Yuting Zhang, PhD

The current benchmark method used by Medicare Shared Saving Program (MSSP) accountable care organizations (ACOs) favors historically high-cost ACOs. To address this problem, the new Next Generation ACO model modifies the benchmarking methodology by incorporating regional and national relative efficiency discounts.  We analyzed the effect of a benchmark on ACO savings under the original benchmark and these 2 adjusted benchmarking techniques.

We simulated different benchmarking methodologies and their effects on ACO savings.

 Using 2012 to 2014 Medicare spending data and MSSP ACO performance data, we applied county-level and national-level benchmark adjustments following the Next Generation ACO benchmarking method. We implemented both adjustments simultaneously and varied the adjustment caps between 0% and 5% to observe how the adjustment magnitude would affect savings for each ACO. We computed the average change in the newly calculated saving rate and earned shared savings for each $1000 change of the original benchmark, adjusting for ACO characteristics.

The Next Generation benchmark adjustments can reduce the correlation between ACO earned savings and historical spending, but the proposed adjustment levels are not large enough. By simulating different adjustment levels, we conclude that a combined national and regional discount cap of 3.5% best removes the correlation between benchmark and MSSP ACO performance.

Under the current MSSP, a higher benchmark level is associated with a better chance to save and larger earned savings. Although the Next Generation ACO benchmark adjustment is effective, the adjustment should be larger to completely remove the correlation between prior performance and an ACO’s savings.

Medicare offers 3 main types of accountable care organization (ACO) programs: Pioneer ACOs, initiated in 2011; Medicare Shared Savings Program (MSSP), initiated in 2012; and Next Generation ACOs, initiated in 2016. The great majority of Medicare ACOs are MSSP ACOs; as of 2016, there are 434 active MSSP ACOs serving over 7.7 million Medicare beneficiaries, but only 9 Pioneer ACOs and 21 Next Generation ACOs.

MSSP ACOs earn shared savings if the realized spending is less than a required percentage, ranging from 2% to 3.9% of the target spending.

 This benchmark, which is established at the start of a 3-year period, adjusts for patient risk scores and increases annually following a national spending trend, but makes no adjustment based on the relative historic spending level. As a result, providers who have achieved cost efficiency may be discouraged from entering or continuing to participate in an ACO payment model due to its reliance on historical performance. Several previous studies show that under the current benchmarking methodology, ACOs that earned more savings were located in high spending regions and had high spending relative to other ACOs in the same region.

In an attempt to fix this problem, the new Next Generation ACO model modifies the benchmarking methodology by incorporating regional and national relative efficiency discounts, so ACOs that have already attained cost efficiency compared with their region will have a more favorable discount.

 The Next Generation ACO also includes a standard discount of 3% and an adjustment based on ACOs’ quality scores. Since the inclusion of the 3% standard discount would not affect the relationship between saving and benchmark, and the quality adjustment has a phase in period in which it is not active, these features are not included in our analysis.

We used 2012 to 2014 Medicare spending data and MSSP ACO performance data to analyze the effect of a benchmark on ACO savings under 3 benchmark schemes: current MSSP benchmark and 2 benchmarks following the Next Generation ACO model, one with regional adjustments and another with national spending adjustments. The light purple bars in the 

 (panel A) show that the higher the current benchmark level, the higher the ACO saving rates. The darker colored bars in the Figure (panel A) show that this positive relationship weakens when incorporating the benchmarking adjustments. However, the regional and national adjustment levels are not large enough to remove the correlation. By simulating different adjustment levels, we concluded that a combined national and regional discount cap of 3.5% best removes the correlation between benchmark and MSSP ACO performance.

MSSP ACO public use files, from 2012 to 2014, contain information on ACOs’ spending in the first 2 performance years, as well as characteristics of assigned beneficiaries, including county of residence, Medicare enrollment type, and risk score.

 We combined this data with county-level fee-for-service (FFS) Medicare expenditures to determine each ACO’s spending level relative to the national and region levels. Our final dataset contained 329 ACOs. We dropped ACOs that stopped participating in MSSP and Golden Life Healthcare LLC—an outlier with a 44% saving rate; including this outlier does not substantially change results.

Outcomes are 2 ACO-level saving measures: 1) saving rate, defined as benchmark expenditures minus actual expenditures as a percentage of benchmark expenditures; and 2) earned shared savings per capita, defined as per-capita shared savings for ACOs that simultaneously exceeded their minimum saving rate and met the program’s quality standard.

Following the Next Generation ACO benchmarking method, we create 2 adjusted benchmarks: 1) county-level efficiency adjustment and 2) national efficiency adjustment. Specifically, if an ACO’s risk-adjusted historical per-capita baseline is higher (or lower) than the risk-adjusted county-level FFS baseline, the ACO’s county-adjusted benchmark is equal to the original benchmark decreased (or increased) up to 1%. Similarly, if an ACO’s historical baseline is higher (or lower) than the national level, the national adjusted benchmark is equal to the original benchmark decreased (or increased) up to 0.5%. We separately analyzed whether either adjustment can remove the relationship between historical spending and savings. Separating regional and national adjustments allowed us to compare them explicitly.

We then implemented both adjustments simultaneously, as in the Next Generation ACO model, and varied the discount caps between 0% and 5% to observe how the adjustment magnitude affects savings for each ACO. We computed the average change in the newly calculated saving rate and earned shared savings per capita for each $1000 change of the original benchmark, adjusting for ACO characteristics, including number of beneficiaries served in the ACO, average county spending level, the ACO’s patient demographic and enrollment characteristics, starting date indicators, and year indicators. Through this simulation, we determined the discount levels that best remove the positive relationship between benchmark and performance.

The purple bars in the Figure show the average saving rates of all ACOs and average earned shared savings per capita among ACOs that achieved saving by decile of the original benchmark amount. The higher the benchmark level, the larger the savings, confirming that the level of the benchmark has a large effect on both the extensive margin (whether or not an ACO saves) and the intensive margin (how much an ACO saves).

Regional and National Efficiency Adjustments Are Somewhat Effective

We recalculated the saving rate and earned shared savings per capita for each ACO under the 2 adjusted benchmarks. The first adjusted benchmark takes into account only the regional efficiency adjustment, while the second benchmark takes into account only the national efficiency adjustment. The Figure (panel A) shows the average saving rate by decile of the original benchmark when the saving rate is calculated using the original and the 2 adjusted benchmarks. The Figure (panel B) presents the relationship between these benchmarks and earned shared savings per capita. Both adjustments slightly mitigate the correlation between the original benchmark and the recalculated outcomes. The regional efficiency adjustment is more effective than the national efficiency adjustment, but this seems to be caused by the larger magnitude of the regional adjustment. Under both adjusted benchmarks, the positive relationships between the recalculated outcomes and the original benchmark are still present.

