Document B5mvKBdXr9pdGLRMy9ja39Ew8

American Journal of Epidemijlot.v Copyright o 1988 by The Johns Hoplun> L niversity School of Hsgiene and Public Health All rights reserved V..| l-jfi. No 2 Printed m l1 S A ST0 I 6 h 9 ! 0 VALIDATION OF WORK HISTORIES OBTAINED FROM INTERVIEWS GREGORY G. BOND,' KENNETH M. BODNER.' WENDY SOBEL.' RITA J SHELLENBERGER.-' and GEORGE H. KLOKES- Bond, G. G. (The Dow Chemical Co., Midland, Ml 48674), K. M. Bodner, W. Sobel, R. J. Shellenberger, and G. H. Flores. Validation of work histories obtained from interviews. Am J Epidemiol 1988;128:343-51. There have been few attempts to assess the validity of occupational histories obtained from interviews. A nested case-control study of lung cancer conducted in 1985 among a cohort of 19,608 male workers at a Texas chemical production facility provided such an opportunity. Telephone interviews were attempted with 308 lung cancer patients and 588 matched controls or their next of kin in order to gather information on tobacco use, diet, places of residence, and occupations. Interview data from 734 respondents (143 subjects and 591 next of kin) were then compared with the records of work area assignments maintained by the company and with exposure profiles developed by an industrial hygienist. Re spondents recalled 48.4 per cent of all documented work area assignments, but only 2.6 per cent of the chemical agents judged as likely exposures. Recall of usual work area assignment was 70.8 per cent. Among the factors found to have most influenced recall were the number and durations of assignments, a subjec tive assessment of the quality of the interview, and the relationship of the interview respondent to the subject. The elapsed time from job assignment to interview was also found to affect recall of the longest assignment of the subject. epidemiologic methods; occupational diseases; questionnaire; retrospective studies; sensitivity and specificity (epidemiology) Questionnaires are often the sole source of occupational exposure data for epidemi ologic studies, particularly for many of the large institution- or population-based casecontrol studies of cancer. The validity of underlying data depends on accurate recall by respondents. iMisclassification of sub jects can lead to over- or underestimates of risk, depending on whether cases are im pacted differently than controls (1, 2). There have been few attempts to validate work histories obtained by interview. Al though a number of investigators have tried to confirm occupational data by contacting reported employers (3-8), such studies can estimate only the degree of overreporting of jobs and cannot evaluate biases due to underreporting. Baumgarten et at. (9) re cently compared employer names obtained from interviews with those recorded in a Canadian government data bank. Although the authors found good agreement between the two sources (82 per cent), only a recent 13-year time period was examined. Fur- Received for publication May 13. 1987. and in final form September 28, 1987. 1 The Dow Chemical Company, Epidemiology, Health and Environmental Sciences, 1803 Building, Midland. MI 48674, (Reprint requests to Dr. Gregory G. Bond.) J Dow Chemical U.S.A.. Texas Operations, Building B-101, Freeport, TX The authors gratefully acknowledge support for this study from local and corporate management of The Dow Chemical Company In particular, they thank Drs. R. R. Cook, W. A. Fishbeck, and J. H. Saunders for administrative support. They also thank <J. B. Cartmill, G. A. ChaLeauneuf, A. Dang. D, Diaz, A. A. Grzegorczyk, and B. J. Lasich for technical assistance. ST0 I 649 I I 344 BOND ET AL. thermore, the authors noted the limitations facility, including dates of job changes. of using employer name as a surrogate mea Since these records were updated at the sure of occupational exposure data. time of assignment changes, they present A nested case-control study of lung can accurate accounts of job mobility that are cer conducted among a large cohort of suitable for validation purposes. chemical workers (10, 11) provided the op The two measures used to characterize portunity to assess the validity of occupa employee exposures, work area and chemi tional histories obtained by interview. This cal or physical exposure agent profiles (13-- report examines the frequency of recall er 15), were developed independently. For the rors and their relation to specific charac first, 50 homogeneous work area classifi teristics of the study design and interview cations were condensed from more than 400 respondents. One advantage of this study departmental titles used by the personnel over previous investigations was the oppor office by grouping departments with similar tunity to assess recall accuracy of jobs held exposure conditions. For the second, chem as long ago as 40 years. In addition, the ical and physical agent profiles were devel dependent variables were specific work area oped for each subject by one of the authors assignments and chemical or physical ex (G. H. F.), a certified industrial hygienist. posure agents, measures particularly appli Work history records for each person were cable to case-control studies designed to reviewed by the industrial hygienist, who identify occupational hazards. was blinded to case or control status, and all agents each worker was likely to have Materials and methods Lung cancer study encountered on any job were listed. Every