AA A yi + yz + ... + vn n But if y i t* y 2 t . * / yn then AA A yi + y 2 + ... + yR _ will give a biased estimate of yi,*y2, ... and yR. To illustrate this point we selected 4 toxicants at random from Table 1: Toxicant Acrolein Carbaryl Carbofuran Chlordane Ratio 0.76 - yj 0.99 = y 2 0.20 - is A 0.56 = y4 Under the assumptions described above (i.e., the y toxicants have different expected values) using the geometric mean of these 4 values to derive one correction factor for all four toxicants will give biased results. The geometric 15 CMA. 049466 mean of the four ratios -..as computed to be 0.538. Using this value as the correction factor for all four toxicants will result in consistently under-estimating the measured concentration of Acrolein and Carbaryl, and consistently over-estimating the measured concentration of Carbofuran. Under these circumstances a far more accurate method of predicting the measured concentration would be to use the . 4 individual ratios as estimates of the correction factors for the 4 toxicants. If, however, we can show that n11 4 correction factors do indeed come from the same population, then the estimate (pi + -,i2 +. ts + {I.,) is the best estimate of p (the common expected value) and it would be appropriate to use a common correction factor. In the light of these observations, we strongly recommend that -the standard analysis of variance methods be used to, test the hypothesis that the mean correction factors for all toxicants arc equal. If this hypothesis is rejected, then we can use various other statistical analyses to identify which toxicants or group of toxicants have significantly different correction factors. A series of correction factors can then be derived to predict the standardized concentrations. The same methods can also he used to investigate the 'variation of correction factors with parameters other than toxicants, such as hard and soft water, methods of measuring LC5rUr., etc. 10 CMA, 049467 5. Estimation of Flow Through/Stntic Correction Factors We feel that the data contained in Table 3 is an inadequate data base from which to estimate flow through/ static correction factors, for the following reasons: 1) Only 12 toxicants are represented, the majority of which are pesticides. Thus, the correction factor derived from these data, may not be representative of other non pesticide compounds. 2) There is a wide variation in correction factor estimates even within the toxicants listed in Table 3, reflected by the difference in functional form of the associated regres sion equations. Me suggest that a larger and more representative data base be used to derive these correction factors. 17 CMA 049468 6. Sensitivity Factors The use of one sensitivity factor to represent all compounds for various species of fish, m Tables 4, and 8 is' justified only under the assumption that the variance of LC^q values, and >1ATC values over different species of fish is independent of toxicant. We suggest that this hypothesis be investigated using standard statistical methods for investigating homogenerity of variance, before making this assumption. If this hypothesis is not rejected, the best estimate of the common logarithmic variance is J3J. N-k a given toxicant, k, is the number of individual toxicants listed in a given Table, and N is the total number of observations. We suggest that under these circumstances, the above estimate be used in place of the "Average logarithmic variance" proposed by KPA in their computation of the sensitivity factor. If this hypothesis is rejected then we suggest using a scries of sensitivity factors associated with groups of toxicants of similar variance. These comments also apply to the sensitivity factor;: derived in Tables 5 and 9, for invertebrates. 