Document XOxgqv9B28QXQg2R6k45Do1Zd

Jun 02 OS 04:lOp P* 1 Exposure Assessment for a Study of Workers Exposed to Acrylonitrile. Ill: Evaluation of Exposure Assessment Methods Patricia Ann Stewart*'*John N. Zey,B Richard HomungJ3, Robert F. Herrick,nc Mustafa Dosemeri,A Dennis Zaebstf and Linda M. Pattern*^ ^-Environmental Epidemiology Branch, National Cancer Institute, Executive Plaza North, Room 418, Rockville, Maryland 20892; ^Notional Institute for Occupational Safety and Health, 4676 Columbia Porkwoy, Cincinnoti, Ohio 46226; ^Current address; Harvard University, School of Public Health. Department of Environmental Health, 665 Huntington Avenue, Boston, Massachusetts 02115-6069; "Current address; Nolionol Institutes of Heolth, Women's Heolth Imtiotive, 7550 Wisconsin Avenue, Bethesda, Maryland 20892; To whom correspondence should be addressed Retrospective exposure assessment in epidemiologic Studies is depen dent on the availability ofhistorical monitoring results, yet rarely arv there sufficient results to rely upon them exclusively. An approach h described that has a formal structure for developing exposure esti mates for an epidemiologic study using a variety ofmethods depend ing on the information available. The approach identifies criteria for determining what data are needed for each method and the hierarchy for using the methods. The estimation methods include; (1) calculat ing a mean from the monitoring results of ajob: (2) identifying ho mogeneously exposedjobs and using the mean ofthe measurements for the jobs as the estimate; (3) applying a ratio of the measurement means oftwojobs in one operation to a thirdjob in another operation to estimate a fourth; (4) weighting by rime various areas or personal (short or full-shift) measurements representing tasks or locations; <5) taking a deterministic approach that modifies a more recent exposure by estimates of exposure modifiers to reflect how changes in the workplace affected exposures; and (6) using professional judgment. The homogeneous exposure group, ratio, rime-weighted, and deter ministic approaches were evaluated for bias, precision, accuracy, and correlation. First, a subset of the monitoring results was removed from the entire data set for use as referent values. Estimates were developed using the four methods without the removed data and compared with the referent value. On average, the estimates tended to overestimate the measurements in the racio method (bi as -- 77%) and underestimate them in the time-weighted method (--24%). The average difference between the means and the esti mates using he homogeneous exposure group method and the deterministic method w zero. The imprecision was threefold to fourfold for the ratio and deterministic methods and 1,5fo)d for the rime-weighted and homogeneous exposure group methods. Cor relations between the estimates and the measurement means ranged between 0.60 and 0.75 and were statistically significant. When the methods were evaluated using only the same cells, the homogeneous exposure group method performed best, followed by the deterministic method and then the ratio method. The rimeweighted average method could not be evaluated because of the lack ofmeasurement dan for these cells. 1996 A1H. Stewart, P.a,; Zey, J.N.; HOrnunG, R.; Herrick, R.F.; DosEmKi, M.; ZaST, 0.; Potterm, l~M.: Exposure Assessment for a Stujy of Workers Exposed to Acryionitriie. 111. Evaiuatton of Exposure Assessment Mmoos. apfi. Occur. Environ. Myc, 11(11); 1312-1321; 1996. uantitative retrospective exposure assessment is a difficult Qtask in occupational epidemiologic studies, because rarely are sufficient monitoring results available for all jobs over all the years in the study. Various approaches have been taken to estimate exposures, but only two, to our knowledge, have compared quantitative exposure estimates to measurement data.'1--? The failure to conduct such an evaluation in other studies is not surprising, however, because if sufficient data were available to hilly validate the exposure estimates, the estimates themselves would be unnecessary. Nonetheless, val idation of estimates is crucial to the field. We are engaged in a cohort mortality study of workers exposed to acrylonitrile (AN),'1'4? The cohort has been em ployed in eight plants that started manufacturing AN, acryl amide. and acrylic fibers or resins between 1952 and 1965Each of the companies conducted air monitoring since the late 1971 Is, but only one monitored before this time. Even after 1977. a large number ofjobs were not monitored (about 3200) compared with the number ofjobs in the study (3500). Thus, u was necessary to develop a large number of estimates both for the unmonuored jobs and the unmomtored time periods to allow the inclusion of all study subjects in the mortality anal ysis. Tills report describes the methods used to estimate his torical exposures and an evaluation of cheir performance. Methods All companies in the study started conducting personal mon itoring in the late 1970s, Accompanying documentation main tained by the plant generally included job title, department, date, duration of the sample, an evaluation of the sample's representativeness of typical exposures, and type of sample (area or personal and time-weighted average or short-term), but not all plants retained the same information. All data were computerized (either by the company or by the study inves tigators) and put into a standard format. There were about 18,000 monitoring results. The quantity ol monitoring data available on the jobs in the study varied 1312 APPL.OCCUP.ENVIRON.HYr,. 11(11) - NOVEMBER 1*1% 1ii47-322X/%/1 111-U12S15,<K!