possible to separate the effects of o-toluidine and aniline because these expo-sures occurred in the same area and both chemicals were likely to have the bladder as a target organ (68).
4.4. Measurement Error and Misclassification
Exposure measurement error is an inherent part of epidemiological studies because of the way information is obtained. Table 2 summarizes the sources of information for epidemiological studies and their advantages and limita-tions. No source of information can be considered absolutely accurate; even
‘‘objective’’ measurements such as levels of a contaminant in the environ-ment or in biological fluids may be affected by sampling method, biological variability, or laboratory error (69). The terms ‘‘measurement error’’ and
‘‘misclassification’’ both refer to any discrepancy between the true value of a variable x and its measured value z, although the term misclassification is more often used with categorical variables and measurement error with continuous variables (69). Errors may be systematic or random; systematic errors refer to errors that are not distributed randomly around the true value (69). Both systematic and random errors may be ‘‘differential’’ or
‘‘nondifferential’’ with respect to disease status. In nondifferential misclassi-fication, the probability and=or direction of misclassification differs between those with disease and those without, as might occur, for example, if an interviewer who knew the health status of the subjects and the study hypo-theses probed more intensely when asking about these exposures in case compared to control interviews. Nondifferential misclassification can intro-duce serious bias in the study results, but can often be avoided by a good study design, i.e., blinded assessment of the study variables (69). Prior to the late 1980s, it was thought that nondifferential exposure measurement error or misclassification would bias studies toward the null (i.e., in the direc-tion of finding no effect), but it is now recognized that there are excepdirec-tions to this rule (71).
Where possible, epidemiologic studies try to minimize measurement error and also to estimate its magnitude. For example, in classifying subjects with respect to current smoking status, self-reported data may be compared to the serum cotinine level; for self-reported exposures, data reported at two different times may be evaluated for consistency; in studies where laboratory analyses are done, blinded split sample analysis and spiked samples with known standard compounds may be used to estimate laboratory error.
Table 2 Commonly Used Assessments for Exposures in Epidemiological Studies Source of exposure data Advantages=limitations
Measurement of substance in biological samples
Good methods available for measuring recent exposures (i.e., cotinine and current smoking status) or retrospective exposure to chemical or physical agents with long half-lives (i.e., organochlorines such as DDT) Time period of interest in cancer studies is
often 20–30 years prior to onset of disease;
some shorter half-lived exposures can be detected in stored sera or urine, if available Accessible samples in living individuals
(blood, urine, buccal cells, etc.) may not reflect exposure at the target tissue of interest (i.e., asbestos in lung tissue) Studying intermediate markers such as
hemoglobin or urothelial cell adducts may yield information about biologically effective dose (84)
Interview data Able to gain information about a wide range of risk factors throughout life, which is
generally not possible with any single record source
Interview data are dependent on the accuracy and completeness of participant recall, may lack detail on specific exposures, such as chemicals or medications, and may be influenced by recall bias
Medical records May be an excellent source for confirming self-reported medically related exposures;
medical records are especially valuable for identifying cohorts for follow-up of medical exposures, i.e., children exposed to
diethylstilbestrol during pregnancy
Location and retrieval of medical records may be difficult in retrospective studies;
investigator may need patient’s permission to access medical records
Work history records Generally a good source to identify individuals for cohort studies, provided that they are complete; may or may not contain detailed information about jobs held, which is needed
(Continued)
Ecological studies: There are enormous variations in the incidence of some cancers worldwide; correlations between site-specific cancer incidence and dietary and other risk factors may lead to potential clues about cancer etiology. Studies looking at such correlations on a population level are termed ecological studies; the unit of observation is a group and not an indi-vidual.
Cross-sectional (or prevalence) studies: These studies examine risk fac-tors and the presence of disease or disease markers simultaneously. In the area of cancer risk assessment, cross-sectional studies may provide valuable information on exposure, including biological indicators such as levels of DNA adducts.
Proportionate mortality ratio (PMR) studies: In some instances, it is not possible to enumerate the entire population at risk for a cohort study, but it is possible to identify deaths that have occurred, for example, from a union-based pension plan. Proportionate mortality ratio studies compare the proportion of deaths by cause in the study and the referent populations, with appropriate control for age, calendar time, gender and race.
Table 2 Commonly Used Assessments for Exposures in Epidemiological Studies (Continued )
Source of exposure data Advantages=limitations
to provide a more detailed assessment of exposure
Industrial hygiene monitoring data
Only in rare instances, such as film badge data available for radiation workers, is it possible to reconstruct exposure data for individuals based on their own measured exposures
Exposure reconstruction in cohort studies is based on industrial hygiene samples to characterize exposures by work area, rather than on sampling results for individuals;
sampling data are often incomplete and may not cover specific jobs or time periods;
assumptions are made to extrapolate data to these periods; often there are no air
sampling results available for early decades of operation
Exposure reconstruction for case-control studies is usually based on information provided by the study subject combined with
(Continued)
Case–cohort studies: A variant of the nested case–control study design, in which cases occurring in a study cohort are compared to a sample of the whole cohort (which may include some cases).
6. METHODS FOR COMBINING THE RESULTS