Document rpNzgk7x7X0DZmLVpov7vM3wa

,... D American Petroleum instiiute 1220 L Street, Northwest Washington, D.C. 20005 j~ Mary Burr PaJcton, Ph.D. Heath Sc!enllsl (202) SB2-8338 (202) 682-8270 (FAX) June 17, 1993 Vernon M. Chinchilli, Ph.D. Center for Biostatistics and Epidemiology College of Medicine Pennsylvania State University Hershey, PA 1'7033 Dear Vern: Enclosed at long- last is the analysis of the Pliofilm data set by Bob Spirtas and Charlie Brown. The Benzene Task Forcewould greatly appreciate your reviewing it incisively, on its -own merits and with respect to the analysi~ we did with ENVIRON. While waiting, I've been exploring why our results differ from Kenny Crump's new results (manuscript enclosed). As is laid out in the table of estimated additional leukemia deaths, the "discrepancies" of concern are: o ENVIRONS estimated risks are generally lower than Crump's; o contrary to expectations, for a given analysis, ENVIRON's estimates of risk using the Paustenbach exposure estimates are somewhat hiqhe:~7 than those~gotten using the Crump and Allen exposure estimates. As summarized in the next table, I've done a number of additional proportional hazards analyses on the Pliofilm data varying the criteria used to stratify the cohort. The first pag~ (A) is for runs with all 15 cases (i.e., including the female), as used in the ENVIRON-analysis; the second page (B) is for runs --on the "white mal-es" (non-female, non-black) only, which matches the data set Crump used. The stratification criteria applylng to a row are listed in the left-most box. The-three main column are for the cumulative exposure estimate used (Rinsky, Crump & Allen, or Paustenbach), each with a column for the slope (g), the standard error of the slope, and its p-value. The possible stratification criteria were: An equal opportunity employer BP-00019140 D- R/S: loc: DOS: DOB(o): DOB(s): race and sex (partitioned as black vs. non-black, female vs. non-female) plant location at St. Marys or Akron date of first exposure, with decades cut at 1940, etc. for St. Marys and at 1936, etc. for Akron date of birth, with decades cut offset to match decades for DOS by location date of birth, with decades cut the same regardless of location Al are the results in the second ENVIRON paper. B6 would be-most comparable to the data set used by Crump, but his model does use age-specific mortality rates i It will be noted (A vs. B) that the inclusion of the female leukemia case and her few stratum mates has virtually no impact on the estimates. If location is omitted as a stratification yariable (analyses 4, 5, and 6), the results with the Paustenbach exposures or with the Crump and Allen exposures are very sirmlar. ; On page 33 of his manuscript, Crump notes that our doseresponse model is exponential and "consequently, although approximately linear at low exposures, exhibits upward curvature. " He has asked whether the proportional hazardsanalysis can be run to make it a linear model directly comparable with his linear models. It can't be, can it? Crump's second point there is that, because date of first employment (DOS-) is likely to he correlated with benzene exposure, stratification on this variable would tend to dampen the measured effect of benzene exposure. The findings for the sets of stratification criteria without DOS wou~d tend to support this assertion in that they are all more significant and have steeper Bs-than those for Al reported in our paper no matter which set of exposure estimates is used. Beeause of thepronounce elevation of the p-values, I'm inclined :to believe that the model in #1 is over-parameterized by including both DOS and location. We were endeavoring to mimic Rinsky_Js matching criteria, but I now think #3 (location without DOS) might have been a better choice. Knowing what we do about the marked qualitative difference between St. Marys and Akron, the stratification on location is desirable, -but Crump may have- a point in asserting that DOS is too c-losely related to exposure. The correlations between year of first work~nd the cumulative exposure measures (see computer pTintouts) do turn out to be significant, particularly for Akron; interestingly, in St. Marys the cumulative .Rinsky exposure estimates- do not show such a correlation. I have run some correlations on the va-rious measures sf exposure with~ each other with wetside-dryside and "white mal-e" BP-00019141 partitions, which I have also included for your perusal. ~s my notations indicated, the perfect explanation of our inverted order of potencies for Crump & Allen vs. Paustenbach exposures did not jump out at me. Perhaps the numbers will say more to you. As I noted in the first of the ENVIRON papers in discussing this- apparent inversion (see page 10; page 16, footnote 10; and page 24, Table VII), the mean increase in the workers' exposure estimates with the Paustenbach estimates over those with the Crump and Allen estimates was relatively greater in the non-cases than it was in the cases. This is further documented in the two attached printouts, which show this again and that this had not been the case for the Crump and Allen estimates vs. the Rinsky estimates. Increase over lower exposure estimates Non-cases Crump and Allen vs. Rinsky 111%. 