Excessive hospital utilization and the fragmented nature of the US healthcare system are considered to be major contributors to the nation’s comparatively high healthcare costs.1 Given that a disproportionate share of costs associated with hospital utilization is generated by a relatively small segment of the population,2 high utilizers represent an important target for efforts to minimize the excessive costs and harm to patients who are associated with fragmentation of healthcare use. Many high utilizers have multiple chronic conditions (MCC) and have frequent hospitalizations and readmissions.3–6 Although most patients with MCC are over the age of 65 years, recent data from IMS Health show that total spending on the privately insured population under 65 years of age is concentrated among those with MCC, who accounted for 63% of inpatient admissions and comprised over 3 quarters of the top 1% of patients in terms of overall cost.7 Another source of high hospital utilization is patients with behavioral health diagnoses.8,9
Providers are under increased pressure to reduce hospital admissions, and payers are investing in care management programs with the goal of reducing preventable hospital use. Medicare and some private carriers are basing their payments to hospitals and hospital-based providers in part on their ability to curb excessive utilization and readmission. Patient-centered medical homes and Accountable Care Organizations (ACO’s) are geared particularly at high utilizers and those with MCC and are designed to foster cooperation between hospital and ambulatory care settings with the goal of stabilizing patients and reducing hospital use.10 Most of these care management activities are directed at complex patients.
In this study, we present the first detailed empirical analysis of the prevalence and characteristics of fragmented inpatient hospital use in a cohort of high utilizers, most of whom have MCC. We categorize fragmented hospital use based on primary diagnosis in the first admission. We then estimate correlates of fragmented use in a Poisson regression model that predicts the number of hospitals used based on the number of chronic conditions, total admissions, and hospital market concentration. We conclude with a discussion of implications for cost, outcomes, and the care management of patients with chronic conditions, in addition to new questions to be addressed more thoroughly in future work.
Our analysis focuses on New Jersey, which is a densely populated state of 8.8 million residents. The state is racially and ethnically diverse. Approximately 60% of the population is non-Hispanic white, 13% is non-Hispanic Black, 8% non-Hispanic Asian, and about 18 % are Hispanic. Nearly one quarter of the New Jersey population is under 18 years of age, whereas nearly 14% is over 65 years of age. Approximately 21% of the New Jersey’s population aged 18–64 years is uninsured, about 71% has private coverage, whereas the rest have Medicaid or some other government coverage.11 There were 80 acute care hospitals in 2007, 11 of which are teaching facilities. New Jersey is an appropriate setting for a study of multiple hospital use, as the state has long been identified as having above-average hospital utilization and costs. New Jersey led the nation in 2009 in the Dartmouth Atlas’ hospital care intensity index, based on inpatient days and inpatient physician admissions among chronically ill Medicare Beneficiaries.12
The primary data source is 2007–2010 New Jersey Uniform Billing hospital discharge data. Hospitals submit claims quarterly to the NJ Department of Health, which edits and standardizes claims and retains them in a centralized database. The data include diagnoses, procedures, patient demographics, and discharge disposition. Under special arrangement with the New Jersey Department of Health, LinkKing software was used to link patients’ hospital records over time. LinkKing uses a combination of deterministic and probabilistic methods to link records primarily on the basis of patient name, date of birth, and Social Security Number. These primary patient identifiers were supplemented with secondary identifiers that record patient sex, race, ethnicity, and residential zip code. These secondary identifiers (referred to as “flex variables” in LinkKing terminology) provide additional information to maximize the rate of true positive linkages and minimize the rate of false positives.13 Although hospital administrative data do not distinguish well between charity care and “self-pay” patients, our analysis addresses this shortcoming by linking all records to the New Jersey Charity Care subsidy program files to identify care, which qualified as free or reduced care provided to low-income (up to 200% of the federal poverty level) uninsured residents. The payer identified at the index admission was used for the analysis, although approximately 14% of patients switch payers during the period. To accurately measure exposure to the risk of subsequent admissions, the discharge file was also linked with New Jersey mortality records to exclude from the analysis individuals who died during the study period. Robustness checks showed that the exclusion of this population did not affect results.
