Scale Medicare per-enrollee spending to enrollees outside nursing homes

changelog.d/medicare-noninstitutional-spending.fixed.md

@@ -0,0 +1 @@
+Scale Medicare spending per enrollee to enrollees who do not live in nursing homes in medicare_gross_cost, medicare_cost, cbo_medicare_benefit and qmb_cost_sharing, since the survey data exclude nursing home residents.

policyengine_us/parameters/calibration/gov/hhs/medicare/noninstitutionalized_spending_ratio.yaml

@@ -0,0 +1,63 @@
+description: PolicyEngine multiplies Medicare spending per enrollee by this ratio of Medicare spending per enrollee who is not a long-stay nursing home resident to Medicare spending per enrollee, because the household survey data exclude people living in nursing homes.
+metadata:
+  unit: /1
+  period: year
+  label: Medicare noninstitutionalized spending ratio
+  reference:
+  # Medicare spending per enrollee (per_capita_cost) averages over every
+  # enrollee, including nursing home residents, whom the Current Population
+  # Survey does not sample. In 2022, among beneficiaries in fee-for-service
+  # Medicare, annual Part A and Part B program spending averaged $31,200 for
+  # long-stay nursing home residents and about $11,300 for other
+  # beneficiaries (June 2025, file page 9), and long-stay residents were 1.7
+  # percent of beneficiaries (October 2024, file page 5; 1.7 percent again in
+  # 2023, June 2025, file page 8). A long-stay resident had been in a nursing
+  # home for more than 90 days in at least one month of the year. So spending
+  # per other beneficiary relative to spending per beneficiary is
+  # 1 / (1 + 0.017 * (31_200 / 11_300 - 1)) = 0.971. The preliminary October
+  # 2024 slide described long-stay residents' spending as about two times
+  # higher; the June 2025 report's dollar figures, almost three times as
+  # high, replace that description.
+  #
+  # The ratio is measured for fee-for-service Parts A and B and applied to
+  # all Medicare spending per enrollee, including Medicare Advantage, Part D
+  # and administration. With fee-for-service enrollment shares instead (62
+  # percent of long-stay residents and 53 percent of others in 2023, June
+  # 2025, file page 13) it is 0.966; pricing Medicare Advantage at the
+  # long-stay residents' risk score ratio (2.12 / 1.01, October 2024, file
+  # page 5) gives about 0.975. The long-stay share was 2.6 percent in 2013
+  # (June 2025, file page 8), but MedPAC publishes relative spending for 2022
+  # only, so the 2022 ratio applies to every year from 2013, the first year of
+  # per_capita_cost.
+  - title: "Medicare Payment Advisory Commission, Report to the Congress, June 2025, Chapter 5, Medicare beneficiaries in nursing homes #page 8-9,13"
+    href: https://www.medpac.gov/wp-content/uploads/2025/06/Jun25_Ch5_MedPAC_Report_To_Congress_SEC.pdf#page=8
+  - title: Medicare Payment Advisory Commission, Medicare beneficiaries in nursing homes, October 2024 public meeting
+    href: https://www.medpac.gov/wp-content/uploads/2023/10/Beneficiaries-in-Nursing-Homes-MedPAC-Oct-2024-SEC.pdf#page=5
+  # The Current Population Survey samples housing units and some
+  # noninstitutional group quarters; institutional group quarters such as
+  # nursing homes have never been part of its sample. The census counts
+  # nursing facility patients at the facility, but counts residents of
+  # assisted living facilities and of housing for older adults, such as
+  # retirement communities, at the residence where they live and sleep most
+  # of the time (sections C.11(c), C.11(d) and C.12).
+  - title: U.S. Bureau of Labor Statistics, Handbook of Methods, Current Population Survey, Design
+    href: https://www.bls.gov/opub/hom/cps/design.htm
+  - title: "U.S. Census Bureau, Final 2020 Census Residence Criteria and Residence Situations, 83 FR 5525 #page 10-11"
+    href: https://www.govinfo.gov/content/pkg/FR-2018-02-08/pdf/2018-02370.pdf#page=10
+  # The Medicare Current Beneficiary Survey's long-term care facilities
+  # include retirement homes, personal care facilities and assisted and
+  # foster care homes as well as nursing homes (Appendix B). Excluding
+  # everyone living in one at the fall interview gives a larger gap: in
+  # 2013, community residents' Medicare benefit payments per capita were
+  # 0.92 of the average over all beneficiaries (Table 4.1 over Table 1.1).
+  - title: Centers for Medicare and Medicaid Services, Characteristics and Perceptions of the Medicare Population, 2013, Tables 1.1 and 4.1 and Appendix B
+    href: https://www.cms.gov/research-statistics-data-and-systems/research/mcbs/downloads/2013_characteristics_and_perceptions.zip
+  # CBO also removes institutionalized enrollees' benefits before assigning
+  # Medicare's average cost to survey respondents (file page 46). It holds
+  # the institutional share of Medicare benefits other than the low-income
+  # subsidy at its 2012 level from 2013 (file page 52), where this parameter
+  # instead holds the ratio of spending per person constant.
+  - title: "Congressional Budget Office, Working Paper 2018-07, How CBO Adjusts for Survey Underreporting of Transfer Income in Its Distributional Analyses #page 46,52"
+    href: https://www.cbo.gov/system/files/2018-07/54234-workingpaper.pdf#page=46
+values:
+  2013-01-01: 0.971