Controlling for ACO characteristics, the correlation between the original benchmark and the saving rate was 0.14 (

 = .00). After incorporating the national and regional adjustments, the correlations decreased to 0.13 (

 = .02), respectively. All the evidence suggests that the adjustment level currently being used in the Next Generation ACO model is effective, but not large enough to remove the correlation between benchmark and ACO performance.

To see what level of adjustment best removes the correlation, we varied the discount caps of these 2 adjustments and displayed the average change in the newly calculated savings rate for each $1000 change of the original benchmark in the 

 (panel A) and newly calculated earned savings per capita in the Table (panel B). In the Table, both adjustments are implemented simultaneously. The darker cells correspond to scenarios with a stronger and more statistically significant relationship between recalculated outcomes and benchmarks, while the whiter cells indicate a weaker relationship with less statistical significance. At the current Next Generation ACO implementation level—1% cap for regional adjustment and 0.5% cap for national adjustment&mdash;an increase of $1000 in the original benchmark is related to a 0.52% increase in the savings rate and a $19.17 increase in earned savings per capita. This is not trivial, considering the average savings rate is 0.54% and the average shared savings per capita is $79.80 in our sample.

The analysis reveals that a combined national and regional discount cap of 3.5% best removes the correlation between benchmark and ACO performance under an ACO model. At this level, the absolute value of correlation is the smallest and is statistically insignificant. If the combined cap is too small, the program rewards ACOs with high benchmarks; if the combined cap is too large, the relationship becomes negative and saving becomes harder to achieve for high-benchmark ACOs. Finally, as shown from the diagonal shape of the white cells in the Table, both adjustments have similar effects on the correlation, and the combined adjustment level is more important than the relative size of each adjustment.

The benchmark calculation can greatly influence realized savings for ACOs. The current MSSP benchmark is set to be ACO-specific and is based on the prior 3 years’ performance in order to encourage participation from high-cost ACOs.

It is essential to have a program that gives all ACOs a fair chance to succeed. Under the current benchmark system, the incentives to join MSSP are weaker for ACOs with a lower benchmark, as it is harder for them to achieve savings. Under a 2-sided risk model, which MSSP was originally scheduled to transition to, the incentive to participate in an ACO program is even weaker because ACOs face a larger risk. Moreover, the current MSSP benchmark is scheduled to be recalculated every 3 years, so previous high-cost, high-performance providers will eventually have a low benchmark. Before the recalculation occurs, attrition may be the best choice for ACOs that would receive a low updated benchmark.

Our study explores 2 benchmark adjustments that are implemented in the Next Generation ACO model and designed to remove the relationship between the benchmark level and an ACO’s success. The regional efficiency adjustment makes savings harder to achieve for ACOs that are performing poorly in their region. Similarly, the national efficiency adjustment makes savings harder to achieve for ACOs that spent more than the national spending average in the current performance year, lowering the correlation between high-cost regions and MSSP success.

At the current Next Generation ACO adjustment level, both adjustments can partially remove the relationship between an ACO’s spending level and the realized savings. However, a stronger adjustment is needed to completely remove the correlation. Our analysis revealed that the level of the combined national and regional discount cap is more important and the respective level of each adjustment does not greatly affect results. It is not surprising that regional and national efficiency adjustments have similar effects, given that more than 80% of the ACOs are affected by these 2 adjustments in the same way: if an ACO spent more than the regional average, it is also more likely to have spent more than the national average. We recommend implementing a benchmark adjustment with a combined cap of 3.5%. At this level, the positive relationship between historical spending and savings is completely removed, making savings equally attainable for all ACOs, independent of their previous spending level.

In summary, the MSSP benchmark is calculated based on the ACO’s previous 3 years’ claims data, making savings hard to achieve for historically low-cost ACOs. We note that a higher benchmark level is associated with a better chance to save and larger earned savings per capita. We also show, when adjusting benchmarks using the adjustment methods in the Next Generation ACO model, that the relationship between benchmark and saving is weakened; however, a benchmark is still an important determinant of saving. We tested different levels of regional efficiency and national efficiency adjustments and determined that a combined adjustment cap of 3.5% best removes the positive relationship between benchmark and ACO performance. Although the Next Generation ACO benchmark adjustment is effective at reducing the correlation between prior poor performance and an ACO’s savings, the adjustment should be larger to completely remove the positive relationship.

Lulu Liu, MA, and Richard Svoboda, MA, made equal contributions to the article.

Department of Economics (LL, RS) and Department of Health Policy and Management (YZ), University of Pittsburgh, Pittsburgh, PA.

Yuting Zhang gratefully acknowledges funding from the Commonwealth Fund (Grant 20150380). The sponsors played no role in the study conduct, data analysis, or report.

The authors report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article.

Concept and design (LL, RS, YZ); acquisition of data (LL, RS, YZ); analysis and interpretation of data (LL, RS, YZ); drafting of the manuscript (LL, RS, YZ); critical revision of the manuscript for important intellectual content (LL, RS, YZ); statistical analysis (LL, RS); obtaining funding (YZ); administrative, technical, or logistic support (LL, RS, YZ); and supervision (YZ).

Yuting Zhang, PhD, Department of Health Policy and Management, University of Pittsburgh, 130 De Soto St, Crabtree Hall A664, Pittsburgh, PA 15261. E-mail:

1. Next Generation Accountable Care Organization Model. CMS website. https://www.cms.gov/Newsroom/MediaReleaseDatabase/Fact-sheets/2016-Fact-sheets-items/2016-01-11.html. Published November 1, 2016. Accessed June 1, 2016.

2. Methodology for determining shared savings and losses under the Medicare Shared Savings Program. CMS website. https://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/sharedsavingsprogram/Downloads/ACO_Methodology_Factsheet_ICN907405.pdf. Published April 2014. Accessed June 1, 2016.

3. Heiser S, Colla C, Fisher E. Unpacking the Medicare shared savings proposed rule: geography and policy. 

 website. http://healthaffairs.org/blog/2015/01/22/unpacking-the-medicare-shared-savings-proposed-rule-geography-and-policy/. Published January 22, 2015. Accessed May 1, 2016.

4. McClellan M, Kocot SL, White R. Early evidence on Medicare ACOs and next steps for the Medicare ACO program (updated).

website. http://healthaffairs.org/blog/2015/01/22/early-evidence-on-medicare-acos-and-next-steps-for-the-medicare-aco-program/. Published January 22, 2015. Accessed August 25, 2016.