We define our study cohort as all patients who had an index inpatient admission in either 2007 or 2008 and at least one additional inpatient stay within 2 years of their index admission. The existence of MCC was defined through use of the HCUP Chronic Condition Indicator, which classifies ICD-9 diagnosis data into chronic and nonchronic categories.14 As compared with approximately 620,000 individuals who had one hospital stay during this period, the study population was older, more likely to be male, and considerably more likely to have MCC. We excluded those under 18 years of age and patients who changed their zip code of residence during the period. Admissions that resulted from a transfer were excluded, as were admissions to psychiatric hospitals, specialty hospitals, or long-term care facilities. There were 10 hospital closures in New Jersey during the study period (5 in 2007, 4 in 2008, and 1 in 2009.) To account for potential bias arising from hospital closures, we excluded all patients who had at least 1 admission at a hospital that closed during the study period. This resulted in the exclusion of 12,419 patients, or 3.5% of the cohort. We also excluded patients who died within 24 months of their index admission. This resulted in the exclusion of 55,704 patients, or about 16% of the cohort. We estimated our analytical models both with and without these excluded populations and found the results were robust to the changes. The final study population used for the analysis had 291,147 patients. This cohort of patients had 894,092 admissions during the 2-year period.
We included several independent variables that we hypothesized to be correlated with fragmentation of hospital use among the cohort. Patient characteristics extracted from the index admission discharge record include age, sex, race/ethnicity, marital status, median household income within patients’ zip codes, and primary payer. The existence of MCC was defined through use of the HCUP Chronic Condition Indicator, which classifies ICD-9 diagnosis data into chronic and nonchronic categories.14 Patient clinical information was also obtained from the data, including ICD-9-CM diagnosis codes. The Major Diagnostic Categories, which are derived from the ICD-9-CM codes, were used to classify the principal diagnosis into one of the 25 mutually exclusive categories. Using the available software from the Agency for Healthcare Research,15–18 we controlled for hospital market structure by constructing a zip code level Herfindahl-Hirschman Index (HHI), which is a sum of squares of market shares (ie, proportions of all discharge from a specific zip code) for all the hospitals.19 An HHI of near zero indicates that there are many hospitals with small shares serving the zip code, whereas a value of 1 indicates that only one hospital serves that zip code. Hence, an increase in HHI value indicates a more concentrated market and a decrease in competition among providers. Finally, we define a change in degree of market competition by the difference of yearly HHI value between patient’s index admission and the last admission in the 24-month window.
The correlates of fragmented hospital use are assessed through use of a Poisson specification, where the total number of admissions is the measure of exposure. The dependent variable measures the number of additional hospitals admitted after the first admission. Using exploratory data analysis, our count data shows some underdispersion (ie, there is less variability in the data than expected under the regular Poisson model).20 A Poisson regression model (with robust “sandwich” estimates of SEs to adjust for underdispersion) was used to estimate the expected number of additional hospitals (other than the hospital utilized at the index admission) that admitted patients during the 24-month window.21 As a robustness check, the analysis was also completed with and without the excluded populations (ie, patients who had an admission at a closed hospital, who died during the period), and results were not sensitive to the exclusions.
Table 1 provides descriptive information about the study population, defined as those with at least 2 inpatient admissions within a 24-month period. As is consistent with other groups of high hospital utilizers, this population has a disproportionate share of female patients (61%) and those over 65 years (46%) of age as compared with the population as a whole. The descriptive data suggest that there are differences in the likelihood of fragmented utilization by age, sex, race, marital status, and income, as well as by major diagnostic categories and comorbidities. To provide some perspective on the distribution of the cohort in terms of total utilization, Table 1 also shows that approximately 58% of the members of the study cohort had no more than 2 admissions during the 24-month period. About 13% had >5 or more admissions.
Figure 1 shows the distribution of fragmentation by the number of total admissions. As can be seen, for those with only 2 admissions, approximately 25% of patients used 2 different hospitals. Not surprisingly, as the number of admissions increases, the proportion of patients who utilize only 1 facility declines. For those patients with 15+ admissions in the 2-year period, only about 45% used the same hospital for all of their admissions and >20% used ≥3 hospitals.
Table 2 shows patient characteristics at the index admission. About 6% of index hospitalizations are due to injury, mental health, or substance abuse and about 12% are due to childbirth. The great majority of this population of high utilizers has MCC, as nearly 80% are identified as having ≥2 chronic conditions at the index admission. The prevalence of various comorbidities in the index admission is also listed in Table 2. Hypertension is identified as a comorbidity in >45% of cases, diabetes without chronic complications is listed approximately 17% of the time, and chronic pulmonary disease appears in about 15% of cases. Fluid and electrolyte disorder is also listed in nearly 15% of cases. Table 2 also shows the primary payer used in the index admission. As is consistent with the relatively elderly nature of this cohort, Medicare is the payer for 45% of index admissions, whereas Medicaid is the payer <6% of the time.
