| LMERConvenienceFunctions-package {LMERConvenienceFunctions} | R Documentation |
Model Selection and Post-Hoc Analysis for (G)LMER Models
Description
The main function of the package is to perform backward selection of fixed effects, forward fitting of the random effects, and post-hoc analyses using parallel capabilities. Other functionality includes the computation of ANOVAs with upper- or lower-bound p-values and R-squared values for each model term, model criticism plots, data trimming on model residuals, and data visualization. The data to run examples is contained in package LCF_data.
Details
| Package: | LMERConvenienceFunctions |
| Type: | Package |
| Version: | 3.0 |
| Date: | 2020-09-27 |
| License: | GPL-2 |
| LazyLoad: | yes |
Author(s)
Antoine Tremblay, Statistics Canada, and Johannes Ransijn, University of Copenhagen
Maintainer: "Antoine Tremblay, Statistics Canada" <trea26@gmail.com>
References
Baayen, R.H. (2008). Analyzing Linguistic Data. A Practical Introduction to Statistics Using R. Cambridge, UK: Cambridge University Press.
Baayen, R.H., Davidson, D.J. and Bates, D.M. (2008). Mixed-effects modeling with crossed random effects for subjects and items. Journal of Memory and Language, 59, 390–412.
Newman, A.J., Tremblay, A., Nichols, E.S., Neville, H.J., and Ullman, M.T. (2012). The Influence of Language Proficiency on Lexical Semantic Processing in Native and Late Learners of English. Journal of Cognitive Neuroscience, 25, 1205–1223.
Newman, A.J., Tremblay, A., Neville, H.J., and Ullman, M.T. (In preparation). The relationship between proficiency and ERP components evoked by grammatical violations in native and late learners of English.
Pinheiro, J.C. and Bates, D.M. (2000). Mixed Effects Models in S and S-Plus. New York: Springer.
Quene, H., & van den Bergh, H. (2008). Examples of mixed-effects modeling with crossed random effects and with binomial data. Journal of Memory and Language, 59, 413–425. doi: 10.1016/j.jml.2008.02.002.
Symonds, M.R.E and Moussalli, A. (2011). A brief guide to model selection, multimodel inference and model averaging in behavioural ecology using Akaike's information criterion. Behavioral Ecology and Sociobiology, 65, 13–21. doi: 10.1007/s00265-010-1037-6
Tremblay, Antoine. (2009). Processing Advantages of Lexical Bundles: Evidence from Self-paced Reading, Word and Sentence Recall, and Free Recall with Event-related Brain Potential Recordings. Ph.D. Dissertation. University of Alberta, Edmonton, Canada.
Tremblay, A. and Tucker B. V. (2011). The Effects of N-gram Probabilistic Measures on the Processing and Production of Four-word Sequences. The Mental Lexicon, 6(2), 302–324.
See Also
bfFixefLMER_F.fnc;
bfFixefLMER_t.fnc;
ffRanefLMER.fnc;
fitLMER.fnc;
mcposthoc.fnc;
summary.mcposthoc;
pamer.fnc;
mcp.fnc;
relLik;
romr.fnc;
plotLMER.fnc;
plotLMER3d.fnc;
plotDensity3d.fnc;
plotRaw3d.fnc;
perSubjectTrim.fnc;
cn;
f;
cd;
cdf;
cdup.
