|
| 1 | +from typing import Any, Dict, Optional |
| 2 | + |
| 3 | +import doubleml as dml |
| 4 | +import numpy as np |
| 5 | +import optuna |
| 6 | +import pandas as pd |
| 7 | +from doubleml.did.datasets import make_did_CS2021 |
| 8 | + |
| 9 | +from montecover.base import BaseSimulation |
| 10 | +from montecover.utils import create_learner_from_config |
| 11 | + |
| 12 | + |
| 13 | +class DIDMultiTuningCoverageSimulation(BaseSimulation): |
| 14 | + """Simulation study for coverage properties of DoubleMLDIDMulti with hyperparameter tuning.""" |
| 15 | + |
| 16 | + def __init__( |
| 17 | + self, |
| 18 | + config_file: str, |
| 19 | + suppress_warnings: bool = True, |
| 20 | + log_level: str = "INFO", |
| 21 | + log_file: Optional[str] = None, |
| 22 | + ): |
| 23 | + super().__init__( |
| 24 | + config_file=config_file, |
| 25 | + suppress_warnings=suppress_warnings, |
| 26 | + log_level=log_level, |
| 27 | + log_file=log_file, |
| 28 | + ) |
| 29 | + |
| 30 | + # Additional results storage for aggregated results |
| 31 | + self.results_aggregated = [] |
| 32 | + |
| 33 | + # Calculate oracle values |
| 34 | + self._calculate_oracle_values() |
| 35 | + |
| 36 | + # tuning specific settings |
| 37 | + # parameter space for the outcome regression tuning |
| 38 | + def ml_g_params(trial): |
| 39 | + return { |
| 40 | + "n_estimators": trial.suggest_int("n_estimators", 100, 200, step=50), |
| 41 | + "learning_rate": trial.suggest_float( |
| 42 | + "learning_rate", 1e-3, 0.1, log=True |
| 43 | + ), |
| 44 | + "min_child_samples": trial.suggest_int( |
| 45 | + "min_child_samples", 20, 50, step=5 |
| 46 | + ), |
| 47 | + "max_depth": 5, |
| 48 | + "lambda_l1": trial.suggest_float("lambda_l1", 1e-3, 10.0, log=True), |
| 49 | + "lambda_l2": trial.suggest_float("lambda_l2", 1e-3, 10.0, log=True), |
| 50 | + } |
| 51 | + |
| 52 | + # parameter space for the propensity score tuning |
| 53 | + def ml_m_params(trial): |
| 54 | + return { |
| 55 | + "n_estimators": trial.suggest_int("n_estimators", 100, 200, step=50), |
| 56 | + "learning_rate": trial.suggest_float( |
| 57 | + "learning_rate", 1e-3, 0.1, log=True |
| 58 | + ), |
| 59 | + "min_child_samples": trial.suggest_int( |
| 60 | + "min_child_samples", 20, 50, step=5 |
| 61 | + ), |
| 62 | + "max_depth": 5, |
| 63 | + "lambda_l1": trial.suggest_float("lambda_l1", 1e-3, 10.0, log=True), |
| 64 | + "lambda_l2": trial.suggest_float("lambda_l2", 1e-3, 10.0, log=True), |
| 65 | + } |
| 66 | + |
| 67 | + self._param_space = {"ml_g": ml_g_params, "ml_m": ml_m_params} |
| 68 | + |
| 69 | + self._optuna_settings = { |
| 70 | + "n_trials": 200, |
| 71 | + "show_progress_bar": False, |
| 72 | + "verbosity": optuna.logging.WARNING, # Suppress Optuna logs |
| 73 | + } |
| 74 | + |
| 75 | + def _process_config_parameters(self): |
| 76 | + """Process simulation-specific parameters from config""" |
| 77 | + # Process ML models in parameter grid |
| 78 | + # Process ML models in parameter grid |
| 79 | + assert ( |
| 80 | + "learners" in self.dml_parameters |
| 81 | + ), "No learners specified in the config file" |
| 82 | + |
| 83 | + required_learners = ["ml_g", "ml_m"] |
| 84 | + for learner in self.dml_parameters["learners"]: |
| 85 | + for ml in required_learners: |
