A quick explanation of the columns:

• Language = Type of signed language
• Filename = The file annotation was performed in. A proxy for participant. 1 file = 1 participant.
• NarrativeReferent = The referent, also containing information status information in the prefix.
• Animacy = Doesn’t denote human. Also has “both” for when there are double referents.
• Activation = information status
• SignType = coocurrance of refering behaviors.

Clean up $NarrativeReferent coding

• First manually check for typos and alternate naming by listing out the levels of $NarrativeReferent (This has been commented out for the markdown), then change them to the preset values. (This will change as new files (languages) are added).

# levels(master$NarrativeReferent)
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyf"] <- "boy"
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyfNEW"] <- "boyNEW"
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyfREIN"] <- "boyREIN"
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyf?"] <- "boy?"
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyf?NEW"] <- "boy?NEW"
levels(master$NarrativeReferent)[levels(master$NarrativeReferent)=="boyf?REIN"] <- "boy?REIN"
# ^^ these aren't actually typos, for the record

• Subset for the referents that are actually a part of this pre-analysis (Note, this does not include uncertain referents at this time, but uncommenting out the referents with question marks would achieve this purpose.)

master1 <- subset(master, NarrativeReferent %in% 
                    c("boy","boyNEW","boyREIN",#"boy?","boy?NEW","boy?REIN",
                    "dog","dogNEW","dogREIN",#"dog?","dog?NEW","dog?REIN",
                    "frog","frogNEW","frogREIN",#"frog?","frog?NEW","frog?REIN",
                    "bees","beesNEW","beesREIN", #"bees?","bees?NEW","bees?REIN",
                    "owl","owlNEW","owlREIN",#"owl?","owl?NEW","owl?REIN",
                    "deer","deerNEW","deerREIN",#"deer?","deer?NEW","deer?REIN",
                    "jar","jarNEW","jarREIN",#"jar?","jar?NEW","jar?REIN",
                    "boots","bootsNEW","bootsREIN",#"boots?","boots?NEW","boots?REIN",
                    "window","windowNEW","windowREIN",#"window?","window?NEW","window?REIN",
                    "hive","hiveNEW","hiveREIN",#"hive?","hive?NEW","hive?REIN",
                    "rock","rockNEW","rockREIN")) #,"rock?","rock?NEW","rock?REIN",))
master1$NarrativeReferent <- factor(master1$NarrativeReferent) #remove empty levels
levels(master1$NarrativeReferent) #list levels

##  [1] "bees"       "beesNEW"    "beesREIN"   "boots"      "bootsNEW"  
##  [6] "boy"        "boyNEW"     "boyREIN"    "deer"       "deerNEW"   
## [11] "deerREIN"   "dog"        "dogNEW"     "dogREIN"    "frog"      
## [16] "frogNEW"    "frogREIN"   "hive"       "hiveNEW"    "hiveREIN"  
## [21] "jar"        "jarNEW"     "jarREIN"    "owl"        "owlNEW"    
## [26] "owlREIN"    "rock"       "rockNEW"    "rockREIN"   "window"    
## [31] "windowNEW"  "windowREIN"

Clean up $SignType coding

• First we need to separate the occurance of each sign type into its own vector. To do this we have a custom function named “clean” I forget who wrote this and I need to credit them somehow.

clean <- function(input) {
  names = names(input)
  for (i in 1:nrow(input)) {
    row = input[i,]
    split = unlist(strsplit(as.character(row$SignType), "[.]"))
    components = length(split)
    input[i,"component.count"] = components
    for (j in 1:components) {
      # take the component from the sign type, i.e. "ds" and turn it into a valid R
      # identifier prefixed with "type.", i.e. "type.ds". "?" is not valid in R
      # identifiers, so it gets replaced with ".q", so "mouth?" becomes "type.mouth.q".
      name = split[j]
      #name = paste("type", name, sep=".") #i removed this line because it cluttered the column names
      name = sub("\\.?\\?$", ".q", name, perl=TRUE)
      if (!name %in% names) {
        input[name] = 0
        names <- c(names, name)
      }
      input[i,name] = input[i,name] + 1
    }
  }
  input
}

• Now we will create a reference vector for the observations and run the function, merging the results back into our master dataframe.

list <- 1:nrow(master1)
master1 <- cbind(list=list, master1)
temp.frame <- subset(master1, select = c("list","SignType"))
cleaned <- clean(temp.frame) 
colnames(cleaned)[1] <- "list"
master.df1 <- merge(cleaned, master1, all.x = T)

• Now view the typos which are apparent as unique vectors within the new data frame. This has been commented out for the markdown
• Rename and remove as appropriate.
• Also separate the weak (non-dominant) hand coding.

