Developing Models to Predict Persistent High-Cost Cases in Florida Medicaid

Abstract

This paper examined individual characteristics associated with being a high-cost case in multiple years for Medicaid-covered health care services. In addition, the accuracy of models that predict future persistent high-cost cases was examined. Florida Medicaid claims from 2005 to 2010 were used to examine characteristics, diagnoses, and services associated with individual costs being in the top 1% of recipients. Regression models were estimated with diagnoses and service use in a base year used to predict future high-cost cases. Several different perspectives were used that focus on predicting current year high-cost cases based on prior persistence, predicting future persistence of high costs, and a combination of using past persistence to predict future persistence. Average annual costs for persistent high-cost cases were more than $140,000. Overall, models were predictive of future high-cost cases. The receipt of intermediate case facility (mental retardation) services was the strongest predictor of future high-cost cases. Inpatient, outpatient, pharmacy, and nursing home services, along with diagnoses, all provided important information for predicting high-cost cases. Diagnosis-based models in conjunction with prior costs can predict future high-cost cases with a high degree of accuracy. However, given that many high-cost cases reside in intermediate care facilities, it is not clear that such individuals would benefit from intensive case management. Service use patterns in prior years, diagnoses, and prior costs should all be used to identify individuals who may benefit from intensive case management. (Population Health Management 2015;18:467–476)

Introduction

A high proportion of health care expenditures are concentrated among a small percentage of individuals. Residents of the United States who were in the top decile of health care costs in 2002 accounted for 64% of all expenditures, while the bottom 50% of the distribution accounted for only 3% of all expenditures.¹ Health care expenditures are high for individuals with persistent conditions. Forty-one percent of individuals in the top decile in one year remained in the top decile in the following year,² and 5 conditions (heart disease, cancer, trauma, mental disorders, and pulmonary conditions) accounted for nearly one third of all expenditures.³ The concentration of spending on relatively few people and conditions is found among both the privately and publicly insured. For example, the 24% of the Medicaid population who are disabled or elderly account for 63% of expenditures.⁴

Consequently, several studies have examined methods to identify individuals at risk for high costs.^5,6,7 Such studies suggest that although prior cost is a predictor of future cost, diagnosis-based models improve prediction of individuals at risk for high future expenditures. Individuals predicted to be high cost by diagnosis-based models had higher future costs than a prior cost model, and were more likely to have a disease commonly targeted by disease management programs (eg, diabetes, congestive heart failure, asthma, depression).⁶ In an exception to the general finding, one study that examined behavioral health costs found prior cost to be a better predictor of future high costs.⁸ Diagnosis-based models may have superior predictive power because prior costs can be high because of acute care needs that will not persist over time.

Given their predictive power, diagnosis-based models are frequently used to risk adjust capitated payments for Medicaid and Medicare managed care organizations and prescription drug plans.^9,10,11 For example, the Medicare Advantage program uses the CMS-HCC (Centers for Medicare & Medicaid Services—Hierarchical Condition Categories), the Medicare part D program uses the CMS-RxHCC model, and several states use the Chronic Illness and Disability Payment System or Adjusted Clinical Groups models to risk adjust payments to managed care organizations. The state of Florida has used the MedRx model since introducing Medicaid Reform in 2006, which differs from diagnosis-based models in that it uses prescription drug claims instead of diagnosis to assign diseases and predict future costs.

In addition to being used for Medicare and Medicaid payment systems, case managers can use information from risk-adjustment models to prospectively identify individuals for case review to determine whether they may benefit from additional intervention programs⁶ High-cost cases may offer opportunities for improvements in quality of care and for cost reduction. For example, Kaiser Permanente implemented an Assertive Community Treatment program with identified high-cost users of behavioral health care.¹² Key to implementing such a program is being able to identify who is at risk for high costs.

The importance of identifying individuals at risk for high costs is apparent. However, prior research has focused on a short time frame for examining costs (1 or 2 years) and/or focused on behavioral health expenditures only. In addition, there are well-documented challenges to using diagnosis-based models or prior costs to predict high-cost cases. Diagnosis-based models assume homogeneity within diagnostic groups with the goal of predicting the average costs for a diagnosis. Consequently, individuals who have relatively mild cases of a disease typically associated with very high costs may be incorrectly identified as a high-cost case.⁵ In addition, individuals with a serious case of a disease typically associated with lower costs may not be identified as a high-cost case despite having persistent high costs. Thus, a diagnosis-based model can systematically misclassify some individuals as high-cost cases, while missing others that should be classified as high cost. However, prior costs methods also can misclassify individuals. For example, prior costs can be high because of a high-cost hospitalization in one year and may not continue to be high in future years. Thus, both diagnosis-based and prior cost methods have shortcomings in identifying high-cost cases.

This paper advances the literature by looking at a 6-year time frame (2005–2010) using Florida Medicaid data to identify persistent high-cost cases. Instead of using a prior cost or diagnosis-based model, a combination of the 2 models is used to improve predictive power. Also, rather than looking at how characteristics in a single year predict costs for the following year, 4 different perspectives are used to examine persistence of high costs where persistence is defined as being a high-cost case in multiple years. First, similar to the prior literature, a model was developed to predict which individuals who were high-cost cases in 2009 remained a high-cost case in 2010. Second, the author considers whether the persistence of high costs in the preceding 5 years is predictive of being a high-cost case in 2010. Third, the author examined the persistence of high costs in the following 5 years among individuals who were high-cost cases in 2005. Fourth, the perspectives were combined to examine the predictive power of models that use the persistence of high costs in the preceding 3 years to predict the persistence of high costs in the following 3 years.

