Qualitative Comparative Analysis (QCA): a practical guide

Equifinality. Several distinct causal paths can lead to the same outcome. A regression coefficient assumes one additive pathway; QCA expects and reports multiple, and a result showing three different routes to success is a substantive finding rather than a modelling failure.

Conjunctural causation. Conditions work in combination, not in isolation. A condition may be irrelevant on its own and decisive in the presence of two others. QCA is built to express that; interaction terms in a regression express it badly beyond two or three variables.

Asymmetry. The conditions that produce an outcome are not simply the negation of those that prevent it. QCA analyses presence and absence separately, and the two analyses regularly identify different conditions. Treating them as mirror images is one of the commonest errors.

Underneath all three sits set theory: a case is a member of the set of "high-performing sites" to some degree, and analysis asks about necessity (the outcome only occurs where this condition is present) and sufficiency (wherever this combination is present, the outcome follows).

QCA was introduced by Charles Ragin in The Comparative Method (1987) and extended to fuzzy sets in Fuzzy-Set Social Science (2000). The practical reference for both variants is Rihoux and Ragin's edited Configurational Comparative Methods — see Crisp-Set Qualitative Comparative Analysis and Qualitative Comparative Analysis using Fuzzy Sets.

The standard technical treatment, and the source for the diagnostic thresholds most reviewers apply, is Schneider and Wagemann's Set-Theoretic Methods for the Social Sciences.

Its evaluation use was argued most influentially by Blackman and colleagues in Using Qualitative Comparative Analysis to understand complex policy problems. The variants in general use are csQCA (crisp sets: conditions are present or absent), fsQCA (fuzzy sets: graded membership between 0 and 1) and mvQCA (multi-value). fsQCA is now the default for evaluation work because most real conditions are matters of degree.

A foundation funded a community health-worker programme in 24 districts. Twelve reached the retention target; twelve did not. The evaluation question was what distinguished them.

Five conditions were selected from prior work and site knowledge: supervisory visit frequency, payment reliability, whether workers were selected by the community, distance to the referral facility, and whether a district health officer had championed the programme. Each was calibrated with substantive anchors — payment reliability, for instance, at full membership where no payment had been more than two weeks late in a year, full non-membership where delays exceeded six weeks more than twice, crossover at a single month-long delay.

Necessity analysis found payment reliability necessary at consistency 0.94 — no district hit the retention target without it — and relevance was acceptable, since a third of districts lacked it.

The truth table with a frequency threshold of 1 and a consistency threshold of 0.83 produced two sufficient paths: reliable payment and community selection and frequent supervision; or reliable payment and an official champion and short referral distance. Solution consistency 0.88, solution coverage 0.83.

The reading was that reliable payment was the floor, and above it two different arrangements worked: strong community ownership with close supervision, or strong institutional backing where the health system was physically accessible. Distance mattered in the second path and not the first. One deviant case — reliable payment, community selection, frequent supervision, and no outcome — turned out on return to the field to have lost its entire cohort to a mining-sector recruitment drive, which was reported as a boundary condition rather than folded into the model.

• Calibrating from the sample distribution. Anchors set at the median or at quartiles encode nothing substantive and make the results an artefact of the sample.
• Too many conditions. More than seven with a moderate number of cases guarantees limited diversity, and the solution then rests mostly on assumptions about configurations never observed.
• Reporting the parsimonious solution without saying so. It is the most compact and the most assumption-laden. Silence about which solution is being shown is a serious reporting failure.
• Assuming symmetry. The analysis of the negated outcome must actually be run. It frequently produces different conditions.
• Reading paths as variables. A condition appearing in one path is not "a driver of the outcome" in general; it operates in that combination.
• Never returning to the cases. Without case-level interpretation, QCA is Boolean algebra over a spreadsheet.
• Hiding the thresholds. Frequency and consistency cut-offs change the solution. Undeclared, they invite the suspicion that they were tuned.

QCA is sensitive to case selection and to calibration in a way that is uncomfortable for a formal method: small, defensible changes to anchors can alter the solution, which is why robustness checks across plausible alternative calibrations and thresholds are now expected rather than optional.

It does not handle measurement error gracefully — a single miscoded case can eliminate a necessity claim — and it says nothing about temporal order unless conditions are constructed to encode it.

It also does not establish mechanism. A solution path states that a combination is regularly associated with the outcome; why it produces it is a separate question, which is why Schneider and Rohlfing's Combining QCA and Process Tracing has become a standard follow-on design: QCA to find the configurations, process tracing within selected cases to establish what actually happened.

Process tracing is the standard complement, verifying mechanisms within cases the QCA identifies.

Case-based impact evaluation shares the comparative-case logic without the formal set-theoretic apparatus, and is the better choice below roughly ten cases.

Realist evaluation asks a similar configurational question in prose rather than in sets; QCA can be read as its formalisation where enough comparable cases exist.

Contribution analysis addresses a different problem — the strength of a claim about one programme — rather than patterns across many.

Minimisation itself needs dedicated software — the R packages QCA and SetMethods, or the fsQCA application — and no general qualitative tool substitutes for them.

The work a qualitative platform does sit upstream and downstream. Upstream: the condition scores usually have to be derived from case documents, interviews and site reports, which is a systematic document-analysis job across dozens of sources, and calibration is only defensible if each score traces back to the evidence behind it. Downstream: interpreting solution paths and investigating deviant cases means returning to that same material with a specific question. Evidano supports QCA as a named methodology on the case-material side; the set-theoretic analysis belongs in the dedicated packages.