Load Packages and Helper Functions

Packages.

rm(list=ls())
library(ggplot2)
## Warning: package 'ggplot2' was built under R version 3.3.2
library(plotrix)
## Warning: package 'plotrix' was built under R version 3.3.2
#suppressPackageStartupMessages(library(dplyr))
suppressPackageStartupMessages(library(plyr))
## Warning: package 'plyr' was built under R version 3.3.2
library(bootstrap)
## Warning: package 'bootstrap' was built under R version 3.3.2
suppressPackageStartupMessages(library(lme4))
## Warning: package 'lme4' was built under R version 3.3.2
suppressPackageStartupMessages(library(lmerTest))
## Warning: package 'lmerTest' was built under R version 3.3.2
library(pbkrtest)
## Warning: package 'pbkrtest' was built under R version 3.3.2
#detach("package:lmerTest", unload=TRUE)

library(knitr)
## Warning: package 'knitr' was built under R version 3.3.2
theme_set(theme_bw())
opts_chunk$set(fig.width=8, fig.height=5, 
                      echo=TRUE, warning=FALSE, message=FALSE, cache=TRUE)

Helper Functions.

SummarySE

This function can be found on the website “Cookbook for R”.

http://www.cookbook-r.com/Graphs/Plotting_means_and_error_bars_(ggplot2)/#Helper

It summarizes data, giving count, mean, standard deviation, standard error of the mean, and confidence interval (default 95%).

  • data: a data frame
  • measurevar: the name of a column that contains the variable to be summarized
  • groupvars: a vector containing names of columns that contain grouping variables
  • na.rm: a boolean that indicates whether to ignore NAs
  • conf.interval: the percent range of the confidence interval (default is 95%)
summarySE <- function(data=NULL, measurevar, groupvars=NULL, na.rm=FALSE,
                      conf.interval=.95, .drop=TRUE) {
    require(plyr)

    # New version of length which can handle NA's: if na.rm==T, don't count them
    length2 <- function (x, na.rm=FALSE) {
        if (na.rm) sum(!is.na(x))
        else       length(x)
    }

    # This does the summary. For each group's data frame, return a vector with
    # N, mean, and sd
    datac <- ddply(data, groupvars, .drop=.drop,
      .fun = function(xx, col) {
        c(N    = length2(xx[[col]], na.rm=na.rm),
          mean = mean   (xx[[col]], na.rm=na.rm),
          sd   = sd     (xx[[col]], na.rm=na.rm)
        )
      },
      measurevar
    )

    # Rename the "mean" column    
    datac <- rename(datac, c("mean" = measurevar))

    datac$se <- datac$sd / sqrt(datac$N)  # Calculate standard error of the mean

    # Confidence interval multiplier for standard error
    # Calculate t-statistic for confidence interval: 
    # e.g., if conf.interval is .95, use .975 (above/below), and use df=N-1
    ciMult <- qt(conf.interval/2 + .5, datac$N-1)
    datac$ci <- datac$se * ciMult

    return(datac)
}

SummarySEwithin

This function can be found on the website “Cookbook for R”.

http://www.cookbook-r.com/Graphs/Plotting_means_and_error_bars_(ggplot2)/#Helper

From that website:

Summarizes data, handling within-subject variables by removing inter-subject variability. It will still work if there are no within-subject variables. Gives count, un-normed mean, normed mean (with same between-group mean), standard deviation, standard error of the mean, and confidence interval. If there are within-subject variables, it calculates adjusted values using the method from Morey (2008).

  • data: a data frame
  • measurevar: the name of a column that contains the variable to be summarized
  • betweenvars: a vector containing names of columns that are between-subjects variables
  • withinvars: a vector containing names of columns that are within-subjects variables
  • idvar: the name of a column that identifies each subject (or matched subjects)
  • na.rm: a boolean that indicates whether to ignore NA’s
  • conf.interval: the percent range of the confidence interval (default is 95%)
summarySEwithin <- function(data=NULL, measurevar, betweenvars=NULL, withinvars=NULL,
                            idvar=NULL, na.rm=FALSE, conf.interval=.95, .drop=TRUE) {

  # Ensure that the betweenvars and withinvars are factors
  factorvars <- vapply(data[, c(betweenvars, withinvars), drop=FALSE],
    FUN=is.factor, FUN.VALUE=logical(1))

  if (!all(factorvars)) {
    nonfactorvars <- names(factorvars)[!factorvars]
    message("Automatically converting the following non-factors to factors: ",
            paste(nonfactorvars, collapse = ", "))
    data[nonfactorvars] <- lapply(data[nonfactorvars], factor)
  }

  # Get the means from the un-normed data
  datac <- summarySE(data, measurevar, groupvars=c(betweenvars, withinvars),
                     na.rm=na.rm, conf.interval=conf.interval, .drop=.drop)

  # Drop all the unused columns (these will be calculated with normed data)
  datac$sd <- NULL
  datac$se <- NULL
  datac$ci <- NULL

  # Norm each subject's data
  ndata <- normDataWithin(data, idvar, measurevar, betweenvars, na.rm, .drop=.drop)

  # This is the name of the new column
  measurevar_n <- paste(measurevar, "_norm", sep="")

  # Collapse the normed data - now we can treat between and within vars the same
  ndatac <- summarySE(ndata, measurevar_n, groupvars=c(betweenvars, withinvars),
                      na.rm=na.rm, conf.interval=conf.interval, .drop=.drop)

  # Apply correction from Morey (2008) to the standard error and confidence interval
  #  Get the product of the number of conditions of within-S variables
  nWithinGroups    <- prod(vapply(ndatac[,withinvars, drop=FALSE], FUN=nlevels,
                           FUN.VALUE=numeric(1)))
  correctionFactor <- sqrt( nWithinGroups / (nWithinGroups-1) )

  # Apply the correction factor
  ndatac$sd <- ndatac$sd * correctionFactor
  ndatac$se <- ndatac$se * correctionFactor
  ndatac$ci <- ndatac$ci * correctionFactor

  # Combine the un-normed means with the normed results
  merge(datac, ndatac)
}

normDataWithin

This function is used by the SummarySEWithin function above. It can be found on the website “Cookbook for R”.

http://www.cookbook-r.com/Graphs/Plotting_means_and_error_bars_(ggplot2)/#Helper

From that website:

Norms the data within specified groups in a data frame; it normalizes each subject (identified by idvar) so that they have the same mean, within each group specified by betweenvars.

  • data: a data frame
  • idvar: the name of a column that identifies each subject (or matched subjects)
  • measurevar: the name of a column that contains the variable to be summarized
  • betweenvars: a vector containing names of columns that are between-subjects variables
  • na.rm: a boolean that indicates whether to ignore NAs
normDataWithin <- function(data=NULL, idvar, measurevar, betweenvars=NULL,
                           na.rm=FALSE, .drop=TRUE) {
    #library(plyr)

    # Measure var on left, idvar + between vars on right of formula.
    data.subjMean <- ddply(data, c(idvar, betweenvars), .drop=.drop,
     .fun = function(xx, col, na.rm) {
        c(subjMean = mean(xx[,col], na.rm=na.rm))
      },
      measurevar,
      na.rm
    )

    # Put the subject means with original data
    data <- merge(data, data.subjMean)

    # Get the normalized data in a new column
    measureNormedVar <- paste(measurevar, "_norm", sep="")
    data[,measureNormedVar] <- data[,measurevar] - data[,"subjMean"] +
                               mean(data[,measurevar], na.rm=na.rm)

    # Remove this subject mean column
    data$subjMean <- NULL

    return(data)
}

myCenter

This function outputs the centered values of an variable, which can be a numeric variable, a factor, or a data frame. It was taken from Florian Jaegers blog: https://hlplab.wordpress.com/2009/04/27/centering-several-variables/

From his blog:

  • If the input is a numeric variable, the output is the centered variable

  • If the input is a factor, the output is a numeric variable with centered factor level values. That is, the factor’s levels are converted into numerical values in their inherent order (if not specified otherwise, R defaults to alphanumerical order). More specifically, this centers any binary factor so that the value below 0 will be the first level of the original factor, and the value above 0 will be the second level.

  • If the input is a data frame or matrix, the output is a new matrix of the same dimension and with the centered values and column names that correspond to the colnames() of the input preceded by “c” (e.g. “Variable1” will be “cVariable1”).

myCenter= function(x) {
  if (is.numeric(x)) { return(x - mean(x, na.rm=T)) }
    if (is.factor(x)) {
        x= as.numeric(x)
        return(x - mean(x, na.rm=T))
    }
    if (is.data.frame(x) || is.matrix(x)) {
        m= matrix(nrow=nrow(x), ncol=ncol(x))
        colnames(m)= paste("c", colnames(x), sep="")
    
        for (i in 1:ncol(x)) {
        
            m[,i]= myCenter(x[,i])
        }
        return(as.data.frame(m))
    }
}

selectCenter

This function provides a wrapper around myCenter allowing you to center a specific list of variables from a dataframe.

  • x: data frame
  • listfname: a list of the variables to be centered (e.g. list(variable1,variable2))

The output is a copy of the data frame with a column (always a numeric variable) added for each of the centered variables. These columns are labelled with each column’s previous name, but with “.ct” appended (e.g., “variable1” will become “variable1.ct”).

selectCenter= function(x, listfname) 
{
    for (i in 1:length(listfname)) 
    {
        fname = as.character(listfname[i])
        x[paste(fname,".ct", sep="")] = myCenter(x[fname])
    }
        
    return(x)
}

get_coeffs and get_coeffs_lmerTest

The get_coeffs function allows us to inspect just particular coefficients from the output of an lme model (as produced by the summary() function in the lme4 package) by putting them in table.

  • x: the output returned when running an lmer or glmer (i.e. an object of type lmerMod or glmerMod)
  • list: a list of names of the coefficients to be extracted (e.g. c(“variable1”,“variable1:variable2”))

The get_coeffs_lmerTest function is identical except that instead of using the summary package from lme4 it uses the summary package from lmerTest using Kenward-Rodger approximation for denominator degrees of freedom and thus also including p values)

get_coeffs <- function(x,list){(kable(as.data.frame(summary(x)$coefficients)[list,],digits=3))}

get_coeffs_lmerTest <- function(x,list){(kable(as.data.frame(summary(x,ddf="kenw")$coefficients)[list,],digits=3))}

filter2

This is a function which filters a column of data removing values which are some number of standard deviations above/below the mean for that participant, possibly in some condition/subcondition.