| 86 | + assert ml in learner, f"No {ml} specified in the config file" |
| 87 | + |
| 88 | + def _calculate_oracle_values(self): |
| 89 | + """Calculate oracle values for the simulation.""" |
| 90 | + self.logger.info("Calculating oracle values") |
| 91 | + |
| 92 | + self.oracle_values = dict() |
| 93 | + # Oracle values |
| 94 | + df_oracle = make_did_CS2021( |
| 95 | + n_obs=int(1e6), dgp_type=1 |
| 96 | + ) # does not depend on the DGP type |
| 97 | + df_oracle["ite"] = df_oracle["y1"] - df_oracle["y0"] |
| 98 | + self.oracle_values["detailed"] = ( |
| 99 | + df_oracle.groupby(["d", "t"])["ite"].mean().reset_index() |
| 100 | + ) |
| 101 | + |
| 102 | + # Oracle group aggregation |
| 103 | + df_oracle_post_treatment = df_oracle[df_oracle["t"] >= df_oracle["d"]] |
| 104 | + self.oracle_values["group"] = df_oracle_post_treatment.groupby("d")[ |
| 105 | + "ite" |
| 106 | + ].mean() |
| 107 | + |
| 108 | + # Oracle time aggregation |
| 109 | + self.oracle_values["time"] = df_oracle_post_treatment.groupby("t")["ite"].mean() |
| 110 | + |
| 111 | + # Oracle eventstudy aggregation |
| 112 | + df_oracle["e"] = pd.to_datetime(df_oracle["t"]).values.astype( |
| 113 | + "datetime64[M]" |
| 114 | + ) - pd.to_datetime(df_oracle["d"]).values.astype("datetime64[M]") |
| 115 | + self.oracle_values["eventstudy"] = df_oracle.groupby("e")["ite"].mean()[1:] |
| 116 | + |
| 117 | + def run_single_rep(self, dml_data, dml_params) -> Dict[str, Any]: |
| 118 | + """Run a single repetition with the given parameters.""" |
| 119 | + # Extract parameters |
| 120 | + learner_config = dml_params["learners"] |
| 121 | + learner_g_name, ml_g = create_learner_from_config(learner_config["ml_g"]) |
| 122 | + learner_m_name, ml_m = create_learner_from_config(learner_config["ml_m"]) |
| 123 | + score = dml_params["score"] |
| 124 | + in_sample_normalization = dml_params["in_sample_normalization"] |
| 125 | + |
| 126 | + # Model |
| 127 | + dml_model = dml.did.DoubleMLDIDMulti( |
| 128 | + obj_dml_data=dml_data, |
| 129 | + ml_g=ml_g, |
| 130 | + ml_m=None if score == "experimental" else ml_m, |
| 131 | + gt_combinations="standard", |
| 132 | + score=score, |
| 133 | + in_sample_normalization=in_sample_normalization, |
| 134 | + ) |
| 135 | + # Tuning |
| 136 | + dml_model_tuned = dml.did.DoubleMLDIDMulti( |
| 137 | + obj_dml_data=dml_data, |
| 138 | + ml_g=ml_g, |
| 139 | + ml_m=None if score == "experimental" else ml_m, |
| 140 | + gt_combinations="standard", |
| 141 | + score=score, |
| 142 | + in_sample_normalization=in_sample_normalization, |
| 143 | + ) |
| 144 | + dml_model_tuned.tune_ml_models( |
| 145 | + ml_param_space=self._param_space, |
| 146 | + optuna_settings=self._optuna_settings, |
| 147 | + ) |
| 148 | + |
| 149 | + # sort out oracle thetas |
| 150 | + oracle_thetas = np.full(len(dml_model.gt_combinations), np.nan) |
| 151 | + for i, (g, _, t) in enumerate(dml_model.gt_combinations): |
| 152 | + group_index = self.oracle_values["detailed"]["d"] == g |
| 153 | + time_index = self.oracle_values["detailed"]["t"] == t |
| 154 | + oracle_thetas[i] = self.oracle_values["detailed"][group_index & time_index][ |
| 155 | + "ite" |
| 156 | + ].iloc[0] |
| 157 | + |
| 158 | + result = { |
| 159 | + "detailed": [], |
