Model Checks
Both checks run the published model for the selected outcome — the same reference specification the simulator pages use — and interrogate it from different angles. One asks whether a scenario estimate depends heavily on that fitted logit; the other maps the conditional dependency structure among its predictors. Use either on its own, or both together. Narrow to a subgroup and the network instead discovers that subgroup’s own model, which is the more useful question there; to see the published specification refit on a subgroup, build it on the Build page and run the check from there.
One fitted explanation, examined from two additional angles. The published logit remains the reference for the estimates shown elsewhere, and neither check replaces it or becomes a second reference model: the Triangulation Check re-estimates a stated scenario through other structures — a propensity route, and the fitted network where it has a comparable estimate — while the network view itself examines how the same predictors are conditionally organized under a different model structure.
Triangulation Check
How dependent is the estimate on model form?
Take a scenario — set one or more factors — and the check estimates the same scenario structurally different ways: through propensity-score subclassification built from the same measured predictors rather than the active fitted outcome model, and — where the fitted Bayesian network gives a scenario factor a direct path to the outcome — off that network as well, under the same reporting rule the simulator uses. This is Mosteller’s triangulation: one quantity estimated from several differing structures, where agreement is itself evidence. The routes differ in form, but they share the sample, measured controls, and important observational limitations.
- When they agree, one concern narrows: the result appears less dependent on the active fitted logit model. A third route that dissents lowers the verdict; one that agrees does not raise it.
- When they diverge, read it cautiously — small sample, weak overlap, residual imbalance, or model dependence.
Bayesian Network Analysis
BETAHow are the selected factors conditionally related?
Take the same predictors as the published model and examine how they are conditionally connected under a constrained Bayesian network. The network provides a different structural representation of the measured variables, and its what-if view shows model-implied changes under stated settings. Those outputs are exploratory properties of the fitted network, not identified intervention effects. The what-if view reports a change two ways: holding the other measured factors at each record’s own values, which matches how the published model’s scenarios are reported, and allowing them to respond within the fitted network, which is larger wherever the graph routes a factor to the outcome through another.
It adds information that ranked predictor lists do not show. It displays which of those variables are connected to the outcome in the fitted network and whether those connections are represented as direct or indirect paths through other measured variables. That direction belongs to the fitted network structure; it does not establish temporal or causal direction.
- Connected factors show stronger model-implied relationships with the outcome under the fitted network.
- Factors without a modeled path may still add information in the logit; the two structures can place the same measured factor differently.
How they work together
One tests the number; the other tests the interpretation.
Does a specific predicted change survive a second route with weaker assumptions about the outcome? Agreement narrows the concern that the number is an artefact of the model’s form.
Maps the conditional dependency structure among the published model's predictors, and whether each is represented as reaching the outcome directly or indirectly. That direction belongs to the fitted network; it does not establish causal direction.
Predictive importance is not network placement. A factor can add substantial information in the logit yet have no modeled path to the outcome in the fitted network. A large logit contrast beside little or no network-implied change therefore reflects the two representations answering different questions rather than a contradiction.
A model that looks similar under both has survived two additional forms of scrutiny, but neither check is a pass/fail test and agreement is not proof. Neither resolves omitted measurement, nonresponse, or other vulnerabilities shared by the underlying data.