attempt was made to include all agents that would have presented a reasonable poten A nested case-control study was recently tial for exposure, and considered such fac conducted among a cohort of 19,608 male tors as amount handled, handling proce chemical workers who had one or more dures, and vapor pressure of the material. years of service at a Texas production fa In the nested study, telephone interviews cility and who were followed from 1940 were attempted with 308 cases and 588 through 1980 (10, 11). These workers had matched controls or their next of kin in been identified as possibly being at in order to gather data on potential confound- creased risk of lung cancer (12), for which ers. Among other items, the questionnaire an occupational etiology was considered. included a series of open-ended questions Cases were defined as those subjects who designed to elicit a chronologic listing of all met the criteria for inclusion in the cohort jobs, department assignments, and agent (11, 12), who died before 1981, and whose exposures pertinent to the subject's career death certificates listed primary lung can at the facility. In order to confirm the iden cer anywhere on the document. Two con tity of the subject and to solicit cooperation, trol groups, one a decedent and one a "liv the respondent was reminded that the sub ing" series, were chosen from among all ject had worked for the company, and the other cohort members who did not have dates of first and last employment were lung cancer. Controls were individually supplied prior to beginning the interview. matched to cases on race, sex, year of birth, Respondents were prompted to list chemi and year hired. cal exposures; however, at no time did the Employees' documented work history interviewers suggest job titles, depart records maintained by the company served ments, or specific agents. (A copy of the as the starting point for categorizing work questionnaire may be obtained from the place exposures. The records supplied de authors upon request.) An attempt was mographic data and identified details of made to interview the person expected to each job and department assignment at the be most knowledgeable about the subject s T 0 I 61 9 I 2 validation of interview-based work histories 345 during his adult lifetime. The subject him such detail may not be required to classify self was preferred, or, if he was deceased or exposures correctly, we grouped depart otherwise incapacitated, a spouse, an off ment titles listed by the respondents into spring, a sibling, another relative, or a the 50 work areas described earlier. Where friend was contacted, in that order. Most in the 50 areas a subject had been classified of the interviews (89 per cent) were com as having worked was then checked against pleted by one of three interviewers, with the company records using the records as a the remainder conducted by two supervi standard. An employee may have contrib sors. The interview sessions ranged in uted data to multiple work areas, but was length from five minutes for partial inter counted only once for each area. views to 94 minutes and averaged about 28 Accuracy of respondent recall was mea minutes. There was essentially no differ sured by two dependent variables: recall of ence in means for length of the interview all work area assignments listed in the com between the cases (27.3 minutes) and the pany records and recall of exposures to controls (decedents = 26.3 minutes; living chemical or physical agents. A variation of controls = 31.3 minutes). the recall of "usual" work area assignments Validation study (defined as the one work area to which the subject had been assigned for the longest Interviews were successfully completed continual period of time) was also used. with 734 (82 per cent) respondents between False positives, in which the respondent March 13 and August 27, 1984. Failure to reported assignments or agents not derived locate the subject or a suitable proxy was from the record, were rare; thus, specificity the major reason for a less-than-complete of the interview data can be regarded as response, since participation among those nearly 100 per cent. The focus of the analy contacted for interview was excellent (96 sis was therefore the sensitivity rate, i.e., per cent). In comparing the participants the degree to which respondents correctly with the nonparticipants, we found that the recalled assignments or agent exposures as former were more likely to have been longer they appeared in the documented company term employees and to have been hired record. more recently than the latter. Other factors Study results were tabulated and ana examined (age at hiring, year of death, ed lyzed by packaged procedures in the Statis ucational level, and job category) revealed tical Analysis System (16). Since it was only negligible differences. suspected that the recall of the respondents Work histories obtained by interview might be influenced by a variety of char were coded according to the same scheme acteristics, sensitivity rates were examined as that used for documented company work for various subsets of the sample of sub histories and were input into the computer jects. The relative strength with which using a display management system, incor these factors may have affected recall was porating a number of quality assurance further evaluated by logistic regression steps and edit checks (11). with