18 CMA 049469 tt. i rcaretion error and PicpoKaliun ol Error No attempt has been made by EPA to estimate the prediction error associated with each correction factor, and the propogation of errors using several correction factors. Table 8 gives the logarithmic standard deviation of the cor rection factors from Tables 1, 2, 3, 6, 7 and 10. It also lists the geometric mean and the 957= confidence limits for each correction factor. The last column in Table 8 gives the average percentage error pf prediction in using a given correction factor. This value is computed in the following 'manner [4J: A Let x = LC5q Cl) where LC^q is the standardized LC^q values using the estima ted correction factor from the appropriate table. A Let u = In (2) If the error in prediction, using the value In LC^q to estimate the actual value of In LC^q is denoted by Ax, then: A / Ax - In LC50 - In LC5Q (3) From (1) and (2) we have: u x=e and ax = xau (4) 19 CMA 049470 Defining aA and a U as estimates oi Ax, ana au wc uulu from (4) the relationship . And OX 3 X ffu ox = ou x Defining error in x (LC^q) which results from using the correction factor as o , the percentage error in LC^q is cA /x which is equal to aU . It should be noted that the estimate of the logarithmic variance given in the above table is the minimum estimate of this variance, under the assumptipn that each indivi dual ration in the original tables has zero error. Thus, the average percentage prediction error shown is the mini mum average error. 20 CMA 049471 Prediction Errors Associated With Correction Factors Table " of FFA Docu ment . Correction factor In S: \ndard deviation Geonio.tr ic mean 957. Con fidence limita 7. error 1, predictin' 1 Measured/ 0.3236 0.77 (0.40. 32.36 Nomina 1 1.45) 2 (Fish) LC 96 he 24 LC50 / LSo LLCtsu96 ^ LLCU3048 ic 96 he 72 So / 50 0.4953 0.66 0.2465 0.1152 0.81 0.92 (0.2$. 1.74) (0.50, 1.31) (0.73. 1.15) 49.5? 24.65 11.52 3 (Fish) Flou Through/ Static '0.6701 0.71 (0.19, 2.64) 67.01 6 (InvertcbraLes) LC 96 he 24 LSo / LL50 LC 96/ IC 48 LU30 / L oO LC 98 / I C 72 LL50 / iAj50 0.S270 0.6168 0.7122 7 ( Invu'Li'- b rates) Flo\; Through/ Static 10 V.C/LC,jQ 1.3/66 0.6001 0.26 0.43 0.55 (0.05, 1.32) (0.13, 1.44) (0.14, 2.2?) 82.70 61.63 71.22 1.1 0.4 4 (0.0/. 16.34) (0.11 , 1 . 69) 137.66 6S.61 A A a/ino -------- 21 E .1 i ikH '-'I''.' ^ *-'-- -- -- ... *- *jf-*--^ -- - * -* ^AV- - - * j- - m`~ m*Lu. Wt ' Es t nr.au iun 01 i ro'pc>;.u.. iua The error generated by using more than one correction fac- ` tor is derived by using the following relationship; Let let LC LC 50 50 (e) (s) be be the the existing IjC50 value and standardized LC^ value obtained by multiplying LC^q^ by n correction factors lTJ, R; . . .Rn , Then LC 50 (s) In LC 50 (s) Ra-R2 ...Rn.LC50(e) In LC50(e) + In ITT + In ill + ... + In Fn 12 Define the variance of In R, to be o3 , ...In R^ to be o n, and assume that In Ri, ...In R are independent random vari ables with expected value In Ri, In R2, ... In R^, Then var. In LC^q^5^ = ++ an (2) Data in Tables 1, 2 and 3 were used to estimate the predic tion error generated by standardizing an LC^q value mea sured at 24 hours, using static methods at nominal concen tration to an LCca value measured at 96 hours using flouthrough methods and measured concentration. In this example, we can see from Table 8, that: oi = 0.3236 cr2 = 0.4953 a 3 = 0.6701 22 CMA 049473 Substituting these valm ; of in equations (1) and (2), we find that the minimum average error in prediction under these conditions was found to be 907-. Data in Tables 6 and 7 wore used to estimate the prediction error generated by standardizing an LC^q value measured at 24 hours using static methods, to an LC^q value measured at 96 hours using flow-through methods (for invertebrates). The minimum average error in prediction under these condi tions was found to be 161%. 23 CMA. 049474 REFERENCE 1. Finney, D.J., 1964. Statistical Method in Biological Assay. Second Edition. (Charles Griffin & Co. Ltd., London) pp. 37-43 Draper, N.R. , and Smith, 11. 1966. Applied Regression Analysis. (John Wiley & Sons, Inc. New York)- pp. 1-35 Zar, Jerrold H., 1974. Biostatistical Analysis. (Prentice-Hall, Inc.) pp. 135-136 4. Bevington, Philip R. 1869. Data Reduction and Error Analysis for the Physical Sciences. (McGraw Hill BooH Co., New York), pp. 56-65 24 CMA 049475 CMA EXHIBIT B COMMENTS ON WATER QUALITY CRITERIA EPA APPENDIX C Federal Register March 15, 1979 pages 15974-15980 and ATTACHMENTS IV - VI CMA 049476 The comments to the Health Guidelines will be broken into five sections. 1.0 2.0 3.0 4.0 '5.0 A Proposal for Assessing The Extrapolation Models Specific Comments Conclusion Literature Cited Risk 1.0' A PROPOSAL FOR_ ASSESSING RISK We do not agree with the policy statement that only the one hit-model should be used for risk assessment of carcinogens. Such a rigid policy precludes the use of alternate models which may best describe the experimental data. We would like to propose a general framework for the problem of making risk assessments of carcinogens. 