“Deputy CMS Administrator Kimberly Brandt with a major update on the hospice fraud crackdown in California … 1,076 hospices have been removed from the Medicare program in California alone... these are hospices that are no longer under review. They’re not on a watch list. They are GONE; thrown out of the program entirely—unable to bill Medicare, unable to touch another taxpayer dollar, unable to take advantage of another vulnerable beneficiary.”

Think about this: one in six hospices may be removed from the Medicare program. How will that affect the proposed SSVI cutpoints? CMS needs to consider the impact of hospices that may be stealing Medicare beneficiary identities, including from deceased people or beneficiaries in long-term care. Put simply, deceased people may appear in hospice claims, but they will not generate hospital bills or other Medicare spending that is counted as non-hospice spending.

So let’s look at the data through a thought experiment.

Analyses of the 2024 SSVI Excel spreadsheet show the impact on the distribution of non-hospice spending, as shown in Table 1. I asked Claude AI to conduct a simulation comparing the current full distribution with a scenario in which all Los Angeles County hospices were removed. I then conducted two additional simulations: one in which another 500 or 1,000 hospices were removed with a distribution similar to Los Angeles County hospices, and another in which 500 or 1,000 hospices with lower non-hospice spending, similar to Los Angeles County hospices, were removed.

Methods

The observed rows in Table 1 are computed directly from the FY2024 Service and Spending Variation Index (SSVI) provider-level file, using each hospice’s total non-hospice Medicare spending; the first row covers all scored hospices and the second excludes every hospice in Los Angeles County. The simulated rows project the effect of removing an additional 500 or 1,000 hospices resembling the Los Angeles cluster.

The simulation reconstructs a synthetic spending distribution calibrated to reproduce the file’s published percentiles, mean, and maximum, together with an “LA-like” removal signature tuned to reproduce the observed shift when Los Angeles County is excluded; the reconstruction reproduces the observed Los-Angeles-excluded percentiles to within a few percent. That signature is applied under two definitions and at two magnitudes. “Matched” removes hospices drawn from the Los Angeles spending profile — a low-tilted spread rather than only the lowest values. “Lowest-spending” removes the smallest spenders and bounds the largest upward shift the removal could produce; the true effect of a continued purge is expected to fall between the two. Under matched removal the shift scales roughly linearly with the number removed.

Table 2 reports the seven boundaries that divide positive-spending hospices into the eight equal groups CMS uses to assign the non-hospice-spending score (0 to 8 points). These boundaries are derived from the reconstructed distribution and are intended to show the direction and magnitude of movement, not to reproduce CMS’s official thresholds. All figures use total, not per-beneficiary, non-hospice spending, so differences reflect a combination of agency size and per-patient utilization.

The impact of this analysis depends on how CMS chooses to use the SSVI. However, because CMS has encouraged Medicare beneficiaries and others to use these public data when making health care decisions, changes in the underlying hospice population could have practical consequences for how the measure is interpreted

.

Well, let me tell you a story about Goldilocks CHC.

In 2017, Vitas settled a lawsuit for $75 million over its use of CHC. Vitas is a large hospice program, so you would expect it to have a high proportion of CHC days — but what rate of CHC is the right rate? But the lawsuit was not just about the high rate of CHC; it was that Vitas was providing incentives to enroll ineligible patients and pressuring staff to increase the level of CHC to boost profits. Vitas entered a five-year Corporate Integrity Agreement that ended in 2022. It is disappointing to read Michael Williams’s article in The Capitol Forum (see https://thecapitolforum.com/chemed-hospice-subsidiary-vitas-healthcare-overutilizing-intensive-services-at-medicares-expense-company-has-long-history-of-improper-billing-practices/) noting high rates of CHC, a patient’s family stating that visits did not occur.

So what is this “Goldilocks CHC”?

I looked at Care Compare data between 2019 and 2024. Here is the difficulty: we don’t know what is “just right,” and there are important patient and policy concerns about both “too little” and “too much” CHC.

Let’s tackle “too little”

Among hospices with an ADC of at least 50, 815 out of 2,138 (38.1%) hospices did not provide a single CHC or GIP day — across more than 69 million days of hospice services. Potential hypotheses: 1) outstanding care; 2) some patients may be suffering because needed services are not being provided; and/or 3) these hospices are not enrolling terminally ill persons.

If I were a state surveyor, I would carefully review charts of persons who died in that hospice, and would probably pull, if possible, the sample from CAHPS hospice surveys where the respondent stated they did not receive “enough” help managing symptoms. The advantage of the state survey is that you could speak with bereaved family members to get their perspective on the quality of care, which may reveal discrepancies with the chart documentation. From the viewpoint of an MAC, I would ask for an ADR (additional documentation request) on a sample of CHC cases to verify their eligibility.

“Too much” is a difficult case

Take the example of one Vitas CCN in Florida that bills for almost 9% of all CHC days in the USA. But that hospice has an ADC of over 10,000 patients, so it is expected to bill for a lot of CHC days. So what are the possible measures to examine “too much” CHC/GIP for November 25 public release of Care Compare Data?

Percentage of CHC days 2.5%

Percentile overall 95.0%

Percentage CHC days over 8.69%

National number of CHC days

Observed to Expected 2.73 so this is 2.5 / 0.92 = 2.73

Percentile with those hospices

In the tenth size decile 84.5%

All we can say from these measures is that, even accounting for Vitas’s size, it is an outlier in the provision of CHC and GIP. Given its low rate of GIP, the majority is CHC. So the difficult next step should be a state survey to ensure quality of patient care, and/or a MAC ADR to ensure appropriate eligibility. The article by Michael Williams of The Capitol Forum interviews staff and family members who raise important concerns that despite the CIA that Vitas remains an outlier in Florida.

One caveat concerns the GIP level of care and health-care transitions. An aggressive policy of targeting GIP outliers could reduce access to freestanding hospice inpatient units, or drive unwanted transitions in the last weeks of life. Prudent policy must account for this potential unintended consequence and hopefully arrive at policy solution that does not result in health care transitions in the last days of life.

So we have quite the conundrum: as a policy maker, what is “too much,” “too little,” and “just right” for CHC?

Dame Cicely Saunders, the founder of the modern hospice movement, conceptualized suffering in terms of its psychological, social, physical, and spiritual dimensions. What is unique about hospice is the multidisciplinary team that includes valued input from health care professionals with a background in each of these domains. In the 1960s, her conceptualization of suffering as not just physical pain was an important contribution. As a physician/scientist, I believe that assessment tools, from surveys to clinical instruments, should honor this important contribution. This led to my fundamental question: what is missing in the HOPE tool, whose name proclaims that it assesses both outcomes and the patient.

Question 1 – What are the domains and content of HOPE?

Note: the entire caregiver domain is one item (A1910, availability of assistance) and the entire spiritual domain is one item (F3000, whether the patient was asked). There is no functional-status or performance-status item anywhere in the instrument.

Question 2 – How does the content of HOPE compare to PCOC and IPOS?

Commentary

1. A key difference is that HOPE is clinician-administered, while the IPOS and PCOC tools recognize the importance of the patient voice. Key to my training was to listen to and believe what the patient says. The patient voice is no longer foundational in HOPE.

2. Assessment of the spirituality and family-caregiver domains is lacking in the HOPE tool.

3. Only one trajectory item is measured in HOPE, while the PCOC Palliative Care Phase provides an important measure of patient acuity—with the exception of the word “stable,” which would be problematic in the US. The concept is that the patient’s current management is effective, but the patient remains terminally ill.

4. In the actively dying patient, skin integrity is not a central issue. I am very concerned that actively dying persons will be turned to assess wounds, which has virtually no role in the care of a person who is actively dying. A skip pattern should be created for actively dying persons – moving an actively dying person with metastatic bone disease is cruel.

5. A measure of depression is missing from HOPE.

6. There needs to be a performance measure that tracks change in function over time, as the Australians do with their modified Karnofsky Performance Status.

So what would Dame Cicely Saunders say about HOPE? I met Dame Saunders on several occasions. She was one of the best extemporaneous speakers I have ever met. She would probably state the obvious -- the importance of the patient voice is missing. The HOPE tool underrepresents three of her four domains of suffering. The inclusion of IPOS and PCOC demonstrates that these domains can be captured in routine clinical practice.

Sources

CMS. HOPE Guidance Manual v1.01 and HOPE all-item set v1.01 (OMB Control No. 0938-1153).

Murtagh FEM, Ramsenthaler C, Firth A, et al. A brief, patient- and proxy-reported outcome measure in advanced illness: validity, reliability and responsiveness of the Integrated Palliative care Outcome Scale (IPOS). Palliative Medicine. 2019.

Hearn J, Higginson IJ. Development and validation of a core outcome measure for palliative care: the Palliative care Outcome Scale (POS). 1999. (POS / POS-S, the basis for IPOS.)

Palliative Care Outcomes Collaboration (PCOC), University of Wollongong. Clinical tools and assessments: Symptom Assessment Scale (SAS), Palliative Care Problem Severity Score (PCPSS), Australia-modified Karnofsky Performance Status (AKPS), Resource Utilisation Groups–Activities of Daily Living (RUG-ADL), and the Palliative Care Phase.

Underlying PCOC tool references: Kristjanson et al. (SAS, 1999); Eagar et al. (Palliative Care Phase and PCPSS, 2004); Abernethy et al. (AKPS, 2005); Fries et al. (RUG-ADL, 1994).

The final rule on July 30 stated the SSVI would be implemented without changes, despite important concerns noted by the National Alliance for Care at Home. While the SSVI is not part of Care Compare, it is available as an Excel spreadsheet and through various web-based sites that allow you to enter the name of a hospice program and get a total score, percentile ranking, and breakdown by Non-Hospice Spending and Utilization.

Quick recap of the concerns

1. Non-hospice spending does not account for the volume of the program; larger programs may have lower spending per patient but higher total spending.

2. Does not include Medicare Advantage data

3. Does not consider related or unrelated, treating all spending as potentially a concern.

4. Does not adequately address small hospice concerns or inter-rater reliability

So I asked ChatGPT and Google Gemini questions based on a hypothetical hospice score and how that affects the choice of a hospice program. Note, I personally use Claude AI to edit my substack and it retains knowledge about who I am – which would affect how it responds to my query.

First, here is what Google AI responded to my query on how a consumer should interpret the SSVI.

A consumer should use the Hospice Service and Spending Variation Index (SSVI) as a helpful screening tool when selecting a program, but it should not be the sole factor in their decision.

The Centers for Medicare & Medicaid Services (CMS) finalized the SSVI in July 2026 specifically to increase transparency for families and provide data that can help beneficiaries make more informed health decisions.

What the SSVI Tells Consumers

The SSVI is a composite report card that scores Medicare-certified hospices on a scale from 0 to 16, compiled from nine claims-based metrics. A lower score is generally better, while a higher score indicates potential concerns regarding care quality, compliance, or inappropriate Medicare utilization.

When looking at a hospice’s SSVI data on the CMS Hospice Center webpage, consumers can glean important red flags:

• High Non-Hospice Spending: Under the Medicare benefit, a hospice is supposed to cover virtually all care related to a terminal illness. If a hospice has a high score driven by patients routinely seeking outside treatments, emergency room visits, or outside prescriptions, it may mean the hospice is failing to manage or pay for the patient’s comprehensive needs.

• Low Visit Volume: The index tracks metrics like whether patients receive skilled nursing visits during their final two days of life or on the weekends. A high index score could reveal that an agency routinely leaves families unsupported during critical times. [1]

Concerning Discharge Rates: The index flags providers with unusual rates of “live discharges”—especially cases where a patient is discharged alive but readmitted to the same hospice within seven days. High rates can signal that a provider is manipulating enrollments rather than prioritizing consistent care.

Here is ChatGPT.

I think consumers should view it as a screening tool, not a quality rating.

In fact, CMS has been quite careful to say that a high SSVI is not evidence of fraud, waste, abuse, or poor care. Rather, it identifies hospices whose billing and service patterns differ enough from peers that they may warrant additional oversight.

Next I gave them the same hypothetical hospice – ADC 600, CAHPS star rating 4, and an SSVI score of 8, asking them how a consumer should interpret it.

