microimpute 3.1.4__tar.gz → 3.2.0__tar.gz

{microimpute-3.1.4 → microimpute-3.2.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: microimpute
-Version: 3.1.4
+Version: 3.2.0
 Summary: Benchmarking imputation methods for microdata
 Author-email: Vahid Ahmadi <vahid@policyengine.org>, Max Ghenis <max@policyengine.org>, María Juaristi <juaristi@uni.minerva.edu>, Nikhil Woodruff <nikhil.woodruff@outlook.com>
 Project-URL: Repository, https://github.com/PolicyEngine/microimpute
@@ -34,6 +34,7 @@ Requires-Dist: ruff<0.17.0,>=0.16.0; extra == "dev"
 Requires-Dist: mypy<2.0.0,>=1.2.3; extra == "dev"
 Requires-Dist: build<2.0.0,>=1.2.0; extra == "dev"
 Requires-Dist: towncrier>=24.8.0; extra == "dev"
+Requires-Dist: hypothesis<7.0.0,>=6.100.0; extra == "dev"
 Provides-Extra: matching
 Requires-Dist: rpy2<4.0.0,>=3.5.0; extra == "matching"
 Provides-Extra: mdn

{microimpute-3.1.4 → microimpute-3.2.0}/microimpute.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: microimpute
-Version: 3.1.4
+Version: 3.2.0
 Summary: Benchmarking imputation methods for microdata
 Author-email: Vahid Ahmadi <vahid@policyengine.org>, Max Ghenis <max@policyengine.org>, María Juaristi <juaristi@uni.minerva.edu>, Nikhil Woodruff <nikhil.woodruff@outlook.com>
 Project-URL: Repository, https://github.com/PolicyEngine/microimpute
@@ -34,6 +34,7 @@ Requires-Dist: ruff<0.17.0,>=0.16.0; extra == "dev"
 Requires-Dist: mypy<2.0.0,>=1.2.3; extra == "dev"
 Requires-Dist: build<2.0.0,>=1.2.0; extra == "dev"
 Requires-Dist: towncrier>=24.8.0; extra == "dev"
+Requires-Dist: hypothesis<7.0.0,>=6.100.0; extra == "dev"
 Provides-Extra: matching
 Requires-Dist: rpy2<4.0.0,>=3.5.0; extra == "matching"
 Provides-Extra: mdn

{microimpute-3.1.4 → microimpute-3.2.0}/microimpute.egg-info/SOURCES.txt

@@ -38,6 +38,7 @@ tests/test_basic.py
 tests/test_dashboard_formatter.py
 tests/test_data_preprocessing.py
 tests/test_metrics.py
+tests/test_moop_spm_experiment.py
 tests/test_predictor_analysis.py
 tests/test_public_api.py
 tests/test_quantile_comparison.py

{microimpute-3.1.4 → microimpute-3.2.0}/microimpute.egg-info/requires.txt

@@ -19,6 +19,7 @@ ruff<0.17.0,>=0.16.0
 mypy<2.0.0,>=1.2.3
 build<2.0.0,>=1.2.0
 towncrier>=24.8.0
+hypothesis<7.0.0,>=6.100.0
 
 [docs]
 jupyter-book

{microimpute-3.1.4 → microimpute-3.2.0}/pyproject.toml

@@ -7,7 +7,7 @@ build-backend = "setuptools.build_meta"
 
 [project]
 name = "microimpute"
-version = "3.1.4"
+version = "3.2.0"
 description = "Benchmarking imputation methods for microdata"
 readme = "README.md"
 authors = [
@@ -49,6 +49,7 @@ dev = [
     "mypy>=1.2.3,<2.0.0",
     "build>=1.2.0,<2.0.0",
     "towncrier>=24.8.0",
+    "hypothesis>=6.100.0,<7.0.0",
 ]
 