1. Qualitative Assessment. This may be made along the lines suggested under the heading of Effects, p 15976. 2. Quantitative Risk Assessment. Once the qualitative assessment has been made, an initial effort at quan titation is performed. The basic premise is that risk assessment is largely a judgmental process made by scientific experts and will have to be made on a case-by-case basis. For example, if the experimental dose response data is more appropriately fitted by another model then such a model should be used for extrapolation. Epidemiological data should also be considered in assigning the final risk. This will allow the use of all the experimental data in making the judgment as to risk. To do otherwise would be to ignore the data for the sake of the simplicity of using the one-hit model. a. Simplistic Approach. The simple one-hit model as advanced by the Agency has utility in infor mation gathering and for prioritizing chemicals as to which materials need additional data. If we confine our attention of the one-hit model to prioritizing, then it does have utility in the ranking of materials. b. Once further data are obtained, a choice of extrapolation methods should be guided by in dividual facts from all experimental data rele vant to carcinogenicity. Such models as the -1- CMA 049477 multi-hit model, the linenr extrapolation from the most reasonable data set, the pharmacokinetic model, and any other model which has relevancy to the experimental data or emerging theoretical considerations should be used, c. Once the risk in animals has been defined, the considerations to be used in translating to man include the following; i. Metabolic similarities ii. Epidemiological data d. In order to assess the risk in humans, a pro gram should be designed to estimate exposure of the population at risk. With these considerations in mind, the mathematical descrip tion of the one-hit model as advanced by the Agency appears to be adequate. This simple model should be used for prioritization of possible carcinogens and used for risk assessment only when no other data supports a less conservative model. As an example, the comments on the tetrachloroethylene document should be examined. In the comments by Dr. Hartung it is clearly indicated that the one-hit model produces a risk assessment which is contrary to what has actually been observed. In such a case the one-hit model is inadequate and a model which more closely reflects reality should be used for risk assessment. 2.0 THE EXTRAPOLATION MODEL historically the evaluation of safety of non-carcinogens lias consisted of the application of a safety factor or uncertainty factor to an experimentally determined "no observed effect level" (NOEL). While a safety factor approach (1:5000) applied to a NOEL has been proposed as an adequate control for carcinogens (Weil, 1972), this approach has not received wide-spread acceptance. Instead, there has been a significant trend towards the use of various extrapolation models. While there have been extensive arguments pro and con regarding the existence of threshold doses to produce carcinogens, nearly all of the presently proposed extra polation models assume the absence of a threshold. The first ex trapolation leading to a standard can be found in the DEIR Report (NAS, 1972) which used a linear model to estimate the effects of ionizing radiation on man. Most of the data in the report were derived from acute exposures of A-bomb victims or from medical x-ray exposures. In many instances several different extrapola tion models could have been accommodated by the spread, in the data of the DEIR Report. The linear model was chosen because it was the most conservative, i.e. it gave the highest estimates of risk -2- CMA 049478 for any given dose level. However, in the case of experimental data in animals, pure linear models of the type Ft = bd (where Ft is the fractional tumor incidence, d is the daily dose, and b is the slope of the line) do not provide a good fit of the experimental data. Such data have traditionally