Examples
## Not run:
############################################
# Load and format data. #
############################################
library(LCFdata)
data(eeg)
# restrict to electrode Fz and 80--180 ms window
eeg <- eeg[eeg$Time >= 80 & eeg$Time <= 180, ]
eeg <- eeg[, c("Subject", "Item", "Time", "Fz",
"FreqB", "LengthB", "WMC")]
# mean center FreqB
eeg$FreqBc <- eeg$FreqB - mean(eeg$FreqB)
# split FreqBc into 3 categories. Doesn't make sense,
# but it's merely for example
eeg$FreqBdc <- "high"
eeg$FreqBdc[eeg$FreqBc<=quantile(eeg$FreqBc)[3]] <- "mid"
eeg$FreqBdc[eeg$FreqBc<=quantile(eeg$FreqBc)[2]] <- "low"
eeg$FreqBdc <- as.factor(eeg$FreqBdc)
eeg$FreqBdc <- relevel(eeg$FreqBdc, "low")
# mean center LengthB
eeg$LengthBc <- eeg$LengthB - mean(eeg$LengthB)
# mean center WMC
eeg$WMCc <- eeg$WMC - mean(eeg$WMC)
############################################
# Demonstrate plotDensity3d.fnc. #
############################################
plotDensity3d.fnc(x = sort(unique(eeg$WMCc)),
y = sort(unique(eeg$LengthBc)))
############################################
# Demonstrate plotRaw3d.fnc. #
############################################
plotRaw3d.fnc(data = eeg, response = "Fz", pred = "WMCc",
intr = "LengthBc", plot.type = "persp", theta = 150)
############################################
# Analyze data. Demonstrate model #
# selection, and diagnostic plots. #
# Also demonstrate forward fitting #
# of random effects and back fitting #
# of fixed effects. Finally, #
# demonstrate pamer.fnc. #
############################################
library(lme4)
# fit initial model
m0 <- lmer(Fz ~ (FreqBdc + LengthBc + WMCc)^2 + (1 | Subject),
data = eeg)
m1 <- lmer(Fz ~ (FreqBdc + LengthBc + WMCc)^2 + (1 | Subject) +
(1 | Item), data = eeg)
# which model to choose?
relLik(m0, m1)
# choose m1
# check model assumptions
mcp.fnc(m1)
# remove outliers
eeg <- romr.fnc(m1, eeg, trim = 2.5)
eeg$n.removed
eeg$percent.removed
eeg<-eeg$data
# update model
m1 <- lmer(Fz ~ (FreqBdc + LengthBc + WMCc)^2 + (1 | Subject) +
(1 | Item), data = eeg)
# re-check model assumptions
mcp.fnc(m1)
# forward-fit random effect structure (simple for the purposes
# of the example).
m2 <- ffRanefLMER.fnc(model = m1, ran.effects =
c("(0 + LengthBc | Subject)", "(0 + WMCc | Item)"),
log.file = FALSE)
# backfit model m2. In this case, could use bfFixefLMER_t.fnc instead.
m3 <- bfFixefLMER_F.fnc(m2, log.file = FALSE)
# The calls to ffRanefLMER.fnc and bfFixefLMER_F.fnc could
# be replaced by a call to fitLMER.fnc. In this latter case, however,
# bfFixefLMER_F.fnc would be called first, then the random effect
# structure would be forward fitted, and finally teh fixed effects
# would be backfitted again.
m3b <- fitLMER.fnc(model = m1, ran.effects = c("(0 + LengthBc | Subject)",
"(0 + WMCc | Item)"), backfit.on = "F", log.file = FALSE)
pamer.fnc(m3b)
# The results are the same. This may not necessarily be the case
# elsewhere. First forward fitting the random effect structure and
# then backfitting the fixed effects, potentially pruning irrelevant
# random effects, is probably the best approach. Nonetheless, there is
# no hard evidence to this effect.
# check model assumptions
mcp.fnc(m3)
# check significance of model terms
pamer.fnc(m3)
############################################
# Demonstrate mcposthoc.fnc and #
# summary.mcposthoc. #
############################################
# Only the intercept is significant. For purposes of the
# example, let's perform a posthoc analysis on FreqBdc on
# model m2.
m2.ph <- mcposthoc.fnc(model = m2, var = list(ph1 = "FreqBdc"))
# Now check if and how the different levels differ between
# each other. First check high vs mid and high vs low:
summary(m2.ph, term = "FreqBdchigh")
# Then low vs mid (the low vs high row is redundant from the
# above summary):
summary(m2.ph, term = "FreqBdcmid")