5. McWilliams JM, Song Z. Implications for ACOs of variations in spending growth. 

2012;366(19):e29. doi: 10.1056/NEJMp1202004.

6. Next Generation ACO model benchmarking methods. CMS website. https://innovation.cms.gov/Files/x/nextgenaco-methodology.pdf. Published December 15, 2015. Accessed June 1, 2016.

7. Shared savings program accountable care organizations (ACO) PUF. CMS website. https://www.cms.gov/Research-Statistics-Data-and-Systems/Downloadable-Public-Use-Files/SSPACO/index.html. Updated March 8, 2016. Accessed June 1, 2016.

8. Aggregate expenditure and risk score data on assignable beneficiaries by county. CMS website. https://www.cms.gov/Medicare/Medicare-Fee-for-Service-Payment/sharedsavingsprogram/Downloads/Assignable-Beneficiaries-by-County-for-CMS-1644-P.zip. Updated June 29, 2016. Accessed August 25, 2016.

Articles

The Medicare Shared Saving Program benchmark can be improved by following the example of Next Generation accountable care organizations, but with a larger adjustment level.

Using a Modified Next Generation ACO Benchmark Can Improve the MSSP

To examine the relationship between 6-month medication adherence and 1-year downstream heart disease—related readmission among patients who survived a myocardial infarction (MI).

Retrospective, nested case-control analysis of Medicare fee-for-service beneficiaries who were discharged alive post MI in 2008 (n = 168,882).

Patients in the case group had their first heart-disease—related readmission post MI discharge during the 6-to-9-month period or the 9-to-12-month period. We then used propensity score matching to identify patients in the control group who had similar characteristics, but did not have a readmission in the same time window. Adherence was defined as the average 6-month medication possession ratio (MPR) prior to the first date of the time-window defining readmission.

After controlling for demographics, insurance coverage, and clinical characteristics, patients who had a heart-disease—related readmission had worse adherence, with MPRs of 0.70 and 0.74 in the case and control groups, respectively. Odds ratio of MPR ≥0.75 was 0.79 (95% CI, 0.75-0.83) among those with a readmission relative to those without.

Our study shows that better 6-month medication adherence may reduce heart-disease—related readmissions within a year after an MI.

This study demonstrates that patients’ good adherence to essential myocardial infarction (MI) medications is a key factor in preventing readmission occurring 6 months to 1 year after discharge, and suggests that clinicians need to pay attention to outpatient medication adherence after inpatient discharge.

		The probability of having a preventable readmission 6 to 12 months post MI discharge was much lower for beneficiaries with good adherence to all MI drugs compared with those with poor adherence.

		Outpatient physicians should pay more attention to medication adherence among MI survivors to prevent readmission in the first year after discharge.

According to the American Heart Association, 7.9 million Americans have a history of myocardial infarction (MI) and 450,000 deaths occur each year in the United States because of new and recurrent MIs.

 Readmissions after MIs are common and costly. For example, the rates for 30-day all-cause readmission after an MI from 2006 to 2008 in the Medicare population were 22.5%, 24.8%, and 23% among white, black, and Hispanic patients, respectively.

 Previous research on readmissions focused primarily on 30-day all-cause readmission and examined the association between readmissions and hospital-level factors. This focus on 30-day readmissions was stimulated in part due to the Affordable Care Act (ACA), with Congress directing CMS to reduce payments to hospitals with higher than expected 30-day readmission rates. However, as a recent article notes,

 the policy makers’ emphasis on 30-day readmission may be misguided since the majority of readmissions take place outside the 30-day window. Additionally, patient-level factors outside of the hospital’s control may drive a considerable part of hospital readmission rates, especially for longer-term readmission.

One key factor affecting readmission rates is believed to be patients’ compliance with essential medications, including beta-blockers, lipid-lowering agents (statins), aspirin, and (depending on the patient) angiotensin-converting enzyme (ACE) inhibitors or angiotensin II receptor blockers (ARBs). Because extensive clinical trials have shown these drugs to be very effective, they have become important components of lifelong medical therapy for these patients.

 As a result, clinical guidelines now recommend that patients who survive an acute MI receive these medications.

Previous studies have shown that better medication adherence could reduce medical spending,

 in part due to reduced hospitalization and emergency department use. Both adherence and the likelihood of readmission are time-dependent variables, and we expected that the likelihood of readmissions was influenced more by drug adherence closer in time to the reference point than adherence further in the past.

To resolve the issue of time-dependent variables, we used a retrospective, nested case-control study design to examine the relationship between 6-month adherence following a hospitalization for MI and a subsequent readmission that occurs between 6 and 12 months post MI. This method is commonly used in medicine and health service research, but it has not been used to study adherence and readmission. We hypothesize that patients with better 6-month adherence to essential medications used to treat MI are less likely to have readmissions related to heart disease.

The key challenge to studying how adherence affects readmission is that both adherence and the likelihood of readmission are time-dependent variables. For example, to study the effect of 6-month adherence on the readmission rates within 1 year of discharge, we could not have simply compared the 1-year readmission rates among individuals with good or bad 6-month adherence. We needed to ensure that the interval used to measure adherence occurred within a reasonably short period prior to the time we measure readmission, so we could more confidently infer that the adherence affects readmission.

We attempted to address time dependency using a nested case-control design because it allowed us to measure adherence before potential readmission. Specifically, we identified the case group as beneficiaries who had their first heart disease—related readmission during 2 time windows post MI: 6 to 9 months and 9 to 12 months. We chose the follow-up period to be a minimum of 6 months for 2 reasons: first, short-term readmissions (eg, 30-day readmission) are more likely due to hospital factors than to patient medication adherence.

Second, 6 months is the shortest period to meaningfully measure adherence because prescriptions with a 90-day supply are now common.

We then used propensity score matching to identify the control group: Medicare beneficiaries who had similar characteristics to those in the case group, as measured by observed covariates, but who did not have a readmission in the same time window. For both groups, the main exposure variable (adherence) was defined as the average 6-month medication possession ratio (MPR) prior to the first date of the time window defining readmission. The first heart disease—related readmission day for each beneficiary in the case group was selected as the index date in order to focus the following 2 dates: the start date from which we counted back 6 months to define adherence, and the anchor date for the matched individuals in the control group to define adherence.

We used 2008 and 2009 pharmacy and medical claims data for all fee-for-service Medicare beneficiaries with Parts A, B, and D coverage who had an MI in 2008. MI was defined as having at least 1 inpatient claim with a primary or secondary diagnosis code as 410.X1 in 

International Classification of Diseases, Ninth Revision, Clinical Modification

We first identified beneficiaries who a) had an MI in 2008 and were discharged alive (the discharge date was defined as discharge index date); b) were enrolled in a standalone Part D plan continuously for 6 months after the discharge index date; and c) had a heart disease—related readmission (primary and secondary 

 diagnosis codes 390-459). We excluded patients who were readmitted within the first 6 months after discharge or who died.

