PyFixest 0.9.12 (#167)

* pyfixest 0.9.11

* prepare iid inference for poisson #166
* add support for fepois iid errors #169
* bugfix #171

pyfixest/estimation.py

@@ -149,7 +149,8 @@ def fepois(
     """
     # fepois
 
-    Method for estimating Poisson regression models with fixed effects.
+    Method for estimating Poisson regression models with fixed effects. Implements the `pplmhdfe` algorithm from the
+    Stata package of the same name.
 
     Args:
         fml (str): A two-sided formula string using fixest formula syntax.

pyfixest/feols.py

@@ -238,10 +238,19 @@ def vcov(self, vcov: Union[str, Dict[str, str]]):
                 vcov_type="iid",
             )
 
-            sigma2 = np.sum((_u_hat.flatten()) ** 2) / (_N - 1)
-            self._vcov = self._ssc * bread * sigma2
+            if self._method == "feols":
+                sigma2 = np.sum(_u_hat.flatten() ** 2) / (_N - 1)
+            elif self._method == "fepois":
+                sigma2 = 1
+            else:
+                raise NotImplementedError(
+                    f"'iid' inference is not supported for {_method} regressions."
+                )
+
+            self._vcov =  self._ssc * bread * sigma2
 
         elif self._vcov_type == "hetero":
+
             self._ssc = get_ssc(
                 ssc_dict=_ssc_dict,
                 N=_N,
@@ -280,6 +289,10 @@ def vcov(self, vcov: Union[str, Dict[str, str]]):
                 self._vcov = self._ssc * bread @ meat @ bread
 
         elif self._vcov_type == "CRV":
+
+            #import pdb;
+            #pdb.set_trace()
+
             cluster_df = _data[self._clustervar]
             # if there are missings - delete them!
 
@@ -327,8 +340,6 @@ def vcov(self, vcov: Union[str, Dict[str, str]]):
 
                 if _is_iv == False:
                     self._vcov = self._ssc * bread @ meat @ bread
-                # if self._is_iv == False:
-                #    self._vcov = self._ssc * bread @ meat @ bread
                 else:
                     meat = _tXZ @ _tZZinv @ meat @ _tZZinv @ self._tZX
                     self._vcov = self._ssc * bread @ meat @ bread
@@ -338,9 +349,6 @@ def vcov(self, vcov: Union[str, Dict[str, str]]):
                 # if not, either error or turn fixefs into dummies
                 # for now: don't allow for use with fixed effects
 
-                # if self._has_fixef:
-                #    raise ValueError("CRV3 inference is currently not supported with fixed effects.")
-
                 if _is_iv:
                     raise VcovTypeNotSupportedError(
                         "CRV3 inference is not supported with IV estimation."

pyfixest/fepois.py

@@ -78,7 +78,7 @@ def __init__(
         self._check_for_separation()
 
         self._support_crv3_inference = False
-        self._support_iid_inference = False
+        self._support_iid_inference = True
 
         # attributes that are updated outside of the class (not optimal)
         self._N_separation_na = None
@@ -242,6 +242,7 @@ def compute_deviance(_Y, mu):
 
         self._scores = self._u_hat[:, None] * self._weights * X_resid
         self._hessian = XWX
+        self._T = self._weights * X_resid
 
         if _convergence:
             self._convergence = True
@@ -306,7 +307,8 @@ def _check_for_separation(self, check: str = "fe") -> None:
         if check == "fe":
             if not self._has_fixef:
                 pass
-
+            elif (self._Y > 0).all():
+                pass
             else:
                 Y_help = pd.Series(np.where(self._Y.flatten() > 0, 1, 0))
                 fe = pd.DataFrame(self.fe)

pyfixest/utils.py

@@ -77,8 +77,6 @@ def get_ssc(ssc_dict, N, k, G, vcov_sign, vcov_type):
     elif vcov_type in ["iid", "CRV"]:
         if adj:
             adj_value = (N - 1) / (N - k)
-        else:
-            adj_value = 1
     else:
         raise ValueError("vcov_type must be either iid, hetero or CRV.")
 

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "pyfixest"
-version = "0.9.11"
+version = "0.9.12"
 
 
 description = "(Not so) experimental draft package for high dimensional fixed effect estimation. Supports OLS, IV and Poisson regression and a range of inference procedures."

tests/check-crv-diffs-fixest-pyfixest-glm.qmd

@@ -10,7 +10,7 @@ Check that `PyFixest` inference for CRV gets as close to `fixest` as `glm`.
 
 
 ```{python, messages = FALSE, warnings = FALSE}
-import pyfixest as pf
+from pyfixest.estimation import fepois
 from pyfixest.utils import get_data
 
 fml = "Y ~ X1 + X2"
@@ -21,7 +21,7 @@ data_list = []
 for x in range(1000):
 
   data = get_data(100, x, "2", "3", "Fepois").dropna()
-  fit = pf.Fixest(data = data).fepois(fml, vcov = {'CRV1':"group_id"}).fetch_model(0)
+  fit = fepois(fml, vcov = {'CRV1':"group_id"})
   se_list.append(fit._se)
   data_list.append(data)
 

tests/test_vs_fixest.py

@@ -34,9 +34,9 @@
 @pytest.mark.parametrize("N", [1000])
 @pytest.mark.parametrize("seed", [76540251])
 @pytest.mark.parametrize("beta_type", ["1", "2", "3"])
-@pytest.mark.parametrize("error_type", ["1", "2", "3"])
-@pytest.mark.parametrize("dropna", [False, True])
-@pytest.mark.parametrize("model", ["Feols", "Fepois"])
+@pytest.mark.parametrize("error_type",["1", "2", "3"])
+@pytest.mark.parametrize("dropna", [True, False])
+@pytest.mark.parametrize("model", ["Fepois", "Feols"])
 @pytest.mark.parametrize("inference", ["iid", "hetero", {"CRV1": "group_id"}])
 @pytest.mark.parametrize(
     "fml",
@@ -176,8 +176,8 @@ def test_single_fit(N, seed, beta_type, error_type, dropna, model, inference, fm
             warnings.filterwarnings("ignore", category=RuntimeWarning)
             iv_check = feols(fml=fml, data=data, vcov="iid")
 
-        if inference == "iid":
-            return pytest.skip("Poisson does not support iid inference")
+        #if inference == "iid":
+        #    return pytest.skip("Poisson does not support iid inference")
 
         if iv_check._is_iv:
             is_iv = True
@@ -206,10 +206,7 @@ def test_single_fit(N, seed, beta_type, error_type, dropna, model, inference, fm
 
             # relax tolerance for Poisson regression - effective rtol and atol of
             # 5e-05
-            inference_inflation_factor = 100
-            # relax tolerance for CRV inference - effective rtol and atol of 5e-03
-            if isinstance(inference, dict):
-                inference_inflation_factor = 500
+            inference_inflation_factor = 50
 
             try:
                 pyfixest = fepois(fml=fml, data=data, vcov=inference)