We concur with Dr. Poole that issues of control selection must be considered in evaluating the results fromcase-control studies of electromagnetic field-cancer relations (I). However, we do not agree with his implicationthat the issue is unique to random digit dialing or that the success of this method hinges on the adequacy ofanalytic control for socioeconomic status. Virtually all of the sampling frames typically used for control selectionin population-based case-control studies (e.g., households with telephones, birth certificates, town lists, motorvehicle licensing records, Health Care Financing Administration files) are to some extent an incompleteaccounting of the population study base from which cases typically are ascertained. When completeness variesinversely by socioeconomic status, restricting cases to those who are part of the sampling frame reduces thepotential for bias. Since we used random digit dialing as the source of controls, we excluded cases who did nothave a residential telephone, thus reducing substantially the likelihood of bias due to severe discordance insocioeconomic status of the respective source populations.With respect to nonconcurrent selection of controls, our study is similar to that of Preston-Martin et al. (2) inthat our cases were diagnosed from 1984 to 1990, while enrollment of controls did not begin until 1989. In ourstudy (3), we minimized the potential bias arising from partial nonconcurrence of control selection by requiringthat a control, like a case, be a resident of the study area at the time of his or her assigned reference date. Evenso, Dr. Poole correctly states that bias could still arise because some children who were residents of the study areaduring the case diagnosis period could not have been selected for our study because they had moved (or, far lesslikely, had died). For bias to have occurred, the exposure characteristics of the children who migrated out of thearea must have differed meaningfully from that of the sampled control group, and the relative proportion of thepopulation at risk that is missing due to migration must have been fairly substantial. If, however, this proportionwas small (as we suspect it was for the Seattle-Puget Sound region), even marked exposure differences betweenresidents who migrated out and residents who remained in the study area are unlikely to have much impact onour results.We believe that in most population-based case-control studies the key potential problem with control selection,whether by random digit dialing or some other method, is bias that may arise due to differential nonparticipation(4). Although measuring and adjusting adequately for socioeconomic status can control for confounding amongrecruited subjects, only measurement of the exposure characteristics of the nonparticipants provides informationabout the impact of this source of potential selection bias (5). In addition, methodological studies such as we haveconducted also can help address control nonparticipation bias concerns by providing reasonable bounds on theextent of the problem (6). The results of our methodological study showed that differential nonparticipation ofpersons from low-income households is unlikely to result in substantial bias from the inverse relation that existsbetween household income level and high exposure wire code category in the Seattle area (6).Furthermore, selection bias due to nonparticipation depends not only on exposure differences betweenparticipating and nonparticipating controls, but also on exposure differences between participating and nonpar-ticipating cases. In our study, we measured wire code configurations in a sample of both the nonparticipatingcases and controls (3). Inclusion of these data yielded estimates for the five-level wire code scheme that weresimilar to those for participants alone, with one notable exception: the odds ratio relating brain tumor occurrenceto the highest exposure category (very high current configuration) was 0.5 among the participants, but increasedto 0.9 when nonparticipants were included. Although based on small numbers, this substantial change in theassociation for this wire code category occurred because 13.2 percent of nonparticipating cases were exposed tovery high current configuration, compared with 3.3 percent of the participating cases; for nonparticipating and
Read full abstract
Jan 9, 2025
Nicole C Loroña + 5
Just Published
Open Access