Propensity Score Matching Adjustment Variables

To ensure that the case and control groups were comparable, we used propensity score matching. In calculating propensity scores, we first conducted a logistic model with the response variable as having a heart disease—related readmission between 6 and 9 months post discharge; that was the index date (ie, an indicator for case/control groups). In the model, we accounted for individual-level characteristics for demographic, insurance coverage, and health status. The demographic variables included gender, age (by group: aged ≤34 years, 35-50 years, 51-64 years, 65-74 years, 75-84 years, and ≥85 years), and race or ethnic group (non-Hispanic white, non-Hispanic black, Hispanic, Asian, and others/unknown, all relative to white). Part D data have an enhanced Research Triangle Institute Race Code verified by first and last name algorithms, which is much more accurate than the usual race variable found in claims data.

We also included information on whether the beneficiary qualified for Medicare because of disability.

The data included 2 additional insurance variables: whether the beneficiary had Medicaid (was dual-eligible) and/or received a low-income subsidy (LIS). These 2 variables provide information on the beneficiary’s income as well as the amount of co-payment and premium subsidies they receive. The income of the median dual-eligible is 75% of the federal poverty level (FPL), although there is some variation across states, while the incomes of the beneficiaries in the LIS group are below 150% of the FPL. Thus, the income in the non-LIS group is above 150% of the FPL. However, since the majority of beneficiaries in the LIS group also have some Medicaid coverage, we created 3 mutually exclusive income categories: duals, nondual LIS, and non-LIS.

The health status variables included 2 prospective risk scores: the CMS Hierarchical Condition Category score and the analogous score for prescription drugs. These risk scores were calculated using the diagnoses and spending in the year prior to the discharge index date. We calculated the prospective risk scores using prior year diagnoses, except for the 2.3% of the sample who were new enrollees, for whom we used concurrent risk scores based on age and gender.

Risk scores represent a proxy for health status, with higher scores indicating greater severity of illness and higher expected healthcare utilization. We included 2 additional health status indicators: one for a history of prior MI and the other for institutionalization, defined as 90 days of care in a nursing home. Our sample represents community-residing beneficiaries; less than 0.1% spent more than 90 days in nursing home facilities. Because Part D data include drugs used in nursing homes, we did not exclude these individuals.

Finally, the model accounted for zip code level income and education: logarithm of the median household income within the zip code in which the beneficiary lived, and educational level (percentages of residents in the zip code who had not completed high school, had completed high school only, and had attended or completed college).

We used nearest neighbor 1:5 matching with replacement, restricting the difference in propensity scores to be within .001. When using 6 to 9 months post MI as a window for readmission, we had 7264 cases and 68,074 potential controls. We were able to match 6864 cases. Our final sample included these 6864 cases and 33,726 controls. (We had 404 cases with fewer than 5 controls). When using 9 to 12 months post MI as the window for readmission, we had 5377 total cases and 61,500 potential controls. We were able to match 5184 cases and our final sample includes these 5184 cases and 25,843 controls.

We calculated 6-month adherence by tracing back from the readmission date for 6 months. Because the control group did not have a readmission date, we traced back from the anchor date, defined above, for 6 months to determine 6-month adherence. In addition, we constrained beneficiaries to be continuously enrolled in a stand-alone Part D plan during the 6-month MPR measurement periods. As a result, the total sample size dropped from 40,590 to 40,245 for the 6-to-9-month readmission window, and from 31,027 to 30,265 for the 9-to-12-month readmission window.

Two adherence measures were created: 1 for betablockers and the other for all 3 subclasses used to treat MI (beta-blockers, statins, and ACE inhibitors/ARBs). We measured adherence for only beta-blockers separately because the 6-month persistent use of beta-blockers post MI is a Healthcare Effectiveness Data and Information Set (HEDIS) quality measure. We did not measure adherencefor aspirin because aspirin is usually purchased over the counter and is not included in Part D claims data.

Adherence was measured by MPR, defined as the ratio of days of supply of medication the patient had in possession (numerator) over the number of days in the measurement period multiplied by the number of medications prescribed (denominator). Since we include several different medications into 1 MPR measurement, we only count medications in the denominator after the initial prescription following hospital discharge. For example, suppose a patient filled her first beta-blocker prescription on January 1, 2008, and her first ACE inhibitor prescription on February 1, 2008. Her MPR would be the number of betablocker pills in the first month divided by 30, while her MPR for the subsequent months would be her number of beta-blocker pills plus the number of ACE inhibitor pills divided by 60.

We then defined 2 indicators for good adherence with 2 commonly used thresholds.

 The first indicator: 1 = MPR ≥0.80; 0 = otherwise. The second indicator: 1 = MPR ≥0.75; 0 = otherwise. We excluded any days in the hospital, for any non—heart-related reason, from the denominator when calculating the MPR because medications used in the hospital cannot be observed in Part D event data (excluding or not excluding hospital days gives very similar results, because we were only measuring 6-month adherence). Prescriptions filled during any nursing home stay, however, can be observed in Part D data and so were included.

We tested the distribution of MPR and found it was not normally distributed. Thus, after propensity score-matching, we performed 2 types of regressions: a rate logistic regression and a binary logistic regression. In the rate logistic regression, the dependent variable is 6-month MPR ranging from 0 to 1. In the binary logistic regression, the dependent variable is the indicator for good adherence—we tested separately for MPR ≥0.8 and MPR ≥0.75. Robust standard errors were used to adjust for the dependence within each matched group. In both models, the key independent variable was the indicator for being in the case or control group&mdash;that is, whether one had a heart disease—related readmission during the time window. We also controlled for all the covariates used in calculating propensity scores, and we then applied a doubly robust procedure in estimating covariate effects in the main outcome models (the 2 types of logistic regression models). The resulting estimators give unbiased estimates of covariate effects when either propensity core or main outcome model is correctly specified, thus allowing the analyst 2 opportunities for obtaining accurate results (doubly robust). The assumption of no unmeasured confounders is still required.

Finally, with this study design, the regression results provide the probability of having good adherence conditional on whether one had a readmission. However, the probability of having a readmission conditional on whether one is a good adherent or not is the core policyrelevant variable. Thus, we used Bayesian probability theory to calculate the probabilities of having readmission given whether one was a good adherent or not. We calculated these separately for each time window and adherence measure.

