SabreR

SabreR just released an update. It is another software package that can estimate multivariate multilevel model (other options are aML, MCMCglmm, etc.). They seem to also have a book dedicated to the software, which be worth checking out.

It will be great if the author can incorporate some plotting function into the package.

MCMCglmm

This R packages estimates Generalised Linear Mixed Models via MCMC. It provides a number of random error distributions and can be used for multivariate multilevel models (simultaneous equation model).

I will do some tests and compare the results to other packages.

This package has the potential to become the ideal modeling tool for multilevel and multiprocess analysis for Bayesians, just as aML and Sabre for non-Bayesians. I have been hoping the new JAGS can have much improved performance with similar models, but I don't know when the new version (2.0) will be out. I will be interesting to conduct a benchmark test between aML, Sabre, GLLAMM, MCMCglmm, WinBUGS, and JAGS on some complicated multilevel multiprocess statistical models.

Unlike aML and Sabre, MCMCglmm seems to be under active development.

Here are some Bugs code for joint modeling

http://www.biostat.umn.edu/~brad/software.html

Sabre and SabreR

I discovered another multilevel multiprocess software, Sabre (and its R implementation SabreR). It seems to be very efficient and well integrated with R. It is also possible to run it (a parallel version) on a cluster. Looks very promising. I should be able to use it for serious exploration in a day or two.

Unlike aML, Sabre is still under active development, which means that new features, such as more finite mixture modeling capability, hopefully may be incorporated into future releases.

Home page: http://sabre.lancs.ac.uk/
Tutorial: http://sabre.lancs.ac.uk/short_course.html

Intel Fortran and aML

I am happy to discover that aML can be successfully compiled on Ubuntu using the non-commercial Intel Fortran compiler v.11.

Starting value for aML

Feeding aML good starting values is crucial for fast convergence. I followed the suggestions offered by the aML manual for a while. Now I no longer need to struggle with the manual: it becomes a fun game that I am really good at.

The secret is to build up complicated models piece by piece, use estimates from the simpler models as the starting values of the more complicated ones; and use stepwise procedure to get good estimates for the intercepts, then fixed effects, then random effects.

Predict, simulate, and graph aML results

The multilevel and multiprocess software aML is quite amazing. It handles a much greater number of models than most other packages, and it is fast to reach convergence given suitable starting values. However, it does not have lots of facilities for results presentation.

I have used this software for several of my own research. I think it will be highly desirable to streamline the communication between aML and Stata or R. I can think of the following steps:

1. Estimate a mode in aML;
2. Extract point estimates and variance/covariance from the output file using Python;
3. Transfer the extracted results to R;
4. Predict, simulate, and graph the results with Zelig or other facilities built into R.
   

aML reported random effects

The random effects that aML reports are standard deviation and correlation, not variance and covariance.