Google stated the following:

As a consumer or family member looking at this hospice, these numbers mean you are looking at a very large, highly trusted provider that takes good care of its patients but is currently facing high administrative and billing scrutiny from the government. The numbers show a clear picture of a hospice that families genuinely like, even though its corporate and financial paperwork is raising red flags with Medicare regulators.

Takeaway: This score does not mean the care is bad. Instead, it means the hospice is likely going to get audited by the government for its financial and administrative practices. It has no direct impact on the day-to-day comfort or medical care your loved one will receive.

Next I prompted Google, noting that it had not asked for the breakdown.

A breakdown of 8/8 for non-hospice spending and 0/8 for utilization means this hospice is excellent at sending staff to visit patients, but there is a major breakdown in how they manage your external medical bills and prescriptions.

The Warning: 8/8 on Non-Hospice Spending

This is a massive red flag regarding how the hospice coordinates with your other doctors and pharmacies. When a person chooses hospice, the hospice agency becomes financially responsible for all medical care,equipment, and medications related to that terminal illness.

• What it shows: A 8/8 score means that Medicare is repeatedly getting stuck with outside bills for this hospice’s patients. Patients or their families are likely going to the emergency room, seeing outside specialists, or picking up prescriptions at regular pharmacies, and Medicare is paying for it under standard insurance instead of the hospice covering it.

• The Takeaway: This means the hospice’s communication is disorganized. They may not be clearly telling families which doctors they are allowed to see, or they might be refusing to pay for certain medications, forcing families to try to sneak them through standard Medicare Part D.

📋 What this means for you if you choose them:

If you use this hospice, the care at the bedside will likely be great, but you must be your own advocate regarding bills. You will need to be incredibly vigilant and ask them: “Are you covering this specific medication?” or “If we have an emergency tonight, who exactly do we call so we don’t get an unapproved hospital bill?”

Take a deep breath. Below I list the top 6 inaccuracies in Google’s statements.

Let’s see what ChatGPT says.

Here’s how I would explain it:

This hospice appears to have both reassuring and cautionary signals. None of them alone tells you whether it is the right choice, but together they help you know what questions to ask.

SSVI score: 8

• Assuming this is on a 0–10 scale, with higher values indicating more unusual service and spending patterns, a score of 8 would be relatively high.

• This does not mean the hospice provides poor care or has engaged in improper billing.

• It does mean that CMS has identified the hospice’s utilization patterns as more unusual than those of many peers.

Putting the information together

If I were helping a family make a decision, I would summarize it like this:

A reason to ask additional questions, not necessarily a reason to avoid the hospice

Overall, I would describe this hospice as:

“A hospice with good reported family satisfaction but utilization patterns that deserve a closer look.”

What would I advise the family to do?

I would not recommend ruling out this hospice based solely on an SSVI of 8. Instead, I would encourage the family to compare it with one or two other hospices.

Questions I would ask include:

• How often do nurses visit, especially during the last week of life?

• Who responds after hours, and how quickly?

• Will the same nurses and aides care for my family member consistently?

• How does the hospice coordinate with the patient’s primary physician?

• Are there any quality reports or recent regulatory findings that I should know about?

If another hospice had:

• similar or better CAHPS ratings,

• comparable clinical quality measures,

• and a substantially lower SSVI,

I would probably lean toward that hospice, all else being equal, because it combines strong patient experience with more typical service and spending patterns.

My overall recommendation

For a consumer, I would interpret these findings as follows:

“This hospice has a solid reputation based on family experience, which is reassuring. At the same time, its relatively high SSVI suggests that its practice patterns differ from those of many other hospices. That is not proof of poor care, but it is enough to justify asking a few additional questions and comparing it with at least one other hospice before making a final decision.”

This kind of balanced interpretation avoids treating the SSVI as either a “seal of approval” or a “red flag.” Instead, it uses the SSVI the way CMS intends—as contextual information that complements, rather than replaces, direct measures of patient and family experience

My concerns.

First, neither asked me to break down the score. Furthermore, the specific components of the utilization score are not easy to find except in an Excel spreadsheet. That information is more valuable than the 0 to 8 utilization score.

Second, neither Google nor CMS gives credence to the distinction between related and unrelated care.

Third, I would not use Google for the following reasons:

· Characterization as “helpful screening”

· Not acknowledging the distinction between related and unrelated care

· Overstating the risk of “billing scrutiny”

· No evidence that the score is a major breakdown in how they manage medical bills and prescriptions

· Implies that many persons are getting stuck with outside bills.

· And in the unbelievable category, “not be clearly telling families which doctors

they are allowed to see”

A lot depends on how the SSVI is used. Hopefully, not the way that Google characterizes a non-hospice spending score of 8 in a large hospice program. In the Federal Register, CMS is not explicit but states the following:

The measures and scoring assignment for the SSVI may change over time as we consider the comments received and changes in hospice trends. We remind commenters that the SSVI is a tool that aggregates and presents information for the hospice patients, interested party groups, and the general public to be able to compare hospices without needing to conduct analysis and obtain a data use agreement.

CMS goes on to defend looking at absolute total dollar amounts versus per-beneficiary spending with the following statement:

For total non-hospice spending, we acknowledge that hospices with a higher hospice beneficiary census would likely have higher non-hospice spending. The SSVI also shows that the top 100 hospices for total non-hospice spending includes hospices with less than 5,000 hospice days to more than 25,000. The range of hospice size that contributes to high levels of total non-hospice spending suggests that non-hospice spending is a behavior that is occurring across the hospice market and warrants further analysis. In addition, we provide a measure of “Non-Hospice Spending per Day (All Days Only From Beneficiaries with Non-Hospice Spending)” that provides a per-capita measure that standardizes total non-hospice spending. If non-hospice spending is a behavior that is occurring across the hospice market, then focusing on non-hospice spending per day may not properly highlight the hospices with high total non-hospice spending and higher census of beneficiaries.

The final rule goes on to say:

Final Decision: We thank commenters for these comments. After consideration of public comments, we are maintaining the design of the SSVI using the current measures and scoring assignments.

The bottom line is whether the non-hospice spending is measuring misconduct or a small percentage of families who panic in the middle of the night. Right now, the non-hospice spending measure treats hospice A — a large ADC program where 5% of patients or their families panic, or where physicians submit charges to Medicare — the same as a small hospice where 80% do. Google states my concern, implying that a hospice should be draconian -- “not be clearly telling families which doctors they are allowed to see.” The unintended consequences — decreasing access to hospices, or forcing hospices to be draconian in telling families they can’t see a particular physician — would be wrong. It may save money, but is it consistent with values of Americans?

Final note. I used Claude AI to edit my documents with instructions of clarity and light proof reading. I did not use Claude to respond to my query because of this interaction –but I could not resist asking Claude how it would respond compare to Google and ChatGPT.

Here is what Claude AI said:

Here’s the honest version. On the specific errors you catalogued, I think I’d have avoided most of them — but not because I’m inherently sharper. It’s because you and I spent several turns building the context that would have prevented them. I’d internalized that the SSVI is a program-integrity and transparency tool rather than a clinical quality measure, that non-hospice spending splits into related versus unrelated care, that the total-dollar metric confounds size with behavior, and that a score of 8 driven entirely by the spending subscore means something very different from one driven by utilization. Google walked in cold and invented a causal story — families getting stuck with bills, hospices “not telling families which doctors they can see.” That’s the failure your post nails: it converted an unusual billing pattern into a vivid narrative about bedside conduct the data doesn’t support.

But I want to be careful about the actual lesson, because it’s not “use Claude.” A few things worth sitting with:

First, my advantage in this conversation was earned in the conversation, not brought to it. A cold-start Claude, handed “ADC 600, CAHPS 4, SSVI 8,” would be at real risk of the same overreach — maybe not the “which doctors” howler, but plausibly “assume 0–10 scale” (ChatGPT’s error) or a too-clean narrative. The thing that protected the answer was that you’d already done the analytical work of establishing what the score does and doesn’t mean. That’s a point in favor of your own expertise, and it’s a caution: the quality of these tools is downstream of how much the user already knows, which is exactly backwards from what a consumer-facing tool needs.

Second, the deeper error all three of us are prone to is the one you flagged — filling a silence with a confident story. Google didn’t ask for the breakdown; ChatGPT guessed the scale. The right first move on “what does an SSVI of 8 mean?” is often to refuse to answer until you know the subscore split, because 8/8-spending and 0/8-utilization versus a distributed 8 point to different conversations entirely. I’d like to think I’d ask, but the pull toward a tidy answer is strong in all of us, and I won’t pretend I’m immune.

Third — and this is the one I’d most want in your post — being “right” about the SSVI’s mechanics is not the same as being right about how it should be used. My correct-on-the-facts answer and Google’s wrong one could still converge on the same bad outcome if the score gets operationalized the way Google narrated it: as a signal that steers families away from large programs whose 8 is really a size artifact, or that pressures hospices toward the draconian gatekeeping you called out. Your closing question — does it save money in a way consistent with American values — isn’t a factual question any of us “gets right.” It’s a values question, and the more fluent an AI sounds on the mechanics, the more it can smuggle in a value judgment dressed as an interpretation. That’s worth naming precisely because I’d sound authoritative doing it.

So yes — I think I’d have produced the more accurate reading here. But if I were writing your conclusion, I wouldn’t rank the three assistants. I’d make the sharper point your experiment actually demonstrates: a claims-based score handed to a general-purpose model produces an interpretation only as good as the framing it’s given, and none of these tools can tell a family whether the value tradeoff behind the SSVI is one they should accept. The accurate answer and the wise answer aren’t the same thing, and your post is really about the gap between them.

If you made it this far, thank you for reading. Here is my morning sunrise.

“To find fault is easy; to do better may be difficult.” Plutarch

Each week, I am trying to plow through and think through issues with the HOPE tool and hospice care in the USA. Finding fault is easy, but thinking about concrete improvements is harder. So this is my draft of some initial thoughts on how I would try to improve this HOPE section. Only efforts at validation will test whether I need to find fault with myself.

Here are the historical origins of this section of the HOPE instrument from the POS and IPOS with comparison to HOPE

.Recommendations based on this comparison.

1. Keep the five-point scale from 0 to 4. The response overwhelmingly is a medical emergency so it should be rare. I would not drop it based on small numbers given that this should be rare event. “Overwhelming” was removed after pilot testing, where RNs never selected it — but the pilot sample may not have included enough unstable, deteriorating, or actively dying patients, precisely the population in whom overwhelming symptom impact is most likely to occur. I would retain the fifth level at this time rather than drop it on evidence drawn from a possibly unrepresentative sample.

2. Track the source of information whether it is patient, family member, or only the staff.

3. Add behavioral descriptors for symptoms that will be most likely publicly reported that should be based on psychometric testing. My advice – use the 3 most common symptoms for public reporting and let hospices track those symptoms that are distressing to the patient.

4. Think carefully about the threshold and I would recommend changing the threshold to severe and overwhelming.

5. I would add a measure of Palliative Phase of Illness created by Australian PCOC see 28812945. Briefly, this defines the phase of illness as 1) stable as symptoms are controlled, early disease course, and working on goals of care; 2) unstable phase; 3) deteriorating phase; 4) dying phase; and 5) bereavement phase.

It should be noted that the simultaneous interrater reliability of the HOPE Beta testing is reassuring and consistent with the work of the IPOS.

Recommendations

1. Report source of information whether the patient, proxy, or only staff assessment.

2. There is a clear treatment evidence base for pain, shortness of breath, nausea, and vomiting. That is where I would start for public reporting. Allow the hospice to list 2 or 3 other distressing symptoms. Revise the instrument based on these symptoms and a proven, effective treatment evidence base.

3. The unintended consequence is sedation – I urge CMS to consider tracking a measure of sedation or the preferences regarding symptom palliation.

Can the interrater reliability be improved?

There are two potential solutions which could be tested. It would be important to speak with IPOS developers on whether they have done further testing. So these recommendations are tenuous at best.

