Model Checks
Auto-Build picks a model for you — the factors that best account for an outcome. Before you trust it, two independent checks interrogate that model from different angles. One tests whether an estimate holds up; the other tests whether the factors it selected actually move the outcome. Use either on its own, or both together.
Triangulation Check
Can you trust the estimate?
Take a scenario — change one or more factors — and the check re-derives the predicted change a second, independent way: a propensity model built live from the same predictors that leans on the outcome much less directly. Two methods that rest on different assumptions.
- When they agree, the estimate is unlikely to be an artifact of one modeling choice.
- When they diverge, read it cautiously — small sample, weak overlap, residual imbalance, or model dependence.
Bayesian Network Key Driver Analysis
BETAWhich factors can you act on?
Take the same factors Auto-Build selected and ask a different question of them: which ones actually drive the outcome when you act on them, and which only ride along with it — tracking the outcome without driving it. The network estimates each factor's effect as a do()-style intervention, and a what-if simulator lets you set factors and read the projected change in the outcome.
It does two things the ranked predictor lists can't. It confirms which of the key predictors Auto-Build selected are real levers — not just correlated with the outcome — and it adds direction: whether a factor feeds the outcome directly or only indirectly, through other factors. A total complement to the main analyses.
- Drivers move the outcome when you act on them — the levers you can act on.
- Riders track the outcome but acting on them does nothing — they ride along, they don't drive it.
How they work together
The two checks look at different halves of the same model, using two different methods. The Triangulation Check (propensity-score matching) validates an estimate — does a specific predicted change survive a second, assumption-independent method? The Key Driver Analysis (a Bayesian network) validates the structure — are the factors Auto-Build leaned on actually drivers, or just riders? One tests the number; the other tests the interpretation.
An Auto-Build model that clears both is one whose headline estimate holds up under an independent method and whose top factors are levers you can act on rather than riders. The key driver analysis also adds something the main predictor analyses don't — direction: whether each factor reaches the outcome directly or only indirectly, through other factors. Neither check rebuilds the model — they interrogate the one Auto-Build already found.