Multi-model Inference For Discrete Choice Model With Network Dependence

Discrete choice model is an important tool for analyzing and studying binary choice variables.It recovers the relationship between the probability of specific behavioral choice of individuals and related explanatory variables under certain conditions and has been widely used in different areas of scientific research.In addition,many macroeco-nomic problems can also be modeled by discrete choice model,so the discrete choice model has also become an important research method in macroeconomics.Moreover,because individuals in social network are not independent and exposed to network de-pendence,they tend to be influenced by other nodes when they make decisions in the network.This kind of interaction among individuals forms a complex network struc-ture.Especially in the social network dependence,if the dependency is ignored,it will lead to the problematic of estimation and prediction.Therefore,it is necessary to con-sider the network dependence in the discrete choice model.In recent years,multi-model inference has been extensively studied by statisticians and econometrists.By conduct-ing model selecting for the discrete choice model with network dependence,we can avoid building a too simple or too complicated model.We aim to study the multi-model inference in binary discrete choice model with network dependence,including model selection and prediction.In order to simplify the calculation,this thesis employs approximated paired maximum likelihood estimation(APML)to estimate the param-eters of binary discrete choice model with network dependence.Two model selection procedures are proposed based on K-fold cross validation for quadratic loss function and Kullback-Leibler(KL)loss function,respectively.When all candidate models are misspecified,under mild regularity conditions,we prove that the proposed model se-lection instruments are asymptotically efficient in terms of quadratic loss or KL loss.Simulation studies are employed to assess the finite sample performance of the pro-posed methods and the results show that when all candidate models are misspecified,as sample size n increases,both Ln((?))/inf1≤s≤S Ln(s)and KLn((?))/inf1≤s≤SKLn(s)tend to approach 1,where(?)and(?)are the labels selected by quadratic loss based cross validation or KL loss based cross validation,respectively.Whereas when the true data generating process(DGP)is nested in the set of candidate models,the proposed methods can identify the true model with high proportion,and the proportion of correct selection by the method based on KL loss function is higher than that based on quadratic loss function.