/Ki C 1W6 ASH i'll 5l(47-J22X(y*)002a3-C .Tim HP PIR fl4 i 1 flp P.p APW-OCCUt*. ENVIR0N.HYC. 1I(J) NOVEMBER 1996 Study of Twfcm Exponed to Acryldnumle 1313 considerably over time and among plants. For some of thejobs directly involved in making the product, there were over 100 foil-shift, personal monitoring results between 1977 and 1983 (the close of foe collection of work history records). In con trast, a few jobs supporting the production operations (e-g., maintenance, quality control, the materials handlers for the AN monomer, environmental control, utilities, engineering, and research and development) had a variable number of results, depending on thejob, the plant, and the exposure level. Administrative, non-AN-cxposed, and other support opera tions (c.g,, the materials handling department not handling the monomer) were rarely monitored- Area measurements were available in three plants, but in only one of these (a fiber plant) had they been collected prior to 1977. The approximately 4(MX) area measurements in this plant were short-duration samples and existed back to 1963 (start-up of the plant was in 1958). The area measurements collected in the other two plants included both shore-duration and full-shift results. Prior to determining what estimation procedure should be used, a review of the published exposure assessment literature was conducted. Because so many jobs were missing measure ments, calculation of mean exposures for all jobs was not possible (e.g.. Reference 5). Grouping jobs into a small num ber of occupational groups so that mean exposures could be calculated for all groups/** derivations of estimates using area measurements weighted by rime.f7,8* or use of statistical modcl$<1,8* was not possible prior to 1977 in seven of the plants because processes and controls had changed over the years and measurements for these conditions were nonexistent. In addi tion, the fiber plant that had collected earlier area measure ments was substantially different from the other two fiber Operations. It was a continuous wet operation, whereas the other two dried the polymer before making the fiber. Thus, extrapolation to the othet two plants was inappropriate. There were also no other parallel chemicals used in the processes that had been measured and could be used to estimate exposure levels.w Moreover, no single estimation approach could be identified that would produce estimates for all cells. As a result, it was decided to use a variety of methods, the selection of the mechod for any particular job/year combination being deter mined by the available information, by defined criteria, and by how well each method performed in che validation exercise where the predicted exposures were compared with the actual measurements (described below). The original intent of the investigators was to develop an estimate for each job/department/plant combination for every year. Many of the years, however, had so few monitoring results that it was decided to combine years, where appropri ate, to form time period cells. Information was collected from interviews with long-term employees at each of the plants on when major changes occurred that were likely to have affected exposure levels in the workplace. The dates of these changes provided boundaries around rime periods, which were made up of years in which exposure levels were likely to have been similar. The periods may have been composed of a single year or may have included several years, and they established che unit oflime for which the exposure estimates were developed. Performing a retrospective exposure assessment usually has two components due to the availability ofmonitoring data and supporting documentation. First, because it is rare that all the jobs have been monitored in a company, the industrial hy gienist must develop at least one exposure estimate (generally the most recent) for each ofthe jobs in the study. The first step in this study, therefore, was to develop baseline exposure estimates for all cells from monitoring results wherever possi ble. Cells for which means could not be calculated were evaluated to determine if they could be estimated using the homogeneous exposure group (HEG), ratio, or time-weighted average (TWA) method. For all remainingjobs that still had no estimate, professional judgment was used to assign an estimate for a single cel). The second component in exposure assessment is to develop estimates through rime for the remaining empty cells. In this study this was done by modifying the baseline estimates, using estimates of exposure modifiers to complete all remaining empty cells. This was called the deterministic method- All estimates in this study were plant/departmcnt/ job/rime period specific. Development of Baseline Estimates Moon of the Monitoring Measurements The best estimates of the true exposures experienced by the study subjects when cumulative exposure is the primary ex posure measure of interest arc the arithmetic means of the measurements.(,,,) Arithmetic means were therefore calculated in this study. Means based on small numbers, however, can be easily affected by a few outliers that weigh more heavily in the calculation of the mean than appropriate, Criteria were there fore needed that ensured, to the extent possible, that the means reflected true exposures. Three variables were evaluated that could have influenced the accuracy of the mean estimates: the duration of the monitoring result, the representativeness of the monitoring result to typical exposures (as indicated in the company records), and the number of monitoring results used in the calculation of the mean. Sample duration was identified in the documentation ofthe measurement data in five plants. Documentation from two other plants indicated whether the result represented an 8-hour TWA. a ceiling, or a peak sample. Personnel from the eighth company indicated to the study investigators that generally full-shift samples were taken; therefore, al! results obtained from this company were assumed to be full-shift in duration. Over 80 percent of all measurements were greater than 6 hours in duration. Because the less than 4-hour samples were generally task oriented rather than full-shift oriented, these measurements were excluded. The 4- to 6-hour measurements appeared to be a mixture of both full-shift and task measure ments. These results were also excluded, however, because they contributed only slightly to the database (8%). A mini mum 6-hour duration was therefore selected to ensure that the measurements were representative of foil-shift exposures. (Shorter duration measurements were available using ocher methods; see below.) Many of the companies designated the measurements as typical or atypical ofnormal operating conditions. Exclusion of che atypical measurements was not considered appropriate because they were