 matching = [

microimpute-3.2.0/tests/test_moop_spm_experiment.py

@@ -0,0 +1,477 @@
+"""Property tests for the MOOP-imputation SPM poverty experiment.
+
+The experiment lives in ``experiments/moop_spm``. These tests use small
+synthetic data, so they run without the CPS ASEC download. They check
+the invariants the published comparison relies on:
+
+- Fold assignment partitions households and keeps each household whole.
+- Poverty is recomputed exactly from resources, MOOP and thresholds, and
+  it never falls when any unit's MOOP rises.
+- People whose unit is poor before MOOP stay poor under any imputation.
+- Weighted group rates, replicate SEs and weighted quantiles match direct
+  reference computations.
+- Cell pools hold only same-cell donor spenders, and the nearest-donor
+  search agrees with a brute-force search.
+- The logistic-participation variants follow their stated rules.
+- Census's MOOP and Medicare premium imputations both count as allocated.
+- Extracted CSVs that differ from the pinned zip are extracted again.
+- The fold cache misses whenever the environment, microimpute's source
+  or the model inputs change.
+"""
+
+import argparse
+import os
+import zipfile
+
+import numpy as np
+import pandas as pd
+import pytest
+from hypothesis import HealthCheck, given, settings
+from hypothesis import strategies as st
+
+from experiments.moop_spm import run as run_module
+from experiments.moop_spm.cps_asec import allocation_flags, extract_verified
+from experiments.moop_spm.evaluate import (
+    RATIO_BANDS,
+    evaluate,
+    group_rates,
+    poverty_flags,
+    ratio_band_index,
+    replicate_se,
+    weighted_quantile,
+)
+from experiments.moop_spm.methods import (
+    CELL_LEVELS,
+    PARTICIPATION_CUTOFF,
+    FoldData,
+    _key,
+    assign_cells,
+    fit_participation,
+    logit_family,
+    nearest_donors,
+    participation_probability,
+)
+from experiments.moop_spm.run import (
+    assign_folds,
+    build_config,
+    collect_folds,
+    config_hash,
+    frame_sha256,
+)
+
+SETTINGS = settings(
+    max_examples=40,
+    deadline=None,
+    suppress_health_check=[HealthCheck.too_slow],
+)
+
+
+def synthetic_spm(seed: int, n_units: int = 60):
+    """Persons and units frames with the columns ``evaluate`` reads."""
+    rng = np.random.default_rng(seed)
+    unit_ids = np.arange(1, n_units + 1) * 7
+    threshold = rng.uniform(20_000, 40_000, n_units)
+    moop = np.where(rng.random(n_units) < 0.1, 0.0, rng.gamma(1.0, 5_000, n_units))
+    pre_moop = threshold * rng.uniform(0.5, 3.0, n_units)
+    units = pd.DataFrame(
+        {
+            "moop": moop,
+            "unit_weight": rng.uniform(100, 3_000, n_units),
+            "pre_moop_ratio": pre_moop / threshold,
+            "any_allocated": rng.integers(0, 2, n_units),
+        },
+        index=pd.Index(unit_ids, name="SPM_ID"),
+    )
+    sizes = rng.integers(1, 4, n_units)
+    person_unit = np.repeat(np.arange(n_units), sizes)
+    n_persons = len(person_unit)
+    persons = pd.DataFrame(
+        {
+            "SPM_ID": unit_ids[person_unit],
+            "A_AGE": rng.integers(0, 90, n_persons),
+            "PRDISFLG": rng.choice([-1, 1, 2], n_persons),
+            "coverage": rng.choice(
+                [
+                    "medicare",
+                    "employer",
+                    "direct",
+                    "medicaid",
+                    "military",
+                    "uninsured",
+                ],
+                n_persons,
+            ),
+            "SPM_MEDXPNS": moop[person_unit],
+            "SPM_RESOURCES": (pre_moop - moop)[person_unit],
+            "SPM_POVTHRESHOLD": threshold[person_unit],
