Technical Appendix. This appendix provides more details about patient identification, consent, randomization,

Transcription

1 Peikes D, Peterson G, Brown RS, Graff S, Lynch JP. How changes in Washington University s Medicare Coordinated Care Demonstration pilot ultimately achieved savings. Health Aff (Millwood). 2012;31(6). Technical Appendix This appendix provides more details about patient identification, consent, randomization, and estimation of program impacts. A. Patient Identification, Consent, and Randomization This section describes (1) StatusOne s algorithm to identify Medicare beneficiaries for participation in the program, (2) the project s approach to informed consent and review by an institutional review board, (3) the process Mathematica Policy Research used to randomize beneficiaries to the care management and usual care groups, (4) the time periods and eligibility for the research sample and follow-up, and (5) the comparability of care management and usual care group beneficiaries before randomization. 1. Algorithm to Identify Eligible Medicare Beneficiaries StatusOne used a proprietary algorithm to identify Medicare beneficiaries eligible for participation in the program. It ran this algorithm on two years of Washington University physicians claims to identify patients who were likely to require a hospitalization in the next year. StatusOne reported that a rough proxy for its algorithm is patients who had (1) claims for two or more of six conditions (diabetes, congestive heart failure, chronic obstructive pulmonary disease, asthma, neoplasms, or renal disease); OR (2) two or more hospitalizations in the prior year; OR (3) two or more emergency room visits in the prior year and one or more of the six conditions. However, this proxy definition identifies beneficiaries who are substantially healthier on average than those who actually enrolled. 1 1

2 2. Informed Consent and Approval by Institutional Review Boards The secretary of the U.S. Department of Health and Human Services, acting through the Centers for Medicare & Medicaid Services (CMS), determined that the overall demonstration and evaluation met all criteria under both the Common Rule and National Institutes of Health s Exemption 5 for exemption from approval by an institutional review board for research and demonstration projects on public benefit and service programs. Mathematica Policy Research has a federal-wide Assurance of Protection for Human Subjects for demonstrations conducted by governmental agencies. Although neither the legislation nor the U.S. Department of Health and Human Services required certification of review by an institutional review board for this exempt research, each of the Medicare Coordinated Care Demonstration programs decided on its own whether to claim the exemption or to seek approval of its protocols from its local institutional review board. Washington University obtained approval from its institutional review board. All study participants provided written informed consent; Washington University allowed proxies to provide consent when necessary and appropriate. 3. Process for Randomizing Enrollees to the Care Management and Usual Care Groups CMS contracted with Mathematica to evaluate the demonstration; Mathematica designed and ran the randomization process for each program. Washington University intake staff identified eligible patients, invited them to participate, and transmitted patient information from consenting beneficiaries to Mathematica s secure study website. Mathematica checked the information for previous enrollment, completeness, and validity of information and then randomized eligible applicants within each program to the care management or usual care group in a 1:1 ratio. The assignment was carried out using a sequence of randomly selected, concealed, strings of assignments drawn (with replacement) from the set of 28 possible strings of two, four, or six assignments that each had equal numbers of care management and 2

3 usual care assignments. This approach ensured that no runs of more than six consecutive assignments to the same group would occur. The website returned the random assignment result within seconds to the program. Because of the nature of the intervention, no individuals were blinded to the group to which participants were randomized. 4. Time Periods and Eligibility for Research Sample and Follow-Up The research sample for impacts before the redesign included beneficiaries who enrolled in the program between its start on August 1, 2002, and February 28, 2005, and who met CMS s demonstration-wide eligibility criteria for at least one month during the period before the redesign (August 1, 2002, through February 28, 2006). The demonstration-wide eligibility criteria for a patient-month to be included in the sample required the beneficiary to be alive at least part of the month, have fee-for-service Medicare coverage, be enrolled in Medicare Parts A and B, and have Medicare as the primary payer for medical expenses. The research sample for impacts after the redesign includes beneficiaries who enrolled in the program between its start on August 1, 2002, and July 31, 2007, and met the demonstration-wide eligibility criteria for at least one month after the redesign (March 1, 2006, through July 31, 2008). The sample cutoff point in each time period ensured that every beneficiary potentially had at least 12 months of follow-up. The higher-risk subgroup met Washington University s eligibility criteria and had two or more hospitalizations in the two years before randomization. 5. Comparability of Care Management and Usual Care Group Beneficiaries Before Randomization The care management and usual care beneficiaries in the research samples had no statistically significant baseline differences before or after the redesign, demonstrating that randomization successfully led to balanced care management and usual care groups on observable characteristics. Appendix Table 1 shows the comparability of baseline characteristics 3