actual measured exposures. Including them with measurements taken under typical conditions was also troublesome, because in calculating a mean, they might be given more weight than appropriate. For example, ifa job was jun us us u*9; i up P* 1314 FA Suwui a J, APPL.OCCUP. ENVtRON.HYG. I1(U) NOVEMBER 1996 monitored every day during a shutdown operation lasting 2 weeks, and was otherwise monitored monthly, the measure ments taken over the 14 days ofthe shutdown would contrib ute 54 percent of the total (14/14+12) measurements in a year, which is considerably more than the true contribution of these measurements (14/365 days or 4%) to an annual average. An examination of the measurements found that fewer than 20 percent of the measurements were taken under unusual conditions. Ofthese, 60 percent were taken during shutdowns, start-ups, or major repair of the operating units, events that usually occurred once a year or less. The frequency ofsome of these events could be estimated, so that they were weighted in the calculation of the mean according to their frequency of occurrence over a year (e.g.. 4% in the above example). When frequency could not be determined (e.g., spills, upsets, no information), the results were given the same weight as che typical measurements. Although not the perfect solution, chis was considered acceptable because these other measurements comprised only about 5 percent of the measurements used m the analysis. Finally, the minimum number ofmeasurements necessary to develop a cell mean was evaluated by looking at the coefficient ofvariation (CV). CVs were calculated for all cells with at least five measurements and then pooled to derive an average CV. This calculation was also done for cells with at least six, seven, eight, nine, and ten measurements. The mean CVs changed very Utdc (less than 30% from the lowest to the highest CV) regardless of the number of measurements. Six was therefore arbitrarily selected as the minimum number of measurements required for a Cell. (Shorter duration measurements were avail able for other methods; see below.) Ratio Method In this method, the assumption was made that similar jobs in similar environments in different plants are likely to have similar exposure levels relative to other jobs in the Same environment. This assumption appears to have validity based on work by Eiscn ei alTu> In their report on granite workers they presented exposure data from seven granite sheds. For each shed, the exposure estimates of six jobs were presented. Although the particular environmental conditions were not presented in the report, it appears from those data that the relationship of exposure levels between two jobs in one shed approximated that of the samejobs in another shed, despite the wide range of exposure levels among sheds. That study pro vides che underlying assumption that (EMONi,')(EMONi-|) ----------------- ------------- L (I) where; EftiOt'r * the exposure estimate derived from the ratio method for job i'j' f|JON * the estimate derived from the mean of the monitoring measurements method jobs I *= operations or plants The criteria for selecting the jobs for this method were as follow*; (1) the tasks being performed by job, in planes j and j' were similar, as were those by job,.; (2) the amounts of time spent by job, performing those tasks were similar in plants j and j', as were those ofjob,.; (3) the sources of exposure for the two jobs (i and i') in plants j and j* were Similar; and (4) the controls influencing those sources were similar. Because of the requirement of similar controls, the estimate for job, and job/ in plant were from the same time period, and those ofjob, and job,, in planty were from the same time period, but the estimates for plants j and j' did not have to be from the same time period. HEG Method Another approach used to estimate exposure was based on che concept of HGs3,2> Jobs identified as being similar were grouped, and the measurements for those jobs were used to calculate a mean value. The criteria for grouping these jobs were that; (1) they had to have the same sources of exposure (i.e., they Were in the same operating unit and plane); (1) the tasks being performed were similar; (3) the time spent per forming those tasks was similar; (4) the controls influencing those sources were the same (resulting in all the measurements being in the same time period); and (5) there had to be at least six measurements of at least 6 hours in duration on the jobs within the group. Time-Weighted Average {TWA} In this method, a TWA was calculated from short-duration or area measurements or from other job estimates and weighted by the time spent at those concentrations by C,c, + C3t3... + Cnt,, Etwa t, + tj .., t,, (2) where: Etwa. = the estimate derived from the TWA method C; the concentration at location, or performing task, t, ** the minutes the job spent ot location, or tosk, Three different sources of exposure concentrations (C.) were available; (1) a concentration entered by the industrial hygien ist based on professional judgment, which was usually 0.00, designating the exposure level experienced when working in a nonexposed office; (2) a mean of area samples or a mean based on less than six personal samples or on samples that were less than 6 hours in duration (These data had been excluded from the mean of the monitoring measurements method); and (3) a previously developed estimate for another job or a mean of other jobs* estimates. For example, an engineer who spent I hour a day in an operating unit was given the same concen tration for chat hour as an engineer who spent 8 hours a day in the unit. The time estimates were either specifically identified from interviews or they were estimated by the study investi gators based on an understanding of the tasks, Estimates derived from these four methods were considered to be of higher confidence, and these methods were used to develop the estimates wherever possible, for the jobs to which these methods could not be applied, a single estimate was developed, usually the most recent, using one of two profes sional judgment techniques: Jun n? flR 04: 1 Ip p- 4 APPL.OCCUP. liNVIRON.HYG. 11(11) NOVEMBER. 