+            "weight": rng.uniform(100, 3_000, n_persons),
+        }
+    )
+    persons["SPM_POOR"] = (
+        persons["SPM_RESOURCES"] < persons["SPM_POVTHRESHOLD"]
+    ).astype(int)
+    replicates = persons["weight"].to_numpy()[:, None] * rng.uniform(
+        0.5, 1.5, (n_persons, 160)
+    )
+    replicate_frame = pd.DataFrame(
+        replicates, columns=[f"repwgt_{i}" for i in range(1, 161)]
+    )
+    persons = pd.concat([persons, replicate_frame], axis=1)
+    return persons, units
+
+
+@SETTINGS
+@given(
+    households=st.lists(st.integers(1, 40), min_size=5, max_size=120),
+    n_folds=st.integers(2, 6),
+    seed=st.integers(0, 2**31),
+)
+def test_folds_partition_households(households, n_folds, seed):
+    series = pd.Series(households)
+    folds = assign_folds(series, n_folds, seed)
+    assert set(np.unique(folds)) <= set(range(n_folds))
+    # Every household lands in exactly one fold.
+    assert (series.groupby(folds).nunique().sum()) == series.nunique()
+    per_household = pd.Series(folds).groupby(series.to_numpy()).nunique()
+    assert (per_household == 1).all()
+    # Household counts per fold differ by at most one.
+    household_counts = series.groupby(folds).nunique()
+    assert household_counts.max() - household_counts.min() <= 1
+    assert (assign_folds(series, n_folds, seed) == folds).all()
+
+
+@SETTINGS
+@given(seed=st.integers(0, 2**31), bump_seed=st.integers(0, 2**31))
+def test_poverty_recomputation_and_monotonicity(seed, bump_seed):
+    persons, units = synthetic_spm(seed)
+    reported = poverty_flags(persons, units, units["moop"].to_numpy())[0]
+    assert (reported == (persons["SPM_POOR"] == 1)).all()
+
+    bump = np.random.default_rng(bump_seed).uniform(0, 20_000, len(units))
+    higher = poverty_flags(persons, units, units["moop"].to_numpy() + bump)[0]
+    # More MOOP can only push people into poverty.
+    assert (higher >= reported).all()
+
+    zero = poverty_flags(persons, units, np.zeros(len(units)))[0]
+    assert (zero <= reported).all()
+    pre_moop_poor = persons["SPM_ID"].map(units["pre_moop_ratio"]).to_numpy() < 1.0
+    # Units poor before MOOP stay poor under any nonnegative MOOP.
+    assert zero[pre_moop_poor].all()
+    assert higher[pre_moop_poor].all()
+
+
+@SETTINGS
+@given(seed=st.integers(0, 2**31))
+def test_group_rates_match_np_average(seed):
+    rng = np.random.default_rng(seed)
+    n = 50
+    flags = rng.random(n)
+    weights = rng.uniform(0.1, 10, (n, 3))
+    groups = rng.random((n, 4)) < 0.5
+    groups[0] = True  # keep every group nonempty
+    rates = group_rates(flags, groups, weights)
+    for g in range(groups.shape[1]):
+        for w in range(weights.shape[1]):
+            expected = np.average(flags[groups[:, g]], weights=weights[groups[:, g], w])
+            assert rates[g, w] == pytest.approx(expected)
+    assert ((rates >= 0) & (rates <= 1)).all()
+
+
+@SETTINGS
+@given(
+    base=st.floats(-5, 5),
+    deviations=st.lists(st.floats(-1, 1), min_size=160, max_size=160),
+    scale=st.floats(-3, 3),
+)
+def test_replicate_se_formula(base, deviations, scale):
+    estimates = np.array([base] + [base + d for d in deviations])
+    expected = np.sqrt(4 / 160 * np.sum(np.square(deviations)))
+    assert replicate_se(estimates) == pytest.approx(expected)
+    assert replicate_se(scale * estimates) == pytest.approx(abs(scale) * expected)
+    assert replicate_se(np.full(161, base)) == 0
+
+
+@SETTINGS
+@given(
+    values=st.lists(st.floats(-1e6, 1e6), min_size=1, max_size=60),
+    q=st.floats(0.01, 0.99),
+)
+def test_weighted_quantile_equal_weights(values, q):
+    values = np.array(values)
