Measurement Issues in EnvironmentalEpidemiologyMaureen Hatchl and Duncan Thomas21Columbia University School of Public Health, 600 West 168th Street, New York, NY 10032; 2University ofSouthern California, School of Medicine, Department of Preventive Medicine, 1420 San Pablo Street, LosAngeles, CA 90033

This paper deals with the area of environmental epidemiology involving measurement of exposure and dose, health outcomes, and important con-founding and modifying variables (including genotype and psychosocial factors). Using examples, we illustrate strategies for increasing the accuracyof exposure and dose measurement that include dosimetry algorithms, pharmacokinetic models, biologic markers, and use of multiple measures.Some limitations of these methods are described and suggestions are made about where formal evaluation might be helpful. We go on to discussmethods for assessing the inaccuracies in exposure or dose measurements, including sensitivity analysis and validation studies. In relation to mea-surement of health outcomes, we discuss some definitional issues and cover, among other topics, biologic effect markers and other early indicatorsof disease. Because measurement error in covariates is also important, we consider the problems in measurement of common confounders andeffect modifiers. Finally, we cite some general methodologic research needs. - Environ Health Perspect 101 (Suppl 4):49-57 (1993).

Key Words: Biologic markers, dose, environment, exposure, mathematical modeling, measurement error, psychosocial factors, sensitivity analysis,susceptibility

Measuring EnvironmentalExposure and DoseConceptsEnvironmental exposures can occur as a

result of contact with a variety of elements(air, water, soil) that, in turn, influence thepathways for exposure (inhalation, inges-tion, dermal). Individuals' interactionswith these elements are complex, andtherefore it is not surprising that exposure

assessment and dose estimation are formi-dable challenges to those investigating thehealth effects of environmental agents.The concepts of exposure and dose have

been elaborated in a series of recent publica-tions issued by the Board on EnvironmentalStudies and Toxicology of the NationalAcademy of Sciences (1,2). The term expo-

sure refers to the concentration of an agent at

the boundary between an individual and theenvironment as well as the duration of contactbetween the two, but dose refers to theamount actually deposited or absorbed in thebody over a given time period. Althoughinternal dose is the ideal measure from the sci-entific standpoint, regulation can deal onlywith external exposures, and therefore one

may want to measure both exposure and dose.

This manuscript was prepared as part of the Environ-mental Epidemiology Planning Project of the HealthEffects Institute, September 1990 - September 1992.

*Author to whom correspondence should beaddressed.

This work was supported in part by grant R01-HD24659 from the National Institute for Child Healthand Human Development.

Individuals' exposures may be modifiedby factors such as activity patterns, whichdetermine encounters with various sourcesof exposure; bioavailability of the agent intime and place; and the rate at which expo-sure occurs (e.g., a relatively constant rateversus a variable rate). From a given expo-sure, a person's resultant dose will dependon host characteristics, such as age, sex, andmetabolism. It also will reflect the suscep-tibility of target tissue at the time of expo-sure; any shielding provided by the body(e.g., the placenta, the blood-brain barrier)or modulation by buildings that attenuateexposure to electric fields and gamma radi-ation but can be a source of exposure toradon; and the effect of concurrent expo-sures, such as cigarette smoking or medica-tions. In addition, only partcular componentsof the dose may be relevant to health effects.For calculating dose-response relationships,this biologically effective dose is whatought to be quantified. But in manyinstances it may be difficult to define whatthe biologically effective dose is, much lessmeasure it. In any event, the definition istime-dependent and subject to change alongwith the state of scientific knowledge, just asmeasurement capabilities change with newtechnology. Epidemiologists undoubtedlyneed to prepare for a new generation ofstudies in which measurement of variableswill involve data at the level of the gene. Acommitment of resources, such as talent andfunding, could improve the state of the art inexposure and dose assessment and potentiallyyield better estimation of exposure-response

relationships and more effective measuresof environmental protection.

In the past, the methods used to assignexposures in environmental health studieswere quite crude, and to some extent theystill are (e.g, pesticide usage patterns, resi-dence near a point source of pollution).Even in studies where disease has beenascertained at the individual level, exposuremeasures may be ecologic in nature andbased on average levels for a group. Whenthe group is defined in geographic terms,exposure levels might be estimated fromvalues recorded by environmental samplingin a subject's general vicinity. However,recent research has shown that correlationssometimes are weak between readings fromarea monitors and subjects' exposures mea-sured using personal monitors (3), whichare presumed to relate more closely to thetrue dose. Discrepancies between readingsfrom personal and areawide samples canresult from heterogeneity of exposures,from poor placement of samplers (e.g., airmonitors at elevations well above thebreathing zone), or from failure to takeaccount of human activity patterns andother sources of exposure.

Exposure monitoring systems can be andare being improved, however. Newerapproaches include sampling the microenvi-ronments where exposure principally occurs,including indoor environments (e.g., bed-rooms and living rooms in studies of radonand electric and magnetic fields), as well astotal exposure monitoring in which allpotentially relevant microenvironments are

49Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

HATCHAND THOMAS

sampled (4,5). The latter approach is par-ticularly important for ubiquitous com-pounds like the polycyclic aromatichydrocarbons. To some extent, personalexposure monitoring is also beginning to beincorporated into environmental healthstudies. In addition to these attempts toimprove externally derived measures ofexposure, efforts are being made to estimateinternal dose using strategies like empiricaldosimetric modeling, pharmacokineticmodeling, and biologic markers.

Such efforts are important. The failure toassign individual exposure and dose accu-rately leads to measurement errors with con-sequent effects on measures of association(and, ultimately, risk assessments) that willdiffer depending on whether the error is ran-dom or systematic and whether the unit ofanalysis is the individual or the group.Systematic error in exposure measurementcan introduce bias either toward or awayfrom the null. Random error tends to biasresults toward the null, although exceptionsto the rule can be found in unusual circum-stances (6). For ecologic studies in whichexposure is a binary variable derived fromcombinations of individual observations, therule stating random error generally biasesresults toward the null may not hold (7).

Given the consequences of error in esti-mating exposure, it is important to try toincrease accuracy of measurement at thedesign stage of a study. How, then, doesan investigator decide when the use of asurrogate exposure measure (i.e., an error-prone measure) is acceptable, and when itis not? Rosner et al. have shown (8) thatfor correlations between surrogate and truemeasures of exposure less than 0.8, theodds ratios estimated by logistic regressionwill differ markedly for the surrogate andthe true exposure measure, while much lessbias will occur when correlations betweenthe two measures are 0.8 or greater. In vivotibia lead levels measured by X-ray fluores-cence have been proposed as a good surro-gate for cumulative blood lead levels on thebasis of a correlation coefficient of 0.84(9). For dietary exposures, however, thecorrelation between food frequency ques-tionnaires and less error-prone methods(food records, measurements in food orbiological samples) is only around 0.5 (10);yet food frequency questionnaires continueto be applied in large-scale studies, onlyoccasionally with correction of risk esti-mates for error in measurement. On theother hand, the failure to find a correlation(actual coefficients not given) between cur-rent adipose tissue or serum dioxin levelsand surrogate measures of past exposure to

Agent Orange in Vietnam (11,12) affecteda decision not to conduct further researchusing exposure surrogates based on trooplocation and herbicide spraying records.These examples underscore the need to beexplicit about criteria for acceptable surrogatemeasures, as well as the need to take errorinto account when surrogates are used, evenwhile emphasizing the development of betterapproaches to exposure-dose assessment.

