How CausalMediation compares

CausalMediation estimates modern mediation contrasts after identification: interventional TE / NDE / NIE under modified treatment policies, natural and organic effects when admissible, controlled direct effects, and recanting-twin / path-specific summaries. Intermediate confounders (moc) are first-class. Super Learner, ShiftPolicy, and fold helpers come from CausalTargeted.jl; graphs and certificates from CausalDynamics.jl.

R packages crumble, medoutcon, and medRCT are the closest conceptual analogues (not API identity; see Naming). Ananke covers some mediation-adjacent targets in Python; LMTP without mediators stays in CausalTargeted / R lmtp.

Choose CausalMediation when you want Julia-native mediation with typed moc, shared IdentificationResult hand-off, and the same Super Learner stack as LMTP.

Prefer crumble / medoutcon when the analysis pipeline is already R end-to-end, or you need a specialised option this package deliberately does not claim (full GPU Riesz nets, every survival flavour).

Stack overview: ECOSYSTEM_COMPARISON.md.

flowchart LR
  subgraph juliaStack [Julia_typed_pipeline]
    G[Graph_CDM] --> ID[identify_certificate]
    ID --> MED[CausalMediation]
    ID --> LMTP[CausalTargeted_LMTP]
    ID --> PLOT[DAGMakie]
  end

Legend

Mark	Meaning
`Yes`	First-class, documented
`Partial`	Possible with glue or a limited API
`—`	Not in that package’s usual scope
`Unique`	Strong differentiator here

Versus R and Python (mediation)

Capability	CausalMediation	R	Python
Interventional TE/NDE/NIE + continuous MTP	Yes	Yes (`crumble` RI, `medoutcon`)	Partial (Ananke)
Intermediate confounding (`moc`)	Yes	Yes (`crumble`, `medoutcon`, `medRCT`)	Partial
Natural effects (empty `moc`)	Yes	Yes (`crumble` `"N"`)	Partial
Organic effects	Yes	Yes (`crumble` `"O"`)	—
Recanting-twin / path-specific	Yes	Partial (`crumble` `"RT"`)	—
Controlled direct effect	Yes	Yes (VanderWeele / related)	Partial
Typed ID certificate → estimate	Unique	Partial	Partial
Cross-fit Super Learner (shared with LMTP)	Yes (via CausalTargeted)	Yes (sl3 + crumble)	Partial
LMTP without mediators	— (use CausalTargeted)	Yes (`lmtp`)	Yes (Ananke)
Full parity with every `crumble` option	— (deliberate)	Yes	—

Julia neighbours

Package	Role
CausalDynamics.jl	`MediationQuery`, `identify`, `IdentificationResult.moc`
CausalTargeted.jl	SL, `ShiftPolicy`, LMTP; soft mediation façades
DAGMakie.jl	Optional DAG figures
TMLE.jl	Point-treatment CM / ATE (not mediation)

What is distinctive here

Typed hand-off — plan_mediation / spec_from_identification carry adjustment, mediators, and moc from the certificate into estimation
Effect families as types — InterventionalMediation, NaturalMediation, … rather than opaque string brands
Shared nuisances with LMTP — same lean Super Learner profiles and shift policies as CausalTargeted

What we deliberately do not claim

Full parity with every option in R crumble / medoutcon (GPU Riesz nets, all survival mediation flavours). Continuous MTP one-step uses an outcome-residual augmentation of the nested-MC plugin (see Methods); binary contrasts use the fuller EIF path.

The CDCS book walks identify → estimate → display end to end.