4 for demographics, diagnoses, and prior cost and service use. There is only one statistically significant difference at the p < 0.10 level of 32 comparisons made, fewer than the 3 that would be expected by chance. Appendix Table 2 shows the same comparisons of baseline characteristics for the subgroup of enrollees that was at higher risk of future hospitalizations. The table shows only 4 significant differences of the 32 comparisons made, slightly more than the 3 expected from chance alone. The four differences are: (1) percentage of enrollees in the pre-redesign research sample with a diagnosis of stroke, (2) percentage of enrollees in the post-redesign sample who are non-white Hispanic, and (3-4) percentage of enrollees in both the pre- and post-redesign samples with a cancer diagnosis. These and other characteristics were controlled for in the regression analysis used to estimate differences in outcomes between the care management and usual care groups. B. Estimation of Program Impacts This section describes (1) an analysis showing that selective attrition by the time the redesign occurred does not contribute to the estimate of the program impacts; (2) the method used to estimate impacts; (3) our rationale for using p < 0.10, rather than the traditional p < 0.05, as the cutoff for statistical significance; and (4) power calculations showing that the lack of impacts found before the program redesign were unlikely to be due to low statistical power to detect impacts. 1. Analysis of Selective Attrition One possible concern about the finding that the program had impacts after the redesign, but not before it, is that selective attrition rather than a true difference in program impacts might have driven the results. However, the impact estimates use an intent-to-treat design that substantially limits the possibility that selective attrition drives the results. All sample members remained in the analysis regardless of whether they actually received care management. Two 4

5 possible sources of selective attrition could have created bias in the impact estimates: (1) differential survival of care management and usual care group members in the research sample before the redesign and (2) program impacts during the period before the redesign on the percentage of enrollees after the redesign who met the demonstration-wide criteria for being in the research sample (which required that a beneficiary be alive, have fee-for-service coverage, be enrolled in Medicare Parts A and B, and have Medicare as the primary payer). Appendix Table 3 shows the percentage of care management and usual care group beneficiaries in the sample before the redesign who survived to March 1, 2006, the start of the post-redesign period. The differences in survival rates between the care management and usual care groups were small (1.0 and 0.3 percentage points for all and for higher-risk enrollees, respectively) and were not statistically significant (p = 0.60 and 0.91, respectively). We next examined the research sample, by also considering whether those who had survived to the postredesign period also met the demonstration-wide eligibility criteria. The difference was small and not significantly different from zero (data not shown). These data indicate that the program did not affect the percentage of study enrollees who were still alive or met the other demonstration-wide eligibility criteria at the time the post-redesign period began, indicating that these potential sources of selective attrition are not biasing the impact estimates. 2. Method for Estimating Impacts We used multivariate regressions to assess program impacts on hospitalizations and costs (with and without program fees). The covariates, specified in a detailed design document, increase the precision of the impact estimates and adjust for any chance baseline differences in the care management and usual care groups observable characteristics. 2 We used this regression model to estimate impacts: 5