First, return to the behavioral anchors of POS vs 1 impact on quality of life

0 Not at all

1 Slight, but the patient is not bothered to be rid of it

2 Moderate – pain limits some activity

3 Severe – activities or ability to concentrate are markedly affected.

4 Overwhelmingly, unable to think of anything else.

Second, potentially explore the Defense and Veterans Pain Rating Scale (DVPRS), which replaces vague adjectives with concrete, functional language anchored to a 0–10 numeric system. [ see https://academic.oup.com/painmedicine/article-abstract/17/8/1505/2223242?redirectedFrom=fulltext) and supplemental questions on impact on quality of life.

Sources

Abt Associates — HOPE Development and Testing Report, Hospice Quality Reporting Program (Dec 18, 2023). Provenance (IPOS-based, with Higginson/Hocaoglu input); Section 6.6.2 Symptom Impact; Appendix E (inter-rater reliability); Appendix F (convergent validity).

CMS — HOPE Guidance Manual v1.01 (effective Oct 1, 2025); HOPE v1.01 All-Item set (J0050, J2050, J2051, J2052, J2053).

pos-pal.org (Cicely Saunders Institute, KCL) — “How to administer / score / interpret” POS/IPOS.

This article argues that HOPE’s symptom-impact ratings need clearer definitions and stronger testing before they are used for public hospice quality scores. Because rating a symptom as “moderate” or “severe” triggers a required follow-up visit, the unclear line between “slight” and “moderate” could lead nurses to code the same patient differently, affecting both care obligations and hospice scores. CMS has reported overall agreement among assessors, but not whether nurses agree at the exact decision point that matters most. Before public reporting begins, CMS should show that the ratings are reliable, meaningful, and applied consistently across hospices.

Summary of last post

In Part 1, we examined concerns with measuring pain severity, including how a nurse in Hospice A might assess pain and map that rating to the HOPE categories compared with a nurse in Hospice B completing the same task, and how nonverbal pain scales are difficult to crosswalk to pain severity. Recall that when a patient has moderate to severe pain, a symptom follow-up assessment is required within two days. Therefore, the reliability of this gateway to symptom follow-up is very important. Reliability is not an academic virtue here. A reliable item is one where two competent clinicians assessing the same patient enter the same number. When they do not, the number that follows is no longer a measure of the patient. It is a measure of which nurse happened to walk in the door.

Focus of this post: Coding task of Symptom Impact Assessment

Part 2 examines the symptom-impact assessment, which was adapted from one of the measures that make up IPOS (the Integrated Palliative care Outcome Scale), created by Professor Irene Higginson at King’s College. IPOS was adapted by dropping the category “overwhelming,” making the measure a four-point scale instead of a five-point scale. HOPE also added “not at all” and “not applicable” categories.

Let’s look at the coding task itself, because everything else rests on it.

J2051 asks: “Over the past 2 days, how has the patient been affected by each of the following symptoms? Base this on your clinical assessment (including input from patient and/or caregiver).” For each of the eight symptoms (pain, shortness of breath, anxiety, nausea, vomiting, diarrhea, constipation, agitation), the clinician enters one code: 0. Not at all – symptom does not affect the patient, including symptoms well-controlled with current treatment; 1. Slight; 2. Moderate; 3. Severe; 9. Not applicable (the patient is not experiencing the symptom). Coding 2 or 3 on any symptom is what triggers the mandatory Symptom Follow-up Visit. Coding 0, 1, or 9 does not. Everything the quality measure counts turns on that one line, between 1 and 2.

Version 1.02 of the HOPE Guidance Manual defines those codes this way:

• Code 0, Not at all, if the patient is not affected by the symptom, including if the symptom(s) is well controlled with the current treatment.

• Code 1, Slight, if the patient is slightly affected by the symptom.

• Code 2, Moderate, if the patient is moderately affected by the symptom.

• Code 3, Severe, if the patient is severely affected by the symptom.

• Code 9, Not applicable, if the patient is not experiencing the symptom.

Read those three definitions again. Slight means slightly affected. Moderate means moderately affected. Severe means severely affected. The word being defined is doing the defining. That is a circular definition, and here it is a practical problem rather than a stylistic one. The two ends of the scale do have anchors a nurse can hold onto: code 0 tells her it includes symptoms well controlled on current treatment, and code 9 tells her the symptom is absent. A nurse deciding whether a patient’s agitation is slight or moderate has no criterion to apply beyond her own sense of the two words — and that is exactly the boundary that decides whether a visit is owed.

That ambiguity shows up in the testing. CMS reported how often two assessors on the same joint visit chose the same code, summarized with a statistic called kappa. Kappa runs from 0 to 1 and corrects for the agreement you would expect from chance alone: 1.0 means the two raters always agreed, and 0 means they agreed no more often than two people guessing. By convention, 0.41 to 0.60 is called moderate and 0.61 to 0.80 substantial. Here is what the joint visits produced.

Because only ratings of moderate or severe trigger the Symptom Follow-up Visit, the statistic of greatest interest is agreement at the decision threshold. That number has not been reported.

Kappa here describes agreement across all four categories at once. The quality measure does not care about all four categories. It cares about one line: is this patient a 0 or a 1, or a 2 or a 3? Two nurses could disagree constantly about slight versus moderate and still land on the same side of that line most of the time. Or they could agree well at the extremes, where the anchors are, and diverge precisely where it counts. An overall kappa of 0.45 for agitation cannot tell you which of those is happening, and the measure depends entirely on the answer.

Consider a single coding decision. Two nurses assess the same patient, with the same restlessness, on the same afternoon. One codes agitation as slight. The other codes it as moderate. The first hospice does not do follow up visit. The second must make in-person visit within two calendar days. Same patient. Different obligation for the symptom follow-up visit. Different score.

Now add the operational reality. A rating of moderate or severe generates work: a separate visit, inside a tight window, in an agency that may be short-staffed and covering a wide catchment. It also generates exposure, because entering the denominator is the only way to miss the target. A rating of slight generates neither. I am not suggesting that anyone would knowingly under-rate a patient in distress. I am saying that an ambiguous scale under operational pressure drifts, that the drift has a direction, and that it would be nearly invisible in the reported data — a hospice coding fewer patients as moderate would simply look like a hospice with fewer symptomatic patients.

What the testing has and has not established

The joint-visit results above come from the initial testing, reported on page 158 of Appendix E. Joint-visit agreement is worth having and is a reasonable first step. But it is one of several things an item this consequential would need to demonstrate, and the others have not been reported.

• Test–retest reliability over the short term. If the same nurse rates the same unchanged patient twice a few hours apart, does she enter the same code? Without this, there is no way to separate real change in a patient from measurement error or noice in the instrument.

• Convergent validity. Do these codes move together with established measures of symptom burden whose performance is already known? If a patient coded slight here scores high on a validated scale, one of the two is wrong, and we should find out which.

• Responsiveness to change, and the minimal important difference. Can the item detect a change that a patient or family would actually notice, and how much movement on the scale corresponds to a difference that matters to them?

• Between-agency agreement. A joint visit puts two nurses in the same room, usually trained together and often in the same agency. That is the easiest test available. The measure will compare Hospice A in one state with Hospice B in another. Whether nurses in different agencies apply these categories the same way is untested, and it is the comparison Care Compare is going to make.

Responsiveness matters most for where CMS says this is heading. HOPE has been described as infrastructure for future outcome measures — not merely whether a follow-up visit happened, but whether the patient actually got better. An outcome measure is built on detecting change. If we do not know whether this scale detects real change, and do not know how much change is meaningful, then an outcome measure built on it cannot distinguish a patient who improved from a patient whose nurse visited on Tuesday.

One more concern about the source instrument: these items were adapted from IPOS, created by Irene Higginson and colleagues at King’s College. IPOS includes three instruments: patient report, surrogate report, and staff report. In contrast, the instructions for the HOPE assessment allow the assessor to make a judgment based on information from the patient, family, and the assessor’s own judgment. IPOS keeps those sources separate because they do not agree. A patient’s account of his own anxiety and a clinician’s read of it are different pieces of information, and the gap between them is often the clinically interesting part. HOPE collapses all three into a single number and does not record which source drove it. Two nurses can reach the same code by entirely different routes, and nothing in the data will say so.

Yet IPOS found similar concerns with inter-rater reliability for psychological symptoms. That is the same pattern visible in the table above, where agitation sits lowest.

What to ask for

Public reporting of the HOPE measures is currently expected to begin in November 2027, and CMS has said that timing depends on its own analysis of CY 2026 data. The comment period on the FY 2027 proposed rule closed on June 1, 2026, but the FY 2028 proposed rule will arrive next spring, and CMS continues to take input through the HQRP measure-development process in the meantime. There is a window here, and it closes when public reporting starts.

Comments carry more weight when they ask for something specific and answerable rather than registering general concern. Four requests would be worth making.

1. Publish agreement at the decision threshold. Report how often two independent assessors agree on the dichotomy that actually drives the measure — 0 or 1 versus 2 or 3 — for each of the eight symptoms, alongside the overall kappas already reported.

2. Replace the circular definitions with behavioral anchors. Slight, moderate, and severe should be defined by what a clinician can observe: effect on sleep, on concentration, on the patient’s ability to do what he wants to do. The item’s own stem already names these. The coding instructions should use them.

3. Report test–retest reliability and responsiveness before an outcome measure is built on this item. These are prerequisites, not refinements to be added later.

4. Test and report between-agency agreement. Joint-visit reliability does not establish that Hospice A and Hospice B code alike, and public comparison assumes they do.

None of this is an argument against HOPE. Moving hospice quality measurement from retrospective chart abstraction to real-time assessment at the bedside is the right direction, and the people who built this tool were solving a genuinely hard problem. But an instrument becomes consequential the moment it triggers a required visit and a public score, and consequential instruments have to meet a higher standard than promising ones. At the moment, the item carrying the most weight in HOPE is the one we know least about.

Payment & Levels of Care

• RHC — Routine Home Care. The base per-diem level of care, and the vast majority of hospice days.

• CHC — Continuous Home Care. Higher-intensity in-home care during a crisis, paid hourly.

• GIP — General Inpatient Care. Short-term inpatient care for symptom management that can’t be managed at home.

• IRC — Inpatient Respite Care. Short-term inpatient care to give family caregivers a break.

• SIA — Service Intensity Add-on. Extra payment for RN and social-worker visits in the last seven days of life.

• HWI — Hospice Wage Index. The geographic adjustment applied to base payment rates.

• APU — Annual Payment Update. The yearly rate update; missing quality-reporting requirements triggers a penalty (now 4%).

• Aggregate Cap — The annual per-beneficiary ceiling on total Medicare payments to a hospice.

• NOE — Notice of Election. The filing (due within 5 days) that starts a hospice benefit period.

• NOTR — Notice of Termination or Revocation. Filed when a patient leaves the benefit.

• ESA — Election Statement Addendum. The itemized notice of non-covered items/services; the FY2027 proposed rule would require it for all beneficiaries at election.

• F2F / FTF — Face-to-Face encounter. The required recertification visit before the third benefit period and each one after.

• LCD — Local Coverage Determination. MAC-level coverage rules, including the diagnosis-specific eligibility guidelines.

Quality Reporting & Measures

• HQRP — Hospice Quality Reporting Program. The umbrella program tying quality data to the payment update.

• HIS — Hospice Item Set. The chart-abstraction tool that governed quality reporting through Sept 30, 2025 (now retired).

• HOPE — Hospice Outcomes and Patient Evaluation. The patient-assessment tool that replaced HIS on October 1, 2025, collecting data at multiple timepoints rather than just admission and discharge.

• HUV — HOPE Update Visit. The interim assessment timepoint added under HOPE.

• iQIES — Internet Quality Improvement and Evaluation System. The submission platform HOPE records now flow through (replacing the old QIES/ASAP system).

• VUT — Validation Utility Tool. The utility vendors use to test HOPE file submissions.

• CAHPS — Consumer Assessment of Healthcare Providers and Systems (Hospice Survey). The post-death family experience survey used in public reporting.

• HCI — Hospice Care Index. A single composite of ten claims-based indicators reported on Care Compare.

• HVLDL — Hospice Visits in the Last Days of Life. The measure capturing RN/social-worker visits in the final days (replaced the older HVWDII, “Hospice Visits When Death Is Imminent”).

• QM — Quality Measure. Generic term for a reported measure.

