Working Paper No. 03
Coverage gaps in the Hospice Care Index
Medicare publishes a quality score for 5,411 of 6,852 certified hospices. For the other 1,441 the page is blank. 854 were certified before the measurement window opened, and 735 carry a footnote that does not fit a claims-based measure.
Preprint, not peer-reviewedThree numbers to read first
Abstract
Medicare publishes the Hospice Care Index for 5,411 of 6,852 certified hospices. The remaining 1,441 have no score. Recent certification explains 587 of them. The other 854 operated throughout the full measurement window. The gap concentrates in for-profit providers, at 82.0% scored against 98.1% for non-profit, and in four states, with Nevada at 47.3% unscored and California at 33.8%. Among scored providers, the index compresses at the top of its range: 86.0% score 8 or higher and 38.6% score exactly 10, with mean score rising monotonically with provider age. Of the 1,441 unscored, 735 carry a footnote that describes a sampling condition or an infection confidence interval, neither of which applies to a claims-based index built entirely from Medicare claims already on file. For those providers, the reason for the missing score cannot be recovered from public data.
1. Why a blank score is a finding
A hospice is chosen quickly and under pressure. The decision is often made in a hospital corridor within a day of a terminal prognosis, by a family with no prior knowledge of the field and by a discharge planner working a caseload. Care Compare exists for exactly that moment. It is the only comparison tool most families will ever see.
The Hospice Care Index is the summary number on that page. It condenses ten claims-based indicators into a single score from zero to ten, covering staffing, visit patterns, spending, and live discharge behavior. It is designed to be read in a hurry by someone who will never read the technical manual.
For one in five hospices, that number is absent. The page loads, the provider is listed, and where the score would be there is nothing. A reader has no way to distinguish a hospice that is small and new from one that is small and troubled, and no way to tell which of those situations they are looking at.
One in five certified hospices has no published score
CMS Hospice Provider Data, retrieved July 2026. All 6,852 Medicare-certified hospices.
The measure matters beyond consumer choice. In June 2026 the Department of Justice charged a hospice owner in the Central District of California in a $27.7 million scheme. The alleged mechanism turned on this family of metrics directly. According to the charges, the defendant enrolled patients who were not terminally ill, understood that the percentage of patients discharged from hospice alive is monitored as a fraud indicator, and paid a funeral home employee between $1,000 and $3,000 per record for the identities of recently deceased Medicare beneficiaries in order to bill for them and reduce his outlier data metrics. Charges are allegations and the defendant is presumed innocent.
That case establishes only one thing for present purposes, and this paper claims nothing further from it. Live discharge behavior is treated as an integrity signal by people whose job is detecting fraud. Four of the ten Hospice Care Index indicators are denominated on live discharges. It is therefore worth asking how well the public version of that signal covers the field it is meant to describe.
The answer is that it covers 79.0% of it.
This paper follows the same argument as Working Paper No. 01, which examined nursing home staffing and found that a published quarterly average erased a within-week pattern that mattered operationally. The failure here is different in kind. The Hospice Care Index does not average away a signal. It declines to produce one, for a fifth of the field, and the public explanation for that silence is often unusable.
2. Being new does not explain the gap
The obvious explanation for a missing score is that the provider is new. A hospice certified partway through the measurement window would not have accumulated enough cases to compute a stable index, and CMS would correctly decline to publish one.
That explanation covers 587 of the 1,441. It does not cover the other 854, which were certified before January 1, 2023 and therefore had the entire two-year window available to them.
Most unscored providers were not new
The 1,441 unscored hospices, split by whether they were certified before the measurement window opened on January 1, 2023.
Scoring rates fall as certification volume rises
Share of each certification-year cohort that carries a published score. 2021 and 2022 had complete exposure to the measurement window.
The 2021 and 2022 cohorts were certified before the window opened and still score only 72.4% and 53.6%. This is reported as a pattern, not a cause.
3. The gap is not evenly distributed
Coverage differs sharply by ownership type. Non-profit providers are scored at 98.1%. For-profit providers are scored at 82.0%. The gap is roughly sixteen percentage points.
Coverage by ownership type
Share of certified hospices with a published Hospice Care Index score.
Geography shows the same concentration. Nevada has 47.3% of its certified hospices unscored and California 33.8%, against a national figure of 21.0%. Texas sits at 30.3%. California holds 30% of all certified hospices in the file and 48% of all unscored ones. These associations are reported without causal claims: a state with more recently certified for-profit hospices would produce both patterns for unrelated reasons.
4. Data
Two files from the CMS Provider Data Catalog, plus one reference table, retrieved July 28, 2026.
| File | Dataset ID | Rows |
|---|---|---|
| Hospice Provider Data | fm2n-hjj6 | 459,084 |
| Hospice General Information | yje5-53sw | 6,852 |
| Footnote Crosswalk | y9us-9xdf | 32 |
The unit of analysis is the CMS Certification Number, zero-padded to six characters before joining because the two files differ in leading-zero handling. All thirty-one Hospice Care Index measure codes carry the same measurement window, January 1, 2023 through December 31, 2024, read from the file rather than assumed. No providers were excluded, and territories are retained.
