Comparison of Multiple Regression and Artificial Neural Network Performances in Determining the Order of Importance of Predictors in Educational Research

Abstract

Studies aiming to determine the importance rankings of one or more predictor variables on the predicted variable are frequently encountered in education. Multiple Regression (MR) and artificial neural network (ANN) are widely used in this type of research. The present study compares the predictive importance rank performances of MR and ANN methods. For this purpose, two separate real data sets, in which MR assumptions are met and the predictor variables are continuous or discrete, and simulation data generated by considering the relationships in these data sets were used. Absolute relative bias (ARB) and mean square errors (MSE) were used to compare the methods’ performances. The results of the research showed that the increase in sample size had an improving effect on the ARBs and MSEs of the methods. In addition, if the predictors are continuous, researchers may be advised to choose either MR or ANN. However, in cases where the predictors are discrete and the number of predictors is three or more, the use of ANN is recommended. In order to obtain optimal estimations with both methods, it is recommended that researchers use a sample size of at least 200. 

Keywords: multiple regression analysis, artificial neural networks, continuous predictor, discrete predictor, order of importance, predictive correlational research.