This paper explains our Minimalist / Barefoot approach to coding causal claims in text as simple directed links (“X influenced Y”), developed through large-scale practical coding and now implemented in the Causal Map app. We write it now because, although our previous work motivates causal mapping in evaluation [@powellCausalMappingEvaluators2024], shows how QuIP-style “stories of change” elicit natively causal narrative evidence [@copestakeAttributingDevelopmentImpact2019], demonstrates ToC validation by comparing empirical maps with programme theory [@powellDoesOurTheory2023], and shows that genAI can extract links exhaustively with quotes as a low-level assistant [@powellAIassistedCausalMapping2025a; @powellWorkflowCollectingUnderstanding2025], none of these papers is a standalone, reader-facing account of the coding stance itself.

Abstract

This guide addresses a cluster of tricky but practical problems in causal coding: how to represent oppositeness, sentiment/valence, and “despite” constructions in a way that stays close to ordinary language and remains auditable.

This paper argues that causal mapping can be treated as a serious form of Qualitative Data Analysis (QDA): a disciplined variant in which the primary coding act is not “apply a theme”, but code a causal link (an ordered pair of factor labels) grounded in a quote and source. The resulting dataset is a structured, auditable qualitative model (a network of causal claims) that can be queried using a transparent library of operations (filtering, path tracing, label transforms, bundling). This gives researchers a way to keep qualitative judgement central while making key intermediate products more reproducible, checkable, and scalable—especially when using AI as a constrained, low-level coding assistant rather than a black-box analyst.

Suppose an evaluation team has a corpus of interviews and progress reports, plus (at least) two candidate theories of change (ToCs): an original one and a revised one. A practical question is: which ToC better fits the narrative evidence?

After inductive causal coding (manual or AI-assisted), you typically end up with many overlapping factor labels (“rising prices”, “inflation”, “cost of living increases”, …). Magnetisation (aka soft recoding) is a fast, transparent way to standardise these labels without re-coding the original text: you supply a list of target labels (“magnets”), and each existing label is reassigned to its closest magnet by semantic similarity (using embeddings). Unmatched labels can be kept (default) or dropped.

Abstract

See also: [[000 Minimalist coding for causal mapping]]; [[015 Combining opposites, sentiment and despite-claims]]; [[005 A formalisation of causal mapping]].

Thanks guys, there's lots to like here and lots to agree with. Helping to find what causal hypotheses to focus on... What do you have to say about pathways as opposed to individual links/mechanisms? If we called this approach "quality and rigour in causal mechanisms evaluation" would that miss anything?
Disclaimer: for me, the logic around how links might combine into pathways and what that means for evaluation, that's the most exciting part. e.g. how might this intervention influence an outcome which might be multiple steps downstream of it?