|
| 1 | +from typing import Any, Dict, Optional |
| 2 | +import optuna |
| 3 | + |
| 4 | +import doubleml as dml |
| 5 | +import numpy as np |
| 6 | +import pandas as pd |
| 7 | +from doubleml.irm.datasets import make_irm_data_discrete_treatments |
| 8 | + |
| 9 | +from montecover.base import BaseSimulation |
| 10 | +from montecover.utils import create_learner_from_config |
| 11 | + |
| 12 | + |
| 13 | +class APOSTuningCoverageSimulation(BaseSimulation): |
| 14 | + """Simulation class for coverage properties of DoubleMLAPOs for APO estimation with 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 | + # Calculate oracle values |
| 31 | + self._calculate_oracle_values() |
| 32 | + |
| 33 | + # tuning specific settings |
| 34 | + # parameter space for the outcome regression tuning |
| 35 | + def ml_g_params(trial): |
| 36 | + return { |
| 37 | + 'n_estimators': trial.suggest_int('n_estimators', 100, 200, step=50), |
| 38 | + 'learning_rate': trial.suggest_float('learning_rate', 1e-3, 0.1, log=True), |
| 39 | + 'min_child_samples': trial.suggest_int('min_child_samples', 20, 50, step=5), |
| 40 | + 'max_depth': 5, |
| 41 | + 'lambda_l1': trial.suggest_float('lambda_l1', 1e-3, 10.0, log=True), |
| 42 | + 'lambda_l2': trial.suggest_float('lambda_l2', 1e-3, 10.0, log=True), |
| 43 | + } |
| 44 | + |
| 45 | + # parameter space for the propensity score tuning |
| 46 | + def ml_m_params(trial): |
| 47 | + return { |
| 48 | + 'n_estimators': trial.suggest_int('n_estimators', 100, 200, step=50), |
| 49 | + 'learning_rate': trial.suggest_float('learning_rate', 1e-3, 0.1, log=True), |
| 50 | + 'min_child_samples': trial.suggest_int('min_child_samples', 20, 50, step=5), |
| 51 | + 'max_depth': 5, |
| 52 | + 'lambda_l1': trial.suggest_float('lambda_l1', 1e-3, 10.0, log=True), |
| 53 | + 'lambda_l2': trial.suggest_float('lambda_l2', 1e-3, 10.0, log=True), |
| 54 | + } |
| 55 | + |
| 56 | + self._param_space = { |
| 57 | + 'ml_g': ml_g_params, |
| 58 | + 'ml_m': ml_m_params |
| 59 | + } |
| 60 | + |
| 61 | + self._optuna_settings = { |
| 62 | + 'n_trials': 200, |
| 63 | + 'show_progress_bar': False, |
| 64 | + 'verbosity': optuna.logging.WARNING, # Suppress Optuna logs |
| 65 | + } |
| 66 | + |
| 67 | + def _process_config_parameters(self): |
| 68 | + """Process simulation-specific parameters from config""" |
| 69 | + # Process ML models in parameter grid |
| 70 | + assert "learners" in self.dml_parameters, "No learners specified in the config file" |
| 71 | + |
| 72 | + required_learners = ["ml_g", "ml_m"] |
| 73 | + for learner in self.dml_parameters["learners"]: |
| 74 | + for ml in required_learners: |
| 75 | + assert ml in learner, f"No {ml} specified in the config file" |
| 76 | + |
| 77 | + def _calculate_oracle_values(self): |
| 78 | + """Calculate oracle values for the simulation.""" |
| 79 | + self.logger.info("Calculating oracle values") |
| 80 | + |
| 81 | + n_levels = self.dgp_parameters["n_levels"][0] |
| 82 | + data_apo_oracle = make_irm_data_discrete_treatments( |
| 83 | + n_obs=int(1e6), n_levels=n_levels, linear=self.dgp_parameters["linear"][0] |
| 84 | + ) |
| 85 | + |
| 86 | + y0 = data_apo_oracle["oracle_values"]["y0"] |
| 87 | + ite = data_apo_oracle["oracle_values"]["ite"] |
| 88 | + d = data_apo_oracle["d"] |
| 89 | + |
| 90 | + average_ites = np.full(n_levels + 1, np.nan) |
| 91 | + apos = np.full(n_levels + 1, np.nan) |
| 92 | + for i in range(n_levels + 1): |
| 93 | + average_ites[i] = np.mean(ite[d == i]) * (i > 0) |
| 94 | + apos[i] = np.mean(y0) + average_ites[i] |
| 95 | + |
| 96 | + ates = np.full(n_levels, np.nan) |
| 97 | + for i in range(n_levels): |
| 98 | + ates[i] = apos[i + 1] - apos[0] |
| 99 | + |
| 100 | + self.logger.info(f"Levels and their counts:\n{np.unique(d, return_counts=True)}") |
| 101 | + self.logger.info(f"True APOs: {apos}") |
| 102 | + self.logger.info(f"True ATEs: {ates}") |
| 103 | + |
| 104 | + self.oracle_values = dict() |
| 105 | + self.oracle_values["apos"] = apos |
| 106 | + self.oracle_values["ates"] = ates |
| 107 | + |
| 108 | + def run_single_rep(self, dml_data: dml.DoubleMLData, dml_params: Dict[str, Any]) -> Dict[str, Any]: |
| 109 | + """Run a single repetition with the given parameters.""" |