stepwise model-building procedures Respondents listed 248 separate depart and likelihood ratio statistical tests. The ment titles, 242 jobs, and 96 chemical or seven factors that were regarded as poten physical agents as exposures for the 734 tial determinants of recall are outlined in subjects. When these were compared with table 1. These included characteristics of the 417 department titles and 542 job titles the work history (elapsed time from the listed in company records and the 171 end of an assignment to the interview, cu agents catalogued by the industrial hygien mulative duration of the assignment, and ist for those subjects, it was obvious that number of assignments), the subject (group the respondents were not able to provide membership), and the interview (relation the same level of detail. Recognizing that ship of the respondent to the subject, qual- S TO I 6li 9 13 346 BON'D ET AL Table 1 Percentage of recorded work assignments reported in interviews, by selected characteristics, study of lung cancer among chemical workers in a Texas production facility, 1940-1980 Characteristic No. of work area assignments Usual work ares assignment* Company records Respondents (%) Company records Respondents Relationship of respondent to subject Self Spouse Other Elapsed time from assignment to interview (years) <15 15-29.9 2:30 Duration of assignment (years) <i 1-4.9 2:5 No. of assignments 1 2 to 3 >4 Quality of interview High Adequate Questionable or unsatisfactory Principal interviewer A fl C Group membership Cases Decedent controls Living controls 358 811 429 275 544 785 436 626 542 281 808 515 1.021 467 112 547 531 346 519 526 559 50.6 47.8 47.6 58.9 48.5 44.6 24.8 49.8 65.7 80.4 46.3 34.2 49.6 48.8 35.7 50.1 49.9 47.7 49.5 47.0 48.7 143 361 228 151 269 314 38 296 400 281 343 no 446 229 57 243 259 154 245 251 238 76.2 72.0 65.4 82.1 68.4 67.5 44.7 70 9 73.2 80.4 66.8 59.1 77.1 65.9 40.4 73.3 68.7 74.0 71.8 70.5 70.2 * Defined as work area assignment of tongest duration, according to company records. (Note: Total counts for some characteristics may vary because of missing data.j ity of the interview, and principal inter viewer). Interview quality was a subjective appraisal made by the interviewer regard ing the respondent's willingness and ability to furnish information about the subject. Results Overall, respondents recalled 48.4 per cent of all work area assignments and 70.8 per cent of usual assignments, but only 2.6 per cent of the chemical or physical agents listed as exposures by the industrial hygien ist. Sensitivity rates for recall of work area assignments are shown in table 1 for the sample as a whole and for subsets of sub jects classified by various potential recall determinants. It can be seen that correct responses diminished with the number of jobs and with elapsed time, but increased with longer duration jobs. Better recall was obtained in higher quality interviews, i.e., those in which the respondent was more cooperative and more familiar with the em ployee's career. In the case of usual work assignments, there appeared to be small differences in recall associated with the relationship of the respondent to the subject. These differ ences were not found when all work area assignments were considered. Table 1 also presents recall sensitivity separately for the cases and the two control series. Fewer than three percentage points ST0 ! 649.1 4 . VALIDATION OF INTERVIEW-BASED WORK HISTORIES 347 difference were seen among the groups for reduced. Thus, the relationship of the re recall of either all or usual work assign spondent to the subject was found to be ments, providing little evidence for any de adequately described by the dichotomy gree of bias based on disease state or vital "self' versus "other." Interview quality was status of the study subjects in this investi redefined as "high or adequate" versus gation. "questionable or unsatisfactory," and, for The relative importance of the factors exposures to physical or chemical agents, that determined recall and the dependen the choice of interviewer was regrouped as cies among them were investigated by "interviewer A" versus "others." building prediction models. Moderately Some examples of model estimates for complex logistic models for the seven main recall of work area are given in table 2. The effects listed in table 1 and in first-order procedure identified many of the same de interactions were considered. Model factors terminants noted in table 1 as important were defined somewhat differently than in (or unimportant). The number of assign the crude tabulations. Elapsed time to in ments and interview quality were shown to terview, duration of assignment, and num be critical to recall accuracy of both usual ber of assignments were tested as continu work area assignment and all work assign ous variables. On the basis of statistical ments. No evidence of effects due to inter evaluation of preliminary modeling results, view differences or to subjects' group status the number of levels of some factors was was seen. Table 2 Percentage of recall of recorded work assignments predicted by logistic models, study of lung cancer among chemical workers in a Texas production facility, 1940-1980 Model parameters--all work area assignments Quality of interview Relationship of Duration of respondent (years) No. of assignments 1 23 High or adequate Self 2 62.2 53.7 45.1 5 66.7 59.8 52.4 10 73.6 69.1 64.2 Other 2 65.4 51.9 38.1 5 69.7 58.0 45.2 10 76.2 67.5 57.4 Questionable or unsatisfactory Self Other 2 23.9 29.5 35.8 5 27.2 34.9 42.8 10 34.8 44.7 54.9 2 26.6 28.0 29.4 5 30.6 33.2 35.9 10 38.0 42.8 47.8 High or adequate Model parameters-- usual work area assignments* Elapsed time from assignment u> interview (>earM No. of assignments 12 1 88.5 88.6 5 87.8 869 10 86.7 84.5 3 88.6 86.0 82.1 Questionable or unsatisfactory 1 64.7 64.8 5 62.9 61.1 10 60.7 56.4 * Defined as work area assignment of longest duration according to company records. 