项目名称: 面板数据分位数回归中的模型选择问题研究
项目编号: No.11301391
项目类型: 青年科学基金项目
立项/批准年度: 2014
项目学科: 数理科学和化学
项目作者: 唐炎林
作者单位: 同济大学
项目金额: 22万元
中文摘要: 本课题主要研究面板数据(面板个数n,面板长度T)分位数回归中的模型选择问题。主要研究内容有:一、研究n固定、T发散的情况,通过复合分位数回归(CQR)提高估计效率,并用ALASSO进行变量选择,证明理论性质;二、研究n发散、T固定的情况,利用QIF方法考虑相关结构,结合CQR,提高估计效率,并通过ALASSO进行变量选择,证明理论性质;三、研究n发散、T固定的高维面板数据,利用QIF方法考虑相关结构,采用LASSO和ALASSO两步惩罚的方法,研究降维和模型选择问题;四、研究带删失的情形,考虑到CQR要求误差同分布的局限,利用自适应LASSO对不同分位数水平上的系数差异进行惩罚,确定系数之间确实存在差异的分位数水平,选择正确的模型,提高估计效率。同时,在本课题的研究中,将通过丰富的计算机模拟,检验所提方法的有限样本性质,并将方法应用到实际问题分析中去。
中文关键词: 面板数据;删失数据;分位数回归;模型选择;假设检验
英文摘要: This project mainly studies model selection problems in quantile regressions, with panel data (n as the number of panels, and T as the number of observations in each panel). The main focuses are as follows. I) We study panle data with fix n and diverging T. We increase the efficiency by composite quantile regression (CQR), and proceed to variable selection by adaptive LASSO penalty. We will prove that the penalized estimator enjoys nice asymptotic properties. II) We study panel data with diverging n and fixed T. We will consider the correlation structure within the panel by QIF method, and combining the CQR, the estimator will be more efficient. The adaptive LASSO will be applied for variable selection, and the asymptotic properties will be proved. III) We study high dimensional panel data with diverging n and fixed T. We will consider the correlation structure within the panel by QIF method, and LASSO and adaptive LASSO will be applied for dimension reduction and model selection. IV) We study quantile regression with censoring. Since the CQR is limited to the case in which ther errors have the same distribution, we will apply adaptive LASSO to penalize the differences in the coefficients at different quantile levels, and determine which quantile levels share the same coefficients and which ones do not. Correct
英文关键词: Panel data;Censored data;Quantile regression;Model selection;Hypothesis testing