ENH/BUG: correct predict with offset

statsmodels/miscmodels/ordinal_model.py

@@ -245,21 +245,15 @@ def transform_reverse_threshold_params(self, params):
                                         np.log(np.diff(start_ppf[:-1]))))
         return thresh_params
 
-    def predict(self, params, exog=None):
+    def predict(self, params, exog=None, offset=None):
         """predicted probabilities for each level of the ordinal endog.
 
 
         """
-        if exog is None:
-            exog = self.exog
-        # structure of params = [beta, constants_or_thresholds]
+        # note, exog and offset handling is in linpred
 
-        # explicit in several steps to avoid bugs
-        th_params = params[-(self.k_levels - 1):]
-        thresh = np.concatenate((th_params[:1],
-                                 np.exp(th_params[1:]))).cumsum()
-        thresh = np.concatenate(([-np.inf], thresh, [np.inf]))
-        xb = exog.dot(params[:-(self.k_levels - 1)])[:, None]
+        thresh = self.transform_threshold_params(params)
+        xb = self._linpred(params, exog=exog, offset=offset)[:, None]
         low = thresh[:-1] - xb
         upp = thresh[1:] - xb
         prob = self.prob(low, upp)
@@ -272,10 +266,17 @@ def _linpred(self, params, exog=None, offset=None):
         """
         if exog is None:
             exog = self.exog
-        if offset is None:
-            offset = self.offset
+            if offset is None:
+                offset = self.offset
+        else:
+            if offset is None:
+                offset = 0
+
+        if offset is not None:
+            offset = np.asarray(offset)
+
         if exog is not None:
-            linpred = self.exog.dot(params[:-(self.k_levels - 1)])
+            linpred = exog.dot(params[:-(self.k_levels - 1)])
         else:  # means self.exog is also None
             linpred = np.zeros(self.nobs)
         if offset is not None:

statsmodels/miscmodels/tests/test_ordinal_model.py

@@ -240,6 +240,8 @@ def test_formula_categorical(self):
         assert not hasattr(modf2, "frame")
 
         with pytest.raises(ValueError):
+            # only ordered categorical or numerical endog are allowed
+            # string endog raises ValueError
             OrderedModel.from_formula(
                 "apply ~ pared + public + gpa - 1",
                 data={"apply": np.asarray(data['apply']),
@@ -248,6 +250,41 @@ def test_formula_categorical(self):
                       "gpa": data['gpa']},
                 distr='probit')
 
+    def test_offset(self):
+
+        resp = self.resp
+        data = ds.df
+        offset = np.ones(len(data))
+
+        formula = "apply ~ pared + public + gpa - 1"
+        modf2 = OrderedModel.from_formula(formula, data, offset=offset,
+                                          distr='probit')
+        resf2 = modf2.fit(method='bfgs', disp=False)
+
+        assert_allclose(resf2.params[:3], resp.params[:3], atol=2e-4)
+        assert_allclose(resf2.params[3], resp.params[3] + 1, atol=2e-4)
+
+        fitted = resp.predict()
+        fitted2 = resf2.predict()
+        assert_allclose(fitted2, fitted, atol=2e-4)
+
+        pred_ones = resf2.predict(data[:6], offset=np.ones(6))
+        assert_allclose(pred_ones, fitted[:6], atol=2e-4)
+
+        # check default is 0. if exog provided
+        pred_zero1 = resf2.predict(data[:6])
+        pred_zero2 = resf2.predict(data[:6], offset=0)
+        assert_allclose(pred_zero1, pred_zero2, atol=2e-4)
+
+        # compare with equivalent results frp, no-offset model
+        pred_zero = resp.predict(data.iloc[:6, 1:], offset=-np.ones(6))
+        assert_allclose(pred_zero1, pred_zero, atol=2e-4)
+
+        params_adj = resp.params.copy()
+        params_adj[3] += 1
+        fitted_zero = resp.model.predict(params_adj)
+        assert_allclose(pred_zero1, fitted_zero[:6], atol=2e-4)
+
 
 class TestLogitModelFormula():