Donor-impute CPS clone features

Author: MaxGhenis

changelog.d/impute-cps-clone-features.changed.md

@@ -0,0 +1 @@
+Donor-impute race, Hispanic status, sex, and occupation-based CPS features onto the PUF clone half of the extended CPS so subgroup and overtime-eligibility inputs better align with PUF-imputed incomes.

policyengine_us_data/datasets/cps/extended_cps.py

@@ -22,15 +22,78 @@
 
 logger = logging.getLogger(__name__)
 
-
 def _supports_structural_mortgage_inputs() -> bool:
     return has_policyengine_us_variables(*STRUCTURAL_MORTGAGE_VARIABLES)
 
+# CPS-only categorical features to donor-impute onto the PUF clone half.
+# These drive subgroup analysis and occupation-based logic, so naive donor
+# duplication dilutes the relationship between the clone's PUF-imputed
+# income and its CPS-side demographic/occupation labels.
+CPS_CLONE_FEATURE_VARIABLES = [
+    "is_male",
+    "cps_race",
+    "is_hispanic",
+    "detailed_occupation_recode",
+]
+
+# Predictors used to rematch CPS features onto the PUF clone half.
+# These are all available on the CPS half and on the doubled extended CPS.
+CPS_CLONE_FEATURE_PREDICTORS = [
+    "age",
+    "state_fips",
+    "tax_unit_is_joint",
+    "tax_unit_count_dependents",
+    "is_tax_unit_head",
+    "is_tax_unit_spouse",
+    "is_tax_unit_dependent",
+    "employment_income",
+    "self_employment_income",
+    "social_security",
+]
+
+_OVERTIME_OCCUPATION_CODES = {
+    "has_never_worked": 53,
+    "is_military": 52,
+    "is_computer_scientist": 8,
+    "is_farmer_fisher": 41,
+}
+_EXECUTIVE_ADMINISTRATIVE_PROFESSIONAL_CODES = np.array(
+    [
+        1,
+        2,
+        3,
+        5,
+        7,
+        9,
+        10,
+        11,
+        12,
+        13,
+        14,
+        15,
+        16,
+        18,
+        19,
+        20,
+        21,
+        22,
+        24,
+        25,
+        27,
+        28,
+        29,
+        30,
+        32,
+        33,
+        34,
+    ],
+    dtype=np.int16,
+)
 
 # CPS-only variables that should be QRF-imputed for the PUF clone half
-# instead of naively duplicated from the CPS donor. These are
-# income-correlated variables that exist only in the CPS; demographics,
-# IDs, weights, and random seeds are fine to duplicate.
+# instead of naively duplicated from the CPS donor. Most demographics,
+# IDs, weights, and random seeds are fine to duplicate; the categorical
+# clone features above are rematched separately.
 CPS_ONLY_IMPUTED_VARIABLES = [
     # Retirement distributions
     "taxable_401k_distributions",
@@ -109,6 +172,186 @@ def _supports_structural_mortgage_inputs() -> bool:
 ]
 