# str(master.df1)
master.df1$SignType <- gsub("(CA)", "ca",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(WhDS)", "WHds",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(Whds)", "WHds",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(.gest)", ".gesture",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(.lenxp)", ".lexnp",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(mouthing)", "mouth",master.df1$SignType, perl=TRUE)
master.df1$SignType <- gsub("(mouth ds)", "mouth.ds",master.df1$SignType, perl=TRUE)

master.df1$SignType <- gsub("(WH)", "WH.",master.df1$SignType, perl=TRUE)

master.df1$SignType <- factor(master.df1$SignType) #remove empty levels

• Rerun the clean function now that we have found and fixed the typos.

1. create a dataframe with just the index column and SignType
2. create a dataframe with everything else but SignType
3. Run “clean” function on first dataframe
4. Merge dataframes from (1) and (2) back together

temp.frame2 <- subset(master.df1, select = c("list","SignType"))
master.frame <- subset(master.df1, select =c("list", "Language", "Filename", "ClauseLikeUtterance", "NarrativeReferent", "Animacy", "Activation"))
cleaned2 <- clean(temp.frame2) 
colnames(cleaned2)[1] <- "list"
master.df2 <- merge(cleaned2, master.frame, all.x = T)
master.df2 <- master.df2[ ,!(colnames(master.df2) == "1")] #remove the pesky "1" from the SignType coding
#master.df3 <- master.df2[, -which(names(master.df2) %in% c("WHlex", "WHds", "WHpt","WhDS","lenxp","WHlexnp","CA","gest"))]
master.df3 <- master.df2
# summary(master.df3$WH) #check merge worked correctly
# summary(master.df3$Language) #check merge worked correctly

Clean up $Animacy and $Activation

1. Consolidate the the uncertain coding (.q)
2. Reorder dataframe
3. Remove extra object created in doing all this.
4. Set all vectors to correct classes and drop unused levels
5. Rename some variables for ease later on

#collapse all .q columns to one column (12)
master.df3$q <- master.df3$fs.q + master.df3$ds.q + master.df3$IN.q + master.df3$ca.q + master.df3$mouth.q + master.df3$.q
uncertain <- master.df3[, -which(names(master.df3) %in% c("fs.q", "ds.q", "IN.q","ca.q","mouth.q",".q","is.q","lex.q","pt.q","lexnp.q"))]
# data <- uncertain[c("list","Filename","Language","NarrativeReferent","Animacy","Activation","component.count","lex","mouth","ca","pt","fs","IN","WH","ds","lexnp","is","dl","gesture","q")] #old non-tidy way of selecting columns...
data <- uncertain %>%
  select(list,Filename,Language,NarrativeReferent,Animacy,Activation,component.count,lex,mouth,ca,pt,fs,IN,WH,ds,lexnp,is,dl,gesture,q)
rm(clean,cleaned, cleaned2, list, master, master.df1,master.df2, master.df3, master1, temp.frame,temp.frame2, uncertain)
# levels(data$Activation)
data$Activation <- mapvalues(data$Activation, from = c("NEW", "MAINTENANCE","REIN","?","UNCERTAIN"), to = c("New", "Maintained", "Reintroduced","Uncertain","Uncertain"))
data <- subset(data, !Activation %in% c("Uncertain"))
# levels(data$Animacy)
data$Animacy <- mapvalues(data$Animacy, from = c("ANIMATE", "INANIMATE","BOTH","?"), to = c("Animate", "Inanimate","Both","Uncertain"))
data <- subset(data, !Animacy %in% c("Both"))
data$Animacy <- factor(data$Animacy)
data$Activation <- factor(data$Activation)
data$NarrativeReferent <- factor(data$NarrativeReferent)
# `r nrow(subset(data, Language == "Auslan"))` to report inline