The different perspectives are considered because interventions may have different goals ranging from immediate savings that require prediction of high-cost cases in the following year, to changing the trajectory of expenditures for individuals expected to be high cost in the future, which requires identification of longer term high-cost cases. The development of models that can identify individuals at risk for persistent high costs might enable the development of cost-effective interventions to reduce long-term costs.

Methods

Data

Florida Medicaid enrollment and claims files from January 2005 through December 2010 were the data source for this paper. Critical enrollment data included beneficiary demographics (age, sex, race, and eligibility status), Medicaid coverage periods, and managed care coverage periods. Medicaid payments and International Classification of Diseases, Ninth Revision (ICD-9) diagnoses were available from inpatient, outpatient, and physician settings. (Throughout the paper, costs refer to expenditures by the Medicaid program.) Payments also were available from prescription drug claims. Actual payments were used instead of provider charges or allowed charges. Each sample included Medicaid beneficiaries with fee-for-service coverage for at least 6 months in all 6 years. Six months was selected instead of 12 because some high-cost cases were not enrolled for the full year and the author did not want to exclude them from the study. On the other hand, 6 months of enrollment were required to avoid having individuals with limited enrollment overly influence the distribution of expenditures. Only fee-for-service coverage periods were included because complete managed care data were not available for the time period.

Analytic methods

Descriptive statistics are presented for the entire sample and for individuals who were high-cost cases in any of the 6 years, and for individuals who were high-cost cases in all 6 years. A high-cost case was defined as having costs that are in the top 1% of all recipients in the year.

The first perspective focused on predicting whether an individual was a high-cost case in 2010. Logistic regressions were estimated because the dependent variable was dichotomous. Logistic regression is an appropriate analytical technique to measure the relationship between a categorical dependent variable and 1 or more independent variables as it estimates probability scores that range from zero to 1. Thus, the probability score represents the predicted probability that an event will occur. Ordinary least squares would be problematic as predicted values be less than zero or greater than 1. For the first regression, the sample was limited to individuals who were high cost in 2009. This regression was similar to much of the prior research on predicting high-cost cases using diagnosis-based models. The second regression examined individuals who were persistent high-cost cases, defined as being a high-cost case in every year from 2005 to 2009. For each regression, the probability of being a high-cost case in 2010 was estimated based on individual characteristics and diagnoses: \documentclass{aastex}\usepackage{amsbsy}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{bm}\usepackage{mathrsfs}\usepackage{pifont}\usepackage{stmaryrd}\usepackage{textcomp}\usepackage{portland, xspace}\usepackage{amsmath, amsxtra}\pagestyle{empty}\DeclareMathSizes{10}{9}{7}{6}\begin{document} \begin{align*}High \ cost_{i , t + 1} = X_{it} \beta + \mu_{it} \tag{1}\end{align*} \end{document}

where i denotes individuals and t time, μ_i is normally distributed with mean zero and a homogenous variance, X includes demographics, dichotomous diagnostic variables, and dichotomous service variables. Demographics include age, race (Asian, black, Hispanic, other; white is the reference category [The Florida Medicaid race variable has a number of individuals with unknown race who are likely to be Hispanic. However, instead of simply assuming they are Hispanic, the race variable was coded as “other.”]), sex (male), and eligibility status (Supplemental Security Income [SSI], long-term care). A diagnosis-based model requires that ICD-9 diagnoses be clustered within diagnostic groups. The final model in this paper contained 18 diagnostic categories (see Table 1 for a complete list of diagnostic categories). An individual with comorbidities would have each disorder counted in the model. Service categories associated with physical health diagnoses include inpatient, outpatient, pharmacy, nursing home, hospice, intermediate care facilities (ICF) for mental retardation, and other services. Variables also were included for inpatient, outpatient, and pharmacy services associated with behavioral health diagnoses.

Table 1.

Sample Demographics

	Full sample		Never high cost		Years high cost ≥1 and ≤5		Always high cost
	Mean/%	SD	Mean/%	SD	Mean/%	SD	Mean/%	SD
Age	37.5	29.7	37.5	29.8	38.9	24.8	38.9	17.7
Sex
Male	42.5%		42.4%		51.0%		59.1%
Female	57.5%		57.6%		49.0%		40.9%
Race
Asian	1.0%		1.0%		0.5%		0.4%
Black	26.2%		26.2%		25.7%		22.3%
Hispanic	25.7%		25.9%		13.6%		5.9%
White	36.9%		36.8%		39.3%		56.3%
Other/Not determined	10.2%		10.1%		20.9%		15.1%
Eligibility
LTC	2.7%		2.6%		9.0%		7.1%
SSI	31.5%		30.9%		66.3%		55.9%
TANF	39.6%		40.1%		10.8%		3.5%
QMB, SLMB, QI1	17.6%		17.7%		11.8%		33.3%
Other	8.6%		8.7%		2.1%		0.2%

LTC, long-term care; SSI, Supplemental Security Income; TANF, Temporary Assistance for Needy Families; QMB, qualified Medicare beneficiary; SLMB, specified low-income Medicare beneficiary; QI1, qualified individuals - 1; SD, standard deviation.

When considering individuals who were high-cost cases for multiple prior years, only the service and diagnostic profile from 2009 was included in the regression. This decision was made because of the difficulty in determining how specific service patterns and diagnoses from several years ago should factor into whether an individual is a high-cost case in the current year. The importance of a diagnosis may vary depending on the time since the diagnosis. Although the development of a predictive model is a primary goal of the paper, there also is a goal of developing a model that would be straightforward to implement.