+    result = weighted_quantile(values, np.ones(len(values)), q)
+    assert result in values
+    assert result == pytest.approx(np.quantile(values, q, method="inverted_cdf"))
+
+
+def _cells(rng: np.random.Generator, n: int) -> pd.DataFrame:
+    return pd.DataFrame(
+        {
+            "group": rng.choice(["medicare", "employer", "public"], n),
+            "age_band": rng.choice(["u45", "45_64", "65p"], n),
+            "income_band": rng.choice(["lt1", "1_2", "2_4", "4p"], n),
+            "size_band": rng.choice(["1", "2", "3p"], n),
+        }
+    )
+
+
+@SETTINGS
+@given(seed=st.integers(0, 2**31), min_donors=st.integers(1, 40))
+def test_cell_pools_are_same_cell_donor_spenders(seed, min_donors):
+    rng = np.random.default_rng(seed)
+    donors = _cells(rng, 400)
+    moop = np.where(rng.random(400) < 0.2, 0.0, rng.gamma(1.0, 3_000, 400))
+    recipients = _cells(rng, 50)
+    assignment = assign_cells(donors, moop, recipients, min_donors=min_donors)
+    spender = moop > 0
+    donor_keys = [_key(donors.loc[spender], columns) for columns in CELL_LEVELS]
+    recipient_keys = [_key(recipients, columns) for columns in CELL_LEVELS]
+    for i, (pool, level) in enumerate(zip(assignment.pools, assignment.levels)):
+        assert len(pool) >= min_donors
+        assert (pool > 0).all()
+        in_cell = donor_keys[level] == recipient_keys[level][i]
+        expected = np.sort(moop[spender][in_cell])
+        assert np.array_equal(np.sort(pool), expected)
+        # No finer level had enough donor spenders.
+        for finer in range(level):
+            assert (donor_keys[finer] == recipient_keys[finer][i]).sum() < min_donors
+    assert np.allclose(assignment.means(), [p.mean() for p in assignment.pools])
+
+
+@SETTINGS
+@given(
+    donor_pred=st.lists(st.floats(-50, 50, allow_nan=False), min_size=1, max_size=40),
+    recipient_pred=st.lists(
+        st.floats(-60, 60, allow_nan=False), min_size=1, max_size=20
+    ),
+    k=st.integers(1, 8),
+)
+def test_nearest_donors_matches_brute_force(donor_pred, recipient_pred, k):
+    donor_pred = np.array(donor_pred)
+    recipient_pred = np.array(recipient_pred)
+    nearest = nearest_donors(donor_pred, recipient_pred, k)
+    k_eff = min(k, len(donor_pred))
+    assert nearest.shape == (len(recipient_pred), k_eff)
+    for i, value in enumerate(recipient_pred):
+        assert len(set(nearest[i])) == k_eff
+        got = np.sort(np.abs(donor_pred[nearest[i]] - value))
+        brute = np.sort(np.abs(donor_pred - value))[:k_eff]
+        assert np.allclose(got, brute)
+
+
+def _fold(seed: int) -> FoldData:
+    rng = np.random.default_rng(seed)
+    n_donor, n_recipient = 600, 80
+
+    def features(n):
+        return pd.DataFrame(
+            {
+                "x1": rng.normal(size=n),
+                "x2": rng.normal(size=n),
+                "x3": rng.integers(0, 2, n).astype(float),
+            }
+        )
+
+    X_donor = features(n_donor)
+    logit = 2.0 + 1.5 * X_donor["x1"].to_numpy()
+    spends = rng.random(n_donor) < 1 / (1 + np.exp(-logit))
+    moop = np.where(spends, rng.gamma(1.0, 4_000, n_donor), 0.0)
+    cells_donor = _cells(rng, n_donor)
+    cells_recipient = _cells(rng, n_recipient)
+    return FoldData(
+        X_donor=X_donor,
+        moop_donor=moop,
+        cells_donor=cells_donor,
+        group_donor=cells_donor["group"].to_numpy(),
+        X_recipient=features(n_recipient),
+        cells_recipient=cells_recipient,
+        group_recipient=cells_recipient["group"].to_numpy(),
+    )
+
+
+@settings(max_examples=8, deadline=None)
+@given(seed=st.integers(0, 2**31))
+def test_logit_family_rules(seed):
+    fold = _fold(seed)