19% Smdy of Worker* Export! p> AcryWoilc 13t 3 TABUE 1. Exomples of Significant Changes ond Estimates of Thoir Sizn ond Effect Plant Typo Fiber Fiber Fiber Monomer Monomer Monomer Monomer Resin AD All An AH Change Installed vena on reactors Quality control sample points enclosed in ventilated sample boxes Pilot plant dosed; quality control samples no longer analyzed Began dosed dome loading of tank ears Single seals replaced by double seals/candem seals Routing roofs installed at tank farm Oil added to seating ponds to reduce evaporation Increased preventative maintenance AN Continuous analyzer installed New OSHA standard became effective Hoods added to laboratory Air conditioning added to control room Eitgr Estimate* Size* Effect15 M M L M M/L S S M M M M/L M 1-1.5 10-100 0.9-0.9 2-10 2-5 l-l.l 1.1-3 1-1.3 1.1-2 2-5 2-10 1-5 TH Estimate Size Effect 30 1.3 30 55 30 0.9 30 6 40 3.5 5 1.1 52 20 i.i U) 1.5 10 3,5 20 6 30 3 ATheic represent the estimates developed by the engineer of the uujy. They were used is guidelines for the utduvtn.il hypcniu. lvThese represent the esnmites issigned by the industrial hygienist of the study. They generally represented the midpomc of the engineer's estimate, cPereent of a job's total exposure that the source being affected contributed n> pnor to the change: S = small (1 to 10%): M " medium () 1 to 49%) J. lJrve (50 to 100W), eoneetion factor designating the reduction or increase that the change had on the emissions being released from the sour, e. For example. 1 would indicate no effect and 1.5 would indicate that emissions from a source were 1.5 times higher before the change was in place. Department-Wide Method This method allowed the industrial hygienist to enter one exposure estimate for all jobs in a department. It was generally used for nonexposed departments, such as administration, for example, where 0.00 was entered. Profeaional Judgment In this method the industrial hygienist developed an estimate based on: (1) a mean of personal measurements where fewer than six existed or where the duration of the samples was less than 6 hours; (2) a mean of area measurements; (3) a mean of measurements after 1983; and/or (4) anecdotal information, such as frequency of health effects, odor, task descriptions, etc. Development of Estimates Over Time After all jobs had at least one baseline estimate, all remaining cells were estimated using the deterministic method, which modified baseline estimates based on estimates of the effect that the major changes that occurred in the workplace, or exposure modifiers/1had on exposure levels. Three variables were identified as having an important effect on AN exposure levels: changes in production ratcs.t'1,) changes in the frequency of exposure, and changes in the operation and in engineering controls. Annual production rates were received from the companies in the study, but because data were missing for several years, production races were not included in the model. The frequency of exposure was obtained from the interviews or estimated. For example, production employees usually were given daily exposure, whereas an engineer may have been exposed only 1 day a week. Estimating the effect of the changes in the operation or in engineering controls w a more difficult task. For many changes there were no monitoring data available pnor to and after implementation of the change. A chemical engineer with experience in these types of plants, therefore, compiled mon itoring data, information from the interviews, information from other plants in the study, and data from the published literature. Using these sources and his own experience, he estimated the effect the changes had on exposure levels for two variables; sir.c and effect. Sire was defined as the percent ofthe job's total exposure that the source being affected contributed to pnor to the change, and was indicated as small (1 to 10%), medium (11 to 49%), or large (50 to 100%). Effect was a correction factor indicating the reduction or increase that the change had on the emissions being released from the source, and was generally indicated as a range. Unless information was available that suggested otherwise, the midpoint of the ranges for both uze and effect was used, Table 1 identifies examples of several changes ansi their estimated size and effect. To develop an estimate, the baseline value for an adjacent (generally more recent) time period was used to derive an estimate for an earlier time period as follows: .i 53 [IV0 ~ 2 S,, + 2 S,,/CF,,)](F(y. ,/Fv) (3) where: Edet = the estimate being derived from the deterministic method 8 " the baseline estimate S the size of the source being affected by change, CF,, = the correction factor or effect of change, f = the frequency of exposure {1.0 - 4 to 5 days/week, 0,4 = 2 to 3 days/week, and 0.1 = <2 days/week) y "i the timo period for the baseline value y - 1 = the time period for the estimate (Sec Appendix for a derivation ofthe equation.) Thus, suppose in 1977 the exposure estimate was 2 ppm, and two changes took place that reduced the exposure between 1977 and 1976, The size and effect of these two changes were 10 and 3.5 for .11J rs nn n4 : i ip? p_ T 1316 P.A- Stewart ct al APPL.OCCUP. F.NVIRON.HYG. 11(11) NOVEMBER im one change and 20 and 1.2 for the Other, respectively. The frequency of exposure was daily in both years: Edet - [2/(1 - 0.10 + 0.10/3.5 - 0.20 + 0.20/1.2)] (1/1) - (2/0-90)0) = 2.22 ppm The Eftirr then became the baseline estimate for the next adjacent time period. In addition to this estimate, maximum and minimum estimates were calculated using the extremes of the size and effect values developed by the chemical engineer. These additional estimates provided an estimate of uncertainty about the estimates. Other Estimates The estimates described thus far were estimates of air concen trations. These are typically the principal estimates used in epidemiologic analyses. In some instances, however, these may not be the best estimates to evaluate disease risks. For example, in some jobs respirator use was mandatory, suggesting that an air concentration may not be the best exposure estimate. Exposure estimates were therefore derived that accounted for respiratory protection when respirator use was mandatory for 8 hours a day (Ekes). For thesejobs, a protection factor (PF) was applied to the air concentration estimate () based on Che type of respirator worn (half-mask: PF = 10; full-facepiece: PF = 50; supplied air; PF * 2000) and multiplied by 0.65<lSl to allow for imperfect protection, chat is, Erb - E/PF(0.65) <4> Because the amount of an airborne chemical received by the body depends on the amount inhaled, a third estimate was derived that took into account respirarory rates based on low, medium, and high levels of physical activity. These levels were defined to correspond to ventilation volumes of 750, H5<>, and 2150 cm5/breath, respectively, assuming 15 breaths per minute at all levels.'1^ An adjusted exposure estimate for a job was therefore calculated by multiplying the estimate of the air concentration derived for the job by the appropriate ventila tion volume and the number of breaths taken in 8 hours. Finally, AN can be absorbed dcrmally.