152% Paustenbach vs Crump and Allen 89% 39% I've scrutinized the estimates of cumulative exposure and the differences between those obtained using the Paustenbach estimates and those gotten using the Crump and Allen estimates on an individual worker basis. In several hundred instances, the Paustenbach estimate of cumulative exposure is less than the Crump and Allen estimates (as is the case for five of the 15 leukemia cases). In 53 cases, this deficit in the Paustenbach estimate is more than 100 ppm-years; in five more, it exceeds 1000 ppm-years. Referring back to the exposure matrices themselves, it is clear that, although overall the Paustenbach estimates are higher, on a job- and year-specific basis there are numerous exceptions to this trend. As is borne out in practice, there is no reason to expect that an individual's cumulative exposure derived using the Paustenbach exposures will be higher than that derived using the Crump and Allen estimates. The precise potency estimates gotten by fitting a particular model will depend on the model itself and the exact set of cas.es and non-cases included in the analysis. Although the basic data set from which ENVIRON and Crump worked were identical, they did not use precisely the same data sets in their actual analyses. All of Crump's analyses employed every "white males. The ENVIRON proportional hazards analyses included the female cases and her stratum mates (although as demonstrated above, the small group-of females had virtually no impact on the outcome). More importantly, the exact set of individuals included in a run of ou~ proportional hazards model is determined by the stratification criteria. Because a non-case will contribute information only if he falls in the same stratum with a case who died younger than-he did, as the stratification criteria are reduced, two factors work to include more non-cases in the analysis. First, less finely defined strata will contain more non-cases if strata without a case are merged with ones that do have a case. Second, more non-cases in a stratum with a case will be eligible to contribute information if their stratum is BP-00019142 merged with a stratum containing a case who died at a younger age. Increasing the number of individuals included in an analysis will increase its power to generate significant findings. I next intend to--s-ee whether the use of intermediate cumulative exposures (for controls at the age corresponding to a case's death) in the proportional hazards calculations might account for the differences we had from Crump's results using always the lifetime c~~ulative exposure. Given that most of the cases died at "retirement ages," however, I wonder if this could be the key. With respect to the inversion in results seen between us and Crump for ~he Crump and Allen exposure estimates and the Paustenbach exposure estimates, I feel more comfortable that the differences are a function of our models and not an indication that someone has done something wrong. I'll call you early next week to see what you think we should do about the current version of our the second paper with Risk Analysis. Maybe it would be easier to get together to work out the fine points. Please keep track of your hours. I got your last bill and entered it into the payment system. We appreciate your efforts. Sincerely, enc: Crump vs. ENVIRON pb table + comments Spirtas manu Crump manu startyr corr expo measure corr diff dist Mary Burr Paxton, Ph.D. Health Scientist BP-00019143 COMPARISON OF RISK ESTIMATES FRO:OO: CRUMP AND ENVIRON ANALYSES Additional leukemia deaths per 1000 workers with 45 years occupational exposure to 1 ppm benzene Model Crump: Weighted dose, non-linear outside Additive Multiplicative Paxton et al. : Proportional hazards stratified on decade of birth, plant location, race, sex Decade of starting Pliofilm work Without stratification on start date Exposure Estimates Crump & Allen Paustenbach 1.9 - 0. 72 2.1 0.56 0.26 0.47 0.49 0.64 .-----~- ----..... :- -:=--:. BP-00019144 IMPACT OF STRATIFICATION ON PROPORTIONAL HAZARDS ANALYSES A. Including Female Case c Sr ti rt ae tr ! ai a 1. R/s loc DOB(o) DOS 2. R/S loc DOB{o) 3. R/S loc DOB{s) I ~- R/S DOB{s) 5. R/S DOS I Rinsky I I Exposu~e Estimates Used in Analysis Crump & A;Llen I Paustenbach fS 0.00376 SE . 0,001;36 :p-value 0.0023 g 0.00075 SE 0.00042 p-value 0.0602 fS 0.00150 SE 0.00051 pvalue 0.0015 0.00439 0.00122 0.0001 0.00512 0. 00132 0. 00()1 I 0.00112 0.00038 O.Q007 I 0.00136 0.00043 0.0003 I 0.00171 0.00046 0.0001 0.00199 0.00050 0.0001 0.00488 0.00132 0. 0001 0.00441 0.00134 0.0002 0.00146 0.00042 0.0001 0.00160 0.00054 0.0011 0.00172 I 0.00165 0.00045 I 0.00047 0.0001 0.0001 6. R/S 0.00521 0.00122 O.Ob01 0.00167 0.00037 0.0001 0.00177 0.00039 0.0001 - '-' \ ;I ttl '"0 I 0 0 0...... \...0... .j::o. (Jl IMPACT OF STRATXFXCATXON ON PROPORTIONAL HAZARDS ANALYSES B. "White Males" Only c Sr ti rt ae tr aI i a 1. R/S loc I DOB(o) DOS :2. R/S loc DOB(o) 3. R/S loc DOB(s) 4.. R/S DOB(s) I g Rinsky Exposure Estimates U~ed in Analysis Crump & Allen Paustenbach SE p-value g SE p-value g SE pvalue 0.00376 0.00136 0.0023 0.00075 0.00042 0.0607 0.00148 0.00051 0.0018 I 0.00439 0.001.22 0.0001 .I 0.00112 O.P0038 0.0007 0.00170 0.00047 0.0001 0.00512 0. 00132 0.0001 0.00136 0.00043 0.0003 0.00197 0.00050 0.0001 I 0.00489 0. 00132 0.0001 0.00146 0.00043 0.0001 0.00170 0.00046 0.0001 5. R/S DOS 0.00442 0.00134 0.0002 I 0.00160 0.00054 0.0010 0.00162 0.00047 0.0001 6. R/S 0.00521 0.00122 0.0001 0.00167 0.00037 0.0001 I 0.00175 0.00039 o.pOOl H '''~' ttl '"0 I 0 0 0...... \...0... +-- 0\