面板數(shù)據(jù)模型的懲罰似然變量選擇方法研究

摘要： 本文針對(duì)面板數(shù)據(jù)模型的懲罰似然變量選擇問(wèn)題,比較研究了Lasso、Adaptive Lasso、Bridge和SCAD四種罰函數(shù)的漸近性質(zhì)。模擬結(jié)果驗(yàn)證了在面板數(shù)據(jù)情況下,Adaptive Lasso、Bridge和SCAD的Oracle性質(zhì)同樣成立,且它們?cè)谧兞窟x擇準(zhǔn)確性、參數(shù)估計(jì)精度和模型預(yù)測(cè)精度三方面的效果都優(yōu)于Lasso。為了合理選取調(diào)整參數(shù),本文考慮AIC、BIC、GCV、Cp四種準(zhǔn)則,通過(guò)模擬顯示BIC和GCV的表現(xiàn)通常要優(yōu)于AIC和Cp。作為實(shí)證研究,本文在面板數(shù)據(jù)框架下應(yīng)用懲罰似然方法對(duì)上市公司市盈率影響因素進(jìn)行選擇,以期對(duì)股市投資者做出理性投資決策有一定指導(dǎo)價(jià)值。
This paper focuses on the methods of penalized likelihood variable selection for the panel data model,and discusses and compares the asymptotic properties of Lasso,Adaptive Lasso,Bridge and SCAD. Through simulations, Adaptive Lasso,Bridge and SCAD are confirmed to have the oracle property and perform better than Lasso on variable selection accuracy,parameters estimation precision as well as model prediction precision. In addition,to properly select the tuning parameters,we consider the criteria AIC,BIC,GCV and Cpand indicate by simulations that tuning based on BIC or GCV in general do better than based on AIC or Cp. As an empirical study,we apply the penalized likelihood methods to selection of the influencing factors on price-earnings ratio of listed companies under the framework of panel data,in order to provide some references to stock investors in making rational investment decisions.