Graph Refutation

Graph refutation tests whether the conditional independence constraints implied by the causal graph are consistent with the observed data. It leverages the Local Markov Conditions (LMCs) — each node in the graph should be conditionally independent of its non-descendants given its parents.

If the dataset violates any of these conditional independence constraints, then the assumed causal graph may be incorrect and downstream causal estimates may be unreliable.

The graph refuter uses partial correlation for continuous data and conditional mutual information for discrete data to test each implied conditional independence statement.

from dowhy import CausalModel

model = CausalModel(
    data=data,
    treatment="treatment",
    outcome="outcome",
    graph=graph
)
res = model.refute_graph(k=1, independence_test={"test_for_continuous": "partial_correlation",
                                                  "test_for_discrete": "conditional_mutual_information"})
print(res)

For a detailed walkthrough, see the Graph Conditional Independence Refuter notebook.