Patterns of Healthcare Spending in the Last Year of Life

Abstract

The underlying assumption that healthcare spending skyrockets at the end-of-life may suggest that policymakers should target the last few months of life to control costs. However, spending patterns leading up to death have not been fully examined. We applied a new methodology to administrative claims data for older Medicare beneficiaries who died in 2012 to characterize trajectories of healthcare spending in the last year of life. After adjustment, we identified four unique spending trajectories among decedents: 48.7 percent had High Persistent spending, 29.0 percent had Moderate Persistent spending, 10.2 percent had Progressive spending, and only 12.1 percent had Late Rise spending. High spending throughout the full year before death (approximately half of all decedents) was associated with having multiple chronic conditions but not any specific diseases. These findings suggest that spending at the end-of-life is a marker of general spending patterns often set in motion long before death.

End-of-life care accounts for approximately 30 percent of national Medicare spending and, as such, continues to have the attention of healthcare policymakers, payers, and providers.^1,2 While decedents represent only a small fraction of all Medicare beneficiaries, Medicare spends on average $40 to $50 thousand per decedent in the last year of life compared to only $7 thousand per year for survivors.^1,3,4 Underpinning health policy debates regarding end-of-life treatment intensity,^5,6 advanced directives,^7,8 and hospice care,¹ is the underlying assumption that adults approaching death will exhibit a steep rise in healthcare spending in the final months of life.⁹ However, this notion that healthcare spending skyrockets at the end of life has not been directly examined.

Aging of the US population compounds concerns about the high cost associated with the end-of-life care as the baby boomers enter Medicare.¹⁰ Despite the relevance of healthcare spending for individuals leading up to death to national healthcare reform, little is known about the patterns or trajectories of their spending at the end of life. It should not be a surprise that people who are sick and dying consume high healthcare resources. The question remains whether the trajectories of spending at the end-of-life vary in meaningful ways. Prior work has suggested that certain diseases are associated with a pattern of decline that may be reflected in spending.^11-13 More detailed information about the course of spending near the end-of-life may provide important insights about the drivers of end-of-life spending and shed light on potential strategies to mitigate the costs while preserving high quality care for people who are dying.¹⁴

Therefore we sought to identify trajectories of Medicare spending among older decedents. To examine potential factors related to observed trajectories of end-of-life spending, we also examined how those trajectories are related to number and type of health conditions, and the type of services that account for the spending. We also examined the relationship between spending trajectories and total Medicare spending per enrollee by US state.

Study Data and Methods

Study Population

We used a random sample of Medicare beneficiaries who died in 2012 to identify patterns of spending in the last year of life. Our study used 2011-2012 data from CMS’s Master Beneficiary Summary (MBS), Part B Carrier, MedPAR, Hospice, Inpatient, and Outpatient files. This study received institutional review board approval.

Among all Medicare beneficiaries continuously enrolled in both Medicare Part A and Part B and not enrolled in a Medicare Advantage throughout 2011-2012, we identified those who died using the reported date of death in the 2012 MBS file. Next, we restricted our sample to those aged 6699 years old as of January 1, 2012 that resulted in identification of 1,266,954 decedents. From older Medicare decedents in 2012, we drew a random sample of 100,000 decedents resulting in, after removing those with missing values for sociodemographic measures, a final sample of 99,848 decedents. For each decedent we gathered a variety of measures including sociodemographic characteristics, conditions based on diagnoses on healthcare claims, and health service use.

Study Variables

Healthcare Spending in the Last Year of Life

For each decedent we determined daily healthcare spending (except for prescription medications) in the last year of life. To do so, we used the date of service on each claim relative to the date of death. For measures that spanned multiple days (e.g., an inpatient hospital stay), spending was uniformly distributed over the service interval. Using established methods that account for variation in Medicare reimbursement, we then price standardized healthcare spending.¹⁵ Daily total healthcare spending was then aggregated according to month before death.

