How CausalTargeted compares

CausalTargeted.jl estimates continuous and longitudinal causal parameters after identification: LMTP / MTP grids, positivity and omitted-confounder diagnostics, and small-n Super Learner profiles. Mediation TE / NDE / NIE grids live in CausalMediation.jl (soft façades remain here). Graphs and certificates live upstream in CausalDynamics.jl.

R packages lmtp and crumble are the closest conceptual analogues (not API identity; see NAMING.md). Ananke is the main Python LMTP reference. DoubleML is Neyman-orthogonal and related in spirit, not full LMTP parity.

Choose CausalTargeted when you already have (or can obtain) an IdentificationResult, want Julia-native LMTP/mediation grids with fold caches, and care about small-n defaults and certificates on the estimate.

Prefer lmtp / Ananke when your pipeline is already R or Python end-to-end, or you need a specialised option (e.g. GPU Riesz nets, survival LMTP flavours) that we deliberately do not claim.

Stack overview: ECOSYSTEM_COMPARISON.md.

flowchart LR
  subgraph juliaStack [Julia_typed_pipeline]
    G[Graph_CDM] --> ID[identify_certificate]
    ID --> SIM[simulate_counterfactual]
    ID --> EST[LMTP_mediation]
    ID --> PLOT[DAGMakie]
    SIM --> EST
  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 (targeted estimation)

Capability	CausalTargeted	R	Python
LMTP / continuous MTP δ-grids	Yes	Yes (lmtp)	Yes (Ananke)
Sequential / longitudinal LMTP	Yes	Yes (lmtp)	Yes (Ananke)
Interventional mediation TE/NDE/NIE	Soft façade → CausalMediation	Yes (crumble, tmle3 mediation)	Partial (Ananke)
Cross-fit Super Learner profiles	Yes (lean / rich / small-n)	Yes (sl3 + lmtp/tmle3)	Partial (DoubleML / EconML nuisances)
Point-treatment TMLE (CM/ATE)	— (use TMLE.jl)	Yes (tmle / tmle3)	Partial (Ananke, others)
Consumes upstream graph ID certificate	Unique	Partial (separate packages)	Partial (DoWhy closer)
Positivity atlas / diagnostics	Yes	Partial (lmtp)	Partial
Omitted-confounder sensitivity	Yes	Yes (sensemakr + glue)	Partial (DoWhy)
DiD / g-computation utilities	Yes	Yes (did / g-comp pkgs)	Partial (CausalML / DoWhy)
Full parity with every lmtp/crumble option	— (deliberate)	Yes	—

Julia neighbours

Package	Role
CausalDynamics.jl	`identify`, `IdentificationResult`, CDMs (comparison)
TMLE.jl	Point-treatment CM / ATE / AIE
DAGMakie.jl	Optional DAG figures (comparison)

What is distinctive here

Typed hand-off — plan_mtp / execute_estimand carry identification certificates into estimate metadata
Small-n defaults — learner profiles and checklists aimed at tens to low hundreds of units (see Small-n checklist)
Julia-native EIF grids — no RCall on the estimation path

What we deliberately do not claim

Full parity with every option in R lmtp / crumble (GPU Riesz nets, all mediation estimand flavours, survival LMTP). See Methods.

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