How CausalDynamics compares

CausalDynamics.jl owns identify → intervene → shared-U counterfactual trajectories for structural and discrete-time dynamical models. It composes with Julia neighbours and sits next to familiar R/Python graph tools—without trying to replace every CRAN or PyPI package.

Choose CausalDynamics when you need typed identification certificates, discrete-time CDMs, temporal unrolling, or SciML-backed continuous CDMs in the same language as estimation and plotting downstream.

Prefer dagitty / DoWhy when you want a GUI-first workflow, an existing Python four-step pipeline, or discovery-heavy tooling you already trust (Associations.jl and causal-learn remain better homes for full discovery).

Stack overview (integration as the product): 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

Julia neighbours

CausalDynamics composes with, rather than competes against, the packages below.

Need	Reach for	Why
Learn a DAG from data (PC, OCE)	Associations.jl	Optional weakdep bridge (`infer_pc_graph`, `infer_oce_temporal_spec`); see Associations integration
Learn a DAG from data (legacy Julia PC)	CausalInference.jl	`pcalg`; we use CausalInference for d-separation and backdoor primitives
Solve ODEs / SDEs / UDEs	SciML, UniversalDiffEq.jl	We supply structure and `do(·)` semantics, not integrators — see SciML recipes
Continuous MTP / LMTP estimation	CausalTargeted.jl	Consumes `IdentificationResult`; see its comparison
Doubly robust point-treatment TMLE	TMLE.jl	We identify the adjustment set; TMLE estimates (integration)
Scalable Bayesian inference	RxInfer.jl	Variational backdoor head is a weakdep extension
Tabular causal simulation / estimation	CausalTables.jl	Table-centric; we are trajectory-centric over occasions `t = 1:T`
DAG figures	DAGMakie.jl	Optional plots; comparison

Versus R and Python (graphs / ID / dynamics)

Capability	CausalDynamics	R	Python
d-separation / backdoor / frontdoor / IV	Yes	Yes (dagitty, ggdag)	Yes (DoWhy, causal-learn)
Typed `identify` → certificate	Unique (`IdentificationResult`)	Partial (dagitty objects + glue)	Partial (DoWhy identify result)
Path mediators / minimal mediator cuts	Yes	Partial	Partial
Static SCM + `do` + shared-`U` CF	Yes (`GraphSCM`)	Partial (`simcausal`, limited)	Partial (DoWhy SCM)
Discrete-time CDM trajectories	Unique	— / Partial	— / Partial
Temporal unroll + time-indexed ID	Unique	Partial (manual)	Partial (custom)
Continuous CDM + SciML `do`	Unique	—	Partial (custom SciPy)
Structure discovery (full algorithms)	Partial (bridges)	Yes (pcalg, bnlearn)	Yes (causal-learn)
Broad four-step estimate API	— (estimation is CausalTargeted / TMLE.jl)	Partial	Yes (DoWhy)

What is distinctive here

Trajectories, not single outcomes — simulate returns a CDMTrajectory with the realised exogenous draws retained
Shared-U counterfactual paths — counterfactual replays the same exogenous draws under a different intervention
Time-indexed identification — unroll_temporal_dag turns a lag specification into a static DAG so ordinary backdoor logic covers lagged confounding
Soft interventions — Policy supports treatment rules that react to state
Lean core — hard dependencies are Graphs and CausalInference only

What is deliberately absent

Full causal discovery implementations, differential-equation solvers, and full symbolic do-calculus. Discovery runs in Associations.jl; identification and trajectories stay here. See Scope.

The CDCS book is the long-form narrative companion for these APIs.