Health Conditions in the Last Year of Life

We used the Agency for Healthcare Research and Quality’s Clinical Classification Software (CCS) system for grouping International Classification of Diseases, Ninth Revision Clinical Modification (ICD–9–CM) codes into meaningful disease categories.¹⁶ Decedents had to have one inpatient or two physician outpatient claims seven days apart in the last year of life to be labeled with a specific diagnosis. From the 260 CCS diagnostic categories, we selected those that represent the leading causes of death among older US adults¹⁷ including: cancer (solid tumors versus other types of cancer), cardiovascular disease (congestive heart failure, myocardial infarction, stroke, other cardiovascular disease), and other types of organ failure (respiratory, renal, liver failure) (Appendix, Exhibit 1).¹⁸ Decedents could have more than one of these conditions thus we also counted the total number of conditions from this list for each decedent, aggregated to the following categories: 0 or 1, 2 or 3, versus 4 or more conditions.

Health Service Use and Place of Death

Using date of service, we identified when the health service use occurred in relationship to date of death. In the last year of life, we report the number of outpatient visits to primary care physicians, specialists, and emergency departments identified in the Part B Carrier and Outpatient files. To identify the number of visits to primary care physicians we used Medicare specialty codes on ambulatory claims that correspond to internal medicine, family medicine, general practice, and geriatric medicine and all others were considered medical or surgical specialists.^19,20 Using data from the MedPAR files we identified the total number of days spent as an inpatient, in an intensive care unit, or at a skilled nursing facility (SNF) .

We report several measures that reflect care more specific to end-of-life, including whether on the date of death the beneficiary had been in a inpatient setting reimbursed by Medicare including a hospital stay without ICU, a hospital stay with ICU, or other type of inpatient stay (primarily SNF) according to claim dates. If no claims were on the date of death, the place of death was considered unknown (the decedent likely died in the community or long-term care facility). We also identified whether decedents enrolled in hospice in the last six months of life.²¹ Finally, we determined the use of life-prolonging treatments including: intubation and mechanical ventilation (ICD-9-CM codes: 96.04, 96.05, 96.7×), gastrostomy feeding tube (ICD-9-CM codes: 43.2, 43.11, 43.19, 43.2, 44.32), and hemodialysis (ICD-9-CM code: 39.95).⁵

Analyses

Identification of Spending Trajectories

We used group-based trajectory modeling to identify distinct patterns of Medicare spending among decedents (Appendix, Exhibit 2) .^18,22 According to spending trajectory assignment, we then examined the decedents’ conditions and use of health services. As healthcare spending at the end-of-life varies significantly by sociodemographic characteristics^23-25 all of these models were adjusted for decedent age, sex, and race/ethnicity. While in our primary analysis adjustment for these factors is appropriate, we also examined how sociodemographic characteristics vary across spending trajectories and tested whether the trajectory models were different when fit separately by decedent age, sex, or race/ethnicity.

Across spending trajectories, we used analysis of variance (ANOVA) to compare means, χ²-test to compare proportions, and Kruskal Wallis test to compare medians. A two-sided p-value of less than 0.05 was considered statistically significant. We assumed any missing values to be missing completely at random and all analyses were based on complete case analysis.

Association with State Medicare Spending

To determine the extent to which the relative percent of decedents decedents in each spending trajectories is related to general regional healthcare spending, we examined relationship between the proportion in each spending trajectories and total price, age, sex, and race adjusted Medicare spending per enrollee by US state.²⁰ We used Spearman correlation and simple linear regression to examine associations. For these analyses, only the 45 states with a sample of at least 300 decedents were used. Data management of CMS administrative claims data were conducted using SAS, version 9.4 (Cary, NC) and statistical analyses were performed in Stata MP, version 13.1 (College Station, TX) .²⁶

Limitations

There are several limitations of our study that must be acknowledged. First, our study was limited to examining Medicare spending among older US adults. This limits the generalizability of our findings and thus patterns may vary, for instance, among younger adults and older adults enrolled in managed care. Second, our analyses do not take into account out-of-pocket or medication spending.²⁷ Lastly, we are unable to capture detail on cause of death and other important factors such as functional status and use of advanced directives in administrative data.

Results

Among our study population of 99,848 older decedents the mean age was 83.2 years, 55.5 percent were female, and 88.1 percent were Non-Hispanic White (Appendix, Exhibit 3).¹⁸ The majority of decedents did not die in a Medicare-reimbursed facility (69.2 percent) and 43.0 percent enrolled in hospice.