In the following section, we describe meth-ods designed to reduce error in exposure mea-surement insofar as is currently possible(approaches such as dosimetric modeling,pharmacokinetic modeling, biologic markers,and use of multiple measures), as well asapproaches to assessing the residual uncertain-ties in the estimated dose. Even the best of thecurrent methods will not yield a measure thatis completely error-free, and it is thereforeimportant to recognize and characterize theresidual error in measurement so that it can beconsidered in analysis ofthe data.

Measurement ApproachesExposu or Dose ModdingEstimating a subject's exposure to an envi-ronmental agent involves combining infor-mation about possible sources of exposure(usually obtained from the subject, fromsome other respondent, or from records)with an assessment of the likely degree ofexposure from each source.When an exposure under study is envi-

ronmental, there may be multiple pathwaysby which a person might be exposed and itcan be important to consider all elementsand all routes. For example, residentsdownwind of the Nevada Test Site couldhave been exposed to external gamma radia-tion from the passing fallout cloud itself,from ingesting contaminated milk orvegetables, or, in the case of infants, fromin utero exposures or breast-feeding. Foreach of these pathways, several differentradionuclides might need to be considered.After eliminating pathways that would beexpected to make a negligible contributionto the total dose, one can estimate thelikely dose rate per unit of exposure to eachpathway. In the fallout example, thisinvolved consideration of a) source term,the amount and type of radionuclidereleased; b) the environmental transport,dispersion from the source to sites of depo-sition; c) rate of radioactive decay and envi-ronmental dispersion of the radionudides;d) farm management practices leading tocontamination of dairy cattle or vegetables;e) estimates of the uptake of radionudidesby vegetables and milk; f) distribution of

milk and vegetables to consumers; and g)uptake by the target organ from ingestedradionuclides. To calculate an individual'sdose, this information was then combinedwith extensive questionnaire data on breast-feeding and maternal and individual con-sumption of milk and vegetables at variousages. For some subjects, modificationswere needed to allow for homegrown veg-etables or backyard cows or goats. For sub-jects with incomplete exposure information,distributions of default values specific to theirparticular circumstances (age, sex, location,etc.) were developed. Similar calculationswere performed for each of over 100 nudeartests, and the results then were summed toproduce estimates of each subject's totaldose (13).The process described above is far more

complex than has been the norm in envi-ronmental epidemiology, but it representsthe current state of the art in environmen-tal dose assessment. Less refined, but per-haps less costly, approaches to exposure-dosemodeling (often for households or geo-graphic areas rather than for individuals)have been based on Gaussian-dispersionmodeling of airborne emissions (14-16),hydrogeologic modeling of waterborneexposures (17), and isopleth modeling ofsoil contaminants (18). Assuming thatdosimetry models are reasonably accurate,such approaches should decrease bias aris-ing from measurement error and increaseprecision. Assessment of the validity ofdosimetry models should be made wheneverpossible. For example, an environmental dis-persion model of emissions at the time of theaccident at the Three Mile Island nuclearplant was validated by the readings fromoff-site thermoluminescent dosimeters.

Dosimetric modeling methods are likelyto be used more frequently in future envi-ronmental health studies. A question iswhether the effort required both in terms ofthe information that must be collected fromstudy subjects and/or by environmental sam-pling and the effort involved in developmentof the dosimetric model itself are warrantedby the gain in precision or reduction in biasof the exposure estimates. Information onthis point could be obtained by comparingthe point and interval estimates of associa-tions observed using gold standard dose esti-mates with those that would be obtainedusing cruder methods. Such comparisonscould be made in existing data sets.Understanding when the gains from dosi-metric modeling are substantial and whenthey are only marginal would be useful inestablishing methodologic standards ofpractice.

Environmental Health Perspectives Supplementsl/rInime 101, Supplement 4, December 1993

50

MEASUREMENTISSUES

Some other issues related to dosimetry areexemplified by studies of cancer and electricand magnetic fields (EMFs). The initialhypothesis about EMFs was derived fromobservations showing apparent excesses ofleukemia (and some other cancers) both inchildren living near electric power lines thatwould be expected to generate high magneticfields (19) and in certain dasses of electricalworkers (20). In both the residential andoccupational settings, it has been difficult toestablish whether the magnetic fields are theresponsible agent. While subsequent studieshave demonstrated that certain electricalwiring configurations and certain categoriesof electrical work are associated with higherthan average fields, so far no convincing asso-ciations have been found between leukemiarisk and individuals' exposure to electric ormagnetic fields determined by area measure-ments. No studies using personal dosimetryhave yet been reported.

Four possible explanations are suggestedfor the failure to establish a clear associa-tion between cancer and measured fieldstrengths. First, it may be due to theirextreme variability in space and time. Anynecessarily short-term measurement (24 hror a week in a small number of locations) isa poor surrogate for lifetime dose; underthis explanation, household wiring classifi-cations and job titles may be more stablemeasures of long-term exposure. Second,the failure to detect an association withmeasured fields may reflect a failure tomeasure the biologically relevant parameter(e.g., peaks, transients, resonance betweenstatic and oscillating fields rather than thetime-weighted average). Studies of repro-ductive outcomes, where the period ofexposure is much shorter than for cancerand where there may be a particular timewindow of vulnerability, could help indi-cate whether the discrepancy in associa-tions with wire codes and measured fields isdue to their capturing different time framesor different dimensions of EMFs. A thirdexplanation for the associations of cancerwith wiring configurations, but not withmeasured fields, relates to selection bias(lower selection probabilities for controlsliving near wiring with high current config-urations). Fourth, the surrogate exposuremeasures (wire codes, job titles) may beconfounded by other correlated risk fac-tors. This controversy is still far fromresolved, but consideration of selection biasand possible confounders together withcareful assessment of all potentially salientaspects of electric and magnetic fields andof the variability of the different measure-ments should shed light on the issue.

The EMF example underscores the needfor making multiple measures of exposure.In particular, it argues for continuing toindude surrogate measures along with goldstandard measures in studies of healtheffects until the relations between the sur-rogate and criterion measures are wellunderstood and there is certainty about thetrue gold standard (i.e., until the correctbiologic mechanism is known). Substitutingan incorrect gold standard for a surrogatemeasure can actually increase measurementerror. One analytic approach to usingmultiple measures that has been proposedas a means of increasing validity is torestrict analysis to subjects who are classi-fied as exposed or unexposed by two differ-ent, if imperfect, exposure measures (21).This dearly risks some loss in power sincesubjects with discordant results on the twomeasures are excluded from analysis.Another proposed approach is to estimatethe misclassification probabilities for eachmeasure and from them to estimate theprevalence of exposure (22).Some mention of personal monitors

should also be made. While these do notprovide a measure of resulting body burden,as biologic markers are meant to do, per-sonal monitors may measure the intensity ofan individual's total exposure to airborneagents better than fixed-site area monitors.This is not always the case, however, partic-ularly in studies of long-term exposures orwhere areawide concentrations are fairlyuniform. The TEAM study (TotalExposureAssessment Methodology) conducted by theU.S. Environmental Protection Agency(EPA) found that personal air monitors wereacceptable to subjects from 7 to 85 years ofage (23). Investigators studying effects ofexposure to EMFs and indoor air pollu-tants on children are anxious to developpersonal monitors that can be used withchildren under age seven, including tod-dlers. At present, personal monitors forEMFs are in the form of wristbands andmay not be suitable for very young chil-dren. Technology for personal exposuremonitoring is still evolving, but it willrarely be feasible to apply personal expo-sure monitoring to all subjects and all rele-vant time periods. Therefore, methodologicapproaches are needed for combining col-lected exposure data with personal samplersand environmental monitors.

