The Net Fiscal Impact Of A Chronic Disease Management P...

The Net Fiscal Impact Of A Chronic Disease Management Program: Indiana Medicaid

The Net Fiscal Impact Of A Chronic Disease Management Program: Indiana Medicaid

Healthaffairs

Ann M. Holmes, Ronald D. Ackermann, Alan J. Zillich, Barry P. Katz, Stephen M. Downs and Thomas S. Inui

In 2003 the Indiana Office of Medicaid Policy and Planning implemented the Indiana Chronic Disease Management Program (ICDMP). This paper reports on the fiscal impact of the ICDMP from the state’s perspective, as estimated from the outcomes of a randomized trial. Medicaid members with congestive heart failure (CHF) or diabetes, or both, were randomly assigned by practice site to chronic disease management services or standard care. The effect of the ICDMP varied by disease group and risk class: while cost savings were achieved in the CHF subgroup, disease management targeted to patients with only diabetes resulted in no significant fiscal impact.



--------------------------------------------------------------------------------
CHRONIC ILLNESSES ACCOUNT FOR MORE THAN 75 percent of total U.S. health care spending.1 Employer groups, health care organizations, and payers are turning to chronic disease management (CDM) programs not only to foster self-management and improve quality of care, but also to curb costs.2 These programs typically involve identification of eligible populations, support for patient self-management, and implementation of evidence-based guidelines or improved primary care.3 In the past decade, more than twenty state Medicaid agencies have instituted CDM programs.4 Despite these programs’ rapid expansion, their net fiscal impact is not clear. Virtually all programs have been evaluated using observational study methods, so estimated cost savings may reflect regression to the mean, selection bias, or confounding.5 In a critique of the literature on CDM, the Congressional Budget Office (CBO) urged future studies of disease management programs to adopt more rigorous experimental designs.6
In 2003 the Indiana Office of Medicaid Policy and Planning (OMPP) introduced the Indiana Chronic Disease Management Program (ICDMP), designed for people with diabetes or congestive heart failure (CHF), or both, in its Aged, Blind, and Disabled Medicaid population. Program interventions were developed by the Indiana OMPP using an approach consistent with the Chronic Care Model.7

Two types of interventions were developed: an intensive nurse care management program offered to program members who were determined to be high risk (for high levels of future utilization), and a less intensive telephonic program offered to program members who were determined to be low risk.8 Nurse care management was provided by registered nurses (RNs) who were assigned to review each member’s medical record, to visit members at home to assess needs and help set self-management goals, and to provide follow-up by phone. The nurse care managers (NCMs) did not act in a direct clinical capacity; rather, they provided educational and motivational materials for Medicaid members, made referrals to community resources, and helped facilitate members’ communication with their primary medical providers (PMPs). Telephonic care management was delivered by trained nonclinical care coordinators under the supervision of an RN. The care coordinators, guided by prepared scripts, delivered one to three phone calls in a series of educational and skill-building telephone interventions. The telephone center also mailed low-literacy educational materials.9

In this paper we report the impact of the ICDMP on overall Medicaid spending, including direct programmatic costs, as well as impacts on subsequent claims paid by Medicaid, as assessed in a randomized controlled trial of the program.

Study Data And Methods
Top
Study Data And Methods
Results
Discussion
NOTES


Design and setting. The OMPP staggered the implementation of the ICDMP within two large Indianapolis group-practice populations, to allow a randomized comparison between eligible members, assigned to the program in July 2003, and contemporaneous controls, eligible but not assigned until June 2005.

Individual participants. The OMPP used administrative data from 1 April 2002 through 31 March 2003 to identify eligible Medicaid members with CHF or diabetes, using a combination of International Classification of Diseases, Ninth Revision (ICD-9), codes and drug and medical supply claims. Exclusion criteria were adopted to ensure that all Medicaid members eligible for the ICDMP were served in one of two fee-for-service (FFS) primary care case management programs.10

Allocation. Eligible members were allocated to the ICDMP or control based on their PMP’s practice location. Allocation was complicated by large disparities in practice size. To minimize imbalances in numbers between treatment and control groups, we ordered the fifty practices by number of total eligible Medicaid patients and then randomly assigned the largest practice to ICDMP or control status. The remaining practices were serially assigned, in descending order of size, to either the ICDMP or control group by determining which assignment minimized the sample-size difference between the two groups at each step. A few patients either moved from the telephonic intervention to the nurse care management program or developed the second chronic condition during the evaluation. However, for all analyses of trial outcomes, we retained patients in their initial disease and risk subgroups.

