The high concentration of health care spending, with the 5% most costly patients accounting for nearly 50% of health care expenditures,1 suggests that improving quality and efficiency of care for such patients could yield savings. These high-cost patients, commonly known as “superusers” (SUs), may suffer from inadequate access to high-quality ambulatory care, poor continuity of care during transitions across treatment settings, and health-related sequelae of complex social factors such as homelessness, pointing to opportunities for better care, better health, and lower cost.
The concentration of costs alongside the complex social circumstances and often poor medical care experienced by high-cost individuals have brought into increased focus the potential for SU-targeted care management programs to reduce overall costs,2 and numerous efforts are currently underway to identify effective SU care management strategies.3–6 Despite considerable focus on intervention development, systematic evaluations have revealed mixed findings. Disease management and care coordination programs directed at Medicare patients with high hospitalization risk have often been found to have limited or no effectiveness on hospital utilization or costs.7,8 In contrast, a few recent initiatives aimed at Medicaid high-utilizing patients have demonstrated significant net savings in costs.6 In all of the systematic assessments, it is actively recognized that while programs may generate observed decreases in costs over time, evaluations should take into account the tendency for high costs to naturally decline over time independent of interventions,9 a statistical phenomenon known as regression-to-the-mean (RTM).10–12 Controlling for RTM can be achieved by randomized controlled trials or by using quasi-experimental methods examining costs trends among patients participating in SU initiatives compared with groups of matched patients. The latter method of benchmarking enables evaluations to measure program-attributable changes in costs when randomized clinical trials are infeasible because they are too expensive or time consuming.
For programs without comparison patients, understanding of the magnitude of RTM may help program managers assess any observed pre-post change in costs. One recent study, by Johnson et al,12 found average costs of SUs identified on the basis of high inpatient (IP) utilization decreased by around 60% over 2 years. More generally, studies of cost persistence have shown that only a limited fraction of high-cost patients (eg, top decile) remain high cost over time.1,13–15 Although not documenting its magnitude, these studies are consistent with substantial RTM.
Our study contributes to the literature on high-cost patients by examining in detail the effect of RTM in synthetic cohorts of New Jersey hospital SUs selected on the basis of high IP or emergency department (ED) utilization over a 6-month baseline period (details on utilization thresholds are provided in the Methods section). We examined quarterly changes in average costs of patients in each cohort over 2 years, comparing rates of RTM and the rapidity with which it occurs across patients having high IP or ED utilization. This analysis builds on previous work by Johnson and colleagues by focusing on, in addition to those with high IP use, SUs characterized by high ED use, and subgroups of IP and ED SUs which demonstrates just one, or both types of high use. We find substantial differences in levels and trends of costs between IP SUs and ED SUs, and additional variations by patient demographics, clinical characteristics, and source of insurance.
We used statewide New Jersey uniform billing hospital discharge data as the source for IP hospitalizations and treat-and-release ED visits during our study period of 2009–2011. Under a special arrangement with the New Jersey Department of Health, LinkKing software16 was used to link patients’ hospital records over time. Further, data from the New Jersey charity care program17 and mortality files, based on the death certificate database maintained by the New Jersey Department of Health (that included information on out-of-state deaths), were also linked to the uniform billing files to flag hospital encounters by low-income uninsured patients eligible for charity care reimbursement and to account for patients who died.
Each hospital record provides information on patient demographics, expected primary and secondary payer (that enabled classification of patients by source/availability of health insurance, namely Medicare-Medicaid dual eligibles, other Medicare, Medicaid only, private insurance, charity care, and self-pay/uninsured), clinical characteristics (primary and secondary diagnoses and procedures), patient residential zip code, and hospital charges associated with each discharge. We utilized hospital-specific cost-to-charge ratios from the Agency for Health Care Research and Quality (AHRQ), Healthcare Cost and Utilization Project,18 zip code level rates of individuals with incomes below the federal poverty level from the 2007–2011 5-year American Community Survey,19 and consumer price indices for medical care20 from the Bureau of Labor Statistics, which is used to adjust for medical care inflation over our study period.