been analyzed by probit analysis (Finney, 1971). or logistic analysis (Berkson, 1944), Both of these models are non-threshold models which fit the dose-response data after transformation. -* For the probit analysis the fractional tumor incidence converted to probit (P) by the relationship: is t e d 1L (See Table 1) The dose (d) is transformed to x = log d. After maximum likeli hood solution a probit regression line in the form P = a + bx is developed which can be extended outside the experimental range with the usual caveats. For the logistic analysis the fractional tumor incidence is converted to logits (L) by: Ft L = In --------- (See Table 2) The dose (d) is transformed to x = log d. After a minimum chisquare solution a logistic regression line in the form L = a + bx is developed which can be extended outside the experimental range with the usual caveats. In general, both probit or logistic analyses provide a good fit for experimental data within the actual dose ranges of the experiment. However, when these two models are extrapolated out side the experimental range, they start to diverge significantly (see Table 3). Mantel and Bryan (1961) developed an extrapolation method based on the probit method that provides more conservative es timates of doses required to produce risks of less than speci fied values. In that method they select the upper 99th percent confidence value of an experimental point and extrapolate down ward from that point with a slope of 1 probit/log cycle. This is illustrated in detail in Figure 1 from Mantel and Bryan's publication, copied here without their permission. The method is applied to the data in Table 3 for comparative purposes, with -3- CMA 049479 Table 1 Percent to Probit Conversion Table 0 1 i. 3 4 5 6 7 8 9 0 10 3.718 2.673 3.773 2.946 3.825 3.119 3.374 3.249 ' 3.355 3.920 3.964 3.445 4.006 3.524 4.046 3.595 4 .035 3.659 4.122 20 4.159 4.194 4.228 4.261 4 . 294 4.326 4.357 4.388 4.418 4.447 30 4.476 4.505 4.533 4.561 4.5S8 4.615 4.642 4.669 4.695 4.721 40 4.747 50 5.000 4.773 5.025 4.793 5.050 4.824 5.075 4.849 5.100 4.875 5.125 4.900 5.151 4.925 5.176 4.950 5.202 4.975 5.227 60 5.253 70 5.524 80 5.841 90 6.282 5.279 5.553 5.8 78 6.341 5.305 5.582 5.915 6.405 5.331 5.612 5.954 6.476 5.358 5.643 5.994 6.555 5.3S5 5.674 6.036 6.64 5 5.412 5.706 6.080 6.751 5.439 5. 739 6.126 6.831 5.469 5.772 6.175 7.054 5.4 9 5 5.806 6.22 7 7.327 Extreme Risk -1 10 -7i 1,, 0 - iO -7 10 ' 10-8 > Values Approx. Probit 1.909 1.281 0n .7o3/ 5-t -0.199 -0.612 -*i \Jqa1culated according to polynomial approximations of Abramowitz and Stegjd (196o) Literature Section - Abram and Stegun Handbook of Mathematical Functions (National bureau of Standards, 1963). I ^. t X 1 '* ' 049480 L = In 1-P 0 0 10 _2 . 197 20 -i .386 30 -0 847 40 -0 405 50 0 000 60 0 405 70 0 847 60 1. 386 90 2. 197 1 -4 .595 __ 2 091 -i 325 -0 800 -0 364 0. 040 0. 447 0. 895 1 450 2. 314 2 -3 892 -1 992 -1 266 -0 754 -0. 323 0 030 0 490 0. 944 1. 516 2 442 Table 2 Percent to IngJt Conversion Table 3 -3 .476 -1 901 -1 .208 -0 703 -0 282 0 120 0 532 0. 995 1 586 2. 587 4 -3 178 -1 .815 -1 153 -0 663 -0 241 0. 160 0 575 1. 046 1. 658 2. 752 5 -2 944 -i 735 -i 099 -0 619 -0. 201 0 201 0 619 1 099 1 735 2. 944 6 __ 2 752 -i .658 -i 046 -0 5 75 -0 160 0 241 0 663 1 1 53 1 815 3. 178 7 -2 .537 -i .586 -0 995 -0 532 -0 120 0 232 0 708 1. 208 1. 901 3. 476 8 _ t .442 -i .516 -0 944 -0 490 -0 630 0 323 0 754 1 266 1 992 3. 892 9 _ 2 314 -i 4 50 -0 895 -0 44 7 -0 04 0 0 364 0 800 1 325 n 091 4. 595 Extreme Values Risk logit io-3 io~4 io-5 10^ 6 10~7 10-8 -6.907 -9.210 -11.513 -13.816 -16.118 -18.421 in I CMA 049481 C ui C CMA 0 4 9 4 8 2 Table 3 ILLUSTRATION OR VARIOUS CAS'lC PRO. JLCI'J ON MCOL'I.S. Assume: a dose (d) of I Omg/kg/day of compound A produces a SOS liunor incider.co in man, and a dose of 5mg/kg/duy produces a Lu;:.or incidence of 25%; Then: the probit regression model is P=2.761 + 2.239x; the logistic regression model is L=-3.651 + l.tolx, Khere P=proLit, L=logit, and x=log dose. ________________ fractional fnmo r iu Idcncc i'rob i t Project Ion Projected Doses mg/kg/day 1.0 U j St i c Project ton Mantel Bryan * Projcctions 8=10 d= 5 Single Hit P rojec tions d=10 d=5 0. 50 10.0 10.0 10.0 10.0 12.0 0.45 8.79 8.81 7.50 8.62 10.4 0.40 7. 