diff --git a/.gitignore b/.gitignore index 70d8396d..bded48fc 100644 --- a/.gitignore +++ b/.gitignore @@ -26,3 +26,6 @@ policyengine_uk_data/storage/local_geography_weights.csv.gz # Hypothesis example database (property-based tests) .hypothesis/ + +# Workspace-local hub job notes and tooling +.job/ diff --git a/changelog.d/frs-reported-dividends.fixed.md b/changelog.d/frs-reported-dividends.fixed.md new file mode 100644 index 00000000..0b49741f --- /dev/null +++ b/changelog.d/frs-reported-dividends.fixed.md @@ -0,0 +1,2 @@ +- FRS investment-account dividends are now summed to the person who holds the account. They were keyed on each person's row position, so almost none matched and the base FRS carried about £40m of dividends instead of about £8bn in 2024-25. +- The enhanced FRS keeps those reported dividends instead of replacing them with draws from the SPI income model, which predicts from age, gender and region within earnings groups and gave Universal Credit claimants six-figure dividends (policyengine-uk#1948). The SPI-donor half still carries the SPI income distribution. diff --git a/policyengine_uk_data/datasets/frs.py b/policyengine_uk_data/datasets/frs.py index 711127b5..fae6595b 100644 --- a/policyengine_uk_data/datasets/frs.py +++ b/policyengine_uk_data/datasets/frs.py @@ -48,6 +48,8 @@ # loaders (e.g. LCFS/ETB in `datasets/imputations/consumption.py`) can import # the same value rather than re-defining `* 52` locally and drifting. WEEKS_IN_YEAR = 365.25 / 7 +# Grosses up an amount taxed at source at the basic rate (1 / 0.8). +INVERTED_BASIC_RATE = 1.25 LEGACY_JOBSEEKER_MIN_AGE = 18 HOURS_WORKED_WEEKS_PER_YEAR = 52 @@ -790,6 +792,28 @@ def validate_frs_survey_year(raw_frs_folder, year: int) -> None: ) +def frs_dividend_income(account: pd.DataFrame, person_ids) -> np.ndarray: + """Annual dividends each person reports on FRS investment accounts. + + Gilt-edged stock taxed at source (account type 6), unit and investment + trusts (7) and stocks and shares (8). Amounts taxed at source are grossed + up at the basic rate. Summed to ``person_ids``, which must be the same + ``household_id * 1e3 + person`` keys as ``account.person_id``: keying on + the positional index instead matched almost no accounts, leaving the FRS + with about £40m of dividends rather than about £8bn in 2024-25. + """ + dividends = ( + account.accint * np.where(account.invtax == 1, INVERTED_BASIC_RATE, 1) + ) * ( + ((account.account == 6) & (account.invtax == 1)) # GGES + | account.account.isin((7, 8)) # Stocks/shares/UITs + ) + return np.maximum( + 0, + sum_to_entity(dividends, account.person_id, person_ids) * WEEKS_IN_YEAR, + ) + + def derive_pension_credit_reported_capital(benunit: pd.DataFrame) -> np.ndarray: """Each benefit unit's capital as the FRS records it, for Pension Credit. @@ -1267,8 +1291,6 @@ def determine_education_level(fted_val, typeed2_val, age_val): * 52 ).astype(float) - WEEKS_IN_YEAR = 365.25 / 7 - pe_person["employment_income"] = np.maximum(0, person.inearns) * WEEKS_IN_YEAR pension_payment = sum_to_entity( @@ -1303,8 +1325,6 @@ def determine_education_level(fted_val, typeed2_val, age_val): ) ) - INVERTED_BASIC_RATE = 1.25 - pe_person["tax_free_savings_income"] = np.maximum( 0, sum_to_entity( @@ -1327,19 +1347,7 @@ def determine_education_level(fted_val, typeed2_val, age_val): 0, taxable_savings_interest + pe_person["tax_free_savings_income"].values, ) - pe_person["dividend_income"] = np.maximum( - 0, - sum_to_entity( - (account.accint * np.where(account.invtax == 1, INVERTED_BASIC_RATE, 1)) - * ( - ((account.account == 6) & (account.invtax == 1)) # GGES - | account.account.isin((7, 8)) # Stocks/shares/UITs - ), - account.person_id, - person.index, - ) - * 