The focus of today isn’t on data analysis, so this last part is more of a “show and tell” of some simple things you could do with the data.

Comparing tenure

We left off with creating the yearSince variable. Let’s do some basic analysis on that. We’ll first turn “party” into a factor variable, and use that to look at the composition of the assembly:

# Check party composition
fullAssembly$party <- as.factor(fullAssembly$party)
summary(fullAssembly$party)

Now let’s look at summary statistics and a histogram for the years in which assembly members joined:

summary(fullAssembly$yearSince)
hist(fullAssembly$yearSince)

This looks like a very young assembly with some very old members! Let’s look at the oldest ones:

arrange(fullAssembly, yearSince) |> head()

Is this the same for both parties? Let’s create subsets for parties?

# Create party subsets
assemblyDem <- fullAssembly |> filter(party == "Democratic")
assemblyRep <- fullAssembly |> filter(party == "Republican")

Look at how they differ:

# Compare how long Dems & Rs are in the Assembly
summary(assemblyDem$yearSince)
summary(assemblyRep$yearSince)

This looks quite different! Democrats have some long-serving members, Republicans none! This becomes even clearer when we graph this in parallel violin plots:

ageViolin <-
  fullAssembly |> filter(party == "Republican" |
                            party == "Democratic") |> ggplot(aes(y = yearSince, x = party)) +
  coord_flip()+
  geom_violin(trim=FALSE)+
  geom_boxplot(width=0.1) +
  theme_minimal()
ageViolin

What an interesting difference! We have a couple of research projects right there? Why is this? Does it affect how the two parties behave in the assembly? Is it similar to other states?

Topic modeling

We have all this text! Let’s do some automated text analysis. We’ll use the quanteda library. We first need to tokenize the data. Let’s create a function that performs standard steps for this:

tokenize_bios <- function (data){
  bioTokens <-
    tokens(
      data$biography,
      remove_punct = TRUE,
      remove_numbers = TRUE,
      remove_symbols = TRUE
    )
  bioTokens <-
    tokens_remove(bioTokens, pattern = c(stopwords("en"), stopwords("es")))
  return(bioTokens)
}

Next we can create some world clouds. Quanteda actually has a feature for contrasting within a single word cloud, but we’ll create separate ones:

bio_wordCloud <- function(data){
  bioTokens <- tokenize_bios(data)
  dfmat_bio <- dfm(bioTokens) |>
    dfm_trim(
      min_termfreq = 0.8,
      termfreq_type = "quantile",
      max_docfreq = 0.1,
      docfreq_type = "prop"
    )
  textplot_wordcloud(dfmat_bio)  
}

bio_wordCloud(fullAssembly)
bio_wordCloud(assemblyDem)
bio_wordCloud(assemblyRep)

Hard to spot much of interest in there if we’re honest. Let’s see if we can see more with some LDA topic modeling. Again, let’s create a function to do this, extracting the top terms for all models, allowing us to specify the number of topics:

topicTerms <- function(data, k = 6) {
  # From https://tutorials.quanteda.io/machine-learning/topicmodel/
  bioTokens <- tokenize_bios(data)
  
  dfmat_bio <- dfm(bioTokens) |>
    dfm_trim(
      min_termfreq = 0.8,
      termfreq_type = "quantile",
      max_docfreq = 0.3,
      docfreq_type = "prop"
    )
  
  tmodBio <- textmodel_lda(dfmat_bio, k)
  return (terms(tmodBio, 10))
}


topicTerms(fullAssembly)
topicTerms(assemblyDem)
topicTerms(assemblyRep)

Note that LDA topic models start at a random location: you won’t always get the same results, especially on small corpora such as ours.

The Finished Code

Here is the finalized script of the entire process for you to look through.