|
Abstract:
|
Analysis of the effects of mediating variables is increasingly common in clinical trials and epidemiology. Methods used to do this in the social sciences, such as structural equation models, are less familiar to biostatisticians than linear regression techniques, but the use of least squares to estimate the mediated effect has not been extended beyond simple models with a single independent and one or two mediating variables. Here, the usual least squares estimates for the mediated effect are extended to any multiple, multivariate or generalized least squares regression. The variance in a new, more precise form and variance estimates, including an unbiased estimate, are derived. While no fault is found with existing methods, least squares can be applied to analysis of mediation more broadly than is generally recognized, and this may lead to wider use of tests for effect mediation.
|