5. Method
Every step is a filter, a join, or a count, so that a reader who disagrees can name the step. Scored providers are those whose value at measure code H_012_00_OBSERVED parses as a number. Provider attributes are joined from the general information file. Certification date is compared against the window start to test the exposure explanation. Certification cohorts, footnote decomposition, ownership and state composition, and the distribution of the scored population are each handled as a separate cross-tabulation.
The analysis fits no model, estimates no effect, and tests no hypothesis about cause. Every result is a count, a share, or a cross-tabulation.
6. The score distribution is compressed
Among the 5,411 scored providers, the index clusters at the top of its range. No provider scores below 3. The effective range in the current refresh is 3 to 10, not 0 to 10.
More than a third of scored hospices tie at the maximum
Distribution of Hospice Care Index scores among the 5,411 scored providers.
86.0% of scored providers score 8 or higher. A measure where more than a third of providers share the maximum score cannot distinguish among them.
7. Score rises monotonically with provider age
Among scored providers, mean HCI rises with the number of years a provider was certified before the measurement window opened. The association is consistent with several explanations, including selection effects, learning effects, and unmeasured characteristics that correlate with both tenure and performance. This paper reports the association and does not adjudicate between them.
Mean score rises with every additional age bracket
Mean Hospice Care Index score by years certified before the measurement window opened.
8. A sampling footnote on a measure with no sampling step
Every unscored provider carries exactly one footnote on the composite measure code, and the four conditions sum to 1,441 with no remainder. Footnote 1, too few cases to report, is self-explanatory and covers 682 providers. Footnote 8, an infection confidence interval condition drawn from hospital-acquired infection measures, appears on 58 hospice scores with no apparent applicability.
A sampling footnote applied to a fully enumerated measure
The four footnote conditions carried by the 1,441 unscored providers.
Footnote 2 describes a sampling condition. The HCI has no sampling step: CMS confirms hospices are automatically 100% compliant on this measure because it is calculated entirely from claims already on file. The footnote does not describe a condition that can occur.
The Care Compare footnote vocabulary is shared across every provider setting on the platform: acute hospitals, skilled nursing facilities, home health agencies, hospices, and others. A footnote valid in one setting is available in all. Footnote 2 was built for survey-based or sampled measures and applied here, at scale, to a fully enumerated claims-based one. The reason a family looking up one of those 677 hospices sees no score is not recoverable from the published footnote.
Two hypotheses the data ruled out
Rejected: the suppression shelter. The hypothesis was that a hospice suppressing its live discharge rate would drop out of four indicators denominated on live discharges while retaining the other six, producing an artificially inflated composite. This is false. Of 1,385 providers missing at least one live discharge indicator, zero carry an overall HCI score. Missing any indicator removes the provider from the composite entirely.
Demoted: the zero floor as an integrity signal. Three of the four live discharge indicators have a 10th percentile of exactly zero. This is primarily a small-denominator artifact: the median denominator across all four indicators is 48 live discharges, so a single event is roughly 2%, and zero is an ordinary outcome for a small provider. This is retained as a precision limitation, not an integrity signal.
9. What this does and does not support
Supported
1,441 of 6,852 certified hospices have no published HCI score in the current data refresh.
854 of those were certified before the measurement window opened.
Unscored share is 21.0% nationally, 47.3% in Nevada, 33.8% in California, and 30.3% in Texas.
For-profit providers score at 82.0% against 98.1% for non-profit.
86.0% of scored providers score 8 or higher and 38.6% score exactly 10, with no provider below 3.
Mean HCI rises monotonically with provider age at window start.
Footnote 2 appears on 677 hospices despite the HCI having no sampling step for it to describe.
The suppression shelter hypothesis was tested against the data and rejected.
Not supported
No claim about fraud prevalence. No provider is named or ranked.
No claim that an unscored hospice is deficient. Most reflect case volume or window exposure and are ordinary.
No causal claim about why for-profit providers score lower or why some states have more unscored providers.
No claim linking geographic concentration to enforcement activity. That relationship was not tested.
No evaluation of hospice care quality itself, only of the public information signal.
No audit of CMS internal detection, which is outside scope and likely more capable than anything published.
The footnote finding is a mismatch in disclosure vocabulary, not an error in the index calculation.
10. Reproduction
Three scripts, run in sequence: fetch_hospice_pdc.py discovers and downloads the source files; analyze_hci.py extracts the score distribution, live discharge indicators, and footnote decomposition; analyze_coverage.py joins certification date and ownership type to the scored and unscored split. Python 3.9, pandas, requests.
The footnote-2 argument is a document claim, not a computation, and rests on the HCI measure specifications: the FY 2022 final rule and the February 2026 HQRP Current Measures List, both of which confirm the measure is calculated entirely from claims already on file with no separate submission or sampling step.
zatara-research-data/03-hospice-coverage-gaps
Cite this paper: