Learn R Programming
languageR (version 1.0)

plotLMER.fnc: plot a mer object

Description

Plot partial effects of a (generalized) linear mixed-effects model fit with lmer. For gaussian models, 95% highest posterior density credible intervals can be added.

Usage

plotLMER.fnc(model, xlabel = NA, xlabs = NA, ylabel = NA, ylimit = NA, 
   ilabel = NA, fun = NA, pred = NA, control = NA, ranefs = NA, n = 100, 
   intr = NA, mcmcMat = NA, lockYlim = TRUE, addlines = FALSE, 
   withList = FALSE, cexsize = 0.5, linecolor = 1, addToExistingPlot = FALSE, 
   verbose = TRUE, ...)

Arguments

model
a mer model object
xlabel
label for X-axis (if other than the variable name in the original model formula)
xlabs
character vector with labels for X-axes in multipanel plot (if other than the variable names in the original model formula); if used, xlabel should not be specified
ylabel
label for Y-axis (if other than the variable name of the dependent variable in the original model formula)
ylimit
range for vertical axis; if not specified, this range will be chosen such that all data points across all subplots, including HPD intervals, will be accommodated
ilabel
label for the interaction shown in the lower right-hand margin of the plot, overriding the original variable name in the model formula
fun
a function to be applied for transforming the dependent variable, if NA, no transformation is applied; for models with family = "binomial", fun is set to plogis by default; this can be disabled by setting
pred
character string with name of predictor; if specified, a single plot will produced for the partial effect of this specific predictor
control
a two-element list list(predictor, val) specifying a predictor the value of which has to be set to val in the partial effect plot(s); the predictor name should be exactly as specified in names(model@fixef)
ranefs
a four-element list Group, Level, Predictor, specifying a random-effect Group (e.g. Subject), a level (e.g., S10) and a value (e.g., LogFrequency) for which partial effects have to be calibrated; impleme
n
integer denoting number of points for the plot, chosen at equally spaced intervals across the empirical range of the predictor variable
intr
a list specifying an interaction to be graphed; obligatory arguments are (1) the name of the interaction variable, followed by (2) a vector of values for that variable, followed by (3) the position for interaction labels ('"beg"', '"mid"', or '"en
mcmcMat
an MCMC matrix as produced by mcmcsamp or pvals.fnc()$mcmc for the input model; when this matrix is specified, an attempt is made to add MCMC-based HPD intervals to the predicted values in the plot. Works only for R 2.6.
lockYlim
logical specifying whether all subplots should have the same range of values for the vertical axis; if TRUE, this range will be chosen to accomodate all fitted values including HDP intervals for all predictors across all plots
addlines
if TRUE, adds line(s) between levels of same factor(s)
withList
logical, if TRUE, a list will be output with all data frames for the subplots
cexsize
character expansion size (cex) for additional information in the plot for interactions
linecolor
color of lines in the plot, by default set to 1 (black)
addToExistingPlot
default FALSE, if set to TRUE, plot will be added to previous plot, but only if pred is specified
verbose
if TRUE (default), effect sizes and default transformations are reported
...
further graphical parameters to be passed down; warning: col, pch, lty and cex will often generate an error as they are internally already fully specified for specialized subplots

Value

  • A plot is produced on the graphical device.

Details

When no predictor is specified, a series of plots is produced for the partial effects of each predictor. The graphs are shown for the reference level for factors and are adjusted for the median value for the other numerical predicors in the model. Interactions are not shown. The user should set up the appropriate number of subplots on the graphics device before running plotLMER.fnc().

Instead of showing all predictors jointly, plotLMER.fnc() can also be used to plot the partial effect of a specific predictor. When a specific predictor is specified (with pred = ...), a single plot is produced for that predictor. In this case, the intr argument can be used to specify a single second predictor that enters into an interaction with the selected main predictor. Polynomials have to be fitted with poly(..., degree, raw=TRUE) and restricted cubic splines with rcs() from the Design package.

When an MCMC object (as produced by the MCMC code in lme4, version 1.0 or higher) is specified (mcmcMat = ...), an attempt is made to plot MCMC HPD intervals for the predicted values. This code is under development. For each MCMC sample, the posterior values of the fixed effects are used to predict the dependent variable for a range of its values, with other predictors in the model held constant (factors at the reference level, covariates at their medians). For each of the predicted 'posterior' values, the HPD interval is obtained, and added to the graph.

References

The 'danish' dataset in the example section is contributed by Laura Winther-Balling, see Winther-Balling, L. and Baayen, R. H., Morphological effects in auditory word recognition: Evidence from Danish, Language and Cognitive Processes, in press.