<,7) Because the ab sorption rate in humans was not found in the literature, a dermal exposure score was calculated by multiplying the fre quency of dermal exposure, arbitrarily selected as 5 for fre quent (more chan once a day) and 1 for infrequent (less than once a day) by the concentration of AN in the liquid. Method Evaluation Evaluation as to how well the estimation methods performed is crucial for the interpretation of an epidemiologic study. In this study, however, there were no full-shift, personal moni toring data prior to 1977 with which to evaluate how well the assessment methods performed during the earlier years of in terest, There were, however, monitoring data available after 1983, the last year for which estimates were developed. In some plants, up to 10 years of monitoring data were available. It was decided to test the estimation methods using these data where possible. Four methods could be tested: the ratio, the HEG. the TWA, and the deterministic methods. To test these methods, a subset of the cells for which there were monitoring results was removed from the estimation process. The assessment methods were then applied to the remaining cells to develop estimates, which were compared with the corresponding mea surement means. All possible job/time period cells were used for all estimates combined and for type of operation (fiber, monomer, and resin). In the ratio method all sets offour jobs in the ratio that met the criteria listed Tot this method were identified. One of the four jobs was selected randomly to be the estimated job. The means of the measurements for the remaining three jobs were used to calculate the estimate for the fourth, using Equation 1. The estimate was compared with the mean of the measure ments of that job. All possible combinations of four jobs svere used (i.e., a single job may have been used in comparison with several different jobs). In the HEG method, at least three jobs, each with measure ments, were identified as being in an HEG using the criteria described eariier ft>T this method. One job was randomly selected from the jobs in the HEG and removed. The mean of the remaining measurements was the estimate and the mean value of the measurements of the removed job was the referent value. For the TWA method, jobs with mean of the monitoring measurements estimates were identified. Wherever possible, estimates were developed for these cells using area measure ments, short-duration measurements, or estimates from other jobs, weighted by time. These estimates were compared with the mean ofthe monitoring measurements values. In the deterministic method, all jobs with at least two estimates calculated from the mean of the monitoring mea surements method were identified. Frequency ofexposure was not relevant here because it designated number ofdays exposed per week, which is not relevant to a full-shift measurement. The estimate of the job in the most recent time period (e.g., 1987) was modified using the estimates of size and effect generated by the chemical engineer to derive an estimate for each time period hack to 1977. The estimates were compared with the referem values (i.c,, the means of the measurements) when they existed. Statistical Methods Bias, precision, and accuracy were calculated for each meth od,i'"> Bias is the average difference between the measurement means and the estimates. Precision (a more tncuirive word would be imprecision, and therefore this term is used here) is the standard deviation of those differences. Accuracy is the square root of the sum of the bias squared and the imprecision squared. The closer these three values are to zero, the better the estimates are. Relative bias, imprecision, and accuracy were calculated by dividing each of these values by the mea surement means. To determine how well the ranking of the estimates compared with the ranking of the referent values, Spearman rank correlation coefficients were calculated. All possible estimates for each ofthe methods were included in the analysis to sec how well the methods performed overall. To | . ; . ' I .Tun O? HR 114 1 1 p p. R APPL-OCCUP. ENVmON-HYC. 11(11) NOVEMBER 2996 "*" Study of Workers to Acrylonitrile 131/ TABLE 2. Bko, Irepreoaion, Accunxy, and Correlation of Three bcpojwe A.T^.,,lnt j?mA SEB j?.c Bia, Rfl fiiu tt (ppm) <PPm) (ppm) (ppm) <%> Ratio 51 0,93 0.14 1.64 0.71 77 HEG 86 0.72 0.14 0-73 <0,01 0 TWA 32 0.62 0.02 0.47 -0.15 -24 Deterministic 177 1.34 0.25 1.33 0.01 1 ^Arithmetic mean of the measurement*. "'Standard error. of the estimates. l`Speanan rank correlation coefficient. *p < 0.001. Fp < 0,005. Imp (ppm) 3.82 1.08 1,03 3.17 Rd Imp <%> 412 149 166 236 Ace (ppm) 3,88 J .OR l.CH 3.17 Rfl Acc <%\ 419 149 168 TV, *** 0.76 0.65" 0.59* 0,58E establish which of the three methods was preferable when the data allowed selection ofmore than one method, only cells for which estimates were derived for the ratio. HEG, and deter ministic methods were included in the calculation of the statistics. None of these cells could be derived from the TWA method, so this method was not evaluated. Results A random sample of one job from each of 51 secs of four jobs was used for evaluating the ratio method (Tabic 2). The overall mean of the Cells based on the mean of monitoring results was 0.93 ppm. whereas the overall mean of the estimates for those