Pharmacokinetic ModelingPharmacokinetic modeling is an approach todosimetry that incorporates informationabout the internal pharmacologic processesthat ensue once an agent reaches the portal(s)

of entry into an individual's body (24).These include uptake into the circulation;distribution within the body; and metabo-lism, storage, and elimination. These mod-els can be simple, involving only one bodycompartment, or complex, involving mul-tiple body compartments. In either case,compartmental rate relationships are usedin the model's equations to estimate con-centrations at critical tissues. Such modelsare also useful as guides to temporally rele-vant and efficient ambient sampling (24).Pharmacokinetic modeling of exposure anddose may be viewed as a counterpart tobiologically based disease models.

Biologic MarkrsBecause of the difficulty of obtaining accu-rate and unbiased exposure informationfrom study subjects and the difficulty of esti-mating the doses that such exposures mightproduce, there has been great interest in thedevelopment of biologic markers. Thesemay be defined as "cellular, biochemical, ormolecular alterations that are measurable inbiological media, such as human tissue, cells,or fluids" (25). If used appropriately, bio-logic markers allow for considerableimprovement in measurement of dose.First, they may obviate the errors arisingfrom subjects' lack of knowledge, memoryfailure, biased recall, or deliberate misinfor-mation (26). Second, even when subjectreports of exposure are accurate, individualsmay vary considerably in uptake and han-dling of a material; the error introduced bysuch individual variation can be reduced orremoved by using markers that provide anestimate of the dose to a particular individ-ual. Third, some markers can be used todetect biological interactions between theexposure of interest and critical tissues;DNA adducts are an example of this type ofmarker. In studying environmental tobaccosmoke, for instance, one can-in additionto asking about maternal smoking duringpregnancy-actually measure smoking-related DNA adducts in placentae (27) and,where the fetus is lost, in critical organs suchas fetal lung or liver (28). Another advan-tage of biologic markers is that generallythey give a quantitative, or at least semi-quantitative, estimate of dose. They alsocan serve as the gold standard for otherinformation sources, thus providing a basisfor error allowance procedures in studiesthat rely on less accurate exposure measuresdue to the cost of the marker.

Other Biologic DosimetersCertain signs or symptoms can also beviewed as biologic dosimeters. For example,

Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

51

HATCHAND THOMAS

in the cohort of atomic bomb survivors, ithas been reported that subjects with a his-tory of epilation have a 2.5-fold steeperdose-response curve for leukemia thanthose without (29). This can be inter-preted either as an indicator of their greaterradiosensitivity or as an indicator of mises-timation of their doses, perhaps as a resultof differences in shielding not accounted forby available dosimetry data.To be useful in environmental epidemiol-

ogy studies, a biologic exposure markershould be dearly better than anamnestic dataor environmental measures; should allow fordifferentiation between exposure levels;should be applicable on a large scale; or if toocostly for large-scale use, should at least beacceptable to subjects in a validation sub-study. Before markers are used in epidemio-logic research, their sensitivity and specificityshould be known from both the laboratoryand epidemiologic perspectives; reproducibil-ity of results within and between laboratoriesmust also be known; and, very importantly,the particular time frame they reflect and dur-ing which they can be measured in vivo mustbe established (25) so that they provideinterpretable data regarding time and dose.At present, few exposure markers satisfy

these requirements. Some markers may pro-vide a record of cumulative exposure (e.g.,bone lead measurement, mercury or cocainemeasurements in hair), but most can assessonly relatively recent exposures. Studies ofbiologic markers that use a case-controldesign and a cross-sectional marker of expo-sure can be difficult to interpret because ofambiguity about the temporal sequence ofthe marker and the disease [e.g., whetherselenium levels in breast cancer cases arecause or consequence (30)]. Indeed, suchstudies can be misleading. Vineis andCaporaso (31) have described how acase-control study nested in a cohortallowed Wald and his colleagues (32) tomake use of the time between initial collec-tion of specimens from members of thecohort and subsequent onset of cancer todarify the time order in the relationship withblood retinol. Although analysis consideringonly the early cases of cancer suggested thatblood retinol might be protective, ultimatelyit was apparent that some metabolic changeassociated with the disease was acting toreduce retinol levels, rather than vice versa.In addition to such problems in interpreta-tion, biological measurements are oftencostly to perform. Furthermore, the need toobtain specimens can reduce the cooperationof subjects and introduce the potential forselection bias to occur through initial refusalor later attrition, although these problems

are probably not insurmountable if they areanticipated and addressed.

Use ofMultiple MeasuWhen the biological basis of an associationis poorly understood, it can be very helpfulto have various types of exposure measure-ments available. Or, as mentioned previouslyin connection with personal exposure moni-toring, it may be necessary to rely on anothersource of exposure information for portionsof the study period. The obvious approach isto analyze each type of measurement sepa-rately, but there may be merit in combiningthem into an index, if only to reduce mea-surement error. Complications can arise if allmeasurements are not available on the samesubjects. Any associations observed might bedue to differences in the measurements or todifferences in the subgroups of subjects forwhom the measurements are available. In astudy of childhood leukemia and electric andmagnetic fields, London et al. (33) reportedthe results separately for various summaries of24-hr bedroom dosimetry, spot measure-ments at various locations, and wiring config-urations. However, drawing on all of thesedata, they also developed regression modelsfor magnetic fields at various locations basedon attributes of the wiring and used the val-ues predicted by these models as the time-weighted average fields for all houses lived in.Thus, predicted values were used both toreplace existing measurements and to imputemissing values. The rationale behind theapproach is to avoid the loss of informationand possible selection bias associated withrestricting analysis to subjects with data for allmeasurements made (34). One alternative isto retain measurements where they exist andto impute only the missing values, leavingopen the possibility of stratifying on dataquality in the analysis. Other approachesundoubtedly can be devised, and it would bedesirable to compare their validity using datasets in which exposure-response relationshipsare well understood and where more thanone measure ofexposure exists.

