[Repo Assist] fix: return p_value=1 when refuter simulations have zero variance (closes #807)

dowhy/causal_refuter.py

@@ -157,6 +157,17 @@ def perform_normal_distribution_test(estimate, simulations: List):
     mean_refute_values = np.mean(simulations)
     # Get the standard deviation for the simulations
     std_dev_refute_values = np.std(simulations)
+
+    # Degenerate case: all simulations returned the same value (zero variance).
+    # Division by zero would produce NaN. Return p_value=1 to indicate that the
+    # original estimate is indistinguishable from the simulation distribution.
+    if std_dev_refute_values == 0:
+        logger.warning(
+            "All simulation values are identical (std = 0). "
+            "The normal-distribution p-value is undefined; returning p_value=1."
+        )
+        return 1.0
+
     # Get the Z Score [(val - mean)/ std_dev ]
     z_score = (estimate.value - mean_refute_values) / std_dev_refute_values
 

tests/test_causal_refuter.py

@@ -4,7 +4,7 @@
 
 from dowhy.causal_estimator import CausalEstimate
 from dowhy.causal_identifier.identified_estimand import IdentifiedEstimand
-from dowhy.causal_refuter import CausalRefuter
+from dowhy.causal_refuter import CausalRefuter, perform_normal_distribution_test
 
 
 class MockRefuter(CausalRefuter):
@@ -24,3 +24,14 @@ def test_causal_refuter_bootstrap_test():
     simulations = np.random.normal(0, 1, 5000)
     pvalue = refuter.perform_bootstrap_test(estimator, simulations)
     assert pvalue > 0.95
+
+
+def test_normal_distribution_test_zero_std_returns_one():
+    """Regression test for issue #807: when all simulations have the same value (std=0),
+    perform_normal_distribution_test must return 1.0 instead of NaN."""
+    estimate = CausalEstimate(None, None, None, 5.0, None, None, None, None)
+    # All simulations return the same value as the estimate → std=0, p-value should be 1.0
+    simulations_identical = [5.0] * 20
+    p_value = perform_normal_distribution_test(estimate, simulations_identical)
+    assert not np.isnan(p_value), "p-value must not be NaN when std of simulations is 0"
+    assert p_value == 1.0, f"Expected p_value=1.0 for zero-variance simulations, got {p_value}"