Narrative Synthesis: integrating studies you cannot pool

It is a synthesis method inside a systematic process. The searching, screening, and appraisal are exactly as systematic as in a meta-analytic review; only the integration step differs.

It is not a narrative review. A narrative review selects sources at the author’s discretion; narrative synthesis inherits its studies from documented criteria and must account for all of them — including the inconvenient ones.

It is not failed meta-analysis. Choosing narrative synthesis because heterogeneity makes pooling meaningless is a methodological decision to defend, not a shortfall to apologise for. Pooling incommensurable studies produces a precise average of nothing.

The touchstone is Popay and colleagues’ Guidance on the Conduct of Narrative Synthesis in Systematic Reviews (ESRC Methods Programme), which organises the work into four elements: a theory of how the intervention works, a preliminary synthesis, exploration of relationships within and between studies, and an assessment of the synthesis’s own robustness.

Reporting now has a dedicated standard: Campbell and colleagues’ Synthesis without meta-analysis (SWiM) in systematic reviews: reporting guideline, nine items covering how studies were grouped, which synthesis metric was used, and how heterogeneity was investigated. SWiM exists because reviews claiming “narrative synthesis” so often reported none of that.

For the mechanics of describing effects without pooling, the Cochrane Handbook’s chapter on summarising study results (Wiley) covers the defensible metrics — direction of effect, standardised summaries, and structured tabulation.

A review of community health worker programmes for hypertension control included 21 studies: six trials, nine quasi-experiments, six controlled before–after designs, with outcomes measured five different ways. Pooling was ruled out — the protocol’s heterogeneity thresholds were exceeded on design and outcome alike — and the SWiM checklist was adopted for reporting.

The theory of change (task-shifting frees clinician time and increases contact frequency; contact frequency drives adherence; adherence drives control) generated the groupings. The preliminary synthesis tabulated all 21 studies with effect directions; eighteen pointed the same way, but the exploration stage did the real work: effects concentrated where community health workers had medication-adjustment authority, and vanished where they only educated. Two negative trials were both authority-free programmes — a pattern, not an embarrassment.

The robustness element removed the five weakest studies (per appraisal) and the pattern held; it also noted that all three studies from one funder shared an outcome measure that inflated apparent effects, and quarantined claims resting on it. The published conclusion was mechanism-specific and quality-bounded — precisely what vote counting could never have produced.

• Vote counting. “Twelve of eighteen studies were positive” ignores size, precision, and quality; SWiM explicitly warns against it, and it remains the most common failure.
• Ordering by author or year. A synthesis grouped by nothing analyses nothing.
• Appraising quality and then ignoring it. If weak and strong studies read identically in the text, the appraisal was decoration.
• Patterns without named studies. Every “effects were larger in rural settings” needs its supporting and contradicting studies listed.
• Calling a narrative review a narrative synthesis. Reviewers check for the systematic scaffolding; the wrong label invites the wrong standard and then fails it.

Without pooling there is no confidence interval on the headline claim; the precision of a narrative conclusion is inherently qualitative, and readers who want a number will be frustrated.

The method is more exposed to reviewer judgement than meta-analysis — grouping choices and pattern-reading both admit discretion — which is why the theory of change and named-studies discipline exist, and why two teams can still reasonably differ.

And it is slower to write than it looks: structured integration of 21 incommensurable studies is harder than feeding 21 commensurable ones to a pooling routine.

A systematic literature review is the container; narrative synthesis is one of its synthesis engines.

Meta-synthesis integrates qualitative findings, where “effect direction” gives way to interpretive translation.

Realist synthesis also asks why effects vary, but organises the whole review around mechanisms rather than treating them as an exploration stage.

The tabulation and grouping stages are structured extraction — fields applied across PDFs — and AI-assisted analysis does that quickly and audibly: Evidano returns each extracted characteristic with the passage it came from, so the effect-direction table can be verified without re-reading every study.

The exploration element is different: reading patterns across studies against a theory of change is analytic judgement. Software can retrieve and juxtapose; deciding that authority, not education, is the active ingredient remains the reviewer’s claim to defend.