Other Issues in Measurementof ExposureTakingAccount of Critical Periods forExposureA principal problem in environmental epi-demiology has been that the inaccuracy inmeasurement generally (although notalways) operates in the direction of overes-timating exposure and therefore underesti-mates risk or perhaps misses health effectsaltogether. For example, when assigningthe same level of exposure to all 1000 resi-

dents living within five miles of a toxicdump site when only 100, say, were trulyexposed and the other 900 were either unex-posed or exposed at very low levels, onewould be certain to calculate an observedrelative risk for exposure that would belower than the true risk. Hence the impor-tance of increasing the accuracy of exposuredefinitions and measurement is obvious.Rothman and Poole have pointed out (35)that it is also important to use informationon critical periods for exposure, either inthe design phase of a study, in the analysisphase, or in both. For example, in a studyof Down's syndrome, parental exposuresoccurring after the fertilization period arepresumably irrelevant to the outcome; infact, there is mounting evidence that mostcases of Down's are traceable to errors atthe time of the first meiotic division in thematernal germ cell (36). By removing allexposures that are not of biologic conse-quence from the estimate of association,one can expect the magnitude of the esti-mated association to increase. Moreover,information on known critical periodsmight be used to test whether an associa-tion appears to be spurious. If an associa-tion were found not only during the criticalperiod but also for exposure during noncriticalperiods, then the association might be due torecall bias, or it could be reflecting autocorrela-tions in exposure status. Multivariate analysisof the effects ofexposure in various critical andnoncritical periods could, in principle, over-come this problem, provided there are enoughexposed subjects with different temporalpatterns ofexposure to be informative.

TakinAccount ofMiration In andOutofxpsd AreasThe problem of in- and out-migration isfrequently raised as an issue in interpretingresults of studies that define exposure interms of time and place. Although severalstudies have considered the effects of popu-lation migration on the validity and preci-sion of estimated associations betweenexposure and disease (37) and have describedwhen and in what direction bias is likely toarise, these issues are still not understoodwell. Perhaps more simulations or empiri-cal demonstrations are needed to improvethe general level of comprehension aboutthe effects of population mobility on geo-graphic studies. In the case of specificstudies, it would help to know somethingabout duration of residence or at least age-specific duration patterns in an area. Onerecent suggestion is to estimate by variousmeans the fraction (f) of time spent by asubject in a particular place and to assign for

Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

52

MEASUREMENTISSUES

the remaining fraction (1-f ) the averageexposure for some total referent area (38).

A ing Past EPosunA major problem in many environmentalhealth studies is the difficulty of estimatingpast exposures when only present-day mea-surements are available. Often, some data onsubjects' past exposures can be obtained byquestionnaire or review of existing records.For example, in occupational studies, payrollrecords are used to assemble a job history.The use ofrecords from years past to establishexposure status has the important advantageofobviating recall bias, although it may intro-duce its own problems (e.g., missing recordsor less specificity in records from early years).Estimating the actual historical exposure lev-els is more difficult than simply classifyingexposure status, and it often involves a largedegree of judgment. Clearly, the more his-torical data there are on variation in exposurelevels over time and place, the better. Studyofsuch patterns ofvariation can suggest mod-els for predicting exposures at times for whichno measurements are available. For example,in a study of salivary tumors and dental X-rays, Preston-Martin et al. (39) reviewed 58studies that described doses from various pro-cedures at various times and, while takinginto account the dates of introduction ofnewtechnologies, used regression analysis todevelop models for the expected dose as afunction of calendar year. In occupationalsettings, the subjective experience of long-service workers has been used to comparecurrent exposures with those in the distantpast. Similar strategies (i.e., tracking techno-logical developments, use of knowledgeableinformants) need to be applied in the assess-ment of past environmental exposures. Forexample, in a case-control study of colorectalcancer and water chlorination among womenteachers in New York State, Lawrence et al.(40) used current water sampling in conjunc-tion with records from water treatment plantscovering the previous 20 years in a mathe-matical model to estimate cumulative expo-sures to chloroform in drinking water athome and at work.

Uses ofEisting

One limitation on assessing past environ-mental exposures is that reviews of existingdata bases at the national and state levelsrepeatedly have found them to be inade-quate for epidemiologic purposes because ofinsufficient data points to assess variability, lackof a standardized Quality Assessment/QualityControl protocol, incomplete geographic cover-age, and missing information (41). Efforts

are underway to modify the major air andwater data bases to make them more usefulfor future environmental health studies.

Existing environmental data banks couldalso be used to define strata within whichto conduct sample surveys. Surveys ofindividuals within these ecological exposuregroupings would help document humanactivity patterns and could indicate the dis-tribution of exposure and important con-founding or effect-modifying variables ineach stratum. Potentially, such stratified-sample surveys might provide the basis forconstructing an environment-exposurematrix similar to the job-exposure matricesused in occupational studies. Such expo-sure matrices are generally assumed to havea "Berkson error" structure (42), in whichthe average of the true doses for all subjectsin an exposure assignment group is equalto the assigned value. As a consequence, ifthe true dose-response is linear, the esti-mated slope of a linear relationship will notbe biased toward the null.

Estimating Dose UncertaintiesA major concern among environmental epi-demiologists is the influence of errors in expo-sure estimates on associations with disease andmethods ofdaling with such errors. The bestcure for this problem is to avoid measurementerror in the first place. When this is not feasi-ble (and it often may not be, particularly ininvestigating common source exposures suchas toxic dump sites), it is helpful to be able toquantify the direction and magnitude of theerrors. This can be done in a number ofways,induding a) validation studies on a subset ofthe study sample or a pilot sample to comparethe measurements to be made in the field witha gold standard, b) replication of measure-ments to assess within-subject variability, c)multiple types of measurements to assessvalidity, and d) sensitivity analysis to esti-mate the influence of various unknowns oruncertain parameters on the estimateddoses. The goal might be either to describethe distribution of exposure errors acrossthe population (or subgroups there of) orto obtain an estimate of the precision ofeach subject's exposure assignment.

Because a gold-standard assay is often notfeasible for use in the field (because of cost,time, acceptability, etc.), validation studiesusually must be limited to a relatively smallnumber of subjects. The resulting estimates oferror distributions may be imprecise (43),although this will be less of a problem if thedata are treated as continuous and if parame-ters for sensitivity and speificity do not haveto be estimated (8). Nonetheless, sample sizesfor validation studies that are needed to insure

good estimates of the error rates in field mea-surements should be calculated carefully.Other considerations are to insure that themeasurement error process in the sample usedfor validating the field measure is similar tothat in the target population for the fullstudy and to avoid selection bias in the vali-dation study, which might arise if require-ments associated with use of the goldstandard measure are very demanding andparticipation rates are consequently low. Inthe New Jersey case-control study of radonand lung cancer among women, in-homeradon measurements were obtained for only40% of the houses targeted, and smokingrates differed among those with measuredand unmeasured homes, raising the possibil-ity of selection bias (44). If data on diseaseare collected on validity study participants,potential selection bias can be examined bytesting for heterogeneity in the risk estimates.

Replicate measurements are useful fordescribing repeatability (45) but cannotassess other components of error, such as sub-jects' tendency to consistently overreport orunderreport exposures. Having differenttypes ofmeasurements available may be moreuseful in estimating misdassification proba-bilities, even if none of the measures is error-free. See, for instance, Hui and Walter'smaximum likelihood method for estimatingerror rates with two independent assessmentsofexposure (22).