same cells was 1.64 ppm. The estimates therefore overesti mated the measurements by 0-71 ppm or about 80 percent. The relative imprecision was about 400 percent. Overall, the estimates were accurare to within about four times the mea surement means. The Spearman correlation was 0.76. which was statistically significant (p < 0.001). The average of the mean ofthe measurements in the 86 cells in the HEG method was 0.72 ppm, with the mean ofthe estimates being 0,73 ppm, The relative bias was zero. The imprecision was 150 percent and overall accuracy was 150 percent. The Spearman coeffi cient was statistically significant at 0.65 (p < 0.001). The TWA method had overall bias of25 percent, the relative imprecision was about 165 percent, and the relative accuracy was about 170 percent. The correlation was 0.59 (p < 0-001). The deterministic method had low bias (mean of the measure ments * 1.34 ppm; mean ofthe estimates " 1.33 ppm; relative bias = 1%). The imprecision and overall accuracy were both about 240 percent and the correlation was 0.58 (p < 0.001). The methods were examined by type of operation. The mean of the estimates using the ratio method (Table 3) was higher than the mean of the monitoring results (2.% versus 1.48 ppm. respectively) in the fiber operations. This resulted in a relative bias of 100 percent and a relative imprecision of 380 percent. Overall accuracy was about 400 percent and the Correlation coefficient was 0.87 (p < 0,001). The estimates for the monomer operations were only slightly higher than the monitoring results (means = 0,56 and 0.48 ppm, respectively). The relative bias was 18 percent, relative imprecision 139 percent, and relative accuracy 140 percent. The correlation was lower than for the fiber operations (monomer r = 0.48). although still significant (p < 0.01). Estimates could not be developed using this method for the resin operation because there was no comparable operation to use in the ratio. The results of the HEG method by cype of operation are found in Table 4, The means of the estimates for the fiber, monomer, and resin plants were fairly close to the means of the measurements (relative biases of--9. 25. and 12%, respective ly). The imprecision and accuracy were 220 percent or less. The correlation between the measurements and the estimates was excellent to moderate for the fiber and monomer plants (r " 0,80 and 0.47. p < 0.01, respectively), but poor for the resin plant (0,31. p > 0.05). This low correlation may be due to the narrow range of exposures in this piant (mean = 0-23, S = 0,00). The TWA method underestimated the measurements by 20 and by 56 percent when k was evaluated for the fiber and monomer operations, respectively (Table 5). There were too few measurements (n 31 2) for the resin plant to evaluate the method. Relative imprecision and accuracy for the two oper ations was about 75 and 125 percent for the two operations, respectively. The correlations were 0.57 (p < 0.01) and 0.43 (p > o.o5). TABLE 3. Bios, Imprecision, ond Accuracy of the Ratio Method by Type of Occupation SE Bias Rel Bias Imp n (ppm) (ppm) (ppm) (ppm) <%> (ppm) fiber 23 1.48 0.26 2.96 1.48 ion 5.61 Monomer 28 0.48 0,07 0.56 0.09 18 0.66 Abbreviations art explained in Table 2. *p < 0.001. "p c 0.01. Re! Imp (%> 380 139 Acc (ppm) 5.80 0,67 Rel Acc <*/.) 393 140 r 0,87A 0.48 Jun [IF UH U*ii 1 Fp p. V 131# P.A. Stewart il. APPL.OCCUP. EKVtftON.HYC. 11(11) NOVEMBER 1996 TABLE 4, &ien, Imprecision, Aeturocy, ond Correlation of the HG Method by Typo of Operation Plant *- SE *,, Bias Rel Bias Imp (ppm) (ppm) (PP) (ppm) <%> (ppm) Fiber 36 1.22 0.30 1.11 --0.11 Monomer 33 0.44 0.12 0.55 0.11 Resin 17 0.23 0.06 0.26 0.03 -9 25 12 138 0.97 0.30 Abbreviations are explained in Table 2. *p < 0,001. "p < 0.01. Re) Imp <%) 114 219 128 Ace (ppm) 1.39 0.97 0.30 Rd Ace <%> 114 221 129 r 0.80* 0.47*' 031 j I Tbc estimates for the deterministic method 'were, on aver age, higher than the measurements in the fiber and resin operations, whereas they were lower than the measurements for the monomer plants (Table 6). The magnitude of the relative bias, on the other hand, was similar for the fiber and monomer plants (18 and --32%, respectively), but large for the resin plant (170%), Relative imprecision and relative accuracy were over twice as high for the monomer and resin plants as compared with the fiber plants. The correlation in all opera tions was moderate (r -- 0.38 to 0,58). To evaluate what hierarchy should be used when more than one method was possible, only the cells for which estimates were developed using all methods were evaluated (Table 7), There were 32 estimates that the ratio, the HEG, and the deterministic methods had in common, but none of these could be developed using the TWA method. The ratio method developed estimates that were much higher than tbc measurements (2.70 versus 1.57 ppm, respectively). The rela tive bias was 7J percent, the relative imprecision was 322 percent, and the relative accuracy was 330 percent. The cor relation coefficient was 0.63 (p < 0.001). The HEG and the deterministic methods also overestimated the measurements, but by much less, with relative biases of 6 and 32 percent, respectively. The relative impredsions and accuracies of these two methods were about and 90 and 120 percent, respectively, and the correlations were moderate (0.65, p < 0,001, and 0.54. p < 0.01, respectively). Discussion This report describes the exposure assessment methods used in an epidemiologic study to assess historical exposures when measurement data were not available. Because measurement data varied in quantity by job and by year, several methods were developed to allow the best use ofthe data. The methods were then evaluated to ensure that they developed reasonable estimates. Most of the methods performed acceptably, Relative bias was 77 percent for the ratio method, 24 percent for the TWA method, and almost zero for the HEG and the deterministic methods. Relative imprecision was reasonable for the latter three methods (150. 