  • im: the input matrix (a data frame)
  • svn: a list of the names of factors to be group by (subject name + one or more conditions)
  • fn: the name of the column containing the data to be filtered
  • lim: how many standard deviations above/below the mean to filter

The function returns an input matrix identical to the input matrix but with an additional columns giving the group means and the “filtered” data

filter2 = function(im, svn, fn, lim)
{
  ## work out means lisfor each subject for each word

x = list()
y = ""

for(n in svn) x=append(im[n],x)
for(n in svn) y=paste(y,n,sep="_")

means = aggregate(im[fn], by = x, mean, na.rm=T)
head(means)
nocols = dim(means)[2]
colnames(means)[nocols] = "means"

sds = aggregate(im[fn], by = x, sd, na.rm=T)
head(sds)
nocols = dim(sds)[2]
colnames(sds)[nocols] = "sds"

gs = merge(means,sds)

## because if there is just one value it doesn't have a stand deviation and don't want to just disregard all of these
gs$sds[is.na(gs$sds)] = 0 

gs$max = gs$means + lim*gs$sds
gs$min = gs$means- lim*gs$sds

im2 = merge(im,gs, sort=F)


im2[paste(fn,"filt",sep="")] = im2[fn]
cn= dim(im2)[2] ## get colnumber (last one added)

im2[,cn][im2[,fn]> im2$max] = ""

im2[,cn][im2[,fn]< im2$min] = ""

im2[,cn]= as.numeric(im2[,cn])

 
names(im2)[names(im2)=="means"] = paste("mean", y, sep="_") 
names(im2)[names(im2)=="sds"] = paste("sd", y, sep="_") 
names(im2)[names(im2)=="max"] = paste("max", y, sep="_") 
names(im2)[names(im2)=="min"] = paste("min", y, sep="_") 

return(im2)
}

Load Datasets

train = read.csv( "training.csv")
discrim = read.csv( "discrim.csv")
lexd = read.csv( "lexd.csv")
engintro = read.csv( "englishintro.csv")

Training

Data Preparation

Check amount of missing data

There should be 320 trials * 10 sessions per participant. Thus for adults there should be 41 adults * 3200 making 1.31210^{5} trials and for children there should 52 children * 3200 trials making 1.66410^{5} trials.

round(1- table(train$agegroup)/c(3200*41, 3200*52),2)
## 
## adult child 
##  0.02  0.06

Select data for analyses

In order to make a fair comparison across conditions, we only look at trials using the two talkers who were also used in the low variability (one talker) condition.

train2 = subset(train, talker == "female1" | talker == "female2")
train2$talker = factor(train2$talker)
train2.child = subset(train2, agegroup =="child")
train2.adult = subset(train2, agegroup =="adult")

Plot means (for Figure 3)

Adults

train2.adult.participantmeans = aggregate(accuracy ~ participant + session + condition, na.rm = T, data = train2.adult, FUN = mean)
train2.adult.groupmeans <- summarySEwithin(train2.adult.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

train2.adult.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker, na.rm = T, data = train2.adult, FUN = mean)

train2.adult.groupmeans.talker1 <- summarySEwithin(subset(train2.adult.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

train2.adult.groupmeans.talker2 <- summarySEwithin(subset(train2.adult.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

train2.adult.groupmeans$accuracy = (train2.adult.groupmeans.talker1$accuracy + train2.adult.groupmeans.talker2$accuracy)/2

# Now plot using the standard error for the confidence intervals

epsilon = 0.02 #how big the hat (bar at the top/bottom) is
with(subset(train2.adult.groupmeans, condition == "highvar"),
plot(as.numeric(session), accuracy, type = "p", col = "black",pch = 15, xlab = "Session", ylab = "Proportion correct", xlim = c(1,10), ylim = c(0.5, 1)) +
  segments(as.numeric(session), (accuracy - ci) ,as.numeric(session), (accuracy + ci), col="black") +
  segments(as.numeric(session)-epsilon, (accuracy - ci) ,as.numeric(session)+epsilon, (accuracy - ci), col="black") +
  segments(as.numeric(session)-epsilon, (accuracy + ci) ,as.numeric(session)+epsilon, (accuracy + ci), col="black"),
     )
## numeric(0)
# Add low variability to the same plot (using "lines" rather than plot)

with(subset(train2.adult.groupmeans, condition == "lowvar"),
lines(as.numeric(session), accuracy, type = "p", col = "grey45", pch = 15, xlab = "Session", ylab = "Proportion correct", xlim = c(1,10), ylim = c(0.5, 1)) +
  segments(as.numeric(session), (accuracy - ci) ,as.numeric(session), (accuracy + ci), col="grey45") +
  segments(as.numeric(session)-epsilon, (accuracy - ci) ,as.numeric(session)+epsilon, (accuracy - ci), col="grey45") +
  segments(as.numeric(session)-epsilon, (accuracy + ci) ,as.numeric(session)+epsilon, (accuracy + ci), col="grey45"),
     )  
## numeric(0)
legend('topleft','groups', c("Low Variability","High Variabiltiy"), col=c("grey45","black"), pch = c(15,17))

Children

train2.child.participantmeans = aggregate(accuracy ~ participant + session + condition, na.rm = T, data = train2.child, FUN = mean)
train2.child.groupmeans <- summarySEwithin(train2.child.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for participant for each talker (average mean values for each talker); error calculations as done over the whole participant set.

train2.child.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker, na.rm = T, data = train2.child, FUN = mean)

train2.child.groupmeans.talker1 <- summarySEwithin(subset(train2.child.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

train2.child.groupmeans.talker2 <- summarySEwithin(subset(train2.child.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

train2.child.groupmeans$accuracy = (train2.child.groupmeans.talker1$accuracy + train2.child.groupmeans.talker2$accuracy)/2

# Now plot using the standard error for the error bars

epsilon = 0.02 #how big the hat (bar at the top/bottom) is
with(subset(train2.child.groupmeans, condition == "highvar"),
plot(as.numeric(session), accuracy, type = "p", col="black",pch = 15, xlab = "Session", ylab = "Proportion correct", xlim = c(1,10), ylim = c(0.5, 1)) +
  segments(as.numeric(session), (accuracy - ci) ,as.numeric(session), (accuracy + ci), col="black") +
  segments(as.numeric(session)-epsilon, (accuracy - ci) ,as.numeric(session)+epsilon, (accuracy - ci), col="black") +
  segments(as.numeric(session)-epsilon, (accuracy + ci) ,as.numeric(session)+epsilon, (accuracy + ci), col="black"),
     )  
## numeric(0)
# Add low variability to the same plot (using "lines" rather than plot)

with(subset(train2.child.groupmeans, condition == "lowvar"),
lines(as.numeric(session), accuracy, type = "p", col="grey45",pch = 15, xlab = "Session", ylab = "Proportion correct", xlim = c(1,10), ylim = c(0.5, 1)) +
  segments(as.numeric(session), (accuracy - ci) ,as.numeric(session), (accuracy + ci), col="grey45") +
  segments(as.numeric(session)-epsilon, (accuracy - ci) ,as.numeric(session)+epsilon, (accuracy - ci), col="grey45") +
  segments(as.numeric(session)-epsilon, (accuracy + ci) ,as.numeric(session)+epsilon, (accuracy + ci), col="grey45"),
     )  
## numeric(0)
legend('topleft','groups', c("Low Variability","High Variabiltiy"), col=c("grey45","black"), pch = c(15,17))

Statistial Analyses

Our approach was to only inspect models for effects and interactions between the experimental variables where there are clear predictions. For the training data, this is the case for both main effects of condition (high/low variability) and session and the interaction between them. We therefore begin by inspecting these effects in the model. Wherever we do find reliable effects, we then check to see whether they are qualified by an interaction with talker. Where this is the case, we check to see if the effect holds for each of the talkers separately.

Adults

train2.adult = selectCenter(train2.adult, list("condition", "session", "talker"))
lmer.adult <- glmer(accuracy ~ 1 
                    + (condition.ct * session.ct * talker.ct)
                    + (session.ct||participant),
                    data = train2.adult, control = glmerControl(optimizer = "bobyqa"), family = binomial)
lmer.adult.coeff = kable(summary(lmer.adult)$coefficients, digits = 3)
get_coeffs(lmer.adult, c("(Intercept)", "condition.ct", "session.ct", "condition.ct:session.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 3.062 0.152 20.197 0
condition.ct 1.626 0.254 6.406 0
session.ct 0.349 0.031 11.208 0
condition.ct:session.ct 0.202 0.052 3.857 0

There are main effects of training session and condition and a reliable interaction between condition and session. This reflects improvement across sessions and overall better performance in the low variability condition which increases across sessions.

Check interactions with talker

Check interactions with talker for each of these effects

get_coeffs(lmer.adult, c("talker.ct", "condition.ct:talker.ct", "session.ct:talker.ct", "condition.ct:session.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 2.107 0.284 7.411 0.000
condition.ct:talker.ct 1.727 0.452 3.817 0.000
session.ct:talker.ct 0.260 0.062 4.182 0.000
condition.ct:session.ct:talker.ct 0.255 0.105 2.442 0.015

There is a main effect of the control variable talker (reflecting greater accuracy with one talker: female1 0.84 female2 0.94 )

There are also interactions with talker for all of the reliable experimental effects. To break down the interactions, we look at each talker separately.

train2.adult.t1 = selectCenter(subset(train2.adult, talker == "female1"), list("condition", "session", "talker"))
lmer.adult.t1 <- glmer(accuracy ~ 1
                       + (condition.ct * session.ct)
                       + (session.ct|participant),
                       data = train2.adult.t1, control = glmerControl(optimizer = "bobyqa"), family = binomial)
kable(summary(lmer.adult.t1)$coefficients, digits = 3)
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.940 0.170 11.403 0.000
condition.ct 0.720 0.285 2.526 0.012
session.ct 0.212 0.020 10.412 0.000
condition.ct:session.ct 0.069 0.035 1.978 0.048 
train2.adult.t2 = selectCenter(subset(train2.adult, talker == "female2"), list("condition", "session", "talker"))
lmer.adult.t2 <- glmer(accuracy ~ 1 
                       + (condition.ct * session.ct)
                       + (session.ct|participant),
                       data = train2.adult.t2, control = glmerControl(optimizer = "bobyqa"), family = binomial)
kable(summary(lmer.adult.t2)$coefficients, digits = 3)
Estimate Std. Error z value Pr(>|z|)
(Intercept) 4.139 0.265 15.642 0.000
condition.ct 2.473 0.442 5.600 0.000
session.ct 0.485 0.064 7.526 0.000
condition.ct:session.ct 0.334 0.108 3.096 0.002

These analyses show that the reliable effects of condition and session hold for both talkers. The interaction holds for talker 2 (couldn’t look at this for talker 1).

Children

Note that full model did not converge, even with interactions between slopes removed. We therefore removed correlations between slopes.

train2.child = selectCenter(train2.child, list("condition", "session", "talker"))
lmer.child <- glmer(accuracy ~ 1 
                    + (condition.ct * session.ct + talker.ct)
                    + (session.ct||participant),
                    data = train2.child, control = glmerControl(optimizer = "bobyqa"), family = binomial)
get_coeffs(lmer.child, c("(Intercept)", "condition.ct", "session.ct", "condition.ct:session.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.064 0.142 7.469 0.000
condition.ct 0.374 0.239 1.564 0.118
session.ct 0.223 0.032 7.055 0.000
condition.ct:session.ct 0.103 0.054 1.924 0.054

There was a reliable main effect of session, reflecting improved performance across sessions, but no reliable main effect of condition. There was also a near reliable interaction between session and condition. Inspecting the figure above, this seems to reflect the fact that the difference between conditions emerges only in the second half of training.