Sensitivity analyses can take a number offorms. The basic idea is to consider a rangeof plausible values for each of the unknownsin the exposure assignment process. If thereare only a few unknowns, one might considereach of them and evaluate their influence oneither the individual exposure assignments orthe final dose-response relation. If there aremany, one can estimate the distribution ofassigned doses, either analytically or byMonte Carlo simulation. The latter approachwas used in the studies around the NevadaTest Site because of the complexity of thedosimetry algorithm. Components of uncer-tainty that were considered indude the sourceterm, environmental transport, farming prac-tices and distribution, and default values forindividuals' missing data. A series of sensitiv-ity analyses were also carried out on a mathe-matical model that estimated the relativegeographic distribution of exposure to acci-dent emissions at Three Mile Island by exam-ining variations in modeling assumptions fortheir effect on the base case (46). Parametersconsidered were the source term, the degreeof plume rise, wind shifts, and residual errorweighting. In addition, a Bayesian analysiswas used to quantify uncertainty about thetime-release pattern.

Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

53

HATCHAND THOMAS

Measuring Outcome ofEnvironmental ExposuresDdeinitioa Iues

As strong effects of environmental exposure

have been identified and dealt with, envi-ronmental epidemiology increasingly hasbecome a search for weaker associations. Itis all the more important, therefore, to

improve measurement of outcome throughcareful definition and avoidance or reduc-tion of error (35). In defining study endpoints, the aim should be to specify thehealth outcome of interest as precisely as

possible in order to avoid further dilutionof a weak association through inclusion ofirrelevant cases. In fact, it may be desirableto consider subgroups of disease that are

etiologically homogeneous and that are

believed to be responsive to the exposure ofinterest on the basis of theory or priorobservations (e.g., certain histopathologictypes of lung cancer and radon; leukemiatypes and subtypes with ionizing radiationand EMFs). This can present something ofa dilemma, however, because statisticalpower for examining subgroups is likely to

be low unless the difference in effect sizeamong subgroups is sufficient to offset thereduced sample size.The virtues of lumping versus splitting

frequently come up for discussion in thecontext of studies of congenital anomalies.It is unlikely that an exposure would affectall types of congenital defects. With mater-nal cocaine use during pregnancy, forexample, defects involving vascular disrup-tion seem to be implicated. However, a

biological basis for positing subgroups ofinterest is often lacking; empirical Bayesianapproaches may be useful in helping to for-mulate relevant subgroupings. In any

event, the numbers in particular case

groups are likely to be small for all but a

few categories. If sufficiently large seriescannot feasibly be accrued in a single study,multisite (even multinational) projects mayneed to be mounted, or more reliance mayneed to be placed on meta-analyses com-

bining results from several studies. Whichof these strategies to pursue should be dis-cussed by groups of investigators studyingthe same exposure, and their potentialfunding sources.

Disease outcomes in environmental epi-demiology can be measured on a continu-ous scale or categorically as incident or

prevalent cases or as deaths. Incidence dataare usually preferable for investigating eti-ology since prevalence or mortality datamay be influenced by factors affectingduration of disease and survival as well as

those relating to cause. However, inci-dence data are often less easily accessedthan mortality data, and they can be sub-ject to artifactual variations in ascertain-ment-as a result of screening programs,for example. Whether incidence or mortal-ity is the more reliable indicator of healthstatus and in what age groups it is reliablehave been discussed extensively but notresolved. See, for example, the recentpapers by Doll (47) and by Davis et al.(48) about cancer time trends. It might behelpful to have a set of recommendedapproaches for trend analysis that weredeveloped by a group of dispassionatemethodologists. For etiologic studies, inci-dence data seem conceptually superior;when mortality data are used, considera-tion needs to be given to accounting forinfluences on survival since these mightcorrelate with exposure.

In some areas of research, such as repro-duction and development, different out-comes can occur depending on the timingand dose of exposure. In such circum-stances, it may be important to examineseveral end points. Extending population-based registration systems to cover moreoutcomes than cancer and birth defects andto cover more geographic areas potentiallycould be useful for environmental studies inseveral respects: in identification of cases, invalidation of self-reported information, andin ascertaining disease status of migrants.

Biologic Effect Markers and OtherEarly Indicators ofDiseawBiologic effect markers potentially have anumber of advantages as study end points,particularly if they are strongly prognosticof disease in ways not explained by avail-able exposure information-for example,by reflecting susceptibility or the action ofcofactors (26). While some effect markersare actually subclinical events (e.g., bio-chemical tests of occult pregnancy loss),often markers of effect correlate onlyweakly with disease. Serum alpha-fetopro-tein is a useful marker for liver cancer aswell as a prenatal marker for neural tubedefects. Markers that are not as clearly pre-dictive of risk, particularly at the individuallevel, can lead to problems of interpretationand to needless anxiety for those individu-als found to have elevated levels. The pre-mature application of a poorly standardizedcytological assay on a group of already con-cerned residents at Love Canal is a case inpoint. Calls have been made repeatedly tocarry out longitudinal studies, in experi-mental animals and humans, that will mea-sure the positive predictive value of such

markers before applying them in field stud-ies; but these have been largely ignored.The Scandinavian countries, however, havemounted a collaborative prospective studyof cancer in a cohort of 3190 individualswho have been tested for sister chromatidexchanges (SCEs), structural chromosomeaberrations, or both. A report based on a13-year follow up of 800 subjects in theFinnish portion of the data (49) found amoderate, statistically significant positiveassociation between cancer risk and chro-mosome aberrations (SMR = 2.65; 95% CI1.2, 5.0); there was a positive trend (SMR =2.06; 95% CI 0.8, 4.2) for SCEs. Additionalprospective studies of this kind are needed toestablish the relationships between markersand disease in order to assure their appro-priate use and interpretation. In addition,determining when a marker could serve asthe basis for preventive health measuresdirected at a distal end point such as canceris an important issue; see Prentice (50) fora useful discussion of this and a proposedoperational criterion for surrogate responsevariables.

Other potential advantages of biologiceffect markers are their use in classifyingdisease more precisely and in suggestingmechanisms of action, such as those relat-ing to susceptible subpopulations. Forexample, biologic markers that distinguishslow from fast acetylators have indicatedthat the enzyme N-acetyltransferase playsan important role in bladder cancersinduced by exposure to aromatic amines(51,52). Methodologic needs in the areaof effect markers include attention tosources of variability, both biological andlaboratory-related, and to logistical issues,such as how to achieve reasonable partici-pation rates when the effect marker requiresa demanding regimen. Three current studiesof early pregnancy loss illustrate this latterproblem. Two of the studies ask participantsfor daily urine samples. The third studyuses a modified specimen collection schemerequiring urine samples only twicemonthly, at the beginning of menses.Preliminary data indicate higher responserates for the study with the simplified col-lection protocol. Whether the variabilityin enrollment is due to the differingdemands on study subjects or to other vari-able aspects of the three studies (such as theperceived salience of the topic in the targetpopulation) is not known. Systematicresearch is needed to determine how toachieve cooperation in studies that use bio-logic markers and how to provide for calcu-lating or estimating the extent andmagnitude of selection bias.

Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

54

MEASUREMENTISSUES

Subdinical End PintsWhat role should physiologic changes (e.g.,nerve conduction velocity, T-cell subsets,sperm count) have in environmental healthassessments? It has been argued that func-tional alterations and nonspecific symptomsare likely to be more frequent consequences oflow-level environmental exposures thanfrank disease (53). However, baseline rates

and normal ranges for such end points maybe lacking. Objective methods of assess-

ment to remove the potential for biasedrecall may be at an early stage of develop-ment, and interpretation of results in terms

of risk to groups and to individuals fre-quendy is problematic, particularly as assay

improvement allows for discriminatingfunction more and more minutely. Thesemethodologic limitations can be addressed-semen evaluation is a case in point (althoughthe clinical significance of altered semen

quality is still not clear-cut)-however,substantial time and effort will be required.

