Those elements have certainly changed. Enrollment in the Medicare hospice benefit increased from 9% in 1992 to 23% in 2000 (11). Between those years, the percentage of hospice enrollees without cancer increased from 24% to 49%, the percentage of enrollees in nursing homes increased from 11% to 36%, and the percentage of those older than 79 years of age increased from 35% to 47% (11, 12). The Balanced Budget Act of 1997 made the Medicare hospice benefit more flexible and long lasting, instituted prospective payments for after-hospital home and skilled-nursing facility care, and decreased hospital payments for unusually short stays before postacute care. These policies increase the desirability of hospice enrollment, especially for patients without cancer.

We estimated the effects of hospice enrollment on national Medicare expenditures during the last year of life among persons who died of conditions other than cancer and made more recent estimates of effects for persons who died of cancer. Examination of expenditures in the last year of life directly addresses the influence of hospice on Medicare program costs at the end of life and avoids serious limitations of previous studies that involved matching enrollees and nonenrollees by duration of hospice enrollment (8). Our design, analytical methods, and measures address selection bias, matching, and generalizability limitations of previous studies (2-4, 8, 13, 14).

Methods

Top Methods Results Discussion Author & Article Info References

Design

In this retrospective cohort study, we used linear regression models to estimate adjusted mean Medicare payments in the last year of life that were associated with hospice enrollment. Hospice enrollees were matched to nonenrollees by using poststratification (15) with strata formed by age and diagnosis group (cancer or noncancer, or principal condition). Within strata, we adjusted for propensity to use hospice, gender, race, enrollment in Medicaid, urban setting, illness duration, comorbid conditions, consistently low use of Medicare, nursing home residence, and year of death. Data sources include denominator and claims files from Medicare.

Setting and Participants

The sample (n = 245 326) comprises all decedents from the Medicare standard national 5% sample who had fee-for-service coverage, were older than 67 years of age, died between 1 January 1996 and 31 December 1999, and had at least 36 months of continuous Part A and B Medicare coverage before death. We excluded decedents who were eligible for Medicare on the basis of end-stage renal disease or disability and those who resided outside the United States.

Outcome Measures

Our primary outcome measure was Medicare payments to providers, adjusted for inflation to 1999 and summed overall and by type of service (hospital inpatient, skilled-nursing facility, home health, hospice, outpatient facility, and physician or supplier). We exclude co-insurance, copayments, and deductibles because our focus was Medicare program expenditures. Measures of volume and intensity of service use in the last year of life were mean days in hospital, mean days in the intensive care unit, and mean hospice payments per diem. Timing of hospice entry was defined by days in hospice.

Covariate Measures

Hospice enrollees comprised decedents with any hospice claim, including the 10% of those discharged before death or who had gaps in hospice enrollment. Information on age at death, gender, race, and Medicaid enrollment was obtained from the denominator file. Race was categorized as white or nonwhite. Any state buy-in during the last year of life indicated Medicaid enrollment. Urban setting was categorized as metropolitan, urbanized, or rural (16). Categories for unusually many ( 5) or few (1 or none) comorbid conditions (based on the number of Charlson comorbid conditions) (17, 18) were used to control for the variation in illness burden among decedents.

Low use of Medicare (within the 25th percentile of total expenditures consistently for 24 months before death) was used to control for selection and data sampling biases. Consistently low use may indicate selection bias related to patient preference for less aggressive care. If hospice enrollees generally wanted less aggressive care, the associated costs would be less even without hospice. Consequently, hospice savings would be overstated without controls for consistently low use (3). Consistently low use of Medicare may also reflect data sampling bias for patients with coverage from other insurance programs (for example, veterans) or care from non-Medicare providers and may imply a lower chance of referral to hospice under Medicare.

Categories of illness duration (Table 1) were used to control for variations in opportunity to enroll in hospice, because some time is needed to arrive at a prognosis and enroll in hospice. Duration of illness was calculated as number of days between the date of death and earliest diagnosis of a principal condition.

