How useful is linear regression analysis in detecting the existence of dose-response relationships in large-scale epidemiologic studies when only a fraction of the population is sensitive? The case of methylmercury

The existence of a dose response in epidemiologic studies is generally dete rmined from the linear regression slope after controlling for covariates, T his approach assumes the entire population is equally sensitive to the toxi cant and that response is a function only of dose and a random error functi on. However, sensitive subpopulations have been identified for a variety of toxicants possibly including methylmercury (MeHg). The study of MeHg expos ure in the Seychelles islands has failed to find significant effects (dose- response slope not significantly different from zero) while other studies h ave found such effects. Using data on the error function in developmental t est scores and MeHg exposure distributions from that study, and assuming pl ausible dose-response relationships for sensitive subpopulations, we conduc ted Monte-Carlo simulations of the power of linear regression analysis to d etect a dose-response relationship from the total sample (n = 700), and to compare dose-response slopes in the total and sensitive populations. Linear regression did not reliably detect a dose-response relationship for most s cenarios when sensitives were 5% of the total and for some scenarios when s ensitives were 10% of the total. We also found that the dose-response slope for the total population underestimated the sensitive dose-response slope in all cases by about an order of magnitude. These findings may have import ant implications for detection and quantification of dose-response relation ships from epidemiologic studies. (C) 2001 Academic Press.