6 where = the outcome for beneficiary i during the follow-up period. We estimated program effects on three outcomes: annualized hospitalizations (number per year), average monthly Medicare Parts A and B expenditures over the follow-up period, and average total Medicare expenditures per month with and without the care coordination fees paid by Medicare. = a binary variable that equals 1 if the beneficiary was randomly assigned to the care management group and 0 if assigned to the usual care group. = a vector of prespecified control variables, assessed at the time of enrollment, for individual i. These control variables are listed in the exhibits. We calculated the outcomes used in the regressions by analyzing each enrollee s Medicare Part A and B claims for the study period. We used the following equation to calculate a beneficiary s annualized rate of hospitalizations: For Medicare Part A and B expenditures (cost in dollars per beneficiary per month), we summed all Part A and B expenditures on the claims during the study period and divided by the beneficiary s number of observed months of follow-up (that is, the number of months the beneficiary was alive and met the demonstration-wide eligibility criteria). Care management fees paid to Washington University by CMS were recorded for each enrollee on special G-coded Medicare claims. To determine the total fees CMS paid for an enrollee, we summed all of these G-coded expenditures during the study period and divided by the number of observed months of follow-up for that enrollee (regardless of whether or not the beneficiary actively participated in the demonstration program that whole time). 6

7 The chronic condition control variables used in the regressions indicate whether a patient had a chronic condition at the time of enrollment. Chronic conditions were based on claims in the year or two before enrollment, as defined by the Chronic Condition Warehouse, version 1.5, with the exception of cancer, which was defined as having one or more inpatient or two or more hospital outpatient or carrier claims in the prior year for International Classification of Diseases, 9th edition, codes 140 to 208 (all cancers except skin cancer). Chronic conditions were not mutually exclusive. Section 4 of this Appendix describes the sample and follow-up cutoffs we used. We estimated impacts for all enrollees by including in the regressions all beneficiaries who were alive and met CMS s demonstration-wide criteria for at least one month during the followup period. We estimated impacts for the higher-risk subgroup by running separate regressions for the subset of all enrollees who, according to their Medicare claims, had experienced two or more hospitalizations in the two years before enrolling in the demonstration. To reflect the number of patient-months in the sample, the outcomes were weighted in proportion to the number of follow-up months during which each sample member was alive and met CMS s demonstration-wide requirements. Weights were calculated separately for the care management and usual care groups. We estimated the values for the coefficients in the regression model using ordinary least squares regressions. We ran 12 regressions (three outcomes for two follow-up periods (before and after the redesign) for two enrollee groups (all enrollees and higher-risk enrollees). The results from each of those regressions are presented in Appendix Tables Rationale for Using a p < 0.10 Cutoff for Statistical Significance In this paper, we consider the program to have had a statistically significant impact on hospitalizations or costs if the p-value for the care management versus usual control group 7

8 differences in a two-tailed test is less than 0.10, rather than the traditional A p < 0.10 level was used because we were concerned with type II as well as type I errors; that is, we wanted to have more confidence than a 0.05-level significance test provides that our test criteria would not result in a high probability of incorrectly concluding that the program had no effects if in fact it did have moderate-sized effects. We used examination of related outcomes to assess whether differences significant at only the p < 0.10 level were likely to be due to chance or true effects. That is, if the impact on hospitalizations is large and statistically significant at the 0.05 level, an estimate of effects on costs that is statistically significant at the 0.10 level but not the 0.05 level is likely to be a true effect, not a statistical anomaly. This is the same approach used in an earlier paper describing the impacts of the Medicare Coordinated Care Demonstration projects overall. 3 It should also be noted that the critical value for a two-tailed test at the 0.10 level is the same as that for a one-tailed test at the 0.05 level. Given that our policy interest is solely in whether the intervention can lower hospitalizations and Part A and B expenditures, a one-tailed test for these outcomes is appropriate. 4. Power Calculations for the Impact Estimates Before Program Redesign To address the concern that we might have found no effects before the redesign due to low statistical power, rather than a true lack of impacts, we calculated the power of the impact estimates during that period. With more than 1,000 beneficiaries assigned to both the care management and usual care groups, sample sizes were substantially larger than most studies of coordinated care. Although possible, it is unlikely that the lack of impacts seen before the redesign was due to low statistical power, rather than a true lack of program impacts. The tests had a 77 and 72 percent chance of detecting a program impact for hospitalizations and costs, respectively, for all enrollees, had the true impacts been at least as large as the point estimates we found after the redesign (assuming a two-tailed test and a p < 0.10 cutoff). Therefore, although 8