Principal conditions in the last year of life were derived from expenditures and principal diagnoses recorded on claims. We adapted a plurality of physician expenditures method that places patients in leading cause-of-death categories as defined by the National Center for Health Statistics (19). Adaptation was necessary for 3 reasons. First, among causes of death derived by using the expenditure plurality method, 3 (pneumonia or influenza, accidents or adverse effects, and septicemia) are often secondary to underlying serious and chronic illnesses, and 6% of cases could not be classified. Second, we aimed for consistency with previous research that selected patients with cancer on the basis of having any cancer diagnosis in claims (3, 4). Finally, for decedents with cancer, we identified subgroups that were homogeneous in terms of survival and disease course to control for confounders associated with those factors and to allow comparison of estimated cost effects of hospice in the cancer cohort.

Consequently, after we classified decedents into condition groups by using the expenditure plurality method, we reclassified those who died of pneumonia or influenza, accidents or adverse effects, septicemia, or unclassified causes by using the plurality of principal diagnoses among all last-year-of-life claims. We then assigned decedents with a cancer classification from the expenditure plurality or the diagnosis plurality method to the cancer cohort. Finally, we divided decedents with cancer into 7 subgroups by plurality of cancer diagnosis (lung, other aggressive types of cancer with median survival <1 year [20], all other types of cancer with metastases, and all other types of cancer without metastases) and timing of diagnoses (incident [first diagnosed in the last year] or prevalent [first diagnosed before the last year]). Because nonaggressive, nonmetastatic types of cancer are fairly indolent, making incident cases uncommon, we combined patients with incident and prevalent cancer. The validity of this method is supported by consistency between the distribution of decedents among principal conditions and distributions by cause of death reported by the National Center for Health Statistics and National Mortality Follow Back Survey (19, 21).

We derived and validated an indicator of nursing home residence from physician claims in which place-of-service or evaluation and management codes indicated encounters that took place at a nursing home or skilled nursing facility. Application of these classification rules to Medicare Current Beneficiary Survey claims resulted in a value of 0.78 (95% CI, 0.76 to 0.80) between our measure and Medicare Current Beneficiary Survey facility residence status (22).

To mitigate potential selection bias, we calculated hospice use propensity scores within each cohort by using logistic regression (23, 24). The propensity score for a decedent is the predicted likelihood of hospice use conditioned on factors that are known to correlate with hospice use (3, 4, 11, 25, 26) (gender, race, Medicaid enrollment, primary condition, comorbid conditions, and urban setting) and an indicator for residence in states with high or low rates of hospice use (20% to 29% and 30% greater or less than the norm among the sample) to account for combined influences of health service supply, provider practice patterns, selection bias, and geopolitical variations that affect service supply or health coverage (for example, Medicaid hospice and home health payments, charismatic leaders, and innovative non–hospice-based end-of-life care programs).

Statistical Analysis

Interaction effects observed among age, principal condition, race, and gender prompted us to perform stratified analyses to avoid masking true effects. Within each stratum of age and cancer or noncancer status, or age and principal condition, linear regression was used to further control for propensity to use hospice, gender, race, Medicaid enrollment, urban setting, comorbid conditions, low use of Medicare, duration of illness, nursing home residence, and year of death.

We estimated the effects of hospice enrollment on Medicare program expenditures in the last year of life, overall and by service type (hospital inpatient, skilled-nursing facility, home health, outpatient facility, and physician or supplier). We modeled untransformed expenditures, expenditures truncated at the 5th and 95th percentile of the strata-specific distributions (that is, more extreme values set to those percentiles), and log-transformed expenditures and present effects from the truncated models because they moderate the disproportionate influence of very high and low use decedents while estimating mean effects in the original dollar scale. Adjusted mean ratios for Medicare program expenditures in the last year of life (hospice enrollees or nonenrollees) less than 1.0 indicated savings, and those greater than 1.0 indicated added costs for hospice enrollees compared with nonenrollees. Adjusted differences between mean expenditures incurred by hospice enrollees and nonenrollees, by service type within age and cancer or noncancer stratum, show the relation of hospice enrollment to patterns of service use and expenditures. Variations in service use and timing of entry to hospice between condition cohort and age strata inform interpretation of results.

