### Supplementary materials for
### Martin Hilpert and David Correia Saavedra
### The unidirectionality of semantic changes in grammaticalization:
### An experimental approach to the asymmetric priming hypothesis


### Analysis of mturk priming experiment
### Data structures in mturk.RData:
### cleanRT = the data set
### modC3full = the first model we fitted
### modC3 = the pruned model that we present in the paper
### modC3small = the second model we present, based on a subset of the data 


# load libraries and relevel the lexical vs. grammatical variable
library(languageR) 
library(lmerTest)
cleanRT$lex.gram<-relevel(cleanRT$lex.gram, "lex") # get lex as the reference level



## model with all control variables
modC3full<-lmer(cleanRT$ReacTime.ln~ cleanRT$Priming*cleanRT$lex.gram + cleanRT$LemFreq..ln + cleanRT$Handedness + cleanRT$age + cleanRT$Gender + cleanRT$LemFreq..ln:cleanRT$Priming +(1|cleanRT$Worker.ID) + (1|cleanRT$Stimulus.ID))
summary(modC3full)


## model with all control variables plus random slope
modC3fullslope<-lmer(cleanRT$ReacTime.ln~ cleanRT$Priming*cleanRT$lex.gram + cleanRT$LemFreq..ln + cleanRT$Handedness + cleanRT$age + cleanRT$Gender + cleanRT$LemFreq..ln:cleanRT$Priming +(1|cleanRT$Worker.ID) + (1|cleanRT$Stimulus.ID) + (1 + cleanRT$Trial.Number|cleanRT$Worker.ID))
summary(modC3fullslope) # lots of warnings



# the pruned model
modC3<-lmer(cleanRT$ReacTime.ln~ cleanRT$Priming*cleanRT$lex.gram + cleanRT$LemFreq..ln +(1|cleanRT$Worker.ID) + (1|cleanRT$Stimulus.ID))
summary(modC3) # Table 3
anova(modC3, modC3full) # pruned model no worse than full model


# the pruned model with random slope
modC3slope<-lmer(cleanRT$ReacTime.ln~ cleanRT$Priming*cleanRT$lex.gram + cleanRT$LemFreq..ln +(1|cleanRT$Worker.ID) + (1|cleanRT$Stimulus.ID)+ (1 + cleanRT$Trial.Number|cleanRT$Worker.ID))
summary(modC3slope) # again, lots of warnings
anova(modC3, modC3slope) # slope model no better than slimmer model




# Figure 2
boxplot(cleanRT$ReacTime~cleanRT$alphCondition, notch=T, col="grey", ylab="reaction times in ms", names=c("primed gram", "unprimed gram", "primed lex", "unprimed lex") ) # boxplot overview with logged RTs


# model with only eight high-freq elements
smallRT<-subset(cleanRT, Element %in% c("about", "come", "get", "go", "have", "help", "need", "use")) # creating the data subset

# the model based on the subset
modC3small<-lmer(smallRT$ReacTime.ln~ smallRT$Priming*smallRT$lex.gram + smallRT$LemFreq..ln +(1|smallRT$Worker.ID) + (1|smallRT$Stimulus.ID))
summary(modC3small) # Table 4

# Boxplot that shows reaction times across conditions, like Figure 2
boxplot(smallRT$ReacTime~smallRT$alphCondition, notch=T, col="grey", ylab="reaction times in ms", names=c("primed gram", "unprimed gram", "primed lex", "unprimed lex") ) # boxplot overview with logged RTs