Biostatistics Journal Club: Linear models for characterizing measurement uncertainty and comparing evaluators: a review, with applications 

This journal club, led by Mark Vangel, PhD, Massachusetts General Hospital, considers two models useful for assessing measurement uncertainty. The first of these, one-way random effects ANOVA, possibly unbalanced and with unequal variances, was first proposed by W.G. Cochran in 1938. Although the one-way model is of course one of the simplest of all linear models, the heteroscedastic model is not often considered, and it has some interesting characteristics, including the possibility of multimodal likelihood. Various approaches for weighted means follow from this model, as well as the well-known meta-analysis method of Dersimonian and Laird. The second class of models considered is a two-way mixed-effects cross-classification, with unequal variances, useful for comparing measurements of the same items by several instruments or evaluators. This approach was first proposed by F.E. Grubbs in 1948, and hence the estimates are often referred to as “Grubbs estimators”. Both of these models are special cases of mixed-model ANOVA, and hence they admit a unified presentation. Although not of much current research interest, they remain useful. Because of the specialized nature of the models, and the accommodation of unequal variances, they are typically not discussed in general texts on linear models. The objective of this talk is to present these methods in a unified way, to discuss their historical development, and to demonstrate their utility through examples.

Speaker:

Mark Vangel, PhD, Massachusetts General Hospital

Required Reading:

Estimation of a Common Mean and Weighted Means Statistics