+    out = logit_family(fold, 4, np.random.default_rng(seed))
+    prob = participation_probability(
+        fit_participation(fold.X_donor, fold.moop_donor), fold.X_recipient
+    )
+    cells = assign_cells(fold.cells_donor, fold.moop_donor, fold.cells_recipient)
+    means = cells.means()
+    below = prob < PARTICIPATION_CUTOFF
+    for method, draws in out.items():
+        assert draws.shape == (4, len(prob))
+        assert (draws >= 0).all(), method
+    assert np.allclose(out["logit_cell_mean"][0], np.where(below, 0, means))
+    assert np.allclose(out["logit_expected_value"][0], prob * means)
+    assert (out["logit_reg_mean"][:, below] == 0).all()
+    assert (out["logit_group_mean"][:, below] == 0).all()
+    assert (out["logit_threshold_hotdeck"][:, below] == 0).all()
+    for i, pool in enumerate(cells.pools):
+        for method in ("logit_draw_hotdeck", "logit_threshold_hotdeck"):
+            positive = out[method][:, i][out[method][:, i] > 0]
+            assert np.isin(positive, pool).all()
+        drawn = out["logit_draw_cell_mean"][:, i]
+        assert np.all((drawn == 0) | np.isclose(drawn, means[i]))
+
+
+@settings(max_examples=10, deadline=None)
+@given(seed=st.integers(0, 2**31))
+def test_evaluate_invariants(seed):
+    persons, units = synthetic_spm(seed, n_units=80)
+    bump = np.random.default_rng(seed + 1).uniform(0, 10_000, len(units))
+    imputations = {
+        "no_moop": np.zeros((1, len(units))),
+        "higher": (units["moop"].to_numpy() + bump)[None, :],
+        "same": np.tile(units["moop"].to_numpy(), (3, 1)),
+    }
+    results = evaluate(persons, units, imputations)
+    groups = results["groups"]
+    methods = results["methods"]
+    nonempty = np.array(results["group_person_counts"]) > 0
+    assert np.allclose(np.array(methods["same"]["diff"])[nonempty], 0)
+    assert methods["same"]["newly_poor_share"] == 0
+    assert methods["same"]["newly_nonpoor_share"] == 0
+    assert (np.array(methods["no_moop"]["diff"])[nonempty] <= 1e-12).all()
+    assert (np.array(methods["higher"]["diff"])[nonempty] >= -1e-12).all()
+    band = groups.index(["pre_moop_ratio", RATIO_BANDS[0][2]])
+    if nonempty[band]:
+        for method in methods.values():
+            assert method["rate"][band] == pytest.approx(1.0)
+
+
+@SETTINGS
+@given(
+    ratios=st.lists(
+        st.one_of(
+            st.floats(-5, 20, allow_nan=False),
+            st.sampled_from([1.0, 1.25, 1.5, 2.0, 3.0]),
+        ),
+        min_size=1,
+        max_size=50,
+    )
+)
+def test_ratio_band_index_matches_band_masks(ratios):
+    ratios = np.array(ratios)
+    index = ratio_band_index(ratios)
+    for b, (low, high, _) in enumerate(RATIO_BANDS):
+        mask = (ratios >= low) & (ratios < high)
+        assert np.array_equal(index == b, mask)
+
+
+def test_allocation_flags_cover_moop_and_premium():
+    persons = pd.DataFrame(
+        {
+            "I_MOOP": [-1, 0, 1, 2, 3, 0, 0],
+            "I_MCPREM": [-1, -1, -1, 0, -1, 2, 0],
+        }
+    )
+    flags = allocation_flags(persons)
+    assert flags["allocated_moop"].tolist() == [0, 0, 1, 1, 1, 0, 0]
+    assert flags["allocated_premium"].tolist() == [0, 0, 0, 0, 0, 1, 0]
+
+
+def test_extract_verified_replaces_altered_copies(tmp_path):
+    zip_path = tmp_path / "release.zip"
+    with zipfile.ZipFile(zip_path, "w") as archive:
+        archive.writestr("a.csv", "x,y\n1,2\n")
+        archive.writestr("b.csv", "z\n3\n")
+    out = tmp_path / "data"
+    out.mkdir()
+    extract_verified(zip_path, out, ["a.csv", "b.csv"])
+    assert (out / "a.csv").read_text() == "x,y\n1,2\n"
+
+    # A matching copy stays in place.
+    os.utime(out / "b.csv", ns=(1, 1))
