Qualitative Data Analysis

Moderate · Data & Analysis track · ~35 min hands-on + readings and quiz

← Back to the Data & Analysis track

Use AI to assist coding and themes — without outsourcing your interpretation.

What you’ll be able to do

  • Use AI to propose codes and group themes in open-text data
  • Build and apply a codebook, then verify agreement
  • Weigh the debate over AI in qualitative work and protect confidentiality

Overview

AI can suggest codes, group themes, and summarize patterns in interview or open-text data quickly. But qualitative scholars warn that outsourcing interpretation strips out the reflexivity that gives the method its meaning. Use AI to assist your reading, not replace it.

Transcripts are often sensitive. De-identify them and use an approved tool — confidentiality comes first.

Practice activities

Activity 1 · Moderate — Draft a codebook from open-text

Time ~18 min · Tools ChatGPT Edu (approved tool for this content)

Goal. Build a codebook from qualitative responses while keeping your interpretation central.

Setup. Use a small, non-sensitive set of open-ended responses. You can use the free-text “description” field from a 311 Service & Information Requests slice, or the eight sample walkability comments below:

“Cars fly down our street, no stop signs.” · “Sidewalks are cracked and hard with a stroller.” · “I feel safe walking to the train in the morning.” · “No crosswalk near the school, kids dart across.” · “Too dark at night, the lights are out.” · “Love the new bump-outs, easier to cross.” · “Trash and parked cars block the corner.” · “The bus stop has nowhere to sit or shelter.”

Steps.

  1. Draft a codebook:

    Propose an initial qualitative codebook for these walkability comments: 5–7 codes, each with a name, a one-sentence definition, and an example quote. Note any response that could fit more than one code.

  2. Revise with your own judgment — rename, merge, or split codes. You own the meaning.

  3. Apply it back:

    Using my revised codebook, code each comment and show your reasoning for each.

Expected result. A codebook (codes + definitions + examples) that you have revised.

Check your work. Do the codes capture what respondents meant, or just what was easy to label? Adjust accordingly.

Common pitfalls. AI flattens nuance and imposes its own frame — treat it as a first pass, not the analysis. De-identify any real transcripts and use an approved tool.

Stretch (optional). Ask AI to surface a theme across codes, then find a quote that complicates that theme.

Activity 2 · Moderate — Code a new batch and check agreement

Time ~13 min · Tools ChatGPT Edu

Goal. Apply the codebook and check reliability against your own coding.

Setup. Use this second batch of 8 walkability comments (different from Activity 1):

“The new bike lane is great but now there’s nowhere to load groceries.” · “My wheelchair can’t get over the curb at 5th and Spring Garden.” · “Drivers run the red at Broad and Lehigh every single day.” · “We need more benches — older people can’t walk far without resting.” · “The crossing signal doesn’t last long enough to get across safely.” · “I’d walk to work but there’s no sidewalk on the last two blocks.” · “The school zone signs are hidden behind a tree branch.” · “I feel safer since they added the speed bumps on my block.”

Steps.

  1. Have AI code them:

    Using this codebook, code each of these new responses and show which code(s) apply and why.

  2. Independently code the same eight yourself.

  3. Compare:

    Here is the AI’s coding and mine. Where do we disagree, and what’s the likely reason?

Expected result. A coded batch plus a short list of disagreements.

Check your work. Disagreements should point to fuzzy code definitions — tighten them.

Common pitfalls. High agreement can be false comfort if the codebook is shallow. Your reading is the standard, not the AI’s.

Stretch (optional). Refine the codebook from the disagreements and re-run the batch.

Activity 3 · Moderate — Apply Your Codebook Across Multiple Files with Claude Cowork

Time ~20 min · Tools Claude Cowork

Goal. Apply the codebook you built in Activities 1–2 consistently across both batches of walkability comments stored as separate files — the way real qualitative data actually arrives — while keeping every interpretive decision yourself.

Setup. Create a folder on your computer called walkability-coding. Save Batch 1’s eight comments as batch1.txt (one comment per line) and Batch 2’s eight comments as batch2.txt. Add a third file, codebook.txt, containing your codes and their definitions from the earlier activities (e.g., traffic safety, infrastructure condition, accessibility, personal safety/lighting). Open Claude Cowork and connect the walkability-coding folder. Unlike pasting into a chat window, Cowork reads the files directly — no copy-paste, and the same codebook is applied to every file.

Steps.

  1. Ask Cowork to code both files against your codebook and combine the results:

    Read codebook.txt, then apply those codes to every comment in batch1.txt and batch2.txt. Produce one combined table with columns: source file, comment, suggested code(s), and a one-line rationale for each. Use only codes from my codebook — do not invent new ones. These are suggestions for my review, not final coding decisions.

  2. Read the combined table against your own coding of at least four comments — two from each batch. Where Cowork disagrees with you, ask it to explain its rationale, then decide yourself. Your reading wins; the table is a first pass, not an answer key.

  3. Ask it to surface its own uncertainty:

    Go back through both files and flag any comment where you were not confident which code applied, or where two codes seemed equally plausible. List those separately with the competing codes and why each could fit.

Expected result. One combined coded table covering all sixteen comments across both files, plus a shorter flagged list of ambiguous comments (e.g., “The crossing signal doesn’t last long enough” could be traffic safety or accessibility) queued for your judgment.

Check your work. Compare Cowork’s codes against your own on your sample — do you agree at least most of the time, and can you articulate why where you don’t? Every low-confidence comment gets read and coded by you personally; never accept a flagged item on the tool’s rationale alone.

Common pitfalls. Treating the combined table as finished coding — it’s a draft that you adjudicate, exactly as in Activities 1 and 2. And the data warning matters more here, not less: this workflow is fine because these comments are fabricated and non-sensitive, but Claude and Claude Cowork are not Drexel-supported and are approved for Low Risk Data only. Connecting a folder of real interview transcripts to Cowork would hand sensitive data to an uncleared tool — real transcripts must be de-identified and analyzed in an approved, Drexel-cleared tool, full stop.

Stretch (optional). Add a batch3.txt with five new comments you write yourself, rerun the same prompt, and see whether the codebook holds up or a comment surfaces the need for a new code — your call to make.

Check your readiness

Answer these, then check — your score suggests whether to dive in or skim the readings first.

Useful resources