Causal

Skill: /causal — OpenCausalInf Causal Inference Toolkit

Purpose

Multi-mode skill for causal inference when experiments aren't possible. Helps users estimate treatment effects from observational data with explicit assumption checking, sensitivity analysis, and mandatory caveats. Uses coded helpers from helpers/stats/experiment_stats/causal/.

When to Use

Invoke as /causal [mode] or trigger on causal inference intents:

• "Did this feature actually cause the improvement?"
• "We can't run an experiment, but..."
• "Was this change responsible for the metric movement?"
• "Can we measure the impact retroactively?"

Modes

/causal select

Purpose: Walk the method selection decision tree and recommend a causal method. Agent: agents/causal/causal-method-selector.md Flow:

1. Ask 4-6 diagnostic questions:
   • Can you randomize? → Route to /experiment design
   • Do you have a comparison group?
   • Do you have pre-treatment data?
   • Are there observable confounders you can measure?
   • How many time periods do you have?
2. Recommend: Pre-Post, DiD, PSM, Regression Adjustment, or "not feasible"
3. Output: recommended method + confidence level + rationale Checkpoint: Method confirmation (Type C — user must confirm before analysis)

/causal analyze

Purpose: Run the selected causal method on data. Agent: agents/causal/causal-analyzer.md Flow:

1. Read selected method from previous step or user input
2. Dispatch to appropriate helper:

   from helpers.stats.experiment_stats.causal import (
       pre_post_analysis, did_basic, propensity_match,
       regression_adjust,
   )
   # Method routing:
   # "pre_post" → pre_post_analysis(pre, post, covariates)
   # "did"      → did_basic(df, outcome, treat, post)
   # "psm"      → propensity_match(df, treat, covariates, outcome)
   # "regression" → regression_adjust(df, outcome, treatment, covariates)

3. Generate charts (treatment effect, balance plots for PSM, event study for DiD)
4. Output: working/causal_analysis_results.json

/causal check

Purpose: Run assumption checks for the selected method. Agent: agents/causal/causal-assumption-checker.md Flow:

1. Identify which assumptions apply to the selected method:
   • DiD: Parallel trends, no anticipation, stable composition
   • PSM: Common support, balance (SMD < 0.1), positivity
   • Pre-Post: No concurrent events, trend stability
   • Regression: All confounders included, correct specification
2. Run quantitative checks:

   from helpers.stats.experiment_stats.causal import (
       check_parallel_trends, check_common_support,
       balance_table,
   )

3. Output: per-assumption PASS / WARNING / FAIL verdicts Checkpoint: Any FAIL (Type C) → present options: adjust method, add caveats, or abort

/causal sensitivity

Purpose: Test how robust the estimate is to unmeasured confounding. Agent: agents/causal/causal-sensitivity.md Flow:

1. Run sensitivity analysis based on method:

   from helpers.stats.experiment_stats.causal import rosenbaum_bounds, e_value
   # PSM: rosenbaum_bounds(treated_outcomes, control_outcomes)
   # All: e_value(risk_ratio, ci_lower)

2. Translate to plain language: "An unmeasured confounder would need to be X times stronger than anything we measured to explain away this result."
3. Output: sensitivity report