The comparisons of characteristics between the case and control groups before and after propensity scorematching are summarized in 

 (9-12 months post MI cohort). Most characteristics are comparable between case and control groups after matching. For the characteristics that have statistically significant differences between the cases and the controls, the magnitude of the differences is small. In the 6 to 9 months cohort, about 15% of our sample were younger than 65 years; 43% were female; and 14% had a history of MI prior to 2008.

Effects of 6-month adherence on downstream readmission.

) regressions confirm that patients with downstream heart disease-related readmissions had poorer adherence for beta-blockers, ACE inhibitors/ ARBs, and statins prior to readmission. These results are after adjustment of all the patient-level and zip code level covariates discussed above. In particular, in the 6-to-9-months cohort, as shown in Table 3, the rate regressions demonstrate that the MPRs for all MI drugs were lower among those with a downstream readmission, 0.70 in the case group relative to 0.74 in the control group after adjustment for all covariates. Adherence to beta-blockers was slightly better, with means of 0.75 and 0.78 in the case and control groups, respectively. The results from the analyses of our 9 to 12 months data were similar.

Similarly, logistic regressions shown in Table 4 demonstrate that the odds ratio of having MPR ≥0.80 for all MI drugs was 0.78 (95% CI, 0.74-0.83) in the case group relative to that of the control group; the analogous odds ratio for beta-blockers was 0.84 (95% CI, 0.79-0.89) (Table 4). The results were similar when we used MPR ≥0.75 as a threshold for good adherence. Results are robust in the sensitivity analysis in the 9 to 12 months cohort.

Probability of having a heart disease—related readmission conditional on adherence status.

 report the probability of having good adherence conditional on whether one had a readmission. These numbers were estimated from the logistic regressions after adjustment of the variables used in the propensity score matching. The last 2 columns of Table 5 report the probability of having a readmission conditional on whether one was a good adherent or not, converted using Bayes’ rule. For example, the probability of having a readmission 6 to 9 months post MI discharge was 8.6% for beneficiaries with good adherence (MPR ≥0.75) for all MI drugs, whereas the probability was 10.7% for those beneficiaries with an MPR <0.75.

We found that patients with better adherence to MI drugs had lower downstream heart disease—related readmission rates, after controlling for observed patient-level and zip code level characteristics. Using the Medicare population who had a recent history of MI, our study demonstrates that medication adherence is an independent factor associated with lower downstream readmission after adjustment for other patient characteristics.

Few studies have examined readmission after 30-day post discharge. One study by Jencks and colleagues noted that among all Medicare fee-for-service beneficiaries in 2003 and 2004, the accumulative rate of rehospitalizations within 3 months postdischarge regardless of medical conditions was 34%; the accumulative rate within 6 months was 47.9%; and within 12 months postdischarge, 59.4%.

 Our study shows that better medication adherence can potentially reduce readmission and may therefore save money for Medicare.

Our study has some limitations. First, because there was not random assignment to the case and control groups, we used a propensity score matching method. However, there may have been unobserved variables that were related to both adherence and readmission, which may have biased our estimates. Second, our matching algorithm was not perfect; even after matching, the case and control groups differed significantly on risk scores, but the magnitude was small. Third, there may be some errors with measuring adherence using claims data: we cannot observe patients who were prescribed a drug but refused to take it, and we cannot observe all contraindictions to the studied drugs (eg, we cannot see measures such as low left ventricular ejection fraction, a contraindication for ACE inhibitors). In those cases where we saw a prescription filled in the claims for an ACE inhibitor, we assumed that the physician who prescribed the drug had evaluated the patient’s profile for contraindications. We also cannot distinguish whether an MI is an ST-segment elevation or not, which is associated with higher early mortality and morbidity. Finally, claims data do not have individual level demographic variables such as education and income, but we used zip code level data.

Beginning in 2013, hospitals began receiving decreased Medicare payments if they had higher than expected 30-day readmission rates for MI, heart failure, and pneumonia. Under ACA, the list of conditions will be expanded. We found that patient compliance to medication treatment is a significant independent predictor of readmission occurring between 6 and 12 months post MI discharge. This implies that patients need to be monitored past the typical 30-day time period regarding their medication use.

We would like to thank Shang-Hua (Sean) Wu, The UCLA Center for Health Policy Research for his excellent programming assistance.

Department of Health Policy and Management, Graduate School of Public Health (YZ, SHB, JRL) and Department of Medicine, School of Medicine (C-CHC), University of Pittsburgh, Pittsburgh, PA; Department of Preventative Medicine, College of Medicine, University of Tennessee Health Science Center (CMK), Memphis, TN.

National Institute of Mental Health (No. RC1 MH088510) and Agency for Healthcare Research and Quality (No. R01 HS018657).

Dr Zhang received grants from the National Institute of Mental Health and the Agency for Healthcare Research and Quality to fund this project. Drs Hyon Baik, Chang, Kaplan, and Lave report no relationship or financial interest with any entity that would pose a conflict of interest with the subject matter of this article. Authorship Information: Concept and design (YZ, SHB, C-CHC, CMK, JRL); acquisition of data (YZ, SHB, JRL); analysis and interpretation of data (YZ, SHB, C-CHC, CMK, JRL); drafting of the manuscript (YZ, SHB, C-CHC, CMK, JRL); critical revision of the manuscript for important intellectual content (YZ, C-CHC, JRL); statistical analysis (YZ, SHB, C-CHC, CMK); obtaining funding (YZ); administrative, technical, or logistic support (YZ); supervision (YZ).

Yuting Zhang, PhD, Department of Health Policy and Management, University of Pittsburgh, 130 De Soto St, Crabtree Hall A664, Pittsburgh, PA 15261. E-mail: 

1. Roger VL, Go AS, Lloyd-Jones DM, et al; American Heart Association Statistics Committee and Stroke Statistics Subcommittee. Heart disease and stroke statistics—2011 update: a report from the American Heart Association. 

Better outpatient medication adherence reduces the likelihood of readmission after a recent myocardial infarction.

Medication Adherence and Readmission After Myocardial Infarction in the Medicare Population

 To evaluate the effects of the Medicare Part D coverage gap on pharmacy use among a national sample of Medicare beneficiaries and on medication adherence among 2 subsamples with heart failure and/or diabetes.  

Pre-post design, with comparison group and propensity score weighting.  

We used a 5% random sample of elderly Medicare beneficiaries enrolled in stand-alone Part D plans in 2007. The comparison group had full coverage in the gap, whereas the 2 study groups had either no coverage or generic-only coverage in the gap. Main outcomes included probability of filling a prescription, monthly pharmacy spending and number of prescriptions filled, and adherence measured by medication possession ratios.

