Fix gaussian glm se (#1078)

* make env work

* fix docs

* bring back benchmarks

* use github resource

* only bread for iid inference

* delete run.py

* refine testing strategy

* refine

Author: s3alfisc

pyfixest/estimation/fegaussian_.py

@@ -94,15 +94,6 @@ def _get_theta(self, mu: np.ndarray) -> np.ndarray:
     def _get_V(self, mu: np.ndarray) -> np.ndarray:
         return np.ones_like(mu)
 
-    def _vcov_iid(self):
-        _N = self._N
-        _u_hat = self._u_hat
-        _bread = self._bread
-        sigma2 = np.sum(_u_hat.flatten() ** 2) / (_N)
-        _vcov = _bread * sigma2
-
-        return _vcov
-
     def _get_score(
         self, y: np.ndarray, X: np.ndarray, mu: np.ndarray, eta: np.ndarray
     ) -> np.ndarray:

tests/test_feols_feglm_internally.py

@@ -49,12 +49,15 @@ def test_ols_vs_gaussian_glm(fml, inference, dropna):
     check_absolute_diff(fit_ols._u_hat[0:5], fit_gaussian._u_hat[0:5], tol=1e-10)
     check_absolute_diff(fit_ols._scores[0, :], fit_gaussian._scores[0, :], tol=1e-10)
 
-    if inference != "iid":
+    if inference == "iid":
         # iid inference different: follows iid-glm; just the bread and not bread x sigma2
+        scaling_factor = fit_ols._vcov[0, 0] / fit_gaussian._vcov[0, 0]
+        # Check that all elements follow the same scaling
         check_absolute_diff(
-            fit_ols.se().xs("X1"), fit_gaussian.se().xs("X1"), tol=1e-10
+            fit_ols._vcov, scaling_factor * fit_gaussian._vcov, tol=1e-10
         )
-        check_absolute_diff(fit_ols._vcov[0, 0], fit_gaussian._vcov[0, 0], tol=1e-10)
+    else:
+        check_absolute_diff(fit_ols._vcov, fit_gaussian._vcov, tol=1e-10)
 
 
 @pytest.mark.skip("Fixed effects are not yet supported.")