全部版块 我的主页
论坛 计量经济学与统计论坛 五区 计量经济学与统计软件 HLM专版
1721 6
2014-04-15
Are there standard algorithms (as opposed to programs) for doing hierarchical linear regression? Do people usually just do MCMC or are there more specialized, perhaps partially closed form, algorithms?
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

全部回复
2014-4-15 05:02:08
There's Harvey Goldstein's iterative generalized least-squares (IGLS) algorithm for one, and also it's minor modification, restricted iterative generalized least-squares (RIGLS), that gives unbiased estimates of the variance parameters.

These algorithms are still iterative, so not closed form, but they're computationally simpler than MCMC or maximum likelihood. You just iterate until the parameters converge.

  • Goldstein H. Multilevel Mixed Linear-Model Analysis Using Iterative Generalized Least-Squares. Biometrika 1986; 73(1):43-56. doi: 10.1093/biomet/73.1.43
  • Goldstein H. Restricted Unbiased Iterative Generalized Least-Squares Estimation. Biometrika 1989; 76(3):622-623. doi: 10.1093/biomet/76.3.622
  • Stephen W. Raudenbush, Anthony S. Bryk. Hierarchical linear models: applications and data analysis methods. (2nd edition) Sage, 2002.
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

2014-4-15 05:02:58
The lme4 package in R uses iteratively reweighted least squares (IRLS) and penalized iteratively reweighted least squares (PIRLS). See the PDF's here:

http://rss.acs.unt.edu/Rdoc/library/lme4/doc/index.html
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

2014-4-15 05:03:37
Douglas Bates and Steven Walker have created a GitHub project whose goal is to use pure R code to implement the PIRLS algorithm above. github.com/lme4/lme4pureR. If you consider the base lmer() function in the lme4 package of R you would normally have to read through a whole bunch of C++ code understand the implementation of PIRLS in lmer() (which may be challenging for those of us not so well versed in C++ programming). –  Chris
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

2014-4-15 05:05:23
Another good source for "computing algorithms" for HLM's (again to the extent that you view them as similar specifications as LMM's) would be:

McCulloch, C., Searle, S., Neuhaus, J. (2008). Generalized Linear and Mixed Models. 2nd Edition. Wiley. Chapter 14 - Computing.
Algorithms they list for computing LMM's include:

EM algorithm
Newton Raphson algorithm
Algorithms they list for GLMM's include:

Numerical quadrature (GH quadrature)
EM algorithm
MCMC algorithms (as you mention)
Stochastic approximation algorithms
Simulated maximum likelihood
Other algorithms for GLMM's that they suggest include:

Penalized quasi-likelihood methods
Laplace approximations
PQL/Laplace with bootstrap bias correction
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

2014-4-15 05:06:17
If you consider the HLM to be a type of linear mixed model, you could consider the EM algorithm. Page 22-23 of the following course notes indicate how to implement the classic EM algorithm for mixed model:

http://www.stat.ucla.edu/~yuille/courses/stat153/emtutorial.pdf
二维码

扫码加我 拉你入群

请注明:姓名-公司-职位

以便审核进群资格,未注明则拒绝

点击查看更多内容…
相关推荐
栏目导航
热门文章
推荐文章

说点什么

分享

扫码加好友,拉您进群
各岗位、行业、专业交流群