| Objective To explore the applicability including advanteges and disadvanteges of multiple imputation (Ml) and other missing-data handling methods in simulated data with missing values.Methods Propensity score (PS) and predictive mean matching (PMM) methods in Ml as well as Ad hoc methds and conditional mean imputation method were used and compared in monotone missing pattern about contious variables in cross-sectional data. In monotone missing pattern about categorical variables in cross-sectional data, the logistic regression method in Ml and Ad hoc methods were empoyed and compared. Markov Chain Monte Carlo (MCMC) model in Ml compared with Ad hoc methods were used in arbitrary missing pattern about contious variables in cross-sectional data, and MCMC model in Ml compared with Ad hoc methods and LOCF were used in longitudinal contious data with missing values. Finally, to impute some variables with missing values in a survey of maternal and children health, the MCMC method in Ml was employed.Results Ad hoc methods works as missing rate is less than 10% in all kinds of data in this research, and the conditional mean imputation method is appropriate when missing rates are from10% to 20% of monotone missing pattern about contious variables in cross-sectional data. While missing rates are from 10% to 20% of monotone missing pattern about contious variables in longitudinal data, LOCF is a better choise. To deal with the contious variables in cross-sectional data with no more than 60% missing-rate or the categrical variables in cross-sectional data with no more than 40% missing-rate or the longitudinal contious data with no more than 50% missing-rate, Ml is the best choise of these methods, otherwise neither is appropriate.Conclusion Although some traditional methods have kind of advantages in some special condition about treating missing data, Ml is able to solve a variety of problems in missing data sets and improve the statistical power.
|
PDF Full Text Request