Missing data in two-way analysis of variance.

We previously described a regression approach to analysis of variance computations that permitted analysis of unbalanced designs and experiments with missing data in two-way (or higher) analyses of variance Am. J. Physiol. 255 (Regulatory Integrative Comp. Physiol. 24): R353-R367, 1988. That approach can only be used correctly under a set of narrow, and relatively uninteresting, circumstances. In fact, in the example we worked, we extended that approach beyond its intended scope and incorrectly computed F statistics for testing hypotheses about the main effects of strain of rat and nephron site in a study of renal Na(+)-K(+)-adenosinetriphosphatase. This paper presents the correct approach, which can be generalized to most situations likely to be encountered when two-way, or higher, analyses of variance are used.