Measuring Confounders andEffect ModifiersEffect on Risk Esima ifIndequatel Controlied

A confounding variable is one that, if not con-

trolled appropriately, will tend to distort theexposure-disease association. For example,when studying whether household exposure to

radon is a cause of lung cancer, one should beconcerned about the possible confoundingeffect of smoking. Smoking is dearly a majorrisk factor for lung cancer. Ifhouses with highradon levels are more likely to be inhabited bysmokers, then this would produce an apparentrelationship between radon and lung cancer

even if there were no causal effect. The con-

verse also could happen; if smokers tended to

live in low-radon houses, then one might failto find an association between radon and lungcancer if it really were present.The strategies commonly used by epi-

demiologists to control confounding induderestriction (e.g., to nonsmokers), matching,or statistical adjustment. All of theseapproaches presume that the confoundingvariable has been correctly measured.Greenland (54) has pointed out that errors

in measurement of a confounding variablewill tend to cause partial loss of an abilityto eliminate confounding bias; for example,if the true odds ratio (adjusted for the true

confounder) is 2.0 and the crude odds ratio(unadjusted) is 4.0, then the odds ratioadjusted for an incorrectly or crudely mea-

sured confounder might be 3.0. Thisintermediate outcome can only be countedupon in a case in which the errors in mea-

suring the confounder are random (unre-lated to exposure or disease status); in othercases, the adjusted odds ratio could be fur-ther from the truth than the unadjusted oddsratio. Kupper (55) has shown that an inac-curate surrogate confounder can produceseriously misleading inferences.A factor like smoking, in addition to

being a confounder, could also act as aneffect modifier-that is, a variable thatmodifies the strength of the associationbetween exposure and disease. A majorquestion in the radon literature is whetherthe joint effects of smoking and radonexposure are multiplicative, additive, orsome intermediate possibility. If they actadditively, for example, then radon expo-sure would produce the same additionalrisk of lung cancer in smokers and non-smokers; but because lung cancer is rare innonsmokers, it would follow that radonexposure might account for a much largerproportion of lung cancers in that group.Conversely, if the two exposures act multi-plicatively, the proportional increase inlung cancer rates due to radon exposurewould be the same in smokers and non-smokers; but because of the higher rates insmokers, the absolute increase would belarger in smokers. This issue therefore hasimportant risk assessment and public healthpolicy implications. Again, Greenland (54)has shown that errors in measurement of acovariate can distort its modifying effectand possibly introduce an apparent interac-tion where none exists. Diet and cookinghabits in relation to aflatoxin exposure, andshowering habits in relation to radon areadditional examples of potentially importantconfounding or effect-modifying variables inenvironmental epidemiology.

Approaches to Measuring CommonConfounders and ModifiersThe implications of the previous sectionare that careful measurement of strong con-founders or modifiers should be given asmuch attention as the exposure and diseasevariables. It follows that some of the sameapproaches discussed in the sections onmeasurement of exposure and disease, suchas use of multiple measures and biologicmarkers, will pertain here as well.

Continuing with the example of smoking,it is not sufficient simply to classify subjectsby their present status as current, former, ornever smokers. As long as smoking is a riskfactor for the disease under study, one usu-ally tries to obtain information on at leastthe ages at starting and stopping and theaverage daily amount of smoking. Thesedata can be used to compute pack-years (the

product of amount and duration), which isa stronger predictor of lung cancer riskthan current status. In some other cases,however, such a product term may actuallyincrease error. Better yet, nonlinear multi-variate models could be used to allow forthe joint effects of age at starting, durationand intensity of smoking, and time sincequitting. Other modifying factors mightinclude changes in level of smoking overtime, use of filter cigarettes, and depth ofinhalation. However, incorporating multi-ple modifying factors into an analysis needsto be done with considerable thought toproduce models that are biologically plausi-ble. Routine inclusion of interaction termsin a multiple logistic regression analysis canproduce models in which ex-smokers even-tually become at lower risk than neversmokers, or light smokers have the samedependence on duration or age at start asheavy smokers. Use of general risk modelsbased on biologically plausible theories is anattractive altemative.Even the most complete smoking history

is still likely to be misclassified, and theerrors might well be related to the exposureor disease variables under study. In anoccupational study of radon exposure andlung cancer, for example, miners with lungcancer might preferentially underreporttheir smoking histories to avoid prejudicinga compensation claim. For these reasons,there has been great interest in developingunbiased methods of assessing potentialconfounders. Biological measures, such asurinary cotinine for smoking or 4-amino-biphenyl-DNA adducts, are very attractivefor this purpose. Other approaches werediscussed above, in the section on exposuremeasurement. The disadvantage of mostof these methods is that they measure onlyrecent exposure and lifetime exposure willstill be misclassified. The development ofmethods for combining information fromdifferent types of measurements could bevery useful. Also discussed previously in theexposure measurement section, and equallyrelevant here, is the need to assess andallow for measurement error in con-founders and effect modifiers wheneverpossible. Therefore, consideration shouldbe given to mounting validation substudiesto quantify measurement error in importantcovariates.

SuscptibiityVariation within a population in sensitivityto an exposure of interest can be substan-tial. Khoury et al. (56) estimated the pro-portion of susceptible individuals in thepopulation for cigarette-induced cancers at

55Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

HATCHAND THOMAS

several sites; the proportions varied from<1% for oral and esophageal cancer up to13% for cancer of the lung. Bias in riskestimates will arise if individuals with simi-lar exposures but different susceptibilitiesare treated the same. There are a numberof epidemiologic designs for assessing sensi-tivity to environmental exposures. As ameasurement problem, the central issue iswhether the marker for sensitivity being exam-ined is a measurement of the genotype itsef,some host characeristic, or family history.The ability to classify genotypes directly

has profound implications for identifyingsensitive individuals. The obvious diffi-culty is that there are millions of geneticloci, for which only a relatively small num-ber have probes available and only a fewmight be relevant to any particular disease.Thus, some prior knowledge that a locushas a role in the disease process is essentialbefore embarking on a search for interac-tions with possible environmental expo-sures. Even so, the information for identifyinggenetically susceptible individuals mayinvolve invasive and costly tests.

Recognition of phenotypically distinguish-able subgroups of the population that havedifferent baseline risks of disease or sensitivi-ties to environmental exposures can thereforebe very useful for public health protection.The measurement issues that arise here areessentially no different from those for anyother effect modifier, as discussed above.

