An aberration in the manner in which controls are selected may manifest itself as an unusual pattern of exposure among subsets of controls if the faulty selection does not apply equally to all segments of the study base. Often we know from previous research that exposure prevalence varies by subgroup, e.g., men tend to drink more alcohol than women, White smokers tend to smoke more heavily than African-American smokers, leisure-time physical activity is greater among persons of higher socioeconomic status. If some erroneous method of selection has been applied that is similarly problematic for all subgroups of controls, defined by gender, race, age, etc., then the pattern of exposure prevalence across those markers of exposure may be as expected. If, however, the problems in selection are more extreme for some groups than others, or simply affect subgroups differentially, we will observe patterns of exposure comparing subsets of controls that deviate from those that would normally be expected.
To evaluate this possibility, the pattern of exposure among controls must be examined to determine whether it conforms to expectations based on external knowledge of patterns among subgroups. For this exercise to be helpful, there must be some basis for such expectations, ideally empirical evidence of exposure patterns from previous surveys. Even reasonably justified intuitive expectations may be helpful as a benchmark, however, recognizing that deviations between our expectations and the data may be a result of our intuition being incorrect. Health-related behaviors such as diet, alcohol and tobacco use, physical activity, and preventive health behaviors, are frequently considered in population surveys. The predictors of such attributes or behaviors often include social and demographic characteristics such as age, race, education, occupation, or location of residence. Confirming the presence of expected patterns among the controls lends support to the contention that the controls have been properly constituted, as well as some evidence that the exposure was accurately measured.
For example, if we chose controls for a study of physical activity and my-ocardial infarction among women through driver's license rosters, our sampling frame might be quite suitable for younger women, but could be increasingly ineffective with advancing age. As people age, and particularly as they age and become more physically impaired, they may be less inclined to maintain a drivers' license. If the older age groups were increasingly different from the source population in that age range, we might see an aberrant pattern in which physical activity levels did not decline with advancing age among the controls and perhaps even rose with advancing age. This would run counter to the expected patterns of declining physical activity with advancing age, suggesting that we had obtained a sample that was deviant among older age groups.
An empirical application of this strategy comes from a study of serum lycopene (an antioxidant form of carotenoid found in fruits and vegetables) in relation to the risk of prostate cancer (Vogt et al., 2002). A multicenter case-control study was conducted in the late 1980s in Atlanta, Detroit, and 10 counties in New Jersey. Controls were chosen through random-digit dialing for men under age 65 and through the Health Care Financing Administration records for men age 65 and older. Among a much larger pool of participants, 209 cases and 228 controls had blood specimens analyzed for lycopenes. Serum lycopene was inversely associated with risk of prostate cancer and found to be lower among African-American controls as compared to white controls (Table 5.4). To corroborate the plausibility of lower levels among African Americans (who experience a markedly higher risk of prostate cancer generally), the authors examined pertinent data from the National Health and Nutrition Examination Survey. In fact, there is strong confirmatory evidence that African Americans in the United States do have lower lycopene levels than whites across the age spectrum (Fig. 5.1). Other methodological concerns aside, this pattern provides evidence in support of having enrolled reasonably representative African-American and white men into the case-control study.
Internal comparisons could, of course, reveal the patterns that would be expected based on prior information, but still have stratum-specific and overall exposure prevalences that are disparate from that in the study base. If we recruited our controls for the study of physical activity and myocardial infarction by random digit dialing, and had a resulting preference for women who stayed at home across the age spectrum, we might well over-sample physically inactive women with some fraction of such women unable to maintain employment due to limited physical ability. The patterns by age might still be exactly as expected, but with a selectively inactive sample within each stratum and therefore a biased sample overall. Nonetheless, for at least some hypothesized mechanisms of selection bias, we would expect the extent of it to vary across strata of other exposure and disease predictors, and for those candidate pathways, examination of exposure prevalence across subgroups may be useful.
Was this article helpful?