Follow up in a model where session is replaced by a binary division into “first half” “second half”

train2.child$testhalf =  2
train2.child$testhalf[train2.child$session < 6] = 1

train2.child = selectCenter(train2.child, list("condition", "testhalf", "talker"))
lmer.child.2 <- glmer(accuracy ~ 1 
                      + (condition.ct * testhalf.ct * talker.ct)
                      + (testhalf.ct||participant),
                      data = train2.child, control = glmerControl(optimizer = "bobyqa"), family = binomial)
get_coeffs(lmer.child.2, c("(Intercept)", "condition.ct", "testhalf.ct", "condition.ct:testhalf.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.994 0.125 7.920 0.000
condition.ct 0.317 0.224 1.412 0.158
testhalf.ct 1.045 0.139 7.514 0.000
condition.ct:testhalf.ct 0.483 0.215 2.248 0.025

There is a reliable interaction between condition and testhalf. We explore this by running the model again but with a separate slope for condition for each half of testing.

train2.child$testhalf = factor(train2.child$testhalf)

## Run the model again but removing the main effect of condition and the condition by testhalf interaction and replacing with separate effects of condition for each test half.  

lmer.child.2.v2 <- glmer(accuracy ~ 1 
                + condition.ct : testhalf
                + (condition.ct * testhalf.ct * talker.ct)
                - condition.ct
                - condition.ct : testhalf.ct
                + (testhalf.ct||participant),
                data = train2.child, control = glmerControl(optimizer = "bobyqa"), family = binomial)

## Check  that this is the same model as above with the same number of dfs  
anova(lmer.child.2, lmer.child.2.v2)
## Data: train2.child
## Models:
## lmer.child.2: accuracy ~ 1 + (condition.ct * testhalf.ct * talker.ct) + (testhalf.ct || 
## lmer.child.2:     participant)
## lmer.child.2.v2: accuracy ~ 1 + condition.ct:testhalf + (condition.ct * testhalf.ct * 
## lmer.child.2.v2:     talker.ct) - condition.ct - condition.ct:testhalf.ct + (testhalf.ct || 
## lmer.child.2.v2:     participant)
##                 Df    AIC    BIC logLik deviance Chisq Chi Df Pr(>Chisq)
## lmer.child.2    10 106402 106496 -53191   106382                        
## lmer.child.2.v2 10 106402 106496 -53191   106382     0      0          1
get_coeffs(lmer.child.2.v2, c("condition.ct:testhalf1", "condition.ct:testhalf2"))
Estimate Std. Error z value Pr(>|z|)
condition.ct:testhalf1 0.077 0.241 0.320 0.749
condition.ct:testhalf2 0.560 0.229 2.442 0.015

This shows that there is a reliable effect of condition only in the second half of training.

Check interactions with talker

Note that we were not able to include interactions with talker in the original model (with session as a continuous variable) so cannot look to see if the reliable effect of session and the (near) realiable session by condition interaction are qualified by an effect of talker

However we can look at the model with session replaced by test-half.

get_coeffs(lmer.child.2, c("talker.ct", "testhalf.ct:talker.ct", "condition.ct:testhalf.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 0.185 0.218 0.851 0.395
testhalf.ct:talker.ct 0.189 0.284 0.666 0.506
condition.ct:testhalf.ct:talker.ct 0.732 0.339 2.162 0.031

There was no reliable effect of talker and the effect of test half isn’t qualified by an interaction with talker. However the interaction between condition and testhalf is qualified by an interaction with talker. We broke this down by looking to see whether there was a reliable effect of condition in each test half for each talker.

lmer.child.2.v3 <- glmer(accuracy ~ 1 
                  + condition.ct : testhalf : talker
                  + (condition.ct * testhalf.ct * talker.ct)
                  - condition.ct 
                  - condition.ct : testhalf.ct
                  - condition.ct: talker.ct
                  - condition.ct : testhalf.ct : talker.ct
                  + (testhalf.ct||participant),
                  data = train2.child, control = glmerControl(optimizer = "bobyqa"), family = binomial)

anova(lmer.child.2, lmer.child.2.v3)
## Data: train2.child
## Models:
## lmer.child.2: accuracy ~ 1 + (condition.ct * testhalf.ct * talker.ct) + (testhalf.ct || 
## lmer.child.2:     participant)
## lmer.child.2.v3: accuracy ~ 1 + condition.ct:testhalf:talker + (condition.ct * 
## lmer.child.2.v3:     testhalf.ct * talker.ct) - condition.ct - condition.ct:testhalf.ct - 
## lmer.child.2.v3:     condition.ct:talker.ct - condition.ct:testhalf.ct:talker.ct + 
## lmer.child.2.v3:     (testhalf.ct || participant)
##                 Df    AIC    BIC logLik deviance Chisq Chi Df Pr(>Chisq)
## lmer.child.2    10 106402 106496 -53191   106382                        
## lmer.child.2.v3 10 106402 106496 -53191   106382     0      0          1
get_coeffs(lmer.child.2.v3, c("condition.ct:testhalf1:talkerfemale1", "condition.ct:testhalf2:talkerfemale1", "condition.ct:testhalf1:talkerfemale2", "condition.ct:testhalf2:talkerfemale2"))
Estimate Std. Error z value Pr(>|z|)
condition.ct:testhalf1:talkerfemale1 -0.048 0.307 -0.157 0.875
condition.ct:testhalf2:talkerfemale1 0.090 0.333 0.272 0.786
condition.ct:testhalf1:talkerfemale2 0.219 0.290 0.756 0.450
condition.ct:testhalf2:talkerfemale2 1.090 0.285 3.828 0.000

We only have an effect of condition for the more intelligible talker (female 2) in the second half of training.

Adult/Child Comparison

Since adults are at ceiling in the low-variablity condition, we look at the high variability condition only.

train2.highvar = selectCenter(subset(train2, condition == "highvar"), list("session", "talker", "agegroup"))

lmer.highvar.age <- glmer(accuracy ~ 1 
                    + session.ct * talker.ct * agegroup.ct                 
                    + (session.ct||participant),
                    data = train2.highvar, control = glmerControl(optimizer = "bobyqa"), family = binomial)

get_coeffs(lmer.highvar.age, c("agegroup.ct", "session.ct:agegroup.ct"))
Estimate Std. Error z value Pr(>|z|)
agegroup.ct -1.008 0.217 -4.644 0.000
session.ct:agegroup.ct -0.050 0.049 -1.020 0.308

There is a main effect of age group but no interaction between age group and session.

Check interactions with talker

get_coeffs(lmer.highvar.age, c("talker.ct", "talker.ct:agegroup.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 0.173 0.216 0.802 0.423
talker.ct:agegroup.ct -1.071 0.435 -2.462 0.014

The main effect of age group is qualified by an interaction with talker, so we look to see if the benefit for adults over children holds for both talkers:

lmer.highvar.age.v2 <- glmer(accuracy ~ 1 
                      + agegroup.ct : talker
                      + session.ct * talker.ct * agegroup.ct                 
                      - agegroup.ct
                      - talker.ct : agegroup.ct
                      + (session.ct||participant),
                      data = train2.highvar, control = glmerControl(optimizer = "bobyqa"), family = binomial)

anova(lmer.highvar.age.v2,lmer.highvar.age)
## Data: train2.highvar
## Models:
## lmer.highvar.age.v2: accuracy ~ 1 + agegroup.ct:talker + session.ct * talker.ct * 
## lmer.highvar.age.v2:     agegroup.ct - agegroup.ct - talker.ct:agegroup.ct + (session.ct || 
## lmer.highvar.age.v2:     participant)
## lmer.highvar.age: accuracy ~ 1 + session.ct * talker.ct * agegroup.ct + (session.ct || 
## lmer.highvar.age:     participant)
##                     Df   AIC   BIC logLik deviance Chisq Chi Df Pr(>Chisq)
## lmer.highvar.age.v2 10 39940 40026 -19960    39920                        
## lmer.highvar.age    10 39940 40026 -19960    39920     0      0          1
get_coeffs(lmer.highvar.age.v2, c("agegroup.ct:talkerfemale1", "agegroup.ct:talkerfemale2"))
Estimate Std. Error z value Pr(>|z|)
agegroup.ct:talkerfemale1 -0.465 0.304 -1.532 0.126
agegroup.ct:talkerfemale2 -1.537 0.311 -4.941 0.000

This suggests that the benefit for adults over children is only reliable for the more intelligible talker.

English Introduction

Data Preparation

Select data for analyses

engintro.child = subset(engintro, agegroup == "child")
engintro.adult = subset(engintro, agegroup == "adult")

Get means for Table 4

Adults

engintro.adult.participantmeans = aggregate(accuracy ~ participant + session + condition, na.rm = T, data = engintro.adult, FUN = mean)

engintro.adult.groupmeans <- summarySEwithin(engintro.adult.participantmeans, measurevar="accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for the imbalance in talkers (rather than the overall mean per participant, get the mean values for each participant for each talker and average those; error calculations remain as done over the whole participant set

engintro.adult.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker, na.rm = T, data = engintro.adult, FUN = mean)

engintro.adult.groupmeans.talker1 <- summarySEwithin(subset(engintro.adult.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

engintro.adult.groupmeans.talker2 <- summarySEwithin(subset(engintro.adult.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

engintro.adult.groupmeans$accuracy = (engintro.adult.groupmeans.talker1$accuracy + engintro.adult.groupmeans.talker2$accuracy)/2

kable(engintro.adult.groupmeans, digits = 2)
condition session N accuracy accuracy_norm sd se ci
highvar post 22 0.98 0.99 0.10 0.02 0.04
highvar pre 22 0.81 0.82 0.10 0.02 0.04
lowvar post 18 1.00 0.99 0.08 0.02 0.04
lowvar pre 18 0.82 0.81 0.08 0.02 0.04

Children

engintro.child.participantmeans = aggregate(accuracy ~ participant + session + condition, na.rm = T, data = engintro.child, FUN = mean)
engintro.child.groupmeans <- summarySEwithin(engintro.child.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for the imbalance in talkers (rather than the overall mean per participant, get the mean values for each participant for each talker and average those; error calculations remain as done over the whole participant set

engintro.child.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker, na.rm = T, data = engintro.child, FUN = mean)
 
engintro.child.groupmeans.talker1 <- summarySEwithin(subset(engintro.child.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

engintro.child.groupmeans.talker2 <- summarySEwithin(subset(engintro.child.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

engintro.child.groupmeans$accuracy = (engintro.child.groupmeans.talker1$accuracy + engintro.child.groupmeans.talker2$accuracy)/2

kable(engintro.child.groupmeans, digits = 2)
condition session N accuracy accuracy_norm sd se ci
highvar post 28 0.88 0.88 0.15 0.03 0.06
highvar pre 28 0.50 0.50 0.15 0.03 0.06
lowvar post 24 0.92 0.92 0.09 0.02 0.04
lowvar pre 24 0.47 0.47 0.09 0.02 0.04

Statistical Analyses

Note we do not do statistical analyses on adult data as they are close to ceiling, even at pre-test.

Children

engintro.child = selectCenter(engintro.child, list("session", "condition", "talker"))
lme.engintrochild <- glmer(accuracy ~ 1 
                           + session.ct * condition.ct *  talker.ct
                           + (session.ct|participant), 
                           control=glmerControl(optimizer = "bobyqa"), data = engintro.child, family = binomial)
get_coeffs(lme.engintrochild, c("session.ct", "condition.ct", "session.ct:condition.ct"))
Estimate Std. Error z value Pr(>|z|)
session.ct -3.191 0.261 -12.232 0.000
condition.ct 0.331 0.285 1.161 0.246
session.ct:condition.ct -0.946 0.505 -1.874 0.061

There was a reliable main effect of session, but not of condition. There was a marginal interaction between session and condition.

Check interactions with talker

get_coeffs(lme.engintrochild , c("talker.ct", "session.ct:talker.ct", "session.ct:condition.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct -0.098 0.281 -0.348 0.728
session.ct:talker.ct 0.138 0.496 0.279 0.780
session.ct:condition.ct:talker.ct -0.555 1.001 -0.555 0.579

There was no main effect of talker and neither the main effect of condition nor the interaction between condition and session was qualified by an interaction with talker.

