|
| 1 | +from dataclasses import dataclass |
| 2 | +import typing as tp |
| 3 | + |
| 4 | +from jaxtyping import Float |
| 5 | +import numpy as np |
| 6 | +import pandas as pd |
| 7 | +from pandera import ( |
| 8 | + Check, |
| 9 | + Column, |
| 10 | + DataFrameSchema, |
| 11 | +) |
| 12 | +from rich import box |
| 13 | +from rich.progress import ( |
| 14 | + Progress, |
| 15 | + ProgressBar, |
| 16 | + track, |
| 17 | +) |
| 18 | + |
| 19 | +from causal_validation.validation.placebo import PlaceboTest |
| 20 | +from causal_validation.validation.testing import ( |
| 21 | + RMSPETestStatistic, |
| 22 | + TestResult, |
| 23 | + TestResultFrame, |
| 24 | +) |
| 25 | + |
| 26 | +RMSPESchema = DataFrameSchema( |
| 27 | + { |
| 28 | + "Model": Column(str), |
| 29 | + "Dataset": Column(str), |
| 30 | + "Test statistic": Column(float, coerce=True), |
| 31 | + "p-value": Column( |
| 32 | + float, |
| 33 | + checks=[ |
| 34 | + Check.greater_than_or_equal_to(0.0), |
| 35 | + Check.less_than_or_equal_to(1.0), |
| 36 | + ], |
| 37 | + coerce=True, |
| 38 | + ), |
| 39 | + } |
| 40 | +) |
| 41 | + |
| 42 | + |
| 43 | +@dataclass |
| 44 | +class RMSPETestResult(TestResultFrame): |
| 45 | + """ |
| 46 | + A subclass of TestResultFrame, RMSPETestResult stores test statistics and p-value |
| 47 | + for the treated unit. Test statistics for pseudo treatment units are also stored. |
| 48 | + """ |
| 49 | + |
| 50 | + treatment_test_results: tp.Dict[tp.Tuple[str, str], TestResult] |
| 51 | + pseudo_treatment_test_statistics: tp.Dict[tp.Tuple[str, str], tp.List[Float]] |
| 52 | + |
| 53 | + def to_df(self) -> pd.DataFrame: |
| 54 | + dfs = [] |
| 55 | + for (model, dataset), test_results in self.treatment_test_results.items(): |
| 56 | + result = { |
| 57 | + "Model": model, |
| 58 | + "Dataset": dataset, |
| 59 | + "Test statistic": test_results.test_statistic, |
| 60 | + "p-value": test_results.p_value, |
| 61 | + } |
| 62 | + df = pd.DataFrame([result]) |
| 63 | + dfs.append(df) |
| 64 | + df = pd.concat(dfs) |
| 65 | + RMSPESchema.validate(df) |
| 66 | + return df |
| 67 | + |
| 68 | + |
| 69 | +@dataclass |
| 70 | +class RMSPETest(PlaceboTest): |
| 71 | + """ |
| 72 | + A subclass of PlaceboTest calculates RMSPE as test statistic for all units. |
| 73 | + Given the RMSPE test stats, p-value for actual treatment is calculated. |
| 74 | + """ |
| 75 | + |
| 76 | + def execute(self, verbose: bool = True) -> RMSPETestResult: |
| 77 | + treatment_results, pseudo_treatment_results = {}, {} |
| 78 | + datasets = self.dataset_dict |
| 79 | + n_datasets = len(datasets) |
| 80 | + n_control = sum([d.n_units for d in datasets.values()]) |
| 81 | + rmspe = RMSPETestStatistic() |
| 82 | + with Progress(disable=not verbose) as progress: |
| 83 | + model_task = progress.add_task( |
| 84 | + "[red]Models", total=len(self.models), visible=verbose |
| 85 | + ) |
| 86 | + data_task = progress.add_task( |
| 87 | + "[blue]Datasets", total=n_datasets, visible=verbose |
| 88 | + ) |
| 89 | + unit_task = progress.add_task( |
| 90 | + f"[green]Treatment and Control Units", |
| 91 | + total=n_control + 1, |
| 92 | + visible=verbose, |
| 93 | + ) |
| 94 | + for data_name, dataset in datasets.items(): |
| 95 | + progress.update(data_task, advance=1) |
| 96 | + for model in self.models: |
| 97 | + progress.update(unit_task, advance=1) |
| 98 | + treatment_result = model(dataset) |
| 99 | + treatment_idx = dataset.ytr.shape[0] |
| 100 | + treatment_test_stat = rmspe( |
| 101 | + dataset, |
| 102 | + treatment_result.counterfactual, |
| 103 | + treatment_result.synthetic, |
| 104 | + treatment_idx, |
| 105 | + ) |
| 106 | + progress.update(model_task, advance=1) |
| 107 | + placebo_test_stats = [] |
| 108 | + for i in range(dataset.n_units): |
| 109 | + progress.update(unit_task, advance=1) |
| 110 | + placebo_data = dataset.to_placebo_data(i) |
| 111 | + result = model(placebo_data) |
| 112 | + placebo_test_stats.append( |
| 113 | + rmspe( |
| 114 | + placebo_data, |
| 115 | + result.counterfactual, |
| 116 | + result.synthetic, |
| 117 | + treatment_idx, |
| 118 | + ) |
| 119 | + ) |
| 120 | + pval_idx = 1 |
| 121 | + for p_stat in placebo_test_stats: |
| 122 | + pval_idx += 1 if treatment_test_stat < p_stat else 0 |
| 123 | + pval = pval_idx / (n_control + 1) |
| 124 | + treatment_results[(model._model_name, data_name)] = TestResult( |
| 125 | + p_value=pval, test_statistic=treatment_test_stat |
| 126 | + ) |
| 127 | + pseudo_treatment_results[(model._model_name, data_name)] = ( |
| 128 | + placebo_test_stats |
| 129 | + ) |
| 130 | + return RMSPETestResult( |
| 131 | + treatment_test_results=treatment_results, |
| 132 | + pseudo_treatment_test_statistics=pseudo_treatment_results, |
| 133 | + ) |
0 commit comments