From correlation to cause: what changes when we intervene?
Pearl’s ladder separates three questions: what we observe, what happens when we act, and what would have happened under a different action. CausalGraphOS focuses on intervention questions, using evidence and explicit assumptions to assess how outcomes may change.
Rung 2 is our focus: understanding the effects of action.
Rung 1—learning patterns and predicting outcomes from observed data—is well served by mature machine learning. Those predictions are useful, but they do not by themselves show what might change if a team intervenes.
What might change if we act?
Rung 2 asks how an outcome may change when a team changes a policy or takes a specific action. For example: if a lender changes an advance cap, would repayment outcomes change, and for whom?
This is CausalGraphOS’s focus. It helps data science teams define the intervention and outcome, bring evidence and domain knowledge together, and estimate effects when the data and assumptions allow. Analysts can compare adjusted estimates with observed patterns, test how conclusions shift under different assumptions, and make evidence gaps and uncertainty visible. The result supports review; it does not answer every policy question automatically.
When can observed data estimate an intervention effect?
If a set of measured variables Z satisfies the back-door criterion for X and Y, an adjustment formula can identify the effect of setting X to x:
P(Y | do(X = x)) = Σz P(Y | X = x, Z = z) P(Z = z)This is not a universal correction. It depends on a valid causal graph, sufficient adjustment variables, and data that meet the assumptions. Do-calculus provides rules for deriving such results from a causal graph.
Sources: Judea Pearl, Causality: Models, Reasoning, and Inference, 2nd ed. (2009); Judea Pearl and Dana Mackenzie, The Book of Why (2018).
What would have happened otherwise?
Evaluating an intervention means comparing the observed outcome with a credible estimate of what would have occurred without it. Timing, selection, confounding, and missing data can all affect that comparison.
CausalGraphOS helps analysts keep evidence and assumptions in view, test how conclusions respond to them, and make a defensible interpretation. It does not turn every correlation into a causal claim.
Bring us the intervention you need to evaluate.
We’re shaping CausalGraphOS with design partners working on consequential decisions.