Confirm semiotic strategies are different

1. Convert semiotic strategies to long format
2. Run mixed effect model to see whether the freq of different strategies varies by activation, animation or language while accounting for individual variation.
3. Plot data

data_long2 <- melt(data,
        # ID variables - all the variables to keep but not split apart on
    id.vars=c("Language", "Activation","Animacy","Filename", "NarrativeReferent"),
        # The source columns
    measure.vars=c("lex", "mouth", "ca", "pt", "fs", "IN","WH", "ds", "lexnp", "is", "dl" ,"gesture", "q" ),
        # Name of the destination column that will identify the original
        # column that the measurement came from
    variable.name="strategy",
    value.name="occurance")
# head(data_long2)
long.table2 <- ddply(data_long2, c("Filename", "Language", "Activation","Animacy","strategy"),summarise, occurance = round(mean(occurance),4))
# long.table2

long.table2$occ.c <-  long.table2$occurance - mean(long.table2$occurance) #center
long.table2$occ.z <- long.table2$occ.c / sd(long.table2$occurance) #standardize
contrasts(data$Animacy) <- contr.sum(2)
contrasts(data$Animacy)
# head(long.table2)
long.table2$Activation2 <- factor(long.table2$Activation, levels = c("Maintained", "Reintroduced","New"))
# levels(long.table2$Activation2)
long.table2$Language2 <- factor(long.table2$Language, levels = c("Auslan","SSL","NSL","ISL", "FinSL"))
# levels(long.table2$Language2)
# head(long.table2)

#ideal model
# occ.m0 <- lmer(occ.z ~ strategy * Language2 + Activation + Animacy + (1+strategy*Language2|Filename), data = long.table2) #careful this takes about > 2hours 
# summary(occ.m0)

occ.m1 <- lmer(occ.z ~ strategy * Language2 + Activation + Animacy + (1+Language2|Filename), data = long.table2) 
summary(occ.m1)

Plot distribution of semiotic strategies

Heatmaps are my new favorite plot type

Auslan PCA

auslan.pca <- PCA(auslan.occurances, quali.sup = 1, graph = F) #run the pca
auslan.pca$eig #print eigen values
barplot(auslan.pca$eig[,2], names = 1:nrow(auslan.pca$eig), xlab = "Components", ylab = "Percentage of explained variance")

• 7 principle components accounts for 84.4451369 of the varience.
• We therefor choose this cut off point for the hierarchical clustering analysis.

However the bartlett test still finds the whole correlation frame to be significantly co-correlated. Unsure what to do with these mixed messages.

Auslan HCPC

• Using the eigen values we specify a new PCA.
• Inspect variables factor map (PCA) to find clustering. Cut off at 80% (PC should explain 80% of variance) In this case this means specifying 7 clusters
• Use PCA graph to eyeball number of clusters.
• Input max values to HCPC and graph.

auslan.df <- auslan.occurances #rename dataframe because will be changing labelling (rownames)
rownames(auslan.df) <- auslan.occurances$NarrativeReferent
auslan.df <- subset(auslan.df, select = -c(NarrativeReferent))
par(mfrow = c(1,1))
auslan.pca2 <- PCA(auslan.df,  ncp = 7, graph = T)

auslan.hcpc <- HCPC(auslan.pca2, max = 3,graph = F)

Characteristics of each cluster

Most central (prototypical) members of each cluster

auslan.hcpc$desc.ind$para

## Cluster: 1
##        jar       bees    jarREIN windowREIN        dog 
##  0.8255535  1.0604249  1.1469012  1.4797074  1.7129512 
## -------------------------------------------------------- 
## Cluster: 2
##  boyREIN      boy  owlREIN 
## 1.701857 3.976011 4.555127 
## -------------------------------------------------------- 
## Cluster: 3
## beesREIN  dogREIN frogREIN  rockNEW  beesNEW 
## 1.590161 1.703463 1.818677 1.870236 2.071075

The center of the cluster is 0. Referents closest to the center of the cluster are listed first.

Top semiotic behaviors in each cluster

auslan.table

##    cluster coding cluster.mean overall.mean p.value
## 2        1     IN        0.357        0.258   0.025
## 1        1     wh        0.271        0.167   0.010
## 7        1  mouth        0.135        0.433   0.000
## 5        1    lex        0.130        0.336   0.000
## 3        1     fs        0.054        0.178   0.012
## 11       2     ca        0.479        0.148   0.008
## 9        2     dl        0.218        0.032   0.000
## 10       2   gest        0.002        0.000   0.002
## 8        3  mouth        0.844        0.433   0.000
## 6        3    lex        0.560        0.336   0.001
## 4        3     fs        0.372        0.178   0.001
## 12       3  lexnp        0.108        0.044   0.024

NSL PCA

NSL.pca <- PCA(NSL.occurances, quali.sup = 1, graph = F) #run the pca
NSL.pca$eig #print eigen values
barplot(NSL.pca$eig[,2], names = 1:nrow(NSL.pca$eig), xlab = "Components", ylab = "Percentage of explained variance")

• 6 principle components accounts for 81.8981726 of the varience.
• We therefor choose this cutoff point for the hierarchical clustering analysis.