71 7.75 5.58 7.37 8.83 0. 35 6. 73 6.77 4.12 6.22 7.49 5.83 5.86 2.99 5.15 6.20 0.25 5.00 5.00 2.12 5.00 4.15 5.00 0. 20 '21 4.17 1.44 3.40 3.22 3.88 0.15 3.45 ! 3.35 0.92 2.17 2.34 2.32 0. 10 2.68 2.50 0.52 1 .23 1.52 1.83 0.05 1.84 1.56 0. 23 0.53 0.74 0.89 0.01 U) J 10 A* 0. 91 0.416 0.218 0.5 5 0.123 -2 3.0 x 10 0.047 8.1 x 10-3 -3 1.9 x 10 0.11 1.9 x 1, 0,,-2 4.5 x 10 -3 0.14 -2 1.4 x 10 1.4 x 10 -3 0.17 1.7 x 10 -2 1.7 x 10 -3 10-5 0.124 7.0 x 10,"3 5.4 x 10 1.3 x io~3 1.4 x io~4 1. 7 x i(T4 ^0 O1 -7 10 10 -8 7.5 it 10-2 -2 4.8 x 10 -2 3.1 x 10 1.6 x 10~3 -4 3.8 x 10 -5 9.0 x 10 1.8 x 10'4 -5 6.3 x 10 , -5 2.4 x 10 4.2 x 10-4 1. 5 x ,10,,-4 5.8 x 1, 0 -4 1 .4 x 10~5 -6 1.4 x 10 -7 ] . 4 x 10 1. 7 x io'5 1.7 x 10 -6 1.7 x ,10,,-7 * 'These y arfel-Ar'an nroiections originate direct ly from, th e date-resnons e cur v e . The use of a 9 9/ conf idence level would significantly decrease the dose estirr;a:G5 t. t ,e ser 1 i his ""abl ' juT-'.i: us to drop them : nf experience it should . actual!" simpler net: rd "[to fact. the 1 m connection with the ' vrrrpriate method is t .Id :cf|utxc a somewhat, \r:::!o for completeness, t'f. -:-,bduy of inversions, ! -ocriure just discussed -,m i 1:), are the results .,-rrr.e into mice. 12 an* :->ed'r may refer to the -' are on inversions. At o roomie combined A: tlu. middle four levels '"n.r.r 'cmiUs end. finally, all yielded 100 percent '.-;t three columns show, r" id be i separate are only one combined . Fe. each such result, . at the 22 percent assurfd tables or calculated The norma! deviate cor'm, te.bles of the normal column 7 shows the Fee nta.'htnumi for this, .,vLT-aU calculated "safe" >n- a_i to make impractical l.v humans. But wc nre .d if at)}- compounds arc -. a^ro, this is one of them, d ani'.!;. As of the c.tperi' . n (>n the vertical scale, i is shown as negative pr'-tin: the outcomes at shown bv arrow. The u probit bar fitted to the ; \L CANCir. INS7ITUTS -- r* Figure 1 TESTtNO Of C.tr.CtNOUZ.VS !( Tablc 2.--Illustration of metoodoloey for cet'-rrr.icnc the "uln" dose from reu at several dcse icsels; data from Br.an nod ibinLia (/;; * Dofc n */ mouse Lo^ cost P.'SUlt No of tumors No of mice Comb'.-ied result No of tlrr.ors No of m11e 3HD P V J1 u ? 99 assur ance Cor:*spotc..-; Dorns! de s ;ac Cal It: (1 : tuT k? C/; tn o ooc:;t 0 con?rd 0 CO 105 0 0030 0 007$ 0 01 jo o c:c: 0 OO'.'j 0 105 0 05 0 30 1. 0 (7) 6 3S8-10 6 900-10 7. .'01-10 7, jo:-n 7. S93-IO a i04-io 3 40wlO 8 703-10 0 007-10 9 39S-10 o e^o-io 10. 000-10 (3) 0/7O 0'4 1 0. 10 0' 10 3'17 6/1 8 13 :o 17 :i :i :i :i :i :i :i :o,:o (4) 0, !,iS 0,70 0,39 0.10 3 17 5MS 13 00 17,21 -- -- (5) 0 OO's 0 OsJ 6 0 114 1 o 0 4-0 o ::o 0 S7I a ois -- -- -- (6) - 1. 501 -1. it; -1 :oj -0 7-0 -0 C30 f 0, 610 - 1. 131 t i. 7:s -- _ __ "" f; o *. . ? C. "> * 2 ;. 1. 0, 1. 71 L 4! " data. Above the first S data potolr. the tmnr!"s shown correspond the nia^mum P values of table 1. Extrapolation, with the slope of Dormal deviate per comnum log to the over-all calculated "saf^'' vs1 is indicated by a broken line. All iriaugtes, other than the one from w. TtiT*rir,e'nc 1 --Estimation of the "mfc" dose from test Jesuits mth s eircui" saethj tchislanthrfn*. at aewral dost levels. At cneh test lesrl both the obv percentage rrsponsc and so upper limit, 00 percent assurance, baa'd "n eontu data are ibown. Solid tint is the msuouin litelibood protm line fitted to `data The "safe" le-. ! of 0 10*' n; per mou." is is tfcr, irv.tsnrc etiitr.a'.r.t extrapolation wstb the conservative slops of 1 normal deviate per lo; from upper limit on P at the second dose level. vol. :t, no. acqust mi i -7- a.1 i i. j-*' "a se CMA. 049483 -- the difference that the extrapolation is made from the central value rather than an upper 99th percent confidence value. The method combines two "safety factors" by extrapolating at a very shallow slope which is lower than normally encountered in em pirical data, and by extrapolating from an upper 99% confidence value. The Mantel-Bryan extrapolation model has received only limited acceptance, possibly due to the unknowns associated with the assumptions used in the extrapolation process. More recently, sev.e&al variants of the "single hit" model have been used as the basis for extrapolation models. The basis of this model is the