摘要
文章采用独立样本M-H算法、逐分量M-H算法和切片Gibbs算法,计算Logistic回归模型的后验分布。在蒙特卡洛模拟中,采用每隔15步抽样一次的方法来降低自相关性。通过绘制直方图、路径图、自相关图等来比较三种算法,分析每种算法的优缺点。结果表明:在先验分布都选取正态分布的前提下,三种算法均具有可行性。随着样本量增大,切片Gibbs算法和独立样本M-H算法估计效果相对较差,逐分量M-H估计效果较好,并且采用Lasso算法进行变量选择可以提高抽样效率。
This paper uses independent sample M-H algorithm,M-H algorithm of single-component and slicing Gibbs algorithm to calculate the posterior distribution of Logistic regression model,and then employs the methods of sampling every 15 steps to reduce the autocorrelation in Monte Carlo simulation.Finally,the paper compares the three algorithms by drawing histogram,path diagram and autocorrelogram,etc,and analyzes the advantages and disadvantages of each sampling method.The results show that the three algorithms are feasible under the premise that the prior distributions all select normal distributions,that with the increase of sample size,the estimation effect of slicing Gibbs algorithm and independent sample M-H algorithm is relatively poorer,while the effect of M-H algorithm of single-component is relatively better,and that the sampling efficiency can be improved by using Lasso algorithm for variable selection.
作者
王纯杰
戚顺欣
张洪阳
Wang Chunjie;Qi Shunxin;Zhang Hongyang(School of Mathematics and Statistics,Changchun University of Technology,Changchun 130012,China;Sohool of Mathematics and Systems Sciences,Nanning Normal University,Nanning 530001,China)
出处
《统计与决策》
CSSCI
北大核心
2020年第22期14-18,共5页
Statistics & Decision
基金
国家自然科学基金资助项目(11671054
11571051)。