However the bartlett test still finds the whole correlation frame to be significantly co-correlated. Unsure what to do with these mixed messages.

NSL HCPC

• Using the eigen values we specify a new PCA.
• Inspect variables factor map (PCA) to find clustering. Cut off at 80% (PC should explain 80% of variance) In this case this means specifying 7 clusters
• Use PCA graph to eyeball number of clusters.
• Input max values to HCPC and graph.

NSL.df <- NSL.occurances #rename dataframe because will be changing labelling (rownames)
rownames(NSL.df) <- NSL.occurances$NarrativeReferent
NSL.df <- subset(NSL.df, select = -c(NarrativeReferent))
par(mfrow = c(1,1))
NSL.pca2 <- PCA(NSL.df,  ncp = 6, graph = T)

NSL.hcpc <- HCPC(NSL.pca2, max = 3,graph = F)

Characteristics of each cluster

Most central (prototypical) members of each cluster

NSL.hcpc$desc.ind$para

## Cluster: 1
##      bees   jarREIN      deer       jar       dog 
## 0.9618113 1.0964162 1.2447966 1.3369165 1.7729490 
## -------------------------------------------------------- 
## Cluster: 2
##  boyREIN      boy  dogREIN 
## 0.551440 1.036889 1.081964 
## -------------------------------------------------------- 
## Cluster: 3
##  frogREIN windowNEW    dogNEW   owlREIN   beesNEW 
## 0.8143005 1.1826362 1.3241244 1.8657728 1.9044351

The center of the cluster is 0. Referents closest to the center of the cluster are listed first.

Top semiotic behaviors in each cluster

NSL.table

##    cluster coding cluster.mean overall.mean p.value
## 1        1     ds        0.683        0.513   0.010
## 7        1  mouth        0.284        0.503   0.000
## 4        1     wh        0.226        0.143   0.047
## 9        1    lex        0.148        0.366   0.000
## 3        1     is        0.029        0.017   0.030
## 5        1     dl        0.000        0.003   0.035
## 11       2     ca        0.530        0.139   0.001
## 6        2     dl        0.033        0.003   0.000
## 8        3  mouth        0.927        0.503   0.000
## 10       3    lex        0.756        0.366   0.000
## 2        3     ds        0.290        0.513   0.046

# FinSL Clustering Analysis

• Summarize occurance of each sign type by referent
• Run barlett test
• Check eigen levels for HCPC

FinSL <- subset(data, Language == "FinSL")
FinSL.occurances <- ddply(FinSL, c("NarrativeReferent"),summarise, 
                  lex = mean(lex),lexnp = mean(lexnp), fs = mean(fs), mouth = mean(mouth), 
                  dl = mean(dl), ds = mean(ds), pt = mean(pt), is = mean(is), 
                  gest = mean(gesture), ca = mean(ca),IN = mean(IN), 
                  wh = mean(WH), q = mean(q)) 
# FinSL.occurances <- subset(FinSL.occurances, select = -c(fs)) #remove those with low instances
FinSL.cortab <- round(cor(FinSL.occurances[, 2:(ncol(FinSL.occurances))], method="spearman"),2)^2 #spearman correlation has no assumptions about the distribution of the data
FinSL.cortab

