Prediction vs. Causation in DAGs

Prediction and causation are fundamentally different in causal diagrams (DAGs). Prediction asks "given X, what can we expect to see?" and works purely with observed patterns—like how ice cream sales predict sunscreen purchases because both increase in summer. Causation asks "if I change X, will Y actually change?" and requires understanding the real mechanism behind the pattern, not just the correlation. DAGs help separate these by showing which variables truly influence others versus which merely share patterns. Try this: Before assuming one thing causes another, ask what else might explain the pattern.

How Prediction Differs from Causation

Prediction models work by finding patterns in data without needing to know why those patterns exist. A machine learning algorithm can predict customer churn accurately by spotting that people who call support twice often cancel—but this doesn't mean support calls cause cancellations. The real cause might be a defective product that generates both support calls and cancellations. DAGs make this distinction explicit by drawing arrows only for true causal relationships, not mere correlations.

Why DAGs Separate the Two

In a DAG, arrows represent direct causal influence. If ice cream sales predict sunglasses purchases, a DAG would show both as children of a common parent: temperature or season. This structure explains why they're correlated (summer increases both) but clarifies that neither causes the other. If you intervened to increase ice cream sales, sunglasses purchases wouldn't budge—yet a naive prediction model might suggest they would. The DAG's structure prevents this mistake by making the causal assumptions visible and testable.

Real-World Examples

In medicine, blood pressure predicts heart disease risk, but lowering blood pressure doesn't always prevent heart attacks—the causal pathway may involve inflammation, stress hormones, or arterial plaque. A DAG helps identify whether blood pressure is a cause, a consequence, or just a marker. In business, employee satisfaction predicts productivity, but improving satisfaction through free snacks won't boost productivity if the real cause is good management. DAGs force you to specify what you'd actually change and what would truly respond, preventing costly interventions that target symptoms instead of causes.

Practical Implications

When building predictive models, you can use any correlated variable—proxies like "season" or "temperature" work fine. But when making decisions or designing interventions, you must identify true causal relationships. DAGs help by forcing you to articulate your causal assumptions: which variables would stay the same if you intervened, which would change, and what hidden common causes might exist. This distinction saves resources by preventing actions that target correlations rather than causes.

Causal DAGs (Directed Acyclic Graphs) formalize the distinction between associational and causal inference through the do-calculus and structural causal models. Pearl's framework shows that prediction relies solely on observational distributions, while causal claims require interventional distributions denoted by the do-operator P(Y | do(X)). [1, 5] High-dimensional prediction models can achieve excellent accuracy while completely misrepresenting causal structure, as "spurious" correlations improve predictions but invalidate causal conclusions. [2]

DAG-based causal conclusions depend heavily on correctly specifying the graph structure—miss a single confounder or backdoor path, and causal estimates become biased. Prediction models don't require this causal specification and may be more robust when assumptions are uncertain. Additionally, some modern machine learning systems (like deep neural nets) achieve superhuman prediction without any interpretable causal structure, suggesting that for pure forecasting, causal understanding may be unnecessary overhead. Finally, DAGs typically assume stable, well-defined causal relationships, which breaks down in complex systems with feedback loops or emergent properties.