Role of the Funding Sources

The funding sources had no role in the analyses or interpretation of the study findings. The manuscript was reviewed by the Centers for Medicare & Medicaid Services (formerly Health Care Financing Administration) to ensure the confidentiality of patients and providers.

Results

Top Methods Results Discussion Author & Article Info References

The sample was 57% female, 10% nonwhite, 72% metropolitan, 13% urbanized, and 15% rural; 24% were Medicaid enrollees (Table 1). Characteristics varied by condition cohort and age. The noncancer cohort comprised 73% of all decedents and had an overall rate of hospice use of 10%, compared with 38% for cancer decedents. Hospice use increased with age in the noncancer cohort and decreased with age in the cancer cohort. Compared with the cancer cohort, the noncancer cohort was older and more likely to be female, use Medicaid, have a longer duration of illness, have more coexisting conditions, live in a nursing home, and have consistently low use of Medicare services in the 24 months before death. In both cohorts, the percentage of women, those using Medicaid, nursing home residents, and those with consistently low use of Medicare increased with age.

Unadjusted Medicare program expenditures and associated expenditure ratios (hospice enrollees to nonenrollees) also varied by age and condition cohort (Table 2). Expenditures decreased with age in every category. Total expenditures were higher at every age for the cancer cohort than the noncancer cohort. Ratios increased with age from 1.04 to 1.22; the average was 1.13 for all ages and all conditions. These unadjusted ratios suggest that use of hospice is associated with additional costs to the Medicare program and that added costs increase with age. Comparison of ratios between the noncancer cohort (1.13 to 1.28) and cancer cohort (0.89 to 0.99) suggests that additional Medicare program costs associated with hospice are a function of hospice use in the noncancer cohort.

Similar findings were derived from estimated Medicare program expenditures that were adjusted for propensity to use hospice, gender, race, use of Medicaid, urban setting, duration of illness, comorbid conditions, consistently low use of Medicare, nursing home residence, and year of death (Table 3). Overall adjusted ratios increased with age from 0.98 to 1.16 and averaged 1.04 for all ages and conditions, suggesting that hospice incurs additional costs to the Medicare program. Comparison of ratios in the noncancer cohort (1.05 to 1.22 [average, 1.11 for all ages]) with those in the cancer cohort (0.95 to 1.06 [average, 0.99 for all ages]) suggests that additional costs are a function of hospice use by the noncancer cohort and the oldest old within the cancer cohort.

Adjusted expenditure ratios by principal condition imply that conditions indicative of multiple organ system failure (kidney disease, diabetes, cerebrovascular disease, chronic obstructive pulmonary disease, and heart disease) have less effect on added costs associated with hospice than do conditions indicative of dementia or frailty (dementia and all other noncancer diseases). In the cancer cohort, savings associated with hospice were higher for incident and more aggressive cancers. Ratios increased with age for all conditions, indicating that hospice savings decrease and added costs increase with age.

To illustrate pattern variations in use of services by condition cohort and age stratum (Figure), we examined differences between estimates for adjusted mean expenditures among hospice enrollees and nonenrollees by type of expenditure (total, hospital inpatient, skilled-nursing facility, home health, outpatient facility, or physician or supplier). The 95% CIs imply that estimated differences are relatively precise, and most are statistically significant (Table 4). The average hospice enrollee without cancer incurs Medicare costs that are about $2579 more in the last year than those of the average nonenrollee, and their additional costs increase with age from $1356 at 68 to 79 years of age to $3725 at 85 years of age or older. The average hospice enrollee with cancer appears to save Medicare $648 in the last year of life compared with the average nonenrollee. The estimated savings of $1703 among enrollees with cancer who are 68 to 79 years of age more than offset the additional cost of $1193 among those 85 years of age or older. For decedents in the cancer cohort who were younger than 85 years of age, the cost of hospice is offset by savings in all other expenditures except outpatient facility use. In contrast, the estimated savings among the noncancer cohort for hospital inpatient, skilled-nursing facility, outpatient facility, and physician or supplier services do not offset the costs of hospice and increased spending for home health care among hospice enrollees.

View larger version (24K): [in this window] [in a new window]   Figure. Estimated effect of hospice on adjusted mean Medicare program expenditures in the last year of life by type of Medicare expenditure within condition cohort and age stratum.

