The Discussion On Solutions Of Multicollinearity In Multilinear Regression Models

Regression analysis is a technology widely used in statistic analysis and forecast fields,such as business administration,economy,society,physic and bioscience etc.In Regression analysis,the phenomenon of multiple correlation between variables often seriously influences the parameter estimation,enlarges the error of models,and damages the stability of models.So,elimination of multiple correlation is very important in parameter estimation of regression analysis.In order to solve the problem of multiple correlation between the multiple variables in multiple regression analysis,three methods are commonly used:Ridge Regression,Principal Component Regression(PCR) and Partial Least Square Regression(PLS).Based on references,the basic theories and properties are firstly introduced and extended in the paper;PLS is compared with ordinary least squares regression,ridge regression and principle components regression ;In the ridge regression we adopt a method of searching ridge parameter ;We substitute weighted sum of squares for sum of squares what filter eigen value's method;The theories of PLS is further explored:it is firstly analyzed from theories and proof whose data are unsuitable to be dealt with using PLS,and a improved method is proposed,validated by case;as PLS model still contains all the original variables,the variables selection is discussed through Path Analysis.