Helper Functions for Regression Analysis.

reghelper

The reghelper R package includes a set of functions used to automate commonly used methods in regression analysis. This includes plotting interactions, calculating simple slopes, calculating standardized coefficients, etc.

Version 1.1.2 has been released. The most recent stable release is available on CRAN, and can be installed like so:

install.packages("reghelper")

You can also install the stable release from Github:

install.packages("devtools")
devtools::install_github("jeff-hughes/[email protected]")

If you would like to install the latest development version, you can do so with the following code:

install.packages("devtools")
devtools::install_github("jeff-hughes/reghelper")

Installation Issues

Networked computers can sometimes result in installation issues, as the install_github function sometimes has difficulty with networked directories. If this happens to you, use the .libPaths() function to find the path to your R libraries. That will likely give you a path starting with two backslashes, but you will need to convert that to a path starting with a drive letter (e.g., ‘C:’, ‘D:’). From there, use the following code:

install.packages("devtools")
devtools::install_github("jeff-hughes/reghelper", args=c('--library="N:/path/to/libraries/"'))

Obviously, change the path to the path where your R libraries are stored.

Summary of helper functions

• beta Calculates standardized beta coefficients.
• build_model Allows variables to be added to a series of regression models sequentially (similar to SPSS).
• ICC Calculates the intra-class correlation for a multi-level model.
• cell_means Calculates the estimated means for a fitted model.
• graph_model Easily graph interactions at +/- 1 SD (uses ggplot2 package).
• sig_regions Calculate the Johnson-Neyman regions of significance for an interaction.
• simple_slopes Easily calculate the simple effects of an interaction.

The table below shows the current types of models for which each function has been implemented:

Removal of beta() function for multilevel models (nlme and lme4)

Strategies for how to center Level 1 variables in multilevel models has been a topic of discussion among methodologists, with a variety of approaches suggested. Two of the most common approaches are grand-mean centering vs. group-mean (or within-cluster) centering (Bryk & Raudenbush, 1992; Enders & Tofighi, 2007). This latter approach could actively be impaired by standardizing the variable after applying the group-mean centering strategy, potentially leading to misinterpretation of the model effects. Some researchers have suggested that Level 1 effects should be standardized within groups (i.e., using the group mean and group standard deviation; Schuurman et al., 2016), but this is not a universally accepted practice and strategies vary. Thus, there is no agreed-upon method of standardizing the random effects of a model, and applying one approach in an R function could have deleterious impacts on interpretation.

Standardizing the fixed effects is less contentious, but there is no restriction on what level variables can be found in the model, and in some cases, variables may be included as both fixed and random effects. Other complications in interpretation could arise if a model includes cross-level effects (cf. Aguinis, Gottfredson, & Culpepper, 2013), custom covariance structures, etc. Because of this, there is no real “default” approach for which variables to standardize in a multilevel model that will be suitable for all model structures, and the researcher should give careful consideration to this based on the structure of their particular model, and relevant theoretical considerations (Enders & Tofighi, 2007; Nezlek, 2012).

Because of this, defining a beta() function in a consistent way for multilevel models (nlme and lme4), in a way that would not lead to potential misinterpretation and/or unexpected behaviour, proves difficult. While the initial implementation of this function aimed to standardize just the fixed effects and the outcome variable, this proved insufficient—in cases where the same variable appears as a fixed effect and random effect, what should happen to this variable? Surprising users with models where some variables end up standardized and others are not is not a suitable result.

As a consequence, the beta() function was deprecated in v0.3.6 and later removed. It will still remain implemented for lm, glm, and aov models. Researchers who wish to obtain standardized effects for their multilevel models should manually standardize the variables that are appropriate given their model structure and theory. However, consideration should be given to the centering strategy used in their model as well. For further discussion on these issues, please see the references below.

References

Aguinis, H., Gottfredson, R. K., & Culpepper, S. A. (2013). Best-practice recommendations for estimating cross-level interaction effects using multilevel modeling. Journal of Management, 39(6), 1490-1528. http://dx.doi.org/10.1177/0149206313478188

Bryk, A. S., & Raudenbush, S. W. (1992). Hierarchical linear models: Applications and data analysis methods. Sage Publications.

Enders, C. K., & Tofighi, D. (2007). Centering predictor variables in cross-sectional multilevel models: A new look at an old issue. Psychological Methods, 12(2), 121-138. http://dx.doi.org/10.1037/1082-989X.12.2.121

Nezlek, J. B. (2012). Multilevel modeling for psychologists. In H. Cooper (Ed.), APA handbook of research methods in psychology (Vol. 3). (pp. 219-241). American Psychological Association. http://dx.doi.org/10.1037/13621-011

Schuurman, N. K., Ferrer, E., de Boer-Sonnenschein, M., & Hamaker, E. L. (2016). How to compare cross-lagged associations in a multilevel autoregressive model. Psychological Methods, 21(2), 206–221. https://doi.org/10.1037/met0000062