9 the program might have had impacts of modest size on hospitalizations and costs, it is unlikely that there were true program impacts during the period before the redesign that were as large as the estimates observed after the redesign. Endnotes 1 Archibald N, Schore J, Brown R, Peikes D, Orzol S (Mathematica Policy Research, Princeton, NJ). The Washington University Medicare Coordinated Care Demonstration Program after one year. Final report. Baltimore (MD): Centers for Medicare & Medicaid Services (US); Contract No.: (09). 2 Brown R, Aliotta S, Archibald N, Chen A, Peikes D, Schore J. Research design for the evaluation of the Medicare Coordinated Care Demonstration. Princeton (NJ): Mathematica Policy Research; 2001 Feb Peikes D, Chen A, Schore J, Brown R. Effects of care coordination on hospitalization, quality of care, and health care expenditures among Medicare beneficiaries: 15 randomized trials. JAMA Feb 11;301(6):

10 List of Appendix Tables Exhibit 1 (table) Appendix Table 1: Baseline Characteristics of the Care Management Versus Usual Care Group Enrollees in the Research Samples for Estimating Washington University Program Impacts Before and After the Program s Redesign SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Exhibit 2 (table) Appendix Table 2: Baseline Characteristics of the Higher-Risk Enrollees in the Care Management Versus Usual Care Groups in the Research Samples for Estimating Washington University Program Impacts Before and After the Program s Redesign SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Exhibit 3 (table) Appendix Table 3: Survival of the Research Sample Before the Redesign at the Start of the Post-Redesign Period SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Exhibit 4 (table) Appendix Table 4: Parameter Estimates (and t statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Hospitalizations SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Exhibit 5 (table) Appendix Table 5: Parameter Estimates (and I statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Part A and B Expenditures, Not Including Program Fees (dollars per beneficiary per month) SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Exhibit 6 (table) Appendix Table 6: Parameter Estimates (and I statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Part A and B Expenditures, Including Program Fees (dollars per beneficiary per month) SOURCES: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. 10

11 Exhibit 1 Appendix Table 1: Baseline Characteristics of the Care Management Versus Usual Care Group Enrollees in the Research Samples for Estimating Washington University Program Impacts Before and After the Program s Redesign Research Sample Before Redesign Research Sample After Redesign Care Management Group (n = 1,078) Usual Care Group (n = 1,066) Care Management Group (n = 1,087) Usual Care Group (n = 1,079) Enrollment Period 8/1/2002 2/28/2005 8/1/2002 7/31/2007 Follow-up Period 8/1/2002 2/28/2006 3/1/2006 7/31/2008 Age Age < 65 years 26.2% 27.7% 27.6% 28.3% Age >= 85 years 10.9% 8.3% 10.0% 8.9% Sex: Male 43.9% 46.2% 42.6% 44.6% Race Race: Black, non-hispanic 38.3% 35.8% 39.7% 38.5% Proportion with Selected Demographic Characteristics and Diagnoses Race: Hispanic 0.2% 0.1% 0.4% 0.1% State Part B Buy-in (proxy for Medicaid coverage) 20.9% 19.9% 22.1% 20.9% Coronary artery disease 64.7% 63.2% 62.8% 60.9% Congestive heart failure 48.7% 46.3% 43.2% 42.6% Diagnosis (not mutually exclusive) Average Individual Medical Use During the Prior Year Diabetes 39.1% 42.1% 38.7% 40.9% Chronic obstructive pulmonary disease 26.5% 25.5% 23.5% 22.1% Cancer 29.4% 26.3% 25.3% 22.6% Stroke 11.1%* 8.8% 10.4% 8.5% Depression 24.5% 22.7% 25.0% 23.3% Dementia 10.5% 8.5% 9.8% 8.8% Number of annualized hospitalizations 1.70/year 1.67/year 1.67/year 1.62/year Monthly Medicare Part A and B $2,365/month $2,450/month $2,294/month $2,220/month expenditures ($) Sources: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Notes: Diagnoses are medical conditions noted on Medicare claims during the one or two years before randomization. Cancer excludes skin cancer. A person meets the demonstration s insurance requirements before or after the redesign period if, for at least one month during that period, he or she is alive, enrolled in fee-for-services, has Medicare Parts A and B, and has Medicare as the primary payer for medical expenses. * = p < 0.10, ** = p < 0.05 for a two-tailed test comparing the care management and the usual care groups. 11

