Model Averaging For Logistic Regression With Fragmentary Data

Nowadays,many researchers are committed to using statistics and machine learning methods to dig out effective information from the massive data on the Internet.”Fragmentary data” has become more and more common in various fields: the source channels of data are diverse,but the data of each channel may be missing during the acquisition process.When the data from multiple channels are combined,the data matrix shows”fragmentation”.Regarding ”fragmentary data”,when the dependent variable is continuous,some scholars have proposed a model averaging method based on linear models to deal with it.In this paper,this method is extended to the case of binary dependent variables,and the model averaging method for logistic regression with ”fragmentary data” is established.Specifically,we build different logistic regression models as candidate models for different response modes.Each model has its own estimation.Furthermore,we use CC data(no missing samples)to select weights.The final output of the method is the weighted average of all estimation.We proved the asymptotic optimality of the selected weights and verified the effectiveness of the method through numerical simulation and real data analysis.