Causal reasoning · 2/6

Look for common causes

Recognize pre-intervention confounding and compare overall with within-group results.

About 15–20 minutes. The following cases are fictional.

Record your first judgment

Eighty percent of AI users are experienced, versus 20% of nonusers. Experienced staff finish 14 tasks either way; beginners finish six. What happens overall?

My initial judgment, confidence and reasons

Understand the judgment process

Prior skill can influence both adoption and output. It is a candidate common cause: skill → adoption and skill → output. Background knowledge must support these assumed arrows.

Users average 0.8×14+0.2×6=12.4; nonusers average 0.2×14+0.8×6=7.6. The overall gap is 4.8 despite zero within-skill gaps. The interactive model changes composition while holding within-skill performance fixed.

Stratification addresses measured, appropriately defined common causes, not all confounding. Adjusting for every available variable can be wrong: a post-intervention mediator may carry part of the effect, and conditioning on a shared outcome can create bias.

Change composition, change the gap?

Experienced staff complete 14 tasks a day and beginners six in either group. With JavaScript, adjust the mix: 80% versus 20% gives 12.4 versus 7.6; 50% versus 50% gives ten each.

Checks you can perform

Map intervention, outcome and candidate common causes.
Check which factors precede intervention.
Inspect counts and outcomes within strata.
Compare with common weights and note unmeasured factors.

Guided practice

If both groups are 50% experienced, what are their means? Does this establish that real AI tools have no effect?

My analysis and checking plan

Explore the explanation

Both average ten. Composition explains this deliberately constructed model; real tools may differ. Task difficulty and motivation may also remain unmeasured.

Apply it in a new context

Club members score higher but also scored higher before joining. Name a common cause and a better comparison.

My independent response

Compare after attempting

Prior study habits may affect both membership and scores. Compare similar baseline students and contemporaneous changes; matching still cannot remove unmeasured motivation.

Preserve a revision record

What changed your judgment? What still supports it? When will you review?

This deterministic synthetic model has no sampling uncertainty. Adjustment does not automatically identify a causal effect.

Check whether you supplied inspectable evidence, a feasible next step and revision conditions. Revealing an answer is not mastery.

Source and scope

Hernán & Robins — Causal Inference: What If (chapters 1–3, 7–8)

The source provides conceptual or methodological background. Cases, steps and exercises are authored here; they are not the original experiment or evidence of this course’s effectiveness.

I have completed this section; uncheck to undo

Completion is a personal record, not proof of mastery or certification.