A comparison of NLMIXED and ADMB-RE: Booth et al example



Code and data
SAS: epil.sas

AD Model Builder: epil.tpl and epil.dat

Files needed to reproduce results for starting value 1: epil.zip.

Navigation
ADMB-RE home
Otter Research

SAS PROC NLMIXED and ADMB-RE are currently the only software packages that allow full flexibility in formulation of generalized linear mixed models, and automatic fitting by maximum likelihood. It is therefore of interest to compare the performance of these packages with respect to the following criteria:
  • Numerical accuracy
  • Computational speed
The epilepsy data considered in Venables and Ripley Modern applied statics with S are used. Booth et al. (2005) fit a negative binomial loglinear mixed models to these data using NLMIXED. We consider variations over their model, focusing on different number of random effects and correlation structures.

Standard GLM-notation is used to express linear predictors. Please refer to Booth et al. (2005) for a description of the covariates involved.
# Parameters
Distribution : Negative binomial 1
Fixed effects part : Base*trt+Age+Visit 6
Random effects part : Base 3
Total = 10
The model has two correlated random effects (intercept and slope of covariate Base). SAS and ADMB-RE give almost identical results. Parameter estimates and run times for 2 different starting values are given: here. Both models are run with 10 point Gauss Hermite quadrature.

Conclusion

SAS NLMIXED and ADMB-RE gives almost identical parameter estimates, but ADMB-RE is approximately 4 times faster.