+    extract_verified(zip_path, out, ["b.csv"])
+    assert (out / "b.csv").stat().st_mtime_ns == 1
+
+    # Altered copies, of the same or a different size, come back.
+    (out / "a.csv").write_text("x,y\n9,2\n")
+    (out / "b.csv").write_text("z\n3\n4\n")
+    extract_verified(zip_path, out, ["a.csv", "b.csv"])
+    assert (out / "a.csv").read_text() == "x,y\n1,2\n"
+    assert (out / "b.csv").read_text() == "z\n3\n"
+
+
+def _cache_inputs(n_units: int = 12):
+    rng = np.random.default_rng(0)
+    index = pd.Index(np.arange(n_units) * 3 + 1, name="SPM_ID")
+    units = pd.DataFrame(
+        {
+            "moop": rng.gamma(1.0, 1_000, n_units),
+            "household": np.arange(n_units) // 2,
+        },
+        index=index,
+    )
+    X = pd.DataFrame({"x": rng.normal(size=n_units)}, index=index)
+    cells = pd.DataFrame({"group": ["employer"] * n_units}, index=index)
+    folds = assign_folds(units["household"], 3, 0)
+    args = argparse.Namespace(
+        folds=3, draws=2, seed=0, skip_forests=False, max_households=None
+    )
+    return args, units, X, cells, folds
+
+
+def _stub_fit(value: float, folds: np.ndarray, calls: list):
+    def fit(fold: int) -> dict[str, np.ndarray]:
+        calls.append(fold)
+        n = int((folds == fold).sum())
+        return {
+            "qrf": np.full((2, n), value),
+            "_timings": np.array([[0.0]]),
+            "_timing_names": np.array(["qrf"]),
+        }
+
+    return fit
+
+
+def test_cache_misses_when_environment_or_microimpute_changes(tmp_path, monkeypatch):
+    args, units, X, cells, folds = _cache_inputs()
+    base_env = run_module.environment()
+    monkeypatch.setattr(run_module, "environment", lambda: base_env)
+    tag = config_hash(build_config(args, units, X, cells, folds))
+
+    calls: list = []
+    first, _ = collect_folds(tmp_path, tag, folds, 3, 2, _stub_fit(17.0, folds, calls))
+    assert calls == [0, 1, 2]
+    assert (first["qrf"] == 17.0).all()
+
+    # Same fingerprint: every fold comes from the cache.
+    calls.clear()
+    again, _ = collect_folds(tmp_path, tag, folds, 3, 2, _stub_fit(99.0, folds, calls))
+    assert calls == []
+    assert (again["qrf"] == 17.0).all()
+
+    # A new microimpute version, or edited microimpute source, refits.
+    newer = {**base_env, "microimpute": base_env["microimpute"] + ".post1"}
+    monkeypatch.setattr(run_module, "environment", lambda: newer)
+    new_tag = config_hash(build_config(args, units, X, cells, folds))
+    assert new_tag != tag
+    fresh, _ = collect_folds(
+        tmp_path, new_tag, folds, 3, 2, _stub_fit(99.0, folds, calls)
+    )
+    assert calls == [0, 1, 2]
+    assert (fresh["qrf"] == 99.0).all()
+
+    monkeypatch.setattr(run_module, "environment", lambda: base_env)
+    monkeypatch.setattr(run_module, "package_sha256", lambda package: "edited")
+    assert config_hash(build_config(args, units, X, cells, folds)) != tag
+
+
+def test_cache_key_tracks_model_inputs_and_row_keys():
+    args, units, X, cells, folds = _cache_inputs()
+    tag = config_hash(build_config(args, units, X, cells, folds))
+    assert config_hash(build_config(args, units, X, cells, folds)) == tag
+
+    changed = units.copy()
+    changed.iloc[0, 0] += 1.0
+    assert config_hash(build_config(args, changed, X, cells, folds)) != tag
+
+    rekeyed = [frame.copy() for frame in (units, X, cells)]
+    for frame in rekeyed:
+        frame.index = frame.index + 1
+    assert config_hash(build_config(args, *rekeyed, folds)) != tag
+
+    assert config_hash(build_config(args, units, X, cells, folds[::-1])) != tag
+    assert frame_sha256(units) != frame_sha256(units.rename(columns={"moop": "m"}))