Enrollment criteria of ongoing SU care demonstration projects6,21,22 guided our selection of utilization thresholds and related time periods for defining SU cohorts. We first selected 2 nonmutually exclusive cohorts of SUs based on their utilization over a 6-month “baseline” period July–December 2009: patients with ≥2 IP stays (denoted IP SUs) and those with ≥6 ED visits (ED SUs). Within these 2 groups, we further examined 3 mutually exclusive and exhaustive subsets of SU individuals. These are comprised of: (a) individuals who we denote as “IP SU only” (patients with 2+ IP stays but fewer than 6 ED visits); (b) those who we consider to be “ED SU only” (patients with 6+ ED visits but fewer than 2 IP stays); and (c) those who were “IP and ED SU” (patients with 2+ IP stays and 6+ ED visits). We further defined as non-SUs, the cohort of individuals who had 1 IP stay or 1 ED visit, the minimum utilization required to be included in this hospital database. Because many SU interventions serve adults with long-term health conditions, all of our cohorts are limited to patients over age 18 years with ≥2 diagnosed chronic conditions, that is, patients with zero or 1 chronic condition are excluded from the study. Further, as is typical of SU initiatives, we excluded patients who had a principal diagnosis of maternity or cancer, and also excluded from the analysis those who died during the baseline period. Our final analytic sample comprised 289,060 patients during the baseline period.
Our outcome was average quarterly hospitalization costs for the baseline period (July–December 2009), and each of the 8 follow-up quarters over January 2010–December 2011. Hospital cost was calculated by applying hospital-specific cost-to-charge ratios to discharge-related charge amounts, and expressed in 2011 dollars.
We examined the association of several patient-related factors with SU costs, adjusting for patient demographics, source/availability of health insurance, clinical factors, and poverty rates at zip codes where patients resided. We utilized algorithms based on the CMS Chronic Conditions Data Warehouse23 to characterize patients based on whether they had 2, 3, or 4+ chronic conditions. The AHRQ Clinical Classification Software24 was used to identify patients diagnosed with behavioral health (BH) conditions including a diagnosis of mental health or substance use disorder. We separately identified patients with a principal diagnosis of severe mental illness (SMI) based on the national comorbidity survey replication25 and subsequent work by Coffey et al.26 These patients experienced functional and social impairment that hampered their regular activities for at least 88 days out of the year and were most likely to have a diagnosis of psychoses, bipolar disorder, drug dependence, obsessive compulsive disorder, dysthymia (chronic depression), or related diagnoses.
We measured RTM as the percentage decrease in quarterly average IP and ED costs between baseline and each of 8 follow-up quarters. We investigated whether SU patients with SMI, other BH conditions, or greater numbers of chronic conditions have less RTM in costs over time, that is, more of cost persistence that is defined as 100% −RTM. We also considered additional indicators of patient vulnerability defined by zip code level poverty and source of health insurance. We conducted descriptive, bivariate, and also multivariate analysis of costs over baseline and 8 follow-up quarters utilizing the framework of generalized estimating equation (GEE) with a gamma distribution, logarithmic link.27,28 We assumed an exchangeable correlation structure as unobserved individual factors related to health status and behavior are expected to remain unchanged over our relatively short study period.
Table 1 shows demographic, clinical, and health insurance information relating to the 6 cohorts of patients during the baseline period and changes in average quarterly costs and cohort size over the subsequent 8 follow-up quarters. Eleven percent of the 289,060 patients at baseline died during the follow-up period. Compared with non-SU patients, the 5 types of SU patients are more likely to live in high-poverty zip codes, have higher prevalence of BH conditions, and, with the exception of ED SU-only individuals, have a higher burden of chronic conditions. ED SUs compared with IP SUs are much less likely to be elderly and more likely to be covered by Medicaid, charity care, or be uninsured. They are also more likely to have BH diagnoses, but less likely to have ≥4 chronic health conditions.