Spending Trajectories

We identified four distinct trajectories of spending in the last year of life including what we refer to as: High Persistent, Moderate Persistent, Progressive, and Late Rise spending (Exhibit 1). Nearly half of decedents were classified as having High Persistent spending in the last year of life (48,605 out of 99,848) characterized by having high initial and steadily increasing spending throughout the last year of life. Approximately 10 percent of decedents exhibited a Progressive pattern where spending started relatively low but increased steeply throughout the time period. Moderate Persistent spending (29.0 percent of decedents) was characterized by moderately high spending initially followed by dip and then increase in the last four months of life. Lastly, in 12.1 percent of decedents spending was very low up to four months before death and then increased exponentially – Late Rise spending. There were small differences in percentage of age, sex, and race of decedents across trajectories (Appendix, Exhibit 3).¹⁸

EXHIBIT 1. Spending trajectories of Medicare decedents in the last year of life.

Source: Authors’ analyses of Medicare data

Note: Solid lines represent observed trajectories and dashed lines represent predicted trajectories that are adjusted for age, sex, and race/ethnicity.

There was substantial variation in absolute dollars of Medicare spending in the last year of life across the four spending trajectories (Exhibit 2). The High Persistent Spenders had significantly greater spending than other trajectories (p-value < 0.001) – for instance the median total Medicare spending in the last year of life among High Persistent Spenders was $59,394 compared to $37,036 among Progressive Spenders. Late Rise Spenders had the steepest curve but substantially lower total Medicare spending in the last year of life (median spending $11,166). U.S. States total Medicare spending per enrollee was strongly associated with the percentage of decedents in the High Persistent Spenders category (Spearman rs = 0.78, p-value < 0.001) by US state (Appendix, Exhibit 4).¹⁸

EXHIBIT 2. Mean and median total Medicare spending in the last year of life according to spending trajectory.

Source: Authors’ analyses of Medicare data

Note: Kruskal Wallis test used to compare median total spending across trajectories.

Spending Trajectories According to Health Conditions

Among decedents with the same health condition, the percentage of decedents in each spending trajectory varied little; however, we observed large differences when examined by the total number of conditions (Exhibit 3). Decedents with fewer conditions (i.e., those with either one or no conditions) more often had Late Rise spending (21.6 percent compared to 6.3 percent among those with four or more conditions) and Moderate Persistent spending (36.6 percent compared to 21.3 percent among those with four or more conditions), p-value < 0.001. Decedents with two or three conditions had the highest proportion of Progressive spending (31.2 percent). Decedents with four or more conditions had the largest proportion of High Persistent spending (62.6 percent)

EXHIBIT 3. Distribution of spending trajectories in the last year of life according to health conditions.

Source: Authors’ analyses of Medicare data

Abbreviations: CHF, congestive heart failure

Use of Health Services According to Spending Trajectory

High Persistent Spenders had greater use of all health services except hospice in the last year of life (Exhibit 4). Most notable were approximately twice the number of outpatient visits to specialists (mean of 6.3 visits versus less than 3 among other spending trajectories), higher number of inpatient days, and days in SNFs, p-value < 0.001 for all. Progressive Spenders exhibited some similarities to High Persistent Spenders in terms of use of inpatient health services but were most likely to use hospice (p-value < 0.001). Late Rise Spenders had significantly lower use of both outpatient and inpatient health services in the last year of life. High Persistent Spenders were more likely to have used life-prolonging treatments than people in other spending trajectories (i.e., use of mechanical ventilation, feeding tube, and hemodialysis). Place of death was similar across spending trajectories – the only noticeable difference being higher proportion of deaths occurring during a hospital stay that included ICU use among Late Rise Spenders.

EXHIBIT 4.

