Handling Of Missing Data, Bayesian Model

In the medical study, the issue of missing data frequently confronts the researchers when they abstracted data from the patient's medical records. The clinical missing data implicate information of great value of a medical diagnosis and study, so how to deal with the missing of medical data problem becomes an important research subject. In the realistic, the phenomenon of missing data comes from many reasons. Toward the different background, taking different methods of dealing with missing data will have a great impact on statistical analysis work. The traditional methods of repairing missing data value are to delete the medical records that containing the missing data, and to replace all missing data by "zero", and replace missing numeric data by the mean value of the complete data. However, it always can't get the satisfied analysis results. In view of the different missing data problems, Little and Rubin defined three different kinds of missing data mechanism as following:First, missing data completely at random (ideally, MCAR); second, Missing data at random (ideally, MAR); Third, the missing of information (information missing, IM).The purpose of this article is to study how to use methods of repairing medical data value under missing of information (IM) mechanism. We used Markov Monte Carlo random data simulations to examine the properties of three Bayesian model, that by placing a simple prior distribution upon the variable that containing the missing data. In general, the paper also discussed two kinds of multivariable logistic regression model of the structure of data packing method. Finally, as a comparator, we also examined a method of completely observation data analysis model and the other method of equated missing data on a variable with absence of conditions. The results showed that the bias and mean square error of each method depended upon the missing rate of related variable and the missing data mechanism. None of each method has a well imputed effect. However, assuming that firstly by placing a simple prior distribution upon variables, and placing a uniform prior distribution upon the parameter of this prior distribution, it resulted in lower relative bias than the other repairing methods in the majority of situations, and took a great insignificant to medical research. Finally, we take a case analysis by analyzing the influence of risk factors on patients with heart disease mortality to examine each method.