Methodological issues in modeling the relationship between low-level lead exposure and infant development: examples from the Boston Lead Study.

This paper addresses several methodological issues relevant to an assessment of the association between low-level lead exposure and early development. In particular, we discuss methods for choosing, from a large pool of candidates, the covariates to control when estimating this association. We examine the issue of confounding and explain why adjusting increased, rather than decreased, the estimate of the association between blood lead level and development at 6 months of age in our sample. A step-by-step description of our strategy for model building is presented. Finally, we demonstrate the robustness of the findings by showing that the magnitude and standard error of the estimated lead effect is not affected appreciably by the method of selecting covariates to be controlled for or by the characterization of lead as a continuous, ordinal, or dichotomous variable. Although these issues arose in the course of analyses of data collected by the Boston lead study (D. Bellinger, H. Needleman, A. Leviton, C. Waternaux, M. Rabinowitz, and M. Nichols (1984), Neurobehav. Tox. Teratol., 6, 387-402), they apply to other current prospective lead studies as well.