12 Exhibit 2 Appendix Table 2: Baseline Characteristics of the Higher-Risk Enrollees in the Care Management Versus Usual Care Groups in the Research Samples for Estimating Washington University Program Impacts Before and After the Program s Redesign Higher-Risk Research Sample Before Redesign Higher-Risk Research Sample After Redesign Care Management Group (n = 641) Usual Care Group (n = 593) Care Management Group (n = 624) Usual Care Group (n = 577) Enrollment Period 8/1/2002 2/28/2005 8/1/2002 7/31/2007 Follow-up Period 8/1/2002 2/28/2006 3/1/2006 7/31/2008 Age Age < 65 years 28.1% 31.7% 28.8% 31.9% Age >= 85 years 10.1% 9.6% 9.0% 10.7% Sex: Male 43.2% 43.3% 42.3% 42.5% Race Race: Black, non-hispanic 37.9% 38.6% 38.6% 40.9% Proportion with Selected Demographic Characteristics and Diagnoses Race: Hispanic 0.2% 0.0% 0.5%* 0.0% State Part B Buy-in (proxy for Medicaid coverage) 20.9% 21.2% 22.8% 22.9% Coronary artery disease 72.5% 71.2% 71.2% 71.1% Congestive heart failure 58.7% 58.3% 52.7% 54.8% Diagnosis (not mutually exclusive) Medical Use During the Year Before Randomization Diabetes 41.3% 42.8% 40.7% 44.7% Chronic obstructive pulmonary disease 30.6% 30.4% 28.8% 26.7% Cancer 28.4%** 22.4% 24.2%** 18.7% Stroke 14.5%* 11.3% 13.5% 11.8% Depression 29.5% 29.0% 30.8% 30.2% Dementia 13.3% 10.3% 11.4% 11.4% Number of annualized hospitalizations 2.54/year 2.66/year 2.62/year 2.71/year Monthly Medicare Parts A and B $3,278/month $3,513/month $3,315/month $3,361/month expenditures ($) Sources: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Notes: Diagnoses are medical conditions noted on Medicare claims during the one or two years before randomization. Cancer excludes skin cancer. Higherrisk enrollees met Washington University s eligibility criteria and had two or more hospitalizations in the two years before randomization. *= p < 0.10, ** = p < 0.05 for a two-tailed test comparing the care management and the usual care groups. 12

13 Exhibit 3 Appendix Table 3: Survival of the Research Sample Before the Redesign at the Start of the Post-Redesign Period Percentage (and number) of Enrollees in the Research Sample Before the Redesign That Were Alive on 3/1/2006 Care Management Group Usual Care Group Difference (p-value) All Enrollees 73.8% (796 of 1,078) 74.8% (797 of 1,066) -1.0% (0.60) Higher-Risk Enrollees 68.8% (441 of 641) 68.5% (406 of 593) 0.3% (0.91) Sources: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. Notes: The research sample before the redesign includes all beneficiaries who enrolled in the Washington University demonstration from August 1, 2002, to February 28, Higher-risk enrollees met Washington University s eligibility criteria and had two or more hospitalizations in the two years before randomization. 13

