Fix IG LL

CHANGELOG.rst

@@ -7,13 +7,22 @@
 Changelog
 =========
 
+3.2.3 - 2026-03-18
+------------------
+
+**Bug fix:**
+
+- Fixed incorrect call in :meth:`~glum.InverseGaussianDistribution.log_likelihood`. The previous implementation always returned NaN.
+
+
 3.2.2 - 2026-03-17
 ------------------
 
 **Bug fix:**
 
 - Fixed incorrect formula in :meth:`~glum.CloglogLink.inverse_derivative2`. This affected observed information matrix computation and robust/clustered standard errors for models using the complementary log-log link.
 
+
 3.2.1 - 2026-03-16
 ------------------
 

src/glum/_distribution.py

@@ -1145,7 +1145,7 @@ def deviance(self, y, mu, sample_weight=None) -> float:  # noqa D
         y, mu, sample_weight = _as_float_arrays(y, mu, sample_weight)
         sample_weight = np.ones_like(y) if sample_weight is None else sample_weight
 
-        return tweedie_deviance(y, sample_weight, mu, p=3.0)
+        return inv_gaussian_deviance(y, sample_weight, mu, dispersion=1.0)
 
     def unit_deviance(self, y, mu):  # noqa D
         return numexpr.evaluate("y / (mu**2) + 1 / y - 2 / mu")
@@ -1205,7 +1205,7 @@ def log_likelihood(self, y, mu, sample_weight=None, dispersion=None) -> float:
         if dispersion is None:
             dispersion = self.dispersion(y, mu, sample_weight)
 
-        return tweedie_log_likelihood(y, sample_weight, mu, 3.0, float(dispersion))
+        return inv_gaussian_log_likelihood(y, sample_weight, mu, float(dispersion))
 
     def dispersion(  # noqa D
         self, y, mu, sample_weight=None, ddof=1, method="pearson"

src/glum/_functions.pyx

@@ -69,10 +69,10 @@ def normal_log_likelihood(
     floating dispersion,
 ):
     cdef int i  # loop counter
-    cdef floating sum_weights  # helper
 
     cdef int n = y.shape[0]  # loop length
     cdef floating ll = 0.0  # output
+    cdef floating sum_weights = 0.0  # helper
 
     for i in prange(n, nogil=True):
         ll -= weights[i] * (y[i] - mu[i]) ** 2
@@ -203,13 +203,15 @@ def gamma_log_likelihood(
     floating dispersion,
 ):
     cdef int i  # loop counter
-    cdef floating ln_y, sum_weights  # helpers
+    cdef floating ln_y  # helper
 
-    cdef int n = y.shape[0]  # loop length
-    cdef floating ll = 0.0  # output
     cdef floating inv_dispersion = 1 / dispersion
     cdef floating normalization = log(dispersion) * inv_dispersion + lgamma(inv_dispersion)
 
+    cdef int n = y.shape[0]  # loop length
+    cdef floating ll = 0.0  # output
+    cdef floating sum_weights = 0.0  # helper
+
     for i in prange(n, nogil=True):
         ln_y = log(y[i])
         ll += weights[i] * (inv_dispersion * (ln_y - log(mu[i]) - y[i] / mu[i]) - ln_y)
@@ -287,22 +289,23 @@ def inv_gaussian_log_likelihood(
     const_floating1d mu,
     floating dispersion,
 ):
-    cdef int n = y.shape[0]  # loop length
     cdef int i  # loop counter
-    cdef floating sum_weights  # helper
-    cdef floating ll = 0.0  # output
 
+    cdef floating scale = 1 / (2 * dispersion)
     cdef floating sq_err  # helper
-    cdef floating inv_dispersion = 1 / (2 * dispersion)  # helper
+
+    cdef int n = y.shape[0]  # loop length
+    cdef floating sum_weights = 0.0  # helper
+    cdef floating ll = 0.0  # output
 
     for i in prange(n, nogil=True):
 
         sq_err = (y[i] / mu[i] - 1) ** 2
 
-        ll -= weights[i] * (inv_dispersion * sq_err / y[i] + log(y[i]) * 3 / 2)
-        sum_weights -= weights[i]
+        ll -= weights[i] * (scale * sq_err / y[i] + log(y[i]) * 3 / 2)
+        sum_weights += weights[i]
 
-    return ll + sum_weights * log(inv_dispersion / M_PI)
+    return ll + sum_weights * log(scale / M_PI) / 2
 
 def inv_gaussian_deviance(
     const_floating1d y,

src/glum/_link.py

@@ -110,7 +110,7 @@ def __init__(self, power):
     def __eq__(self, other):  # noqa D
         return isinstance(other, self.__class__) and (self.power == other.power)
 
-    def __tweedie__repr__(self):  # noqa D
+    def __tweedie_repr__(self):  # noqa D
         return self.__class__(self.power)
 
     @property

tests/glm/test_distribution.py

@@ -1,6 +1,7 @@
 import numpy as np
 import pytest
 import scipy as sp
+import scipy.stats
 import tabmat as tm
 
 from glum._distribution import (
@@ -540,6 +541,34 @@ def test_binomial_deviance_dispersion_loglihood(weighted):
     np.testing.assert_approx_equal(ll, -3.365058)
 
 
+@pytest.mark.parametrize("dispersion", [1, 2])
+def test_inv_gaussian_deviance_loglihood(dispersion):
+
+    n = 1_000
+    rng = np.random.default_rng(42)
+
+    mu = rng.uniform(1, 2, size=n)
+
+    y = rng.wald(mu, 1 / dispersion, size=n)
+
+    candidate = InverseGaussianDistribution().log_likelihood(y, mu, None, dispersion)
+
+    target = scipy.stats.invgauss.logpdf(
+        y, mu * dispersion, scale=(1 / dispersion)
+    ).sum()
+
+    np.testing.assert_approx_equal(candidate, target)
+
+    candidate = InverseGaussianDistribution().deviance(y, mu)
+
+    target = 2 * (
+        scipy.stats.invgauss.logpdf(y, y).sum()
+        - scipy.stats.invgauss.logpdf(y, mu).sum()
+    )
+
+    np.testing.assert_approx_equal(candidate, target)
+
+
 @pytest.mark.parametrize("weighted", [False, True])
 def test_negative_binomial_deviance_dispersion_loglihood(weighted):
     # y <- c(0, 1, 0, 1, 0)