theory that a tumor can be produced by the interaction of a single molecule with a susceptible target mole cule in the organism. In its simplest terms this model can be presented as (Hoel et. a_l., 1975) : where R has been termed potency and the other terms are as defined previously. Then, for very low doses this relationship becomes linear, so that Ft = Rd A variant of this model has been adopted by the U.S.E.P.A. to extrapolate carcinogenic data to man. The model attempts to introduce various correction factors for the duration of the test, the size of the test animal, and the incidence of tumors in con trol animals. It should be noted that the single hit model often docs not provide as good a fit of the experimental data as the probit or logistic models (see example in Table 3). The National Academy of Sciences Safe Drinking Water Committee proposed the use of a multi-step model for use in extrapolating carcinogenicity data (NAS, 1977). This model was developed by Armituge and Doll (1961) and is presented by: Ft = 1 - e_ o + /l/ d + ^ -f * * * * * ^ A ^ J It is clear that there are many extrapolation models, and that these models result in significantly different extrapolations (Table 3). The extrapolations at risk levels of 1:10^ arc grossly different and exceed any observed differences between species and also exceed any differences observed for any experimental designs. It should be noted that each one of these extrapolation models is conceptually based on very simplistic foundations. None of the models considers the multiplicity of causative and repair factors that influence the final occurrence of a hit. A valid extrapolation model should consider at a minimum the causative factors, metabolism, -8- CMA. 049484 immunology, repair and individual variability. Since the balance between these factors may differ drastically for different chemi cals it is that a single extrapolation'model cannot be effective to cover all chemicals. If there were such a theoretical model, then it would certainly be much more complex than the single-hit model.. 3.0 SPECIFIC COMMENTS Page 15974. Under the section Objective (I) it is stated that ideally one would want to derive criteria for absolute safety. However, the determination of absolute safety is in itself a logical impossibility. The Agency makes this same point in the preamble, consequently the statement on p 15974 should be deleted as an objective. While some of the aspects of this problem are covered in the discussion under objectives, they are not suffi ciently clear. The difficulties in establishing safe or absolutely safe levels are inherent in the process of safety evaluations. Most scientific investigations seek to determine that specific findings have really occurred and that they are not due to the vagaries of chance. While this is the exact approach taken during the evaluation of environmental and toxicological processes, the application of data gathered in these investigations to safety evaluations require significant change in the handling of the data. In addition to data which demonstrates statistically sig nificant effects, it is necessary to attempt evaluation of data which demonstrate no effect. To demonstrate that a "no-measured effect level" is real and not due to statistical variability is difficult. The difficulties are escalated when one considers that all studies are conducted in model test systems. The results from such studies are used to estimate the chances that no unto ward effects will be seen in different species, under different circumstances and from much larger populations. Under such cir cumstances, it becomes clearly impossible to prove the negative, that is to prove absolute safety. Thus, safety evaluations are inherently judgmental and only the relative level of safety can be described, but absolute safety cannot be guaranteed. Page 15975. Types of Criteria (II) The use of the terms "stochastic effects" versus "non-sto chastic effects" is misleading, especially since they are de fined in terms of the occurrence of a threshold and the absence of a threshold. Such distinctions should be made strictly in terms of threshold and non-threshold models rather than confusing the issue with the terms of stochastic versus non-stochastic, which in themselves do not