Relative to the comparison group, beneficiaries without drug coverage in the gap reduced the number of prescriptions filled per month by 16.0% (95% confidence interval [CI], 15.5%-16.5%); those with generic drug coverage in the gap reduced it by 10.8% (95% CI, 10.3%-11.4%). Most of the reduction was attributable to reduced use of brand-name drugs. Beneficiaries with heart failure reduced adherence to heart failure drugs by 3.6% (95% CI, 2.9%-4.2%) and beneficiaries with diabetes reduced antidiabetic medication adherence by 10.3% (95% CI, 9.4%-11.3%).  

Medicare beneficiaries reduced medication use (mainly brand-name drugs) after entering the coverage gap. This result suggests that while beneficiaries’ financial burden would continue because of the coverage gap, the gap would not result in a large reduction in medication adherence for essential drugs for diabetes and heart failure.  

 We examined the effect of the coverage gap on a national sample of Medicare beneficiaries as well as those with heart failure and/or diabetes. Although beneficiaries’ financial burden in the coverage gap will continue until 2020, the gap will not result in a large reduction in medication adherence for essential drugs.

		Medicare beneficiaries reduced medication use (mainly brand-name drugs) after entering the coverage gap.

		Beneficiaries with generic drug coverage in the gap reduced their drug use less.

		However, medication adherence was reduced to a much smaller degree for essential drugs used to treat heart failure and diabetes.

Medicare Part D offers prescription drug coverage for Medicare beneficiaries. Since the program’s inception in 2006, many beneficiaries have obtained improved drug coverage and increased their use of medications.

 The standard Part D benefit in 2007 included an initial $265 deductible, a period in which beneficiaries paid 25% of drug costs between $265 and $2400, a coverage gap in which they paid 100% of costs until their total out-of-pocket spending reached a catastrophic limit of $3850, and a catastrophic coverage period in which they paid 5%.

A major concern is the large coverage gap in the standard Part D design. About one-third of all Medicare beneficiaries enter this coverage gap each year.

 Faced with a sudden increase in medication costs, beneficiaries might cut back on their medication use and reduce medication adherence for the essential drugs.

That, in turn, might put their health at risk, leading to subsequent increases in hospitalizations and medical spending.

 Because of the high cost of Medicare and the potential implications for health outcomes, it is important to determine whether the coverage gap leads to reduction in medication adherence for essential drugs.

A few studies have examined the effects of the coverage gap on medication use, but the majority of these studies used pharmacy data either from 1 local Medicare-Advantage Part D plan

 One recent study used Medicare Part D data to examine the effects of the coverage gap among 2 subsamples with hypertension and hyperlipidemia.

 They found that the coverage gap wasassociated with reduction in use of and adherence to medications for hypertension and hyperlipidemia. Another recent study looked at aged Medicare beneficiaries with cardiovascular disease enrolled in a standalone Part D plan (PDP) or a retiree drug plan that was administered by CVS Caremark.

 They found that beneficiaries with no financial assistance during the coverage gap were more likely to discontinue their cardiovascular drugs.

In this study, we evaluated the effects of the coverage gap on medication spending and use among aged Medicare beneficiaries (those 65 years or older) in the national Part D data. In addition, we selected 2 subsamples of those with heart failure and/or diabetes to evaluate the gap effect on medication spending and adherence. We selected these 2 conditions because they are common in Medicare beneficiaries and because the medication treatment is very important— premature discontinuation may lead to increases in subsequent spending on medical care.

We determined how beneficiaries with these chronic conditions responded to the coverage gap compared with the general Medicare population.

Our analyses were guided by a number of hypotheses: (1) Medicare beneficiaries will decrease all their medication use regardless of drug classes once they enter the gap— and that the decrease will be higher for those with no drug coverage in the gap than for those with generic coverage in the gap.

However, those who anticipate going through the gap and entering the catastrophic period will not reduce their medication use in the gap.

 (2) Beneficiaries will decrease their use of all brand-name drugs more than generic drugs, and if beneficiaries have generic drug coverage they will shift from brand-name to generic drugs.

 (3) Among those with heart failure and/or diabetes, adherence to cardiovascular and diabetic drugs will decrease in the coverage gap.

We obtained beneficiary demographic and enrollment information, plan benefits, prescription drug events, and medical claims for a 5% random sample of all aged Medicare beneficiaries who were continuously enrolled in PDPs in 2007. We identified 2 subpopulations of beneficiaries: those with heart failure and/or diabetes. Each subpopulation was identified by 2 criteria: (1) having a diagnosis for the condition before January 1, 2007, defined by 2006 Centers for Medicare & Medicaid Services Chronic Condition Warehouse indicators (

; and (2) use of at least 1 medication for the condition in the initial coverage period so we could ensure that beneficiaries were on medications of interest before entering the coverage gap (

  provides the list of medications). The study design was approved by the institutional review board at the University of Pittsburgh.

Although the standard Part D benefit includes the 4 phases described in the first paragraph of this study, some companies offering PDPs modified the design and offered either “actuarially equivalent” or enhanced plans. In 2007, 72% of stand-alone PDPs had the standard coverage gap, 27% of PDPs had some generic coverage in the gap, and fewer than 1% offered coverage for both brand-name and generic drugs.

 In addition, beneficiaries with low incomes who are eligible for subsidies had a lower copayment throughout the year and did not face a coverage gap.

Using this variation in drug coverage, we identified 2 study groups based on the type of coverage they had in the gap (no coverage, generic only) and 1 comparison group. The comparison group consisted of beneficiaries who had either 12-month Medicaid coverage or 12-month low-income subsidies (LIS group). Thus, the drug coverage for the comparison group did not change throughout the year; that is, they did not face the gap in coverage. Even though the LIS group was different from the 2 study groups in socioeconomic characteristics, they could still serve the purpose of controlling for the underlying time trends in medication use, because their drug coverage did not change while the study groups had a sudden increase in drug costs once entering the coverage gap.

We used a pre-post study design with a comparison group to assess the impact of the coverage gap on medication and medical care use. This approach uses a difference-in-difference estimate, comparing the pregap and within-gap change for each study group with the pregap and within-gap change in the comparison group. This design does not require that study and comparison groups have the same baseline characteristics. As long as different groups have similar baseline trends, we could get unbiased results.

 We tested the baseline pregap trends in overall medication use across 3 groups, and they were similar.

We hypothesized that beneficiaries who went through the coverage gap and entered the catastrophic coverage period would not reduce their medication use in the gap.

 We ran regressions to test this hypothesis and found that those entering the catastrophic period did not change their medication use in the gap (

). Thus, for this study we primarily focused on beneficiaries who entered the gap but did not go through it. That is, each individual in our study sample had a pregap and a within-gap period.