Figure 1 shows average quarterly costs at baseline and in each of the 8 follow-up quarters for each of 4 SU cohorts and all patients (the IP SU-only cohort and the non-SU cohort are not shown as their costs were almost identical to the IP SU cohort and all-patient cohort, respectively). The slope of the lines in Figure 1 reflects the magnitude of RTM in each cohort. The IP SU cohort experienced the greatest RTM by the end of the follow-up period (69.7%), whereas the ED SU-only cohort experienced the least RTM (8.4%). The ED SU cohort experienced a low RTM as well, at 38.3%. Although ED SUs had lower average cost during the baseline period, these differential trends resulted in their having an average cost at the end of the follow-up period that was 12% higher than IP SUs.
The lower baseline costs and RTM for ED SUs was driven by the ED SU-only subgroup, which had the lowest baseline cost ($3889) and lowest RTM (8.4%) among all the SU cohorts. The remaining individuals among ED SUs, those who were both IP and ED SU, were distinctly different. They had the highest baseline costs among all SU groups and a RTM (46.4%) that was just above that of ED SUs (38.3%). The non-SU cohort (not shown here, but see Table 1) exhibited a level of costs that was slightly lower compared with all patients and also a declining trend in costs that reflects a decreased likelihood of further hospitalizations in subsequent quarters.
We also replicated this analysis after dropping patients who died at any time during the follow-up period to account for the possibility that at least a part of the high costs at baseline or in the early follow-up quarters among SUs may be due to end-of-life costs, leading to the sharp decrease in average cohort costs after the death of such individuals. This exclusion had little impact on either the level of cost or RTM in any of the cohorts. The overall trend and quarterly changes in costs when analysis was restricted to only those patients who survived over the 8 quarters is similar to that in our original sample (see Fig., Supplemental Digital Content 1, http://links.lww.com/MLR/B196). In addition, though baseline costs are indeed slightly lower when only surviving patients are included, we see very similar rates of RTM (see Table, Supplemental Digital Content 2, http://links.lww.com/MLR/B197).
Table 2 reports findings from our adjusted analysis of hospitalization costs based on the GEE model and sheds further light on average cost levels and how quickly RTM occurs among SU cohorts. It reports the persistence percentage (100% −RTM) and actual decrease in costs corresponding to each of the independent variables. The percentages relating to each of the 8 follow-up quarters thus equal average costs in that quarter expressed as a percentage of baseline costs. On the basis of the ratio of the corresponding RTM values (calculated as 100% minus coefficient estimates) at the end of the first and eighth quarters, we calculated that IP SUs experience the effects of RTM earlier, with 82.0% (100%–45%/100%–33%) of their overall decrease in costs occurring in the first quarter of follow-up. In contrast, the ED SU cohort experienced only 60.5% of their overall decrease in costs during the same period of time, and IP and ED SUs 65.9%. By the fourth quarter, all cohorts experienced at least 97% of their decrease in costs. Patients with BH conditions, 4+ chronic conditions, those who were dual eligible, or Medicaid beneficiaries faced higher levels of hospitalization costs, with the presence of 4+ chronic conditions making the greatest difference to costs.
Table 3 further illustrates the relationship of patient characteristics to cost levels and RTM across specific cohorts of SUs. Considering all patients, baseline and follow-up costs are greater when SMI, any BH, or 4+ chronic conditions are present, but the specific associations differ among SU groups. Consistent with the multivariate results reported above, though the presence of SMI increased the risk of being classified as an IP SU by 3 times and as ED SU by 16 times (data not shown), among SUs the incremental effect of SMI on costs is not evident. Presence of any BH diagnosis increases baseline costs among ED SUs, whereas having 4+ chronic conditions has a positive effect on costs for all categories of SUs. BH diagnosis was positively associated with RTM for ED SU, but negatively associated with RTM for IP SUs and when all patients were considered.