##          lex  lexnp     fs  mouth     dl     ds     pt     is   gest
## lex   1.0000 0.1521 0.0324 0.6724 0.0016 0.0064 0.0676 0.0729 0.0009
## lexnp 0.1521 1.0000 0.0324 0.0169 0.0004 0.0081 0.0025 0.0784 0.0324
## fs    0.0324 0.0324 1.0000 0.0100 0.0100 0.0100 0.0121 0.0225 0.0049
## mouth 0.6724 0.0169 0.0100 1.0000 0.0289 0.0324 0.0529 0.0625 0.0484
## dl    0.0016 0.0004 0.0100 0.0289 1.0000 0.0064 0.0169 0.0961 0.0576
## ds    0.0064 0.0081 0.0100 0.0324 0.0064 1.0000 0.1089 0.0441 0.0049
## pt    0.0676 0.0025 0.0121 0.0529 0.0169 0.1089 1.0000 0.0036 0.0196
## is    0.0729 0.0784 0.0225 0.0625 0.0961 0.0441 0.0036 1.0000 0.0081
## gest  0.0009 0.0324 0.0049 0.0484 0.0576 0.0049 0.0196 0.0081 1.0000
## ca    0.0049 0.0196 0.0169 0.0036 0.3969 0.0361 0.1024 0.0169 0.0049
## IN    0.1089 0.0000 0.0400 0.1849 0.0529 0.0676 0.0025 0.0169 0.0025
## wh    0.0961 0.0025 0.0009 0.1089 0.0009 0.1764 0.0009 0.0004 0.0169
## q     0.0484 0.0036 0.0400 0.1764 0.2116 0.0324 0.0484 0.0000 0.0324
##           ca     IN     wh      q
## lex   0.0049 0.1089 0.0961 0.0484
## lexnp 0.0196 0.0000 0.0025 0.0036
## fs    0.0169 0.0400 0.0009 0.0400
## mouth 0.0036 0.1849 0.1089 0.1764
## dl    0.3969 0.0529 0.0009 0.2116
## ds    0.0361 0.0676 0.1764 0.0324
## pt    0.1024 0.0025 0.0009 0.0484
## is    0.0169 0.0169 0.0004 0.0000
## gest  0.0049 0.0025 0.0169 0.0324
## ca    1.0000 0.0121 0.0289 0.3025
## IN    0.0121 1.0000 0.0004 0.0049
## wh    0.0289 0.0004 1.0000 0.0064
## q     0.3025 0.0049 0.0064 1.0000

FinSL.occurances2 <- subset(FinSL.occurances, select = -c(NarrativeReferent))
FinSL.corr <- round(cor(FinSL.occurances2), 1)
ggcorrplot(FinSL.corr, hc.order = TRUE, 
           type = "lower", 
           lab = TRUE, 
           lab_size = 3, 
           method="circle", 
           colors = c("tomato2", "white", "springgreen3"), 
           title="Correlogram of FinSL Semiotic strategies", 
           ggtheme=theme_bw)

FinSL.cortab <- as.data.frame(round(cor(FinSL.occurances[, 2:(ncol(FinSL.occurances))], method="spearman"),2)^2)
FinSL.cortab$Index <- apply(FinSL.cortab, 1, function(x) paste(colnames(FinSL.cortab)[which(x > .3 & x < .9)], collapse = ", "))
FinSL.cortab.report <- FinSL.cortab$Index[FinSL.cortab$Index != ""]
FinSL.cortest <- cortest.bartlett(FinSL.occurances[, 2:(ncol(FinSL.occurances))]) #if p sig, then co-correlated

## R was not square, finding R from data

• Check that variables are intercorrelated. Following Levshina:353 absolute correlations should be between .3 and .9
• These semiotic strategies were correlated with another between 0.3 - 0.9: “mouth, lex, ca, dl, q, ca.”
• Yet the Bartlett’s test found significant co-correlated between stragegies: p < 0.00000. Therefor we continue with the PCA.

FinSL PCA

FinSL.pca <- PCA(FinSL.occurances, quali.sup = 1, graph = F) #run the pca
FinSL.pca$eig #print eigen values
barplot(FinSL.pca$eig[,2], names = 1:nrow(FinSL.pca$eig), xlab = "Components", ylab = "Percentage of explained variance")

• 7 principle components accounts for 83.8887325 of the varience.
• We therefor choose this cutoff point for the hierarchical clustering analysis.

However the bartlett test still finds the whole correlation frame to be significantly co-correlated. Unsure what to do with these mixed messages.

FinSL HCPC

• Using the eigen values we specify a new PCA.
• Inspect variables factor map (PCA) to find clustering. Cut off at 80% (PC should explain 80% of variance) In this case this means specifying 7 clusters
• Use PCA graph to eyeball number of clusters.
• Input max values to HCPC and graph.

