Meta-Analytic Combining of Related Regressions

Robert L. Wolpert, Chengchang Li
Duke Univeristy

Apr 25 1993

We consider the problem of synthesizing evidence about the relationship between an explanatory variable and a response variable from related regression experiments, when one or more confounding variables are included in each experiment. If each experiment reports the values of the same confounding variables, the synthesis is shown to be routine. If they each employ difference confounding variables, we show how additional model structure and assumptions will again allow for a synthesis, and we identify the statistics that must be reported from each study to permit it. Finally, we show that intuitively appealing methods for deriving consensus or pooled regression estimates can be seriously misleading if they are drawn only from the reported regression coefficients (rather than from the whole data set) and if the different experiments sue different models. the motivating problem is that of assessing the effect of environmental lead on infant mental development, on the basis of reports from several current and recent longitudinal studies.


PDF icon 1993-16.pdf