roberta.perera

Title of QTUG presentation:

The Effects of Collinearity on Composite Power in Structural Equation Modeling and Regression.

List of Author and co-authors for QTUG presentation:

Robert A. Perera & Scott E. Maxwell

Abstract:

The effects of collinearity on the variance of parameter estimates are well known in multiple regression and SEM. Extreme collinearity causes large increases in the variance of parameters leading to low power and decreased precision. The multivariate distribution of parameter estimates has received less attention; as the covariance of the predictors becomes high, the covariance of the parameters becomes highly negative. This implies researchers often reject one null hypothesis, but rarely reject multiple null hypotheses; in a regression model with two predictors, finding a significant result for at least one of the regression coefficients would often occur, but rarely for both simultaneously. Researchers should strive for high composite power (correctly rejecting both null hypotheses).

In order to examine composite power in SEM and regression, a simulation was conducted. Results revealed that the probability of finding at least one significant predictor can be higher than 0.90, while the probability of correctly rejecting both null hypotheses is lower than 0.15. Although regression produces biased results, it exhibited more composite power than fitting the corresponding SEM model. Due to the increased variance in the SEM model, type III errors (rejecting the null hypothesis in the wrong direction) were found to be more prevalent.