Merge pull request #352 from batuhanovski/ssm-hyperparameter-tuning

SSM: nonignorable nuisance tuning aligned with estimation; update tests and manual tuning utils

Author: SvenKlaassen

doubleml/irm/ssm.py

@@ -428,86 +428,114 @@ def _nuisance_tuning(
     ):
         x, y = check_X_y(self._dml_data.x, self._dml_data.y, force_all_finite=False)
         x, d = check_X_y(x, self._dml_data.d, force_all_finite=False)
-        # time indicator is used for selection (selection not available in DoubleMLData yet)
         x, s = check_X_y(x, self._dml_data.s, force_all_finite=False)
 
         if self._score == "nonignorable":
             z, _ = check_X_y(self._dml_data.z, y, force_all_finite=False)
-            dx = np.column_stack((x, d, z))
-        else:
-            dx = np.column_stack((x, d))
 
         if scoring_methods is None:
             scoring_methods = {"ml_g": None, "ml_pi": None, "ml_m": None}
 
-        # nuisance training sets conditional on d
-        _, smpls_d0_s1, _, smpls_d1_s1 = _get_cond_smpls_2d(smpls, d, s)
-        train_inds = [train_index for (train_index, _) in smpls]
-        train_inds_d0_s1 = [train_index for (train_index, _) in smpls_d0_s1]
-        train_inds_d1_s1 = [train_index for (train_index, _) in smpls_d1_s1]
-
-        # hyperparameter tuning for ML
-        g_d0_tune_res = _dml_tune(
-            y,
-            x,
-            train_inds_d0_s1,
-            self._learner["ml_g"],
-            param_grids["ml_g"],
-            scoring_methods["ml_g"],
-            n_folds_tune,
-            n_jobs_cv,
-            search_mode,
-            n_iter_randomized_search,
-        )
-        g_d1_tune_res = _dml_tune(
-            y,
-            x,
-            train_inds_d1_s1,
-            self._learner["ml_g"],
-            param_grids["ml_g"],
-            scoring_methods["ml_g"],
-            n_folds_tune,
-            n_jobs_cv,
-            search_mode,
-            n_iter_randomized_search,
-        )
-        pi_tune_res = _dml_tune(
-            s,
-            dx,
-            train_inds,
-            self._learner["ml_pi"],
-            param_grids["ml_pi"],
-            scoring_methods["ml_pi"],
-            n_folds_tune,
-            n_jobs_cv,
-            search_mode,
-            n_iter_randomized_search,
-        )
-        m_tune_res = _dml_tune(
-            d,
-            x,
-            train_inds,
-            self._learner["ml_m"],
-            param_grids["ml_m"],
-            scoring_methods["ml_m"],
-            n_folds_tune,
-            n_jobs_cv,
-            search_mode,
-            n_iter_randomized_search,
-        )
+        # Nested helper functions
+        def tune_learner(target, features, train_indices, learner_key):
+            return _dml_tune(
+                target,
+                features,
+                train_indices,
+                self._learner[learner_key],
+                param_grids[learner_key],
+                scoring_methods[learner_key],
+                n_folds_tune,
+                n_jobs_cv,
+                search_mode,
+                n_iter_randomized_search,
+            )
 
-        g_d0_best_params = [xx.best_params_ for xx in g_d0_tune_res]
-        g_d1_best_params = [xx.best_params_ for xx in g_d1_tune_res]
-        pi_best_params = [xx.best_params_ for xx in pi_tune_res]
-        m_best_params = [xx.best_params_ for xx in m_tune_res]
+        def split_inner_folds(train_inds, d, s, random_state=42):
+            inner_train0_inds, inner_train1_inds = [], []
+            for train_index in train_inds:
+                stratify_vec = d[train_index] + 2 * s[train_index]
+                inner0, inner1 = train_test_split(train_index, test_size=0.5, stratify=stratify_vec, random_state=random_state)
+                inner_train0_inds.append(inner0)
+                inner_train1_inds.append(inner1)
+            return inner_train0_inds, inner_train1_inds
+
+        def filter_by_ds(inner_train1_inds, d, s):
+            inner1_d0_s1, inner1_d1_s1 = [], []
+            for inner1 in inner_train1_inds:
+                d_fold, s_fold = d[inner1], s[inner1]
+                mask_d0_s1 = (d_fold == 0) & (s_fold == 1)
+                mask_d1_s1 = (d_fold == 1) & (s_fold == 1)
+
+                inner1_d0_s1.append(inner1[mask_d0_s1])
+                inner1_d1_s1.append(inner1[mask_d1_s1])
+            return inner1_d0_s1, inner1_d1_s1
 