The third regression focused on predicting persistent high-cost cases for the following 5 years. Among individuals who were high-cost cases in 2005, the probability of individuals remaining high-cost cases in each year between 2006 and 2010 was examined. Once again, the predictive accuracy of the models was assessed by sorting individuals into quartiles based on the predicted probability of being a persistent high-cost case. For each quartile, the proportion of individuals who were high-cost cases in the year(s) was determined. The fourth perspective used the persistence of prior high costs to predict the persistence of future high-cost cases. In this case, the author examined individuals who were high-cost cases in 2005–2007 and predicted who remained a high-cost case in 2008 through 2010.

Predictive accuracy of diagnosis-based models is typically assessed by sorting samples into groups and comparing predicted costs with actual costs in each group. For the regression models in this paper, the coefficients and individual characteristics were used to generate the predicted probability of remaining a high-cost case. The sample was sorted into quartiles based on the predicted probability of remaining a high-cost case. For each quartile, the author determined the proportion of individuals who actually remained high-cost cases. It was expected that individuals with a higher predicted probability of remaining high-cost cases should be more likely to actually remain high-cost cases. Service utilization and average service costs also are reported for individuals in each quartile.

Results

There were 512,631 people with at least 6 months of Medicaid enrollment in all 6 years between 2005 and 2010. Descriptive statistics are in Table 1.

Table 1 also distinguishes between individuals who were not high-cost cases in any of the 6 years, individuals who were high-cost cases in some years, and individuals who were high-cost cases in each of the 6 years. Individuals in the top 1% of expenditures tended to be Medicaid eligible because of SSI or long-term care as opposed to Temporary Assistance for Needy Families, were more likely to be white and less likely to be Hispanic, and were more likely to be male.

Table 2 contains descriptive statistics for annual service use and expenditures by specific service categories. Average annual expenditures for the sample were $7218. The average is higher than the average for all Medicaid recipients because of the requirement in this study that individuals remain enrolled in Medicaid for all 6 years. As is typical, the distribution of expenditures was highly skewed with the median of $710 being much less than the mean. The 99^th percentile for annual expenditures was $89,149 and individuals with expenditures during the year higher than the 99^th percentile were classified as a high-cost case in that year. In all, 97% of Medicaid beneficiaries (498,676/512,631) in the sample were not high-cost cases in any of the 6 years, 2% of beneficiaries were high-cost cases in at least 1 year, and 0.6% were high-cost cases in all 6 years. Despite being a small proportion of the sample, beneficiaries who were high cost in at least 1 year comprised 19% of total Medicaid expenditures for the sample, and individuals who were high cost in all 6 years comprised 12% of all expenditures.

Table 2.

Annual Medicaid Expenditures 2005–2010

	Full sample	Never high cost	Ever high cost	Always high cost
	%	%	%	%
Utilization
Physical health
Inpatient	11.6%	11.0%	33.3%	24.6%
Outpatient	76.2%	75.6%	96.0%	97.0%
Pharmacy	48.7%	47.9%	78.3%	74.4%
Hospice	0.2%	0.2%	1.2%	0.3%
Nursing home	3.8%	3.4%	20.2%	5.9%
ICF - Mental retardation	0.5%	0.0%	5.8%	60.2%
Behavioral health
Inpatient	1.8%	1.6%	10.8%	1.5%
Outpatient	22.6%	21.8%	54.0%	34.8%
Pharmacy	23.9%	22.9%	61.7%	67.1%
Hospice	0.0%	0.0%	0.1%	0.0%

Expenditures	Mean	SD	Mean	SD	Mean	SD	Mean	SD
Physical health
Inpatient	554	5199	340	2076	8875	27,836	5950	28,010
Outpatient	2409	9932	1779	5761	20,871	32,040	38,943	68,824
Pharmacy	864	9082	715	2638	4610	18,498	11,630	105,061
Hospice	66	1812	57	1625	475	5617	105	2707
Nursing home	1577	9575	1338	8580	11,420	25,595	5678	25,728
ICF—Mental retardation	566	8326	14	975	5143	21,958	72,612	64,917
Behavioral health
Inpatient	143	2610	75	1214	3239	15,513	252	3891
Outpatient	641	4323	485	2997	6251	16,649	6209	22,955
Pharmacy	317	1321	277	1198	1765	3263	1746	3279
Hospice	10	664	10	656	26	1012	0	0
Total payments	7218	21,303	5149	12,249	63,123	48,230	143,347	111,056
Observations (person-year)	3,075,786		2,992,056		64,932		18,798
People	512,631		498,676		10,822		3133

Notes: The sample contains 1 observation per person per year.

ICF, intermediate case facility; SD, standard deviation

Among all recipients in the sample, 76% used physical health outpatient services and 23% used behavioral health outpatient services. Other high-use service categories include physical health pharmacy (49%) and behavioral health pharmacy (24%). Only 4% of recipients used nursing home services, while nursing home expenditures averaged $1577 across all recipients (including users and nonusers).

Among those who were high-cost cases in all 6 years, 97% of the person-year observations included physical health outpatient services, 60% ICF-mental retardation services, 74% physical health pharmacy services, and 67% behavioral health pharmacy services. Average annual costs were $143,347 for high-cost cases, including $38,943 for physical health outpatient, $72,612 for ICF-mental retardation, and $5678 for nursing home services. Individuals in the top 1% of costs tended to have very high utilization of physical health services. Total expenditures for the 3133 individuals classified as high-cost cases in all 6 years were over $2.6 billion dollars ($2,694,636,906).