In our study population, the number of different hospitals to which patients were admitted, in addition to the hospital used on the first admission, ranged form 0 to 9. Results from the Poisson model in Table 3 show that patients with MCC were more likely than others to use multiple hospitals. The relationship between the number of chronic conditions and fragmentation is nonlinear, as can be seen by the fact that the coefficient on the category 5+ chronic conditions is smaller than the other categories. This reflects the fact that those with 5+ chronic conditions are disproportionately likely to be over the age of 65 years. Two thirds of those with 5+ chronic conditions are 65+ years of age, as compared with 33% of those with fewer than 5 chronic conditions, and fragmentation is less likely among the elderly population. To facilitate the interpretation of the relationship between the number of chronic conditions and fragmentation, we transformed the estimated coefficients from Table 3 into incidence rate ratios (IRR). For those with 1 chronic condition, the IRR was 1.149, which means holding other variables constant, fragmentation was 14.9% greater among patients with 1 chronic condition relative to those with no chronic conditions. For patients with 2–4 chronic conditions, the IRR was 1.151 (IRRs) and for those with ≥5 chronic conditions, the IRR was 1.108.
A behavioral or injury diagnosis in one of the first 2 admissions was also associated with fragmentation. Men, adults under 65 years of age, non-Hispanic whites, singles, and the privately insured were significantly more likely to use more hospitals, as were patients living in zip codes with low household income. In addition, hospital market concentration was negatively associated with fragmentation, and a measure of the change in the degree of concentration during the period of the study had the expected effect. In other words, more concentrated hospital markets had less fragmentation, and markets that became more competitive over time had an increase in fragmentation.
Our analysis shows that fragmented hospital use, defined as multiple hospital episodes across different facilities, is very common. More than 25% of our cohort of high utilizers had admissions at >1 hospital during the 2-year study period. Fragmentation was more likely among men, the middle aged, and the privately insured. Those with a first or second hospitalization for childbirth were least likely to use more than one hospital, whereas those with a first or second hospitalization for a behavioral health diagnosis were most likely to use >1 hospital. As shown in Table 3, among individuals with MCC, who comprise the majority of our study population, the relative risk of fragmented hospital use was approximately 1.14.
Our results suggest that the extent to which an admission is planned (eg, childbirth vs. injury) may affect the degree of fragmentation of hospital use, with hospital use being less fragmented in cases where admissions are more likely to be planned. Our results also show that although those with hospitalizations for injury and behavioral health have a greater likelihood of fragmentation, these hospitalizations do not play a major role in multihospital use, as patients with these diagnoses are a relatively small share of the overall cohort. The type of multiple hospital utilization, which has the greatest potential to create inefficiency, or even to harm patients, may be when patients visit multiple facilities for the same or closely related conditions. As has been shown here, this behavior comprises a significant component of the multiple hospital utilization observed in this cohort. This behavior is very difficult for any particular provider to manage and may result in additional costs without corresponding improvements in outcomes. Inefficient care resulting from lack of coordination is one of the problems that ACOs, patient-centered medical homes, and related reforms are designed to address. However, multiple hospital use is challenging for ACOs, as attributed patients are not prohibited from seeking care elsewhere. Fragmented hospital use can be a particular problem for ACOs that serve patients with MCC. Expanding the scale of ACOs to include more hospitals may enable them to improve care management of patients with MCC. Such an expansion, however, would have to be balanced against the potential for higher prices and other anticompetitive behavior that might result from extremely large ACO arrangements.
There are a number of steps that ACOs can take to minimize negative consequences of multiple hospital use by patients with MCC. Strengthening patient-centered medical home designs for patients with MCC can improve disease management and potentially reduce symptoms that may result in hospitalization. This may include strengthening patient management and follow-up and potentially increasing access to a health coach or physician extender. At the payer level, changes in benefit design could create disincentives for patients to visit multiple hospitals. For example, in the commercial market, tiered networks incentivize patients to use ACO providers.
