Abstract

The application of multiple regression and path analysis is discussed in regard to the exclusive use of the beta coefficients. Beta is one of the possible ways of controlling for the effects of the remaining predictors, others are part and partial correlation, part and partial covariance. A typology is developed for the difference between total and controlled effect. With this instrument (for controlled effect = beta) it can be shown under which conditions the sum of the total explaining power (r2y,xi) and the sum of the additional explaining power (Part correlation2) do exceed or not Multiple R2. By the criteria of sign rule and of the relative advantages of the decomposition of total effect in controlled effect and rest effect it is argued that beta is still the best way (of those considered) of control. Next some problems of the otherwise recommended LISREL approach are discussed: Estimation, data-fitting versus model test, model-fitting, standardization. Finally some parallels between models for metric and models for non-metric data are specified.