Table 3 illustrates the persistence of high-cost cases among individuals who were high-cost cases in 2005. Nearly half of all individuals who were high-cost cases in 2005 remained high-cost cases for the following 5 years. Similarly, while not included in the table, 77% of individuals who were high-cost cases in 2009 remained high-cost cases in 2010. Among individuals who were high-cost cases in 2008 and 2009, 87% remained high-cost cases in 2010. And 92% of individuals who were high-cost cases in the prior 3 years remained high-cost cases in 2010, while 93% of individuals who were high-cost cases in the prior 4 years and 94% of those who were high-cost cases in the prior 5 years remained high-cost cases in 2010.

Table 3.

Persistence of High-Cost Cases

	N	%
High cost cases in 2005
Remained high cost through:
2006	4628	72.7%
2007	4057	63.7%
2008	3565	56.0%
2009	3325	52.2%
2010	3133	49.2%

Table 4 contains the regression results. The first specification estimated the probability of being a high-cost case in 2010 among individuals who had been high-cost cases in 2009. In all, 77% of 7432 individuals remained high-cost cases in 2010. Odds ratios (ORs) are presented to indicate the relationship between each variable and the likelihood of remaining a high-cost case. Overall, predictive accuracy for each model is addressed in Table 5. As the name implies, the OR is the ratio of odds; for example, the OR for females is computed as: [(# of women who remain high cost/# women who do not remain high cost)/(# of men who remain high cost/# men who do not remain high cost)]. Females were less likely to have high costs in 2010 (OR=0.79, CI 0.70–0.90), as were older individuals (OR=0.99, CI 0.98–0.99). Individuals eligible for Medicaid because of SSI were more likely to have high costs than individuals eligible for other reasons. Several service categories were negatively associated with the likelihood of being a high-cost case in the following year including behavioral health inpatient and physical health inpatient services. Nursing home services and ICF-mental retardation services were associated with being a high-cost case in the next year. Nervous system disorders and digestive disorders were associated with remaining a high-cost case, while respiratory system disorders were associated with a lower probability of being a high-cost case.

Table 4.

Logistic Regression Results Predicting High-Cost Cases

	High cost prior year			High cost prior 5 years			High cost next 5 years			High cost next 3 years
	Odds ratios			Odds ratios			Odds ratios			Odds ratios
	Estimate	95% CI		Estimate	95% CI		Estimate	95% CI		Estimate	95% CI
Age	0.98	0.98	0.99	0.98	0.97	0.99	0.98	0.97	0.98	0.98	0.98	0.99
Behavioral Health Utilization
Inpatient	0.29	0.23	0.35	0.28	0.14	0.55	0.26	0.19	0.34	0.41	0.27	0.61
Outpatient	0.96	0.81	1.13	0.61	0.40	0.93	0.81	0.68	0.97	0.78	0.62	0.98
Pharmacy	1.06	0.90	1.24	0.92	0.63	1.34	1.38	1.18	1.60	1.15	0.96	1.38
Physical Health Utilization
Inpatient	0.56	0.47	0.65	0.75	0.50	1.13	0.51	0.43	0.61	0.84	0.66	1.05
Other	0.37	0.14	0.99	>999.9	<0.001	>999.9	1.92	0.77	4.75	3.37	1.01	11.28
Outpatient	0.69	0.42	1.15	<0.001	<0.001	>999.9	1.00	0.63	1.58	0.71	0.36	1.37
Pharmacy	0.92	0.74	1.13	1.11	0.68	1.83	1.15	0.77	1.71	1.00	0.78	1.28
Nursing Home	1.31	1.09	1.58	1.17	0.67	2.05	1.53	1.20	1.96	1.50	1.07	2.12
Hospice	0.90	0.58	1.39	0.43	0.08	2.26	0.90	0.43	1.89	0.58	0.16	2.15
ICF—Mental Retardation	31.40	22.08	44.63	13.47	7.82	23.18	14.31	11.39	17.98	8.31	6.10	11.32
Diagnosis (ICD-9)
Infectious diseases 001–139	1.02	0.88	1.17	1.15	0.80	1.65	1.03	0.89	1.20	0.99	0.82	1.21
Neoplasms 140–239	0.83	0.69	1.01	0.70	0.41	1.18	0.55	0.43	0.70	0.68	0.49	0.95
Endocrine diseases 240–279	1.07	0.92	1.25	0.99	0.69	1.42	0.98	0.85	1.14	1.04	0.86	1.27
Diseases of blood 280–289	0.97	0.84	1.13	1.06	0.72	1.55	1.04	0.87	1.23	1.07	0.85	1.34
Mental disorders 290–319	1.00	0.84	1.20	1.24	0.79	1.97	0.99	0.84	1.16	0.92	0.75	1.13
Nervous system 320–359	1.30	1.13	1.50	1.13	0.79	1.63	1.65	1.43	1.89	1.22	1.01	1.46
Sense organs 360–389	1.11	0.98	1.26	1.25	0.90	1.73	1.13	0.99	1.29	1.17	0.99	1.39
Circulatory system 390–459	0.98	0.84	1.15	0.93	0.65	1.34	0.76	0.65	0.88	0.93	0.76	1.14
Respiratory system 460–519	0.81	0.70	0.94	1.08	0.74	1.59	1.06	0.91	1.24	1.20	0.98	1.47
Digestive system 520–579	1.18	1.02	1.36	0.95	0.67	1.36	1.18	1.02	1.37	1.42	1.17	1.72
Genitourinary system 580–629	1.03	0.89	1.20	1.03	0.70	1.50	0.91	0.78	1.06	0.96	0.78	1.17
Pregnancy complication 630–679	0.64	0.30	1.33	>999.9	<0.001	>999.9	0.28	0.07	1.15	1.80	0.43	7.45
Skin diseases 680–709	0.90	0.79	1.04	1.16	0.80	1.67	0.88	0.76	1.02	0.91	0.75	1.11
Musculoskeletal 710–739	0.95	0.82	1.09	1.02	0.73	1.44	1.12	0.97	1.29	0.88	0.73	1.06
Congenital anomalies 740–759	0.96	0.81	1.15	0.83	0.56	1.24	1.05	0.89	1.26	0.93	0.74	1.18
Perinatal conditions 760–779	1.42	0.93	2.18	0.62	0.32	1.20	0.32	0.25	0.42	0.57	0.39	0.85
Ill-defined conditions 780–799	0.93	0.76	1.13	1.27	0.81	2.01	1.16	0.97	1.38	1.07	0.87	1.32
Injury/poisoning 800–999	0.99	0.85	1.15	1.66	1.05	2.60	0.92	0.78	1.08	1.11	0.89	1.37
Eligibility
SSI	1.38	1.16	1.64	0.94	0.60	1.46	1.39	1.14	1.69	1.07	0.81	1.41
LTC	0.95	0.72	1.25	1.09	0.36	3.25	1.04	0.78	1.37	0.85	0.59	1.24
Race (ref: white)
Asian	0.70	0.34	1.46	0.45	0.09	2.43	0.88	0.34	2.30	0.93	0.30	2.94
Black	0.99	0.84	1.16	1.17	0.76	1.81	1.01	0.85	1.19	1.05	0.85	1.30
Hispanic	0.85	0.70	1.03	1.02	0.59	1.77	0.89	0.70	1.13	1.04	0.76	1.43
Other race	1.01	0.84	1.21	1.01	0.63	1.61	1.18	0.97	1.43	1.22	0.94	1.58
Female (ref:male)	0.79	0.70	0.90	0.87	0.64	1.20	0.90	0.79	1.02	0.89	0.75	1.05
Observations	7432			3325			6367			4057