An additional consideration is the potential for electronic health records and regional health information exchanges, which are under development in many areas, may reduce fragmentation and the inefficiency associated with multiple facility use. For example, hospitals in a region may choose to share information about their admissions, discharges, and transfers (ADT feeds) without consolidating further. These types of information sharing arrangements have the potential to reduce inefficiency and improve quality of care delivered to patients with multiple conditions, although significant gains from advances in HIT have yet to be realized.21
Finally, increased awareness is required among patients about potential adverse consequences of fragmented hospital care. It may be difficult for patients with MCC to actively seek to reduce the number of hospitals they use, but they should be encouraged to consult with their providers about the issue. This “patient-centered” focus would be consistent with emerging efforts by ACOs to integrate shared decision making and self-care with the enhanced care coordination practices to improve management of complex conditions. As patients with fragmented use are most likely to be individuals with MCC, these individuals should be the first and primary focus of these education efforts.
There are a number of limitations to the current analysis that should be noted. First, the hospital billing data used in our analysis do not allow us to distinguish between multiple hospital use because of provider referral versus patient choice. In addition, although we have excluded transfers from our analysis because we believe they are unlikely to reflect problematic or avoidable multihospital use, it is possible that this is not the case. In addition, in this analysis, we were not able to distinguish between hospitals that share a common network or other organizational structure versus those that do not. The former category may share electronic medical records or other information. As noted previously, we are not able to comment on how the multiple hospital use is related to patient outcomes and to the quality of care. The assignment of payer for the initial admission is a simplification, and in this analysis we do not assess how payer changes may have affected fragmentation. Finally, we cannot be sure how well the experience of New Jersey may generalize to the rest of the nation.
Future work is required to learn more about the extent to which multiple hospital use affects quality of care and outcomes for patients with MCC. Fragmented utilization may be a response to poor quality care, but fragmentation may create additional quality problems stemming from poor coordination among providers. The impact of fragmented utilization may differ for different patient populations and in the inpatient as compared with the emergency department context. Studies of emerging ACOs should consider how their structures relate to actual patterns of multiple hospital use among high utilizing patients. For certain patient populations, particularly those with MCC, trends in the prevalence of multiple hospital use should potentially be an outcome measure in evaluations of ACOs and other care coordination strategies.
The authors thank Ping Shi of the Center for Health Statistics, New Jersey Department of Health, for help with file linkage.
1. Emanuel EJ.Where are the health care cost savings?JAMA.2012;307:39–40.
3. Bush H.Tackling the high cost of chronic diseases.Hosp Health Netw.2012;October:34–40.
4. Bush H.Caring for the costliest.Hosp Health Netw.2012;86:25–29.
5. Feigenbaum P, Neuwirth E, Trowbridge L, et al..Factors contributing to all-cause 30-day readmissions: a structured case series across 18 hospitals.Med Care.2012;50:599–605.
6. Jencks SF, Williams MV, Coleman EA.Rehospitalization among patients in the Medicare fee-for-service program.N Engl J Med.2009;360:1418–1428.
7. .Healthcare spending among privately insured individuals under age 65, February 2012. Available at: http://www.imshealth.com
. Accessed November 12, 2013.
8. Raven M, Carrier ER, Lee J, et al..Substance use treatment barriers for patients with frequent hospital admissions.J Subst Abuse Treat.2010;38:22–30.
9. Walley AY, Paasche-Orlow M, Lee E, et al..Acute care hospital utilization among medical inpatients discharged with a substance use disorder diagnosis.J Addict Med.2012;6:50–56.
10. Lewis VA, Larson BK, McClurg AB, et al..The promise and peril of accountable care for vulnerable populations: a framework for overcoming obstacles.Health Aff.2012;31:1777–1785.
13. Campbell KM.The Link King: Record Linkage and Consolidation Software v6.4. Available at: http://www.the-link-king.com
. Accessed November 12, 2013.
16. Elixhauser A, Steiner C, Harris DR, et al..Comorbidity measures for use with administrative data.Med Care.1998;36:8–27.
17. Southern DA, Quan H, Ghali WA.Comparison of the elixhauser and Charlson/Deyo methods of comorbidity measurement in administrative data.Med Care.2004;42:355–360.
18. Li B, Evans D, Faris P, et al..Risk adjustment performance of charlson and elixhauser comorbidities in ICD-9 and ICD-10 administrative databases.BMC Health Serv Res.2008;8:12PMCID: PMC2267188.
19. Scherer FM, Ross D.Industrial Market Structure and Economic Performance.1990:3rd Ed.Boston, MA:Houghton Mifflin.
20. Guangyong Z.A modified poisson regression approach to prospective studies with binary data.Am J Epi.2004;159:702–706.
21. Colin CA, Trivedi PK.Regression Analysis of Count Data, Vol. 30.1998;Cambridge University Press.