165, and 240%, respectively), but high for the ratio method (400%). Correlation between the measure ments and the estimates was moderate (for the HEG, TWA. and deterministic methods, r ^ 0.5H to 0.65) to excellent (ratio method, r = 0.76). These statistics were generated using every possible cell for each method. In many cases, however, the study investigators were able to develop estimates using more than one method, so that the methods were also evaluated using the same subset ofjob/rime periods. (The TWA method could not be evaluated because no estimates could be developed for any of the cells that were developed for the other three methods.) Bias and imprecision were found to be lowest for the HEG method (6 and 88%. respectively), and ir had a correlation equal to or higher than the other two methods (r = 0.65). The HEG method, there fore, was the method of choice when the criteria described earlier for using means ofthe measurements wore not met. The deterministic method performed better than the ratio method, so it was the next preferable method. The results of the evaluation were examined by type of operation to determine where the estimation methods per formed poorly. Of the ten method/operation evaluations, six estimates overestimated the measurements and four underesti mated the measurements. There was no consistency by type of operation, although the estimates averaged higher than the measurements for both sets of resin operation comparisons. The monomer operation generally had a lower bias, impreci sion, and accuracy than did the fiber operations, but the average differences were slight. The methods did much worse for the resin operation. The fiber operation had, on average, higher correlations than the monomer or resin operations, but all were modcrace. TABLE 5. Bias, Imprecision, Accuracy, and Correlation of the TWA Method by Type of Operation SE i. Bias Rel Bias Imp Re) Imp n (ppm) (ppm) (ppm) (ppm) (V.) (ppm) <%> Fiber 22 0.80 0.65 0.64 -0.16 -20 Monomer 8 0.27 0.25 0.13 -0.15 -56 0.58 0.34 73 126 Abbreviations are explained in Table 2. Resin plant results are not presented; because n - 2. *p > 0,01. Ace (ppm) 0.61 0.37 Rel Aec <%> 76 137 r 037* 0.43 jun UiL UB U't; l./'p APPL.OCCUP. ENVIRON.HYG. 11(H) NOVEMBER 1W6 Study of Wofkcn txposed [f> Acryionimie 1319 TABLE 6. Bio, Imprecision, Accuracy, and Correlation of the Deterministic Method by Typo of Operation Plant SE Bias Rel Bias Imp Rel Imp a (ppm) (ppm) (ppm) (ppm) <%) (ppm) <%) Fiber 74 1.48 0.18 1.75 0.26 18 Monomer 69 1.72 0.61 1.18 -0.54 -32 Resin 34 0.28 0.04 0.75 0.47 170 1.81 4.66 0.83 122 271 301 Abbreviations ire explained in Table 2. *p < 0.001. Acc (ppm) 1.83 4.69 0.95 Rel Aec <%) 123 273 346 r 0.38* 0.55* 0.58* Where large differences occurred between the estimates and the measurements, almost all (88%, n - 14) occurred when the measurement data appeared to be unusually low or high when compared with means in other years or with means of other similarjobs in the same year, for example, in onejob the mean of the measurements in 1978 was 1.83 ppm: in 1979, 8.04 ppm: in 1980, 0-41 ppm: and in 1981, 0.56 ppm. (All these means were based On more than ten measurements,) The three methods for which estimates were developed for this job underestimated the 8.04 ppm measurement (the ratio estimate was 1.61 ppm; the HEG estimate, 4.06 ppm; and the deter ministic estimate, 2.28 ppm). These values arc more in line with the measurement means of the other years. In one sense this is reassuring, in that the methods produced estimates that appeared reasonable, even when the referent value appeared unreasonable. In another sense, however, this difference is troublesome, because it points out the problem of using mea surement data as the referent value. Measurements are generally considered to represent truth, but because they arc generally few in number, they can be influenced by a few unrepresentative values. An attempt was made to reduce such an influence by requiring at least six 6-hour or more measurements, with nonrepresentative mea surements generally weighing less in the calculation of the mean than the sample results taken under typical conditions. It may be that this approach was not totally successful, however. One explanation for these oudying data is that the measure ments may have been taken on days that were typical, but on the high end of the typical scale (i.e., day-to-day variability), There was, however, no way to determine chis possibility- from the data. The data could also represent undocumented aty pical conditions. There are few studies evaluating historical exposure assess ment methods with which to compare these results. Homung a af.O) compared the results of a regression model with mea surement data in a study of ethylene oxide workers and found a relative bias of32 percent, a relative precision of 105 percent, and a relative accuracy of 109 percent (Homung. R., personal Communication). Investigators of a man-made mineral fiber study simulated historical environmental conditions and found that their estimates were about fourfold the measurements taken under simulation.^5 Two studies evaluated assessments made of current exposures using professional judgment. One found approximate relative biases of 200 and -4 percent (calculated from published data) yvhen the raters estimated exposures without and with monitoring data, respectively.(,9> A second study found correlations of 0,6? to 0,73 between exposure estimates and measurements of methylene chloride and 0.12 to 0.29 for styrene.Overall, our results compare favorably with these other studies. In conclusion, methods were described that svere used to estimate historical data when measurements were nonexistent. Four methods were evaluated by comparing estimates derived from these methods with measurements. All performed satis factorily, although three (the HEG, TWA, and deterministic methods) gave more accurate results than the fourth (the ratio method). Recommendations When developing historical exposures, the estimation proce dures described m tho report should be considered when measurements are nonexistent or limited and more rigorous methods, such as regression models, are nor possible. These procedures require fewer monitoring data than more rigorous methods and can be used with descriptive information. At tempts should be made, however, to evaluate their perfor mance. Acknowledgment The authors thank Dave Chesnut, Mary Ann Hcyer. and Winner Ricker of IMS. The advice of Dr. Roy Shore, Dr. TABLE 7. Bios, Imprecision, Accuracy, ond Correlation of the Ratio, HEG. and Deterministic Methods Using the Some Estimates Method S xw Bias Rel Bias Imp Rel Imp Acc Rel Acc n (ppm) (ppm) (ppm) (ppm) (%) (ppm) <%) (ppm) <%) Rario 32 1.57 0.29 2.70 M3 71 5.07 322 5.19 330 HEG 32 1.57 0.29 1.67 (1,10 6 1.39 8K 1.39 89 Deterministic 32 1,57 0.29 2.08 0.51 32 1.88 119 1.94 123 Abbreviations arc explained in Table 2. \ < 0.001. l,p < 0.01. ff 0.63A 0.65A 0.54 13 P.A. Stewart et a!L APPL.OCCUP.ENVIRQN.HYC 51(1 i) NOVEMBER 19% Richard Monson, Dr. Robert Harris Dr. Carol Rice, and Dr. Katharine Hammond is also gratefully acknowledged. References 1. Homurtg, R.W.; Greifc, A.L.; Srayner. C.T.; ct al,: Statistical Model Tor Predicting Retrospective Exposure to Ethylene Oxide in art Occupational Mortality Study. Am. J. Ind. Med. 25:825 836 (1994). 