FinSL.df <- FinSL.occurances #rename dataframe because will be changing labelling (rownames)
rownames(FinSL.df) <- FinSL.occurances$NarrativeReferent
FinSL.df <- subset(FinSL.df, select = -c(NarrativeReferent))
par(mfrow = c(1,1))
FinSL.pca2 <- PCA(FinSL.df,  ncp = 7, graph = T)

FinSL.hcpc <- HCPC(FinSL.pca2, max = 3,graph = F)

Characteristics of each cluster

Most central (prototypical) members of each cluster

FinSL.hcpc$desc.ind$para

## Cluster: 1
##     window windowREIN       hive        jar    jarREIN 
##   1.140464   1.307465   1.368188   1.589189   1.611950 
## -------------------------------------------------------- 
## Cluster: 2
##       boy   boyREIN   dogREIN       dog 
## 0.7944186 1.4300421 1.5308766 1.7525523 
## -------------------------------------------------------- 
## Cluster: 3
##  frogREIN   beesNEW windowNEW   owlREIN    boyNEW 
## 0.6867219 1.6537180 1.6548275 1.7366753 2.3579961

The center of the cluster is 0. Referents closest to the center of the cluster are listed first.

Top semiotic behaviors in each cluster

FinSL.table

##    cluster coding cluster.mean overall.mean p.value
## 4        1     ds        0.653        0.512   0.046
## 3        1     IN        0.392        0.281   0.010
## 9        1    lex        0.325        0.504   0.000
## 7        1  mouth        0.302        0.479   0.001
## 1        1     wh        0.170        0.110   0.007
## 5        1     dl        0.000        0.015   0.021
## 11       2     ca        0.626        0.149   0.000
## 6        2     dl        0.119        0.015   0.000
## 8        3  mouth        0.856        0.479   0.000
## 10       3    lex        0.850        0.504   0.000
## 2        3     wh        0.023        0.110   0.022

SSL PCA

SSL.pca <- PCA(SSL.occurances, quali.sup = 1, graph = F) #run the pca
SSL.pca$eig #print eigen values
barplot(SSL.pca$eig[,2], names = 1:nrow(SSL.pca$eig), xlab = "Components", ylab = "Percentage of explained variance")

• 5 principle components accounts for 80.0348481 of the varience.
• We therefor choose this cutoff point for the hierarchical clustering analysis.

However the bartlett test still finds the whole correlation frame to be significantly co-correlated. Unsure what to do with these mixed messages.

SSL HCPC

• Using the eigen values we specify a new PCA.
• Inspect variables factor map (PCA) to find clustering. Cut off at 80% (PC should explain 80% of variance) In this case this means specifying 7 clusters
• Use PCA graph to eyeball number of clusters.
• Input max values to HCPC and graph.

SSL.df <- SSL.occurances #rename dataframe because will be changing labelling (rownames)
rownames(SSL.df) <- SSL.occurances$NarrativeReferent
SSL.df <- subset(SSL.df, select = -c(NarrativeReferent))
par(mfrow = c(1,1))
SSL.pca2 <- PCA(SSL.df,  ncp = 5, graph = T)

SSL.hcpc <- HCPC(SSL.pca2, max = 3,graph = F)

Characteristics of each cluster

Most central (prototypical) members of each cluster

SSL.hcpc$desc.ind$para

## Cluster: 1
##    deerNEW windowREIN     window       rock        jar 
##  0.6799906  0.9348407  1.0278956  1.0747697  1.0785145 
## -------------------------------------------------------- 
## Cluster: 2
## windowNEW  beesREIN  frogREIN  bootsNEW   beesNEW 
## 0.9748434 0.9848163 1.0773566 1.6910901 1.8270525 
## -------------------------------------------------------- 
## Cluster: 3
##  owlREIN  boyREIN      dog      boy   owlNEW 
## 1.368911 2.172187 2.185436 2.644263 3.543535

The center of the cluster is 0. Referents closest to the center of the cluster are listed first.

Top semiotic behaviors in each cluster

SSL.table

##    cluster coding cluster.mean overall.mean p.value
## 3        1     ds        0.672        0.525   0.033
## 1        1     IN        0.401        0.269   0.006
## 5        1     wh        0.298        0.185   0.043
## 8        1  mouth        0.188        0.473   0.000
## 6        1    lex        0.163        0.401   0.000
## 9        2  mouth        0.933        0.473   0.000
## 7        2    lex        0.901        0.401   0.000
## 11       3     ca        0.448        0.168   0.001
## 12       3     pt        0.248        0.106   0.005
## 10       3     dl        0.193        0.058   0.001
## 4        3     ds        0.185        0.525   0.006
## 14       3  lexnp        0.172        0.053   0.036
## 13       3     fs        0.076        0.020   0.015
## 2        3     IN        0.043        0.269   0.008

ISL PCA

ISL.pca <- PCA(ISL.occurances, quali.sup = 1, graph = F) #run the pca
ISL.pca$eig #print eigen values
barplot(ISL.pca$eig[,2], names = 1:nrow(ISL.pca$eig), xlab = "Components", ylab = "Percentage of explained variance")

• 6 principle components accounts for 84.5028666 of the varience.
• We therefor choose this for the hierarchical clustering analysis.