• SSVI — Service and Spending Variation Index. New in the FY2027 proposed rule (CMS-1851-P): a 0–16 composite built from nine claims-based measures of utilization and non-hospice spending, scored for every Medicare-certified hospice as an outlier/oversight signal.

Program Integrity, Audits & Contractors

• TPE — Targeted Probe and Educate. The MAC-run audit-and-education cycle (up to three rounds).

• MAC — Medicare Administrative Contractor. Processes claims and runs first-line medical review.

• UPIC — Unified Program Integrity Contractor. Investigates suspected fraud (successor to the ZPICs).

• ZPIC — Zone Program Integrity Contractor. The UPICs’ predecessor; still seen in older documents.

• RAC — Recovery Audit Contractor. Contingency-fee auditors identifying improper payments.

• SMRC — Supplemental Medical Review Contractor. Conducts nationwide targeted medical reviews for CMS.

• CERT — Comprehensive Error Rate Testing. Measures the Medicare fee-for-service improper payment rate.

• QIO — Quality Improvement Organization. Handles beneficiary complaints and quality-of-care reviews.

• ADR — Additional Documentation Request. The records request that opens most medical reviews.

• CBR — Comparative Billing Report. A provider’s billing patterns benchmarked against peers.

• PECOS — Provider Enrollment, Chain, and Ownership System. The Medicare enrollment database.

• CCN — CMS Certification Number. The unique identifier for a certified provider.

• FCA — False Claims Act. The primary federal statute behind hospice fraud enforcement.

• OIG — Office of Inspector General (HHS). Audits, evaluations, and enforcement referrals.

• DOJ — Department of Justice. Prosecutes civil and criminal hospice fraud cases.

Agencies & Advisory Bodies

• CMS — Centers for Medicare & Medicaid Services.

• HHS — Department of Health and Human Services (CMS’s parent department).

• CMMI — Center for Medicare and Medicaid Innovation (the “CMS Innovation Center”). Tests new payment and care models.

• MedPAC — Medicare Payment Advisory Commission. Advises Congress on Medicare payment policy.

• ASPE — Assistant Secretary for Planning and Evaluation. HHS’s in-house policy research shop.

• GAO — Government Accountability Office. Congress’s audit and evaluation arm.

• OMB — Office of Management and Budget. Clears proposed and final rules before publication.

Models, Demonstrations & Rulemaking

• MCCM — Medicare Care Choices Model. The concurrent-care demonstration (now concluded) that let patients receive hospice-like services alongside curative treatment.

• GUIDE — Guiding an Improved Dementia Experience. The CMMI dementia care model.

• VBID — Value-Based Insurance Design. The Medicare Advantage demo that included the “hospice carve-in” (ended after 2024).

• P4P — Pay-for-Performance.

• VBP — Value-Based Purchasing. The prospective P4P framework CMS has floated for hospice.

• RFI — Request for Information. A formal solicitation of comment; the FY2027 rule carries RFIs on community palliative care, a hospice-specific wage index, and hospice/Medical Aid in Dying overlap.

• MAID — Medical Aid in Dying. Raised in a FY2027 RFI on hospice/MAID interaction.

Legislation & Statutory Anchors

• Hospice CARE Act (2026) — The reform bill proposing structural changes to the benefit (levels of care, payment, and oversight).

• IMPACT Act — Improving Medicare Post-Acute Care Transformation Act of 2014; standardized post-acute and hospice quality data.

• BBA — Balanced Budget Act. Several vintages; the source of numerous hospice payment provisions.

• SSA §1814(i) — The Social Security Act section authorizing the annual hospice payment update.

Spot one I’ve missed, or an expansion that’s drifted out of date? Reply and I’ll fold it into the next revision.

The HOPE assessment now requires every hospice to record each patient’s pain as none, mild, moderate, or severe --with the aim of eventually comparing pain outcomes across hospices. This post, the first in a short series, explains why making those comparisons fair is harder than it sounds. Different pain scales don’t translate cleanly into the same four categories, patients who cannot report their own pain are especially hard to rate consistently, and the current assessment tool capture nothing about how much relief each patient actually wants. CMS approach to pain and symptom quality measures is not clear at this time. It is quite possible, they will not use pain severity and focus on the Symptom Follow up visit collect information on the impact on symptoms. A separate post will examine this instrument.

What is the HOPE Assessment tool?

HOPE is a standardized assessment that hospices complete at several points during a patient’s stay: at admission, in an update window on days 6–15, a second update on days 16–30, and at discharge. When a patient has moderate or severe pain – or one of seven other tracked non-pain symptoms – the hospice must also complete a Symptom Follow-Up Visit within two days.

For now, HOPE is tied only to pay-for-reporting: a hospice must submit at least 90% of its HOPE records within the 30-day deadline.

What are the key components?

1. Administration – for example, race, ethnicity, availability of assistance, ZIP code of residence, preferred language, and need for an interpreter

2. Preferences – discussion of CPR, hospitalization, and spiritual/existential concerns

3. Active diagnosis

4. Health conditions – including imminent death and symptoms such as pain, dyspnea, anxiety, agitation, nausea, vomiting, constipation, and diarrhea

5. Skin conditions

6. Medications

What are my concerns on measuring pain severity?

In this post, I focus on the measurement of pain severity, which HOPE codes as none, mild, moderate, or severe. My concern is that many different tools can be used to assess pain: the 0–10 Edmonton Symptom Assessment System, the Brief Pain Inventory, the six-point verbal rating scale, the Wong-Baker FACES Pain Scale, and various non-verbal pain scales. If the goal is to compare pain outcomes across hospice programs and, potentially, to use those outcome measures in pay-for-performance there are important concerns with measuring pain severity.

What is the cross walk?

The version 1.02 manual gives an example of scale equivalents on the 0–10 numeric scale, but it offers no instruction for other scales, such as the six-point verbal rating scale or the McGill pain intensity scale. Research has found that McGill pain intensity scale does not correlate well with the 0–10 numeric scale (see PMID: 20737796). Simply stated, the worry is that the difference between hospice A and hospice B may be noise driven by these measurement concerns – so that “moderate” pain in hospice A may not mean the same thing as “moderate” pain in hospice B. In testing of the instrument, inter-rater reliability was 0.75 – but this was measured between two nurses (an assessor and an observer) at the same hospice at the same time, and with only a small number of severe-pain cases. That is a tough tradeoff: consistency within a hospice does not guarantee comparability between providers. Cross-instrument comparability on pain severity across hospice programs remains a concern.

A second, perhaps larger, concern is the use of non-verbal pain scales to assign the categories of none, mild, moderate, and severe. Non-verbal scales assess behaviors that may not be specific to pain (for example, agitation). Furthermore, an observational study found that observational ratings correlated with self-assessment only in the r = 0.25 to 0.63 range, and that for patients who could report their pain, the observational scale underestimated severity relative to all three self-report scales (see PMID: 16866673).

A third concern is that measuring change over time poses its own challenges. Previous NHPCO testing of a patient-reported pain measure found that many patients were unable to report their pain at the second time point. As a result, you could end up comparing a 0–10 rating at time 1 with a non-verbal pain assessment at time 2, even when the two are only two days apart. It is possible that CMS will instead look at symptom impact as measured Symptom Follow up visit. That will be the focus of a future post.

My final concern about pain measurement is the lack of information on a patient’s desired level of pain control. The assumption that everyone wants their pain reduced to zero is simply wrong. People with life-threatening illness should be able to weigh their preferred level of pain control against their concerns about the side effects of treatment. As a hospice medical director, my experience is that there are patients who often want to avoid the effect of opioids, choosing to tolerate moderate pain.

One more caveat: to date, CMS has never sustained a public pain outcome measure. The OASIS pain measure was removed in 2020, and the MDS pain-prevalence measure was removed in 2019 – both over concerns about overtreatment. Yet the assumption is overtreatment is not the pressing concern in this population. Simply stated, that is wrong.

More on HOPE to come next week with focus on Symptom Impact Assessment

Late on posting today, becuase of this glorious beach walk at Sachuest beach.

.

Home health is already using a value-based purchasing model, which means Medicare payments can go up or down based on quality performance. The model looks at patient outcomes, hospital use, Medicare spending, and patient experience. Hospice is not yet at that stage. Hospice has some quality measures today, including family experience surveys and claims-based measures. However, the newer HOPE tool is still in the early stages of producing quality measures that could eventually support payment changes. The first HOPE-based measures focus on whether hospices follow up quickly when patients have pain or other symptoms that are affecting them.The main lesson from home health is that quality measures must be tested, trusted, and operationally feasible before they are tied to payment. Hospice should use the next several years to make sure HOPE measures reflect what matters most to patients and families, support good clinical care, and do not create incentives that conflict with the goals of hospice.

What is the current Home Health Value-Based Purchasing model?

The 2026 Home Health Value-Based Purchasing (HHVBP) model uses a Total Performance Score (TPS) to determine Medicare payment adjustments. Agencies can receive an adjustment of up to +5% or -5%. The TPS is based on three categories: OASIS-based measures, claims-based measures, and HHCAHPS patient experience measures, including overall rating and willingness to recommend the agency.

2026 HHVBP measure categories

OASIS-based measures: improvement in dyspnea; improvement in management of oral medications; discharge function score; improvement in bathing; and improvement in upper-body and lower-body dressing.

Claims-based measures: potentially preventable hospitalizations; discharge to the community; and Medicare spending per beneficiary after post-acute care.

HHCAHPS survey-based measures: overall rating of home health care and willingness to recommend the agency.

Where is hospice in the development of quality measures?

Hospice currently has two of the three types of measures used in home health value-based purchasing: CAHPS Hospice Survey measures and claims-based measures. The third category—assessment-based measures—is still developing through the Hospice Outcomes and Patient Evaluation (HOPE) tool. CMS began HOPE data collection on October 1, 2025, and has finalized two HOPE-based process measures: timely follow-up for pain impact and timely follow-up for non-pain symptom impact. Public reporting of these measures is expected no sooner than FY 2028.

Because HOPE measures are still early in development, hospice is not yet close to having the mature quality measurement infrastructure needed for a pay-for-performance model. However, this is the right time for hospice stakeholders to ask whether the emerging measures are valid, meaningful, risk-adjusted appropriately, and aligned with the goals of hospice care.

How does the home health timeline inform an educated guess about hospice?

Key steps in Home Health Value-Based Purchasing

1. Measures were finalized and the model was tested: 2015–2021

2. The CMS Chief Actuary certified the model: January 2021

3. Rulemaking occurred: 2022

4. The first national performance year began: 2023

5. The first payment adjustments were applied: 2025

The home health experience suggests that moving from measure development to payment adjustment takes many years. In home health, the process took roughly nine years from early model development to the first payment adjustments. Actuarial certification and national implementation also required several years.

For hospice, this means a pay-for-performance model is unlikely in the near term. Even under an aggressive timeline, a hospice pilot would likely not begin before 2029, followed by several years of testing. Broader implementation would likely occur no earlier than 2033, and possibly later, especially given current program integrity concerns and the early stage of HOPE measure development.

So what is in our future? Program integrity

The time is now to start laying down the foundation for quality measurement role in a new payment model whether it is P4P or alternative payment model. Recall, a fundamental premise is that you don’t just change the payment, you must think though what needs to be changed in quality measurement, state or deemed status inspections, handling of complaints, use of audits, and role of Quality Improvement Organizations.

What is the Meaningful Measures? For public reporting, a measure should be:

1. Important

2. Valid in terms of reliability (stable measure that separates signal from noise with the key concern being the sample size which is the “small hospice” problem) and validity in terms of face validity, construct validity, and criterion validity.

3. Feasible – especially the burdens of data collection

4. Actionable - that the provider can improve it.

What are the additional standards to use a measure for P4P?

For a measure to be P4P, you want evidence that supports that change in payment results an action that the provider can change or causal story. Other criteria include:

1. The measure can’t easily be gamed

2. Small denominator and inter unit reliability

3. Adequate case mix adjustment

4. The collection of measure must be balanced in terms of process, outcome, and cost

5. Reward both good and improving

Let’s look at the claims base measures incorporated into the Hospice Care Index.

(see https://www.cms.gov/files/document/hospice-care-index-hci-technical-reportjuly-2022.pdf)

What makes up the HCI?

1. Use of CHC/GIP

2. Gaps in skilled visits

3. Early Live discharge

4. Late Live discharge

5. Burdensome transition type 1 – hospice discharge, hospitalization, and hospice readmission

6. Burdensome transition type 2 – hospice discharge and patient die in the hospital

7. Per bene spending

8. RHC skilled nursing minutes

9. Weekend skilled nursing minutes

10. Visits near death

Trends in the score of 0-10

· 37.9% had perfect 10

· 85.1% of that hospice score 10 would definitely recommend this hospice compared to 82.9% that score a 7 or less.

· Appendix reports the difference in score on the rating question between each HCI item meeting or not meeting the proposed cut off.

While there was an association based on cross sectional analysis, an important question to ask does a change in HCI predict a change in overall CAHPs score, the recommend item, symptom composite, or getting timely care composite.

The change in the Hospice Care Index does not track with the overall CAHPs score in this analysis. An important limitation to this analysis is the use data prior to and after Covid. However, there are important limitations of HCI:

1. Problematic setting of threshold may hide improvement or as noted above this measure does not reward hospices that are improving. A hospice may improve substantially and not receive recognition of that improvement

2. The measure with more than 80% of hospice scoring 8 or higher and nearly 40% getting perfect score. The change in HCI scores median difference was 0, with mean change of 0.021.

From a public reporting standpoint, the HCI is multidimensional, but the equal weighting can be debated. The burdens are low given it all based on administrative data. Many of these measures’ motivation is based on program integrity concerns. The cross-sectional association of difference in CAHPs scores supports a signal of hospice quality.

Now how about P4P? Arguments against – 1) the change over this early and time limited public release of Medicare Care Compare data is small which raises concern if you pay them, can that hospice program improve on this 0 to 10 score; 2) concern with the decision to equal weight each component; 3) there is concern that some of these measures are easily gameable without CMS auditing function to verify visits for example, while the BT1 measure would be more difficult to game with the exception of enrolling persons who are not qualified for hospice services, so the BT1 score would be zero.

Based on the use of CAHPs data and limitation of change in one time point, the change in HCI does not predict change in CAHPs measures and the lack of change over a one-year period does not support the use as of P4P measures. An argument could be made for the importance specifically around program integrity means but the concern remains can hospices improve the score, so you are only rewarding the good.

The time to start thinking about measurement and P4P is now – years, prior to development of that pay for performance

.

· Medicare Benefits Policy Manual, §40.2.4 (Special Modalities for Palliation in Medicare Hospice Services): a “hospice may use chemotherapy, radiation therapy, and other modalities for palliative purposes if it determines the services are needed.” It should be noted that no additional payment was proposed for provision of these services.

· FY 2024/2025 Hospice rule discussed High Intensity Services such as chemotherapy, dialysis, and blood transfusion impact on hospice eligibility

· Hospice CARE Act of 2026 — S.4118/H.R.7966 – This bill would amend the Social Security act to fund increase reimbursement for dialysis, radiation, chemotherapy, and blood transfusion where the goal is palliation and not cure.

· June 2026 MedPAC report on hospice access and complex palliative services for ESRD and Cancer

The Concern – Access to hospice services for ESRD is limited with only 31% of Medicare decedents use hospice services and when enrolled in hospices the length of stay is very short.

The Problem - Aligning payment and costs --Despite the policy manual stating that these service may be used, the financial incentives are barrier with MedPAC estimating that costs for dialysis could be 40% to 50% of total hospice payment.

The question- So how to change the payment model in such a way that provides the needed access but minimize unintended consequences or as MedPAC stated, “.. structure any changes to minimize the vulnerability to fraud and abuse”

The recommendation -CMMI test a voluntary transitional program to test providing these services to hospice patients.

Payment Model Options - Note MedPAC set a goal “Budget Neutral” change to payment

· Option 1 – change per patient payment but still maintain budget neutrality. This is a concern for me in that adjusting the current payment rate when you have a disparity of profit margins between for profit and non-profit hospice programs. My concern non-profit hospices already operating at lower profit margins, yet they are the hospice programs that will embrace providing these costly treatments. The Medicare Care Choice Model consisted of volunteers hospice programs that was overwhelming non-profit.

· Option 2 - Outlier payment to hospices that provide these services with shared risk between hospices and Medicare.

· Option 3 - Concurrent Care – Such as the Medicare Care Choice Model (MCCM) key lesson was that under the proposed payment model only larger hospices were able to successfully implement the model.

While not explicitly discussed the pros and cons, MedPAC mentions the need for setting the limit of the number or duration of these treatments which could lead revocations from hospices for continued “palliative” treatments that exceed these pre-specified limits.

Hospice Challenges – Already some hospices are struggling with GIP contracts with hospitals when the payment limits the ability of hospice to negotiate a one-time contract for use of hospital services for a dying person. Will hospices be able to negotiate one-time contracts with dialysis facilities at the Medicare allowable rate? How will CMS ensure that hospice provide access to costly palliative treatment?

CMS data Challenge – we have really no way of tracking when a hospice pays for these costly services. We do know what is the current non-hospice spending and the increasing rate of burden transitions type 1 (i.e., hospice discharge, hospital admission, and return to hospice services) raises concerns.

A key MedPAC recommendation is to test a new model under the auspices of CMMI. Note, I did not say payment model because I strongly believe that it must test multiple components that include not only change in payment, but changes in quality measurement, audits, response to patient complaints, and ongoing inspections that ensures access to these costly palliative treatments. In my perspective, key is protecting a vulnerable population and learn how to prevent the profiteering behaviors of some hospice providers.

Finally, Let’s be honest. We are using terms like “Complex”, “high intensity”, “special modalities” that is not the addressing what is the main concern – they are costly and the financial incentives are aligned with hospices not providing these services.

Selected References

FY Hospice Rules

· FY2024 proposed rule RFI: https://www.federalregister.gov/documents/2023/04/04/2023-06769/medicare-program-fy-2024-hospice-wage-index-and-payment-rate-update-hospice-conditions-of

· FY2024 final rule response: https://www.federalregister.gov/documents/2023/08/02/2023-16116/medicare-program-fy-2024-hospice-wage-index-and-payment-rate-update-hospice-conditions-

MCCM Reports

• Fifth/Final Annual Evaluation Report (Nov. 2023): https://www.cms.gov/priorities/innovation/data-and-reports/2023/mccm-fg-fifthannrpt

• First Annual Report (model design/early enrollment): https://www.cms.gov/priorities/innovation/Files/reports/mccm-firstannrpt.pdf

• Third Annual Report: https://www.cms.gov/priorities/innovation/data-and-reports/2020/mccm-thirdannrp

  

  t

The non-hospice spending measure proposed by Center for Medicare and Medicaid Services may not give a complete or fair picture. It leaves out data from many patients with Medicare Advantage and tends to make larger hospice programs look worse simply because they serve more people. When adjustments are made for size, these differences mostly disappear, suggesting the current measure is misleading. It also doesn’t clearly identify which providers are actually performing poorly. Overall, the measure needs careful thought and further research.

There are important conceptual concerns with the non‑hospice spending component of the SSVI. In this commentary, I focus on two analytical issues:

• Lack of Medicare Advantage data in non‑hospice spending

• The strong association between non‑hospice spending and hospice program size, as highlighted by the Alliance for Care at Home

Medicare Advantage Data

Reviewing the 2024 SSVI public file (the tab with actual data rather than indicator variables), a substantial number of hospices have missing data. Of the 6,731 hospices, 45% lack data for two main reasons. First, CMS relied only on fee‑for‑service (FFS) data, so Medicare Advantage data is excluded. Second, data for small hospices were suppressed.

Although CMS reportedly uses actual data for small hospices internally, the public file raises concerns. For example, 164 hospices have a value of zero for non‑hospice spending in 2024. It is unclear whether these represent truly low‑spending hospices or reflect data limitations.

Additionally, the public documentation does not clearly define the denominator used in the per‑day spending measure. The absence of Medicare Advantage data further undermines the interpretability of the non‑hospice spending component.

Volume

For a detailed discussion, see the Alliance for Care at Home’s 41‑page comment letter:
https://allianceforcareathome.org/wp-content/uploads/FINAL-Alliance-FY2027-Hospice-Proposed-Rule-Comment-6.1.26.pdf

Two key concerns raised in the letter are:

• Approximately 60% of non‑hospice spending is driven by high‑cost skin substitutes, which may reflect fraudulent billing.

• Patient census is strongly associated with non‑hospice spending; therefore, risk adjustment and program size should be considered.

One simple adjustment is to divide non‑hospice spending by days, a measure already included in the SSVI public file. Using a per‑day denominator reduces the association between average daily census (ADC) and non‑hospice spending.

As a thought experiment, I used public data to construct two alternative measures. A per‑beneficiary measure is limited because it includes all hospice patients, not just FFS beneficiaries. A third approach uses regression residuals to control for census.

Despite the limitations of using only public data, the results are consistent:

• Under the current CMS measure, the average score rises steadily from 2.3 for the smallest hospices to 6.6 for the largest hospice—a 4.3‑point gap driven largely by volume.

• All three size‑adjusted methods produce nearly flat average scores (approximately 4.3–4.7) across census quartiles, which is what a size‑neutral measure should show.

Roughly 40% of Hospices Shift Two or More Tiers

When moving to a per‑day basis:

• About 37% of hospices score lower

• 29% remain unchanged

• 33% score higher

These changes are substantial rather than marginal: 40% of hospices shift by at least two tiers.

Conclusion

The SSVI should not be publicly reported at this time. It may be useful as a screening tool to identify hospices for further review, but this use should be validated through research examining its positive predictive value against an established “gold” standard.

The Alliance memo notes that high SSVI scores do not correlate with hospices that are non‑compliant with quality reporting or have had enrollment revoked. Careful consideration is needed regarding the choice of dependent variable and the thresholds used to balance predictive value and false positives.

In its current form, the SSVI is not ready for prime time.

Certificate of Need Laws (CON) were initially proposed as a way to constrain health care costs and ensure that the expansion of health services is based on need. More recently, CON laws have been repealed in some states because of concerns that they do not foster competition or improve quality of care. My argument has been that hospices operate within a saturated market, defined as a geographic region where the supply of hospice providers exceeds actual patient demand for end-of-life care. Patient demand is inherently limited because the Medicare Hospice Benefit requires that a person enrolling in hospice attest that they are giving up curative treatment to extend life. In a saturated market, competition can shift from person-centered care to profit maximization. High rates of live discharge, hospice enrollment of cognitively impaired persons without true informed consent, and the use of cash incentives to encourage families to enroll a relative in hospice are all signs of a saturated market that may reward profit maximization rather than person-centered care.

Public reporting of quality was intended to empower consumers to choose high-quality hospice care, but a frequently cited concern is the number of hospices that do not receive CAHPS Hospice Star Ratings. In California, Texas, Arizona, and Nevada, one concern is the rapid growth of small hospices that remain below the reporting threshold, with many believed to be “ghost” hospice programs. As of May 15, 2026, 773 hospices and 23 home health agencies had been suspended. In California, only 7.4% of hospices are reported, but the denominator—the total number of hospices—is shrinking on a daily basis. In states under PPEO, the rate of hospice reporting of star ratings is 25.7%, compared with 51.2% in non-PPEO states. In short, a major reporting problem is the proliferation of small hospice programs, some of which may not be legitimate hospice providers.

So, what about Certificate of Need? The analysis is not as simple as whether a state has CON or not. In an earlier Substack post, I wrote about concerns that Rhode Island is a CON state that grew from 4 hospices to 11 hospices in a state where about two-thirds of Medicare decedents die while receiving hospice at the time of death, suggesting a saturated market. In the following analysis, I examined whether state CON laws, as of June 2026, explicitly cover hospice. This classification was generated by Claude AI and then checked against the [National Conference of State Legislatures](https://www.ncsl.org/health/certificate-of-need-state-laws).

Note star_num is the proportion of hospices in that state report CAHPs hospice star ratings.