 
+def _clone_half_person_values(data: dict, variable: str, time_period: int):
+    """Return clone-half values for ``variable`` mapped to person rows."""
+    if variable not in data:
+        return None
+
+    values = data[variable][time_period]
+    n_persons = len(data["person_id"][time_period])
+    n_persons_half = n_persons // 2
+    if len(values) == n_persons:
+        return np.asarray(values[n_persons_half:])
+
+    entity_mappings = [
+        ("household_id", "person_household_id"),
+        ("tax_unit_id", "person_tax_unit_id"),
+        ("spm_unit_id", "person_spm_unit_id"),
+        ("family_id", "person_family_id"),
+    ]
+    for entity_id_var, person_entity_id_var in entity_mappings:
+        if entity_id_var not in data or person_entity_id_var not in data:
+            continue
+        entity_ids = data[entity_id_var][time_period]
+        if len(values) != len(entity_ids):
+            continue
+        entity_half = len(entity_ids) // 2
+        clone_entity_ids = entity_ids[entity_half:]
+        clone_person_entity_ids = data[person_entity_id_var][time_period][n_persons_half:]
+        value_map = dict(zip(clone_entity_ids, values[entity_half:]))
+        return np.array([value_map[idx] for idx in clone_person_entity_ids])
+
+    return None
+
+
+def _build_clone_test_frame(
+    cps_sim,
+    data: dict,
+    time_period: int,
+    predictors: list[str],
+) -> pd.DataFrame:
+    """Build clone-half predictor data with available doubled-dataset overrides."""
+    X_test = cps_sim.calculate_dataframe(predictors).copy()
+    for predictor in predictors:
+        clone_values = _clone_half_person_values(data, predictor, time_period)
+        if clone_values is not None and len(clone_values) == len(X_test):
+            X_test[predictor] = clone_values
+    return X_test[predictors]
+
+
+def _prepare_knn_matrix(
+    df: pd.DataFrame,
+    reference: pd.DataFrame | None = None,
+) -> np.ndarray:
+    """Normalise mixed-scale donor-matching predictors for kNN."""
+    X = df.astype(float).copy()
+    for income_var in CPS_STAGE2_INCOME_PREDICTORS:
+        if income_var in X:
+            X[income_var] = np.arcsinh(X[income_var])
+
+    ref = X if reference is None else reference.astype(float).copy()
+    for income_var in CPS_STAGE2_INCOME_PREDICTORS:
+        if income_var in ref:
+            ref[income_var] = np.arcsinh(ref[income_var])
+
+    means = ref.mean()
+    stds = ref.std(ddof=0).replace(0, 1)
+    normalised = (X - means) / stds
+    return np.nan_to_num(normalised.to_numpy(dtype=np.float32), nan=0.0)
+
+
+def _derive_overtime_occupation_inputs(
+    occupation_codes: np.ndarray,
+) -> pd.DataFrame:
+    """Derive occupation-based overtime-exemption inputs from POCCU2."""
+    occupation_codes = np.rint(occupation_codes).astype(np.int16, copy=False)
+    derived = {
+        name: occupation_codes == code
+        for name, code in _OVERTIME_OCCUPATION_CODES.items()
+    }
+    derived["is_executive_administrative_professional"] = np.isin(
+        occupation_codes,
+        _EXECUTIVE_ADMINISTRATIVE_PROFESSIONAL_CODES,
+    )
+    return pd.DataFrame(derived)
+
+
+def _impute_clone_cps_features(
+    data: dict,
+    time_period: int,
+    dataset_path: str,
+) -> pd.DataFrame:
+    """Rematch CPS demographic/occupation features for the clone half."""
+    from policyengine_us import Microsimulation
+    from sklearn.neighbors import NearestNeighbors
+
+    cps_sim = Microsimulation(dataset=dataset_path)
+    X_train = cps_sim.calculate_dataframe(
+        CPS_CLONE_FEATURE_PREDICTORS + CPS_CLONE_FEATURE_VARIABLES
+    )
+    available_outputs = [
+        variable for variable in CPS_CLONE_FEATURE_VARIABLES if variable in X_train.columns
+    ]
+    if not available_outputs:
+        n_half = len(data["person_id"][time_period]) // 2
+        return pd.DataFrame(index=np.arange(n_half))
+
+    X_test = _build_clone_test_frame(
+        cps_sim,
+        data,
+        time_period,
+        CPS_CLONE_FEATURE_PREDICTORS,
+    )
+    del cps_sim
+
+    train_roles = (
+        X_train[["is_tax_unit_head", "is_tax_unit_spouse", "is_tax_unit_dependent"]]
+        .round()
+        .astype(int)
+        .apply(tuple, axis=1)
+    )
+    test_roles = (
+        X_test[["is_tax_unit_head", "is_tax_unit_spouse", "is_tax_unit_dependent"]]
+        .round()
+        .astype(int)
+        .apply(tuple, axis=1)
+    )
+
+    predictions = pd.DataFrame(index=X_test.index, columns=available_outputs)
+    for role in test_roles.unique():
+        test_mask = test_roles == role
+        train_mask = train_roles == role
+        if not train_mask.any():
+            train_mask = pd.Series(True, index=X_train.index)
+
+        train_predictors = X_train.loc[train_mask, CPS_CLONE_FEATURE_PREDICTORS]
+        test_predictors = X_test.loc[test_mask, CPS_CLONE_FEATURE_PREDICTORS]
+        train_matrix = _prepare_knn_matrix(train_predictors)
+        test_matrix = _prepare_knn_matrix(test_predictors, reference=train_predictors)
+
+        matcher = NearestNeighbors(n_neighbors=1)
+        matcher.fit(train_matrix)
+        donor_indices = matcher.kneighbors(
+            test_matrix,
+            return_distance=False,
+        ).ravel()
+        donor_outputs = (
+            X_train.loc[train_mask, available_outputs]
+            .iloc[donor_indices]
+            .reset_index(drop=True)
+        )
+        predictions.loc[test_mask, available_outputs] = donor_outputs.to_numpy()
+
+    if "detailed_occupation_recode" in predictions:
+        occupation_codes = predictions["detailed_occupation_recode"].astype(float).to_numpy()
+        for column, values in _derive_overtime_occupation_inputs(occupation_codes).items():
+            predictions[column] = values
+
+    return predictions
+
+
+def _splice_clone_feature_predictions(
+    data: dict,
+    predictions: pd.DataFrame,
+    time_period: int,
+) -> dict:
+    """Replace clone-half person-level feature variables with donor matches."""
+    n_half = len(data["person_id"][time_period]) // 2
+    for variable in predictions.columns:
+        if variable not in data:
+            continue
+        values = data[variable][time_period]
+        new_values = np.array(values, copy=True)
+        pred_values = predictions[variable].to_numpy()
+        if np.issubdtype(new_values.dtype, np.bool_):
+            pred_values = pred_values.astype(bool, copy=False)
+        else:
+            pred_values = pred_values.astype(new_values.dtype, copy=False)
+        new_values[n_half:] = pred_values
+        data[variable] = {time_period: new_values}
+    return data
+
+
 def _impute_cps_only_variables(
     data: dict,
     time_period: int,
@@ -172,17 +415,16 @@ def _impute_cps_only_variables(
             missing_outputs,
         )
 