For family history as a marker of suscep-tibility to a disease, the basic minimalinformation that needs to be collected isthe identification of the family memberswith the disease and the number, ages, andrelationships of family members at risk.This information should be collected sys-tematically for all first-degree relatives (par-ents, siblings, and offspring), and possiblyfor all second-degree relatives. As theobjective is to examine family history as amarker of sensitivity to an environmentalexposure, every effort should be made toobtain exposure information on all relatives,not just the affected ones.

oschial Stress as Confounder,Modifier~, and Mediator

The psychosocial stress that may be associ-ated with exposure to a perceived environ-mental hazard can potentially confound,mediate, or modify any associations betweenthe exposure and disease. Stress mightoperate indirectly and cause exposed indi-viduals to alter risk behaviors. Stress alsocould have an artifactual association withthe end point of concern because of changesin care seeking, diagnostic practices, or self-reported health states. Alternatively, con-cern about environmental exposures couldcause adverse outcomes other than thosepotentially associated with the perceivedhazard. For example, studies around theThree Mile Island and Chernobyl nuclearplants indicate that the perception of dan-ger can increase distress levels or clinicalstates like anxiety and depression (57,58),irrespective of whether radiation-inducedincreases in cancer actually occur.The issue of stress as a confounder, effect

modifier, mediator, indicator of somemethodologic bias-or even as an exposureor outcome-needs to be explicitly addressedin future environmental epidemiologicresearch conducted on sensitized popula-tions. Some relevant methodology has beendeveloped in studies of communities neartoxic wastes to distinguish between biologiceffects of exposure to hazardous substances atsuch sites and either symptoms of stress oraltered symptom reporting (59,60). Thesepreliminary efforts indude use of a scale tomeasure hypochondriasis and stratifiedanalysis of self-reported symptoms to takeaccount of subjects' perception about thesource of pollution. Environmental epi-demiologists need to learn when and how toaddress the issue of psychosocial stress inorder to clarify interpretation of healtheffects studies and to estimate the importanceof stress in its own right. Considerationshould be given to measuring perceived stressand physiologic indicators of stress as well asto collecting data on methodologicalcovariates such as motivation to participate,

interest in receiving health care, and beliefsabout the exposure in question as a cause ofadverse health effects.

Methodologic Needs andRecommendationsThe aspect of study design that involvesmeasurement of variables is critical, espe-cially in fields like environmental epidemi-ology where the risks from exposure arelikely to be small, difficult to detect, and per-haps not dinically significant, yet may be ofpublic health importance. Methodologicresearch in this area should emphasize thefurther development and application ofdosimetric modeling. Existing data sets rep-resenting a range of research problems withinenvironmental epidemiology could be used toassess the gains from dosimetry algorithmscompared with cruder, more conventionalmethods ofexposure assessment.

Dosimetry models invariably will use acombination of questionnaire data, envi-ronmental measurements, and biologicmarkers; this underscores the need fordevelopment and refinement of methodsfor handling multiple measures. Biologicmarkers themselves, as measures of expo-sure, effect, or susceptibility, are an areawhere additional methodologic developmentwould be desirable.A second important aspect of method-

ologic research relates to sensitivity analysesand other approaches for estimating theuncertainty in measurement of exposureand dose. Included in this category wouldbe validation studies to compare a goldstandard with a more error-prone exposuremeasurement in order to allow for correc-tion of bias in the analysis stage of research.Consideration needs to be given to thecosts and benefits of investigating measure-ment error in the primary study or in a sub-study (which could be carried out intemallyor extemally in relation to the primary study).A final area that deserves attention is mea-surement error in covariates, which can be asimportant as measurement error in theexposure or outcome variables. eg

REFERENCES

1. National Research Council. Biologic markers in reproductive toxi-cology. Washington DC: Nationl Academy Press, 1989.

2. National Research Council. Human exposure assessment for air-borne pollutants: advances and opportunities. Washington, DC:National Academy Press, 1991.

3. Wallace LA, Ziegenfus RC. Comparison ofcarboxyhemoglobin con-centrations in adult nonsmokers with ambient carbon monoxidelevels. J Air Pollution Control Assoc 35:944-949 (1985).

4. Lioy PJ. Exposure analysis and assessment for low-risk cancer agents.Int J Epidemiol 19:S53-S61 (1990).

5. Wallace L Pellizzari E, Hartwell TD, Perritt R, Sparacno C, Whitmore R,Sheldon L, Zelon H. TheTEAM (total exposure assessment methodology)study personal exposures to toxic substances in air, drinking, and breathof400 residents ofNJ, NC, and ND. Environ Res 43:290-307 (1987).

6. Dosemeci M, Wacholder S, Lubin JH. Does nondifferential mis-classification of exposure always bias a true effect toward the nullvalue? Am J Epidemiol 132:746-748 (1990).

7. Brenner H, Savitz DA, Jockel KH, Greenland S. The effects ofnon-differential exposure misclassification in ecological studies. Am JEpidemiol 135: 85-95 (1992).

56 Environmental Health Perspectives SupplementsVolume 101, Supplement 4, December 1993

MEASUREMENTISSUES

8. Rosner B, Spiegelman D, Willett WC. Correction of logistic regres-sion relative risk estimates and confidence intervals for measurementerror: the case of multiple covariates measured with error. Am JEpidemiol 132:734-45 (1990).

9. Somervaille LJ, Chettle DR, Scott MC, Tennant DR, McKiernanMJ, Skilbeck A, Trethowan WN. In vivo tibia lead measurementsas an index of cumulative exposure in occupationally exposed sub-jects. Br J Ind Med 45:174-181 (1988).

10. Willett WC, Sampson L, Stampler MJ, Rosner B, Bain C, WitschiJ, Hennekens CH, Speizer FE. Reproducibility and validity of asemiquantitative food frequency questionnaire. Am J Epidemiol122:51-65 (1985).

11. Kang HK, Watanabe KK, Breen J, Remmers J, Conomos MG, StanleyJ, Flicker M. Dioxins and dibenzofurans in adipose tissue of USVietnam veterans and controls. AmJ Pub Health 81:344-349 (1991).

12. Centers for Disease Control Veterans' Health Studies. Serum2,3,7,8-tetrachlorodibenzo-p-dioxin levels in U.S. Army Vietnam-era veterans. J Am Med Assoc 260:1249-54 (1988).

13. Stevens W, Thomas DC, Lyon JL, Till J, Kerber RA, Simon SL,Lloyd RD, Elghany NA and Preston-Martin S. Leukemia in Utahand radioactive fallout from the Nevada Test Site: a case-controlstudy. J Am Med Assoc 264:585-591 (1990).

14. Hatch M, BeyeaJ, NievesJ, SusserW. Cancer near the Three Mile Islandnuclear plant: radiation emissions. Am J Epidemiol 132:397-412(1990).

15. Brown LM, Pattern LM, Blot WJ. Lung cancer in relation to envi-ronmental pollutants emitted from industrial sources. Environ. Res.34:250-261 (1984).

16. Eifler CW, Stinnett SS, Buffler PA. Simple models for community expo-sure from point sources of pollution. Environ Res 25:139-146 (1981)

17. Wrensch M, Swan S, Murphy PJ, Liscomb J. Hydrogeologic assess-ment of exposure to solvent-contaminated drinking water: preg-nancy outcomes in relation to exposure. Arch Environ Health45:210-216 (1990).

18. Crump KS, Ng TH, Cuddihy RG. Cancer incidence patterns inthe Denver metropolitan area in relation to the Rocky Flats plant.Am J Epidemiol 126:127-135 (1987).

19. Wertheimer N, Leeper F. Electrical wiring configuration and child-hood cancer. Am J Epidemiol 109:273-284 (1979).

20. Milham S. Mortality from leukemia in workers exposed to electri-cal and magnetic fields. N Engl J Med 307:249 (1982).