Data sources. Programmatic costs (direct costs of development, implementation, and administration of the ICDMP) were obtained from data provided by the OMPP and relevant contracted entities. Health care utilization and claims paid by Medicaid were obtained from the OMPP’s fiscal agent.

Analytic methods. Cost inputs for our evaluation included those necessary for the direct provision of the intervention (including nurse care manager and telephonic efforts), Medicaid staff time needed to administer the program, ICDMP contractor resources for information management, and consultative assistance used to develop and support the program.11

Claims paid were calculated from Medicaid administrative data for each person in the randomized trial. Comparison of claims paid between the intervention and control groups required sophisticated multivariable statistical adjustment methods because of the outcomes of the group-level randomization used in our research design and features of the claims data themselves.

To adjust for differences in baseline characteristics that were potentially related to subsequent claims paid, we included covariates in our statistical model. Our covariates control for baseline differences in age, sex, risk status (high or low), number of unique prescription drugs, months of eligibility for Medicaid and Medicare coverage, and differences in the levels of claims paid during the twenty-one-month pre-ICDMP period. A second complication from the group-level randomization method was the creation of possible clustering of observations by practice. To adjust for possible within-practice correlation, we adjusted our variance estimates using generalized estimating equations (GEE).

Two features of the resultant data further complicated the analysis. First, discontinuities in enrollment created unbalanced panels. To control for the effects of such imbalances, particularly any potential seasonal effects, we exactly matched each member’s months in Medicaid in the twenty-one-month evaluation period (1 September 2003 through 31 May 2005) with the corresponding months in the twenty-one-month pre-ICDMP period (1 September 2001 through 31 May 2003). Second, as is typically observed in health care data, the distribution of claims paid was found to be highly skewed. Problems with skewed data are often addressed by using a log transformation of the dependent variable; however, we found that such adjustments were inadequate for the data used in this analysis. Instead, we adopted a generalized linear model with a log link function with a flexible gamma distributional assumption. The difference in the expected costs derived from these equations for the intervention and control groups was used to calculate the effect of the intervention on claims paid by Medicaid.

To evaluate the net costs of the ICDMP, we calculated the per participant per month (PPPM) programmatic costs and differences in claims paid using an "intent-to-treat" framework. Thus, participant months reflect the total number of people eligible for the ICDMP each month, regardless of whether they received any care management intervention.12

Results
Top
Study Data And Methods
Results
Discussion
NOTES


Cohorts at baseline. In June 2003, Indiana Medicaid’s fiscal agent identified 857 people who met eligibility criteria for the study. Thirty-one were excluded from the analysis because they (1) died before September 2003, (2) did not have corresponding months in Medicaid in 2001–2003 and 2003–2005, or (3) were found in retrospective analysis to have either no CHF or diabetes claims or no risk score calculated during April 2002–March 2003. After these exclusions, the analytic data for the trial included 826 people (Exhibit 1).




View this table:
[in this window]
[in a new window]

EXHIBIT 1 Baseline Cohort Characteristics, Study Of The Indiana Chronic Disease Management Program (ICDMP), 2001–2005




Estimates of ICDMP costs. For our base-case analysis, we calculated that the OMPP incurred costs of $57.03 PPPM for nurse care management for high-risk clients (of which 7.3 percent was for administrative support) and $19.54 PPPM for the telephonic intervention for low-risk clients (of which 21.1 percent was for administrative support) (Exhibit 2). These estimates excluded certain start-up costs that were "sunk" in nature and were considered irrelevant for future policy decisions by the OMPP. Because costs to start a new program would be faced by other agencies that might replicate the ICDMP, we performed a sensitivity analysis that also included the start-up costs paid by the OMPP for the development of the ICDMP after 1 July 2003. Because we expensed start-up costs over the twenty-one-month evaluation period in this sensitivity analysis, these estimates overstate start-up costs for new programs in place for longer periods. Including start-up costs, the OMPP’s costs increased to $62.56 PPPM to serve high-risk members and $23.86 PPPM to serve low-risk members. We also considered the impact of capacity constraints on our estimates by assuming that the NCMs and telephone center were operating only at 90 percent capacity. This latter figure was chosen to show the possible impact of unused capacity on results, and it does not imply that service providers were operating either efficiently or inefficiently. When we increased the number of people served to the implied hypothetical maximum, average costs to Medicaid fell to $55.39 PPPM for high-risk members and $18.58 PPPM for low-risk members.



