# Workshop Guide: Causal Inference

## Audience

Analysts, data scientists, researchers, and product teams making causal claims from observational data.

## 60-Minute Agenda

1. 0-10 min: Choose one observational causal question.
2. 10-20 min: Define treatment, control, population, outcome, and time horizon.
3. 20-35 min: Draw the DAG in words.
4. 35-45 min: Choose identification strategy and assumptions.
5. 45-60 min: Peer critique: what would break the claim?

## 90-Minute Agenda

1. 0-10 min: Review the worked identification memo.
2. 10-25 min: Teams draft estimand and decision context.
3. 25-45 min: Teams build DAG, separating confounders, mediators, colliders, and selection mechanisms.
4. 45-60 min: Choose strategy and diagnostics.
5. 60-75 min: Sensitivity and interpretation limits.
6. 75-90 min: Group critique and revision.

## Team Exercise

Each team produces a one-page identification memo with:

- estimand
- DAG in words
- identification strategy
- assumptions
- diagnostics
- sensitivity plan
- interpretation limits

## Discussion Prompts

- What is the strongest unmeasured confounder?
- Which variable is tempting to control for but post-treatment?
- Where is overlap weak?
- Would a randomized prompt or rollout be more credible?

## Facilitator Notes

Do not let teams start with an estimator. The estimator comes after the causal model and identification argument.

Common failure modes:

- treating prediction features as valid controls
- adjusting for mediators or colliders
- ignoring positivity
- reporting correlation as causal effect

## Review Standard

Use `final-assessment.md` as the rubric. A strong memo makes the causal claim criticizable and clearly separates evidence, assumptions, and decision implications.