Use of health services in last year of life according to spending trajectory

	Spending Trajectory
	High Persistent	Moderate Persistent	Progressive	Late Rise
Sample Size, No.	48,605	28,976	10,198	12,069
Place of Death, %
Hospital without ICU stay	7.3	8.3	6.7	7.4	***
Hospital with ICU stay	15.2	17.0	15.0	18.5
Other inpatient	7.4	6.7	8.4	6.2
Unknown	70.1	68.0	70.0	68.0
Hospice Use in Last Six Months, %
Yes	45.9	39.5	52.1	32.6	***
No	54.1	60.5	47.9	67.4
Health Service Use in Last Year
Outpatient services, mean No.
Visits to PCPs	8.7	6.2	5.4	2.1	***
Visits to specialists	6.3	2.5	2.8	0.7	***
ED visits	1.2	0.7	0.9	0.5	***
Inpatient services, mean No.
ICU days	5.7	2.5	4.4	2.2	***
Inpatient days	15.0	6.3	11.2	5.1	***
SNF days	27.1	6.5	18.9	3.1	***
Life-prolonging treatment, % who used
Ventilator	19.3	13.4	15.3	15.6	***
Feeding tube	6.4	2.4	6.1	2.4	***
Hemodialysis	19.3	13.4	15.3	15.6	***

Source: Authors’ analyses of Medicare data

Abbreviations: PCP, primary care physician; SNF, skilled nursing facility; ED, emergency department; ICU, intensive care unit

^***Note: indicates a p-value < 0.001 across spending trajectories. Analysis of variance (ANOVA) used to compare means and χ²-test to compare proportions.

Influence of Sociodemographic Characteristics on Spending Trajectory

While our goal was to create aggregated trajectory patterns for all Medicare decedents, prior literature suggests there may be differences by age, sex, or race in the way decedents use health services at the end of life.^23-25 To test whether spending trajectories differ by these factors models were fit for each sociodemographic group. The main differences were that for decedents older than 75 and for Non-Hispanic Whites compared to Non-Hispanic Blacks or other racial groups, the number of trajectories dropped to three with loss of the Progressive Rise category (Appendix, Exhibits 5-7).¹⁸ Future examination of variability in end-of-life spending trajectories according to sociodemographic characteristics is warranted.

Discussion

In this study we identified four unique Medicare spending patterns among older adults in the last year of life. We found the trajectory with the largest proportion of decedents to be High Persistent spending – decedents who had a high spending pattern throughout the entire last year of life. This finding challenges the idea that healthcare spending skyrockets for those immediately nearing death.²⁸

Overall, High Persistent spenders used more outpatient services but their use of inpatient services was comparable to those who had Progressing spending over the last year of life. Only 12.1 percent of decedents in our sample were assigned the least expensive pattern of end-of-life spending: Late Rise spending. The distribution of trajectories differed little according to specific health diseases, but was strongly related to the total number of chronic health conditions. At the US state level, we found higher total Medicare spending per all enrollees to be associated with a high percent of High Persistent decedents, whereas other trajectories such as Late Rise were associated with lower spending.

Trajectory classification has previously been applied for examining disability and functional decline.^29-31 One prior analysis has profiled decendents,³² yet to our knowledge this is the first study to use group-based trajectory modeling to classify patterns of end-of-life spending. Nearly half of older decedents in our study, however, were high spending throughout the entire last year of life. This finding challenges a widely-held notion that most end-of-life costs accrue in the last few weeks to months and could point to very different solutions for how one would influence end-of-life care.³³ Accounting for nearly half of all Medicare spending, inpatient services are often targeted for reduction by moving people toward more palliative settings.³⁴ Yet efforts focused on reducing reliance on costly inpatient and intensive services only in the last few months of life will likely have less effect on overall costs than might be anticipated.

Specific diseases are associated with certain patterns of decline,^11-13 but we find that diseases are not associated with particular spending trajectories. Instead, it is the number of conditions that appear to be most relevant to end-of-life spending. The importance of multiple chronic conditions for spending aligns with other studies that show the presence of multiple chronic conditions itself, whether at the end-of-life or otherwise, is a critical factor in healthcare spending.^4,35-38

The association between High Persistent Spending with having four or more chronic conditions provides a clue where more effective strategies might be deployed. People in the high spending trajectories also used more ambulatory health services including both primary care and specialty visits. The high frequency of these visits affords the potential for intervening upstream of the hospital among people with multiple conditions and at risk of death. Among these people who have a high burden of chronic illness and functional impairment⁴³ it may be harder for patients, families and providers to recognize when a final turn for the worse is occurring. Improving care coordination, fragmentation across providers, and linkages to supportive services for these high cost, high need beneficiaries in general over a longer time span, rather than focusing on the last months of life, may improve care experience and reduce overall cost and reliance on inpatient services.

