(Powell et al., 2023)

Source: book chapter draft in content/000 Articles/020 !! toc book chapter.md.

• Purpose

• Show how QuIP-style causal mapping can compare an official programme ToC (“their theory”) with empirically coded beneficiary narratives (“our theory”), as a disciplined way to revise ToCs and “middle-level theory”.

• Minimal definition (what a causal map is)

• A causal map is nodes + directed links, where a link means (at minimum) someone believes C influenced E.

• Links need not encode necessity/sufficiency, nor quantified strength/polarity (though those are sometimes added in other approaches).

• QuIP as causal mapping

• Goal-free / (partially) blindfolded elicitation of stories of change reduces confirmation bias.
• “Causal back-chaining” elicits causes, causes-of-causes, etc.

• Coding is inductive and multi-source; maps are then filtered/queried to answer evaluation questions.

• How analysis is actually done

• Global maps are too large; use filters (e.g. theme/keyword searches, distance steps, frequency thresholds).
• Hierarchical coding / zooming out: encode subfactors in factor labels so detailed factors can be rolled up into higher-level factors for readable summary maps.

• Evidence strength is often shown with counts on links (mentions / sources).

• Interpretation pitfalls (explicitly listed)

• Beliefs about causation are not facts about causation: evaluator judgement remains separate.
• Absence of a mentioned link is not evidence of absence (random-walk conversations; negative cases).
• Transitivity trap / context overlap: stitching A→B (source 1) and B→C (source 2) does not justify A→C unless contexts overlap.
• Aggregation/generalisation is non-trivial; counts support confidence but don’t convert to “truth percentages”.