Update of the nlme Package to Allow a Fixed Standard Deviation of the Residual Error
Simon H. Heisterkamp, Engelbertus van Willigen, Paul-Matthias Diderichsen and John Maringwa
, The R Journal (2017) 9:1, pages 239-251.
Abstract The use of linear and non-linear mixed models in the life sciences and pharmacometrics is common practice. Estimation of the parameters of models not involving a system of differential equations is often done by the R or S-Plus software with the nonlinear mixed effects nlme package. The estimated residual error may be used for diagnosis of the fitted model, but not whether the model correctly describes the relation between response and included variables including the true covariance structure. The latter is only true if the residual error is known in advance. Therefore, it may be necessary or more appropriate to fix the residual error a priori instead of estimate its value. This can be the case if one wants to include evidence from past studies or a theoretical derivation; e.g., when using a binomial model. S-Plus has an option to fix this residual error to a constant, in contrast to R. For convenience, the nlme package was customized to offer this option as well. In this paper, we derived the log-likelihoods for the mixed models using a fixed residual error. By using some well-known examples from mixed models, we demonstrated the equivalence of R and S-Plus with respect to the estimates. The updated package has been accepted by the Comprehensive R Archive Network (CRAN) team and will be available at the CRAN website.
Received: 2016-08-25; online 2017-05-10@article{RJ-2017-010, author = {Simon H. Heisterkamp and Engelbertus van Willigen and Paul- Matthias Diderichsen and John Maringwa}, title = {{Update of the nlme Package to Allow a Fixed Standard Deviation of the Residual Error}}, year = {2017}, journal = {{The R Journal}}, doi = {10.32614/RJ-2017-010}, url = {https://doi.org/10.32614/RJ-2017-010}, pages = {239--251}, volume = {9}, number = {1} }