To define pregap and within-gap periods, we first identified the index date as the first day that the beneficiary’s total drug spending reached the coverage gap threshold. The pregap period was defined as January 1, 2007, to the index date. The within-gap period was defined as the time from the first day after the index date until December 31, 2007. The index date was included in the pregap period because almost all prescriptions filled by beneficiaries on that day were subject to copayment levels in the initial coverage period. We compared the change in each outcome in the pregap and withingap periods between each study group and the comparison group: no coverage versus LIS and generic only versus LIS. We used a propensity score weighting mechanism to balance each study group with the comparison group. The propensity score weighting was conducted separately for each group: the general population and each subgroup of those with heart failure and/or diabetes.

For the pregap and within-gap periods separately, we defined 4 outcomes to measure medication use: (1) probability of using a medication (1 = used a medication; 0 = did not use a medication in the study period); (2) the mean number of monthly prescriptions filled per month, defined as the total number of prescriptions standardized by 30-day supply (ie, a prescription with a 90-day supply would count as 3 monthly drugs) divided by the number of months in the study period; (3) mean monthly pharmacy spending per month; and (4) medication adherence measured by medication possession ratio (MPR). For the general population, we focused on the first 3 outcomes. We measured these outcomes for all medications and then for brand-name and generic drugs separately.

For the patients with chronic conditions, we focused on the mean number of monthly prescriptions filled and the MPR. We examined the mean number of monthly prescriptions filled overall as well as the number of monthly prescriptions for disease-specific drugs. We defined the MPR as the proportion of days during a given period (eg, either the pregap or the within-gap period) that a subject had possession of any drugs used to treat the chronic illness. The prescriptions filled on the day the patient entered the coverage gap were included in calculating the within-gap MPRs for 2 reasons: (1) because these drugs were used in the within-gap period and (2) because inclusion of these drugs was consistent with the pregap MPR definition where drugs filled on the first day (January 1, 2007) were included in the calculation. Similarly, to be consistent with the MPR calculation in the pregap period where prescriptions filled before January 1, 2007, were not included, prescriptions filled before entering the gap were not included in the calculation for the withingap adherence. However, the definitions for MPRs should not change the difference-in-difference results because we applied the same rules for the study and comparison groups.

To implement the propensity score weighting mechanisms, we conducted a 2-stage analysis. In the first stage, we ran 2 logistic regression models to predict the probability of being in a study group relative to the comparison group, controlling for age, sex, race, number of Elixhauser comorbidities,

 and Prescription Drug Hierarchical Condition Category, the beneficiary risk adjuster used by Centers for Medicare & Medicaid Services to adjust payment to plans for pharmacy costs.

In the second stage, we ran a difference-in-difference model with the inverse of the propensity score as a weight. This effectively assigned a higher weight to individuals in the comparison group who had characteristics similar to those of individuals in the study group. In this model, the dependentvariable was the difference between within-gap and pregap periods for each previously defined outcome. Because pregap and within-gap measures were likely to be correlated, the advantage of this approach versus using 2 interrelated outcomes is that we could simply eliminate the correlated structure in 2 outcomes. The key independent variable was the indicator for being in the study group relative to the comparison group. All the covariates used in calculating propensity scores were included in the model.

In addition, we controlled for duration of time spent in the coverage gap (

) in the model because the longer beneficiaries stayed in the gap, the more likely it was that they would change their medication use and medical spending. All analyses were conducted using SAS software, version 9.2 (SAS Institute Inc, Cary, North Carolina) and R: A Language and Environment for Statistical Computing, version 2.12 (http://www.r-project.org).

 compares the characteristics between each study group and comparison group for beneficiaries whose pharmacy spending exceeded the coverage gap threshold but did not exceed the catastrophic coverage threshold. All the numbers are after adjustment with propensity score weights. After adjustment, all characteristics were comparable (ie, there was no statistically significant difference at 

 >.05) between each study group and the comparison group.  

Effects of the Coverage Gap on Medication Use and 

presents the effects of the coverage gap on the probability of using a drug, the mean number of monthly prescriptions filled, and the mean monthly pharmacy spending for all medications for the overall population. There are 3 main findings:

First, relative to the comparison group, there were statistically significant reductions in all of the outcomes (probability of using a drug, mean number of monthly prescriptions filled, and monthly pharmacy spending) in both study groups. However, those with no coverage generally decreased their use of medications more than those with generic drug coverage in the gap.

Second, the overall decrease in monthly medications and spending on drugs was primarily due to the decrease in the use of brand-name drugs. For example, those without drug coverage in the gap reduced their overall medication use by 0.85 medication per month (95% confidence interval [CI], 0.82-0.88); 75% of the reduction was accounted for by brandname drugs and 25% by generic drugs. This group decreased its monthly pharmacy spending by $73.15 (95% CI, $71.66- $74.65), of which $66.65 (95% CI, $65.30-$68.01) was for brand-name drugs and $6.40 (95% CI, $5.88-$6.92) was for generic drugs.  

Third, those with only generic coverage in the gap reduced their use of brand-name drugs but did not increase their use of generic drugs. In fact, they decreased their use of generic drugs slightly but negligibly. Among the general population, relative to the comparison group, the generic-only group reduced the number of monthly prescriptions filled by 0.66 (95% CI, 0.63-0.70); almost all of this decrease was attributable to the reduction in brand-name drugs (0.61 [95% CI, 0.59-0.62]).  

Effects of the Coverage Gap Among Patients With Chronic Conditions

Beneficiaries with heart failure and/or diabetes decreased their overall use of medications, and the overall decrease was similar to that found for the general population (

). They also decreased their use of the drugs specific to their conditions. The relative decreases in the condition-specific drugs were similar to those observed for the drugs overall.  

The decrease in medication use was accompanied by a decrease in medication adherence in both groups, but the decrease in adherence was smaller than the reduction in the number of prescriptions filled (

). In addition, relative to the comparison group, the decrease in medication adherence was always larger in the no-coverage group than in the generic-only group. In the heart failure sample, the MPR for heart failure drugs among the no-coverage group dropped from 0.87 in the pregap period to 0.83 in the within-gap period, while the MPR for those with generic-only coverage dropped from 0.87 to 0.86. Although statistically significant, the relative reduction in adherence for heart failure was negligible. The same pattern was observed for those with diabetes, although decrease in medication adherence was larger in this group.  