See Also

See also other utilities in languageR for facilitating work with lmer

Examples

Run this code
#############################################################
# fitting a cosine with a spline (simulated data)
#############################################################

require("Design", quietly=TRUE, character=TRUE)
dfr = makeSplineData.fnc()
table(dfr$Subject)
xylowess.fnc(Y ~ X | Subject, data = dfr)
# the smoother doesn't recognize the cosine function implemented in makeSplineData.fnc()
dev.off()   

dfr.lmer = lmer(Y ~ rcs(X, 5) + (1|Subject), data = dfr)
plotLMER.fnc(dfr.lmer)

# comparison with ols from Design package
dfr.dd = datadist(dfr)
options(datadist='dfr.dd')
dfr.ols = ols(Y~Subject+rcs(X), data=dfr, x=T, y=T)
dfr$fittedOLS = fitted(dfr.ols)
dfr$fittedLMER = as.vector(dfr.lmer@X %*% fixef(dfr.lmer))

# we plot the lmer() fit in blue, the ols() fit in red (both adjusted for
# subject S1), and plot the underlying model in green

plot(dfr[dfr$Subject=="S1",]$X, dfr[dfr$Subject=="S1",]$fittedLMER +
  ranef(dfr.lmer)[[1]]["S1",], type="l", col="blue",
  ylim = range(dfr$y + ranef(dfr.lmer)[[1]]["S1",],
  dfr[dfr$Subject == "S1",]$fittedLMER,
  dfr[dfr$Subject == "S1",]$fittedOLS),
  xlab="X", ylab="Y")   
lines(dfr[dfr$Subject=="S1",]$X, dfr[dfr$Subject=="S1",]$fittedOLS, 
col="red")
lines(dfr[dfr$Subject=="S1",]$X, dfr[dfr$Subject=="S1",]$y+
ranef(dfr.lmer)[[1]]["S1",], col="green")
legend(2,29,c("30+cos(x)", "lmer (S1)", "ols (S1)"), lty=rep(1,3), 
col=c("green", "blue", "red"))





#############################################################
# a model with a raw polynomial: the mcmc credible intervals
# will be different each time this example is run; for large
# numbers of samples, the plotting code will become quite slow
#############################################################

bg.lmer = lmer(LogRT ~ PC1+PC2+PC3 + ReadingScore +
poly(OrthLength, 2, raw=TRUE) + LogFrequency + LogFamilySize +
(1|Word) + (1|Subject)+(0+OrthLength|Subject) +
(0+LogFrequency|Subject), data = beginningReaders)
mcmc = pvals.fnc(bg.lmer, nsim=1000, withMCMC=TRUE)

pars = par()
par(mfrow=c(3,3), mar=c(5,5,1,1))
plotLMER.fnc(bg.lmer, mcmcMat=mcmc$mcmc, fun=exp, ylabel = "RT (ms)")

#############################################################
# a model with an interaction involving numeric predictors
#############################################################

danish.lmer = lmer(LogRT ~ PC1 + PC2 + PrevError + Rank +
ResidSemRating + ResidFamSize + LogWordFreq*LogAffixFreq*Sex +  
poly(LogCUP, 2, raw=TRUE) + LogUP + LogCUPtoEnd + 
(1|Subject) + (1|Word) + (1|Affix), data = danish)
danish.lmerA = lmer(LogRT ~ PC1 + PC2 + PrevError + Rank +
ResidSemRating + ResidFamSize + LogWordFreq*LogAffixFreq*Sex +  
poly(LogCUP, 2, raw=TRUE) + LogUP + LogCUPtoEnd + 
(1|Subject) + (1|Word) + (1|Affix), data = danish,
subset=abs(scale(resid(danish.lmer)))<2.5)
mcmc = pvals.fnc(danish.lmerA, nsim=10000, withMCMC=TRUE)
mcmc$fixed[,1:5]

# plot for reference level of Sex
plotLMER.fnc(danish.lmerA, pred = "LogAffixFreq", 
intr=list("LogWordFreq", round(quantile(danish$LogWordFreq),3), "beg",
list(c("red", "green", "blue", "yellow", "purple"), rep(1,5))), ylimit=c(6.5,7.0),
mcmcMat=mcmc$mcmc, xlabel = "log affix frequency", ylabel = "log RT auditory lexical decision")

