Nelson (2020) — Computational Grounded Theory: A Methodological Framework

TL;DR: The paper that coined “computational grounded theory” — a three-step framework (pattern detection, pattern refinement, pattern confirmation) combining unsupervised machine learning with qualitative deep reading to analyze large text corpora inductively. Influential in the computational social sciences as a principled bridge between quantitative scale and qualitative interpretation. The framework that carlsen-ralund-computational-grounded-theory-2022 later critiques at its foundations.

Problem

By the mid-2010s, computational text analysis had become standard in political science and some branches of psychology, but was struggling to gain traction in qualitative sociology. The reason was epistemological: sociologists interested in meaning, culture, and interpretation found the available computational methods — topic models, word counts, co-occurrence statistics — poorly suited to the interpretive goals of their research.

The question Nelson addresses: can computational methods be integrated into inductive, grounded-theory-style inquiry without betraying the core methodological commitments that make grounded theory valuable? The existing options seemed to force a choice — either abandon scale (use grounded theory and read every document) or abandon interpretation (use computational methods and lose the nuance).

Nelson argues this is a false choice. Computational methods can discover patterns that inform qualitative reading; qualitative reading can interpret those patterns in contextually grounded ways; further computational analysis can validate the resulting categories. Each step leverages the strengths of both approaches.

Approach

Computational Grounded Theory (CGT) has three steps:

Step 1: Pattern detection. Inductive computational exploration of the full text corpus using unsupervised methods — specifically LDA topic models and word scores. The computer operates without prior categories, surfacing patterns the researcher might miss through manual sampling. Nelson is explicit: the goal is not to replace the researcher’s interpretive judgment but to expand what the researcher can see before making interpretive judgments.

The key advantage: computational methods can process corpora of tens of thousands or millions of documents — far beyond what any researcher could read. This is the scale problem that grounded theory cannot solve on its own.

Step 2: Pattern refinement. Return to qualitative deep reading. Identify “paradigmatic” documents — those most representative of the computationally identified patterns — and read them closely. This is where the researcher’s interpretive competence, cultural knowledge, and theoretical insight come in. The computational patterns are hypotheses; the qualitative reading tests and interprets them.

Step 3: Pattern confirmation. Return to computation: use supervised classification or NLP to assess whether the patterns identified in Step 2 can be reliably measured across the full corpus. This step produces validated, reproducible categories that can be used for quantitative analysis.

The CGT framework is designed for text as data — interview transcripts, speeches, open-ended survey responses, field notes — and is explicitly positioned within the sociological tradition rather than computer science or computational linguistics.

AI’s Role

AI (in the form of unsupervised machine learning, specifically LDA topic models) is positioned as a pattern discovery instrument in Step 1 — surfacing structure in large text corpora that guides qualitative inquiry. This is one of the earliest principled frameworks for using computational methods in the service of qualitative interpretation rather than as an alternative to it.

Nelson does not address LLMs (the paper predates their widespread adoption), but CGT’s logic is a direct predecessor to much of the LLM-era qualitative research literature. The distinction between pattern detection (AI’s job) and interpretation (the researcher’s job) runs through frameworks as different as CALM, GAITA, and CAAI.

Epistemological Stance

Interpretivist with a scientific measurement ambition — the same tension that carlsen-ralund-computational-grounded-theory-2022 identify as productive in the CALM framework. Nelson wants both: interpretive depth and reproducible, scalable measurement. CGT is designed to honor both through the division of labor between computational and qualitative methods.

The grounded theory grounding is important: CGT is not quantitative content analysis with a qualitative gloss. The goal is genuinely inductive, theory-building inquiry. The computation serves the interpretive aims, not the other way around.

Rigor and Trustworthiness

The paper provides detailed methodological guidance and illustrative empirical applications (texts from US feminist newsletters). The three-step structure is clearly operationalized, with specific computational methods recommended for each step and specific quality criteria for evaluating the output of each.

The framework’s reproducibility claim is serious: because computational methods produce auditable outputs, CGT analysis can be reproduced and verified in ways that purely manual grounded theory cannot. This is one of CGT’s genuine methodological advantages.

Limitations

Step 1: LDA topic models enforce assumptions that may not hold in practice — a fixed number of topics and uniform topic distribution across the corpus. Carlsen & Ralund demonstrate through simulation that these assumptions cause LDA to produce fused topics (two real topics merged), duplicate topics, and instability across runs. The pattern detection step may not find the right patterns.

Step 2: Reading only paradigmatic documents does not constitute the immersion in the full corpus that genuine qualitative understanding requires. The researcher who has read the model’s top 10 documents per topic is not qualified to interpret the meaning of topics in the way a researcher who has read all 50,000 documents would be.

Step 3: Indirect validation (correlating topic measures with external variables) cannot detect systematic measurement error. If LDA consistently miscodes documents from a particular community, but those miscoded documents still correlate with external variables, the correlation passes — but the measurement is invalid.

These are the three critiques that Carlsen & Ralund’s simulation study demonstrates rigorously. They are not arbitrary objections; they are structural limitations of the framework as specified.

Connections

• computational-grounded-theory — the concept page that summarizes CGT, the Carlsen & Ralund critique, and the CALM alternative
• carlsen-ralund-computational-grounded-theory-2022 — the paper that critiques each of CGT’s three steps and proposes CALM as the replacement; read together as debate and response
• llm-qualitative-research — the modern context for which CGT is historical foundation; CGT’s logic shapes the entire LLM-era AI-TA literature’s division of labor between AI and researcher
• ubellacker-academiaos-2024 — the most automated modern extension of CGT-style thinking; what happens when the human is moved further from the interpretive center
• sinha-gpt4-grounded-theory-2024 — CGT’s logic applied with GPT-4 as the pattern detector; compare the updated approach with the original framework
• epistemic-flattening — the risk that CGT’s Step 1 amplifies: unsupervised models reproduce statistically dominant patterns, not the marginal meanings that careful qualitative reading recovers
• intercoder-agreement — CGT’s Step 3 uses supervised classification to validate categories; the validation approach is indirect (correlational) rather than direct (human-coded test set)
• validity-trustworthiness — the validity critique is the core of the Carlsen & Ralund argument; CGT’s indirect validation cannot detect systematic measurement error