In CDEager/standardize: Tools for Standardizing Variables for Regression in R

knitr::opts_chunk$set(
  collapse = TRUE,
  comment = "#>",
  fig.path = "README-"
)

standardize 0.2.2-9999

Installation

To install the standardize package, call:

install.packages("standardize")

Package use

The standardize package provides tools for controlling continuous variable scaling and factor contrasts. The goal of these standardizations is to keep the regression parameters on similar scales, and to ensure that the intercept (which is the predicted value of an observation when all other coefficients are multiplied by 0) represents the corrected mean (i.e. the predicted value for an observation which is average in every way, holding covariates at their mean values and averaging over group differences in factors). When the predictors are all on a similar scale, there are computational benefits for both frequentist and Bayesian approaches in mixed effects regressions, reasonable Bayesian priors are easier to specify, and regression output is easier to interpret. Take, for example, the ptk dataset included in the package (for which we will create a new ordered factor preheight):

library(standardize)

ptk$preheight <- "Mid"
ptk$preheight[ptk$prevowel == "a"] <- "Low"
ptk$preheight[ptk$prevowel %in% c("i", "u")] <- "High"
ptk$preheight <- factor(ptk$preheight, ordered = TRUE, levels = c("Low",
  "Mid", "High"))

summary(ptk)

Suppose we want to fit a linear mixed effects regression with total consonant duration cdur as the response, place, stress, preheight, the natural log of wordfreq, and speaker-relative speechrate as fixed effects, and random intercepts for speaker. The variables for this regression can be easily placed into a standardized space with the standardize function:

sobj <- standardize(cdur ~ place + stress + preheight + log(wordfreq) +
  scale_by(speechrate ~ speaker) + (1 | speaker), ptk)

sobj

names(sobj)

head(sobj$data)

mean(sobj$data$cdur)
sd(sobj$data$cdur)

mean(sobj$data$log_wordfreq)
sd(sobj$data$log_wordfreq)
all.equal(scale(log(ptk$wordfreq))[, 1], sobj$data$log_wordfreq[, 1])

with(sobj$data, tapply(speechrate_scaled_by_speaker, speaker, mean))
with(sobj$data, tapply(speechrate_scaled_by_speaker, speaker, sd))

sobj$contrasts

sobj$groups

The default settings for standardize have placed all the continuous variables on unit scale, set (named) sum contrasts for all unordered factors, and orthogonal polynomial contrasts with column standard deviations of 1 for all ordered factors. In the case of speechrate, the call to scale_by ensured that rather than simply placing speechrate on unit scale, it was placed on unit scale for each speaker, so that the resulting variable represents speaker-relative speech rate. We can then simply use the formula and data elements of the object returned by standardize to fit the mixed effects regression in this standardized space:

library(lme4)

mod <- lmer(sobj$formula, sobj$data)

summary(mod)

The scaling of the predictors can be controlled through the scale argument to standardize. For example:

sobj <- standardize(cdur ~ place + stress + preheight + log(wordfreq) +
  scale_by(speechrate ~ speaker) + (1 | speaker), ptk, scale = 0.5)

sobj

names(sobj)

head(sobj$data)

mean(sobj$data$cdur)
sd(sobj$data$cdur)

mean(sobj$data$log_wordfreq)
sd(sobj$data$log_wordfreq)
all.equal(0.5 * scale(log(ptk$wordfreq))[, 1], sobj$data$log_wordfreq[, 1])

with(sobj$data, tapply(speechrate_scaled_by_speaker, speaker, mean))
with(sobj$data, tapply(speechrate_scaled_by_speaker, speaker, sd))

sobj$contrasts

sobj$groups

The standardize function works by making use of the function scale from base R, as well as the standardize functions scale_by, named_contr_sum, and scaled_contr_poly. For more details, install the package and see the vignette "Using the standardize Package".

CDEager/standardize documentation built on March 13, 2021, 3:48 p.m.