(exposure prevalence in a case-control study or disease risk in a cohort study). The criteria for defining the groups, the methods of evaluating their representativeness, and the assessment of bias in study results depend on the design of the study. In case-control studies, subjects are selected with respect to disease status. If the selection process generates an over-or under representation of exposed cases or exposed controls, a bias will be produced in the measurement of association between disease status and exposure status. In cohort studies, subjects are selected on the basis of exposure status at a time when all subjects are disease free; they are then followed over many years. If the completeness of follow-up differs depending on exposure and disease status, bias might result.
Specific aspects of the residential childhood cancer studies that might have led to bias (Tables 5-5 and 5-6) are considered here. The discussion provides an inventory of possible sources of bias but does not suggest that such bias actually exists, at least not of the magnitude to explain the previously reported associations or lack of associations. The focus is on four main issues: (1) representativeness of cases in case-control study, (2) representativeness of controls in case-control studies, (3) participation (nonresponse) rates, and (4) differential mobility of study subjects.
Representativeness of Cases Three epidemiologic designs have been used in the studies of residential exposure and childhood cancer. A standard case-control design was used in the first type of study (Wertheimer and Leeper 1979; Fulton et al. 1980; Tomenius 1986; Savitz et al. 1988; Coleman et al. 1989; Myers et al. 1990; London et al. 1991; Olsen et al. 1993). Children with cancer and children without cancer were identified, and their past exposures were estimated. Cases were selected from population-based or hospital-based registries (e.g., birth certificates), by a random-digit-dialing procedure, or from referrals to friends of the cases.
A nested case-control design was used in the second type of study (Feychting and Ahlbom 1993). A cohort of exposed and unexposed individuals was identified. From this cohort, cases and controls were chosen independent of their exposure status and then compared with respect to their exposure status, as in a traditional case-control study.
The third type of study was a historical cohort study (McDowall 1986; Schreiber et al. 1993; Verkasalo et al. 1993). In these studies, investigators identified all persons living near the electric transmission facilities as their cohort and compared their mortality experience to that of the national population.
The case-control studies used registry records to identify all cases diagnosed or dying during a certain time, living in a specified geographic region, and in a specified age range (Table 5-5). Three data sources were used in the studies. Wertheimer and Leeper (1979) and McDowall (1986) used death certificates. One concern with that source is that disease survival would affect identification (e.g., nonfatal leukemia would not be identified). Because diagnosis and treatment