Finally, Figure 2 illustrates differences in levels of cost and RTM among subgroups of IP SUs and ED SUs distinguished by health and socioeconomic factors such as source/availability of insurance and residence in poor zip codes. We define 3 mutually exclusive subgroups of SUs based on these characteristics. First, we define SUs as “more vulnerable” if they lived in a high-poverty zip code and were covered by Medicaid (including as Medicaid-Medicare dual eligibles) or classified as charity care eligible, did not have BH diagnoses, and had 2 or 3 chronic conditions. Next we classify SUs as “most vulnerable” if they differed in terms of having ≥1 non-SMI BH diagnoses and 4+ chronic conditions. For contrast, SUs not living in a high-poverty area, without BH diagnoses, with only 2 or 3 chronic conditions, and covered by Medicare (but not Medicaid) or private insurance are classified as “less vulnerable.”
The most vulnerable IP SU group had the highest overall cost at baseline and experienced much less RTM compared with their less vulnerable counterparts. Consistent with previously described patterns, all ED SU subgroups had lower baseline cost than the corresponding IP SUs, but less RTM. We also find that vulnerability, defined by insurance status and poverty, did not have a substantial effect on costs of IP or ED SUs, but poorer health had a much greater effect on the level of cost in relative terms on ED SUs than IP SUs.
Understanding patterns of RTM among SU populations is important for designing intervention strategies, as there is greater potential for savings among patients with more persistent costs (less RTM). Further, as many SU initiatives lack resources for rigorous evaluation, accounting for the effects of RTM is vital while interpreting program outcomes. Although the estimates from our study cannot substitute for an appropriate comparison group, they inform ongoing projects targeting patients based on high costs at a point in time on the need to adjust for such effects over time. Judging by their high baseline costs alone, IP SUs would seem to offer the greatest potential for savings, but their costs came down due to high RTM. However, mean cost in this cohort still remained elevated well above that of the average hospital patient, reflecting potential cost-saving opportunities, especially if some of their utilization can be prevented,29 for instance, hospital use arising from ambulatory care sensitive conditions. RTM also occurred very quickly in this group suggesting that intervention projects without resources to establish rigorous comparison groups might effectively assess their impact by examining changes in cost beginning several months after program enrollment when the natural reduction in cost due to RTM has leveled off.10
Compared with IP SUs, ED SUs were characterized by lower burden of chronic physical conditions, higher prevalence of SMI, higher likelihood of residence in low-income areas, and higher likelihood of being Medicaid covered or charity care eligible. These findings point to the distinct nature of ED SUs and the need to address factors beyond chronic disease management, including integrating medical, behavioral, and social services and addressing sequelae of poverty such as housing instability. Among ED SUs there was substantial heterogeneity between those who were ED SU only compared with those who were IP and ED SU. IP and ED SUs had greater baseline costs compared with the overall group of IP SUs ($19,088 vs. $16,939), and the difference in costs at end of follow-up period was even higher ($10,231 vs. $5141). This greater differential was due to their low RTM at 46.4%, just above that of ED SUs. The higher cost as well as higher cost persistence in this group of individuals makes them the category of SUs with greatest potential for savings.
Our study also shows that risk factors for higher costs may differ between SUs and hospital patients overall. Although the presence of SMI diagnoses substantially increases the likelihood of being classified as SU, SUs with SMI do not have higher baseline cost or lower RTM. These patterns underscore the importance of incorporating information on SMI in predictive modeling intended to identify future high-cost patients and suggest that it may be efficient early-on to engage patients with SMI in strategies to prevent them from becoming SUs. Finally, our analysis of the effects of indicators of vulnerability shows varying patterns of RTM among IP and ED SUs.