necessarily define the occurrence or absence 9 CMA. 049485 of a threshold. At the present time, the concept of threshold versus non-threshold effects relate largely to the mathematical models applied to the experimental data rather than to the em pirical data themselves. To clarify, it is common to find dose levels at which experimentally no effects are found even for events such as carcinogenicity. To clearly establish whether a threshold does or does not exist on an experimental basis is difficult. Laboratory data which demonstrate no carcinogenic effect at a certain low ..dose level, when such an effect is seen at a high dose level, are always open to the conjectural criticism that the effect might have been seen if more animals had been ex posed. Epidemiological data are also limited in their applica bility because of background rates of incidence of tumors in populations which have apparently had no exposure to a particular chemical. Since the empirical data are not in themselves con clusive in establishing which events are threshold events and which ones are non-threshold events, the selection of those effects which were to be treated by non-threshold models has been limited to one category of effects and that is carcino genicity. Since various defense and repair mechanisms have been identified even in the induction of cancers by chemicals, it has not been clearly established that the same non-threshold model can be applied to the evaluation of very low dose level of all chemical carcinogens. The question of the possible occurrence of thresholds cannot be clearly answered at the present time and should thus be left open. A better understanding of metabolic effects, immunological responses and repair mechanisms will undoubtedly suggest practical threshold phenomena which may relate to key steps of the carcino genic process on a case-by-case basis. The statement on organoleptic effects should be re-worded to read: "it is recognized that criteria based solely on organo leptic effects do not represent satisfactory approximations of low risk levels." The Approach, p 15975 We approve of the Approach in that the four indicated topics need to be addressed, namely, exposure, pharmacokinetics, toxicity and criterion formulations. We especially commend the Agency for their recognition of the importance of pharmacokinetics as a method for assessing the species to species extrapolation and in character izing the mode of toxic action. Clearly, the more we know about the mechanism the more reliable will be the resulting criterion. Page 15976. The refinements of the use of the bioconcentration factor appear to be very useful. And the distinction between those substances which are lipid soluble and those which are non-lipid soluble is very necessary. - 10 - CMA 049436 The reference to the EPA Interagency Cancer Guidelines should indicate that this document is under intensive peer review and the conclusions could change once the review is completed. We agree with the Agency on their recognition that a chemical which has not induced a significant cancer response in humans or experimental animals is not identified, as a possible carcinogen. in this con nection the Agency cites that there are cases where weak carcinogenie'effects are observed but the data cannot be used to estab lish criteria, we would add that bioassay systems that are con ducted at doses greater' than the maximum tolerated dose or by routes other than oral may suggest that the chemical has potential for human cancer but should not be used to establish numerical water.'OualltY Criteria. Pages 15977-15978. The assumption that those chemicals which induce mutations are als likely to induce cancer in animals and therefore presumably in People is only a theory. While the majority of materials which h ave been shown to be carcinogenic in animals are also capable of inducing mutations, it has not been dcmonstrated what proporti cbn of chemicals that are found to be mutagenic will also be fou nd to be carcinogenic in animal tests, In addition, epidemiolog cal evidence for man has failed to provide evidence of carcin ocjjenicity for many materials which have been found to be carcin ogenic in animal tests. In addition, it must be noted that there are a few chemicals which appear to