Multicollinearity Effect in Regression Analysis: A Feed Forward Artificial Neural Network Approach
C. P. Obite *
Department of Statistics, Federal University of Technology, Owerri, Nigeria.
N. P. Olewuezi
Department of Statistics, Federal University of Technology, Owerri, Nigeria.
G. U. Ugwuanyim
Department of Statistics, Federal University of Technology, Owerri, Nigeria.
D. C. Bartholomew
Department of Statistics, Federal University of Technology, Owerri, Nigeria.
*Author to whom correspondence should be addressed.
Abstract
In this study we compared the performance of Ordinary Least Squares Regression (OLSR) and the Artificial Neural Network (ANN) in the presence of multicollinearity using two datasets – a real life insurance data and a simulated data – to know which of the methods, models a highly correlated dataset better using the Root Mean Square Error (RMSE) as the performance measure. The ANN performed better than the OLSR model for all the different ANN models except the models with nine and ten nodes in the hidden layer for the real life data. The network with four hidden nodes was the best model. For the simulated data, the ANN model with two hidden nodes gave us the least RMSE when compared to the OLSR model and the other ANN models in the testing set. The network with two hidden nodes modelled the data very well. In the presence of multicollinearity, ANN model achieves a better fit and forecast than the OLSR.
Keywords: Multicollinearity, ordinary least squares, artificial neural network, root mean square error.