policyengine_us/tests/core/test_cbo_medicare_benefit_series.py

@@ -3,11 +3,12 @@
 CBO counts Medicare in household income at the government's average cost
 per enrollee, net of premiums paid by enrollees, and counts the Part D
 low-income subsidy separately. ``cbo_medicare_benefit`` takes CMS Medicare
-spending per enrollee (``calibration.gov.hhs.medicare.per_capita_cost``) and
-subtracts four series under ``calibration.gov.hhs.medicare``, each an
-aggregate from the Medicare Trustees Reports over Table V.B3's total
-enrollment (people with Part A and/or Part B). From 2015 they come from the
-2026 report:
+spending per enrollee (``calibration.gov.hhs.medicare.per_capita_cost``) less
+the low-income subsidy, scales it to enrollees outside nursing homes
+(``noninstitutionalized_spending_ratio``), and subtracts premium income.
+The four series under ``calibration.gov.hhs.medicare`` are each an aggregate
+from the Medicare Trustees Reports over Table V.B3's total enrollment (people
+with Part A and/or Part B). From 2015 they come from the 2026 report:
 
 - ``premium_income_per_enrollee.part_a``: Table III.B4 premiums from
   voluntary enrollees.
@@ -40,8 +41,10 @@
    standard premium times Part B enrollment plus Table V.E3's aggregate
    income-related premiums reproduces it within 0.5 percent, while the
    standard premium alone falls more than 5 percent short.
-5. Bounds: every series is positive, and net Medicare spending per enrollee
-   stays between 0 and gross spending, in every year from 2013 to 2200.
+5. Bounds: every series is positive, and net Medicare spending per enrollee,
+   (spending - low-income subsidy) x the noninstitutionalized ratio -
+   premiums, stays between 0 and scaled spending, in every year from 2013 to
+   2200.
 6. Extension: from 2036 to 2100 each series grows with Medicare spending
    per enrollee, so each keeps its 2035 share of spending; later years hold
    the 2100 value.
@@ -63,6 +66,7 @@
     "lis": f"{MEDICARE}.part_d_low_income_subsidy_per_enrollee",
 }
 SPENDING = f"{MEDICARE}.per_capita_cost"
+RATIO = f"{MEDICARE}.noninstitutionalized_spending_ratio"
 FIRST_YEAR, LAST_PROJECTED_YEAR, END_YEAR = 2015, 2035, 2100
 YEARS = range(FIRST_YEAR, LAST_PROJECTED_YEAR + 1)
 