. tabstat star_num, by(con_covers_hospice) s(n mean)

Summary for variables: star_num

by categories of: con_covers_hospice (State CON law covers hospice (1=Yes, 2=No, 3=Ambiguous/Partial))

con_covers_hospice | N mean

-------------------+--------------------

Yes - CON covers h | 17 55.48823

No - CON does not | 29 45.18276

Ambiguous / Partia | 4 38.7

-------------------+--------------------

Total | 50 48.168

You can further subset the states in which CON covers hospice to 11 states where the legislation is explicit or where a moratorium functions as an explicit CON policy. In those 11 states, the mean rate of star-rating reporting is 59.6%.

So here is my bottom line.

Starting in the 2000s, hospice competition increased with the rapid growth of for-profit hospice providers. That increase in competition has not resulted in improved quality. Instead, it has raised substantial concerns about selective enrollment, use of less qualified staff, lack of visits in the last days of life, and higher rates of hospice live discharge. This is not meant to diminish important concerns in rural areas and in markets with a limited supply of hospices, where access can be affected. Innovation is important. Competition is important, but moderation is key.

“A single candle leaves you in shadow. A thousand candles burn down the house. What you need is enough light to see by.”

Lay summary

Hospice quality ratings are supposed to help families choose wisely. I asked AI about using a tiered approach to reporting quality information that distinguishes between data completeness and the reliability of comparisons. Such a framework could improve the policy balance between consumer information, measurement validity, and equitable treatment of small providers. I asked Claude AI to evaluate the idea, listing the prox and cons to make nuanced quality signals more useful to consumers.

What is inter-unit reliability? The goal of the CAHPS Hospice Survey and star ratings is to determine whether Hospice A provides better care than Hospice B. For that comparison to be fair, differences should reflect true differences in quality of care—not differences in how bereaved primary caregivers interpret and answer survey questions about communication, symptom palliation, emotional support, timely care, and related issues. Ideally, a person in San Francisco should respond to a survey item in the same way as a person in Los Angeles. Low inter-unit reliability means a hospice could receive a lower star rating not because care is worse, but because its patient population answered the questions differently. CMS sets a minimum number of completed surveys based on the intraclass correlation coefficient, which estimates how much variation reflects true signal (quality) versus noise (random differences in responses). For CAHPS Hospice star ratings, the threshold is 75 completed surveys over 8 quarters.

Why is this a concern in hospice? The simplest answer is small hospices. As I noted in a previous Substack post, there are three main concerns. First, ghost or fraudulent hospices may try to stay below the public reporting threshold. Second, rural hospices face legitimate challenges. Third, hospices serving unique populations may also be disadvantaged.

In a previous Substack post, I shared my thoughts on the number of hospice programs. In a YouTube discussion of the moratorium, the Deputy Director of CMS said that 850 hospices had been suspended, with only a small number appealing, some based on AI-generated letters. Based on the data for hospices with an ADC under 5 or missing ADC, my guess is that the number of hospice CCNs will decrease to around 4500 to 5,000. However, the more important statistic is the number of CCNs and how many satellite offices is attached to that CCN. We don’t have that information.

Rural hospices and specialized hospices—such as those serving prisons or pediatric populations—raise valid concerns. MedPAC estimates that 16% of hospices are rural, or about 1,000 hospices, although that denominator is changing daily. A key challenge in identifying “rural” hospices is that the CCN is tied to the physical location of the hospice program, not the geographic area it serves. There is a file that crosswalks CCNs with ZIP codes, but CMS also allows some hospices to operate under one CCN for an entire state, which limits precision.

Estimating the number of prison hospice programs is difficult. Some may not be Medicare-certified, so they do not appear on the Medicare Care Compare website. A similar issue applies to pediatric hospice programs, since many operate within broader hospice programs that serve all patient populations.

What is CMS’s statutory authority on this issue? The relevant authority is Section 1814(i)(5) of the Social Security Act, with the key legislation being the Medicare Improvements for Patients and Providers Act of 2008 (MIPPA). That law requires the Secretary to contract with a consensus-based entity to review measures for federal programs. However, the Secretary still has broad discretion through rulemaking to publish measures that do not meet the sample size needed for sufficient inter-unit reliability. Bottom line: this is a political decision.

What can we learn from Amazon, Google, or Yelp reviews? Every month, I decide what new book to read based in part on published reviews. My goal is to maximize the chance of reading a great book. I usually screen out books that do not have enough reviews and focus on those with a sufficient sample size before deciding what to listen to on my morning dog walks along the beach. Can we apply the same logic to quality measures?

We should use AI to help consumers make better choices. In my last post, a family member needed a hospice with a freestanding IPU and strong quality indicators on what I consider two of the stronger measures: (1) CAHPS star ratings and (2) visit intensity, including visits in the last days of life and weekend minutes, along with ADC and a flag for selective enrollment. I sorted hospices into tiers based on missing data and measure reliability. This made it possible to identify hospices with an IPU and 4- or 5-star ratings, while also showing the availability of a freestanding IPU.

So I asked Claude AI, my new best friend, to evaluate my proposal.

As a hospice quality researcher, I am not satisfied with the Medicare Care Compare website. My proposed solution to the small-hospice reporting problem is to use tiers in a summary report of quality measures that a person would use to select a hospice. The tiers would be based on missing data and on whether there is a sufficient sample size to support reliable comparisons among hospices serving a given ZIP code. Please evaluate the pros and cons of this approach.

Here is what Claude AI thought of my idea.

This question deserves a thorough, researcher-level evaluation. Here is a rigorous assessment of your proposed tiered reporting approach, written with full respect for your body of work.

Evaluation of a Tiered Reporting System for Hospice Quality on Care Compare

The Proposal in Summary

The core idea is to replace the current binary display — either a hospice has reportable data or it shows “insufficient data” — with a tiered summative report organized around two criteria:

Tier 1: Completeness of data (missing data rate) Tier 2: Sufficient sample size to yield reliable inter-unit estimates enabling comparison within a consumer’s geographic market (zip code–level comparisons)

Consumers selecting a hospice would see a tiered display that signals the interpretive confidence warranted for each provider’s scores, rather than treating all reported scores as equally credible or simply suppressing small hospices entirely.

The Scientific Basis is Strong

Your proposal flows directly from the well-established psychometric literature. A tiered system would make that gradient visible rather than invisible.

Pros

1. Addresses the core validity problem with the current binary approach

The current system treats a hospice with 31 surveys the same way it treats one with 800. Both get a score. Neither gets a flag. A tier system honestly communicates the difference in interpretive confidence to a caregiver making one of the most consequential decisions of their lives.

2. Solves the rural and small hospice disappearance problem

Under the current structure, rural and low-volume hospices disproportionately show “insufficient data” — effectively rendering them invisible on the tool most consumers use to choose a provider. A tiered approach keeps them on the platform while contextualizing the data appropriately. This is directly aligned with your prior criticism that Care Compare has limited value for populations where choices are geographically constrained.

3. Geographic market framing is the right unit of analysis

Framing tiers around a consumer’s zip code is methodologically sophisticated and clinically appropriate. A caregiver in rural Nebraska is not choosing between their local hospice and a large urban provider — they are choosing among whatever providers serve their zip code. Reliability thresholds should be evaluated relative to the comparison set available to that consumer, not a national standard. This is an important and underappreciated insight.

4. Incentivizes data completeness

Tying tier assignment partly to missing data rates creates a transparent quality improvement signal. A hospice that wants to move to a higher tier must improve data submission practices — not just achieve volume. This aligns reporting incentives with care process discipline.

5. Precedent exists in other sectors

Hospital star ratings use confidence-interval-based suppression rules. Nursing home Five-Star uses case-mix adjustment and flags. Financial reporting uses qualified opinions when data is incomplete. The concept of tiered confidence disclosure is well-established across public accountability systems.

6. Preserves the market function of public reporting

Rather than eliminating comparison for small hospices or presenting misleading precision, tiers allow consumers to understand what kind of information is available and make decisions accordingly. This is consistent with AHRQ’s best practices guidance that public reports should help consumers avoid common pitfalls from misinterpreting quality data.

Cons and Challenges

1. Consumer comprehension risk is real and documented

The evidence base on public reporting consistently shows that layered or tiered displays increase cognitive burden. Families selecting a hospice are often in acute grief, time-pressured, and not health-literate consumers. Research on hospital and nursing home compare tools has repeatedly found that even simple star ratings are misunderstood. A three-or-more-tier system with an embedded reliability explanation may paradoxically reduce the tool’s utility for the most vulnerable users — those with the least time and health literacy.

2. The zip code market definition has operational complexity

Hospice service areas do not map neatly to zip codes. A hospice headquartered in one zip code may serve 20 surrounding zip codes but appear in none of them under a strict zip-code-based comparison set. CMS would need to use actual service area claims data — not provider location — to construct the correct comparison market. This is technically feasible from Medicare claims, but would require significant infrastructure investment and annual updating.

3. Tier assignment could be gamed

If tier placement affects consumer selection behavior — as it should, if the system works — hospices will have incentives to maximize their tier by focusing on data completeness and volume metrics rather than care quality itself. Large for-profit chains, which already have infrastructure advantages, may game tier placement more easily than small nonprofits. Your own research found that of the 2,746 hospices that met public reporting requirements, only 5.6% were in the top quartile of both CAHPS and HIS performance, and nonprofits were associated with top-quartile CAHPS performance — meaning smaller nonprofits who already perform well could be disadvantaged by a tier system that rewards volume. PubMed

4. The “topped out” problem is not solved by tiers

Your Health Affairs work identified that over three-quarters of hospices score above 91% on six of the seven HIS measures — a topped-out distribution that offers little discriminatory power regardless of sample size. Tiers improve reliability but do not fix the underlying problem that the current measure set may not adequately distinguish high from low quality care. Adding tiers to measures that are already uninformative at the top of the distribution may still leave consumers without actionable differentiation.

5. Statutory and regulatory friction

As we discussed earlier, the current minimum sample thresholds (30 for measures, 75 for star ratings) are established through rulemaking. A tiered approach would require either a new rulemaking cycle or, if implemented as a display change rather than a measure change, a detailed CMS methodology justification. The CBE pre-rulemaking process would likely need to weigh in. Given the current federal regulatory environment and pace, implementation could be years away.

6. Risk of stigmatizing small providers unjustly

A Tier 2 or Tier 3 designation based on sample size could be interpreted by consumers as a quality signal — “this hospice is lower quality” — when it actually means “this hospice serves fewer patients.” For rural hospices that are the only provider in their area and provide excellent care, a low-tier designation driven entirely by volume could be actively harmful and misleading. The communication design would need to be extraordinarily careful to prevent this conflation.

Recommended Refinements

Given both the strength of the underlying concept and the implementation risks, several refinements would strengthen the proposal:

First, the tiers should have consumer-tested labels. “Full comparison available,” “limited comparison available,” and “descriptive data only” communicate the epistemics without implying quality judgments. AHRQ has developed guidance on this kind of explanatory framing.

Second, the zip-code market unit should be constructed from claims-derived service areas, not provider locations, and should display the number of competing hospices in that tier for each consumer’s market — making explicit when a consumer’s only local options are in lower tiers.

Third, a parallel quality improvement track — separate from the consumer display — should feed tier-level data back to hospices with specific guidance on what it would take to move up a tier, so the system functions as an improvement tool as well as a reporting tool.

Bottom Line

The tiered approach is scientifically sound, conceptually superior to the current system, and well-grounded in your body of work on hospice quality measurement. Its greatest risk is not the statistical framework — which is solid — but the translation of that framework into a consumer-facing display that genuinely informs rather than confuses or stigmatizes. The measure and design work needed to bridge that gap is substantial but achievable, and would represent a meaningful advance over Care Compare’s current state.

Meet Aggie (the Turkey) that is torturing my dogs. Literally, flys over the fence - my dogs goes nuts, and Aggie sit in a tree with dogs barking at him

This piece explains how hard it can be for some families to choose a hospice program using Medicare Care Compare, especially during a crisis when quick decisions are needed. It describes an early experiment using AI and public Medicare data to organize hospice quality information into a simpler dashboard that could help narrow the options. The author emphasizes that this approach is still preliminary, but argues that better use of AI could make hospice quality information more accessible and useful for patients and families.