View this table:
[in this window]
[in a new window]

EXHIBIT 2 Average Indiana Chronic Disease Management Program (ICDMP) Programmatic Costs, By Type Of Expense, 2003–2005




Estimates of effects on claims paid. Unadjusted and adjusted estimates of program effects on claims paid by Medicaid are presented in Exhibits 3 and 4. Because of differences in the usual clinical presentation and natural history of diabetes and CHF, it is not surprising that differences in the relationship between claims paid and covariates could not be adequately captured by a simple shift factor in the regression analysis. So we estimated two separate equations, one for members with diabetes only, and one for all members with CHF (including those with CHF and diabetes). Similarly, we found that complex differences also existed between high-and low-risk groups within each disease group. These subtleties can also be captured by estimating separate equations by risk class, but such partitioning creates small sample sizes, notably for the high-risk CHF group (n = 69). Given these considerations, we report results (1) stratified by disease group and risk class, and (2) pooled over both risk classes within each disease classification.



View this table:
[in this window]
[in a new window]

EXHIBIT 3 Average (Unadjusted) Claims Paid For Indiana Chronic Disease Management Program (ICDMP) Participants And Control Group, 2001–2005






View this table:
[in this window]
[in a new window]

EXHIBIT 4 Multivariable Estimates Of Indiana Chronic Disease Management Program (ICDMP) Effects On Claims Paid By Medicaid, 2003–2005




Disease category. After controlling for baseline differences in age, sex, risk status, prescription drug use, Medicare coverage, and levels of claims paid during the twenty-one-month pre-ICDMP period, we found that average monthly claims paid for all members with CHF were $283.01 lower in the ICDMP group than in the control group (Exhibit 4). Conversely, among members with diabetes, average claims in the ICDMP group were $3.81 higher than in the control group.
Risk strata. For both disease categories, the largest savings in claims paid by Medicaid were for low-risk members, who were offered the telephonic intervention. Average monthly claims paid per low-risk member were $247.11 lower in the ICDMP group than in the control group with CHF and were $3.81 lower per low-risk member in the ICDMP group than in the control group with diabetes. In contrast, savings in claims paid for high-risk members with CHF were more modest (–$149.74); for members with diabetes, claims paid were actually higher for high-risk members in the ICDMP group than in the control group ($143.68).

Sensitivity analysis. The results reported above are subject to two distinct types of uncertainty, one stemming from the choice of methods used in calculating programmatic costs, and another arising from the statistical imprecision in the estimation of claims paid. The effects of both sources of uncertainty are evaluated jointly using sensitivity analysis (Exhibit 5). As a "worst case" scenario, we included all start-up costs incurred after 1 July 2003 and assumed that the effect on claims paid by Medicaid was represented by the lower bound of the 50 percent confidence interval estimated by the multivariate analysis. In this scenario, only the CHF program for low-risk members resulted in overall cost reductions to Medicaid. As a "best case" scenario, we assumed that the ICDMP could operate more efficiently and assumed that the effect on charges paid by Medicaid was represented by the upper bound on the 50 percent confidence interval estimated by the multivariate analysis. In this scenario, only the diabetes program targeting high-risk members failed to achieve savings for Medicaid. Although not all findings proved robust, we can be fairly confident that the ICDMP reduced net overall costs for low-risk CHF participants but increased net overall costs for high-risk diabetes participants.




View this table:
[in this window]
[in a new window]

EXHIBIT 5 Impact Of The Indiana Chronic Disease Management Program (ICDMP) On Costs To Medicaid, Stratified By Patient Type, 2003–2005




Discussion
Top
Study Data And Methods
Results
Discussion
NOTES


This paper reports the impact of the ICDMP on Medicaid spending for members with CHF or diabetes, or both. Because the OMPP supported a staggered implementation in two Indianapolis group practices, we were able to conduct an evaluation within an experimental design. To our knowledge, this study is the first randomized controlled trial of a CDM program in a Medicaid population. In addition, our analysis incorporates estimates of programmatic costs, and it provides a more complete accounting of the fiscal impact of CDM than studies that have focused on utilization effects alone.
Our results suggest the potential cost savings attributable to the ICDMP varied by target condition and whether or not the Medicaid member was considered to be at high risk for future health care use. Significantly reduced overall health care spending was observed in the CHF subgroup, particularly those assigned to receive call-center support. In contrast, no statistically significant impacts on claims paid were found for members with diabetes. Such findings are consistent with prior research suggesting that CHF may be more amenable to disease management interventions than is diabetes, at least in the relatively short term.13

