Generalized Regression Models in Educational Research

Educational researcher workers often conceive of regression in the limited sense of the additive, linear model, which assumes that the independent variables do not interact'and are linearly related to the dependent variable. This paper first enlarges this conception of regression analysis (without using mathematics beyond simple algebra) and shows several essential similarities of regression and analysis of variance in detecting non-additivity (interaction) and non-linearity. Second, it argues that regression has several practical advantages over analysis of variance and that multivariate regression (or canonical correlation analysis) should be used in place of factor analysis in prediction research involving several independent and several dependent variables. Last, the generalized regression models advocated are illustrated in several pieces of recent, substantive research.