greater is the likelihood that confounding or recall bias may explain the association (Wynder, 1987; Khoury et al., 1992a).
Confounding can result in the estimate of the relative risk being biased either toward or away from the null value of 1.0 (no association). By definition, confounding factors have to be risk factors for the disease under consideration. Confounding with respect to reproductive outcomes is a difficult issue. For some outcomes, such as birth defects, there are few suspected, let alone established, risk factors (e.g., maternal age and Down's syndrome). The possibility of unmeasured confounding could therefore play a role in explaining some of the associations reported. Conversely, it is possible that in some situations, confounding could mask a stronger association. The extent of confounding in the studies examined for this chapter is uncertain.
Biased recall of exposure or outcome is another potential problem in reproductive epidemiology studies. Misclassification of exposure to herbicides has been discussed in Chapter 5. Misclassification of outcome can be a problem leading to either an under- or an overestimate of the true relative risk. For certain reproductive outcomes and childhood conditions (e.g., specific birth defects), accurate recall of the outcome may be difficult, especially if the pregnancy or event occurred in the distant past. If both the study and the comparison groups (e.g., veterans versus nonveterans) have similarly flawed recall, the relative estimate will be reduced toward 1.0. If, however, recall differs between the groups, a biased estimate may be obtained. Medical record verification of many reproductive and childhood health conditions is needed to minimize this potential bias.
The statistical power of reproductive epidemiology studies to detect an elevated relative risk, if one exists, should also be borne in mind when interpreting the evidence (see Chapter 5). In a cohort study evaluating herbicide exposure (e.g., among occupational groups or Vietnam veterans) and spontaneous abortion, approximately 266 total pregnancies would have to be studied to detect a doubling of risk (relative risk = 2, alpha = .05, beta = .80). For other outcomes, the sample sizes (exposed and unexposed groups combined) required to detect a doubling of risk would be 656 live births for low birthweight; 2,478 live births for all major birth defects; 16,932 live births for the most common major birth defect; 17,902 live births for chromosomal abnormalities; and 1,856 live births for infant death. Some of the studies reviewed had adequate statistical power for assessment of some of the more common reproductive outcomes such as spontaneous abortion, but power may have been lacking for rarer outcomes such as specific birth defects. When evaluating a given study finding, examination of the confidence interval around the point estimate of the relative risk will provide guidance as to the degree of precision and study size.