Bayesian Network Key Driver Analysis

BETA

A network over the modeled factors. Lines show which factors depend on which; each arrow points to the factor (or the outcome) that depends on the one at its tail — so up vs. down is just the layout, what matters is which way the head points. The % on each line (and its thickness) is that link's confidence — how consistently it holds up across resamples. Factors are ranked by how much the outcome changes when a factor is set to a different value — the two values shown on each row — once the rest of the network is accounted for. The effects are valid under the structure shown; treat thin (low-confidence) links as unsettled. These are the factors Auto-Build selected for this outcome — change the set or refit there.

Reading it against the logit: the network's direct effect lines up with the model's adjusted effect (both hold the other factors fixed), and the direct-vs-indirect split is the same mediation as the gap between a factor's standalone and incremental importance — the network just names the path. Factors both methods flag are the high-confidence drivers; where they disagree it's usually a diagnostic (mediated, suppressed, or — on a thin slice — a real link the network couldn't place).

Outcome

Subgroup

Level

Choose an outcome and any filters above, then click Run analysis to build the network.

Cross-check summary

Does each variable drive the outcome, or merely ride along with it?

—

Does the factor drive the outcome, or just ride along with it? Almost everything tracks the outcome — rises and falls with it — so tracking alone makes nearly every factor look important. What separates a driver from a rider is whether acting on the factor drives the outcome: the Bayesian network's do-effect. A factor that tracks the outcome and drives it is a true driver — the real thing to act on. One that tracks it but doesn't drive it is a rider: it rides along, but acting on it does nothing.

What each verdict means — does it drive, or ride along?

Driver Drives it. It tracks the outcome, and acting on it changes the outcome — the real thing to act on.

Rider Rides along. It tracks the outcome, but acting on it does nothing — it rides along rather than driving.

Minor Neither. Barely tracks the outcome, and acting on it does little.

Chip colour shows the factor's network path: direct indirect no-path

What-if Simulator

Project the outcome if a factor were set for the whole sample.

1Set one or more factors to a level — the rest stay as observed.
2Click Run simulation.
3Read the projected outcome and its change vs the observed base.

Only the drivers from the Bayesian analysis are included here — the factors that actually move the outcome. Riders (they track the outcome but don't move it) and factors with no path to it are left out, since changing them wouldn't shift the projection. Each change is a do() intervention, so the rest of the network responds.

Projected outcome

—

Which Factors Actually Move the Outcome?