64.9 59.3 52.1 ST0 I 649 I 5 348 BOND ET AL. A few differences between the model and Table 3 the crude tabulations were noted. The re lationship of the respondent to the subject had a stronger effect on recall in the all work areas model than in the usual work area model, the reverse of what is apparent from table 1. Although elapsed time to in terview appeared to be a strong factor in table 1, its influence in the all work areas recall model was overshadowed by another factor with which it was highly correlated, Percentage of recorded work assignments expected in interviews by selected work areas, study of lung cancer among chemical workers in a Texas production facility, 1940-1980 Work area Magnesium Construction Power No. of work area assignments Usual work arcs assignment' Company Respond- Company Respondrecords ents (%} records ents(%) 2231 36 46 69.1 30.6 65.2 122t 5 26 82.8 40.0 80.8 duration of assignment. However, for usual assignments, presumably all of consider able and more uniform duration, the elapsed time effect was seen to emerge. A substantial degree of interaction can Styrene Chlorine Maintenance Machine shop Pipe shop Administrative 29 48.3 11 56 46.4 25 143 28.7 42 40 47.5 21 135 55.6 85 204 59.3 112 72.7 60.0 45.2 61.9 64.7 79.5 be seen in table 2, all because of the influ ence of the factor number of assignments on the other independent variables. Only interview quality in the usual work areas model was unaffected. In both work area models, an increase in assignments was associated with poorer recall. Increase in assignments also acted to magnify the ef Vinyl-vinylidene chloride Paint shop Overall 1L 97 1.604 72.7 55.7 48.4 8 52 734 100.0 82.7 70.8 * Defined as work area assignment of longest du ration, according to company records. ""Number of subjects with work area assignment recorded in company work history. t Number of subjects for whom this was the usual fects of elapsed time to interview, the rela tionship of the respondent to the subject, and the duration of assignment, but dimin assignment. Number of total work area assignments (sum of number of assignments held by each subject). ished the advantage of high-quality inter views. terminants. As was found for the work area Recall of some work areas was better analyses, duration of assignment and the than others. Table 3 shows that, according relationship of the respondent to the sub to company records, recall sensitivity var ject affected recall; however, some bias be ied by about 40 percentage points for the cause of differences among interviewers most frequently occurring work area as and the study group membership of the signments of the study sample, and by as subject was evident. The interpretation of much as 60 points when only usual assign these two factors is complicated by a sig ments were considered. To determine how nificant interaction found between them. much of this variation was due to the fac It was also found that recall of some tors already examined, we included work agents was considerably better than that of area as the last term in each of the above others, although none approached the same models. Even after adjustment for the fac levels obtained for work areas. Among the tors in table 2, both models showed signif eight most common agents derived from icant differences among the five most com the company records of the study sample, mon work areas (magnesium, maintenance, the highest recall rates were found for pipe shop, administration, and paint shop) chlorine (10.7 per cent) and asbestos (8.6 and all others. per cent), and the lowest for sulfur dioxide Recall rates for physical and chemical (0.7 per cent) and heat (0.5 per cent). Sig agents were considerably lower than were nificant differences among the agents were those for work assignments, and modeling also noted in the logistic model, even after procedures identified a different set of de- adjustment for the other factors named STO I 64916 validation of interview.based work histories 349 above. Thus, as with work areas, there were particular, since the job assignments of apparently some factors not specifically greatest relevancy to disease incidence are studied that were operating to affect recall. often those that were held sufficiently long Discussion ago for an adequate latency period to have been passed. The results from this validation study The fact that agreement on work areas indicate that a fair amount of underreport was found to be almost identical for cases ing of work histories can be expected when and controls argues against the presence of data are obtained from interview, but that any differential misclassification in these some job characteristics are associated with data. Nondifferential misclassification, better recall than others. Not surprisingly, such as that due to underreporting of ex recall of usual assignments was better than