Comparison Of Imputation Methods Based On Value Prediction

Missing data can be found in various research fields.If it is used without processing,it will cause certain difficulties in model selection and research analysis.Therefore,it's necessary to conduct in-depth analysis of the information contained in the missing data,use observable data to restore the real data situation,and effectively deal with the missing values.The purpose of imputation can generally be divided into three kinds,complementing the original data,parameter estimation and forecasting.Therefore,each method has its own advantages and applicable scenarios in dealing with missing values.Some methods are suitable for restoring real variable data;some have good performance in parameter estimation;some when used for forecasting,due to the deviation in imputation and parameter estimation is neutralized,the data after imputing has a good prediction effect.This paper found that based on the multiple linear regression model,the effect of mean imputation was not ideal when used for the coefficient estimation,but when used for the dependent variable estimation,it was better than most common imputation methods.This paper verifies this new point of view through simulation research and empirical analysis.The simulated data has four different proportions of missing under the three missing mechanisms,and mean imputation and the other six methods are used to fill them.The accuracy of the seven filling methods used in the parameter estimation and variable estimation of the multiple linear regression model is judged by a unified evaluation index.In the empirical analysis,based on the score data of football players in the football game,the same missing simulation method as in the simulation research is used to compare the seven missing value filling effects,and the same conclusions are obtained.In the end,this paper did some preliminary research on the theoretical level.