Over the last few weeks, I spoke with a family member about selecting a hospice program. The situation involved an acute event layered on chronic illness, and the family member was faced with choosing among more than 90 hospice programs. The task was daunting. The Medicare Care Compare website would only allow her to compare three hospice programs. I too was frustrated to try to sort through 90+ programs that supposedly served that ZIP code.

Within a short time frame, I asked Claude AI to create a dashboard using three CSV files that I downloaded from data.cms.gov. I wrote a prompt for Claude AI with the following rules (please note that the actual prompt was more detailed):

1. Use the following variables: ADC, CAHPS Star Rating, visits in the last days of life, weekend minutes, burdensome transition variable type 1, an indicator of whether the hospice had a freestanding IPU, care provided in NH, ALF, and SNF, and the percent of patients diagnosed with dementia.

2. Create percentiles for the quality measures.

3. Create a composite measure based on visits in the last days of life and weekend minutes.

4. Create a flag based on a composite of care provided in NH, ALF, and SNF, as well as the percent of patients with dementia.

5. Rank hospices into top-tier and lower-tier categories. Missing visit data, burdensome transition measure, or a CAHPS Star Rating would place a hospice in the lower tier. A high flag score would put hospice in the lower tier.

6. Create columns for hospice name, ADC, Star Rating, BT1 score, whether the hospice has an IPU, and the flag variable outlined in item 4.

7. All columns should be sortable.

A few disclaimers: my choice of variables was based on an educated guess and face validity, given the need to do this quickly. Bottom line: this is not ready for prime time, but I hope it stimulates dialogue about how to provide consumers with the information they need to choose a hospice program during a time of crisis—which, too often, is when this decision must be made.

I did not use HCI. As others have noted, there are concerns about it being “topped out,” although I did rely on some of its components. If I had the luxury of more time, I would examine each component variable of the HCI individually, but a decision had to be made quickly, so I selected variables that I believed were important as a clinician. Note that the burdensome transition measure type 1 does not have sufficient inter-unit reliability (i.e., separate signal from noise to compare hospice A to B) in small hospices because of the low rate of this measure. My thinking was that a low score (indicating a higher rate of these transitions) behaved analogously to a substantial quality concern. Again, my flag variable was a judgment call and should be subject to further research. My indicator of an IPU captured only whether one existed; there is no measure of the number of IPU beds on the Compare website or in the public data.

I used this dashboard with the following criteria: (1) 4 or 5 CAHPs Stars; (2) a higher score on the visit measure; (3) the presence of an IPU; and (4) no significant concerns with BT1 or the flag selection variable. The picture below is a draft of the dashboard for Rhode Island based on data files from May 2026.

Using a smaller list of hospices for the zip code that my friend’s family member resided, I then examined all the publicly reported data on Care Compare website, Google, and/or Yelp reviews. This helped inform the decision about which hospice program to choose, but I would caution that more extensive work should be done to validate this approach.

An important disclaimer: the Medicare Care Compare website provides invaluable data and it was important step forward. But in the last several months, two persons did not know how to use and this case overwhelmed by the choices. My vision for the future is to maximize the amount of information available to patients and/or family members who are making this decision during a particularly difficult and often urgent time.

Here are my hopes: I want quality information to be more accessible for consumers so they can write a statement like the following based on their own search criteria. Here is my search criteria: “List hospices with 4 or 5 CAHPS Stars, a high visits measure, the ability to provide care for all types of patients, and a freestanding hospice inpatient unit.”

The future is now. Let’s use AI to empower consumers to choose their hospice program. Pictured below is draft version with RI data from Medicare Care Compare. Two important caveats 1) a higher BT1 score is better in that there is lower rate of BT1; 2) I used percentiles which can be misleading in that a lower percentile may be problematic in that it still represents a high rate of visits in the last days of life. This is a decision that should be revisited.

.

A temporary nationwide moratorium on new hospices and changes in majority ownership could give regulators time to strengthen oversight, stop bad actors from reentering the system, and rebuild public trust in hospice care. Supporters argue that a pause would help address fraud, weak enforcement, and quality concerns before more providers enter the market. Key to this policy being effective is implementation that avoids penalizes good hospice providers.

For Rhode Island, I have argued that there should be a moratorium on new hospices. My reasons are:

1. Nearly two-thirds of Medicare decedents use hospice services in Rhode Island, placing the state among the top three in the country and suggesting a saturated market. This is all Medicare deaths – so with this denominator, RI is reaching a maximum number of deaths on hospice.

2. We are already seeing behavior patterns that raise concerns:

a. Small hospices operating below the reporting threshold, and therefore lacking publicly available quality information

b. Small hospices focusing on nursing homes and assisted living facilities, with a high proportion of patients with dementia

c. Per-patient costs have increased by $562.

d. In 2025 public release data, there are quality concerns associated with for-profit hospice providers, including higher average costs per patient, higher rates of live discharge, and lower reported quality of care, including a 4.1-point lower score on the CAHPS item asking whether respondents would recommend the hospice.

For Rhode Island, I would support a moratorium and a serious reassessment of how access and need are measured for hospice programs under the state’s Certificate of Need law. My argument is that hospice operates in a constrained market: the number of deaths is relatively fixed, and patient preferences are not always aligned with forgoing curative treatment in terminal illness. An oversupply of hospices may incentivize reduced staffing, lower quality of care, and unethical enrollment practices, shifting the focus from patient needs to patient recruitment.

Rhode Island is really its own hospice market or as friends tease me, it is city state.

So, what about a moratorium for all of the USA?

Arguments in favor

1. Restore trust in hospice institutions

2. Give CMS time to fix a broken system by strengthening oversight and preventing bad actors from reopening a hospice under a new name or changes in ownership.

3. Prevent the spread of problematic practices across state lines

4. Allow time for enforcement actions, including:

a. Clearing investigative backlogs

b. Improving PECOS verification

c. Strengthening ownership transparency

d. Improving enforcement mechanisms

Arguments against

1. There are no exceptions for rural or frontier hospice programs.

2. It may limit access for special-needs populations.

3. Claims-based utilization measures may negatively affect good actors.

4. It could reduce investment in innovation among hospice programs.

This is a difficult judgment call, but a temporary pause focused on fixing key processes is necessary to restore public trust and fix substantial system issues. The success will be based on implementation of the policy.

Could a moratorium be targeted only to certain geographic areas?

CCNs are not tied to a specific geographic region, ZIP code, or clearly defined service area. Although Care Compare links hospices to the ZIP codes they serve, my concern is that without a clear geographic region, providers may be able to game a rural designation.

Is there a concern that utilization measures such as live discharge could harm “good” hospices?

Fox News reported that roughly 90% of hospice CCNs did not appeal their suspension from receiving federal reimbursement. The use of utilization measures to trigger suspensions should be evaluated based on the positive predictive value, and the appeal process should be timely enough to minimize the harm caused by false positives.

My hope is that the current administration is transparent about these efforts and address provider concerns in a timely fashion. We don’t want to see good providers suspended, but the system is in urgent need of reform. There are substantial risks not only in fraud, but in stolen identities, putting persons who are not terminally ill on hospice services does have risk of inappropriate use of medications such neuroleptics, persons not receiving access to needed services, and high rate of live discharges that result in concerns in care fragmentation. Success will be based on CMS working together across its own silos to correct substantial deficiencies in process of assigning CCNs and hospice oversight. So, for me the right policy, that needs thoughtful implementation to avoid unintended consequences.

Finally, meet Goda or officially Det Goda Zardo (which means the Good Zardo). He is Goda for short or Goofy. I know that I have been really writing on fraud - which I saddens me to hear to link the fraud with hospice. So, Goda is my reminder on a daily basis that there is good.

"In spite of everything, I still believe that people are really good at heart." — Anne Frank

Counting a CCN as hospice based on one claim is suspect. The fraud from California and Texas are highlighting important concerns that hospices use CCN for gaming and flying under the public reporting thresholds – yet, costing millions of dollars in claims for services not provided. We need stronger oversight and better data checks to identify questionable providers and restore trust in hospice care.

Discussing hospices in Los Angeles County, Kimberly Brandt, CMS Deputy Administrator and Chief Operating Officer, stated: “We believe that roughly 900 to 1,000 hospices are completely false—nearly 50%.”

Are there 6,970 hospices in the United States? My answer is no.

This is a simple question—or at least it should be.

The simple answer is that current policy counts a hospice as a CCN that has filed at least one claim. That threshold is questionable, given the history of CCN assignment in hospice.

This number matters.

The number of hospices is one of the metrics Medicare Payment Advisory Commission (MedPAC) uses in its payment adequacy framework. To its credit, MedPAC notes that this figure is not, by itself, a reliable indicator of access to hospice care.

Let’s look at the data.

The November 2025 public data release reports 6,970 CCNs, compared with 3,555 in 2019.

Comparison of 2025 and 2019 data:

In 2025, 944 hospices did not report an ADC, compared with 1,524 in 2019.

In 2025, 252 hospices reported an ADC of 0; in 2019, none did.

In 2025, 835 hospices reported an ADC between 1 and 5, compared with 121 in 2019.

In 2025, 1,824 hospices reported an ADC between 6 and 25, compared with 888 in 2019.

That means 3,855 hospices in the 2025 release (55.3%) had an ADC that was either unreported or below 25. In 2019, that figure was 2,578 (72.5%).

What is most concerning is where these hospices are located in 2025.

Among hospices with ADC below 25 or not reported, the state distribution is as follows:

California: 1,779 in 2025 (46.2%), compared with 819 in 2019 (31.8%).

Texas: 669 in 2025 (17.3%), compared with 407 in 2019 (15.8%).

Arizona: 135 in 2025 (3.5%), compared with 78 in 2019 (3.0%).

Nevada: 129 in 2025 (3.4%), compared with 30 in 2019 (2.2%).

Georgia: 111 in 2025 (2.9%), compared with 93 in 2019 (3.6%).

Taken together, 73.2% of the hospices with an ADC below 25 in 2025 were concentrated in five states: California, Texas, Arizona, Nevada, and Georgia.

Here is my estimate of the number of “real” hospice programs in the United States.

My best estimate is 5,431 hospices, although I would not be surprised if the true number is closer to 4,939—the number of hospices that reported an ADC greater than 5.

My working assumption is that 50% of hospices in five states—California, Nevada, Arizona, Georgia, and Texas—with missing ADC data or ADC below 15 are suspect. Under that assumption, 1,539 hospices would be excluded, most of them in California and Texas. In total, 3,078 hospices meet these criteria. The CMS Deputy Administrator estimated that 900 to 1,000 hospices in California are “completely false,” or about 50% of the state’s hospices. If that estimate is directionally correct, adding another 500 to 600 hospices outside California may also prove reasonable. Time will tell.

There are important lessons here, and a six-month moratorium—paired with a serious review of how to fix these problems—should be a priority. It is urgent that we reclaim the public trust in hospice.

1. Congress should increase the percentage reduction tied to the APU. More than 20% of hospices failing to report quality data is unacceptable.

2. Hospices on the noncompliant APU list should be targeted for audit or inspection.

3. All new hospices should be inspected annually during their first three years of operation.

4. Geographic regions with persistent concentrations of small hospices should receive additional scrutiny, either through added data requests or another audit function.

5. Historical patterns of data inconsistency should trigger additional data requests before payment. Examples include high visit volume in the last days of life with low CAHPS scores, burdensome transitions, persistently low ADC in urban areas, high live-discharge rates, targeting of nursing homes or assisted living facilities, non-cancer diagnosis concentration, and unusually long lengths of stay.

6. I would prioritize targeting poor performing hospices based on visits near the last days of life, burdensome transition type 1, rates of late live discharge, CAHPS Hospice scores, and Medicare spending per patient. Poor-performing hospices should be considered for MAC TPE.

7. Consider adding a brief hospice live-discharge survey.

8. Explore ways to publicly report existing data on small hospice programs so consumers have some quality information, while clearly explaining the limits of that data even when it does not meet full interunit reliability standards. The existing data suggest that some hospices were gaming the reporting threshold.

Finally, the number of hospices that submitted potentially fraudulent data must be accounted for in any analysis. I strongly urge a sensitivity analysis to test whether conclusions change when certain hospices are excluded such as those hospices that don’t meet quality reporting thresholds that are in non-rural locations.