@@ -225,9 +229,11 @@ def test_part_b_premium_income_includes_income_related_premiums():
 @given(st.integers(BACKFILL_YEARS[0], END_YEAR + 100))
 @example(BACKFILL_YEARS[0])
 def test_series_are_positive_and_net_spending_is_between_zero_and_gross(year):
-    deductions = sum(_value(series, year) for series in SERIES)
+    ratio = _read(RATIO, f"{year}-01-01")
+    premiums = sum(_value(series, year) for series in SERIES if series != "lis")
+    net = (_spending(year) - _value("lis", year)) * ratio - premiums
     assert all(_value(series, year) > 0 for series in SERIES)
-    assert 0 < _spending(year) - deductions < _spending(year)
+    assert 0 < net < _spending(year) * ratio
 
 
 @pytest.mark.parametrize("series", SERIES)

policyengine_us/tests/core/test_medicare_cost_part_d_premium.py

@@ -1,10 +1,11 @@
 """The Part D premium that ``medicare_cost`` subtracts.
 
 CMS Medicare spending per enrollee (``calibration.gov.hhs.medicare.
-per_capita_cost``) covers Parts A through D, averaged over every Medicare
-enrollee, gross of the premiums that finance it. ``medicare_cost`` subtracts
-the Part D premium averaged the same way, plus the person's own
-income-related monthly adjustment amount (IRMAA):
+per_capita_cost``, scaled by ``noninstitutionalized_spending_ratio`` to
+enrollees outside nursing homes) covers Parts A through D, with Part D
+averaged over every Medicare enrollee, gross of the premiums that finance it.
+``medicare_cost`` subtracts the Part D premium averaged the same way, plus the
+person's own income-related monthly adjustment amount (IRMAA):
 
     enrollment_share x premiums_per_enrollee + Part D IRMAA
 
@@ -38,7 +39,8 @@
 5. Calendar years: every date in a year reads the January 1 value; between
    the Trustees' five-year enrollment points the share holds the earlier one.
 6. Consumer (Hypothesis, differential): ``medicare_cost`` equals a reference
-   implementation, max(gross - Part A - Part B - share x premium - IRMAA, 0),
+   implementation, max(gross - Part A - Part B - share x premium - IRMAA, 0)
+   with gross = per_capita_cost x ``noninstitutionalized_spending_ratio``,
    for any year and premiums; it lies in [0, ``medicare_gross_cost``], which
    does not change with the Part D premium.
 7. Accounting (Hypothesis): the person's Part D IRMAA lowers
@@ -232,7 +234,10 @@ def _enrollee(year, part_a, part_b, irmaa):
 
 
 def _reference_medicare_cost(year, part_a, part_b, irmaa):
-    gross = _read("calibration.gov.hhs.medicare.per_capita_cost", f"{year}-01-01")
+    date = f"{year}-01-01"
+    gross = _read("calibration.gov.hhs.medicare.per_capita_cost", date) * _read(
+        "calibration.gov.hhs.medicare.noninstitutionalized_spending_ratio", date
+    )
     part_d = _share(year) * _premium(year) + irmaa
     return max(gross - part_a - part_b - part_d, 0)
 

