7 ExposureAssessment Markf. Nieuwenhuijsen Centrefor Researchin Environmental Epidemiology (CREAL), Barcelona, Spain 7.1. INTRODUCTION 7.2. INITIATCONSIDERATIONS OF AN 'Exposure' is a substanceor factor affecting human EXPOSUREASSESSMENT STRATEGY health, either adversely or beneficially. Exposure A major aim of epidemiologicalstudies is to deter- variablesused in practicein epidemiologyusually mine whether or not there is an associationbetween 'true' have to be regardedas approximationsto the a particular substanceof interest,the exposureand exposure of the subjects who are being studied. morbidity and/ormortality. If there is an association, 'true' The accuracy and precision with which it is desirableto be ableto showan exposure-response exposure is being approximated may vary widely (or dose-response)relationship, i.e. a relationshipin 'surrogate' from one exposurevariable to the next. which the rate of diseaseincreases when the level Exposure misclassification and/or measurement of exposure(or dose) increases.This will aid in the error can lead to attenuationin health risk estimates interpretationof such studies. and/or a loss of power. A simple example illustrates Over recentyears there hasbeen increasinginter- the potential effects of exposure misclassification est in the field of exposureassessment causing it to and shows how important a good exposure assess- develop rapidly. We know more than ever to what, ment is. Let's assumea hypotheticalexample of where and how people are exposed and improve- a study with 2000 subjects,of whom 1000 are ments havebeen made to methodsfor assessingthe exposed to an environmental pollutant and the levels of exposure,its variability and the determi- other 1000 are unexposed.In the exposed group, nants.New methods have been developedor newly 200 have developednthedisease, while only 100 applied throughout this field, including analytical, have done so in the unexposedgroup (Table 7.1). measurement, modelling and statistical methods. The relative risk in this caseis 2. However, the risk This has led to a considerable improvement in estimatewould be only 1.5 if 20Voof the subjects exposure assessmentin epidemiological studies, were misclassified (20Vo of the exposed become and therefore improvement in the epidemiological unexposedand vice versa; see numbers in brack- studiesthemselves. ets in Table 7.1). This exampleshows that even if In epidemiology,there are different study designs the exposure misclassification is relatively minor (e.g.cohort, case-control) to assessthe association (20Vo),the effect on the relative risk is consider- between exposure and disease (see Chapter 2). able. Of course,this is a fairly simple example, and All the study designs require exposure estimates the effect of exposure misclassification/measure- or exposure indices to be able to estimate the ment error dependson other factors, including the risk associatedwith the substanceof interest, but type of error model (Classical or Berkson error; they may differ depending on the study design. Nieuwenhuijsen,2003). The design and interpretation of epidemiological Molecular Epidemiology of Chrcnic Diseases, Edited by C. P. Wild, P. Vineis, and S. Garte O 2008 John Wiley & Sons, Ltd 84 EXPOSUREASSESSMENT Thble 7.1 Hypothetical example of the effect of The basic premise of Armstrong's considera- the relativ, exposuremisclassification. The numbers in brackets tions is that it is better, but more expensive, to main studl 'true' 'approxi- are after 20Eoexposure misclassification measure exposurethan to measure Besides mate' exposure.When the correlation between the ti<lnsto be Diseased Non-diseased Total 'approximate' 'true' and the exposure variable is oxposurea Exposed 200(180) (820) 'approximate' 800 1000 high, the loss of power by using the study are, 'true' Non exposed 100(120) 900(880) 1000 rather than the exposure variable is small. and availal Therefore, if the cost per study subject of measur- particularl' Total 300 1700 2000 'approximate' ing exposureis clearly lower than the ple, many 'true' cost ofmeasuring exposure,a study using the use routin 'approximate' studies is often dependenton the exposure assess- measure of exposure will be more subjectsfo ment and therefore needs careful consideration. effrcient. Alternative study designscan be compared tion of ret 'asymptotic Quantification of the relation between exposure by calculating the so-called relative the subjec and adversehuman health effects requires the use efficiency' (ARE), definedas the rario of the sample or biomor of exposure estimates that are accurate, precise, size necessaryto achieveequal power to detect an this case I biologically relevant, apply to the critical exposure association (Armstrong 1996). When r is the cor- of environ 