-        params = {"ml_g_d0": g_d0_best_params, "ml_g_d1": g_d1_best_params, "ml_pi": pi_best_params, "ml_m": m_best_params}
+        if self._score == "nonignorable":
 
-        tune_res = {"g_d0_tune": g_d0_tune_res, "g_d1_tune": g_d1_tune_res, "pi_tune": pi_tune_res, "m_tune": m_tune_res}
+            train_inds = [train_index for (train_index, _) in smpls]
+
+            # inner folds: split train set into two halves (pi-tuning vs. m/g-tuning)
+            inner_train0_inds, inner_train1_inds = split_inner_folds(train_inds, d, s)
+            # split inner1 by (d,s) to build g-models for treated/control
+            inner_train1_d0_s1, inner_train1_d1_s1 = filter_by_ds(inner_train1_inds, d, s)
+
+            # Tune ml_pi
+            x_d_z = np.column_stack((x, d, z))
+            pi_tune_res = []
+            pi_hat_full = np.full(shape=s.shape, fill_value=np.nan)
+            for inner0, inner1 in zip(inner_train0_inds, inner_train1_inds):
+                res = tune_learner(s, x_d_z, [inner0], "ml_pi")
+                best_params = res[0].best_params_
+
+                # Fit tuned model and predict
+                ml_pi_temp = clone(self._learner["ml_pi"])
+                ml_pi_temp.set_params(**best_params)
+                ml_pi_temp.fit(x_d_z[inner0], s[inner0])
+                pi_hat_full[inner1] = _predict_zero_one_propensity(ml_pi_temp, x_d_z)[inner1]
+                pi_tune_res.append(res[0])
+
+            # Tune ml_m with x + pi-hats
+            x_pi = np.column_stack([x, pi_hat_full.reshape(-1, 1)])
+            m_tune_res = tune_learner(d, x_pi, inner_train1_inds, "ml_m")
+
+            # Tune ml_g for d=0 and d=1
+            x_pi_d = np.column_stack([x, d.reshape(-1, 1), pi_hat_full.reshape(-1, 1)])
+            g_d0_tune_res = tune_learner(y, x_pi_d, inner_train1_d0_s1, "ml_g")
+            g_d1_tune_res = tune_learner(y, x_pi_d, inner_train1_d1_s1, "ml_g")
 
-        res = {"params": params, "tune_res": tune_res}
+        else:
+            # nuisance training sets conditional on d
+            _, smpls_d0_s1, _, smpls_d1_s1 = _get_cond_smpls_2d(smpls, d, s)
+            train_inds = [train_index for (train_index, _) in smpls]
+            train_inds_d0_s1 = [train_index for (train_index, _) in smpls_d0_s1]
+            train_inds_d1_s1 = [train_index for (train_index, _) in smpls_d1_s1]
+
+            # Tune ml_g for d=0 and d=1
+            g_d0_tune_res = tune_learner(y, x, train_inds_d0_s1, "ml_g")
+            g_d1_tune_res = tune_learner(y, x, train_inds_d1_s1, "ml_g")
+
+            # Tune ml_pi and ml_m
+            x_d = np.column_stack((x, d))
+            pi_tune_res = tune_learner(s, x_d, train_inds, "ml_pi")
+            m_tune_res = tune_learner(d, x, train_inds, "ml_m")
+
+        # Collect results
+        params = {
+            "ml_g_d0": [res.best_params_ for res in g_d0_tune_res],
+            "ml_g_d1": [res.best_params_ for res in g_d1_tune_res],
+            "ml_pi": [res.best_params_ for res in pi_tune_res],
+            "ml_m": [res.best_params_ for res in m_tune_res],
+        }
+
+        tune_res = {
+            "g_d0_tune": g_d0_tune_res,
+            "g_d1_tune": g_d1_tune_res,
+            "pi_tune": pi_tune_res,
+            "m_tune": m_tune_res,
+        }
 
-        return res
+        return {"params": params, "tune_res": tune_res}
 
     def _sensitivity_element_est(self, preds):
         pass

doubleml/irm/tests/_utils_ssm_manual.py

@@ -273,17 +273,14 @@ def var_selection(theta, psi_a, psi_b, n_obs):
     return var
 
 
-def tune_nuisance_ssm(y, x, d, z, s, ml_g, ml_pi, ml_m, smpls, score, n_folds_tune, param_grid_g, param_grid_pi, param_grid_m):
+def tune_nuisance_ssm_mar(y, x, d, z, s, ml_g, ml_pi, ml_m, smpls, n_folds_tune, param_grid_g, param_grid_pi, param_grid_m):
     d0_s1 = np.intersect1d(np.where(d == 0)[0], np.where(s == 1)[0])
     d1_s1 = np.intersect1d(np.where(d == 1)[0], np.where(s == 1)[0])
 
     g0_tune_res = tune_grid_search(y, x, ml_g, smpls, param_grid_g, n_folds_tune, train_cond=d0_s1)
     g1_tune_res = tune_grid_search(y, x, ml_g, smpls, param_grid_g, n_folds_tune, train_cond=d1_s1)
 
-    if score == "nonignorable":
-        dx = np.column_stack((x, d, z))
-    else:
-        dx = np.column_stack((x, d))
+    dx = np.column_stack((x, d))
 
     pi_tune_res = tune_grid_search(s, dx, ml_pi, smpls, param_grid_pi, n_folds_tune)
 
@@ -295,3 +292,44 @@ def tune_nuisance_ssm(y, x, d, z, s, ml_g, ml_pi, ml_m, smpls, score, n_folds_tu
     m_best_params = [xx.best_params_ for xx in m_tune_res]
 
     return g0_best_params, g1_best_params, pi_best_params, m_best_params
+
+
+def tune_nuisance_ssm_nonignorable(
+    y, x, d, z, s, ml_g, ml_pi, ml_m, smpls, n_folds_tune, param_grid_g, param_grid_pi, param_grid_m
+):
+
+    train_inds = [tr for (tr, _) in smpls]
+
+    inner0_list, inner1_list = [], []
+    for tr in train_inds:
+        i0, i1 = train_test_split(tr, test_size=0.5, stratify=d[tr] + 2 * s[tr], random_state=42)
+        inner0_list.append(i0)
+        inner1_list.append(i1)
+
+    X_dz = np.c_[x, d.reshape(-1, 1), z.reshape(-1, 1)]
+    pi_tune_res = tune_grid_search(s, X_dz, ml_pi, [(i0, np.array([])) for i0 in inner0_list], param_grid_pi, n_folds_tune)
+    pi_best_params = [gs.best_params_ for gs in pi_tune_res]
+
+    pi_hat_full = np.full_like(s, np.nan, dtype=float)
+    for i0, i1, gs in zip(inner0_list, inner1_list, pi_tune_res):
+        ml_pi_temp = clone(ml_pi)
+        ml_pi_temp.set_params(**gs.best_params_)
+        ml_pi_temp.fit(X_dz[i0], s[i0])
+        ph = _predict_zero_one_propensity(ml_pi_temp, X_dz)
+        pi_hat_full[i1] = ph[i1]
+
+    X_pi = np.c_[x, pi_hat_full]
+    m_tune_res = tune_grid_search(d, X_pi, ml_m, [(i1, np.array([])) for i1 in inner1_list], param_grid_m, n_folds_tune)
+    m_best_params = [gs.best_params_ for gs in m_tune_res]
+
+    X_pi_d = np.c_[x, d.reshape(-1, 1), pi_hat_full.reshape(-1, 1)]
+    inner1_d0_s1 = [i1[(d[i1] == 0) & (s[i1] == 1)] for i1 in inner1_list]
+    inner1_d1_s1 = [i1[(d[i1] == 1) & (s[i1] == 1)] for i1 in inner1_list]
+
+    g0_tune_res = tune_grid_search(y, X_pi_d, ml_g, [(idx, np.array([])) for idx in inner1_d0_s1], param_grid_g, n_folds_tune)
+    g1_tune_res = tune_grid_search(y, X_pi_d, ml_g, [(idx, np.array([])) for idx in inner1_d1_s1], param_grid_g, n_folds_tune)
+
+    g0_best_params = [gs.best_params_ for gs in g0_tune_res]
+    g1_best_params = [gs.best_params_ for gs in g1_tune_res]
+
+    return g0_best_params, g1_best_params, pi_best_params, m_best_params