Examination of volume, intensity, and timing of service use indicates that the noncancer cohort tended to use fewer but more intensive hospital and hospice services in the last year of life. On average, patients in the noncancer cohort spent 4 fewer days as a hospital inpatient than did those in the cancer cohort (16 vs. 20 days), yet patients in both cohorts spent an average of 3 days in a hospital intensive care unit. Hospice enrollment accounted for an average decrease of 6 days in hospital and 2 days in the intensive care unit among the cancer cohort but had little effect among the noncancer cohort. Mean per diem hospice expenditures were higher among enrollees without cancer ($155) than those with cancer ($136). Finally, entry to hospice in the last week of life was more prevalent among enrollees without cancer than those with cancer (36% vs. 23%). Because of higher hospice costs and lower expenditures for other services among the noncancer cohort (Table 2), hospice has reduced opportunities to show savings.

The results reported above persisted even after we varied our methods, although these alternative strategies affected the exact estimates. Specifically, population characteristics would not change substantially if the continuous enrollment requirement were 12 rather than 36 months. The model presented has the same overall implications as do models with untransformed or log-transformed expenditures. Overall implications are also consistent with those of models that exclude hospice enrollees with discontinuous stays and models that do not correct for consistently low Medicare use.

Discussion

Top Methods Results Discussion Author & Article Info References

Hospice is cost-neutral to cost-saving for persons who die of cancer and generally yields added costs for those who do not die of cancer. The latter is the fastest growing group of hospice enrollees (11). Overall, hospice users incur an estimated 4% greater costs than do similar patients who do not use hospice. Savings decrease and relative costs increase with age at death. Medicare expenditures near death decrease with age (27-30), whereas hospice expenditures remain relatively constant. Medicare expenditures were lower at every age and mean per diem hospice costs were higher in the noncancer cohort than the cancer cohort. Consequently, the potential for hospice savings was lower for the noncancer cohort.

Our findings confirm and update those of other studies. Studies in patients with cancer in the late 1980s to early 1990s indicate that hospice may have only small effects in the last year of life (2-4). More recent descriptive studies suggest that hospice may increase costs in the last year of life for persons who do not die of cancer (14, 31).

The relative costs of hospice are highest among patients with dementia and relatively nonspecific diagnoses and intermediate among patients with organ system failures. Hospice-related savings are often realized among patients with cancer. This pattern probably reflects differences in service needs (31-37) and certainty of prognosis (38-42) that are associated with 3 major trajectories to death: a short period of obvious decline at the end of life, which is typical of cancer; long-term disability with exacerbations and unpredictable timing of death, which is typical of chronic organ system failure; or persistent decline and deficits in self-care associated with frailty or dementia (37, 43, 44). Effective and reliable care for persons approaching death may require organization and financing of care that match these trajectories (4, 37, 43, 45-60).

Earlier entry to hospice in the noncancer cohort may appear to be a way to reduce added costs associated with hospice care. However, earlier entry may not reduce costs or be achievable. The costs of hospice may exceed the costs of services avoided by earlier entry. Because prognoses in the noncancer cohort are typically uncertain (38-42), patients and their physicians may be unable or unwilling to determine or accept a 6-month prognosis or to forgo "curative treatment for their terminal illness," as required for hospice eligibility.

Our study has some limitations. First, although our methods offer improved control for selection bias and other confounders that limited previous research, some selection bias and confounding inevitably remain. Second, findings pertain to the Medicare program only and do not consider the effect of Medicare's hospice benefit on expenditures in the last year of life by patients and their families, Medicare for caregivers (61, 62), Medicaid, or other public or private payers. Finally, judging the merits of the hospice benefit requires understanding of the effect of hospice on quality of life, the impact of the Medicare hospice benefit on expenditures from all sources, and alternatives for organization and financing of end-of-life care. Table 5 shows key conclusions and implications for policy and future research.

Most Americans are seriously and chronically ill in the years before death. Sustainable comprehensive services are required that ensure comfort, advance planning, closure, and family support, in addition to treatment for medical conditions. Even if hospice care costs somewhat more than conventional care, its comprehensiveness and continuity may merit those costs.