Collapsibility Of Variables And Statistical Inference Of Missing Data

In actual study,when we merge some data,it often appears such a sit-uation:the dominant one in the group is no longer an advantage one in the merged. This kind of phenomenon is the famous Yule-Simpson para-dox.so,study about collapsibility of variables is very necessary.It can help us to point out which variable can be ignored and which one must be considered.It will provide the safeguard to the accuracy of the statistics and analysis.Firstly,this paper gives definitions of simple collapsibility and strong collapsibility of odds ratio over continuous outcome variable and then introduce sufficient and necessary conditions of strong collapsibility of odds ratio.Next we introduce collapsibility of logistic regression coef-ficients with collapsibility of odds ratio similar.We will firstly give def-initions of simple collapsibility、strong collapsibility and consecutive of logistic regression coefficients.Then we give the sufficient and necessary conditions.Finally,we analyze the relationship between the three kinds of collapsibility.In many studies,such as social investigation, medical research, psy-chology experiment research,treatment datum often miss with kinds of rea-sons.For example,respondents refuse to answer and a patient leaves medi-cal test for their own reasons.statistical analysis of missing data have some certain difficulties.In recent years,the research about statistical inference of missing data has become a hot area.We summer up and summarize caus-es of missing data, missing mechanism and model, and present different processing methods under different missing model. Chinese and foreign scholars are used to study the missing data in many different ways.We fo-cus on principal stratification which is pointed out by Rubin.This paper study the missing data due to death and missing with partial compliance with principal stratification.Then,Framework and theory model under this two models will be given.The research of this part can be used for reference by other scholars.