doubleml/irm/tests/test_ssm_tune.py

@@ -9,7 +9,7 @@
 import doubleml as dml
 
 from ...tests._utils import draw_smpls
-from ._utils_ssm_manual import fit_selection, tune_nuisance_ssm
+from ._utils_ssm_manual import fit_selection, tune_nuisance_ssm_mar, tune_nuisance_ssm_nonignorable
 
 
 @pytest.fixture(scope="module", params=[RandomForestRegressor(random_state=42)])
@@ -115,41 +115,73 @@ def dml_ssm_fixture(
     np.random.seed(42)
     smpls = all_smpls[0]
     if tune_on_folds:
-        g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm(
-            y,
-            x,
-            d,
-            z,
-            s,
-            clone(learner_g),
-            clone(learner_m),
-            clone(learner_m),
-            smpls,
-            score,
-            n_folds_tune,
-            par_grid["ml_g"],
-            par_grid["ml_pi"],
-            par_grid["ml_m"],
-        )
+        if score == "missing-at-random":
+            g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm_mar(
+                y,
+                x,
+                d,
+                z,
+                s,
+                clone(learner_g),
+                clone(learner_m),
+                clone(learner_m),
+                smpls,
+                n_folds_tune,
+                par_grid["ml_g"],
+                par_grid["ml_pi"],
+                par_grid["ml_m"],
+            )
+        elif score == "nonignorable":
+            g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm_nonignorable(
+                y,
+                x,
+                d,
+                z,
+                s,
+                clone(learner_g),
+                clone(learner_m),
+                clone(learner_m),
+                smpls,
+                n_folds_tune,
+                par_grid["ml_g"],
+                par_grid["ml_pi"],
+                par_grid["ml_m"],
+            )
 
     else:
         xx = [(np.arange(len(y)), np.array([]))]
-        g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm(
-            y,
-            x,
-            d,
-            z,
-            s,
-            clone(learner_g),
-            clone(learner_m),
-            clone(learner_m),
-            xx,
-            score,
-            n_folds_tune,
-            par_grid["ml_g"],
-            par_grid["ml_pi"],
-            par_grid["ml_m"],
-        )
+        if score == "missing-at-random":
+            g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm_mar(
+                y,
+                x,
+                d,
+                z,
+                s,
+                clone(learner_g),
+                clone(learner_m),
+                clone(learner_m),
+                xx,
+                n_folds_tune,
+                par_grid["ml_g"],
+                par_grid["ml_pi"],
+                par_grid["ml_m"],
+            )
+        elif score == "nonignorable":
+            g0_best_params, g1_best_params, pi_best_params, m_best_params = tune_nuisance_ssm_nonignorable(
+                y,
+                x,
+                d,
+                z,
+                s,
+                clone(learner_g),
+                clone(learner_m),
+                clone(learner_m),
+                xx,
+                n_folds_tune,
+                par_grid["ml_g"],
+                par_grid["ml_pi"],
+                par_grid["ml_m"],
+            )
 
         g0_best_params = g0_best_params * n_folds
         g1_best_params = g1_best_params * n_folds