be carcinogenic in man by epidemiological evaluations, which to date have not been shown to be carcinogenic in laboratory animals, Therefore, it is clear that the extrapolation of data gathered in bacterial systems to experimental animals to humans is Crought with great difficulty, At the present time, the finding of mutagenic effects can only be utilized as a screening test and for the establishment of priorit es for further testing before materials come into contact with m an. To date, the select on of animal models for carcinogenicity testing has been relative ly non-critical and has been largely restricted to the selectio of animal strains which have demonstrated high sensitivity to chemjiical carcinogenesis. in the face of all these uncertainties, eve the animal data should be examined critically on a case-by- case basis. In the discussion s action on stochastic effects and the justification for the se lection of the single-hit model, it is stated "The possibility of threshold mechanisms has given support to other dose response m odels (Guess ct _al. , 1977) for cxtrapolation of cancer risks, su ch as log-probit, logit and multi-stage models. These models ar e nonlinear and result in lower risks - 11 - MA. 049487 than the one-hit model fo r a given low exposure." This statcment is at least partiall y in error and the error is significant, All of the extrapolation models cited above are non-threshold models and, in addition, all of these models have the characteristic that they tend to f it the experimental data better than the single-hit model does. A t extremely low dose levels those models are c-lose enough to linea city that the selection or rejection of one model over the other cannot be made on the basis of experimental data even for such artificial models as mutagenesis in bacterial systems induced by radiation or chemical. Similarly, all of these models can b e readily applied to the epidemiological studies done in man and w ill fit v/ell within the margins of error provided by those studies Therefore, most of the criteria used for the selection of the single-hit model are also applicable to the log-probit, logit and multi-stage models. Thus, the only reason "remaining for the selection of the model is one of policy and the fact that the sin gle-hit model provides greater risk estimates than other models The selection of the extrapolation model to be used requires a great deal more attenti on and justification than has been given in this criterion documen t. The reasons for that will be obvious when one examines in deta il the consequences of selection of various extrapolation mod els. 4.0 CONCLUSION In summary, we feel the alternative approach suggested in our Specific Comments is a more reasoned approach to setting health effects criteria for pote tial carcinogens. The Agency's use of the one-hit model preclud ^s the use of any other model which may better fit the data. CMA urges strong consideration of our alternative approach. - 12 - CMA 049488 5.o literature cited Armitage, P. and R. Doll 1961. Stodastic models for carcino genesis. Proc. 4th Berkeley Symposium on Mathematical Statistics and Probability. 4:19-38 Univ. of California Press. Berkson, J. 1944. Application of the logistic function to bioassay. J. Am. Stat. Assoc. 39:357-365. Doll, R. 1971. The age distribution of cancer: implications for models of carcinogenesis. J. Prog. Stat. Soc. 13:133-166. Finney, D.Z. 1971. Probit Analysis. 3rd ed. Cambridge U. Press, 333 pp. lloel, D.G. e_t a_L. 1975. Estimation of risks of irreversible, delayed toxicity. J. Toxicol. & Environ. Health 1:133-151. Mantel, N. and W.R. Bryan. 1961. "Safety" testing of carcino genic agents. J. National Cancer Inst. 27:455-470. NAS. 1977. Drinking Wa :er and Health. Safe Drinking Water Committee. National Academy of Sciences, Washington, D.C. 939 pp. NAS/NRC. 1972. The effects on populations of exposure to low levels of ionizing radiation. (BEIR Report). Div..of Med. Sci. NAS/NRC, Washington, D.C. 217 pp. Ramsey, J.C. c_t a_l. 1979. Carcinogenic risk assessment: ethylene dibromide. Toxicol. & Appl. Pharmacol. 47:411-414. Weil, C.S. 1972. Statistics vs. safety factors and scientific judgment in the evaluation of safety for man. Toxicol. & Appl. Pharmacol. 21)454-463. - 13 CMA 049489