-    # Build PUF clone test data: demographics from CPS sim (PUF clones
-    # share demographics with their CPS donors), income from the
-    # PUF-imputed values in the second half of the doubled data.
-    n_persons_half = len(data["person_id"][time_period]) // 2
-    X_test = cps_sim.calculate_dataframe(CPS_STAGE2_DEMOGRAPHIC_PREDICTORS)
+    # Build PUF clone test data from the clone half itself, falling back to
+    # the CPS sim for formula variables that are not stored in the dataset.
+    X_test = _build_clone_test_frame(
+        cps_sim,
+        data,
+        time_period,
+        all_predictors,
+    )
     del cps_sim
 
-    for var in CPS_STAGE2_INCOME_PREDICTORS:
-        # Income comes from PUF imputation in the second half.
-        X_test[var] = data[var][time_period][n_persons_half:]
-
     logger.info(
         "Stage-2 CPS-only imputation: %d outputs, "
         "training on %d CPS persons, predicting for %d PUF clones",
@@ -438,11 +680,24 @@ def generate(self):
             dataset_path=str(self.cps.file_path),
         )
 
-        # Stage 2: QRF-impute CPS-only variables for PUF clones.
+        # Stage 2a: donor-impute CPS feature variables for PUF clones.
+        logger.info("Stage-2a: rematching CPS features for PUF clones")
+        clone_feature_predictions = _impute_clone_cps_features(
+            data=new_data,
+            time_period=self.time_period,
+            dataset_path=str(self.cps.file_path),
+        )
+        new_data = _splice_clone_feature_predictions(
+            data=new_data,
+            predictions=clone_feature_predictions,
+            time_period=self.time_period,
+        )
+
+        # Stage 2b: QRF-impute CPS-only continuous variables for PUF clones.
         # Train on CPS data using demographics + PUF-imputed income
         # as predictors, so the PUF clone half gets values consistent
         # with its imputed income rather than naive donor duplication.
-        logger.info("Stage-2: imputing CPS-only variables for PUF clones")
+        logger.info("Stage-2b: imputing CPS-only variables for PUF clones")
         cps_only_predictions = _impute_cps_only_variables(
             data=new_data,
             time_period=self.time_period,