CI, confidence interval; ICD-9, International Classification of Diseases, Ninth Revision; ICF, intermediate case facility; LTC, long-term care; SSI, Supplemental Security Income.

Table 5.

Predictive Accuracy of Logistic Regressions

	Predictive probability quartiles
		1	2	3	4
High cost prior year
Not high cost	N	931	500	219	25	1675
	%	50.1%	26.9%	11.8%	1.4%
High cost	N	927	1358	1639	1833	5757
	%	49.9%	73.1%	88.2%	98.6%
		100.0%	100.0%	100.0%	100.0%
High cost prior 5 years
Not high cost	N	128	50	10	4	192
	%	15.4%	6.0%	1.2%	0.5%
High cost	N	704	781	821	827	3133
	%	84.6%	94.0%	98.8%	99.5%
High cost next 5 years
Not high cost	N	1413	1064	523	234	3234
	%	88.8%	66.8%	32.9%	14.7%
High cost	N	179	528	1069	1357	3133
	%	11.2%	33.2%	67.2%	85.3%
High cost in 2008–2010
Not high cost	N	474	253	127	70	924
	%	46.7%	25.0%	12.5%	6.9%
High cost	N	541	761	887	944	3133
	%	53.3%	75.1%	87.5%	93.1%

Notes: The dichotomous dependent variable denotes the individual is a high-cost case in 2010. For each regression the sample is limited to individuals who were high cost in the prior year, prior 2 years, prior 3 years, prior 4 years, and prior 5 years, respectively.

The second specification in Table 4 focuses on predicting who will remain a high-cost case in 2010 among individuals who have been persistent high-cost cases in the prior 5 years. A total of 94% of the 3325 individuals who were high-cost cases in each of the prior 5 years remained high-cost cases in 2010. Older individuals were less likely to be high-cost cases. Among the services with the strongest negative association with remaining high-cost cases were behavioral health inpatient and outpatient services. ICF-mental retardation services were positively associated with remaining a high-cost case in 2010. Among the 2009 diagnostic groups, only injury/poisoning was associated with remaining a high-cost case in 2010. None of the demographic characteristics were associated with remaining a high-cost case.

The third specification focuses on predicting who will remain a high-cost case in 2006–2010 among individuals who were high-cost cases in 2005. As presented in Table 3, 49% of the 3033 individuals who were high-cost cases in 2005 remained high-cost cases in each of the next 5 years. Older individuals were less likely to remain high-cost cases. Among the services used in 2007 associated with reduced likelihood of being a future high-cost case were behavioral health inpatient, outpatient, and pharmacy services, and physical health inpatient services. Nursing home and ICF-mental retardation services were associated with a greater likelihood. Diagnostic groups positively associated with being a high-cost case included nervous system disorders and digestive disorders, while neoplasms, circulatory system disorders, and perinatal conditions were associated with a lower probability of remaining a high-cost case. Individuals eligible for Medicaid related to SSI were more likely to have high costs than individuals eligible for other reasons.