14 Exhibit 4 Appendix Table 4: Parameter Estimates (and t statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Hospitalizations Variable (all variables are binary unless otherwise noted) Before Redesign (4/2002 2/2006) After Redesign (3/2006 7/2008) All Enrollees (n = 2,144) R 2 = Higher-Risk Enrollees (n = 1,234) R 2 = All Enrollees (n = 2,166) R 2 = Higher-Risk Enrollees (n = 1,201) R 2 = In Care Management Group (-0.48) (-0.86) (-2.06) (-2.60) (vs. Usual Care) Age < (2.44) (3.60) (-0.45) (0.44) Age (-1.31) (-0.80) (-1.75) (-1.26) Age (-0.88) (0.13) (-0.92) (-1.26) Male (-0.75) (-0.22) (-0.55) (-0.90) Originally entitled for Medicare (-2.62) (-3.76) (0.01) (-0.86) due to a disability Had end-stage renal disease (2.74) (1.70) (4.10) (3.10) State Part B buy-in (a proxy for (1.20) (1.27) (2.54) (2.79) Medicaid coverage) Black, non-hispanic (1.37) (1.88) (2.87) (3.46) Hispanic (0.07) (-0.01) (0.80) (1.88) Total Parts A and B costs in (1.39) (-1.29) (-0.20) (-0.43) years prior in 2nd quartile of cost distribution in 3nd quartile of cost distribution (-0.15) (-1.81) (-0.44) (-0.88) in top quartile of cost (1.30) (-0.49) (-0.48) (-0.78) distribution Had congestive heart failure (1.75) (1.76) (0.63) (0.64) Had diabetes (2.64) (2.18) (3.85) (2.75) Had chronic obstructive (3.83) (2.93) (3.30) (2.29) pulmonary disease Had coronary artery disease (1.82) (0.23) (1.37) (-0.40) Had cancer (not skin) (1.42) (1.48) (0.96) (0.88) Had atrial fibrillation (0.80) (0.98) (2.99) (2.67) Had osteoporosis (-0.27) (-0.03) (-0.10) (-0.44) Had rheumatoid arthritis (0.53) (0.38) (-0.04) (0.64) Had depression (0.35) (0.81) (1.66) (1.37) Had a stroke (-0.78) (-1.07) (0.26) (-0.11) Had Alzheimer s/dementia (-0.41) (-1.02) (-1.34) (-1.89) Had chronic kidney disease (3.08) (1.76) (3.84) (2.38) Annualized number of hospital admissions in prior 2 years Any use of home health in prior year Any use of skilled nursing facilities in prior year (13.53) (10.45) (14.27) (11.10) (-0.02) (-0.07) (0.88) (0.26) (1.22) (0.99) (1.57) (1.52) Source: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database. 14

15 Exhibit 5 Appendix Table 5: Parameter Estimates (and t statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Part A and B Expenditures, Not Including Care Management Fees (dollars per beneficiary per month) Before Redesign (4/2002 2/2006) After Redesign (3/2006 7/2008) Variable (at time of randomization) (all variables are binary unless otherwise noted) All Enrollees (n = 2,144) R 2 = Higher-Risk Enrollees (n = 1,234) R 2 = All Enrollees (n = 2,166) R 2 = Higher-Risk Enrollees (n = 1,201) R 2 = In Care Management Group (vs. Usual Care) (0.71) (0.25) (-1.93) (-2.54) Age < (2.13) (2.91) (0.86) (1.64) Age (0.09) (-0.04) (-0.07) (0.53) Age (-0.73) (-0.43) (0.04) (1.02) Male (-0.26) (-0.57) (-0.20) (-0.88) Originally entitled for Medicare due to a disability (-2.01) (-2.99) (-0.85) (-1.68) Had end-stage renal disease (6.34) (4.93) (7.23) (5.74) State Part B buy-in (a proxy for (-0.14) (0.62) (1.90) (2.88) Medicaid coverage) Black, non-hispanic (2.00) (1.90) (2.89) (3.14) Hispanic (-0.44) (-1.17) (0.26) (1.07) Total Parts A and B costs in 2 years prior in 2nd quartile of cost distribution ($) (1.74) (0.72) (0.51) (1.85) in 3nd quartile of cost distribution (2.45) (1.21) (1.90) (0.88) in top quartile of cost distribution (6.16) (4.28) (3.39) (2.84) Had congestive heart failure (1.81) (1.48) (2.08) (2.01) Had diabetes (3.05) (2.62) (4.30) (3.65) Had chronic obstructive pulmonary (2.86) (2.48) (2.30) (1.49) disease Had coronary artery disease (0.29) (-1.11) (-0.48) (-2.63) Had cancer (not skin) (1.96) (2.12) (2.25) (1.97) Had atrial fibrillation (0.14) (0.19) (2.08) (1.94) Had osteoporosis (0.50) (0.86) (0.13) (0.14) Had rheumatoid arthritis (-0.37) (-0.04) (-0.37) (0.16) Had depression (-0.07) (0.07) (0.22) (-0.10) Had a stroke (0.23) (0.06) (0.80) (0.51) Had Alzheimer s/dementia (0.26) (-0.14) (-0.28) (-0.51) Had chronic kidney disease (6.18) (4.34) (5.96) (3.95) Annualized number of hospital (4.98) (3.01) (4.79) (3.53) admissions in prior 2 years Any use of home health in prior year (-0.29) (-0.15) (0.53) 2.37 (0.01) Any use of skilled nursing facilities in prior year (1.48) (1.12) (2.67) (2.74) Source: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database 15