Lexical Decision

Data Preparation

Select data for analyses

Remove practice and nonword trials (which aren’t used in any analyses) and then subset into the adult and child data sets for greek and english primes

lexd = subset(lexd, trialnumber != "practice" & trialtype != "NONWORD")
lexd.adult.english = subset(lexd, agegroup == "adult" & primelang == "english")
lexd.adult.greek = subset(lexd, agegroup =="adult" & primelang == "greek")
lexd.child.english = subset(lexd, agegroup == "child" & primelang == "english")
lexd.child.greek = subset(lexd, agegroup == "child" & primelang == "greek")

Filtering for RT analyses

For each of these data sets, remove the trials that were not answered correctly, and filter the RT data using the filter2 function. This function removes trials where the RT is > or < 2.5 SD from the mean value for that participant in that session. Finally, remove any remaining RTs that are < 200 ms.

flist = c("participant", "session") # list of parameters for calculating means
lim = 2.5 # number of standard deviations 

lexd.adult.greek.RT = subset(lexd.adult.greek, accuracy == 1)
lexd.adult.greek.RT = filter2(lexd.adult.greek.RT, flist, "RT", lim)
lexd.adult.greek.RT$RTfilt[lexd.adult.greek$RTfilt < 200] <- NA

lexd.child.greek.RT = subset(lexd.child.greek, accuracy == 1)
lexd.child.greek.RT = filter2(lexd.child.greek.RT, flist, "RT", lim)
lexd.child.greek.RT$RTfilt[lexd.child.greek$RTfilt < 200] <- NA

lexd.adult.english.RT = subset(lexd.adult.english, accuracy == 1)
lexd.adult.english.RT = filter2(lexd.adult.english.RT, flist, "RT", lim)
lexd.adult.english.RT$RTfilt[lexd.adult.english$RTfilt < 200] <- NA

lexd.child.english.RT = subset(lexd.child.english, accuracy == 1)
lexd.child.english.RT = filter2(lexd.child.english.RT, flist, "RT", lim)
lexd.child.english.RT$RTfilt[lexd.child.english$RTfilt < 200] <- NA

Calculate % trials removed because targets were incorrectly identified as nonwords:

#adult, greek-primes percent incorrect: 
round(100 * (1 - dim(lexd.adult.greek.RT)[1]/dim(lexd.adult.greek)[1]))
## [1] 5
#child, greek-primes percent incorrect: 
round(100 * (1 - dim(lexd.child.greek.RT)[1]/dim(lexd.child.greek)[1]))
## [1] 10
#adult, english-primes percent incorrect: 
round(100 * (1 - dim(lexd.adult.english.RT)[1]/dim(lexd.adult.english)[1]))
## [1] 6
#child, english-primes percent incorrect: 
round(100 * (1 - dim(lexd.child.english.RT)[1]/dim(lexd.child.english)[1]))
## [1] 16

Calculate % of further trials removed due to filtering:

#adult, greek-primes  
round((sum(is.na(lexd.adult.greek.RT$RTfilt))/sum(!is.na(lexd.adult.greek.RT$RT)))*100)
## [1] 3
#child, greek-primes 
round((sum(is.na(lexd.child.greek.RT$RTfilt))/sum(!is.na(lexd.child.greek.RT$RT)))*100)
## [1] 3
#adult, english-primes 
round((sum(is.na(lexd.adult.english.RT$RTfilt))/sum(!is.na(lexd.adult.english.RT$RT)))*100)
## [1] 4
#child, english-primes 
round((sum(is.na(lexd.child.english.RT$RTfilt))/sum(!is.na(lexd.child.english.RT$RT)))*100)
## [1] 4

Get Means and Plots

Greek Primes

Adult Accuracy

lexd.adult.greek.participantmeans = aggregate(accuracy ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.adult.greek, FUN = mean)

lexd.adult.greek.groupmeans <- summarySEwithin(lexd.adult.greek.participantmeans, measurevar = "accuracy", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

kable(lexd.adult.greek.groupmeans, digits = 3)
session primetarget_relationship N accuracy accuracy_norm sd se ci
post related 41 0.973 0.973 0.032 0.005 0.010
post unrelated 41 0.923 0.923 0.048 0.007 0.015
pre related 41 0.974 0.974 0.046 0.007 0.014
pre unrelated 41 0.928 0.928 0.047 0.007 0.015

Adults RT

lexd.adult.greek.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.adult.greek.RT, FUN = mean)

lexd.adult.greek.RT.groupmeans <- summarySEwithin(lexd.adult.greek.RT.participantmeans, measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

kable(lexd.adult.greek.RT.groupmeans, digits = 3)
session primetarget_relationship N RTfilt RTfilt_norm sd se ci
post related 41 1007.995 1007.995 184.290 28.781 58.169
post unrelated 41 1087.945 1087.945 144.334 22.541 45.557
pre related 41 1031.923 1031.923 143.959 22.483 45.439
pre unrelated 41 1142.395 1142.395 157.665 24.623 49.765

Children Accuracy

lexd.child.greek.participantmeans = aggregate(accuracy ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.child.greek, FUN = mean)

lexd.child.greek.groupmeans <- summarySEwithin(lexd.child.greek.participantmeans, measurevar = "accuracy", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

kable(lexd.child.greek.groupmeans, digits = 3)
session primetarget_relationship N accuracy accuracy_norm sd se ci
post related 52 0.907 0.907 0.095 0.013 0.026
post unrelated 52 0.851 0.851 0.099 0.014 0.028
pre related 52 0.946 0.946 0.094 0.013 0.026
pre unrelated 52 0.881 0.881 0.118 0.016 0.033

Children RT

lexd.child.greek.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.child.greek.RT, FUN = mean)

lexd.child.greek.RT.groupmeans <- summarySEwithin(lexd.child.greek.RT.participantmeans, measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

kable(lexd.child.greek.RT.groupmeans, digits = 3)
session primetarget_relationship N RTfilt RTfilt_norm sd se ci
post related 52 1389.938 1389.938 356.514 49.440 99.254
post unrelated 52 1461.169 1461.169 330.132 45.781 91.909
pre related 52 1552.385 1552.385 393.381 54.552 109.518
pre unrelated 52 1612.984 1612.984 269.153 37.325 74.933

English Primes

Adult Accuracy

lexd.adult.english.participantmeans = aggregate(accuracy ~ participant + session + primetarget_relationship + condition, na.rm = T, data = lexd.adult.english, FUN = mean)

lexd.adult.english.groupmeans <- summarySEwithin(lexd.adult.english.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.adult.english.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker + primetarget_relationship, na.rm = T, data = lexd.adult.english, FUN = mean)

lexd.adult.english.groupmeans.talker1 <- summarySEwithin(subset(lexd.adult.english.participantmeans.t12, talker == "female1"), measurevar ="accuracy", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.groupmeans.talker2 <- summarySEwithin(subset(lexd.adult.english.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.groupmeans$accuracy = (lexd.adult.english.groupmeans.talker1$accuracy + lexd.adult.english.groupmeans.talker2$accuracy)/2

kable(lexd.adult.english.groupmeans, digits = 3)
condition session primetarget_relationship N accuracy accuracy_norm sd se ci
highvar post related 22 0.982 0.976 0.036 0.008 0.016
highvar post unrelated 22 0.923 0.917 0.051 0.011 0.023
highvar pre related 22 0.980 0.974 0.033 0.007 0.015
highvar pre unrelated 22 0.902 0.897 0.058 0.012 0.026
lowvar post related 19 0.972 0.977 0.036 0.008 0.017
lowvar post unrelated 19 0.894 0.898 0.067 0.015 0.032
lowvar pre related 19 0.958 0.967 0.037 0.009 0.018
lowvar pre unrelated 19 0.914 0.922 0.045 0.010 0.021

Adult RT

lexd.adult.english.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship + condition, na.rm = T, data = lexd.adult.english.RT, FUN = mean)

lexd.adult.english.RT.groupmeans <- summarySEwithin(lexd.adult.english.RT.participantmeans, measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.adult.english.RT.participantmeans.t12 = aggregate(RTfilt ~ participant + session + condition + talker + primetarget_relationship, na.rm = T, data = lexd.adult.english.RT, FUN = mean)

lexd.adult.english.RT.groupmeans.talker1 <- summarySEwithin(subset(lexd.adult.english.RT.participantmeans.t12, talker == "female1"), measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.RT.groupmeans.talker2 <- summarySEwithin(subset(lexd.adult.english.RT.participantmeans.t12, talker == "female2"), measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.RT.groupmeans$RTfilt = (lexd.adult.english.RT.groupmeans.talker1$RTfilt + lexd.adult.english.RT.groupmeans.talker2$RTfilt)/2

kable(lexd.adult.english.RT.groupmeans, digits = 3)
condition session primetarget_relationship N RTfilt RTfilt_norm sd se ci
highvar post related 22 1151.098 1138.351 199.289 42.489 88.360
highvar post unrelated 22 1234.067 1221.320 220.874 47.091 97.930
highvar pre related 22 1225.341 1212.594 289.115 61.640 128.186
highvar pre unrelated 22 1226.076 1213.328 199.420 42.516 88.418
lowvar post related 19 1097.814 1116.801 93.480 21.446 45.056
lowvar post unrelated 19 1208.202 1231.300 239.688 54.988 115.526
lowvar pre related 19 1138.798 1151.622 129.654 29.745 62.491
lowvar pre unrelated 19 1273.559 1285.869 166.363 38.166 80.184
Plot for Figure 4

Note that we have collapsed across condition here.

# First get the means (collapsed across conditions) to plot

lexd.adult.english.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.adult.english.RT, FUN = mean)

lexd.adult.english.RT.groupmeans <- summarySEwithin(lexd.adult.english.RT.participantmeans, measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.adult.english.RT.participantmeans.t12 = aggregate(RTfilt ~ participant + session + talker + primetarget_relationship, na.rm = T, data = lexd.adult.english.RT, FUN = mean)

lexd.adult.english.RT.groupmeans.talker1 <- summarySEwithin(subset(lexd.adult.english.RT.participantmeans.t12, talker == "female1"), measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.RT.groupmeans.talker2 <- summarySEwithin(subset(lexd.adult.english.RT.participantmeans.t12, talker == "female2"), measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.adult.english.RT.groupmeans$RTfilt = (lexd.adult.english.RT.groupmeans.talker1$RTfilt + lexd.adult.english.RT.groupmeans.talker2$RTfilt)/2

# Now plot

lexd.adult.english.RT.groupmeans$session = relevel(lexd.adult.english.RT.groupmeans$session, "pre")

p = ggplot(lexd.adult.english.RT.groupmeans, aes(x = session, y = RTfilt, fill = primetarget_relationship))
p = p + geom_bar(stat = "identity", position = "dodge", colour = "black", size = .3)
p = p + geom_errorbar(aes(ymin = RTfilt - ci, ymax = RTfilt + ci), width = .4, position = position_dodge(.9))
p = p + xlab("") + ylab("RT(ms)")
p = p + scale_x_discrete(labels = c("pre" = "Pre test", "post" = "Post test"), expand = c(0, 0.6))
p = p + scale_fill_manual(values = c("grey", "white"), name = "Prime target relationship", labels = c("related", "unrelated"))
#p = p + scale_y_continuous(limits = c(0,1.1), expand = c(0, 0))
p = p + coord_cartesian(ylim = c(1000, 1900))

p

Children Accuracy

lexd.child.english.participantmeans = aggregate(accuracy ~ participant + session + primetarget_relationship + condition, na.rm = T, data = lexd.child.english, FUN = mean)

lexd.child.english.groupmeans <- summarySEwithin(lexd.child.english.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.child.english.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker + primetarget_relationship, na.rm = T, data = lexd.child.english, FUN = mean)

lexd.child.english.groupmeans.talker1 <- summarySEwithin(subset(lexd.child.english.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session","primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.groupmeans.talker2 <- summarySEwithin(subset(lexd.child.english.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.groupmeans$accuracy = (lexd.child.english.groupmeans.talker1$accuracy + lexd.child.english.groupmeans.talker2$accuracy)/2

kable(lexd.child.english.groupmeans, digits = 3)
condition session primetarget_relationship N accuracy accuracy_norm sd se ci
highvar post related 28 0.891 0.876 0.103 0.019 0.040
highvar post unrelated 28 0.834 0.819 0.114 0.021 0.044
highvar pre related 28 0.884 0.869 0.086 0.016 0.033
highvar pre unrelated 28 0.820 0.805 0.129 0.024 0.050
lowvar post related 24 0.867 0.889 0.113 0.023 0.048
lowvar post unrelated 24 0.815 0.836 0.094 0.019 0.040
lowvar pre related 24 0.838 0.858 0.084 0.017 0.035
lowvar pre unrelated 24 0.770 0.787 0.132 0.027 0.056