The final specification focuses on predicting who will remain a high-cost case in 2008–2010 among individuals who were high-cost cases in 2005–2007. A total of 77% of the 4057 individuals who were high-cost cases in 2005–2007 remained high-cost cases in 2008–2010. Older individuals were less likely to remain high-cost cases. Among the services used in 2007 associated with reduced likelihood of being a future high-cost case were behavioral health inpatient and outpatient services. Nursing home and ICF-mental retardation services were associated with a greater likelihood. Diagnostic groups positively associated with being a high-cost case included nervous system disorders and digestive disorders, while neoplasms and perinatal conditions were associated with a lower probability of remaining a high-cost case.

Table 5 examines the predictive power of the models. For each regression model the sample was sorted into quartiles based on the predicted probability from the regression coefficients. For each quartile the table reports the proportion of people who actually remained a high-cost case and the proportion who did not remain a high-cost case. Several results are of interest. First, the models distinguish between individuals more and less likely to remain high cost. For example, among individuals who were a high cost case in 2009, 50% of individuals in the lowest predicted quartile remained a high-cost case in 2010, while 98.6% of individuals in the highest predicted quartile remained a high-cost case. Second, persistent high-cost cases in the prior 5 years are very likely to remain high-cost cases, with 84.6% of individuals with the lowest predicted probabilities remaining high-cost cases. Third, individuals in the highest quartile are virtually certain to remain high-cost cases, regardless of the regression model. Fourth, predictive power is lowest when attempting to predict who remained a persistent high-cost case for the following 5 years. Still, more than 85% of the top predicted quartile and 67% of the third predicted quartile remained high-cost cases for the following 5 years. Finally, among individuals who were persistent high-cost cases in 2005–2007, more than 93% of the top quartile remained persistent high-cost cases in 2008–2010.

Table 6 examines services and costs for high-cost cases in 2005–2007 based on predicted probabilities of remaining high cost in 2008–2010. Several findings are of interest. First, mean costs in 2009 were higher for people in the second and third quartiles than for individuals with the highest probabilities of remaining high-cost cases. Second, inpatient costs, typically associated with high acute costs, were not associated with actually remaining a high-cost case. Indeed, higher inpatient costs were more indicative of individuals who would not remain high cost. Third, the most significant determinant of remaining a high-cost case was physical health ICF-mental retardation services. ICF-mental retardation services are comprehensive health care and rehabilitation services provided in residential facilities and promote improved functioning and independence for people with mental retardation. In all, 96% of the third quartile and 100% of the top quartile received ICF-mental retardation services, with 88% and 93% actually remaining high-cost cases, respectively. Fourth, although many individuals received behavioral health services, the average costs were a relatively small proportion of total costs. However, expenditures for behavioral health outpatient services among individuals in the bottom quartile were well above the other 3 groups. Fifth, the majority of individuals have several diagnoses with prevalence rates nearing or exceeding 50% for a majority of the diagnostic groups. Sixth, although the top 2 quartiles were primarily individuals receiving ICF-mental retardation services, the second quartile had the highest physical health outpatient and pharmacy costs.

Table 6.

Service Use and Costs for Individuals Who Were High-Cost Cases in 2005–2007

	Predicted probability quartiles individuals is high cost in 2008–2010
Prior year	1	2	3	4
High cost in 2008–2010	53.3%	75.0%	87.5%	93.1%
Total costs in 2007	127445	165595	118711	128184
Age	33.9	21.0	48.9	37.1
Behavioral Health Utilization
Inpatient	11.8%	0.1%	0.2%	0.0%
Other	0.0%	0.0%	0.0%	0.0%
Outpatient	84.4%	46.2%	16.7%	16.8%
Pharmacy	61.5%	61.0%	52.1%	72.5%
Physical Health Utilization
Inpatient	28.8%	37.1%	15.1%	18.1%
Other	0.8%	0.5%	0.0%	0.0%
Outpatient	99.9%	99.7%	95.8%	93.6%
Pharmacy	75.7%	90.2%	45.8%	80.3%
Nursing Home	12.5%	11.2%	2.4%	0.7%
Hospice	0.9%	0.0%	0.1%	0.0%
ICF - Mental Retardation	0.0%	10.7%	96.4%	100.0%
Behavioral Health Costs
Inpatient	3766	1	4	0
Other	0	0	0	0
Outpatient	42360	13663	1206	594
Pharmacy	3075	1936	448	2146
Physical Health Costs
Inpatient	6543	12871	990	2188
Other	1246	924	0	0
Outpatient	44047	85679	3381	1685
Pharmacy	16680	27643	743	1893
Nursing Home	9252	11087	2681	46
Hospice	476	0	0	0
ICF—Mental Retardation	0	11792	109258	119632
Diagnosis (ICD-9)
Infectious diseases 001–139	39.9%	46.4%	24.1%	42.6%
Neoplasms 140–239	8.1%	5.6%	8.2%	2.5%
Endocrine diseases 240–279	43.2%	37.6%	22.7%	39.2%
Diseases of blood 280–289	19.8%	21.5%	8.1%	22.2%
Mental disorders 290–319	68.3%	49.1%	40.8%	51.9%
Nervous system 320–359	41.1%	67.1%	23.9%	53.4%
Sense organs 360–389	38.0%	44.9%	31.8%	59.6%
Circulatory system 390–459	37.3%	27.1%	20.2%	19.9%
Respiratory system 460–519	37.4%	70.0%	17.3%	29.7%
Digestive system 520–579	29.4%	56.7%	20.3%	45.4%
Genitourinary system 580–629	30.5%	26.8%	21.1%	33.5%
Pregnancy complication 630–679	0.4%	0.6%	0.0%	0.2%
Skin diseases 680–709	30.0%	27.5%	12.0%	25.9%
Musculoskeletal 710–739	39.2%	36.0%	29.1%	32.4%
Congenital anomalies 740–759	14.7%	33.2%	10.7%	12.9%
Perinatal conditions 760–779	7.6%	6.6%	0.1%	0.2%
Ill-defined conditions 780–799	63.8%	84.3%	48.2%	71.3%
Injury/poisoning 800–999	19.5%	20.1%	14.3%	20.2%

