Case Study

Weight Matrix — IatrogenicMedications.bayes

Structural coefficients from the fitted Gaussian CLG. Rows = cause, columns = effect. Blue = positive weight, red = negative. U nodes excluded. Weights are fitted parameters, not elicited estimates.

Why the weight matrix matters

The weight matrix is the skeleton of the structural causal model — the fitted linear coefficients that quantify how strongly each direct cause acts on each effect. Unlike correlation coefficients, these weights have a directional, mechanistic interpretation: holding all other parents fixed, a weight of +1.5 means one unit of cause produces 1.5 units of effect downstream, not merely that the two tend to move together.

The matrix makes model transparency concrete. Every quantitative claim — path effects, sensitivity scores, counterfactual predictions — is a composition of these weights across chains of edges. Inspecting the matrix is the first check that domain knowledge was correctly encoded and that the fitted model behaves as intended.

0.000
positive
negative
no link
darker = stronger magnitude
Weights by magnitude