16 Exhibit 6 Appendix Table 6: Parameter Estimates (and t statistics) for the Control Variables Used in the Multivariate Regressions to Estimate the Impacts of the Washington University Program on Part A and B Expenditures, Including Care Management Fees (dollars per beneficiary per month) Before Redesign (4/2002 2/2006) After Redesign (3/2006 7/2008) Variable (all variables are binary unless otherwise noted) All Enrollees (n = 2,144) R 2 = Higher-Risk Enrollees (n = 1,234) R 2 = All Enrollees (n = 2,166) R 2 = Higher-Risk Enrollees (n = 1,201) R 2 = In Care Management Group (vs. Usual Care) (2.42) (1.40) (-0.59) (-1.67) Age < (2.10) (2.87) (0.84) (1.61) Age (0.10) (-0.03) (-0.08) (0.52) Age (-0.73) (-0.43) 6.23 (0.03) (1.00) Male (-0.27) (-0.58) (-0.22) (-0.91) Originally entitled for Medicare due to a disability (-1.98) (-2.95) (-0.83) (-1.66) Had end-stage renal disease (6.35) (4.94) (7.22) (5.74) State Part B buy-in (a proxy for (-0.16) (0.60) (1.89) (2.86) Medicaid coverage) Black, non-hispanic (1.99) (1.88) (2.90) (3.15) Hispanic (-0.43) (-1.16) (0.24) (1.04) Total Part A and B costs in 2 years prior in 2nd quartile of cost distribution (1.72) (0.73) (0.50) (1.89) in 3nd quartile of cost (2.45) (1.23) (1.91) (0.93) distribution in top quartile of cost (6.15) (4.28) (3.39) (2.85) distribution Had congestive heart failure (1.81) (1.49) (2.09) (2.01) Had diabetes (3.05) (2.62) (4.28) (3.64) Had chronic obstructive pulmonary disease (2.87) (2.49) (2.32) (1.48) Had coronary artery disease (0.29) (-1.12) (-0.47) (-2.63) Had cancer (not skin) (1.96) (2.13) (2.27) (2.00) Had atrial fibrillation (0.12) (0.17) (2.06) (1.90) Had osteoporosis (0.50) (0.84) (0.12) (0.12) Had rheumatoid arthritis (-0.39) (-0.05) (-0.37) (0.16) Had depression (-0.08) (0.06) (0.20) (-0.12) Had a stroke (0.23) (0.07) (0.81) (0.53) Had Alzheimer s / dementia (0.26) (-0.14) (-0.29) (-0.51) Had chronic kidney disease (6.19) (4.34) (5.97) (3.95) Annualized number of hospital (4.99) (3.02) (4.80) (3.55) admissions in prior 2 years Any use of home health in prior year (-0.28) (-0.15) (0.50) (-0.02) Any use of skilled nursing facilities in prior year (1.45) (1.10) (2.64) (2.71) Source: Authors analysis of the Medicare National Claims History File, Standard Analytic File, and Enrollment Database 16