README.md

In cvoeten/buildmer: Stepwise Elimination and Term Reordering for Mixed-Effects Regression

The buildmer package

buildmer is an R package written to simplify the process of testing whether the terms in your lmer (or equivalent) models make a significant contribution to the log likelihood, AIC, BIC, or explained deviance (the latter is not a formal statistical test, but informally tests if the model fit improved by at least an ounce of a percent). The aim of the package is to fully automate model selection, as is already possible for non-mixed regression models and lmerTest models using the step function.

In addition, the package (optionally) determines the order of your predictors by their contribution to the model fit. This is intended to mimic the results you might get from the stepwise functions available in, e.g., SPSS (although SPSS support stepwise elimination only for fixed-effect models...). In sum, the buildmer package aims to take the complex and time-consuming parts of the model fitting procedure out of your hands -- all you need to do is specify your intended maximal model and your dataset, and the package will take care of the rest.

Nonconvergence of models is handled properly by removing a random slope if the maximal model you have specified turns out to not converge. p-values are calculated for you using Wald z-scores. Better alternatives (Kenward-Roger, Satterthwaite) are available at the user's option (if the lmerTest package is installed).

The package supports the fitting of a wide variety of models, if the relevant packages are available. The following buildmer functions make it possible to fit the following types of models: * buildmer: lm, glm, lmer, glmer * buildgamm4: gamm4 (package gamm4) * buildbam: bam (package mgcv) * buildgam: gam (package mgcv) * buildgamm: gamm (package mgcv) * buildglmmTMB: glmmTMB (package glmmTMB) * buildmultinom: multinom (package nnet) * buildmertree: lmertree, glmertree (package glmertree) * buildlme: lme (package nlme) * buildclmm: clmm (package ordinal) * buildmer.nb: glm.nb (package MASS) and glmer.nb (package lme4) * buildcustom: enables you to write your own wrapper function to use the buildmer term-reordering and elimination features with any type of model you want.

Automatic elimination of fixed, random, and/or smooth terms, is possible and enabled by default using the backward (default) or forward stepwise method. Bi-directional elimination is also possible, by passing e.g. direction=c('forward','backward','forward'), although I would not want to recommend doing this.

The intention of the buildmer package is to make your life as simple as:

library(buildmer)
model <- buildmer(Reaction ~ Days + (Days|Subject),lme4::sleepstudy)

In other words, the intention of the package is for you to specify the maximal model that you would like to fit, and let the package worry about whether or not your model converges, and whether or not your effect structure before, during, or after non-significant term removal needs to be passed to gamm4, lmer, or lm (or glm(er) or {g,b}am). More options are available; please see the documentation for details.

If certain terms need to be added together, you can construct your own buildmer formula object by calling tab <- tabulate.formula(formula), modifying the tab to suit your needs, and passing it to the call to buildmer() (or another buildmer fitter) in the formula part, with an extra argument dep passing the dependent variable as a character string. This can be used to force certain parameters to only be evaluated together (see the block column in the tab object); this makes it possible to use buildmer to build models with diagonal covariance structures if using a specially-prepared data frame. See the vignette for details.

Installation

remotes::install_github('cvoeten/buildmer')

Known issues

Bugs will be fixed as they are uncovered. If you notice a bug, please file an issue for it and I will look into it if I have time.

cvoeten/buildmer documentation built on March 3, 2023, 3:25 p.m.