We also observed an interesting pattern in the approximately 10 percent of decedents who had Progressive spending, decedents who spent little early in the year and steadily increased (linearly) to death. Although we can only speculate, this pattern could reflect situations where patients, families and doctors discern more clearly clinical progression since these decedents were most likely to enroll in hospice. This pattern is in contrast to the Late Rise decedents in whom multiple chronic conditions were less frequent and may indicate people who experience a sudden health event that more rapidly leads to death or for which the patient and family opted for a less intensive approach in anticipation of death.

While the trajectories differentiate between individual spending patterns, we found that the proportion of decedents in each trajectory across regions varied as well. Regional differences in level of end-of-life spending are not fully explained by variation in illness^39,40 or patient preferences.⁴¹ The correlation of regional spending at the end-of-life with total state Medicare spending per enrollee suggests that end-of-life care patterns may be a signal for general approaches to intensity of service provision.^5,42 Our findings support this idea by showing that there is two-fold variation in the proportion of decedents who are High Persistent spending across regions and the distribution of spending trajectories is correlated with total population spending.

Insight into the patterns of healthcare spending occurring at the end-of-life has important health policy implications. High Persistent spending that is strongly associated with higher regional spending overall, appears to be set in motion before the last year of life and with having multiple chronic conditions. This finding suggests that health policy strategies focused on improving quality and reducing cost for those immediately approaching death may have less of an impact than originally anticipated.^43,44 Strategies such as Hospice that target those exhibiting a rapid clinical decline (with a prognosis of less than six months) may apply to fewer than 20 percent of older decedents (i.e., potentially those in our analyses with Progressive or Late Rise spending). Therefore, future strategies directed at meeting the care needs for older adults with multiple chronic conditions – not necessarily for those with a poor immediate prognosis – could have the largest impact on national spending.⁴⁵

Appendix, Exhibit 1. Criteria based on the Agency for Healthcare Research and Quality’s Clinical Classification Software used to identify health conditions in last year of life

Condition	Criteria
Cancer, solid tumor	Cancer of specific organ (CCS11-CCS38)
Cancer, other types	Leukemias and myelomas (CCS39 & CCS40) or other malignant neoplasms (CCS41-CCS44)
Congestive Heart Failure	Congestive heart failure (CCS108)
Cardiac Arrest	Acute myocardial infarction (CCS100) or cardiac arrest and ventricular fibrillation (CCS107)
Stroke	Acute cerebrovascular disease (CCS109) or occlusion/stenosis of precerebral arteries (CCS110)
Other Cardiovascular	Heart valve disorders (CCS96); peri-, endo-, and myocarditis, or cardiomyopathy (CCS97); pulmonary heart disease (CCS103); or cardiac dysrhythmias (CCS106)
Respiratory Failure	Chronic obstructive pulmonary disease and bronchiectasis (CCS127); pleurisy, pneumothorax, pulmonary collapse (CCS130); or respiratory failure, insufficiency, arrest (CCS131)
Renal Failure	Chronic renal failure (CCS158) or acute and unspecified renal failure (CCS157)
Liver Failure	Liver disease, alcohol related (CCS150) or other liver diseases (CCS151)

Appendix, Exhibit 2. Model specifications and sensitivity analyses

Model Specifications

Group-based trajectory modeling is an application of finite mixture modeling where the model assumes a priori that the population is composed of a mixture of distinct groups defined by their trajectories. Trajectory groups are conceptually thought of as latent longitudinal strata where population variability is captured by differences across groups in the shapes of the curves. The number of groups is determined in a data-adaptive fashion using model selection criteria. Each subject is assigned a trajectory group based on the probability of membership.

We used an approach developed by Nagin to identify and assign distinct healthcare spending trajectories to decedents.^1-3 As healthcare spending is highly skewed, we used the natural logarithm transformation to normalize spending data. For decedents who had $0.00 spending on a given month, we assigned a value of $1.00 prior to the transformation to facilitate the logarithmic transformation.

