Merge pull request #391 from uriahf/fix/static-interventions-avoided-scaling

Fix static Interventions Avoided per-100 scaling

Author: uriahf

CHANGELOG.md

@@ -2,6 +2,8 @@
 
 <!--next-version-placeholder-->
 
+- Fixed Interventions Avoided to apply the per-100 scaling to the full model expression, including the false-negative penalty term.
+
 ## v0.1.36 (21/08/2026)
 
 - Fixed several binary and time-dependent curve consistency issues, including cutoff-grid endpoints, binary cutoff equality, time-dependent reference prevalence, gains perfect-reference behavior, custom colors, and plot sizing.

src/rtichoke/processing/transforms.py

@@ -684,10 +684,13 @@ def _turn_cumulative_aj_to_performance_data(
         .alias("net_benefit"),
         pl.when(pl.col("stratified_by") == "probability_threshold")
         .then(
-            100 * (pl.col("true_negatives") / pl.col("n"))
-            - (pl.col("false_negatives") / pl.col("n"))
-            * (1 - pl.col("chosen_cutoff"))
-            / pl.col("chosen_cutoff")
+            100
+            * (
+                (pl.col("true_negatives") / pl.col("n"))
+                - (pl.col("false_negatives") / pl.col("n"))
+                * (1 - pl.col("chosen_cutoff"))
+                / pl.col("chosen_cutoff")
+            )
         )
         .otherwise(None)
         .alias("net_benefit_interventions_avoided"),

tests/test_interventions_avoided_scaling.py

@@ -0,0 +1,100 @@
+import numpy as np
+import polars as pl
+import pytest
+
+from rtichoke import (
+    create_decision_curve,
+    plot_decision_curve,
+    prepare_performance_data,
+)
+from rtichoke.processing.plotly_helper_functions import _create_reference_lines_data
+
+
+PROBS = {"model": np.array([0.9, 0.8, 0.7, 0.6, 0.4, 0.3, 0.2, 0.1])}
+REALS = np.array([1, 0, 1, 0, 1, 0, 0, 1])
+THRESHOLDS = [0.25, 0.5, 0.75]
+EXPECTED_IA = [-25.0, 0.0, 25.0]
+OLD_BUGGY_IA = [12.125, 24.75, 37.375]
+EXPECTED_TN = [1.0, 2.0, 3.0]
+EXPECTED_FN = [1.0, 2.0, 3.0]
+
+
+def _performance_rows() -> pl.DataFrame:
+    data = prepare_performance_data(probs=PROBS, reals=REALS, by=0.25)
+    data = data.filter(pl.col("chosen_cutoff").is_in(THRESHOLDS))
+    return data.sort("chosen_cutoff")
+
+
+def test_static_interventions_avoided_matches_r_definition_per_100():
+    """Expected values characterize the current R static IA definition."""
+    rows = _performance_rows()
+
+    np.testing.assert_allclose(rows["true_negatives"].to_numpy(), EXPECTED_TN)
+    np.testing.assert_allclose(rows["false_negatives"].to_numpy(), EXPECTED_FN)
+    actual = rows["net_benefit_interventions_avoided"].to_numpy()
+    np.testing.assert_allclose(actual, EXPECTED_IA)
+    assert not np.allclose(actual, OLD_BUGGY_IA)
+
+
+def test_static_interventions_avoided_count_and_nb_forms_are_equivalent():
+    rows = _performance_rows()
+    prevalence = float(REALS.mean())
+
+    for row in rows.iter_rows(named=True):
+        threshold = float(row["chosen_cutoff"])
+        tn = float(row["true_negatives"])
+        fn = float(row["false_negatives"])
+        n = float(row["n"])
+        net_benefit = float(row["net_benefit"])
+        net_benefit_all = prevalence - (1 - prevalence) * threshold / (1 - threshold)
+        from_counts = 100 * (tn / n - fn / n * (1 - threshold) / threshold)
+        from_nb = 100 * (net_benefit - net_benefit_all) * (1 - threshold) / threshold
+        actual = float(row["net_benefit_interventions_avoided"])
+
+        assert actual == pytest.approx(from_counts)
+        assert from_counts == pytest.approx(from_nb)
+
+
+def test_interventions_avoided_model_and_references_use_per_100_units():
+    rows = _performance_rows()
+    aj = pl.DataFrame({"reference_group": ["population"], "aj_estimate": [0.5]})
+    refs = _create_reference_lines_data(
+        curve="interventions avoided",
+        aj_estimates_from_performance_data=aj,
+        multiple_populations=False,
+        min_p_threshold=0.25,
+        max_p_threshold=0.75,
+    )
+
+    treat_all = refs.filter(pl.col("reference_group") == "treat_all")
+    assert np.allclose(treat_all["y"].to_numpy(), 0.0)
+
+    is_treat_none = pl.col("reference_group") == "treat_none"
+    is_test_threshold = pl.col("x").is_in(THRESHOLDS)
+    treat_none = refs.filter(is_treat_none & is_test_threshold).sort("x")
+    expected_treat_none = [-100.0, 0.0, 100.0 / 3.0]
+    np.testing.assert_allclose(treat_none["y"].to_numpy(), expected_treat_none)
+    np.testing.assert_allclose(
+        rows["net_benefit_interventions_avoided"].to_numpy(), EXPECTED_IA
+    )
+
+
+def test_static_interventions_avoided_public_apis_are_unchanged():
+    data = prepare_performance_data(probs=PROBS, reals=REALS, by=0.25)
+    created = create_decision_curve(
+        probs=PROBS,
+        reals=REALS,
+        decision_type="interventions avoided",
+        by=0.25,
+        min_p_threshold=0.25,
+        max_p_threshold=0.75,
+    )
+    plotted = plot_decision_curve(
+        data,
+        decision_type="interventions avoided",
+        min_p_threshold=0.25,
+        max_p_threshold=0.75,
+    )
+
+    assert created.__class__.__name__ == "Figure"
+    assert plotted.__class__.__name__ == "Figure"