Surprise findings. We were frankly surprised to find a larger program impact in the low-risk than in the high-risk CHF subgroup, a result at odds with existing literature. The fiscal impact literature with which we were familiar is based largely on uncontrolled program evaluations. Our a priori expectations and, therefore our surprise, may reflect the dominance of studies subject to regression-to-the-mean effects that may affect their findings. Such biases not only tend to overstate the potential cost savings for high-risk people with high initial costs, but can also tend to understate potential cost savings for low-risk people with low initial costs. Differences in effectiveness by risk strata may also reflect the difficulty of affecting utilization in very sick populations, or differences in the net impacts of the nurse care manager versus the less intensive telephonic intervention.

Study limitations. Although our research design offers a number of advantages over observational studies, there are also some notable limitations. First, the analysis was based on a relatively small sample. As a result, its overall statistical power to detect meaningful differences in health care costs is modest.

Second, although our group-level randomization procedure worked well to balance sample sizes between ICDMP and control practices, the relatively small numbers of practices and members resulted in imbalances in the baseline characteristics of these two groups. We adopted methods in which statistical controls were used to adjust for several observed differences that otherwise could have affected the estimate of the ICDMP on claims paid, including baseline claims paid.14

Third, our analysis used an intent-to-treat model and did not account for differences in program exposure by Medicaid members in the ICDMP group, nor did it identify the aspects of the intervention that had the greatest impact on costs within that group. Our experimental framework was designed only to detect differences in overall costs for the program. With the sample sizes obtained, it is unlikely that our data could be used to identify program- or participant-specific impacts to any reliable degree. Our goal was to measure the effect of a program as experienced in a community setting, not to assess efficacy in a highly controlled environment. Thus, we did not attempt to adjust our results for limited penetration, including within the telephonic component of the program.15

Fourth, the use of a twenty-one-month follow-up period limited the opportunity for program maturation and longer-term program effects to be realized. This period approximates the biennial time frame used in state-level budget making in Indiana, and thus can be used to assess the short-term budget-neutrality of the ICDMP from the state’s perspective. Although the longer-term impacts on health and costs are certainly of interest to policymakers and others, our analysis, unfortunately, cannot illuminate what these impacts might be.

Fifth, the results of this trial might not be transferable to other settings. Although the ICDMP included the essential elements of any good CDM program, it was developed using local resources and might not be reproducible elsewhere. Also, the effects of the program might not be robust to differences in health care delivery, system practices, or private and public insurance circumstances.

Implications. Our analyses have possible implications for the conduct of health services research, clinical practice, and health policy. We have identified a number of challenges as well as the feasibility of using an experimental design to assess the impacts of chronic disease management. Our results suggest that some of the benefits of chronic disease management found in observational studies may be overstated because of nonexperimental bias, although they confirm that short-term cost savings are more likely to be seen in chronic disease management targeted to CHF patients than patients with diabetes. Perhaps most surprisingly, we found greater fiscal gains from disease management targeted to lower-risk rather than higher-risk patients. Such a finding suggests that concentrating disease management efforts on patients who are already likely to experience high use of services might not be an effective strategy in reducing overall costs.



Editor's Notes

Ann Holmes is an associate professor in the School of Public and Environmental Affairs, Indiana University Purdue University Indianapolis (IUPUI). Ronald Ackermann is an assistant professor of medicine in the Department of Medicine, Indiana University (IU) School of Medicine, in Indianapolis. Alan Zillich is an assistant professor in the School of Pharmacy and Pharmaceutical Sciences, Purdue University, in Indianapolis. Barry Katz is director of the Division of Biostatistics, Department of Medicine, IU School of Medicine. Stephen Downs is an associate professor, director of Children’s Health Services Research, and director of General and Community Pediatrics, in the Department of Pediatrics, IU School of Medicine. Thomas Inui (tinui@iupui.edu) is president and chief executive officer of the Regenstrief Institute in Indianapolis.

An earlier version of this paper was presented at the AcademyHealth Annual Research Meeting, in Seattle, Washington, 25–27 June 2006. This research was funded, in part, by the Indiana Office of Medicaid Policy and Planning (OMPP). The authors acknowledge the contributions of Kathryn Moses, Katie Holeman Shipp, Melanie Bella, Jeanne Labrecque, Gregory Wilson, Roberta Ambuehl, John Barth, Mary Jo Golubski, and Amanda Mizell.