posure and to some study design features recall of all assignments in general. A job (2), has been shown not to change the di is more likely to be considered a usual rection of the risk estimate, but rather to assignment as its duration increases. It was bias it toward the null value (1). It can be found that when only assignments of at demonstrated that the bias is particularly least five years duration were considered, severe when the exposure rate among the recall of all work areas improved from 48.4 controls is high. Since investigators tend to to 65.7 per cent. This approached the over select study populations with inherently all 70.8 per cent recall of usual assignments, high rates of exposure, low recall sensitivity which averaged about 5.9 years of duration. should be perceived as a definite problem. Recall of usual work areas increased only Investigators often make an effort to modestly to 73.2 per cent when assignments match dead controls with dead cases in of less than five years were ignored. Failure order to achieve similar interview circum to recall assignments of short duration is stances, even though this requires that in probably not critical to many study designs terview data be obtained from proxies. if the operation of a dose-response mecha Gordis (17) has argued against the use of nism is expected. In fact, it may be easier dead controls, contending that proxies are to detect an occupational effect among often able to provide data of a quality com- workers exposed for longer periods of time, parable with that supplied by the subjects without the diluting influence of inconse themselves. The results obtained from val quential exposures. idation of work area recall in this study Multiplicity of jobs remains a definite tend to support that contention. Only for concern to occupational epidemiology stud more complicated work histories was the ies based on interview data. It can be ex relationship of the respondent to the sub pected that more exposures will be over ject seen to be a factor. looked by the respondent as the number of Overall, free recall of specific physical or assignments increases. The findings of chemical agent exposures was poor (less this study agree with results obtained by than 3 per cent). This occurred in spite of Baumgarten et al. (9) who found signifi the fact that employees at this facility have cantly better agreement between the report been apprised repeatedly of the materials of the respondent and company records with which'they work as part of their safety among respondents who reported fewer training. This program, which has been than two jobs than among those who re evolving since the early 1960s, involves ported two or more jobs. identification of the materials in the work Elapsed time between the end of an as environment of the employee and discus signment and the interview had a substan sion of chemical toxicity, safe handling tial effect on recall, especially as the num practices, and emergency response proce ber of assignments increased. This consid dures. It should be noted that some consid eration is important to cancer studies in erations may have mediated the effect of S T 0 I 6 !* 9 I 7 350 BOND ET AL. the training program on recall. The major usual assignment per subject. The similar ity (81 per cent) of the exposures of the ities in outcome for usual assignments and subjects in this study predate the formal all assignments lessens any concern for vi inception of the program. Moreover, most olations of assumptions of independence in of the interviews (82 per cent) were con this study. ducted with a surrogate respondent instead Some unique features of this study may of the subject. The results of this investi limit the inferences that can be drawn from gation would argue that, in at least some it. First, the validation of work histories instances, free recall is not a viable tool was limited to one employer. (Although upon which to classify chemical exposures. interviews furnished data on other expo It would be preferable for investigators who sures and occupations, these could not be conduct interview-based occupational validated with company resources.) Work health studies to consult with industrial conditions or demographic characteristics hygienists to develop a specific chemical of employees at other facilities might differ agent exposure classification scheme sufficiently to render different results or to through the use of job titles and work du contribute substantial additional variabil ties. ity to interindustry studies. Some important considerations gleaned The comparisons made were limited to from the modeling process are that the specific work area assignments as well as determinants of recall are numerous, that to chemical and physical agents, as opposed they act and interact in complex ways, and to job titles. This approach was taken be that they are to some degree dependent cause work areas and chemical and physical upon study design. The large differences in agent exposures have been the primary recall of common work areas and chemical variables upon which case-control studies or physical agents, even after adjusting for of workplace hazards at this facility have known factors, hints that many important focused (13-15). To some extent, recall or factors have not been evaluated in this lack of recall may be attributed to the par study. This is further indicated by the fact ticular