However the bartlett test still finds the whole correlation frame to be significantly co-correlated. Unsure what to do with these mixed messages.

ISL HCPC

• Using the eigen values we specify a new PCA.
• Inspect variables factor map (PCA) to find clustering. Cut off at 80% (PC should explain 80% of variance) In this case this means specifying 6 clusters
• Use PCA graph to eyeball number of clusters.
• Input max values to HCPC and graph.

ISL.df <- ISL.occurances #rename dataframe because will be changing labelling (rownames)
rownames(ISL.df) <- ISL.occurances$NarrativeReferent
ISL.df <- subset(ISL.df, select = -c(NarrativeReferent))
par(mfrow = c(1,1))
ISL.pca2 <- PCA(ISL.df,  ncp = 6, graph = T)

ISL.hcpc <- HCPC(ISL.pca2, max = 3,graph = F)

Characteristics of each cluster

Most central (prototypical) members of each cluster

ISL.hcpc$desc.ind$para

## Cluster: 1
##       frog       bees        dog        owl windowREIN 
##   1.121817   1.332203   1.485279   1.718586   1.753860 
## -------------------------------------------------------- 
## Cluster: 2
##  owlREIN      boy  dogREIN deerREIN  boyREIN 
## 1.016299 1.233968 1.617385 1.619432 2.158819 
## -------------------------------------------------------- 
## Cluster: 3
##  hiveREIN  beesREIN    owlNEW windowNEW   rockNEW 
##  1.333214  1.777564  1.788830  1.849798  1.893235

The center of the cluster is 0. Referents closest to the center of the cluster are listed first.

Top semiotic behaviors in each cluster

ISL.table

##    cluster coding cluster.mean overall.mean p.value
## 1        1     IN        0.535        0.336   0.003
## 6        1  mouth        0.148        0.461   0.000
## 3        1     wh        0.121        0.062   0.021
## 8        1    lex        0.074        0.362   0.000
## 4        1     fs        0.057        0.157   0.032
## 10       2     ca        0.662        0.210   0.000
## 2        2     IN        0.113        0.336   0.025
## 12       2     dl        0.015        0.003   0.001
## 7        3  mouth        0.724        0.461   0.000
## 9        3    lex        0.621        0.362   0.000
## 5        3     fs        0.289        0.157   0.003
## 13       3     pt        0.185        0.108   0.022
## 11       3     ca        0.072        0.210   0.028

Mixed effects model

• For activation and animacy on semiotic strategy count.

data$c.cc <- data$component.count - mean(data$component.count) #center
data$z.cc <- data$c.cc / sd(data$component.count) #standardize
contrasts(data$Animacy) <- contr.sum(2)
# contrasts(data$Animacy)
# levels(data$Language2)
# head(data)
cc.m <- lmer(z.cc ~ Activation2 * Animacy + Language2 + (1+Animacy|Filename), data = data)
summary(cc.m)

Yes it appears these patterns are significant. Just as we found for Auslan, when we control for random variation in individuals and referents we find a cross-linguistic pattern for:

1. Maintained referents (estimate: -0.3077103) receive less semiotic strategies than reintroduced (estimate:0.1720772, p < 0.00000) and reintroduced receive less new (estimate: 0.8047572, p < 0.00000).
2. Animate new referents (estimate: -0.1128695, p < 0.00359) recieve less inanimate referents (-0.5025512, p < 0.00359).

• But, these data show that SSL (estimate: 0.102843, p < 0.19596) does not follow these patterns.
• The plots below show that this is because SSL also maintains a difference in the number of semiotic strategies between inanimate (mean = 1.92) and animate (mean = 1.46) for maintained referents, not only new referents.
• Note that ISL, while not significantly different (estimate: 0.1302788, p < 0.09350), is showing a very different pattern in the reintroduced referents inanimate (mean = 1.82) receiving less than animate (mean = 2.49)

Plots