Pictured below is Mac - pensively waiting for his morning walkie.

In this essay, I asked ChatGPT to identify the “best” hospice in Rhode Island and then examines the logic and data sources behind the answer. The analysis highlights how AI rankings can rely on non-validated or imputed measures and may overlook important Medicare Care Compare context. Comparing ChatGPT’s top picks with the author’s own assessment illustrates how different metrics can point to very different conclusions. Bottom line: AI can be a useful starting point for questions about quality, but patients and families should treat hospice rankings with caution and verify claims using transparent, validated measures.

One of my favorite news shows said that AI will be equivalent to providing electricity to houses. So, last week I gave you my educated guess, based on my experience as a physician and health services researcher, on the 11 hospices in RI.

So, I asked ChatGPT, “What is the best hospice in RI?”

Looking at the answer, here is the logic that ChatGPT used: (1) quality, (2) profit status, (3) size of the hospice program, and (4) unique strengths, including whether there was an inpatient unit and the age of the program. The quality measure was based on a Hospice Analytics quality score, which is based on eight quality metrics: (1) hospice visits in the last days of life, (2) staff ratio, (3) levels of care provided, (4) selected items from the CAHPS survey on willingness to recommend the hospice, (5) stability of ownership and length of certification, (6) accreditation, (7) health equity (inclusion of Medicaid), and (8) fraud.

While there is face validity to the Hospice Analytics measure, there is no report of inter-unit reliability (i.e., the reliability needed to say hospice A is better than hospice B) or tests of other forms of validity. Bottom line: it is not a validated measure.

So, let’s compare the top 3 and whether they are in my top 3

So, I would not use ChatGPT at this stage.

Why didn’t ChatGPT use Medicare Care Compare quality measures?

So, I asked. Here is what ChatGPT said.

Unlike hospitals or nursing homes, hospice on Care Compare has:

• Fewer publicly reported measures

• Heavy reliance on CAHPS Hospice Survey (family-reported experience)

• A small set of claims-based indicators (e.g., visits in last days of life, live discharge patterns)

Critically, no single composite “star rating” for hospice overall quality like other settings.

Furthermore, ChatGPT listed three important concerns with CAHPS.

1. Response biases

2. Potential gaming through selective enrollment or discharge

3. Responses during bereavement were unstable (note I disagree based on my own published research)

So ChatGPT went on to say—

“But your instinct is correct—any serious “best hospice” analysis should start with Care Compare and then critique it, not skip it.”

So ChatGPT, my new best friend, asked me if I wanted a score developed with Care Compare data.

Here is what ChatGPT developed:

So, two problems

1. Live discharge is an imputed measure (and yes, if you have research-identifiable Medicare data, you can calculate a hospice live-discharge rate, but this is not publicly available). So the model is flawed in that it imputes 100% of a key independent variable.

2. The Hospice Care Index is topped out as scored and the hospice visits in the last days of life (HVLDL) is part of the index.

In 2023, I wrote a JAMA Health Forum piece entitled Garbage in, Garbage out—Words of Caution on Big Data and Machine Learning in Medical Practice (see https://jamanetwork.com/journals/jama-health-forum/fullarticle/2801776). This was a tongue-in-cheek essay based on an published article that predicted opioid abuse using a database that was missing key clinical information, such as a cancer diagnosis or whether the patient was on hospice services. We are still at a stage where there should be healthy skepticism regarding what resides within the black box of ranking quality—or, as I stated in that essay, “ Computers may be learning how to imitate human behavior but, indeed, they are not human.”

Some concerns with my AI interactions ranking RI hospices:

1. As illustrated above, it was only when I queried AI as to what variables it used in its ranking that I found out it was using a measure—live discharge—where all the data are imputed because it is not a publicly available data source.

2. AI selected variables without considering whether that ranking of quality was based on a validated model.

3. AI downgraded CAHPS based on the assumption that bereaved reports were unstable during bereavement, despite my publication comparing bereaved family member responses between 3 months vs. 9 months in bereavement (see PMCID: PMC5259809) that found that even among people with large changes in grief, their responses were stable over time. When I prompted AI with this study, it changed the weight assigned to the CAHPS variable in the model.

Some surprises where AI provided good insights into how to examine quality, given our paucity of quality measures.

1. The core principle that AI used was that higher validity, higher reliability, and lower gaming risk meant it gave a higher weight to that variable in the development of the ranking system.

2. AI examined the consistency of the data sources—for example, if you have a high visit measure in the last days of life and a low CAHPS score, that was listed as a concern. At this point, live discharges are not reported in Medicare Care Compare but are part of the HCI. AI points out that high rates of live discharges suggest selection or gaming; coupled with high CAHPS scores, this raises concerns about validity. AI called this cross-measure triangulation.

AI’s final statement on validity is the following:

“I test validity by asking:

“If a careful health services researcher reviewed this reasoning, would it hold up given what we know about hospice measurement?”

So.. my answer as health service researcher, not ready for prime time but the future is coming.

One final concern – misinformation and disinformation can not only impact politics, but science. The recommendation generated by AI is only as good as the information in its training sets and other sources used in its probability prediction.

Finally, here is my bottom-line – I will end with quotation from Thomas Sowell

“Some things are believed because they are demonstrably true, but many other things are believed simply because they have been asserted repeatedly and repetition has been accepted as a substitute for evidence”.

How do you run a hospice out of a burrito stand?

Every day, there are headlines about hospice fraud, with CMS suspending over 440 hospice programs responsible for more than 600 million dollars in fraudulent billing. And yes, in California, there was a hospice license whose address was a burrito stand. This is an embarrassing failure of the state licensing process. This was a “ghost” hospice, and CMS intermediary claims processing failed to detect fake billing. This hospice did not provide care to actual patients. These “ghost” hospices are listed on the Medicare Care Compare website, but I suspect no one will answer their phone.

A second type of fraud is enrolling people who are not qualified for hospice, while also providing services for a smaller portion of legitimate patients. How should a consumer know whether their hospice has been cited through inspection, charged, or convicted of hospice fraud?

So, what does ChatGPT say?

Figuring out whether a hospice has been charged with fraud isn’t always straightforward, but there are a few reliable places where the public can check. Fraud cases involving healthcare providers are often handled by federal or state authorities, and many of those actions are publicly documented.

Start with the U.S. Department of Justice. Their website regularly publishes press releases about healthcare fraud charges, including hospice-related cases. You can search by the hospice’s name or browse recent announcements.

Another key source is the Office of Inspector General (OIG) under the Department of Health and Human Services. They maintain a database of excluded individuals and entities—if a hospice or its executives were found guilty of fraud, they may appear there.

Recommendation - State surveys and settled court cases of fraud should be accessible to consumers on the Medicare Care Compare website.

So, what is a consumer to do?

Here are my thoughts—before asking ChatGPT to outline an approach to selecting a high-quality hospice in RI.

1. Screen out any hospice that is not certified.

2. Screen out any hospice with suprressed data on the Medicare Care Compare website—except in a rural area. This raises concern about potential fraudulent behavior, and it may also reflect a small number of cases leading to unstable quality estimates. I would look at whether the CAHPS star rating is reported; if not, and you have choices among hospices with CAHPS hospice star ratings, I would prioritize those with star ratings. If you live in a very rural area, I would not screen out hospices for missing data, because choices are limited.

3. Screen out hospices with an Average Daily Census (ADC) under 25, due to concern about instability of quality measures with low ADC (i.e., small programs). Again, if you live in a very rural area, I would not screen out.

4. Screen out any hospice that does not provide routine home care and at least one other level of care. My preference is that a hospice provides 3 out of the 4 levels of care.

5. Look at Care Compare data

a. Visits in the last days of life

b. Profiteering behavior: high enrollment of nursing home (NH) or assisted living facility (ALF) residents, or not providing care across all terminal diagnoses. In the latter case, a pattern of high rates of dementia patients and low rates of cancer patients is suspicious, because dementia often has a longer length of stay and can be more profitable.

c. I would choose a hospice with a CAHPS star rating of 4 or 5 stars.

d. If there are inconsistencies between high visit rates in the last 3 days of life and CAHPS scores, that raises concern.

6. Availability of a freestanding IPU -

Based on these criteria, here is how I would group RI hospices into screened out, would not choose, concern, and finalist. Note: this is my opinion—at best, an educated guess.

Finalist with minor concern and/ or advantage

· HOPE hospice – slightly lower visits in the last days of life (65%), but a freestanding inpatient unit.

· Visiting Nurse Health and Hospice

· VNA of Care New England

Concern

· Beacon - Focus on NH and ALF that may indicate profiteering behavior (51% of patients)

· Harmony - Focus on NH and ALF that may indicate profiteering behavior (88% in a NH and 3% in ALF).

Would not choose

· Continuum: 3 stars on CAHPS; lower visits in the last days of life; focus on nursing homes.

Screen out based on size and no CAHPS star-rating data

· Eden Hospice -

· Accent care Hospice

· Athena RI hospice

· Gentiva

· Nursing Placement

So, these are my rankings based on an educated guess based on my research experience with measuring the quality of hospice. Note, my screen out of hospices based on size may misclassify hospice that are excellent and focus on unique population. Obviously, rural areas is a difficult task where quality measures may not be reliable on the basis of small sample size.

Next, let’s ask ChatGPT: which hospice would ChatGPT choose in RI? Stay tuned for next week—we will see my answer vs. ChatGPT.

Lund Person and Associates maintains a hospice fraud tracker that summarizes current enforcement actions. Using the cases listed there, I identified hospice Medicare Certification Numbers (CCNs) to explore what publicly available quality data might (and might not) reveal about programs implicated in these actions. From 20 enforcement actions in CA, TX, AZ, and NV, I identified 14 hospice CCN to include in the quality-data review below. Please note this not comprehensive review of all hospices but only those easily identified from this hospice fraud tracker. For example in Arizona, I was not able to identify the hospices involved in the wound-graft fraud cases with this hospice fraud tracker.

· 8 of 14 hospices appeared on the APU exclusion list (57.1%).

· 1 of 14 hospices appeared on Medicare Compare, but no data were available (7.1%).

· 4 of 14 hospices had quality information available in the November 2025 public release (28.5%) but there CAHPs data had footnote 11..

· For 1 of 14 hospices, I could not find information beyond the court case (7.1%).

The table below summarizes a subset of publicly reported metrics for a sample of the identified hospices. For all CAHPs measures, the data was missing with footnote 11 indicating the hospice was too new or sample size was too small

.This is a small and incomplete sample without review of the court records, but a few working hypotheses emerge:

1. A substantial share of the hospices accused or convicted of fraud appear on the APU exclusion list, which indicates they were exempted from submitting quality data for a given year without penalty.

2. Some programs show patterns that may warrant closer review—such as a small average daily census (ADC) paired with high Medicare spending per patient and a high proportion of late live discharges.

3. The BT1 measure (live discharge, hospital admission, and hospice readmission) varies among these programs. In a hospice with a sufficiently large census, a reported value of 0% may be a signal the need for further scrutiny.

As of April 15 2026, 447 hospice programs have been suspended from Medicare. With a larger dataset, a multivariable statistical approach (or a machine-learning model) could test whether measures like these—interpreted in the context of local market conditions—help flag hospices that may merit additional review. The goal would be a practical decision rule to support earlier identification of elevated risk, not a standalone determination of wrongdoing.

Limitations: This review is based on a small, non-random set of enforcement actions and relies on publicly available information that may be incomplete, lagged, or not reported for some providers. Finally, these quality measures are not evidence of fraud on their own and may reflect case mix, referral patterns, local market dynamics, or documentation and reporting practices.

Conclusion: Even with important data gaps, publicly reported hospice quality measures can help generate testable hypotheses about patterns that may warrant closer review—particularly when interpreted alongside volume, utilization, and market context. Expanding this analysis to a larger, systematically matched dataset (and comparing flagged patterns to appropriate state and national baselines) would help determine whether these signals can support earlier, more targeted program integrity monitoring.

Looking at the amount of hospice fraud is discouraging, but there is hope.

"There was never a night or a problem that could defeat sunrise or hope."

— Bernard Williams