52, - ) + pe_person["dividend_income"] = frs_dividend_income(account, person.person_id) pe_person["property_income"] = frs_property_income(person, household) maintenance_to_self = np.maximum( pd.Series(np.where(person.mntus1 == 2, person.mntusam1, person.mntamt1)).fillna( diff --git a/policyengine_uk_data/datasets/imputations/income.py b/policyengine_uk_data/datasets/imputations/income.py index 799b5f35..1e63dd5b 100644 --- a/policyengine_uk_data/datasets/imputations/income.py +++ b/policyengine_uk_data/datasets/imputations/income.py @@ -450,8 +450,9 @@ def impute_over_incomes( # Housing costs (rent, mortgage interest, mortgage capital) used to be # rescaled here by new_income_total / original_income_total across - # INCOME_COMPONENTS. Because FRS dividend_income is near-zero and the - # SPI-trained QRF predicts materially larger dividends, the ratio + # INCOME_COMPONENTS. Because FRS dividend_income was then near-zero (a + # keying error in frs.py, since fixed) and the SPI-trained QRF predicts + # materially larger dividends, the ratio # inflated rent/mortgage by ~2.5× uniformly in the built enhanced FRS # — pushing AHC poverty rates 10–18 pp above HBAI for non-pensioners # (see issue #367). Housing costs now pass through unchanged; their @@ -488,13 +489,11 @@ def impute_income(dataset: UKSingleYearDataset) -> UKSingleYearDataset: Returns: Combined dataset with original data plus synthetic high-income individuals. """ - # Impute wealth, assuming same time period as trained data dataset = dataset.copy() # gift_aid and charitable_investment_gifts are in IMPUTATIONS but are not # columns on the raw FRS build, so initialise them to zero everywhere - # before imputation. Without this, the full-FRS half stays NaN for these - # columns (they're never touched by the dividend-only impute_over_incomes - # call below), and the eventual stacked dataset fails validate(). + # before imputation. The FRS half keeps its reported incomes, so these + # columns must already be populated when the stacked dataset is validated. for column in ("gift_aid", "charitable_investment_gifts"): if column not in dataset.person.columns: dataset.person[column] = 0.0 @@ -507,7 +506,7 @@ def impute_income(dataset: UKSingleYearDataset) -> UKSingleYearDataset: model = create_income_model() - # Impute just dividends on the original, full variable set on the copy + # Impute the full income set on the SPI-donor copy only. zero_weight_copy = impute_over_incomes( zero_weight_copy, @@ -550,11 +549,14 @@ def impute_income(dataset: UKSingleYearDataset) -> UKSingleYearDataset: ) ) - dataset = impute_over_incomes( - dataset, - model, - ["dividend_income"], - ) + # The FRS half keeps its reported dividends. Replacing them with a draw + # from the SPI model, which predicts from age, gender and region within + # earnings groups (whether the person has pay or a trade), gave + # dividends to FRS respondents without regard to their investments or + # benefits, so Universal Credit claimants received them + # (policyengine-uk#1948). The SPI-donor half carries the SPI dividend + # distribution, and calibration to the HMRC dividend targets reweights + # between the two. zero_weight_copy.validate() dataset.validate() diff --git a/policyengine_uk_data/tests/test_frs_dividend_income.py b/policyengine_uk_data/tests/test_frs_dividend_income.py new file mode 100644 index 00000000..7c3809d4 --- /dev/null +++ b/policyengine_uk_data/tests/test_frs_dividend_income.py @@ -0,0 +1,82 @@ +"""FRS dividends land on the person who holds the account (policyengine-uk#1948).""" + +import numpy as np +import pandas as pd +import pytest +from hypothesis import given, settings, strategies as st + +from policyengine_uk_data.datasets.frs import WEEKS_IN_YEAR, frs_dividend_income + + +def test_dividends_are_keyed_on_person_id_not_row_position(): + # Person ids are household_id * 1000 + person number, so they never + # coincide with the positional index the rows happen to sit at. + person_ids = pd.Series([1_001, 1_002, 2_001, 2_002]) + account = pd.DataFrame( + { + "person_id": [1_002, 2_001, 2_001, 2_002, 1_001, 2_002], + # 8: stocks and shares; 7: unit/investment trusts; 6: gilts; 1: bank account + "account": [8, 7, 6, 6, 1, 8], + "accint": [10.0, 4.0, 2.0, 3.0, 50.0, 0.0], + "invtax": [2, 1, 1, 2, 2, 2], + } + ) + dividends = frs_dividend_income(account, person_ids) + weekly = np.array( + [ + 0.0, # 1,001 has only a bank account + 10.0, # 1,002: shares, not taxed at source + 4.0 * 1.25 + 2.0 * 1.25, # 2,001: trusts and gilts taxed at source + 0.0, # 2,002: gilts not taxed at source are excluded; shares pay 0 + ] + ) + assert dividends == pytest.approx(weekly * WEEKS_IN_YEAR) + + +def test_dividends_are_never_negative(): + person_ids = pd.Series([1_001]) + account = pd.DataFrame( + {"person_id": [1_001], "account": [8], "accint": [-5.0], "invtax": [2]} + ) + assert frs_dividend_income(account, person_ids).tolist() == [0.0] + + +@st.composite +def dividend_account_cases(draw): + household_ids = draw( + st.lists(st.integers(1, 50), min_size=1, max_size=8, unique=True) + ) + person_ids = [household_id * 1_000 + 1 for household_id in household_ids] + rows = draw( + st.lists( + st.tuples( + st.sampled_from(person_ids), + st.sampled_from([1, 6, 7, 8, 9]), + st.integers(0, 1_000_000), + st.sampled_from([1, 2]), + ), + max_size=40, + ) + ) + account = pd.DataFrame(rows, columns=["person_id", "account", "pence", "invtax"]) + account["accint"] = account.pop("pence") / 100 + # Neither frame's index can be mistaken for an FRS person identifier. + account.index = np.arange(len(account)) * 7 + 100 + person_ids = pd.Series(person_ids[::-1], index=np.arange(len(person_ids)) * 3 + 10) + return account, person_ids + + +@given(dividend_account_cases()) +@settings(max_examples=100, deadline=None) +def test_eligible_gross_dividends_are_conserved_when_joining_to_people(case): + account, person_ids = case + expected_by_person = {person_id: 0.0 for person_id in person_ids} + for row in account.itertuples(): + if row.account in (7, 8) or (row.account == 6 and row.invtax == 1): + gross_weekly_amount = row.accint / (0.8 if row.invtax == 1 else 1) + expected_by_person[row.person_id] += gross_weekly_amount * WEEKS_IN_YEAR + + dividends = frs_dividend_income(account, person_ids) + + assert dividends.sum() == pytest.approx(sum(expected_by_person.values())) + assert dict(zip(person_ids, dividends)) == pytest.approx(expected_by_person) diff --git a/policyengine_uk_data/tests/test_imputation_source_flags.py b/policyengine_uk_data/tests/test_imputation_source_flags.py index 2df1f9d8..1a324d5e 100644 --- a/policyengine_uk_data/tests/test_imputation_source_flags.py +++ b/policyengine_uk_data/tests/test_imputation_source_flags.py @@ -4,6 +4,9 @@ import numpy as np import pandas as pd +import pytest +from hypothesis import given, settings, strategies as st +from policyengine_uk.data import UKSingleYearDataset class _FakeDataset: @@ -131,3 +134,238 @@ def test_impute_capital_gains_marks_capital_gains_clone_households(monkeypatch): True, ] assert result.household.loc[2:, "household_weight"].eq(1).all() + + +def test_impute_income_keeps_reported_dividends_on_the_frs_half(monkeypatch): + """SPI draws replace incomes only on the SPI-donor copy (policyengine-uk#1948). + + The SPI income model predicts from age, gender and region within earnings + groups, so using it to overwrite the FRS half's dividends gave Universal + Credit claimants dividends unrelated to any investments they reported. + The stage-2 model then trains on the FRS half's reported dividends. + """ + from policyengine_uk_data.datasets.imputations import income as income_module + from policyengine_uk_data.datasets import disability_benefits + from policyengine_uk_data.datasets.imputations import frs_only + + imputed_halves = [] + stage_2_training = [] + + def fake_impute_over_incomes(dataset, _model, output_variables): + imputed_halves.append( + ( + bool(dataset.household["household_is_spi_synthetic"].all()), + tuple(output_variables), + ) + ) + dataset = dataset.copy() + for column in output_variables: + dataset.person[column] = 123_456.0 + return dataset + + monkeypatch.setattr(income_module, "create_income_model", lambda: object()) + monkeypatch.setattr( + income_module, + "subsample_dataset", + lambda dataset, _sample_size: dataset.copy(), + ) + monkeypatch.setattr(income_module, "impute_over_incomes", fake_impute_over_incomes) + + def fake_impute_frs_only_variables(train_dataset, target_dataset): + stage_2_training.append(train_dataset.person["dividend_income"].tolist()) + return target_dataset + + monkeypatch.setattr( + frs_only, "impute_frs_only_variables", fake_impute_frs_only_variables + ) + monkeypatch.setattr( + disability_benefits, + "strip_internal_disability_reported_amounts", + lambda dataset: dataset, + ) + monkeypatch.setattr(income_module, "stack_datasets", _stack_without_remapping) + + dataset = _fake_dataset() + dataset.person["dividend_income"] = [150.0, 0.0] + result = income_module.impute_income(dataset) + + # Only the SPI-donor copy is imputed, with the full income set. + assert imputed_halves == [(True, tuple(income_module.IMPUTATIONS))] + frs_half = result.person.iloc[:2] + assert frs_half["dividend_income"].tolist() == [150.0, 0.0] + assert frs_half["employment_income"].tolist() == [20_000.0, 80_000.0] + assert result.person.iloc[2:]["dividend_income"].eq(123_456.0).all() + # Stage 2 learns from the FRS half's reported dividends, not SPI draws. + assert stage_2_training == [[150.0, 0.0]] + + +def _reported_dividend_dataset(dividends_by_person) -> UKSingleYearDataset: + person_ids = list(dividends_by_person) + household_ids = [person_id // 1_000 for person_id in person_ids] + person = pd.DataFrame( + { + "person_id": person_ids, + "person_household_id": household_ids, + "person_benunit_id": [ + household_id * 100 + 1 for household_id in household_ids + ], + "age": 40, + "gender": "FEMALE", + "employment_status": "FT_EMPLOYED", + "employment_income": 20_000.0, + "self_employment_income": 0.0, + "savings_interest_income": 0.0, + "dividend_income": list(dividends_by_person.values()), + "private_pension_income": 0.0, + "property_income": 0.0, + }, + index=np.arange(len(person_ids)) * 7 + 10, + ) + household_ids = list(dict.fromkeys(household_ids))[::-1] + return UKSingleYearDataset( + person=person, + benunit=pd.DataFrame( + {"benunit_id": [household_id * 100 + 1 for household_id in household_ids]}, + index=np.arange(len(household_ids)) * 3 + 50, + ), + household=pd.DataFrame( + { + "household_id": household_ids, + "household_weight": 2.0, + "region": "LONDON", + }, + index=np.arange(len(household_ids)) * 5 + 20, + ), + fiscal_year=2025, + ) + + +def _stub_income_training(monkeypatch): + """Avoid model training and simulation; retain income draws and real stacking.""" + from policyengine_uk_data.datasets.imputations import income as income_module + from policyengine_uk_data.datasets.imputations import frs_only + + class FakeSimulation: + def __init__(self, dataset): + self.dataset = dataset + + def calculate_dataframe(self, variables): + person = self.dataset.person + regions = self.dataset.household.set_index("household_id").region + return pd.DataFrame( + { + "age": person.age.to_numpy(), + "gender": person.gender.to_numpy(), + "region": person.person_household_id.map(regions).to_numpy(), + } + )[variables] + + class FakeIncomeModel: + def __init__(self): + self.predictions = [] + + def predict(self, inputs): + self.predictions.append(inputs.copy()) + return pd.DataFrame( + 123_456.0, index=inputs.index, columns=income_module.IMPUTATIONS + ) + + def reordered_donor_copy(dataset, _sample_size): + dataset = dataset.copy() + dataset.person = dataset.person.iloc[::-1].copy() + dataset.benunit = dataset.benunit.iloc[::-1].copy() + dataset.household = dataset.household.iloc[::-1].copy() + return dataset + + model = FakeIncomeModel() + monkeypatch.setattr(income_module, "Microsimulation", FakeSimulation) + monkeypatch.setattr(income_module, "create_income_model", lambda: model) + monkeypatch.setattr(income_module, "subsample_dataset", reordered_donor_copy) + monkeypatch.setattr( + frs_only, + "impute_frs_only_variables", + lambda train_dataset, target_dataset: target_dataset, + ) + return income_module, model + + +def _income_halves(dataset): + source = dataset.household.set_index("household_id").household_is_spi_synthetic + is_spi = dataset.person.person_household_id.map(source) + assert is_spi.notna().all() + return ( + dataset.person.loc[~is_spi].set_index("person_id"), + dataset.person.loc[is_spi].set_index("person_id"), + ) + + +def test_account_dividends_survive_imputation_and_real_stacking(monkeypatch): + from policyengine_uk_data.datasets.frs import WEEKS_IN_YEAR, frs_dividend_income + + income_module, model = _stub_income_training(monkeypatch) + dataset = _reported_dividend_dataset( + {2_001: 0.0, 1_002: 0.0, 1_001: 0.0, 2_002: 0.0} + ) + account = pd.DataFrame( + { + "person_id": [1_002, 2_001, 2_001, 1_001, 2_002], + "account": [8, 7, 6, 1, 6], + "accint": [10.0, 4.0, 2.0, 50.0, 3.0], + "invtax": [2, 1, 1, 2, 2], + }, + index=[43, 19, 71, 7, 28], + ) + dataset.person["dividend_income"] = frs_dividend_income( + account, dataset.person.person_id + ) + original_person = dataset.person.copy() + + result = income_module.impute_income(dataset) + frs_half, spi_half = _income_halves(result) + + assert frs_half.dividend_income.to_dict() == pytest.approx( + {2_001: 7.5 * WEEKS_IN_YEAR, 1_002: 10 * WEEKS_IN_YEAR, 1_001: 0, 2_002: 0} + ) + assert frs_half.employment_income.eq(20_000.0).all() + person_offset = original_person.person_id.max() + 1 + assert ( + spi_half.index.tolist() + == (original_person.person_id.iloc[::-1] + person_offset).tolist() + ) + assert spi_half.dividend_income.eq(123_456.0).all() + assert len(model.predictions) == 1 + assert len(model.predictions[0]) == len(original_person) + assert result.person.person_id.is_unique + assert result.benunit.benunit_id.is_unique + assert result.household.household_id.is_unique + assert result.person.person_benunit_id.isin(result.benunit.benunit_id).all() + pd.testing.assert_frame_equal(dataset.person, original_person) + result.validate() + + +@given( + st.dictionaries( + st.integers(1, 50).map(lambda household_id: household_id * 1_000 + 1), + st.integers(0, 1_000_000).map(lambda pence: pence / 100), + min_size=1, + max_size=8, + ) +) +@settings(max_examples=50, deadline=None) +def test_frs_dividend_totals_are_conserved_through_income_imputation( + reported_dividends, +): + # A fresh patch context per example avoids state leaking between property cases. + with pytest.MonkeyPatch.context() as monkeypatch: + income_module, _model = _stub_income_training(monkeypatch) + dataset = _reported_dividend_dataset(reported_dividends) + result = income_module.impute_income(dataset) + + frs_half, spi_half = _income_halves(result) + assert frs_half.dividend_income.sum() == pytest.approx( + sum(reported_dividends.values()) + ) + assert frs_half.dividend_income.to_dict() == pytest.approx(reported_dividends) + assert spi_half.dividend_income.sum() == pytest.approx( + len(reported_dividends) * 123_456 + ) diff --git