To arrive at the optimal number of distinct groups that best fit our data, we constructed models using two to six groups. We also investigated the unique combination of functional forms (i.e., intercept-only, linear, quadratic, versus cubic) that resulted in the best fitting model. The final model was selected using Bayesian Information Criterion.⁴ As healthcare spending at the end of life varies significantly by sociodemographic characteristics^5-7 all of these models were adjusted for decedent age, sex, and race/ethnicity. Age was included in models as a continuous variable and race/ethnicity was included as a nominal categorical variable with Non-Hispanic White as the reference category.

Sensitivity Analyses

Although our primary analysis was adjusted for Medicare decedent sociodemographic characteristics, we also performed sensitivity analyses by fitting trajectory models separately according to age (young versus old), sex, and race/ethnicity (Non-Hispanic White, Non-Hispanic Black, and Other). For these analyses younger Medicare decedents were defined as those between 65 to 74 years of age and older Medicare decedents as those ≥ 75 years of age. For selecting the optimal model in regards to the number of groups and functional forms, we performed the identical approach as for our primary analysis. Results of our sensitivity analyses can be found in Appendix Exhibits 5-7.

Appendix, Exhibit 3. Characteristics of Medicare decedents according to spending trajectory

		Spending Trajectory
Characteristic	All	High Persistent	Moderate Persistent	Progressive	Late Rise
Sample Size, No.	99,848	48,605	28,976	10,198	12,069
Sociodemographic Characteristics
Age, mean years (SD)	83.2 (8.5)	82.4 (8.3)	85.1 (8.2)	83.1 (8.6)	82.0 (9.2)	***
Sex, %
Male	44.5	43.8	41.1	46.5	53.6	***
Female	55.5	56.2	58.9	53.5	46.4
Race and ethnicity, %
Non-Hispanic white	88.1	86.8	91.0	87.6	86.5	***
Non-Hispanic black	8.0	9.0	5.7	8.5	8.6
Other	3.9	4.2	3.2	3.9	4.9
US region of residence, %
Northeast	19.0	19.4	20.1	17.1	16.6	***
Midwest	24.8	23.6	26.5	25.0	25.2
South	40.3	42.3	37.4	40.7	38.7
West	15.9	14.7	16.0	17.2	19.5
Place of Death, %
Hospital without ICU stay	7.6	7.3	8.3	6.7	7.4	***
Hospital with ICU stay	16.1	15.2	17.0	15.0	18.5
Other inpatient	7.2	7.4	6.7	8.4	6.2
Unknown	69.2	70.1	68.0	70.0	68.0
Hospice Use in Last Six Months, %
Yes	43.0	45.9	39.5	52.1	32.6	***
No	57.0	54.1	60.5	47.9	67.4

Source: Authors’ analyses of Medicare data

Abbreviations: SD, standard deviation; ICU, intensive care unit

^***Note: indicates a p-value < 0.001 across spending trajectories. Analysis of variance used to compare means and χ²-test to compare proportions.

Appendix, Exhibit 4. Relationship between percent of High Persistent (A), Moderate Persistent (B), Progressive (C), and Late Rise (D) decedents and total Medicare spending per enrollee by US state

Source: Authors’ analyses of Medicare data

Abbreviations: r_s, Spearman correlation coefficient

Note: Only the 45 states with a sample size of least 300 decedents are plotted. All correlations are highly significant at p-value < 0.001.

Appendix, Exhibit 5. Spending trajectories of young (A) versus old (B) Medicare decedents in the last year of life

Source: Authors’ analyses of Medicare data

Note: Young Medicare decedents defined as 65 to 74 years of age; Old Medicare decedents defined as ≥ 75 years of age. Solid lines represent the observed trajectories and dashed lines represent the predicted trajectories.

Appendix, Exhibit 6. Spending trajectories of male (A) versus female (B) Medicare decedents in the last year of life

Source: Authors’ analyses of Medicare data

Note: Solid lines represent the observed trajectories and dashed lines represent the predicted trajectories.

Appendix, Exhibit 7. Spending trajectories of Non-Hispanic White (A), Non-Hispanic Black (B), versus Other Race and Ethnicity (C) Medicare decedents in the last year of life

Source: Authors’ analyses of Medicare data

Note: Solid lines represent the observed trajectories and dashed lines represent the predicted trajectories.