# Note that none of the levels differ from each other. Indeed,
# the backfitting process indicated that the model only has an
# intercept (i.e., the FreqBc factor variable was not significant).
# Just to show how one would look at posthocs for interactions. Let's
# look at the effect of Length at each FreqB bin:
summary(object = m2.ph, term = "LengthBc")
# Does Length effect different Freq bins? Start with low
# versus mid and high
smry <- summary(object = m2.ph, term = "FreqBdchigh:LengthBc")
# then mid versus low and high
smry <- summary(object = m2.ph, term = "FreqBdcmid:LengthBc")
############################################
# Demonstrate `revived' version of #
# plotLMER.fnc and plotLMER3d.fnc. #
############################################
# Generate plot for Length X Freq with function plotLMER.fnc.
plotLMER.fnc(m2, pred = "LengthBc", intr = list("FreqBdc",
levels(eeg$FreqBdc), "beg", list(1 : 3, 1 : 3)))
# Plotting the Length:WMC interaction with plotLMER3d.fnc. It'll
# take a little bit of time.
plotLMER3d.fnc(m2,"LengthBc","WMCc")
# Plot it a second time to demonstrate caching. You can notice the
# speed-up.
plotLMER3d.fnc(m2,"LengthBc","WMCc")
############################################
# Demonstrate modeling and #
# backfitting of glmer. #
############################################
# Split FreqBc into 2 categories.
eeg$FreqBdc <- "high"
eeg$FreqBdc[eeg$FreqBc<=median(eeg$FreqBc)] <- "low"
eeg$FreqBdc <- as.factor(eeg$FreqBdc)
eeg$FreqBdc <- relevel(eeg$FreqBdc, "low")
# Fit glmer model.
m4 <- glmer(FreqBdc ~ (Fz + LengthBc + WMCc)^2 + (1 | Subject),
family = "binomial", data = eeg)
summary(m4)
pamer.fnc(m4)
# Back fit fixed effects, forward fit random effects, and then
# re-back fit fixed effects. Need to set argument backfit.on to "t".
m5 <- fitLMER.fnc(model = m4, ran.effects = "(0 + LengthBc | Subject)",
backfit.on = "t", log.file = FALSE)
summary(m5)
pamer.fnc(m5)
# Plot the 2-way interaction.
plotLMER.fnc(m5, pred = "Fz", intr = list("LengthBc",
quantile(eeg$LengthBc), "med",list(1:5,1:5)))
# Look at the same plot, but in 3d.
plotLMER3d.fnc(m5, pred = "Fz", intr = "LengthBc")
############################################
# Test backfitting on AIC, #
# BIC, llrt, relLik.AIC, and #
# relLik.BIC. #
############################################
# AIC
m.test <- bfFixefLMER_F.fnc(m2, method = "AIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m2, method = "AIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m4, method = "AIC",
log.file = FALSE)
m.test <- bfFixefLMER_F.fnc(m4, method = "AIC",
log.file = FALSE)
# BIC
m.test <- bfFixefLMER_F.fnc(m2, method = "BIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m2, method = "BIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m4, method = "BIC",
log.file = FALSE)
# llrt
m.test <- bfFixefLMER_F.fnc(m2, method = "llrt",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m2, method = "llrt",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m4, method = "llrt",
log.file = FALSE)
# relLik.AIC
m.test <- bfFixefLMER_F.fnc(m2, method = "relLik.AIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m2, method = "relLik.AIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m4, method = "relLik.AIC",
log.file = FALSE)
# relLik.BIC
m.test <- bfFixefLMER_F.fnc(m2, method = "relLik.BIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m2, method = "relLik.BIC",
log.file = FALSE)
m.test <- bfFixefLMER_t.fnc(m4, method = "relLik.BIC",
log.file = FALSE)
## End(Not run)