causalml

Use this repo skill when the task names CausalML or needs Python operating guidance for causal effect estimation, uplift modeling, propensity/matching, causal trees, optional neural causal models, validation metrics, sensitivity analysis, feature selection, or treatment decision optimization.

Install and environment check

For normal package use, install the package and only the optional extras needed by the requested workflow:

pip install causalml
pip install "causalml[tf]"      # TensorFlow DragonNet only when needed
pip install "causalml[torch]"   # Torch/Pyro CEVAE only when needed
pip install "causalml[jax]"     # JAX DragonNet/CEVAE only when needed

When working from a source checkout, use the equivalent editable install such as pip install -e . or pip install -e ".[tf,torch,jax]" from that checkout.

After installation, run the bundled diagnostic from this skill tree if import status is uncertain:

python scripts/check_env.py
python scripts/check_env.py --backend tf --backend jax
python scripts/check_env.py --backend all --json

The diagnostic checks core imports and only checks optional backends that are explicitly requested.

Route by workflow

Task	Open
Synthetic data, feature matrix preparation, propensity scores, matching, balance tables, CSV matching	sub-skills/data-preparation/
S/T/X/R/DR meta-learners, TMLE, IV/DRIV, estimator API contracts, Polars/DataFrame support, persistence	sub-skills/causal-estimation/
Causal trees, causal forests, uplift trees, uplift forests, tree plots, tree save/load, compiled tree import issues	sub-skills/tree-models/
TensorFlow DragonNet, Torch/Pyro CEVAE, JAX DragonNet, JAX CEVAE, deep-backend setup and smoke checks	sub-skills/deep-models/
AUUC/Qini/RATE/validation metrics, sensitivity analysis, feature selection, SHAP/importances, policy/value optimization, PNS bounds	sub-skills/analysis-and-decision/

Start at the most specific sub-skill. Return here only when routing across workflows or diagnosing cross-cutting install/import failures.

Operating defaults

1. Prefer keyword arguments for estimator calls: fit(X=X, treatment=treatment, y=y, ...). Positional argument order is migration-sensitive in current CausalML.
2. Keep control_name explicit whenever treatment labels are strings or multi-arm.
3. Treat propensity scores as probabilities strictly inside (0, 1); clip extreme scores before X/R/DR/TMLE, validation, and matching workflows when appropriate.
4. For multi-treatment CATE arrays, use the learner's treatment-group metadata (t_groups or _classes) rather than guessing column order.
5. For metrics and policy workflows, separate model prediction columns from feature columns. Many metrics treat every non-reserved column as a model score.
6. Do not rely on old documentation references to causalml.inference.nn; current neural models route through causalml.inference.tf, causalml.inference.torch, or causalml.inference.jax.
7. CPU execution is sufficient for small correctness and wiring checks. Do not claim GPU/accelerator behavior unless the selected backend stack has been installed and verified separately.

Cross-cutting references

• references/troubleshooting.md: install/import, optional dependency, Graphviz/OpenMP, argument-order, backend, and stale-API issues.
• references/repo-provenance.md: source commit, package version, evidence paths, selected coverage, and refresh triggers.
• references/repo-routing-metadata.json: structured router scenario metadata for managed repo-skill import.