Causal mapping is a 50-year tradition used in many disciplines including cognitive psychology, ecology, decision science, management and evaluation.

Causal Map Ltd provides the Causal Map app for evaluation and qualitative research.

Its niche: coding causal claims in text, closer to CAQDAS/NVivo than to a diagramming tool.

The app: app.causalmap.app. Free for unlimited public projects!

→ System dynamics, CLDs, FCMs

→ Causal Map

In this text-coding tradition, a causal map first records what people say causes what.

• Each link belongs to a quote, a source, a context
• Aggregating across sources is a deliberate later step
• The map shows stakeholder thinking, not the shape of the world

A CLD is usually a hypothesis about the system. This map shows the distribution of claims across people.

P and Q’s separate beliefs, then combined. Each link still carries source metadata.

A factor is a short proposition, close to what someone said.

Not “money” as a variable from low to high.

What this minimalist style does not code on the link itself:

• necessity or sufficiency
• non-linearity or polarity
• mediator, moderator, inhibitor roles
• strength or effect size

Bare links plus good labels let analysts:

• aggregate across sources without arguing about what symbols mean
• compare across groups, sites, projects
• summarise with simple counts (citations, sources)
• code thousands of claims at workable speed

It works fully by hand. AI makes it faster at scale.

Hierarchical labels let a dense map zoom out without deleting detail.

Zoom to level 1:

New intervention -> Healthy behaviour

Opposites coding handles pairs like:

• Employed and ~Employed
• Good health and ~Good health
• Eating vegetables and ~Eating vegetables

The ~ marks the opposite pole. Combining opposites rewrites both poles to one canonical label, while keeping polarity metadata.

Coding rarely sits inside a fixed codebook from the start.

Revisit the codebook as coding proceeds. Merge near-duplicates, split overloaded factors, lift to a useful level of abstraction.

Quality questions: are factors at a consistent grain? Do labels mean the same thing across sources? Have AI-coded factors drifted from the intended vocabulary?

Code the smallest defensible claim: this source says X influenced Y.

First quality move: keep the unit small, traceable and inspectable.

Manual or AI, the check is the same: is there a causal claim here, and are the endpoints right?

Check each raw claim before it feeds the map.

Mark what matters: tags, conviction, source reliability, custom columns.

These link-level checks flow into filters and bundle summaries downstream.

A bundle is all claims X→Y across all sources. Either collapse it into one assessed link, or skip it if evidence is too thin.

The workflow:

1. Finish coding; fix on the bundles that survive your filters
2. For each bundle: read the quotes, apply your rubric, judge quality
3. Create one assessed link (e.g. confidence 1-5), or none
4. Switch the view: unassessed claims stay in the database; the assessed map is what you argue from

The app requires a written rubric before assessed links. Deliberately.

Typical scale: 1,000 raw claims → 30 bundles → 25 assessed links

Unassessed view: many raw claims per bundle

Assessed view: one vouched link per bundle

A→B plus B→C does not automatically mean A→C. Contexts may not line up.

Source tracing asks: did anyone tell the whole story from intervention to outcome?

Once a pathway is credible, ask:

• Did this pathway matter?
• Was the contribution large enough to care about?
• How does it compare with alternative explanations?
• Did it produce valuable, harmful or ambiguous consequences?

Counts and maps help. The analyst weighs significance, context and rivals.

The final judgement is about the whole account.

A vignette can draft this whole-map judgement. You check it.

13 countries   692 outcomes   5,430 claims

INTRAC was MEL partner for a 13-country civil-society programme. Outcome Harvesting gave rich data, but not the big picture.

Zero-shot AI coding: no pre-set codebook. Outputs: one programme map, 13 country maps, filtered views by objective and time period.

Burkina Faso, country view

Same database, filtered to Burkina Faso.

Advocacy, training and public mobilisation feed action on rights violations.

Factor and link sizes = citation counts. Arrowhead colour = sentiment (blue positive, red negative).

Every link traces to a quote. Nothing in the visual is inferred.

Coding is half the job. The database under the map is for questions.

Stack filters and you can answer most evaluation questions worth asking.

INTRAC: pathways from interventions to outcomes in a 13-country lobby and advocacy programme.

The method works fully without AI. Minimalist coding is a small, clear task, which suits LLMs.

Validation study (Powell and Cabral, 2025), QuIP corpus, GPT-4 temperature 0:

Composite score: correct endpoints, real causal claim, not hypothetical, correct direction.

Top-level overview maps from AI and human coding come out broadly the same.

It works because the prompt is narrow: “where is a causal claim, and what influences what?”