classification schemes used in this that the per cent variation explained by the study and may differ under another ap models was low in all cases (all work areas, proach. However, many of the work areas R2 = 13 per cent; usual work areas, R2 = 8 imply job titles, e.g., machine shop, paint per cent; physical and chemical agents, R2 shop, electric shop, and pipe shop. Further = 7 per cent), although it is unlikely that more, even chemical production work areas more complex model structures would have imply job titles, since nearly all employees helped. assigned to such areas can be classified as For the response variables of all work operators. area assignment recall and physical or The interviews were administered by chemical agent recall, almost all subjects telephone. Face-to-face interviews were not contributed multiple responses to the feasible because of substantial migration of analysis, whereas the statistical design ig former employees and their next of kin nored these potential intrasubject depend from the area. By necessity, telephone in encies. (We note, however, that this situa terviews are brief, and it may not be possi tion is not dissimilar to the prevalent use ble to obtain the complete attention of the of person-year as the sampling unit in other respondent. The respondent is also less able types of epidemiologic research.) A com to consult with records or other sources to pletely acceptable framework for such an stimulate his or her memory. It might be analysis was not apparent. On the other expected that better agreement may have hand, the analysis of recall of usual work resulted from face-to-face interviews, al area was not burdened by such dependen though this is a matter yet to be confirmed cies, since by definition there is only one by investigation. ST0 ! 649 I 8 VALIDATION OF INTERVIEW-BASED WORK HISTORIES 351 Although this questionnaire employed open-ended questions to elicit complete re sponses, we recognize that the use of closedended questions, such as a list of occupa tions or chemical agents of interest, might have stimulated recall and improved the performance of the respondents. On the other hand, such an instrument may intro duce significant recall bias. This, too, is a matter to be resolved by further study. Finally, there were few demographic data collected on the respondents (as opposed to the subjects). Although one might expect recall to be influenced by such factors as the age, education, or income level of the respondents, Baumgarten et al. (9) did not find that these factors affected agreement in their study. It is clear from this investigation that a real potential for recall bias and exposure misclassifkation exists in interview-based occupational case-control studies and that the effects on risk estimation can be seri ous. While some factors that influence re call can be identified and adjusted for dur ing analysis, others are best handled by careful decision-making during the design stage of a study. Because so few of the determinants of recall have yet been ade quately examined, there is good reason to use caution in the interpretation of such interview-based research. References 1. Copeland KT, Checkoway H, McMichaei AJ, el al. Bias due to misclassification in the estimation of relative risk. Am J Epidemiol 1977;105:488-95. 2. Flegal KM, Brownie C, Haas JD. The effects of exposure misclassification on estimates of relative risk. Am J Epidemiol 1986;123:736-51. 3. Keating E, Paterson DC, Stone CH. Validity of work histories obtained by interview. J Appl Psy chol 1950;34:6-11. 4. Mosel JN, Cozan LW. The accuracy of application blank work histories. J Appl Psychol 1952;36:365- 9. 5. Jonathan G, Moore F, Roberts L. A discussion of technique and an analysis of errors in taking industrial histories in coal miners. Br J Ind Med 1957;14:135-6. 6. Weiss DJ. Davis RV. An objective validation of factual interview data. J Appl Psychol 1960; 44:381-5. 7. Weiss DJ, Davis RV, England DW. et al. Validity of work histories obtained by interview. Minne sota Studies in Vocational Rehabilitation 12, In dustrial Relations Center, bulletin 34, September 1961. 8. Goldstein IL. The application blank: how honest are the responses? J Appl Psychol 1971;55:491-2. 9. Baumgarten M, Siemiatycki J, Gibbs GW. Valid ity of work histories obtained by interview for epidemiologic purposes. Am J Epidemiol 1983; 118:583-91. 10. Bond GG, Flores GH, Shellenberger RJ, et al. Nested case-control study of lung cancer among chemical workers. Am J Epidemiol 19S6;124:5366. 11. Bond GG. Determinants of lung cancer among chemical workers. Ph.D. dissertation. University of Michigan, University Microfilms International, Ann Arbor, MI. 1985. 12. Bond GG, Shellenberger RJ, Fishbeck WA, et al. Mortality among a large cohort of chemical man ufacturing employees. JNCI 1985;75:859-69. 13. Bond GG, Cook RR, Wight PC, et al. A casecontrol study of brain tumor mortality at a Texas chemical plant. J Occup Med 1983;25:377-86. 14. Bond G, Flores G. Exposure classification in nested case-control studies. (Abstract.) Am J Ep idemiol 1984;120:503. 15. Bond GG. Shellenberger RJ. Flores GH. A casecontrol study of renal cancer mortality at a Texas chemical plant. Am J Ind Med 1985;7:123-39. 16. SAS Institute, Inc. SAS user's guide: basics. Ver sion 5. Cary, NC: SAS Institute Inc. 1985. 17. Gordis L. Should dead cases be matched to dead controls? Am J Epidemiol 1982;115:1-5.