policyengine_us/tests/core/test_medicare_noninstitutionalized_spending_ratio.py

@@ -0,0 +1,157 @@
+"""The ratio that scales Medicare spending per enrollee to enrollees outside
+nursing homes.
+
+``calibration.gov.hhs.medicare.per_capita_cost`` averages Medicare spending
+over every enrollee, including nursing home residents, whom the household
+survey data do not sample. ``noninstitutionalized_spending_ratio`` is
+spending per enrollee who is not a long-stay nursing home resident over
+spending per enrollee, from MedPAC's 2022 fee-for-service Part A and B
+spending ($31,200 for long-stay residents, about $11,300 for others; June
+2025 report, chapter 5, file page 9) and the long-stay share of beneficiaries
+(1.7 percent; October 2024 slides, file page 5):
+
+    ratio = 1 / (1 + share x (long-stay spending / other spending - 1))
+
+These tests read parameters only. Tests elsewhere cover the consumers at the
+stored ratio: the YAML tests of ``medicare_gross_cost``, ``medicare_cost``,
+``cbo_medicare_benefit`` and ``qmb_cost_sharing``, and the reference
+implementations in test_medicare_per_capita_cost.py,
+test_medicare_cost_part_d_premium.py and test_cbo_medicare_benefit_series.py,
+which multiply by this parameter.
+
+Invariants:
+
+1. Transcription: the stored value is the formula applied to MedPAC's
+   published figures, rounded to three decimals.
+2. Bounds and constancy: the ratio is strictly between 0 and 1 and is the
+   same on every date from 2013, the first year of ``per_capita_cost``.
+3. Sensitivity: the alternatives documented in the parameter file bracket
+   the stored value: fee-for-service enrollment shares (62 percent of
+   long-stay residents, 53 percent of others) give a lower ratio, and pricing
+   Medicare Advantage at the CMS-HCC risk score ratio (2.12 / 1.01) gives a
+   higher one.
+4. Definition order: excluding everyone living in a Medicare Current
+   Beneficiary Survey long-term care facility, which includes assisted living
+   and retirement homes, gives a lower ratio in every year of the published
+   2006-2013 tables (community-only over all-beneficiary benefit payments per
+   capita), so the nursing home ratio is the smaller of the two corrections.
+5. Differential: CBO's 2006 figures, from a different method and source,
+   imply a ratio per recipient and per survey reporter within 0.01 of the
+   stored value.
+"""
+
+import pytest
+from hypothesis import given
+from hypothesis import strategies as st
+from policyengine_core.parameters.operations.get_parameter import get_parameter
+
+from policyengine_us.system import system
+
+RATIO = "calibration.gov.hhs.medicare.noninstitutionalized_spending_ratio"
+FIRST_YEAR, END_YEAR = 2013, 2100
+
+# MedPAC, June 2025 Report to the Congress, chapter 5 (file pages 8, 9 and
+# 13), and October 2024 meeting slides (file page 5).
+LONG_STAY_SPENDING, OTHER_SPENDING = 31_200, 11_300
+LONG_STAY_SHARE = 0.017
+LONG_STAY_FFS_SHARE, OTHER_FFS_SHARE = 0.62, 0.53
+LONG_STAY_RISK_SCORE, OTHER_RISK_SCORE = 2.12, 1.01
+
+# MCBS Characteristics and Perceptions of the Medicare Population, each
+# year's edition
+# (https://www.cms.gov/research-statistics-data-and-systems/research/mcbs/
+# downloads/<year>_characteristics_and_perceptions.zip), all beneficiaries:
+# Table 1.1 beneficiaries (thousands) and Table 4.1 benefit payments ($
+# millions), total and community only.
+MCBS = {
+    2006: (43_867, 41_599, 305_066, 267_344),
+    2007: (45_046, 42_867, 307_372, 271_509),
+    2008: (46_100, 43_906, 320_032, 282_887),
+    2009: (47_371, 45_053, 335_081, 289_843),
