---

name: analysis-and-decision description: "Score, validate, interpret, and optimize causalml treatment-effect outputs for uplift analysis and treatment decisions." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

causalml analysis and decision router

Use this sub-skill when the task is about evaluating or acting on treatment-effect estimates rather than fitting the primary estimator.

Route here for

• Uplift/CATE ranking metrics: AUUC, Qini, cumulative lift/gain/Qini curves, TOC/RATE, bootstrap intervals, and plotting. See metrics-and-validation.
• CATE validation losses and validation summaries: doubly robust pseudo-outcome loss, plug-in T loss, R-loss, TMLE-based segment gain/Qini, and propensity balance diagnostics. See metrics-and-validation.
• Sensitivity and robustness checks: placebo treatment, random cause, subset data, random replace, selection bias, marginal sensitivity model bounds, and confounding-function summaries. See sensitivity-and-selection.
• Feature selection and interpretation: FilterSelect, meta-learner importances, SHAP summaries/dependence plots, and tree/importances interpretation handoff. See sensitivity-and-selection.
• Treatment decision optimization: PolicyLearner, CounterfactualUnitSelector, CounterfactualValueEstimator, cost/value utilities, uplift-best recommendations, and probability-of-causation bounds. See optimization.

Before using an API

1. Identify the data orientation: outcome column, treatment column, propensity column if any, true treatment-effect column if available, and one or more model prediction columns.
2. Confirm treatment coding. Most metric and policy APIs assume binary 0/1 treatment with 1 as treated and 0 as control; string-labeled or multi-arm data usually needs explicit one-vs-control conversion or APIs with control_name/treatment_names contracts.
3. Decide whether the task is about effect magnitude accuracy or targeting quality:
   • Use DR/T/R losses for CATE magnitude accuracy.
   • Use AUUC/Qini/TOC/RATE for ranking or treatment prioritization.
   • Use value/policy APIs when decisions have costs, benefits, or allowable treatment assignments.
4. If the request requires fitting causal estimators first, route to the estimation, tree, or deep-model sub-skill, then return here for scoring, interpretation, or optimization.

Related sub-skills

• Classical meta-learners, IV/DRIV, TMLE API limits, and serialization: ../causal-estimation/SKILL.md.
• Uplift tree fitting, pruning/fill/save/load, visualization primitives, and tree-specific import issues: ../tree-models/SKILL.md.
• Neural DragonNet/CEVAE backend setup and fit/predict workflows: ../deep-models/SKILL.md.
• Synthetic data, matching, propensity models, and data-contract preparation: ../data-preparation/SKILL.md.

Troubleshooting first stop

For wrong treatment labels, reversed prediction orientation, missing model columns, feature-selection method constraints, optimization shape/order issues, stale API names, or optional plotting/SHAP dependency errors, start with troubleshooting.