This study has several limitations. New Jersey has high average income per capita, an ample supply of hospital beds, and historically generous Medicaid eligibility (including expanding Medicaid under the ACA). As a result, access to care may be better for SUs, limiting generalizability. That said, findings based on statewide analysis offer more generalizability than prior studies restricted to specific health care systems with distinct characteristics12 and also ensures that all hospital utilization within the state is accounted for. However, being solely based on hospital discharges, we are unable to assess total health care costs as there is no information on physician professional services, prescription drugs, and outpatient services. In addition diagnostic information in our study is based on hospital information only. Administrative billing data do not include important information on clinical acuity or the social situation of patients, for example, housing problems or a history of trauma over life course. Finally, we measure cost by adjusting charge data, not actual payments to hospitals. Actual savings will vary depending on payment source.
The popularity of SU care programs seems to be growing. Achieving success in such programs is dependent on effectively targeting patients who are likely to have persistently high costs and are likely to respond to interventions.3,7,8 Our study addressed the cost persistence question, but we could not address questions related to responsiveness to interventions. Nevertheless, understanding the natural tendency of cost among SU populations to moderate over time is critically important for designing and evaluating the impact of SU programs.
The authors thank Jian Tong and Bram Poquette for their help with this project.
8. Brown RS, Peikes D, Peterson G, et al.. Six features of Medicare coordinate care demonstration programs that cut hospital admissions of high-risk patients. Health Aff (Millwood). 2012;31:1156–1166.
9. Melnick GA, Green L, Rich J. House calls: California program for homebound patients reduces monthly spending, delivers meaningful care. Health Aff (Millwood). 2016;35:28–35.
10. Barnett AG, van der Pols JC, Dobson AJ. Regression to the mean: what is it and how to deal with it. Int J Epidemiol. 2005;34:215–220.
11. Press MJ, Scanlon DP, Ryan AM, et al.. Limits of readmission rates in measuring hospital quality suggest the need for added metrics. Health Aff (Millwood). 2013;32:1083–1091.
12. Johnson TL, Rinehart DJ, Durfee J, et al.. For many patients who use large amounts of health care services, the need is intense yet temporary. Health Aff (Millwood). 2015;34:1312–1319.
13. Monheit AC. Persistence in health expenditures in the short run: prevalence and consequences. Med Care. 2003;41(suppl):III53–III64.
14. Hirth RA, Gibson TB, Levy HG, et al.. New evidence on the persistence of health spending. Med Care Res Rev. 2015;72:277–297.
15. Coughlin TA, Long SK. Health care spending and service use among high-cost Medicaid beneficiaries, 2002–2004. Inquiry. 2009–2010;46:405–417.
20. United States Department of Labor. Measuring price change for medical care in the CPI. 2010. Available at: http://www.bls.gov/cpi/cpifact4.htm
. Accessed September 14, 2015.
21. Yedidia MJ, Lontok O, Vaz S, et al.. Confronting Challenges in Health Care Programs Serving High-Needs Patients. New Brunswick, NJ: Rutgers Center for State Health Policy; 2015.
25. Kessler RC, Chiu WT, Demler O, et al.. Prevalence, severity, and comorbidity of 12-month DSM-IV disorders in the National Comorbidity Survey Replication. Arch Gen Psychiatry. 2005;62:617–627.
26. Coffey RM, Houchens R, Chu B-C, et al.. A severity-of-illness classification for mental and substance-use disorders for use with hospital administrative data. 2010. Available at: http://www.hcup-us.ahrq.gov/reports/SOI.jsp
. Accessed September 14, 2015.
28. Liang K-Y, Zeger SL. Longitudinal data analysis using generalized linear models. Biometrika. 1986;73:13–22.
29. Joynt KE, Gawande AA, Orav EJ, et al.. Contribution of preventable acute care spending to total spending for high-cost Medicare patients. JAMA. 2013;309:2572–2578.