| 110 | + # Extract parameters |
| 111 | + learner_config = dml_params["learners"] |
| 112 | + learner_g_name, ml_g = create_learner_from_config(learner_config["ml_g"]) |
| 113 | + learner_m_name, ml_m = create_learner_from_config(learner_config["ml_m"]) |
| 114 | + treatment_levels = dml_params["treatment_levels"] |
| 115 | + trimming_threshold = dml_params["trimming_threshold"] |
| 116 | + |
| 117 | + # Model |
| 118 | + dml_model = dml.DoubleMLAPOS( |
| 119 | + obj_dml_data=dml_data, |
| 120 | + ml_g=ml_g, |
| 121 | + ml_m=ml_m, |
| 122 | + treatment_levels=treatment_levels, |
| 123 | + trimming_threshold=trimming_threshold, |
| 124 | + ) |
| 125 | + # Tuning |
| 126 | + dml_model_tuned = dml.DoubleMLAPOS( |
| 127 | + obj_dml_data=dml_data, |
| 128 | + ml_g=ml_g, |
| 129 | + ml_m=ml_m, |
| 130 | + treatment_levels=treatment_levels, |
| 131 | + trimming_threshold=trimming_threshold, |
| 132 | + ) |
| 133 | + dml_model_tuned.tune_ml_models( |
| 134 | + ml_param_space=self._param_space, |
| 135 | + optuna_settings=self._optuna_settings, |
| 136 | + ) |
| 137 | + |
| 138 | + result = { |
| 139 | + "coverage": [], |
| 140 | + "causal_contrast": [], |
| 141 | + } |
| 142 | + for model in [dml_model, dml_model_tuned]: |
| 143 | + model.fit() |
| 144 | + model.bootstrap(n_rep_boot=2000) |
| 145 | + causal_contrast_model = model.causal_contrast(reference_levels=0) |
| 146 | + causal_contrast_model.bootstrap(n_rep_boot=2000) |
| 147 | + for level in self.confidence_parameters["level"]: |
| 148 | + level_result = dict() |
| 149 | + level_result["coverage"] = self._compute_coverage( |
| 150 | + thetas=model.coef, |
| 151 | + oracle_thetas=self.oracle_values["apos"], |
| 152 | + confint=model.confint(level=level), |
| 153 | + joint_confint=model.confint(level=level, joint=True), |
| 154 | + ) |
| 155 | + level_result["causal_contrast"] = self._compute_coverage( |
| 156 | + thetas=causal_contrast_model.thetas, |
| 157 | + oracle_thetas=self.oracle_values["ates"], |
| 158 | + confint=causal_contrast_model.confint(level=level), |
| 159 | + joint_confint=causal_contrast_model.confint(level=level, joint=True), |
| 160 | + ) |
| 161 | + |
| 162 | + # add parameters to the result |
| 163 | + for res_metric in level_result.values(): |
| 164 | + res_metric.update( |
| 165 | + { |
| 166 | + "Learner g": learner_g_name, |
| 167 | + "Learner m": learner_m_name, |
| 168 | + "level": level, |
| 169 | + "Tuned": model is dml_model_tuned, |
| 170 | + } |
| 171 | + ) |
| 172 | + for key, res in level_result.items(): |
| 173 | + result[key].append(res) |
| 174 | + |
| 175 | + return result |
| 176 | + |
| 177 | + def summarize_results(self): |
| 178 | + """Summarize the simulation results.""" |
| 179 | + self.logger.info("Summarizing simulation results") |
| 180 | + |
| 181 | + # Group by parameter combinations |
| 182 | + groupby_cols = ["Learner g", "Learner m", "level", "Tuned"] |
| 183 | + aggregation_dict = { |
| 184 | + "Coverage": "mean", |
| 185 | + "CI Length": "mean", |
| 186 | + "Bias": "mean", |
| 187 | + "Uniform Coverage": "mean", |
| 188 | + "Uniform CI Length": "mean", |
| 189 | + "repetition": "count", |
| 190 | + } |
| 191 | + |
| 192 | + # Aggregate results (possibly multiple result dfs) |
| 193 | + result_summary = dict() |
| 194 | + for result_name, result_df in self.results.items(): |
| 195 | + result_summary[result_name] = result_df.groupby(groupby_cols).agg(aggregation_dict).reset_index() |
| 196 | + self.logger.debug(f"Summarized {result_name} results") |
| 197 | + |
| 198 | + return result_summary |
| 199 | + |
| 200 | + def _generate_dml_data(self, dgp_params: Dict[str, Any]) -> dml.DoubleMLData: |
| 201 | + """Generate data for the simulation.""" |
| 202 | + data = make_irm_data_discrete_treatments( |
| 203 | + n_obs=dgp_params["n_obs"], |
| 204 | + n_levels=dgp_params["n_levels"], |
| 205 | + linear=dgp_params["linear"], |
| 206 | + ) |
| 207 | + df_apo = pd.DataFrame( |
| 208 | + np.column_stack((data["y"], data["d"], data["x"])), |
| 209 | + columns=["y", "d"] + ["x" + str(i) for i in range(data["x"].shape[1])], |
| 210 | + ) |
| 211 | + dml_data = dml.DoubleMLData(df_apo, "y", "d") |
| 212 | + return dml_data |
0 commit comments