ICD-9, International Classification of Diseases, Ninth Revision; ICF, intermediate case facility

Discussion

Medicaid expenditures are highly concentrated among a relatively small proportion of recipients. Efforts to reduce expenditures might focus on people who are expected to utilize the most services. However, identification of individuals who will use a high level of services in the future is not straightforward. This paper focused on the identification of factors associated with longer term high-cost cases. Instead of looking at 1 or 2 years, this paper examined a 6-year period to look at 3 different perspectives. In general, several findings are consistent regardless of the perspective. First, high costs are very persistent, as a high percentage of individuals are high-cost cases for multiple years. Second, individuals receiving ICF-mental retardation services have persistent high costs. Third, individuals with 1 or more inpatient stays in the base year are less likely to remain high cost in the future. Fourth, high-cost cases have multiple diagnoses.

The models presented in this paper have several potential applications. Several are discussed in the following text based in part on the findings from this paper and in some cases reflecting recommendations by the author.

Past research found that a limited number of conditions (heart disease, cancer, trauma, mental disorders, and pulmonary conditions) accounted for a substantial portion of costs.³ Consequently, many disease management interventions have been designed and tested to improve care and reduce costs for individuals with a high-cost disease. Examples include studies on the cost-effectiveness of interventions for patients with persistent obstructive pulmonary disease,¹³ diabetes,¹⁴ and congestive heart failure,¹⁵ among others. Care coordination programs implemented with Medicare patients with congestive heart failure, coronary artery disease, or diabetes produced mixed results.¹⁶ Fifteen care coordination programs were studied that attempted to provide education to patients by nurses and provide telephone monitoring twice a month. Program effects were modest at best, with most of the 15 programs showing no reduction in costs or hospitalizations and no improvements in adherence.¹⁶ The authors of that study suggest that the interventions were too modest to provide any meaningful effect for complex cases, and that in-person interventions targeted toward moderate-to-severe cases were most likely to improve patient care.

As noted in Table 2, most diagnostic categories were highly prevalent and suggest a high rate of comorbid conditions. Nearly 57% of individuals who were high-cost cases in 2005–2007 received a mental health diagnosis in 2007, while 48% had infectious diseases, 42% had diseases of the nervous system, and 45% had diseases of the sense organs. Such results suggest that it may not be a single disease that generates such high costs, rather it is the combination of several diseases. Thus, efforts to implement in-person case management might focus on cases that involve combinations of diseases to focus on cases with the highest cost implications.

Many individuals who are high-cost cases in a single year will continue to be high-cost cases into the future. As a result, efforts to change the pattern or trajectory of expenditures may have benefits for many years to come. For example, a majority of individuals who were high-cost cases in 2005–2007 remained high-cost cases in 2008, 2009, and 2010. Although the top 2 quartiles are comprised primarily of individuals receiving ICF-mental retardation services, a majority of the lower 2 quartiles will continue to have high costs. Individuals in the lower 2 quartiles receive considerable outpatient services, and individuals in the bottom quartile have particularly high behavioral health outpatient expenses. Thus, a demonstration could be designed that develops an intensive case management program targeted toward individuals who consume a very high level of outpatient services. A review of records would be performed in an attempt to distinguish between individuals who are high users because of the extreme complexity of their disease(s) and individuals who are perhaps higher users of services than their disease profile would suggest. Given the extremely high costs for high-cost cases, a relatively modest reduction in costs would be needed to make such a program cost-effective.

High-cost cases may be at particular risk as more Medicaid programs transition to managed care. Evaluation of managed care programs should include a particular emphasis on high-cost cases as they may be at greater risk for adverse outcomes related to the transition to a gatekeeper system. For example, the enrollment process for high-cost cases should be evaluated to determine whether high-cost cases have adequate access to available plans. In addition, research should determine whether high-cost cases have reduced services after implementation, and if so, whether such reductions reflect appropriate medical management.

Defining the criteria for high cost is somewhat arbitrary, with past research using a range from the top decile² down to the top .1%.⁶ The present study used the top 1% as the cutoff. Although there is no single correct cutoff, using .1% would simply identify individuals receiving ICF-mental retardation as high-cost cases, while a 5% or 10% cutoff would identify a large number (25,000-50,000 people in this sample). Future research should determine an optimal level for defining high-cost cases that would enable the design of cost-effective interventions, and examine which combinations of diagnoses might be most effectively managed among high-cost cases.

The sample was limited to people who were enrolled in the Medicaid program in all 6 years. This inclusion criteria has several potential implications for the distribution of costs and the identification of high-cost cases. For example, people who died would be excluded unless they died at the very end of the sample period. Individuals are often very expensive in the year before death, especially for individuals who die from expensive diseases. As a result, excluding people who died alters the distribution of costs. On the other hand, many people were excluded because they did not meet Medicaid eligibility requirements in all 6 years. Such individuals tend to be less expensive to the program. Indeed, the average cost in the sample was greater than the average cost for all Medicaid beneficiaries, suggesting that, on average, people who were excluded because of eligibility or death were less expensive than people included in the sample.

