BUG: properly assign zero probability to outside-outside first-second…

… choices
jeffgortmaker · Jun 19, 2022 · 98cbf85 · 98cbf85
1 parent 42930bf
commit 98cbf85
Show file tree

Hide file tree

Showing 3 changed files with 10 additions and 5 deletions.
diff --git a/pyblp/markets/market.py b/pyblp/markets/market.py
@@ -1504,6 +1504,8 @@ def compute_micro_dataset_contributions(
                 if weights.shape[2] == 1 + self.J:
                     assert eliminated_outside_probabilities is not None
                     dataset_weights[:, -self.J:, 0] *= eliminated_outside_probabilities.T
+                if weights.shape[1] == weights.shape[2] == 1 + self.J:
+                    dataset_weights[:, 0, 0] = 0
 
             # truncate numerical errors from below by zero
             dataset_weights[dataset_weights < 0] = 0
@@ -1541,6 +1543,9 @@ def compute_micro_dataset_contributions(
                             assert eliminated_outside_probabilities is not None
                             product1[:, -self.J:, 0] *= eliminated_outside_probabilities.T
                             product2[:, -self.J:, 0] *= eliminated_outside_probabilities_tangent_mapping[p].T
+                        if weights.shape[1] == weights.shape[2] == 1 + self.J:
+                            product1[:, 0, 0] = 0
+                            product2[:, 0, 0] = 0
 
                         weights_tangent *= product1 + product2
 

diff --git a/tests/conftest.py b/tests/conftest.py
@@ -396,17 +396,17 @@ def large_blp_simulation() -> SimulationFixture:
             ),
         ),
         MicroMoment(
-            name="diversion interaction",
+            name="unconditional diversion interaction",
             dataset=MicroDataset(
                 name="inside first and second",
                 observations=simulation.N,
-                compute_weights=lambda _, p, a: np.ones((a.size, p.size, p.size)),
+                compute_weights=lambda _, p, a: np.ones((a.size, 1 + p.size, 1 + p.size)),
                 market_ids=[simulation.unique_market_ids[0]],
             ),
             value=0,
             compute_values=lambda _, p, a: (
-                np.tile(p.X2[:, [2]], (a.size, 1, p.size)) *
-                np.tile(p.X2[:, [2]], (a.size, 1, p.size)).swapaxes(1, 2)
+                np.tile(np.c_[np.r_[0, p.X2[:, 2]]], (a.size, 1, 1 + p.size)) *
+                np.tile(np.c_[np.r_[0, p.X2[:, 2]]], (a.size, 1, 1 + p.size)).swapaxes(1, 2)
             ),
         ),
     ])

diff --git a/tests/test_blp.py b/tests/test_blp.py
@@ -1482,7 +1482,7 @@ def test_objective_gradient(
     def test_finite_differences(theta: Array, _: Any, objective_function: Callable, __: Any) -> Tuple[Array, bool]:
         """Test central finite differences around starting parameter values."""
         approximated = compute_finite_differences(lambda x: objective_function(x)[0], theta, epsilon_scale=10.0)
-        np.testing.assert_allclose(approximated.flatten(), exact.flatten(), atol=1e-8, rtol=1e-2)
+        np.testing.assert_allclose(approximated.flatten(), exact.flatten(), atol=1e-8, rtol=1e-3)
         return theta, True
 
     # test the gradient