causal-ml-estimator-selector

name

causal-ml-estimator-selector

description

Selects, audits, and explains causal machine learning workflows for heterogeneous treatment effects and graph-based causal ML. Use for CATE, ITE, uplift modeling, S/T/X/R/DR learners, causal forests, matching, propensity scores, IV/DRIV, CEVAE/DragonNet, causal GNNs, graph neural networks with causal claims, therapeutic perturbation prediction, optimal intervention design, causal disentangled graphs, LLM-enhanced GNN mechanism identification, fault-diagnosis causal subgraphs, treatment targeting, policy personalization, Python CausalML/EconML-style projects, validation of causal-ML claims, and boundary checks between treatment-effect estimation and causal-invariant/stable prediction.

Causal ML Estimator Selector

Use this skill when a project uses machine learning to estimate causal effects, especially heterogeneous treatment effects (CATE), individual treatment effects (ITE), uplift, targeting, or treatment personalization.

Core Principle

Causal ML is not prediction with causal language. Separate:

• causal question;
• estimand;
• identification assumptions;
• treatment assignment process;
• feature set and timing;
• nuisance-model estimation;
• treatment-effect model;
• validation;
• policy or targeting use.

Good predictive performance does not prove causal validity.

Intake

Identify:

• treatment or intervention;
• outcome;
• unit of observation;
• graph object, if present: nodes, edges, edge direction, node states, and graph source;
• data source: experiment, quasi-experiment, or observational data;
• treatment timing;
• available pre-treatment covariates;
• target estimand: ATE, ATT, CATE, ITE, LATE, or uplift;
• assignment mechanism;
• overlap risks;
• hidden-confounding risks;
• intended tool: causalml, econml, dowhy, custom Python, R, Stata, or mixed stack.

Load:

• references/causal-ml-toolkit.md for estimator families, guardrails, and CausalML-specific notes.
• references/causal-gnn.md for graph neural networks with causal, intervention, perturbation, mechanism-identification, or causal-disentanglement claims.
• templates/causal-ml-design-card.md for project setup.
• templates/causal-gnn-audit.md and templates/graph-causal-validation-schema.csv for graph-based causal ML studies.
• templates/causal-ml-validation-plan.csv for validation planning.

If the project is about stable prediction under distribution shift rather than treatment-effect estimation, route the main workflow to prediction-modeling-forge and keep this skill only for causal-language boundary checks.

Workflow

1. Name the estimand.
2. Decide whether the design is experimental, quasi-experimental, or observational.
3. Check whether treatment, outcome, and covariates are temporally ordered.
4. Check overlap and treatment prevalence.
5. Select estimator family:
   • meta-learners for flexible CATE;
   • doubly robust learners when nuisance models and overlap are central;
   • uplift trees/forests for targeting and segmentation;
   • IV/DRIV when treatment is endogenous and a credible instrument exists;
   • causal GNNs when graph structure, intervention targets, mechanism discovery, or graph-level domain shift are central;
   • neural methods only when data scale and validation justify them.
6. Plan cross-fitting or sample splitting when appropriate.
7. Compare learners and report stability.
8. Validate with multiple estimates, synthetic/placebo checks, AUUC/uplift curves, and sensitivity analysis.
9. Translate results as conditional or exploratory unless the identification design is strong.
10. For graph models, audit whether the graph is a true causal graph, a proxy causal graph, a correlation graph, a PPI/GRN/knowledge graph, or a learned graph.

Output Modes

Estimator Recommendation

Estimand:
Data design:
Main identification assumption:
Recommended estimator family:
Why this estimator:
Minimum diagnostics:
Validation plan:
Interpretation limit:

Causal ML Audit

Verdict:
Causal question:
Treatment timing:
Feature leakage risks:
Overlap:
Learner choice:
Validation:
Heterogeneity claim:
Policy use:
Must-fix issues:

Causal GNN Audit

Graph source:
Node state:
Edge meaning:
Causal claim:
Intervention or perturbation:
Validation split:
Counterfactual or held-out mechanism test:
Proxy-graph risk:
Interpretation limit:

Implementation Scaffold

Give package-agnostic pseudocode unless the user asks for a specific library.

1. Build design card.
2. Freeze feature set.
3. Split/cross-fit.
4. Fit nuisance models if needed.
5. Fit treatment-effect estimator.
6. Validate effect stability.
7. Export CATE/ITE estimates with uncertainty and cohort summaries.
8. Write assumptions and limits.

Guardrails

• Do not claim causal discovery unless graph/discovery assumptions are explicitly part of the design.
• Do not include post-treatment variables as features for treatment-effect estimation.
• Do not treat CATE ranking as a welfare policy without cost, constraint, and fairness checks.
• Do not write subgroup effects as confirmatory when they were searched after seeing results.
• Do not use causal ML to rescue an unidentified observational design.
• Do not hide package versions, random seeds, train/test logic, or feature construction.
• Do not treat causal-invariant or stable predictive features as treatment effects unless the project defines a treatment, estimand, and identification strategy.
• Do not call a GNN causal because it uses a graph; the edge semantics, intervention target, and validation design must support the claim.
• Do not treat PPI, GRN, sensor adjacency, or LLM-enhanced node features as a verified causal graph without assumptions and sensitivity checks.