Research that remains interested in identifying high-cost cases could extend the current study in several ways. For example, given the importance of pharmacy costs, additional research should identify specific medication classes associated with high costs. Greater specificity of diagnosis and service use also could be included in the model. Although the models focus on total Medicaid costs, there was a dramatic difference in the type of services received by individuals who remained high cost. Most individuals received physical health outpatient and “other” services, but individuals who remained high cost tended to have much greater physical health other costs while individuals who did not remain high cost tended to have high physical health outpatient costs.

Future research also may address several additional issues. First, prior research has focused on identifying high-cost individuals based on the diagnostic profile from a single year of data. This paper took the same approach, using diagnoses from a single year even when multiple years of persistent high costs were required. Future research may investigate the importance of past diagnoses and their relevance to current high costs. In addition, the lack of managed care data limited the analysis to the fee-for-service program. Managed care organizations may manage high-cost cases differently, and may be more effective at providing coordinated care that limits costs. Of course, an alternative possibility is that managed care organizations limit access to needed care through the use of the gatekeeper system.

As with any study that uses administrative data, there are several shortcomings to the analysis. Administrative data are often subject to data entry errors and diagnostic coding errors. The ability of diagnosis-based models to predict accurately depends on the specificity and completeness of diagnostic data. In addition, the data are limited to individuals in the fee-for-service program as managed care encounters were not available. The use of administrative data also limits generalizability. Medicaid data are examined from a single state, and the estimated models may not generalize to all states or to other payers.

Conclusion

This study developed models that can be used to identify persistent high-cost cases in the Florida Medicaid program. Models were developed based on a combination of individual characteristics, diagnoses, service use, and prior costs. The models had strong predictive power and several potential applications were presented.

Footnotes

Author Disclosure Statement

Dr. Robst declared no conflicts of interest with respect to the research, authorship, and/or publication of this article.

The author received following financial support for the research, authorship, and/or publication of this article: Financial support was provided by the Florida Medicaid Drug Therapy Management Program for Behavioral Health at the University of South Florida, and the Florida Agency for Healthcare Administration, Tallahassee, Florida.

References

Conwell

, Cohen

. Characteristics of People with High Medical Expenses in the U.S. Civilian Noninstitutionalized Population, 2002. Statistical Brief #73. Rockville, MD: Agency for Healthcare Research and Quality; 2005.

Cohen

, Yu

. The Persistence in the Level of Health Care Expenditures Over Time: Estimates for the U.S. Population, 2002–2003. Statistical Brief #124. Rockville, MD: Agency for Healthcare Research and Quality; 2006.

Olin

, Rhoades

. The Five Most Costly Medical Conditions, 1997 and 2002: Estimates for the U.S. Civilian Noninstitutionalized Population. Statistical Brief #80. Rockville, MD: Agency for Healthcare Research and Quality; 2005.

Kaiser Family Foundation. Medicaid moving forward. http://kff.org/medicaid/fact-sheet/the-medicaid-program-at-a-glance-update/. Accessed January 2015 .

Meenan

, O'Keefe-Rosetti

, Hornbrook

, et al. The sensitivity and specificity of forecasting high-cost users of medical care. Med Care., 1999; 37:815–823.

Ash

, Zhao

, Ellis

, Schlein Kramer

. Finding future high-cost cases: comparing prior cost versus diagnosis-based methods. Health Serv Res., 2002; 36(6 pt 2):194–206.

Meenan

, Goodman

, Fishman

, Hornbrook

, O'Keefe-Rosetti

, Bachman

. Using risk adjustment models to identify high-cost risks. Med Care., 2003; 41:1301–1312.

Robst

. Comparing methods for identifying future high-cost mental health cases in Medicaid. Value Health., 2012; 15:198–203.

Kronick

, Gilmer

, Dreyfus

, Lee

. Improving health-based payment for Medicaid beneficiaries: CDPS. Health Care Financ Rev., 2000; 21(3):29–64.

10.

Pope

, Kautter

, Ellis

, et al. Risk adjustment of Medicare capitation payments using the CMS-HCC model. Health Care Financ Rev., 2004; 25(4):119–141.

11.

Robst

, Levy

, Ingber

. Diagnosis-based risk adjustment for Medicare prescription drug plan payments. Health Care Financ Rev., 2007; 28(4):15–30.

12.

Quinlivan

. Treating high-cost users of behavioral health services in a health maintenance organization. Psychiatr Serv., 2000; 51:159–161.

13.

Spencer

, Briggs

, Grossman

, Rance

. Development of an economic model to assess the cost effectiveness of treatment interventions for persistent obstructive pulmonary disease. PharmacoEconomics., 2005; 23:619–637.

14.

Diabetes Prevention Program Research Group. Within-trial cost-effectiveness of lifestyle intervention or Metformin for the primary prevention of type 2 diabetes. Diabetes Care., 2003; 26:2518–2523.

15.

Jerant

, Azari

, Nesbitt

. Reducing the cost of frequent hospital admission for congestive heart failure: a randomized trial of a home telecare intervention. Med Care., 2001; 39:1234–1245.

16.

Peikes

, Chen

, Schore

, Brown

. Effects of care coordination on hospitalization, quality of care, and health care expenditures among Medicare beneficiaries: 15 randomized trials. JAMA., 2009; 301:603–618.