Terminology: Regression, ANOVA, ANCOVA

In every field, it's essential to use the proper terminology in order to be understood and to inspire confidence. In statistics, some analyses can be described correctly in different ways. This can be viewed as liberating or as evidence of a sinister plot according to one's general outlook on life.

An example of this overlap in nomenclature occurs with

These methods are used to model a numerical response variable (such as cholesterol level) in terms of predictor variables. The name of the method depends on whether the predictors are numerical, categorical (for example, rice oil/canola oil/peanet oil), or both.

One simple rule states that

Sometimes additional criteria are applied when both qualitative and quantitative predictors are used.

Most analysts would say the name ANCOVA should be used only when the model does not include interactions between the covariates and the factor of interest. Thus, a strict ANCOVA model is a "parallel slopes" model, so that the regression coefficients for the covariates are the same for all factor levels. When an author says that an ANCOVA model was fitted, assume no allowance was made for an interaction between the factor and covariates unless there is a statement to the contrary.

The name of the analysis is not always the the name of the computer program that performs it. Any ANOVA or ANCOVA can be performed by using an ordinary regression package if one is clever about constructing the proper sets of indicator variables. One can and should report that an ANOVA or ANCOVA was performed even when a regression program is used to do it.

[back to LHSP]

Copyright © 1998 Gerard E. Dallal