Children RT

lexd.child.english.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship + condition, na.rm = T, data = lexd.child.english.RT, FUN = mean)

lexd.child.english.RT.groupmeans <- summarySEwithin(lexd.child.english.RT.participantmeans, measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.child.english.RT.participantmeans.t12 = aggregate(RTfilt~ participant + session + condition + talker + primetarget_relationship, na.rm = T, data = lexd.child.english.RT, FUN = mean)

lexd.child.english.RT.groupmeans.talker1 <- summarySEwithin(subset(lexd.child.english.RT.participantmeans.t12, talker == "female1"), measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.RT.groupmeans.talker2 <- summarySEwithin(subset(lexd.child.english.RT.participantmeans.t12, talker == "female2"), measurevar = "RTfilt", betweenvars = c("condition"), withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.RT.groupmeans$RTfilt = (lexd.child.english.RT.groupmeans.talker1$RTfilt + lexd.child.english.RT.groupmeans.talker2$RTfilt)/2

kable(lexd.child.english.RT.groupmeans, digits = 3)
condition session primetarget_relationship N RTfilt RTfilt_norm sd se ci
highvar post related 28 1486.777 1536.915 378.880 71.602 146.914
highvar post unrelated 28 1574.379 1624.517 274.708 51.915 106.521
highvar pre related 28 1669.582 1719.720 351.204 66.371 136.183
highvar pre unrelated 28 1665.049 1715.187 290.669 54.931 112.710
lowvar post related 24 1494.584 1439.786 331.837 67.736 140.122
lowvar post unrelated 24 1622.993 1562.857 468.758 95.685 197.939
lowvar pre related 24 1811.015 1765.427 444.960 90.827 187.890
lowvar pre unrelated 24 1869.315 1828.268 409.474 83.584 172.906
Plot for Figure 4

Note that we have collapsed across condition here.

# First get the means (collapsed across conditions) to plot

lexd.child.english.RT.participantmeans = aggregate(RTfilt ~ participant + session + primetarget_relationship, na.rm = T, data = lexd.child.english.RT, FUN = mean)

lexd.child.english.RT.groupmeans <- summarySEwithin(lexd.child.english.RT.participantmeans, measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

lexd.child.english.RT.participantmeans.t12 = aggregate(RTfilt ~ participant + session + talker + primetarget_relationship, na.rm = T, data = lexd.child.english.RT, FUN = mean)

lexd.child.english.RT.groupmeans.talker1 <- summarySEwithin(subset(lexd.child.english.RT.participantmeans.t12, talker == "female1"), measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.RT.groupmeans.talker2 <- summarySEwithin(subset(lexd.child.english.RT.participantmeans.t12, talker == "female2"), measurevar = "RTfilt", betweenvars = NULL, withinvars = c("session", "primetarget_relationship"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

lexd.child.english.RT.groupmeans$RTfilt = (lexd.child.english.RT.groupmeans.talker1$RTfilt + lexd.child.english.RT.groupmeans.talker2$RTfilt)/2

# Now plot

lexd.child.english.RT.groupmeans$session = relevel(lexd.child.english.RT.groupmeans$session, "pre")

p = ggplot(lexd.child.english.RT.groupmeans, aes(x = session, y = RTfilt, fill = primetarget_relationship))
p = p + geom_bar(stat="identity", position="dodge", colour="black", size=.3)
p = p + geom_errorbar(aes(ymin = RTfilt-ci, ymax=RTfilt+ci), width = .4, position = position_dodge(.9))
p = p + xlab("") + ylab("RT(ms)")
p = p + scale_x_discrete(labels = c("pre" = "Pre test", "post" = "Post test"), expand = c(0, 0.6))
p = p + scale_fill_manual(values = c("grey", "white"), name = "Prime target relationship", labels = c("related", "unrelated"))
#p = p + scale_y_continuous(limits=c(0,1.1), expand = c(0, 0))
p = p + coord_cartesian(ylim = c(1000, 1900))

p

Statistical Analyses

Adults, Greek, RT

Note that for these analyses - which have a continuous predicted variable and use lmer rather than glmer - p values are obtained using the package lmerTest. A Kenward-Roger approximation is used (computational error occurs when using the default Satthwaite approximation). (Note that “talker”" is not included as variable since there was only one greek talker used for these stimuli)

lexd.adult.greek.RT$primetarget_relationship = factor(lexd.adult.greek.RT$primetarget_relationship) 

lexd.adult.greek.RT = selectCenter(lexd.adult.greek.RT, list("session", "primetarget_relationship"))

lme.greek.adult<- lmer(RTfilt ~ 1
                      + session.ct * primetarget_relationship.ct
                      + (session.ct * primetarget_relationship.ct|participant), 
                      data = lexd.adult.greek.RT, control = lmerControl(optimizer = "bobyqa"), REML = F )
kable(summary(lme.greek.adult, ddf = "kenw")$coefficients, digits = 3)
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 1066.968 23.936 42.066 44.028 0.000
session.ct 37.904 39.445 42.220 0.949 0.348
primetarget_relationship.ct 101.326 13.516 49.702 7.403 0.000
session.ct:primetarget_relationship.ct 17.707 26.409 45.666 0.668 0.507

An effect of prime-target relationship (related = 1014, unrelated = 1114 but no effect of test session and no interaction.

Adult, English, RT

Begin by looking at experimental effects with predictions:

lexd.adult.english.RT$primetarget_relationship = factor(lexd.adult.english.RT$primetarget_relationship)
lexd.adult.english.RT$talker = factor(lexd.adult.english.RT$talker)

lexd.adult.english.RT = selectCenter(lexd.adult.english.RT, list("session", "primetarget_relationship", "talker", "condition"))

lme.english.adult<- lmer(RTfilt ~ 1
                        + session.ct * primetarget_relationship.ct * condition.ct * talker.ct
                        + (session.ct* primetarget_relationship.ct|participant), 
                        data = lexd.adult.english.RT, control = lmerControl(optimizer = "bobyqa"), REML = F )
get_coeffs_lmerTest(lme.english.adult, c("(Intercept)", "session.ct", "primetarget_relationship.ct", "session.ct:primetarget_relationship.ct", "session.ct:primetarget_relationship.ct:condition.ct"))
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 1195.162 29.115 46.830 39.026 0.000
session.ct 42.175 48.643 48.747 0.824 0.414
primetarget_relationship.ct 83.619 20.127 57.046 4.036 0.000
session.ct:primetarget_relationship.ct -37.261 41.760 55.469 -0.854 0.397
session.ct:primetarget_relationship.ct:condition.ct 95.898 83.810 55.473 1.096 0.278

An effect of prime-target relationship (related = 1153 unrelated = 1238) but no effect of test session, no interaction between prime-target relationship and test session, and no three-way interaction between condition, prime-target relationship and test sesssion.

This indicates semantic priming across the langauges, but no indication that it is influenced by training.

Check interactions with talker

Check whether effect of prime-target relationship is qualified by an interaction with talker

get_coeffs_lmerTest(lme.english.adult, c("talker", "primetarget_relationship.ct:talker"))
Estimate Std. Error df t value Pr(>|t|)
talker.ct -47.841 57.930 49.687 -0.782 0.438
primetarget_relationship.ct:talker.ct -71.576 40.291 59.469 -1.721 0.090

No overall effect of talker, but talker did interact marginally with prime-target relationship. Looking at the means, this reflects more priming with the less intelligble talker (female 1)

round(with(lexd.adult.english.RT, tapply(RTfilt, list(primetarget_relationship, talker), mean, na.rm = T)))
##           female1 female2
## related      1163    1143
## unrelated    1284    1193

Child, Greek, RT

lexd.child.greek.RT$primetarget_relationship = factor(lexd.child.greek.RT$primetarget_relationship) 

lexd.child.greek.RT = selectCenter(lexd.child.greek.RT, list("session", "primetarget_relationship"))
lme.greek.child<- lmer(RTfilt ~ 1
                      + session.ct * primetarget_relationship.ct
                      + (primetarget_relationship.ct * session.ct|participant), 
                      data = lexd.child.greek.RT, control = lmerControl(optimizer = "bobyqa") , REML = F)
kable(summary(lme.greek.child, ddf = "kenw")$coefficients, digits = 3)
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 1510.026 43.241 53.549 34.568 0.000
session.ct 136.083 69.209 54.156 1.946 0.057
primetarget_relationship.ct 72.343 29.209 52.715 2.467 0.017
session.ct:primetarget_relationship.ct -25.249 67.967 71.995 -0.368 0.714

There was a significant effect of prime-target relationship (related = 1487 unrelated = 1549), and a marginal effect of test session (pre = 1563 post = 1469). The interaction between prime-target relationship and session was non-significant.

Child, English, RT

lexd.child.english.RT$primetarget_relationship = factor(lexd.child.english.RT$primetarget_relationship)
lexd.child.english.RT$talker = factor(lexd.child.english.RT$talker)

lexd.child.english.RT = selectCenter(lexd.child.english.RT, list("session", "primetarget_relationship", "talker", "condition"))

lme.english.child<- lmer(RTfilt ~ 1
                      + session.ct * primetarget_relationship.ct * condition.ct * talker.ct
                      + (session.ct * primetarget_relationship.ct|participant), 
                      data = lexd.child.english.RT, control = lmerControl(optimizer = "bobyqa"), REML = F)
get_coeffs_lmerTest(lme.english.child, c("(Intercept)", "session.ct", "primetarget_relationship.ct", "session.ct:primetarget_relationship.ct", "session.ct:primetarget_relationship.ct:condition.ct"))
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 1641.142 48.517 57.344 32.416 0.000
session.ct 181.883 74.463 57.593 2.339 0.023
primetarget_relationship.ct 65.165 33.389 47.873 1.945 0.058
session.ct:primetarget_relationship.ct -52.665 66.597 47.363 -0.788 0.434
session.ct:primetarget_relationship.ct:condition.ct -7.548 133.805 47.616 -0.056 0.955

There is a marginal effect of prime-target relationship (related = 1620 unrelated = 1664 and a main effect of test session (pre = 1704 post = 1580

Check interactions with talker

get_coeffs_lmerTest(lme.english.child, c("(Intercept)", "talker.ct", "session.ct:talker.ct", "primetarget_relationship.ct:talker.ct"))
Estimate Std. Error df t value Pr(>|t|)
(Intercept) 1641.142 48.517 57.344 32.416 0.000
talker.ct -70.538 97.110 57.398 -0.696 0.489
session.ct:talker.ct -245.946 149.068 57.665 -1.580 0.120
primetarget_relationship.ct:talker.ct -8.199 66.880 47.941 -0.122 0.903

No effect of talker or interactions involving this variable.

