causal-inference-root-cause_skill

Overview

This skill helps you identify true root causes by distinguishing causation from correlation and separating symptoms from underlying drivers. It provides a step-by-step workflow for generating hypotheses, building causal models, testing causality, and documenting validated conclusions. Use it to design experiments, control confounders, and produce actionable interventions with stated confidence levels.

How this skill works

The skill inspects the effect you want to explain, guides structured hypothesis generation, and maps causal chains and confounders into an explicit model. It recommends appropriate tests (from RCTs to natural experiments and observational checks), evaluates evidence against causal criteria (temporality, mechanism, dose-response, etc.), and produces a documented root-cause report with recommendations and uncertainty. It also provides patterns and tests tailored to incidents, metrics, policy evaluation, and debugging.

When to use it

• Investigating system outages, performance regressions, or production incidents
• Explaining sudden metric changes (conversion drops, churn spikes, engagement shifts)
• Researching treatment effects, policy impacts, or health outcomes
• Debugging complex failures where multiple plausible causes exist
• Designing experiments or analyses to validate causal claims

Best practices

• Define the effect precisely: quantify magnitude, timeline, baseline, and stakeholders
• Generate multiple competing hypotheses and avoid single-cause fixation
• Map confounders and enforce temporal ordering before claiming causation
• Prefer randomized or quasi-experimental tests where possible; use observational checks otherwise
• Document mechanism, evidence strength, alternative explanations, and a confidence score

Example use cases

• Postmortem for an outage: timeline + dependency graph → logs, rollback, prevention plan
• Product metric drop: cohort analysis, A/B test re-check, segmentation to isolate cause
• Policy evaluation: build DAG, apply difference-in-differences or regression discontinuity
• Health outcome research: control confounders, seek dose-response and biological plausibility
• Bug triage: reproduce, isolate code paths, git bisect to distinguish symptom vs root cause