| 160 | + "group": [], |
| 161 | + "time": [], |
| 162 | + "eventstudy": [], |
| 163 | + } |
| 164 | + for model in [dml_model, dml_model_tuned]: |
| 165 | + model.fit() |
| 166 | + model.bootstrap(n_rep_boot=2000) |
| 167 | + for level in self.confidence_parameters["level"]: |
| 168 | + level_result = dict() |
| 169 | + level_result["detailed"] = self._compute_coverage( |
| 170 | + thetas=model.coef, |
| 171 | + oracle_thetas=oracle_thetas, |
| 172 | + confint=model.confint(level=level), |
| 173 | + joint_confint=model.confint(level=level, joint=True), |
| 174 | + ) |
| 175 | + |
| 176 | + for aggregation_method in ["group", "time", "eventstudy"]: |
| 177 | + agg_obj = model.aggregate(aggregation=aggregation_method) |
| 178 | + agg_obj.aggregated_frameworks.bootstrap(n_rep_boot=2000) |
| 179 | + |
| 180 | + level_result[aggregation_method] = self._compute_coverage( |
| 181 | + thetas=agg_obj.aggregated_frameworks.thetas, |
| 182 | + oracle_thetas=self.oracle_values[aggregation_method].values, |
| 183 | + confint=agg_obj.aggregated_frameworks.confint(level=level), |
| 184 | + joint_confint=agg_obj.aggregated_frameworks.confint( |
| 185 | + level=level, joint=True |
| 186 | + ), |
| 187 | + ) |
| 188 | + |
| 189 | + # add parameters to the result |
| 190 | + for res in level_result.values(): |
| 191 | + res.update( |
| 192 | + { |
| 193 | + "Learner g": learner_g_name, |
| 194 | + "Learner m": learner_m_name, |
| 195 | + "Score": score, |
| 196 | + "In-sample-norm.": in_sample_normalization, |
| 197 | + "level": level, |
| 198 | + "Tuned": model is dml_model_tuned, |
| 199 | + } |
| 200 | + ) |
| 201 | + for key, res in level_result.items(): |
| 202 | + result[key].append(res) |
| 203 | + |
| 204 | + return result |
| 205 | + |
| 206 | + def summarize_results(self): |
| 207 | + """Summarize the simulation results.""" |
| 208 | + self.logger.info("Summarizing simulation results") |
| 209 | + |
| 210 | + groupby_cols = [ |
| 211 | + "Learner g", |
| 212 | + "Learner m", |
| 213 | + "Score", |
| 214 | + "In-sample-norm.", |
| 215 | + "DGP", |
| 216 | + "level", |
| 217 | + "Tuned", |
| 218 | + ] |
| 219 | + aggregation_dict = { |
| 220 | + "Coverage": "mean", |
| 221 | + "CI Length": "mean", |
| 222 | + "Bias": "mean", |
| 223 | + "Uniform Coverage": "mean", |
| 224 | + "Uniform CI Length": "mean", |
| 225 | + "repetition": "count", |
| 226 | + } |
| 227 | + |
| 228 | + result_summary = dict() |
| 229 | + for result_name, result_df in self.results.items(): |
| 230 | + result_summary[result_name] = ( |
| 231 | + result_df.groupby(groupby_cols).agg(aggregation_dict).reset_index() |
| 232 | + ) |
| 233 | + self.logger.debug(f"Summarized {result_name} results") |
| 234 | + |
| 235 | + return result_summary |
| 236 | + |
| 237 | + def _generate_dml_data(self, dgp_params) -> dml.data.DoubleMLPanelData: |
| 238 | + """Generate data for the simulation.""" |
| 239 | + data = make_did_CS2021(n_obs=dgp_params["n_obs"], dgp_type=dgp_params["DGP"]) |
| 240 | + dml_data = dml.data.DoubleMLPanelData( |
| 241 | + data, |
| 242 | + y_col="y", |
| 243 | + d_cols="d", |
| 244 | + id_col="id", |
| 245 | + t_col="t", |
| 246 | + x_cols=["Z1", "Z2", "Z3", "Z4"], |
| 247 | + ) |
| 248 | + return dml_data |
0 commit comments