Child, English, Accuracy

lexd.child.english$primetarget_relationship = factor(lexd.child.english$primetarget_relationship)
lexd.child.english$talker = factor(lexd.child.english$talker)

lexd.child.english = selectCenter(lexd.child.english, list("session", "primetarget_relationship", "talker", "condition"))

lme.english.child.acc<- glmer(accuracy ~ 1
                              + session.ct * primetarget_relationship.ct * condition.ct * talker.ct
                              + (session.ct * primetarget_relationship.ct|participant), 
                              data = lexd.child.english, control = glmerControl(optimizer = "bobyqa" ), family = binomial)
get_coeffs(lme.english.child.acc, c("(Intercept)", "session.ct", "primetarget_relationship.ct", "session.ct:primetarget_relationship.ct", "session.ct:primetarget_relationship.ct:condition.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 2.029 0.132 15.355 0.000
session.ct -0.320 0.168 -1.906 0.057
primetarget_relationship.ct -0.708 0.126 -5.596 0.000
session.ct:primetarget_relationship.ct 0.261 0.232 1.123 0.261
session.ct:primetarget_relationship.ct:condition.ct 0.012 0.399 0.030 0.976

There is a reliable main effect of prime-target relationship (related = 0.87 unrelated = 0.81 and a marginal main effect of test session (pre = 0.83 post = 0.85). No other reliable effects.

Check interactions with talker

get_coeffs(lme.english.child.acc, c("(Intercept)", "talker.ct", "primetarget_relationship.ct:talker.ct", "session.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 2.029 0.132 15.355 0.000
talker.ct -0.195 0.260 -0.750 0.453
primetarget_relationship.ct:talker.ct 0.130 0.227 0.575 0.566
session.ct:talker.ct 0.625 0.322 1.945 0.052

There is no overall effect talker and the effect of prime-target relationship isn’t qualified by an interaction with talker. However, the marginal effect of session is qualified by an interaction with talker. This reflects the fact that children only showed improvements from pre to post test with one of the two talkers (female 1 - the less intelligible talker)

female 1 : pre = 0.82 post = 0.89 female 2 :pre = 0.84 post = 0.81

Discrimination

Data Preparation

Select data for analyses

discrim.child = subset(discrim, agegroup == "child")
discrim.adult = subset(discrim, agegroup == "adult")

Get Means and Plots

Adults

Table of Means

discrim.adult.participantmeans = aggregate(accuracy ~ participant + session + condition + word_oldnew + voice_oldnew, na.rm = T, data = discrim.adult, FUN = mean)

discrim.adult.groupmeans <- summarySEwithin(discrim.adult.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval  = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

discrim.adult.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker + word_oldnew + voice_oldnew, na.rm = T, data = discrim.adult, FUN = mean)

discrim.adult.groupmeans.talker1 <- summarySEwithin(subset(discrim.adult.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.adult.groupmeans.talker2 <- summarySEwithin(subset(discrim.adult.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.adult.groupmeans$accuracy = (discrim.adult.groupmeans.talker1$accuracy + discrim.adult.groupmeans.talker2$accuracy)/2

kable(discrim.adult.groupmeans, digits = 3)
condition session voice_oldnew word_oldnew N accuracy accuracy_norm sd se ci
highvar post newvoice newword 22 0.836 0.835 0.193 0.041 0.086
highvar post newvoice oldword 22 0.823 0.821 0.134 0.029 0.059
highvar post oldvoice newword 22 0.800 0.798 0.184 0.039 0.082
highvar post oldvoice oldword 22 0.841 0.839 0.175 0.037 0.078
highvar pre newvoice newword 22 0.709 0.708 0.191 0.041 0.085
highvar pre newvoice oldword 22 0.755 0.753 0.185 0.039 0.082
highvar pre oldvoice newword 22 0.800 0.798 0.186 0.040 0.082
highvar pre oldvoice oldword 22 0.764 0.762 0.186 0.040 0.083
lowvar post newvoice newword 19 0.836 0.833 0.205 0.047 0.099
lowvar post newvoice oldword 19 0.810 0.807 0.186 0.043 0.090
lowvar post oldvoice newword 19 0.791 0.802 0.186 0.043 0.089
lowvar post oldvoice oldword 19 0.776 0.786 0.161 0.037 0.078
lowvar pre newvoice newword 19 0.812 0.812 0.126 0.029 0.061
lowvar pre newvoice oldword 19 0.767 0.765 0.179 0.041 0.086
lowvar pre oldvoice newword 19 0.741 0.749 0.186 0.043 0.090
lowvar pre oldvoice oldword 19 0.752 0.760 0.150 0.034 0.072

Plot for Figure 5

Plot of the difference scores (showing improvement from pre- to post- test)

discrim.adult.participantmeans.2 = reshape(discrim.adult.participantmeans, timevar = "session", idvar = c("participant", "condition", "word_oldnew", "voice_oldnew"), direction = "wide")

discrim.adult.participantmeans.2$accuracy.diff = discrim.adult.participantmeans.2$accuracy.post - discrim.adult.participantmeans.2$accuracy.pre

discrim.adult.groupmeans.diff <- summarySEwithin(discrim.adult.participantmeans.2, measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

discrim.adult.participantmeans.t12.2 = reshape(discrim.adult.participantmeans.t12, timevar = "session", idvar = c("participant", "condition", "word_oldnew", "voice_oldnew", "talker"), direction = "wide")

discrim.adult.participantmeans.t12.2$accuracy.diff = discrim.adult.participantmeans.t12.2$accuracy.post - discrim.adult.participantmeans.t12.2$accuracy.pre

discrim.adult.groupmeans.diff.t1 <- summarySEwithin(subset(discrim.adult.participantmeans.t12.2, talker == "female1"), measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.adult.groupmeans.diff.t2 <- summarySEwithin(subset(discrim.adult.participantmeans.t12.2, talker == "female2"), measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew","word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.adult.groupmeans.diff$accuracy.diff =  (discrim.adult.groupmeans.diff.t1$accuracy.diff + discrim.adult.groupmeans.diff.t2$accuracy.diff)/2 

# Now plot

discrim.adult.groupmeans.diff$word_oldnew = relevel(discrim.adult.groupmeans.diff$word_oldnew, "oldword")
discrim.adult.groupmeans.diff$novelty = paste(as.character(discrim.adult.groupmeans.diff$word_oldnew), as.character(discrim.adult.groupmeans.diff$voice_oldnew))

discrim.adult.groupmeans.diff$novelty = factor(discrim.adult.groupmeans.diff$novelty)
discrim.adult.groupmeans.diff$novelty = factor(discrim.adult.groupmeans.diff$novelty, 
levels = c("oldword oldvoice", "oldword newvoice", "newword oldvoice", "newword newvoice"))

p = ggplot(discrim.adult.groupmeans.diff, aes(x = novelty, y = accuracy.diff, fill = condition))
p = p + geom_bar(stat="identity", position = "dodge", colour = "black", size = .3)
p = p + geom_errorbar(aes(ymin = accuracy.diff-ci, ymax=accuracy.diff+ci), width = .4, position = position_dodge(.9))
p = p + xlab("") + ylab("% increase in correct responses")
p = p + scale_x_discrete(labels=c("oldword oldvoice" = " trained words \n trained voice", 
                                  "oldword newvoice" = " trained words \n untrained voice",
                                  "newword oldvoice" = " untrained words \n trained voice",
                                  "newword newvoice" = " untrained words \n untrained voice"
                                  ), expand=c(0, 0.6))
p = p + scale_fill_manual(values = c("grey", "white"), name = "Condition", labels = c("high-variablity", "low-variablity"))
p = p + coord_cartesian(ylim = c(-.1, .4))

p

Children

Table of Means

discrim.child.participantmeans = aggregate(accuracy ~ participant + session + condition + word_oldnew + voice_oldnew, na.rm = T, data = discrim.child, FUN = mean)
discrim.child.groupmeans <- summarySEwithin(discrim.child.participantmeans, measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

discrim.child.participantmeans.t12 = aggregate(accuracy ~ participant + session + condition + talker + word_oldnew + voice_oldnew, na.rm = T, data = discrim.child, FUN = mean)

discrim.child.groupmeans.talker1 <- summarySEwithin(subset(discrim.child.participantmeans.t12, talker == "female1"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.child.groupmeans.talker2 <- summarySEwithin(subset(discrim.child.participantmeans.t12, talker == "female2"), measurevar = "accuracy", betweenvars = c("condition"), withinvars = c("session", "voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.child.groupmeans$accuracy = (discrim.child.groupmeans.talker1$accuracy + discrim.child.groupmeans.talker2$accuracy)/2

kable(discrim.child.groupmeans, digits = 3)
condition session voice_oldnew word_oldnew N accuracy accuracy_norm sd se ci
highvar post newvoice newword 28 0.746 0.724 0.163 0.031 0.063
highvar post newvoice oldword 28 0.754 0.731 0.148 0.028 0.057
highvar post oldvoice newword 28 0.714 0.692 0.148 0.028 0.057
highvar post oldvoice oldword 28 0.754 0.731 0.148 0.028 0.057
highvar pre newvoice newword 28 0.664 0.642 0.192 0.036 0.075
highvar pre newvoice oldword 28 0.629 0.606 0.196 0.037 0.076
highvar pre oldvoice newword 28 0.686 0.663 0.154 0.029 0.060
highvar pre oldvoice oldword 28 0.668 0.646 0.131 0.025 0.051
lowvar post newvoice newword 24 0.770 0.801 0.167 0.034 0.070
lowvar post newvoice oldword 24 0.748 0.780 0.218 0.045 0.092
lowvar post oldvoice newword 24 0.720 0.738 0.166 0.034 0.070
lowvar post oldvoice oldword 24 0.744 0.759 0.180 0.037 0.076
lowvar pre newvoice newword 24 0.584 0.613 0.206 0.042 0.087
lowvar pre newvoice oldword 24 0.526 0.555 0.153 0.031 0.064
lowvar pre oldvoice newword 24 0.588 0.609 0.176 0.036 0.074
lowvar pre oldvoice oldword 24 0.567 0.580 0.178 0.036 0.075

Plot for Figure 5

Plot of the difference scores (showing improvement from pre- to post- test)

discrim.child.participantmeans.2 = reshape(discrim.child.participantmeans, timevar = "session", idvar = c("participant", "condition", "word_oldnew", "voice_oldnew"), direction = "wide")

discrim.child.participantmeans.2$accuracy.diff = discrim.child.participantmeans.2$accuracy.post - discrim.child.participantmeans.2$accuracy.pre

discrim.child.groupmeans.diff <- summarySEwithin(discrim.child.participantmeans.2, measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

# To correct means for imbalance in talkers, get the mean values for each participant for each talker (average mean values for each talker); error calculations as done over the whole participant set

discrim.child.participantmeans.t12.2 = reshape(discrim.child.participantmeans.t12, timevar = "session", idvar = c("participant", "condition", "word_oldnew", "voice_oldnew","talker"), direction = "wide")

discrim.child.participantmeans.t12.2$accuracy.diff = discrim.child.participantmeans.t12.2$accuracy.post - discrim.child.participantmeans.t12.2$accuracy.pre

discrim.child.groupmeans.diff.t1 <- summarySEwithin(subset(discrim.child.participantmeans.t12.2, talker =="female1"), measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.child.groupmeans.diff.t2 <- summarySEwithin(subset(discrim.child.participantmeans.t12.2, talker =="female2"), measurevar = "accuracy.diff", betweenvars = c("condition"), withinvars = c("voice_oldnew", "word_oldnew"), idvar = "participant", na.rm = FALSE, conf.interval = .95)

discrim.child.groupmeans.diff$accuracy.diff =  (discrim.child.groupmeans.diff.t1$accuracy.diff + discrim.child.groupmeans.diff.t2$accuracy.diff)/2 

