Course 2 · Tier 5
Causal Reasoning and World Modeling
AI finds correlations. Humans build causal models. This course operates at Tier 5 — the deepest level in the series — because causal reasoning is the capacity that most decisively separates human cognition from statistical pattern matching.
DAG Construction
A Directed Acyclic Graph (DAG) is the formal representation of a causal model. Students learn to construct DAGs from domain knowledge — not from data. This is the critical distinction: data can tell you what happened, but only a human can propose why.
The course walks through the full process of building a DAG: identifying variables, specifying directed edges that represent causal claims, and — most importantly — defending the exclusion of edges. Every missing arrow is an assertion that two variables are not directly causally related, and every such assertion must be justified.
- Variable selection from domain expertise, not data mining
- Edge specification as causal claims requiring justification
- Missing edges as testable assumptions
- Iterative refinement through peer critique and empirical challenge
The Backdoor Criterion and Identification
Once a DAG is constructed, the identification layer determines whether a causal effect can be estimated from observational data. The backdoor criterion provides a systematic method: if you can block all backdoor paths between treatment and outcome by conditioning on a set of variables, the causal effect is identified.
This is where the course reaches Tier 5. Students must reason about confounding, collider bias, and mediation — concepts that require genuine counterfactual thinking. No current AI system can construct a defensible causal model from scratch, because doing so requires knowledge of the world that is not contained in the data.
Confounding
A common cause of both treatment and outcome creates a spurious association. Students learn to identify confounders from the DAG structure and condition on them appropriately — or recognize when conditioning introduces new bias.
Collider Bias
Conditioning on a common effect of two variables opens a non-causal path between them. This is counterintuitive and routinely missed by automated analyses. Recognizing collider structures requires understanding the causal story, not just the statistical associations.