# this model has a significant three-way interaction
# for visualization, we can either relevel Sex and refit,
# or make use of the control option. First releveling:

danish$Sex=relevel(danish$Sex, "F")
danish.lmerF = lmer(LogRT ~ PC1 + PC2 + PrevError + Rank +
  ResidSemRating + ResidFamSize + LogWordFreq*LogAffixFreq*Sex +  
  poly(LogCUP, 2, raw=TRUE) + LogUP + LogCUPtoEnd + 
  (1|Subject) + (1|Word) + (1|Affix), data = danish)
danish$Sex=relevel(danish$Sex, "M")
danish.lmerM = lmer(LogRT ~ PC1 + PC2 + PrevError + Rank +
  ResidSemRating + ResidFamSize + LogWordFreq*LogAffixFreq*Sex +  
  poly(LogCUP, 2, raw=TRUE) + LogUP + LogCUPtoEnd + 
  (1|Subject) + (1|Word) + (1|Affix), data = danish)

# Next preparing for using the control option:
#
# names(fixef(danish.lmer))[10]  # SexM
# unique(danish.lmer@X[,10])     # 1 0

par(mfrow=c(2,2))

plotLMER.fnc(danish.lmer, pred="LogWordFreq", ylimit=c(6.5,7.0),
intr=list("LogAffixFreq", quantile(danish$LogAffixFreq), "end"),
control=list("SexM", 0))
mtext("females", line=1.5, cex=0.9)

plotLMER.fnc(danish.lmer, pred="LogWordFreq", ylimit=c(6.5,7.0),
intr=list("LogAffixFreq", quantile(danish$LogAffixFreq), "end"),
control=list("SexM", 1))
mtext("males", line=1.5, cex=0.9)

plotLMER.fnc(danish.lmerF, pred="LogWordFreq", ylimit=c(6.5,7.0), 
intr=list("LogAffixFreq", quantile(danish$LogAffixFreq), "end"))
mtext("females", line=1.5, cex=0.9)

plotLMER.fnc(danish.lmerM, pred="LogWordFreq", ylimit=c(6.5, 7.0),
intr=list("LogAffixFreq", quantile(danish$LogAffixFreq), "end"))
mtext("males", line=1.5, cex=0.9)

par(mfrow=c(1,1))

#############################################################
# calculating effect sizes, defined as max - min
#############################################################

# effect size for a covariate

dfr = plotLMER.fnc(danish.lmerA, pred = "LogCUP", withList=TRUE)
max(dfr$LogCUP$Y)-min(dfr$LogCUP$Y)

# effect size for a factor

dfr = plotLMER.fnc(danish.lmerA, pred = "PrevError", withList=TRUE)
max(dfr$PrevError$Y)-min(dfr$PrevError$Y)


# effect sizes for the quantiles in an interaction plot

dfr = plotLMER.fnc(danish.lmerA, pred = "LogAffixFreq", 
        withList=TRUE,
        intr=list("LogWordFreq", round(quantile(danish$LogWordFreq),3), "beg"))

unlist(lapply(dfr$LogAffixFreq, FUN=function(X)return(max(X$Y)-min(X$Y))))


#############################################################
# plotting an interaction between two factors
#############################################################

danish$WordFreqFac = danish$LogWordFreq > median(danish$LogWordFreq)
danish.lmer2 = lmer(LogRT ~ WordFreqFac*Sex +  
(1|Subject) + (1|Word) + (1|Affix), data = danish)
mcmc = pvals.fnc(danish.lmer2,nsim=1000,withMCMC=TRUE)

plotLMER.fnc(danish.lmer2, pred = "Sex", 
intr=list("WordFreqFac", c("TRUE", "FALSE"), "end", 
list(c("red",  "blue"), rep(1,2))),
ylimit=c(6.5,7.0), cexsize=1.0, addlines=TRUE, mcmcMat=mcmc$mcmc)

#############################################################
# a generalized linear mixed-effects model
#############################################################

dative.lmer = lmer(RealizationOfRecipient ~ 
AccessOfTheme + AccessOfRec + LengthOfRecipient + AnimacyOfRec +
AnimacyOfTheme + PronomOfTheme + DefinOfTheme + LengthOfTheme +
SemanticClass + Modality + (1|Verb), 
data = dative, family = "binomial")

par(mfrow=c(3,4),mar=c(5,5,1,1))
plotLMER.fnc(dative.lmer, fun=plogis, addlines=TRUE)

# with user-specified labels for the x-axis
par(mfrow=c(3,4),mar=c(5,5,1,1))
plotLMER.fnc(dative.lmer, fun=plogis, addlines=TRUE,
xlabs=unlist(strsplit("abcdefghij","")))

par(pars)

Run the code above in your browser using DataLab