'approximate' 'true' period and show a range of exposure levels in the relation between the and the residence. population under study.Furthermore, there is gener- exposure,C, is the basic cost of including a subject or little in ally also a need for the assessmentof confounders, in the study (e.g. related to assessmentof disease spends hir i.e. substancesthat are associatedwith both expo- status) and C" and C, are the cost of measuring may be in 'approximate' 'true' sure and disease and may bias the study results. and exposure,respectively, per exposure. Assessmentof confoundersshould be in similar study subject, then the ARE can be expressedas: difficult ar detail to the assessmentof exposureindices, since often tend measurementerror in confounders may also affect ARE1,: P[(Ci+ C)l(Ci+ C^)] sonal mor the healthrisk estimates(Nieuwenhuijsen 2003). often not i 'approxi- In epidemiology we often deal with large popu- So, when the correlation between the for many 'true' lation sizes, with the population spread over large mate' and the exposure variable is equal to informatic : 'approximate' distances.This may make estimating exposure for 0.5 (l 0.25), then rhe use of the to rely on 'true' the subjectsin the study more diffrcult, for example variable is more effrcient than the use of the A furtl in environmental epidemiological studies, since variable when the total cost of including subjects individual 'approximate' we cannot go out and visit each of the subjects. by measuring the variable is more approach, Therefore, one often relies on some form of mod- than four times lower than the cost of including individual 'true' elling or surrogate of exposure (e.g. distance to subjectsin the study by measuringthe expo- don is m< a source). Small sample sizes, on the other hand, sure variable. Differences in cost of this or even a group app may allow some more refined exposure assess- larger magnitude can easily occur when the choice subpopult ment, such as personal monitoring. The size of the is between doing personal exposuremeasurements groups" I population determines how refined the exposure on all subjects compared to collecting information sure, and assessmentcould be. Increasing the population on sources,habits and/or occupations by question- are obtair 'fixed' size could allow for cruder exposure estimates, naire, and when the cost of including study mental e1 while smaller population sizes will require more subjects(C') is not too high comparedto the cost of may be d refined exposureestimates to have similar statisti- making exposuremeasurements per se. absenceo 'approximate' cal power. Armstrong (1996) has provided a gen- A decision to use exposurevari- or smoket eral framework that can help individual researchers ables needs to be based on knowledge of the source (sr 'true' to decide which measuresof exposureto include in correlation betweenthese variables and the as playin their study in order to obtain maximum statistical exposure, and often, this information needs to exposure power, and which validity and reliability substud- be obtained in a pilot study preceding the full The unde ies to include to assessthe quality of the exposure study. This adds ro the cost of the total study, each exP( assessmentmethods used in the full studv. and Armstrong (1996) also provides guidancefor character EXPOSUREPATHWAYS AND ROUTES 85 )onsidera- the relative allocation of budget for pilot and the ation. A representativesample of members from ensive, to main study. each exposuregroup can be personally monitored, 'approxi- Besides sample size and costs, other considera- either once or repeatedly.If the aim is to estimate rtweenthe tions to be taken into accountin designinga specific mean exposure,the averageof the exposuremeas- rariable is exposureassessment strategy for an epidemiological urements is then assigned to all the members in rroximate' study are, for example, accessibilityto the subjects that particular exposuregroup. Alternatively, other : is small. and availability of tools and measurementmethods, exposure estimatescan be assignedto the groups, rf measur- particularly for historical assessments.For exam- e.g. data from ambient air pollution monitors in er than the ple, many environmental epidemiological studies the area where the subjects live. Ecological and y using the use routinely collected health outcome data from individual estimatescan be combined,e.g. in the I be more subjectsfor whom only postcode(or zip code) loca- case of chlorination by-products, where routinely :compared tion of residenceis available and no contact with collected trihalomethane measurementsproviding ic relative the subject can be made. Questionnaires,personal ecologicalestimates are sometimescombined with the sample or biomonitoring