2. Dodgson,Cherne, J.; Groat, 5.: Estimates ofPast Exposure to Respirable Man-made Mineral Fibers in the European Insulation Wool Industry. Ann. Occup. Hyg. 31:567-582 (1987). 3. Stewart. P.A.; Lcnwnski, D.; White. D.; ec al.: Exposure Assess ment for a Study of Workers Exposed to Acrylonitrile. I. job Exposure Profiles: A Computerized Data Management System. Appl. Occup. Environ. Hyg. 7:820-825 (1992), 4. Stewart, P.A.; Triolo, H.; Zey. J.; et al.: Exposure Assessment for a'Study of Workers Exposed to Acrylonitrile. 11. A Computer ized Exposure Assessment Program. Appl Occup. Environ. Hyg. 10:698-706 (1995). 5. Smith, T.J.; Hammond, S.K.; Laadfow, F., Fine, S: Respiratory Exposures Associated with Silicon Carbide Production: Estima tion of Cumulative Exposures for an Epidemiological Studv. Br, J, Ind. Med. 41:100--108 (1984). ' 6. Scums, N.5.; Moulton, L.H.: Robins. T.G.: et al.: Estimation of Cumulative Exposures for die National Study of Coal Workers' Pneumoconiosis. Appl. Occup. Environ. Hyg. <v.t032~lo4t 0991). 7. Rice. C.; Hams. RJL.; Lumsden, J.C,; Symons, M.J.: Recon struction ofSilica Exposures in the North Carolina Dusty Trades. Am. Ind. Hyg. Assoc. J. 45:689-696 (1984). 8. Demem.J.M.; Hams, R.L.; Symons, M.J.; Shy, C.M.: Exposures and Mortality among ChrysotiJc Asbestos Workers. Part 1; Ex posure Estimates. Am. J. Ind. Med. 4:399-419 (1983). 9. Armstrong. B.G.; Tremblay. C.G.; Cyr. P.; Theriault. GV Estimating the Relationship Between Exposure to Tar Volatiles and the Incidence of Bladder Cancer in Aluminum Smelter Workers. Scand. J. Work Environ. Health 12:486-493 (1986). 10. Seixa>. N.S.; Tobins, T.G.; Mowton, L.H.: The Use of Geomer- tie and Arithmetic Mean Exposure in Occupational Epidemiol ogy. Am. j. Ind. Med. 14:463-477 (1988), 11. Ein, E.A.; Smith, T.J.; Wegman. O.H.; et al.; Estimation of Long Terns Oust Exposure in the Vermont Granite Sheds. Am Ind. Hyg. Assoc. J. 45:89-94 (1984). 12. Com, M.; Esmen. N.A.: Workplace Exposure Zones lor Clas sification of Employee Exposures to Physical and Chemu\il Agents. Am. Ind. Hyg, Assoc. J. 40:47-57 (1979). 13. Schneider, T.; Olsen, L; Jorgcmen. O.; Lauerscn. 13.: Evaluation of Exposure Information. Appl. Occup, Environ. Hyg. 6:47^- 481 (1991). 14. I-Ajncn, N.; Retrospective lndusmal Hypcnc Surveys. Am. Ind. Hyg. Assoc. J. 40:58-65 (1979). 15. Toney, C.R-; liamhart. W.L.; Performance Evaluation of Re spiratory Protective Equipment Used in Paint .Spraying Opera tions. U.S. Department of Health. Education, and Welfare. Cin cinnati. OH (1976). 16. Ta>k Group on Lung Dynamics; Deposition and Retention Modeb for Internal ITosimcov of the Human Respiratory Tract, Health Physics 12:173-207 (1966). ` 17. U.S. Environmental Protection Agency: Health Assessment Doc ument for Acrvlonirrile. EPA--600/8-H2-W7. USEPA, Wash ington. DC (1982). IK. Homung, R.; Statistical Evaluation of Exposure Assessment Strategies. Appl. Occup. Environ. Hyg. 6:516-520 (1991). 19. Hawkins, N.C..; vam, J.S.: Subjective Estimation of Toluene Exposures: A Calibration Study of Industrial Hygienists. Appl. tad. Hyg;. J. 4:61-68 (1989). ' 20. Post, W.; Kromhout. H.; Hccderik, D.; et a).: Scmiquannutivc Estimates of Exposure to Methylene Chloride and Styrene; The Influence of Quantitative Exposure Data. Appl. Occup. Environ. Hyg. 6:197-204 (1991). Appendix The deterministic model was based on the assumption that two important modifiers of exposure levels arc changes in the process or in engineering controls or work, practices, and changes in frequency of exposure. In most ofdie situations in \vhich the deterministic method will be used, measurement data will be available in a more recent time period and the more historical time period is the period being estimated. U is easier to conceptualize the deri vation, however, taking the opposite case (i.t\. assuming mea surement data exist for an earlier time period, and that it is the exposure of the current time period being estimated). The historical exposure is designated as the baseline estimate (B^d) (i.e,, the estimate that will be modified), and the exposure for the current year (y) is designated as the estimate derived from this merhod (r;ni7v). An exposure (whether a mean (M) or an estimate (E)) is a result of numerous sources generating emissions, and can be measured by summing the individual source' of emissions (et). Thus, Bh-n ~ *^<N.(V -n (0 If the total exposure (E or M) is equal to the sum of the emissions of all The sources, size, as a percent. Can be used to quantify the contribution of each source. The percent of emissions contributed by any source (S,) ts equal to the amount of emissions divided by the total emissions or the mean: u ~ c,,v_,/lhv-n (2) and because* the source sizes are percentages, thev must equal 1: ' " 1 (1) If a change affects the level of emissions from a source, the amount of'emissions will decrease by a quantity or a correction factor (CF). Tins term is defined as CF. * e,,, ,/e,, (4) Substituting in Equation 4 with Equation 2. we have CF, - (S.t,._n)(U(v ()/cn. (5) and o,, * (S^-uHlC.-tVCF, (6) and the estimate is therefore equal to the sum of the new emission?: BdeTy " ^CV (7) Where no change occurs at a particular source, CF = i. Thus, the estimate for the more recent time period equals the sum of the emissions that were changed (c,Vl.) and the sum of che emissions that were unchanged (e,<y_ i>,,):