# Now plot

discrim.child.groupmeans.diff$word_oldnew = relevel(discrim.child.groupmeans.diff$word_oldnew, "oldword")
discrim.child.groupmeans.diff$novelty = paste(as.character(discrim.child.groupmeans.diff$word_oldnew), as.character(discrim.child.groupmeans.diff$voice_oldnew))

discrim.child.groupmeans.diff$novelty = factor(discrim.child.groupmeans.diff$novelty)
discrim.child.groupmeans.diff$novelty = factor(discrim.child.groupmeans.diff$novelty, levels = c("oldword oldvoice", "oldword newvoice", "newword oldvoice", "newword newvoice"))

p = ggplot(discrim.child.groupmeans.diff, aes(x = novelty, y = accuracy.diff, fill = condition))
p = p + geom_bar(stat="identity", position = "dodge", colour = "black", size = .3)
p = p + geom_errorbar(aes(ymin = accuracy.diff-ci, ymax = accuracy.diff+ci), width = .4, position = position_dodge(.9))
p = p + xlab("") + ylab("% increase in correct responses")
p = p + scale_x_discrete(labels=c("oldword oldvoice" = " trained words \n trained voice", 
                                  "oldword newvoice" = " trained words \n untrained voice",
                                  "newword oldvoice" = " untrained words \n trained voice",
                                  "newword newvoice" = " untrained words \n untrained voice"
                                  ), expand = c(0, 0.6))
p = p + scale_fill_manual(values = c("grey", "white"), name = "Condition", labels = c("high-variablity", "low-variablity"))
p = p + coord_cartesian(ylim = c(-.1, .4))

p

Statistical Analyses

Adults

discrim.adult = selectCenter(discrim.adult, list("session", "word_oldnew", "voice_oldnew", "talker", "condition", "agegroup", "mean_replay_in_training"))

adult.lmer <- glmer(accuracy ~ 1 
                    + (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                    + mean_replay_in_training.ct 
                    + (word_oldnew.ct * voice_oldnew.ct * session.ct|participant), 
                    data = discrim.adult, 
                    control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 1e5)), family = binomial)

get_coeffs(adult.lmer, c("(Intercept", "condition.ct", "session.ct", "word_oldnew.ct", "condition.ct:session.ct", "session.ct:word_oldnew.ct", "session.ct:voice_oldnew.ct", "condition.ct:session.ct:word_oldnew.ct", "condition.ct:session.ct:voice_oldnew.ct", "condition.ct:word_oldnew.ct:voice_oldnew.ct", "session.ct:word_oldnew.ct:voice_oldnew.ct", "condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.661 0.099 16.854 0.000
condition.ct -0.050 0.193 -0.260 0.795
session.ct -0.520 0.151 -3.440 0.001
word_oldnew.ct -0.148 0.120 -1.234 0.217
condition.ct:session.ct 0.065 0.291 0.223 0.824
session.ct:word_oldnew.ct 0.071 0.248 0.287 0.774
session.ct:voice_oldnew.ct 0.189 0.249 0.761 0.446
condition.ct:session.ct:word_oldnew.ct 0.229 0.462 0.496 0.620
condition.ct:session.ct:voice_oldnew.ct -0.910 0.457 -1.993 0.046
condition.ct:word_oldnew.ct:voice_oldnew.ct -0.005 0.452 -0.012 0.991
session.ct:word_oldnew.ct:voice_oldnew.ct -0.824 0.495 -1.666 0.096
condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct 1.938 0.918 2.111 0.035

There is a reliable effect of test-session but no predicted test-session by condition interaction. There is a test-session by condition by voice-novelty interaction, but this is qualified by a four-way test-session by condition by voice-novelty by word-novelty interaction.

We break down the four way interaction by re-running the model so that we have a slope for the three way test-session by condition by voice-novelty interaction for each of trained and untrained words.

adult.lmer.v2 <- glmer(accuracy ~ 1 
                       + (condition.ct: session.ct:voice_oldnew.ct:word_oldnew)
                       + (condition.ct  * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                       - (condition.ct: session.ct: word_oldnew.ct: voice_oldnew.ct)
                       - (condition.ct: session.ct: voice_oldnew.ct)
                       + mean_replay_in_training.ct 
                       + (word_oldnew.ct * voice_oldnew.ct * session.ct|participant),
                       data = discrim.adult, 
                       control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun=1e5)), family = binomial)
anova(adult.lmer.v2, adult.lmer)
## Data: discrim.adult
## Models:
## adult.lmer.v2: accuracy ~ 1 + (condition.ct:session.ct:voice_oldnew.ct:word_oldnew) + 
## adult.lmer.v2:     (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * 
## adult.lmer.v2:         talker.ct) - (condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct) - 
## adult.lmer.v2:     (condition.ct:session.ct:voice_oldnew.ct) + mean_replay_in_training.ct + 
## adult.lmer.v2:     (word_oldnew.ct * voice_oldnew.ct * session.ct | participant)
## adult.lmer: accuracy ~ 1 + (condition.ct * session.ct * word_oldnew.ct * 
## adult.lmer:     voice_oldnew.ct * talker.ct) + mean_replay_in_training.ct + 
## adult.lmer:     (word_oldnew.ct * voice_oldnew.ct * session.ct | participant)
##               Df    AIC    BIC  logLik deviance Chisq Chi Df Pr(>Chisq)
## adult.lmer.v2 69 3127.5 3548.1 -1494.7   2989.5                        
## adult.lmer    69 3127.5 3548.1 -1494.7   2989.5     0      0          1
get_coeffs(adult.lmer.v2, c("condition.ct:session.ct:voice_oldnew.ct:word_oldnewnewword", "condition.ct:session.ct:voice_oldnew.ct:word_oldnewoldword"))
Estimate Std. Error z value Pr(>|z|)
condition.ct:session.ct:voice_oldnew.ct:word_oldnewnewword -1.879 0.674 -2.789 0.005
condition.ct:session.ct:voice_oldnew.ct:word_oldnewoldword 0.059 0.620 0.095 0.925

The condition by session by voice-novelty interaction is reliable only for untrained words. We further break down the condition by session by voice-novelty interaction for untrained words.

adult.lmer.v3 <- glmer(accuracy ~ 1 + 
                         + (condition.ct: session.ct: word_oldnew: voice_oldnew)
                       + (condition.ct  * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                       - (condition.ct: session.ct: word_oldnew.ct: voice_oldnew.ct)
                       - (condition.ct: session.ct: voice_oldnew.ct)
                       - (condition.ct: session.ct: word_oldnew.ct)
                       - (condition.ct: session.ct)
                       + mean_replay_in_training.ct 
                       + (word_oldnew.ct * voice_oldnew.ct * session.ct|participant),
                       data = discrim.adult, 
                       control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun=1e5)), family = binomial)
anova(adult.lmer.v3, adult.lmer)
## Data: discrim.adult
## Models:
## adult.lmer.v3: accuracy ~ 1 + +(condition.ct:session.ct:word_oldnew:voice_oldnew) + 
## adult.lmer.v3:     (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * 
## adult.lmer.v3:         talker.ct) - (condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct) - 
## adult.lmer.v3:     (condition.ct:session.ct:voice_oldnew.ct) - (condition.ct:session.ct:word_oldnew.ct) - 
## adult.lmer.v3:     (condition.ct:session.ct) + mean_replay_in_training.ct + 
## adult.lmer.v3:     (word_oldnew.ct * voice_oldnew.ct * session.ct | participant)
## adult.lmer: accuracy ~ 1 + (condition.ct * session.ct * word_oldnew.ct * 
## adult.lmer:     voice_oldnew.ct * talker.ct) + mean_replay_in_training.ct + 
## adult.lmer:     (word_oldnew.ct * voice_oldnew.ct * session.ct | participant)
##               Df    AIC    BIC  logLik deviance Chisq Chi Df Pr(>Chisq)
## adult.lmer.v3 69 3127.5 3548.1 -1494.7   2989.5                        
## adult.lmer    69 3127.5 3548.1 -1494.7   2989.5     0      0          1
get_coeffs(adult.lmer.v3, c("condition.ct:session.ct:word_oldnewnewword:voice_oldnewnewvoice", "condition.ct:session.ct:word_oldnewnewword:voice_oldnewoldvoice"))
Estimate Std. Error z value Pr(>|z|)
condition.ct:session.ct:word_oldnewnewword:voice_oldnewnewvoice 0.890 0.504 1.764 0.078
condition.ct:session.ct:word_oldnewnewword:voice_oldnewoldvoice -0.989 0.517 -1.913 0.056

There is a marginal condition by session interaction for the untrained voice which went in the predicted direction, but also a marginal interaction in the opposite direction for the trained voice. In other words, the interaction rests both on a trend towards a greater benefit of high-variability input compared with low-variability input for untrained words-untrained voice items (which is predicted since novelty should aid generalization) and a trend towards a greater benefit of low-variability for untrained words-trained voice items (which is not predicted).

Check interactions with talker

get_coeffs(adult.lmer, c("talker.ct", "session.ct:talker.ct", "condition.ct:session.ct:word_oldnew.ct:talker.ct", "condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 1.678 0.133 12.661 0.000
session.ct:talker.ct -0.590 0.234 -2.528 0.011
condition.ct:session.ct:word_oldnew.ct:talker.ct 0.367 0.919 0.399 0.690
condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:talker.ct 1.886 1.838 1.026 0.305

There is a reliable main effect of talker, reflecting female being overall more intelligible (female1 0.68, female2 0.9).

The main effect of test-session is qualified by an interaction with talker. We look to see if the effect of test-session holds for each talker;

adult.lmer.v4 <- glmer(accuracy ~ 1
                      + session.ct : talker
                      + (condition.ct  * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                      - session.ct : talker.ct
                      - session.ct
                      + mean_replay_in_training.ct 
                      + (word_oldnew.ct * voice_oldnew.ct * session.ct|participant),
                      data = discrim.adult, 
                      control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun=1e5)), family=binomial)
anova(adult.lmer.v4, adult.lmer)
## Data: discrim.adult
## Models:
## adult.lmer.v4: accuracy ~ 1 + session.ct:talker + (condition.ct * session.ct * 
## adult.lmer.v4:     word_oldnew.ct * voice_oldnew.ct * talker.ct) - session.ct:talker.ct - 
## adult.lmer.v4:     session.ct + mean_replay_in_training.ct + (word_oldnew.ct * 
## adult.lmer.v4:     voice_oldnew.ct * session.ct | participant)
## adult.lmer: accuracy ~ 1 + (condition.ct * session.ct * word_oldnew.ct * 
## adult.lmer:     voice_oldnew.ct * talker.ct) + mean_replay_in_training.ct + 
## adult.lmer:     (word_oldnew.ct * voice_oldnew.ct * session.ct | participant)
##               Df    AIC    BIC  logLik deviance Chisq Chi Df Pr(>Chisq)
## adult.lmer.v4 69 3127.5 3548.1 -1494.7   2989.5                        
## adult.lmer    69 3127.5 3548.1 -1494.7   2989.5     0      0          1
get_coeffs(adult.lmer.v4, c("session.ct:talkerfemale1", "session.ct:talkerfemale2"))
Estimate Std. Error z value Pr(>|z|)
session.ct:talkerfemale1 -0.225 0.149 -1.514 0.13
session.ct:talkerfemale2 -0.815 0.226 -3.613 0.00

The effect of session is only reliable for the more intelligible talker (female 2), although means are in the same direction in each case.