21. Marshall JR, Graham S. Use of dual responses to increase validityof case-control studies. J Chron Dis 37:125-136 (1984).

22. Hui SL, Walter SD. Estimating the error rates of diagnostic tests.Biometrics 36:167-171 (1980).

23. Wallace LA. The total exposure assessment methodology (TEAM)study. Washington DC: U.S. Environmental Protection Agency, 1987.

24. Smith TJ. Exposure assessment for occupational epidemiology. AmJ Ind Med 12:249-268 (1987).

25. Hulka B, Wilcosky T, Griffith J. Biological markers in epidemiol-ogy. New York: Oxford University Press, 1990.

26. Stein Z, Hatch M. Biological markers in reproductive epidemiology:prospects and precautions. Environ Health Perspect 74:67-75 (1987).

27. Everson RB, Randerath E, Santella RM, Cefalo RC, Avitts TM,Randerath K. Detection of smoking-related covalent DNA adductsin human placenta. Science 231:54-57 (1986).

28. Hatch M, Warburton D, SantellaR Polycyclic aromatic hydrocarbon-DNA adducts in spontaneously aborted fetal tissue. Carcinogenesis11:1673-1675 (1990).

29. Sposto R, Stram DO, Awa AA. An investigation of random errorsin the DS86 dosimetry using data on chromosome aberrations andsevere epilation (RERF TR 7-90). Hiroshima: Radiation EffectsResearch Foundation, 1990.

30. Meyer F, Verreault R. Erythrocyte selenium and breast cancer risk:brief report. Am J Epidemiol 125:917-919 (1987).

31. Vineis P, Caporaso N. Applications of biochemical epidemiology inthe study ofhuman carcinogenesis. Tumorigenesis 74:19-26 (1988).

32. Wald N, Boreham J, Bailey A. Serum retinol and subsequent riskof cancer. Br J Cancer 54:957-961 (1986).

33. London SK, Thomas DC, Bowman JD, Sobel E, Peters JM.Exposure to residential electric and magnetic fields and risk of child-hood leukemia. Am J Epidemiol 134:923-937 (1991).

34. Rubin J. Multiple imputation for non-response in surveys. NewYork: John Wiley, 1987.

35. Rothman K, Poole C. A strengthening program for weak associa-tions. Int J Epidemiol 17 (Suppl):955-959 (1988).

36. Antonarakis SE, and the Down's Syndrome Collaborative Group.Paternal origin ofthe extra chromosome in trisomy 21 as indicated byanalysis ofDNA polymorphisms. N EnglJ Med 324:872-876 (1991).

37. Polissar L. The effect of migration on comparison of disease rates ingeographic studies in the United States. AmJ Epidemiol 111:175-182(1980).

38. Cohen BL. Ecological versus case-control studies for testing a lin-ear-no threshold dose-response relationship. Intl J Epidemiol19:680-684 (1980).

39. Preston-Martin S, Thomas DC, White SC, Cohen D. Prior expo-sure to medical and dental x-rays related to tumors of the parotidgland. J Natl Cancer Inst 80:943-949 (1988).

40. Lawrence CE, Taylor PR, Trock BJ, Reilly AA. Trihalomethanes indrinking water and human colorectal cancer. J Natl Cancer Inst72:563-68 (1984).

41. Armentrout DW. Overview of EPA major air data bases. In:Environmental epidemiology (Kopfler FC, Craun GF, eds). Chelsea,MI: Lewis Publishers, 1986; 149.

42. Armstrong BG. The effects of measurement errors on relative riskregressions. Am J Epidemiol 132:1176-1184 (1990).

43. Greenland S. Statistical uncertainty due to misdassification: implica-tions for validation substudies. J Clin Epidemiol 41:1167-1174 (1988).

44. Schoenberg J, Klotz J, Wilcox HB, Nicholls GP, Gil-del-Real MT,Stemhagen A, Mason TJ. Case-control study of residential radon andlung cancer among New Jersey women. Cancer Res 50:6520-6524(1990).

45. Brunekreef B, Noy D, Clausing P. Variability ofexposure measurementsin environmental epidemiology. AmJ Epidemiol 125:892-898 (1987).

46. Beyea J, DeCicco J. Reestimating the noble gas releases from theThree Mile Island accident. Health Phys (submitted).

47. Doll R. Are we winning the fight against cancer? An epidemiologi-cal assessment. Eur J Cancer 26:500-508 (1990).

48. Davis DL, Hoel D, Lopez A. International trends in cancer mortal-ity in France, West Germany, Italy, Japan, England and Wales, andthe USA. Lancet 336:474-481 (1990).

49. Sorsa M, Ojajarvi A, Salomaa S. Cytogenetic surveillance ofwork-ers exposed to genotoxic chemicals: preliminary experiences from aprospective cancer study in a cytogenetic cohort. Teratog CarcinogMutagen 10:215-221 (1990).

50. Prentice RL. Surrogate endpoints in clinical trials: definition andoperational criteria. Stat Med 8:431-440 (1989).

51. Cartwright R, Ahmad RA, Barham-Hall D, Glashan RW, RogersHJ, Higgins E, Kahn MA. Role of N-acetyltransferase phenotypesin bladJer carcinogenesis: a pharmacogenetic epidemiologic approachto bladder cancer. Lancet 2:842-845 (1982).

52. LewalterJ, Korallus V. Blood protein conjugates and acetylation ofaro-matic amines. Int Arch Occup Environ Health 56:179-196 (1985).

53. PhillipsAM, Silbergeld EK Health effects studies ofexposure from haz-ardous waste sites-where are we toda? Am J Ind Med 8:1-7 (1985).

54. Greenland S. The effect ofmisclassifcation in the presence of covari-ates. Am J Epidemiol 112:564-569 (1980).

55. Kupper LL. Effects of the use of unreliable surrogate variables on thevalidity of epidemiology research studies. Am J Epidemiol120:643-648 (1984).

56. Khoury MJ, Flanders WD, Greenland S, Adams MJ. On the mea-surement ofsusceptibility in epidemiologic studies. Am J Epidemiol129:183-190 (1989).

57. Dohrenwend BP, Dohrenwend BS, Kasl SV, Warheit GJ, Bardett GS,Chisholm RF, Goldsteen RL, Goldsteen K, Martin JL. Report of thePublic Health and Safety Task Group on Behavioral Effects. In: Staffreports to the President's Commission on the Accident at Three MileIsland Public Health and Safety Task Force. No 052-003-00732-1.Washington, DC: U.S. Government Printing Office, 1979; 257-308.

58. Dickman S. Chernobyl effects not as bad as feared. Nature 351:335(1991).

59. Ozonoff D, Colten ME, Cupples A, Heeren T, Schatzkin A,Mangione T, Dresner M, Colton T. Health problems reported byresidents ofa neighborhood contaminated by a hazardous waste facil-ity. Am J Ind Med 11:581-597 (1987).

60. Roht LH, Vernon SW, Weir FW, Pier SM, Sullivan G, Reed LJ.Community exposure tO hazardous waste disposal sites: assessingreporting bias. AmJEpidemioll122:418-433 (1985).

Environmental Health Perspectives Supplements 57Volume 101, Supplement 4, December 1993