a/policyengine_uk_data/tests/test_legacy_benefit_proxies.py b/policyengine_uk_data/tests/test_legacy_benefit_proxies.py index 64bf5670..54f23f4f 100644 --- a/policyengine_uk_data/tests/test_legacy_benefit_proxies.py +++ b/policyengine_uk_data/tests/test_legacy_benefit_proxies.py @@ -3,6 +3,7 @@ import pytest import policyengine_uk import policyengine_uk_data.datasets.frs as frs_module +from policyengine_uk_data.utils.datasets import sum_to_entity as real_sum_to_entity from policyengine_uk_data.datasets.frs import ( FRS_EMPSTATI_EMPLOYMENT_STATUS, @@ -382,12 +383,21 @@ def calculate(self, variable, year=None): raise KeyError(variable) -def create_single_adult_frs(tmp_path, monkeypatch, empstati=8, with_child=False): +def create_single_adult_frs( + tmp_path, monkeypatch, empstati=8, with_child=False, accounts=None +): monkeypatch.setattr(policyengine_uk, "Microsimulation", FakeMicrosimulation) monkeypatch.setattr(frs_module, "load_take_up_rate", lambda *args, **kwargs: 0.0) monkeypatch.setattr(frs_module, "load_parameter", lambda *args, **kwargs: 0.0) + # The real join for any table with rows; empty tables give zeros. monkeypatch.setattr( - frs_module, "sum_to_entity", lambda values, ids, index: np.zeros(len(index)) + frs_module, + "sum_to_entity", + lambda values, ids, index: ( + real_sum_to_entity(values, ids, index) + if len(values) + else np.zeros(len(index)) + ), ) monkeypatch.setattr( frs_module, @@ -521,8 +531,12 @@ def create_single_adult_frs(tmp_path, monkeypatch, empstati=8, with_child=False) ] ), "oddjob": pd.DataFrame(columns=["person", "sernum", "ojamt", "ojnow"]), - "accounts": pd.DataFrame( - columns=["person", "sernum", "accint", "acctax", "invtax", "account"] + "accounts": ( + accounts + if accounts is not None + else pd.DataFrame( + columns=["person", "sernum", "accint", "acctax", "invtax", "account"] + ) ), "job": pd.DataFrame(columns=["person", "sernum", "deduc1", "spnamt", "salsac"]), "benefits": pd.DataFrame( @@ -598,6 +612,32 @@ def test_create_frs_child_rows_are_child(tmp_path, monkeypatch): } +def test_create_frs_keys_account_dividends_on_person_id(tmp_path, monkeypatch): + # One stocks-and-shares account (type 8) held by the adult. create_frs + # must join it on person_id (household_id * 1000 + person), not on the + # person table's row position, which matches no account id. + accounts = pd.DataFrame( + [ + { + "person": 1, + "sernum": 100, + "accint": 10.0, + "acctax": 0, + "invtax": 0, + "account": 8, + } + ] + ) + person = create_single_adult_frs( + tmp_path, monkeypatch, with_child=True, accounts=accounts + ).person.set_index("person_id") + + assert person.loc[100_001, "dividend_income"] == pytest.approx( + 10.0 * frs_module.WEEKS_IN_YEAR + ) + assert person.loc[100_002, "dividend_income"] == 0 + + def test_create_frs_rejects_unknown_adult_empstati(tmp_path, monkeypatch): with pytest.raises(ValueError, match="EMPSTATI"): create_single_adult_frs(tmp_path, monkeypatch, empstati=12) diff --git a/policyengine_uk_data/utils/datasets.py b/policyengine_uk_data/utils/datasets.py index 2f4cd1b4..60d887d5 100644 --- a/policyengine_uk_data/utils/datasets.py +++ b/policyengine_uk_data/utils/datasets.py @@ -16,10 +16,12 @@ def sum_to_entity(values: pd.Series, foreign_key: pd.Series, primary_key) -> np. Args: values (pd.Series): The values in the non-entity table. foreign_key (pd.Series): E.g. pension.person_id. - primary_key ([type]): E.g. person.index. + primary_key ([type]): The entity's own ids in the same key space as + ``foreign_key``, e.g. person.person_id. A positional index such as + person.index matches the wrong rows or none. Returns: - pd.Series: A value for each person. + np.ndarray: A value for each entry of ``primary_key``. """ return values.groupby(foreign_key).sum().reindex(primary_key).fillna(0).values