ESRC Grant: Sample size, Identifiability and MCMC Efficiency in complex random effect models. - Supporting web page


Sample Size and MLPowSim


The MLPowSim software is now available as version 1.0 Beta 1. We must stress here that MLPowSim is free software and comes with no WARRANTY whatsoever. Below you will find the software and manual along with two accompanying datasets. The manual has been written by William Browne, Mousa Golalizadeh and Richard Parker with the software written by William Browne and Mousa Golalizadeh.

MLPowSim executable file.

MLPowSim manual.

fife.txt file.

fife2.txt file.

A few talks have been given on the topic of sample size calculations in multilevel models.

ESRC research methods 2006 talk

ESRC research methods 2008 talk

Amsterdam 2009 talk


NEW !!! Known Bugs


1. There is a bug in the latest version of MLwiN 2.10 with regard the multivariate normal random number generator that means some MLPOWSIM code will cause MLwiN to crash upon attempted execution. The same code will however work fine in version 2.02. This will affect any files created by MLPOWSIM that contain the command MRAN. Fixed in version 2.11.

2. The 1 level Binomial and Poisson models throw up an error message in MLwiN. This can be cured by replacing the file PRE in the discrete sub-directory of the MLwiN install with the following version PRE. . It has also been observed that on some systems were the user does not have administrator rights the macros run much slower in MLwiN 2.10 for binomial and Poisson models as compared to earlier versions of MLwiN. Fixed in version 2.11

3. In R, the use of the PQL method for non-normal responses has been removed from the lmer function in later version of the function. Thus you may find that the PQL method can therefore not be used and so you will need to choose a different estimation procedure.


Other Multilevel Sample Size software


The PinT software (Tom Snijders, Roel Bosker, and Henk Guldemond) that is used for comparison in the maunal can be downloaded from here.

The ML-DEs software package (Cools, Van den Noortgate & Onghena) that has been developed independently from MLPowSim but which also uses MLwiN and simulation to calculate power calculations for multilevel designs is available from here. We hope to compare MLPowSim with ML-DEs in further work.

The OD (Optimal Design) software package (Steve Raudenbush and colleagues) also looks at multilevel power calculations and in particular cluster randomized designs. It can be downloaded from here.


MCMC Efficiency


The link below gives a current draft (as of 25/08/09) of a new edition of the MCMC estimation in MLwiN manual that includes 5 new chapters (21-25) due to changes achieved in this part of the grant and some additional changes to the existing chapters.

MCMC in MLwiN manual.

If you wish to try out chapters 21-25 with the current version of MLwiN the MCMC options menu item is hidden. To access it type MCSH in the command interface window and ignore the error messages.


Here we give drafts of two papers and a couple of talks on these topics:

The use of simple reparameterisations to improve the efficiency of MCMC estimation for multilevel models with applications to discrete-time survival models

William Browne, Fiona Steele, Mousa Golalizadeh, and Martin Green (to appear in JRSS A)

Reparameterisations paper.

MCMC algorithms for structured multivariate normal models

William Browne, Suna Akkol and Harvey Goldstein (submitted)

SMVN paper.

Presentation to the 6th Amsterdam International Multilevel modelling conference, April 2007

Using SMCMC for normal response multilevel models.

Presentation given in shorter forms at IBC and RSS conferences 2008 and at Cambridge Stats Lab December 2008

Simple methods to improve MCMC efficiency in random effect models.


Identifiability


This topic will be the main focus of the remainder of the grant

Return to: Homepage

This page has been visited times since 20/03/09
View My Stats