Qualitative Data Science: Using RQDA to analyse interviews

Word clouds are a popular method for exploratory analysis of texts. The wordcloud is created with the text mining and wordcloud packages. The transcribed interviews are converted to a text corpus (the native format for the tm package) and whitespace, punctuation etc is removed. This code snippet opens the RQDA file and extracts the texts from the database. RQDA stores all text in an SQLite database and the package provides a query command to extract data.

digital-humanities

library(tidyverse)
library(RQDA)
library(tm)
library(wordcloud)
library(igraph)

openProject("stakeholders.rqda")
interviews <- data.frame(name = RQDAQuery("SELECT name FROM source"),
                         text = RQDAQuery("SELECT file FROM source"))
interviews <- Corpus(VectorSource(interviews$file))

interviews <-  tm_map(interviews, stripWhitespace)
interviews <-  tm_map(interviews, content_transformer(tolower))
interviews <-  tm_map(interviews, removeWords, stopwords("english"))
interviews <-  tm_map(interviews, removePunctuation)
interviews <-  tm_map(interviews, removeNumbers)
interviews <-  tm_map(interviews, removeWords, c("interviewer", "interviewee"))

# Word cloud
set.seed(1969)
wordcloud(interviews, min.freq = 10, max.words = 50, rot.per=0.35, 
          colors = brewer.pal(8, "Blues")[-1:-5])

This word cloud makes it clear that the interviews are about water businesses and customers, which is a pretty obvious statement. The interviews are also about the opinion of the interviewees (think). While the word cloud is aesthetically pleasing and provides a quick snapshot of the content of the texts, they cannot inform us about their meaning.

Topic modelling is a more advanced method to extract information from the text by assessing the proximity of words to each other. The topic modelling package provides functions to perform this analysis. I am not an expert in this field and simply followed basic steps using default settings with four topics.

dtm <- DocumentTermMatrix(interviews)
dtm <- removeSparseTerms(dtm, 0.99)
ldaOut <- LDA(dtm, k = 4)
terms(ldaOut,6)

This code converts the corpus created earlier into a Document-Term Matrix, which is a matrix of words and documents (the interviews) and the frequency at which each of these words occurs. The LDA function applies a Latent Dietrich Allocation to the matrix to define the topics. The variable k defines the number of anticipated topics. An LDA is similar to clustering in multivariate data. The final output is a table with six words for each topic.

This table does not tell me much at all about what was discussed in the interviews. Perhaps it is the frequent use of the word "water" or "think"—I did ask people their opinion about water-related issues. To make this analysis more meaningful, I could perhaps manually remove the words water, yeah, and so on. This introduces a bias in the analysis and reduces the reliability of the topic analysis because I would be interfering with the text.

Numerical text analysis sees a text as a bag of words instead of a set of meaningful words. It seems that any automated text mining needs a lot of manual cleaning to derive anything meaningful. This excursion shows that automated text analysis is not a sure-fire way to analyse the meaning of a collection of words. After a lot of trial and error to try to make this work, I decided to go back to my roots of qualitative analysis using RQDA as my tool.

To use RQDA for qualitative data science, you first need to manually analyse each text and assign codes (topics) to parts of the text. The image below shows a question and answer and how it was coded. All marked text is blue, and the codes are shown between markers. Coding a text is an iterative process that aims to extract meaning from a text. The advantage of this method compared to numerical analysis is that the researcher injects meaning into the analysis. The disadvantage is that the analysis will always be biased, which in the social sciences is unavoidable. My list of topics was based on words that appear in a marketing dictionary so that I analysed the interviews from that perspective.

My first step was to look at the occurrence of codes (themes) in each of the interviews.

## Load and transform data
codings <- getCodingTable()[,4:5]
categories <- RQDAQuery("SELECT filecat.name AS category, source.name AS filename 
                         FROM treefile, filecat, source 
                         WHERE treefile.catid=filecat.catid AND treefile.fid=source.id AND treefile.status=1")
codings <- merge(codings, categories, all.y = TRUE)
head(codings)

## Open coding
reorder_size <- function(x) {
    factor(x, levels = names(sort(table(x))))
}
ggplot(codings, aes(reorder_size(codename), fill=category)) + geom_bar(stat="count") + 
    facet_grid(~filename) + coord_flip() + 
    theme(legend.position="bottom", legend.title=element_blank()) + 
    ylab("Code frequency in interviews") + xlab("Code")

The code uses an internal RQDA function getCodingTable to obtain the primary data. The RQDAQuery function provides more flexibility and enables more complex queries of the data. You can view the structure of the RQDA database using the RQDATables function.

This bar chart helps to explore the topics that interviewees discussed, but it does not help to understand how these topic relate to each other. This method provides better insights into the text than the 'bag of words' approach. Qualitative data analysis provides meaning to a text that no algorithm can discern. Although this meaning is ultimately subjective, proper methodology clarified any biases.

RQDA provides a facility to assign each code to a code category. A network diagram is a good tool to visualise this structure. The igraph package provides extensive functionality to analyse graphs.

Qualitative data analysis can create value from a text by interpreting it from a given perspective. This article is not even an introduction to the science and art of qualitative data science. I hope it invites you to explore RQDA and similar tools.

If you are interested in finding out more about this method, then feel free to read chapter three of my dissertation on customer service in water utilities.

  edges <- RQDAQuery("SELECT codecat.name, freecode.name FROM codecat, freecode, treecode WHERE codecat.catid=treecode.catid AND freecode.id=treecode.cid")

  g <- graph_from_edgelist(as.matrix(edges), directed = FALSE) %>%
    simplify()
  V(g)$name <- gsub(" ", "\n", V(g)$name)

  c <- spinglass.community(g)
  par(mar=rep(0,4))
  set.seed(666)
  plot(c, g, 
       vertex.size = 10,
       vertex.color = NA,
       vertex.frame.color = NA,
       layout = layout.drl)

As seen on R Bloggers