fix: bootstrap ratios from raw saliences, not S-scaled loadings

CITATION.cff

@@ -8,8 +8,8 @@ authors:
     orcid: "https://orcid.org/0000-0003-3715-7012"
 license: BSD-3-Clause
 repository-code: "https://github.com/braincentrekcl/plsdo"
-version: "0.1.1"
-date-released: "2026-05-22"
+version: "0.2.0"
+date-released: "2026-06-05"
 doi: 10.5281/zenodo.20333467
 identifiers:
   - type: doi

plsdo/__init__.py

@@ -1,3 +1,3 @@
 """PLS covariance analysis with statistical testing and visualisation."""
 
-__version__ = "0.1.1"
+__version__ = "0.2.0"

plsdo/core.py

@@ -50,8 +50,8 @@ def fit(self):
         self.xcorr = self.X.T @ self.Y / (self.n_subjects - 1)
         self._decompose()
         self._fix_component_signs()
-        self.u_loadings = self.u * self.s[np.newaxis, :]
-        self.vt_loadings = self.s[:, np.newaxis] * self.vt
+        self.u_loadings = self.u
+        self.vt_loadings = self.vt
         self.x_scores = self.X @ self.u
         self.y_scores = self.Y @ self.vt.T
         self._fitted = True
@@ -141,22 +141,19 @@ def bootstrap(self, n_bootstraps: int = 10000) -> None:
             y_boot = zscore_columns(self.Y[idx, :])
 
             boot_xcorr = x_boot.T @ y_boot / (self.n_subjects - 1)
-            boot_u, boot_s, boot_vt = np.linalg.svd(boot_xcorr, full_matrices=False)
+            boot_u, _, boot_vt = np.linalg.svd(boot_xcorr, full_matrices=False)
 
             # Procrustes: rotate bootstrap Vt to align with reference
             Q, _ = orthogonal_procrustes(boot_vt.T, self.vt.T)
 
-            boot_u_load = boot_u * boot_s[np.newaxis, :]
-            boot_vt_load = boot_s[:, np.newaxis] * boot_vt
-
-            aligned_u_load = boot_u_load @ Q
-            aligned_vt_load = Q.T @ boot_vt_load
+            aligned_u = boot_u @ Q
+            aligned_vt = Q.T @ boot_vt
 
             # No separate sign correction: orthogonal_procrustes returns an
             # unconstrained orthogonal matrix (reflections allowed), so the
             # alignment above already resolves each component's arbitrary sign.
-            u_distribution.append(aligned_u_load)
-            vt_distribution.append(aligned_vt_load)
+            u_distribution.append(aligned_u)
+            vt_distribution.append(aligned_vt)
 
         self.u_se = np.std(np.stack(u_distribution, axis=2), axis=2, ddof=1)
         self.vt_se = np.std(np.stack(vt_distribution, axis=2), axis=2, ddof=1)

tests/test_core.py

@@ -36,18 +36,20 @@ def test_cross_correlation_formula(self, x_array, y_array):
         expected = X.T @ Y / (X.shape[0] - 1)
         np.testing.assert_allclose(model.xcorr, expected)
 
-    def test_loadings_are_scaled_vectors(self, x_array, y_array):
+    def test_loadings_are_raw_saliences(self, x_array, y_array):
+        """Loadings are the singular vectors (saliences) themselves, not
+        scaled by the singular values. Bootstrap ratios are formed from these
+        raw saliences (McIntosh & Lobaugh 2004; Krishnan et al. 2011), so the
+        singular values must not enter and contaminate the bootstrap SE."""
         from plsdo.io import zscore_columns
 
         X = zscore_columns(x_array)
         Y = zscore_columns(y_array)
         model = PLS(X, Y)
         model.fit()
 
-        expected_u_load = model.u @ np.diag(model.s)
-        expected_vt_load = np.diag(model.s) @ model.vt
-        np.testing.assert_allclose(model.u_loadings, expected_u_load)
-        np.testing.assert_allclose(model.vt_loadings, expected_vt_load)
+        np.testing.assert_allclose(model.u_loadings, model.u)
+        np.testing.assert_allclose(model.vt_loadings, model.vt)
 
     def test_scores_are_projections(self, x_array, y_array):
         from plsdo.io import zscore_columns
@@ -462,8 +464,8 @@ def test_single_x_feature(self):
     def test_additive_multifactor_trailing_lv_is_inert(self):
         """An additive K-factor dummy design is rank-deficient by K-1. The
         degenerate trailing latent variable must be harmless: ~zero singular
-        value, non-significant, dropped by filter_lvs, with ~zero loadings.
-        (Guards the design without switching to contrast coding.)"""
+        value, non-significant, dropped by filter_lvs. (Guards the design
+        without switching to contrast coding.)"""
         from plsdo.io import zscore_columns
 
         a = np.repeat(np.arange(3), 4)  # factor A, 3 levels, 12 subjects
@@ -480,7 +482,6 @@ def test_additive_multifactor_trailing_lv_is_inert(self):
         assert model.s[-1] < 1e-8
         assert not model.significant_lvs[-1]
         assert not model.final_lvs[-1]
-        assert np.abs(model.u_loadings[:, -1]).max() < 1e-6
 
     def test_many_group_discriminatory(self):
         """Dummy-coded X with five groups gives min(n_groups, n_y) components."""