Back to CNMA dashboard
MODULE GUIDE

Component network meta-analysis

Decompose treatments into components, and estimate effects for combinations the network can identify.

§ 1

What CNMA does

Component Network Meta-Analysis (CNMA) decomposes multi-component treatments into their constituent building blocks and estimates an effect for each component. From those component effects the model can predict the effect of any combination the network is able to identify — including combinations no trial directly observed.

The canonical example is triple therapy in COPD. Trials of ICS+LABA+LAMA, ICS+LABA, LABA+LAMA, and the three monotherapies form a network that lets CNMA isolate the effect of ICS, LABA, and LAMA separately. The same machinery works for multi-component psychotherapies, drug cocktails, complex care bundles, and other compositional interventions.

§ 2

The additivity assumption

CNMA assumes that the effect of a combination equals the sum of its components' effects on the additivity scale: the log scale for ratio measures such as OR, RR, and HR, and the original analysis scale for absolute and continuous measures such as RD, MD, and SMD [7]. That's a strong methodological claim, and one CoreSR surfaces explicitly rather than hides.

On every analyze pass, three R models fit in sequence: standard NMA via netmeta::netmeta, additive CNMA via netmeta::netcomb, and (when the network is small enough) an interaction CNMA — netmeta::netcomb refitted with forward-selected candidate interaction columns at alpha=0.05. On a disconnected network no standard NMA exists: the additive CNMA is fitted with netmeta::discomb (shared components bridge the sub-networks, and every published estimate is gated on its own identifiability), the Q-difference additivity test is not available, interaction CNMA is not applicable, and an active standard-NMA selection is switched to the additive model with a note. CoreSR treats this as a screening aid: retained interaction terms still need clinical plausibility rather than purely data-driven selection [7]. The Q-difference between standard and additive NMA is your test [1]:

  • p >= 0.10 - additivity supported. Predicted combinations can be considered, but unobserved combinations remain extrapolations that need clinical review.
  • 0.05 <= p < 0.10 - borderline. Consider switching to the interaction model.
  • p < 0.05 - additivity violated. Predicted combinations from the additive model are unreliable; the GRADE-CNMA rubric automatically downgrades the additivity-violation domain by 1 level (2 levels at p < 0.01).

You'll see a disclosure modal the first time you view results, conditioned on the actual Q-diff bucket. If a re-analysis changes the bucket, the disclosure re-fires.

§ 3

Identifiability — what can be estimated

Not every component can be estimated. A component column in the C-matrix is identifiable if it is linearly independent of the others. When two components always co-occur (e.g. one never appears without the other), netcomb cannot separate their effects.

The C-matrix editor runs an exact-arithmetic rank check on every cell toggle. Flagged components surface in the row rail as amber and trigger the "Show what would identify it" deep-dive — a plain-language explanation plus four affordances: merge with a co-occurring component, drop the flagged component, find an additional trial that separates them, or proceed and accept that the component will appear as "not identifiable" in plots and is excluded from GRADE.

§ 4

Extrapolation distance

Every predicted combination carries an extrapolation distance: the count of unordered component pairs in that combination that never co-occurred in any contributing trial. Distance 0 means every pair has been observed; the combination is interpolated. Distance ≥ 1 means the combination is extrapolated — its effect is built from additivity over component effects that aren't all directly comparable.

The GRADE-CNMA rubric treats extrapolation distance as a 7th deterministic-judge domain: distance 0 ⇒ not serious; distance 1 ⇒ serious (1 level down); distance ≥ 2 ⇒ very serious (2 levels down). The extrapolation-distance badge ED stays visible on every appearance of the combination — league table, combinations list, GRADE entity row, and SoF — so the user always knows whether a prediction is an interpolation or an extrapolation.

§ 5

GRADE-CNMA: three entity types

GRADE-CNMA grades three entity types in a dependency-ordered rail:

  1. Components — five standard domains (RoB, inconsistency, indirectness, imprecision, publication bias) plus the locked additivity-violation domain. Components downgrade network-wide whenever Q-diff is significant.
  2. Combinations — inherit worst-of-each-domain from their constituent components, then add a locked extrapolation-distance domain and any additional additivity penalty for combinations containing a retained interaction pair.
  3. Contrasts — worst-of-each-domain from the two combinations being compared, plus the contrast's own imprecision (computed from contrast-level CI vs MID).

Each domain carries a source-label badge (Lock = deterministic judge, Calculator = computed fact, Sliders = CoreSR heuristic, CoreSR mark = AI-derived, User = reviewer override). Locked domains cannot be overridden by the reviewer — to change them you change the underlying analysis (switch model, edit C-matrix). Everything else accepts a typed-rationale override.

§ 6

Async interaction CNMA

Forward-selection of interaction terms scales O(K²) in components (one netcomb refit per co-occurring pair). For networks with K > 12 components, interaction CNMA is decoupled from the synchronous run: standard NMA + additive CNMA always finish; the interaction fit is queued and notifies you when it completes. Candidate interactions are reviewed as clinically plausible signals, not automatic proof that all possible interactions should be searched. The R script never fabricates results when the compute genuinely cannot finish in time.

§ 7

Status: provisional rubric

Provisional GRADE-CNMA rubric - GRADE/CINeMA are not yet formally extended to component-specific estimates. The rate-down thresholds follow Petropoulou 2023 [1] for additivity and the unified imprecision framework of Hultcrantz 2025 [2] for CI-vs-MID comparison. Welton 2009 [3] introduced the early CNMA formulation; Rücker 2020 [4] implemented it in the netmeta R package. Veroniki 2026 [7] highlights that GRADE/CINeMA have not yet been formally extended to component-specific estimates, so CoreSR keeps component certainty labels explicitly provisional.

References

  1. [1] Petropoulou M, et al. Res Synth Methods 2023;14(2):265-279 (component NMA; additivity Q-diff)
  2. [2] Hultcrantz M, et al. J Clin Epidemiol 2025 (unified imprecision framework)
  3. [3] Welton NJ, et al. J Clin Epidemiol 2009;62:857-864 (early CNMA formulation)
  4. [4] Rücker G, et al. Biom J 2020;62:808-821 (netmeta package implementation)
  5. [5] Brignardello-Petersen R, et al. J Clin Epidemiol 2023;160:151-159 (GRADE Concept Paper 5 — intransitivity)
  6. [6] van Buuren S, Groothuis-Oudshoorn K. J Stat Softw 2011;45(3):1-67 (mice multiple imputation, Rubin's rules pooling)
  7. [7] Veroniki AA, Wolfe DM, Hutton B, Schwarzer G, McIsaac DI, Straus SE, Jackson D, Tricco AC. BMJ 2026;393:e083075 (practical CNMA guidance; additivity scale; interaction and reporting cautions). doi:10.1136/bmj-2024-083075