1

Someone who I consider an expert and very knowledgeable in mixed models has referred to Grimm, Ram, & Estabrook's "Growth Modeling: Structural Equation and Multilevel Modeling Approaches", recommending nlme over lme4. Specifically, they state that lme4 will perform poorly on non-normal outcomes. I have searched around online and have been struggling to find papers, blogs, etc. on this topic on when one mixed modeling R package may outperform the other.

Does anyone have insight on this topic? Does lme4 truly perform worse on non-normal outcomes? If so, why? If not, does one package perform better than the other package for mixed modeling, under certain conditions? When might one expect different results between the packages, or better results for one package over another?

kjetil b halvorsen
  • 63,378
  • 26
  • 142
  • 467
JElder
  • 189
  • 6
  • 1
    Please say more about what you mean by "non-normal outcomes." Do you mean non-normal outcomes of the type typically handled by generalized linear models (binomial, Poisson counts, etc.)? Or do you mean when the errors in a linear regression model aren't actually distributed normally? – EdM Oct 08 '21 at 17:31
  • I am not the source, so I can't be 100% certain-- but the way it was communicated to me the former-- regarding non-normal outcomes-- and not the latter-- about residuals. So this person described lme4 as performing poorly with non-normal outcomes, such as in a generalized linear model, not non-normal residuals. – JElder Oct 08 '21 at 23:15
  • Grimm book states, "Throughout this book we discuss the nlme procedure (over lme procedure) because of its ability to fit inherently nonlinear models. The lme4 package is a newer package for fitting linear and nonlinear mixed effects models in R and is able to fit mixed-effects models to non-normal outcomes (an advantage over nlme); however, nlme is more flexible when it comes to fitting inherently nonlinear models and its programming is more straightforward. For these reasons, we focus on nlme instead of lme4." – JElder Oct 08 '21 at 23:19
  • I was told by person I'm corresponding with that this is a typo, this person has corresponded with others in the field regarding the issue, and that it is lme4 that performs worse for non-normal outcomes. Any thoughts? Feedback? – JElder Oct 08 '21 at 23:20
  • 1
    Does this answer your question? [How to choose nlme or lme4 R library for mixed effects models?](https://stats.stackexchange.com/questions/5344/how-to-choose-nlme-or-lme4-r-library-for-mixed-effects-models) So far as I understand, the `nlme` package cannot handle generalized linear models at all. I don't know what's meant by "inherently nonlinear models," and many find `lme4` syntax to be easier. The `nlme` package provides more flexibility in specifying error covariance structures and for generalized least squares; maybe that's what's meant. – EdM Oct 09 '21 at 13:46
  • It's possible that by "inherently nonlinear models" the authors were speaking about models where the _predictors_ have an association with outcome that can't be translated into a model that is linear in regression coefficients. The `nlme` name stands for "_nonlinear_ mixed effects," and can fit such models. The `lme4` package also handles "inherently nonlinear models" in that sense. I don't have any experience to know if one package or the other works "better" for such nonlinear models. – EdM Oct 09 '21 at 15:14

0 Answers0