+    2010: (48_454, 46_270, 365_098, 322_150),
+    2011: (50_186, 47_969, 374_252, 330_907),
+    2012: (51_723, 49_816, 338_395, 299_144),
+    2013: (53_574, 51_568, 352_943, 312_626),
+}
+
+# CBO 2006, in billions of dollars. Working Paper 2018-07: 93 percent of
+# enrollees reported Medicare in the survey and 6 percent of recipients were
+# institutionalized (file pages 45-46); Medicare benefits other than the
+# low-income subsidy after CBO's adjustments, $352.4, and the subsidy, $10.3
+# (file page 51). CBO (2013), The Distribution of Federal Spending and Taxes
+# in 2006: $32 of Medicare spending for people in long-term care institutions
+# (file page 53). 2007 Medicare Trustees Report, Table II.B1: 2006 benefits,
+# $402.0, and Part A premiums, $2.6, which CBO subtracts (file page 52).
+CBO_REPORTING_RATE, CBO_INSTITUTIONALIZED_SHARE = 0.93, 0.06
+CBO_ADJUSTED_BENEFITS, CBO_LOW_INCOME_SUBSIDY = 352.4, 10.3
+CBO_INSTITUTIONAL_SPENDING = 32
+TRUSTEES_BENEFITS, TRUSTEES_PART_A_PREMIUMS = 402.0, 2.6
+
+
+def _ratio(share, relative):
+    return 1 / (1 + share * (relative - 1))
+
+
+def _stored(date):
+    return get_parameter(system.parameters, RATIO)(date)
+
+
+def test_value_is_the_medpac_formula():
+    expected = _ratio(LONG_STAY_SHARE, LONG_STAY_SPENDING / OTHER_SPENDING)
+    assert _stored("2022-01-01") == round(expected, 3) == 0.971
+
+
+@given(st.integers(FIRST_YEAR, END_YEAR + 100), st.integers(1, 12))
+def test_ratio_is_a_constant_share_below_one(year, month):
+    value = _stored(f"{year}-{month:02d}-01")
+    assert 0 < value < 1
+    assert value == _stored("2022-01-01")
+
+
+def test_documented_alternatives_bracket_the_value():
+    relative = LONG_STAY_SPENDING / OTHER_SPENDING
+    ffs_share = (
+        LONG_STAY_SHARE
+        * LONG_STAY_FFS_SHARE
+        / (
+            LONG_STAY_SHARE * LONG_STAY_FFS_SHARE
+            + (1 - LONG_STAY_SHARE) * OTHER_FFS_SHARE
+        )
+    )
+    ffs_ratio = _ratio(ffs_share, relative)
+    # Price Medicare Advantage enrollees at the risk score ratio, with other
+    # beneficiaries' spending the same in both programs.
+    groups = {
+        "long_stay_ffs": (LONG_STAY_SHARE * LONG_STAY_FFS_SHARE, relative),
+        "long_stay_ma": (
+            LONG_STAY_SHARE * (1 - LONG_STAY_FFS_SHARE),
+            LONG_STAY_RISK_SCORE / OTHER_RISK_SCORE,
+        ),
+        "other": (1 - LONG_STAY_SHARE, 1),
+    }
+    average = sum(count * cost for count, cost in groups.values())
+    mixed_ratio = 1 / average
+    assert round(ffs_ratio, 3) == 0.966
+    assert round(mixed_ratio, 3) == 0.975
+    assert ffs_ratio < _stored("2022-01-01") < mixed_ratio
+
+
+@pytest.mark.parametrize("year", MCBS)
+def test_mcbs_facility_definition_gives_a_larger_gap(year):
+    total_people, community_people, total_pay, community_pay = MCBS[year]
+    mcbs_ratio = (community_pay / community_people) / (total_pay / total_people)
+    # The stored ratio starts in 2013, so compare every year to its value.
+    assert 0.9 < mcbs_ratio < _stored("2022-01-01")
+
+
+def test_cbo_2006_figures_imply_a_similar_ratio():
+    benefits = TRUSTEES_BENEFITS - TRUSTEES_PART_A_PREMIUMS - CBO_LOW_INCOME_SUBSIDY
+    assert benefits == pytest.approx(389.1)
+    per_recipient = (1 - CBO_INSTITUTIONAL_SPENDING / benefits) / (
+        1 - CBO_INSTITUTIONALIZED_SHARE
+    )
+    per_reporter = CBO_ADJUSTED_BENEFITS / benefits / CBO_REPORTING_RATE
+    assert round(per_recipient, 3) == 0.976
+    assert round(per_reporter, 3) == 0.974
+    for implied in (per_recipient, per_reporter):
+        assert implied == pytest.approx(_stored("2013-01-01"), abs=0.01)