Source
TL;DRAn influential framework (Nelson 2020) combining unsupervised topic modeling with qualitative reading — but one that rests on three flawed assumptions, per Carlsen & Ralund's critique. Their replacement, CALM, keeps humans as the interpretive ground truth.

The original framework (Nelson 2020)

Nelson coined “computational grounded theory” (CGT) to align unsupervised machine learning with grounded theory’s inductive, data-driven spirit. Three steps:

  1. Pattern discovery — unsupervised models (typically LDA topic models) find clusters in the corpus without researcher-imposed categories
  2. Pattern refinement — researcher reads paradigmatic documents per topic and interprets what the model found
  3. Pattern confirmation — indirect validation by correlating topic measures with external variables

The appeal: scalable, seemingly inductive, requires minimal manual reading. Popular in computational social science.

Why it doesn’t work (Carlsen & Ralund 2022)

(carlsen-ralund-computational-grounded-theory-2022) systematically critiques all three steps:

Discovery fails: LDA enforces uniform topic sizes and fixed topic counts — real corpora don’t look like this. The result is fused, duplicated, and unstable topics. Word co-occurrence doesn’t reliably map to semantic meaning; you can’t trust unsupervised models to find the right categories.

Minimal immersion fails: Reading only model-selected “paradigmatic” documents cannot qualify a researcher to interpret meaning within a community. Paradigmatic cases are only meaningful relative to extensive prior reading. CGT shortcuts the very immersion that makes grounded theory valid.

Indirect validation fails: Correlating a topic measure with another variable cannot detect systematic measurement error. If the model consistently miscodes documents, external correlations will still appear. Only direct validation — human coding of a random sample — catches this.

CALM: the replacement framework

Carlsen & Ralund’s Computer Assisted Learning and Measurement (CALM):

Stage What happens Who leads
Discovery HSBM/word clustering generates candidate categories and search terms Computer proposes; human decides
Interpretation Search terms retrieve documents; researcher reads extensively, writes memos, builds coding scheme Human
Classification Coding scheme applied; supervised ML scales it Human codes examples; computer scales
Validation Direct validation against human-coded test set Human
Measurement Validated classifier applied to full corpus Computer

The key move: the computer solves the rarity and scale problem (finding enough relevant cases for saturation, then scaling classification), but the human remains the interpretive ground truth throughout.

Relationship to LLM-assisted research

CALM was designed before the LLM era, but its logic applies directly. Modern LLM-based approaches (like bijker-chatgpt-qca-2024) that have LLMs generate coding schemes and code data wholesale are essentially the computer-led paradigm CGT represents — they inherit the same epistemological risks. The CALM corrective: use LLMs to surface candidate themes and assist discovery, while insisting on human grounding and direct validation.

See also

What links here