female 1 : pre = 0.66 post = 0.7 female 2 :pre = 0.87 post = 0.93

Children

Note that a model containing correlations between slopes didn’t converge, these are removed in the model below (ie using the “(x||participant)” syntax)

discrim.child = selectCenter(discrim.child, list("session", "word_oldnew", "voice_oldnew", "talker", "condition", "agegroup", "mean_replay_in_training"))

child.lmer <- glmer(accuracy ~ 1 
                    + (condition.ct  * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                    + mean_replay_in_training.ct 
                    + (word_oldnew.ct * voice_oldnew.ct * session.ct||participant),
                    data = discrim.child, 
                    control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun=1e5)), family = binomial)

get_coeffs(child.lmer, c("(Intercept", "condition.ct", "session.ct", "word_oldnew.ct", "condition.ct:session.ct", "session.ct:word_oldnew.ct",
                         "session.ct:voice_oldnew.ct", "condition.ct:session.ct:word_oldnew.ct", "condition.ct:session.ct:voice_oldnew.ct", "condition.ct:word_oldnew.ct:voice_oldnew.ct", "session.ct:word_oldnew.ct:voice_oldnew.ct", "condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 0.870 0.080 10.941 0.000
condition.ct -0.249 0.160 -1.552 0.121
session.ct -0.671 0.099 -6.774 0.000
word_oldnew.ct -0.027 0.072 -0.378 0.705
condition.ct:session.ct -0.490 0.198 -2.478 0.013
session.ct:word_oldnew.ct -0.250 0.145 -1.730 0.084
session.ct:voice_oldnew.ct 0.216 0.154 1.405 0.160
condition.ct:session.ct:word_oldnew.ct 0.073 0.290 0.251 0.802
condition.ct:session.ct:voice_oldnew.ct -0.020 0.308 -0.064 0.949
condition.ct:word_oldnew.ct:voice_oldnew.ct 0.085 0.290 0.294 0.769
session.ct:word_oldnew.ct:voice_oldnew.ct -0.134 0.300 -0.445 0.656
condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct -0.042 0.602 -0.070 0.944

There was no reliable main effect of word-novelty. There was a reliable effect of test-session, indicating an effect of training (pre-test 0.62, post test 0.74). There was a marginal interaction between word-novelty and test session (untrained words 0.63 –> 0.74, untrained words 0.6 –> 0.75 However there is reliable interaction between test-session and condition. This is the reverse to the predicted direction - greater improvement in the low variability condition (0.18) than in the highvariablity condition (0.08)

Breaking this down:

discrim.child = selectCenter(discrim.child, list("session", "word_oldnew", "voice_oldnew", "talker", "condition", "agegroup", "mean_replay_in_training"))

child.lmer.v2 <- glmer(accuracy ~ 1
                       + condition.ct : session
                       + (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * talker.ct)
                       - condition.ct : session.ct
                       - condition.ct
                       + mean_replay_in_training.ct 
                       + (word_oldnew.ct * voice_oldnew.ct * session.ct||participant),
                       data = discrim.child, 
                       control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 1e5)), family=binomial)
anova(child.lmer, child.lmer.v2 )
## Data: discrim.child
## Models:
## child.lmer: accuracy ~ 1 + (condition.ct * session.ct * word_oldnew.ct * 
## child.lmer:     voice_oldnew.ct * talker.ct) + mean_replay_in_training.ct + 
## child.lmer:     (word_oldnew.ct * voice_oldnew.ct * session.ct || participant)
## child.lmer.v2: accuracy ~ 1 + condition.ct:session + (condition.ct * session.ct * 
## child.lmer.v2:     word_oldnew.ct * voice_oldnew.ct * talker.ct) - condition.ct:session.ct - 
## child.lmer.v2:     condition.ct + mean_replay_in_training.ct + (word_oldnew.ct * 
## child.lmer.v2:     voice_oldnew.ct * session.ct || participant)
##               Df    AIC    BIC  logLik deviance Chisq Chi Df Pr(>Chisq)
## child.lmer    41 4920.5 5180.1 -2419.2   4838.5                        
## child.lmer.v2 41 4920.5 5180.1 -2419.2   4838.5     0      0  < 2.2e-16
##                  
## child.lmer       
## child.lmer.v2 ***
## ---
## Signif. codes:  0 '***' 0.001 '**' 0.01 '*' 0.05 '.' 0.1 ' ' 1
get_coeffs(child.lmer.v2, c("condition.ct:sessionpost", "condition.ct:sessionpre"))
Estimate Std. Error z value Pr(>|z|)
condition.ct:sessionpost -0.003 0.192 -0.018 0.986
condition.ct:sessionpre -0.494 0.184 -2.683 0.007

There is a reliable difference between the conditions at pre-test (low variability: 0.56, high-variability: 0.66) but although children in the low-variability have caught up by post-test, they have not overtaken the high-variability condition (low variability: 0.74, high-variability: 0.74).

Check interactions with talker

get_coeffs(child.lmer.v2, c("talker.ct", "condition.ct:session.ct:talker.ct", "session.ct:word_oldnew.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 0.801 0.074 10.820 0.000
condition.ct:session.ct:talker.ct -0.358 0.308 -1.160 0.246
session.ct:word_oldnew.ct:talker.ct -0.167 0.300 -0.556 0.578

There is a reliable main effect of talker, reflecting female 2 being overall more intelligible (female 1 0.6, female 2 0.76. )

The interactions between word-type and session, and condition and session do not interact with talker. However, since the benefit of low-variability is unexpected, particularly for untrained voices, we consider whether the pattern is ordinal across counterbalances. Looking at the difference scores by talker:

discrim.child.talker = aggregate(accuracy ~  session + condition + talker + voice_oldnew, na.rm = T, data = discrim.child, FUN=mean)

discrim.child.talker.2 = reshape(discrim.child.talker, timevar = "session", idvar = c("condition", "talker","voice_oldnew"), direction = "wide")

discrim.child.talker.2$accuracy.diff = discrim.child.talker.2$accuracy.post - discrim.child.talker.2$accuracy.pre
kable(discrim.child.talker.2, digits = 3)
condition talker voice_oldnew accuracy.post accuracy.pre accuracy.diff
1 highvar female1 newvoice 0.686 0.568 0.118
3 lowvar female1 newvoice 0.691 0.509 0.182
5 highvar female2 newvoice 0.814 0.725 0.089
7 lowvar female2 newvoice 0.827 0.600 0.227
9 highvar female1 oldvoice 0.664 0.604 0.061
11 lowvar female1 oldvoice 0.623 0.473 0.150
13 highvar female2 oldvoice 0.804 0.750 0.054
15 lowvar female2 oldvoice 0.841 0.682 0.159

This suggests that for both counterbalances, for both trained and untrained voices, there is more improvement in the low variability condition.

Age Group Comparison

discrim = selectCenter(discrim, list("session", "word_oldnew", "voice_oldnew", "talker", "condition", "agegroup", "mean_replay_in_training"))

age.lmer <- glmer(accuracy ~ 1 
            + (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * agegroup.ct * talker.ct)
            + mean_replay_in_training.ct 
            + (word_oldnew.ct * voice_oldnew.ct * session.ct||participant),
            data = discrim, 
            control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun = 1e5)), family=binomial)

get_coeffs(age.lmer, c("(Intercept)","agegroup.ct", "session.ct:agegroup.ct", "condition.ct:session.ct:agegroup.ct",  "session.ct:word_oldnew.ct:agegroup.ct", "session.ct:voice_oldnew.ct:agegroup.ct", "condition.ct:session.ct:word_oldnew.ct:agegroup.ct", "condition.ct:session.ct:voice_oldnew.ct:agegroup.ct", "session.ct:word_oldnew.ct:voice_oldnew.ct:agegroup.ct", "condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:agegroup.ct"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.203 0.060 19.890 0.000
agegroup.ct -0.807 0.126 -6.427 0.000
session.ct:agegroup.ct -0.172 0.170 -1.015 0.310
condition.ct:session.ct:agegroup.ct -0.537 0.340 -1.580 0.114
session.ct:word_oldnew.ct:agegroup.ct -0.316 0.261 -1.211 0.226
session.ct:voice_oldnew.ct:agegroup.ct 0.034 0.267 0.128 0.898
condition.ct:session.ct:word_oldnew.ct:agegroup.ct -0.174 0.522 -0.333 0.739
condition.ct:session.ct:voice_oldnew.ct:agegroup.ct 0.876 0.534 1.642 0.101
session.ct:word_oldnew.ct:voice_oldnew.ct:agegroup.ct 0.506 0.521 0.969 0.332
condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:agegroup.ct -1.995 1.038 -1.922 0.055

Main effect of age-group (children 0.68, adults 0.79)

No reliable interaction between age-group and test-session and no higher interaction with any combination of condition, word-novetly or voice-novelty. There was however, a near reliable five-way interaction (reflecting different factors in the models above)

Check interactions with talker

get_coeffs(age.lmer, c("talker.ct", "agegroup.ct:talker.ct", "condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:talker.ct"))
Estimate Std. Error z value Pr(>|z|)
talker.ct 1.171 0.068 17.283 0.000
agegroup.ct:talker.ct -0.825 0.138 -5.974 0.000
condition.ct:session.ct:word_oldnew.ct:voice_oldnew.ct:talker.ct 1.103 0.999 1.104 0.269

There was a effect of talker. Moreover the main effect of age group was qualified by an interaction with talker (the five way interaction was not). Breaking down talker by age group interaction:

age.lmer.v2 <- glmer(accuracy ~ 1 
                     + agegroup.ct:talker
                     + (condition.ct * session.ct * word_oldnew.ct * voice_oldnew.ct * agegroup.ct * talker.ct)
                     - agegroup.ct: talker.ct
                     - agegroup.ct
                     + mean_replay_in_training.ct 
                     + (word_oldnew.ct * voice_oldnew.ct * session.ct||participant),
                     data = discrim, 
                     control = glmerControl(optimizer = "bobyqa", optCtrl = list(maxfun=1e5)), family = binomial)
anova(age.lmer, age.lmer.v2)
## Data: discrim
## Models:
## age.lmer: accuracy ~ 1 + (condition.ct * session.ct * word_oldnew.ct * 
## age.lmer:     voice_oldnew.ct * agegroup.ct * talker.ct) + mean_replay_in_training.ct + 
## age.lmer:     (word_oldnew.ct * voice_oldnew.ct * session.ct || participant)
## age.lmer.v2: accuracy ~ 1 + agegroup.ct:talker + (condition.ct * session.ct * 
## age.lmer.v2:     word_oldnew.ct * voice_oldnew.ct * agegroup.ct * talker.ct) - 
## age.lmer.v2:     agegroup.ct:talker.ct - agegroup.ct + mean_replay_in_training.ct + 
## age.lmer.v2:     (word_oldnew.ct * voice_oldnew.ct * session.ct || participant)
##             Df    AIC  BIC  logLik deviance Chisq Chi Df Pr(>Chisq)
## age.lmer    73 7987.2 8492 -3920.6   7841.2                        
## age.lmer.v2 73 7987.2 8492 -3920.6   7841.2     0      0          1
get_coeffs(age.lmer.v2,c("(Intercept)","agegroup.ct:talkerfemale1", "agegroup.ct:talkerfemale2"))
Estimate Std. Error z value Pr(>|z|)
(Intercept) 1.203 0.060 19.890 0.000
agegroup.ct:talkerfemale1 -0.395 0.131 -3.017 0.003
agegroup.ct:talkerfemale2 -1.219 0.155 -7.877 0.000

The difference between age groups holds for each talker.