Medicare beneficiaries decreased their medication use once they entered the gap—and the decrease in use was consistently higher for those who had no coverage in the gap than for those with generic drug coverage in the gap. We hypothesized that beneficiaries would reduce their use of brand-name drugs after entering the coverage gap and if they had generic drug coverage, they would shift their brand-name drugs to generic drugs. We did observe that beneficiaries reduced their use of brand-name drugs substantially more than their use of generic drugs, but those with generic drug coverage did not switch from brand-name to generic drugs. This is partially because those in the generic-only group tended to use more generic drugs in the initial coverage gap period.  

We adjusted the time spent in the gap in our model so the beneficiaries spending less time in the gap contributed less to the regression results. In addition, we conducted sensitivity analysis by excluding those beneficiaries who stayed in the gap for less than 1 month or less than 2 months. The results were robust.  

Entering the Medicare Part D coverage gap does not result in a large reduction in medication adherence for essential drugs.

Effects of Medicare Part D Coverage Gap on Medication Adherence

Medicare Part D has a drug coverage gap, which imposes risks for discontinuing medications, particularly in mental health disorders where drug costs are high. However, some beneficiaries have generic drug coverage in the gap.

To examine the health outcomes and cost-effectiveness of generic-drug coverage compared with no gap coverage in patients with bipolar disorder and schizophrenia.

Markov model—based cost-effectiveness analysis using identical hypothetical cohorts to examine drug coverage strategies.

The incremental cost-effectiveness of Part D coverage strategies was estimated, using differences in medical costs and quality-adjusted life years between plans. Coverage strategy—specific costs and hospitalization rates were obtained from 2007 Medicare data, adjusted for age, sex, race, and health status.

 When comparing generic-only coverage with no gap coverage, generic-only coverage cost less and was more effective than no gap coverage, due mainly to lower hospitalization rates. In sensitivity analyses, generic-only coverage continued to be favored over no gap coverage unless generic coverage costs increased

3% in bipolar disorder and >5% in schizophrenia; generic coverage in the gap was also favored in probabilistic sensitivity analyses.

In Medicare Part D, generic drug coverage was cost saving compared with no coverage in bipolar disorder and schizophrenia while improving health outcomes. Policy makers and insurers might consider generic-only coverage, rather than no gap coverage, to both conserve healthcare resources and improve health.

Optimal drug benefit design can decrease drug and nondrug spending while also improving health outcomes.

		The Medicare Part D prescription drug benefit has a coverage gap, which has a greater impact on patients with bipolar disorder and schizophrenia.

		A cost-effectiveness analysis of Part D plans is unconventional but can inform managed care decision making.

		Generic drug coverage in the gap is cost saving compared with no gap coverage (the standard Part D design) and improves health outcomes.

Medicare Part D offers prescription drug coverage for Medicare beneficiaries. Since the program’s inception in 2006, many enrollees have benefited from improved drug coverage and increased medication use.

 However, a major concern is the large coverage gap in the standard Part D design, where beneficiaries pay 100% of medication costs out-of-pocket. About one-third of all Medicare beneficiaries enter this coverage gap each year,

and once there, they often reduce medication use,

 which may lead to increases in hospitalization and medical spending.

The coverage gap is an even larger concern for Medicare beneficiaries with severe mental disorders such as bipolar disorder and schizophrenia. First, they are much more likely to enter the gap: 62% of aged beneficiaries with bipolar disorder and 56% of those with schizophrenia entered the gap in 2007.

 Second, if they discontinue psychotropic medications, they may relapse to more severe episodes and require psychiatric hospitalization.

Third, they experience high rates of comorbid chronic physical conditions such as heart disease and diabetes, which can be exacerbated by untreated mental illness and increase morbidity, medical spending, and mortality.

The standard Part D benefit in 2007 included 4 phases: (1) an initial $265 deductible; (2) a period in which beneficiaries paid 25% of drug costs between $265 and $2400; (3) a coverage gap in which they paid 100% of costs between $2401 and $3850, where they reached their total out-of-pocket spending catastrophic limit; and (4) a catastrophic coverage period where they paid 5% of costs.

 Although the standard Part D benefit includes these 4 phases, some companies offering Part D drug plans modified the design and offered either “actuarially equivalent” or enhanced plans. In 2007, for example, 72% of stand-alone Part D plans had the standard coverage gap, 27% offered coverage for generic drugs used in the gap, and fewer than 1% offered coverage for both brandname and generic drugs.

In addition, the Medicare Part D drug benefit includes substantial premium and cost-sharing subsidies for Medicare beneficiaries with low incomes, which come directly from the federal government. Each year, roughly 50% of beneficiaries enrolled in Part D plans receive a low-income subsidy (LIS).

 Due to more generous benefits, those with LIS are not exposed to the coverage gap even when their pharmacy spending reached the coverage gap threshold.

 Enrollment in LIS is not due to self-selection, because only beneficiaries who qualify due to low income can choose LIS and most eligible beneficiaries do enroll in LIS. For this reason, this group is very different from non-LIS beneficiaries.

Our objective in this analysis was to examine differences in health outcomes and costs between coverage groups in patients with bipolar disorder and schizophrenia. We were most interested in differences between the no-gap-coverage and the generic-coverage groups, because plan self-selection could be an issue and policies governing the plans’ availability and affordability could affect health costs and outcomes for beneficiaries entering the coverage gap. For example, sicker patients could more often select the more generous generic coverage, while patients in the less generous no-gap-coverage plans might defer care more often and have worse health outcomes as a result. Beneficiaries in those plans are also likely to be more comparable to each other than to beneficiaries in LIS plans, where prominent differences in income and education level occur. In the primary analysis, we evaluated the costeffectiveness of generic coverage compared with no coverage. In a secondary analysis, we also included the LIS plan to complete the comparisons.

We obtained demographic and enrollment information, plan benefits, prescription drug events, and medical claims for Medicare beneficiaries with schizophrenia and/ or bipolar disorder (100% of non-dual eligibles and a 20% random sample of dual eligibles) who were continuously enrolled in Part D plans in 2007. We identified patients with each condition based on the existence of 2 or more inpatient or outpatient claims for schizophrenia (

International Classification of Diseases, Ninth Revision

, code 295) or bipolar disorder (codes 296.0, 296.1, 296.4-2.96.8, 301.11, 301.13) between January 1, 2007, and December 31, 2007. Because their treatment patterns might differ, we separately evaluated those who qualified for Medicare due to disability (denoted hereafter as disabled) or due to age (denoted as aged). The study design was approved by the institutional review board at the University of Pittsburgh.

Health outcomes included mortality and all-cause hospitalization rates. Cost outcomes were drug costs and total medical costs. The drug costs included Part D plan payment before rebates, beneficiary out-of-pocket spending, and subsidy amount. The total medical costs included inpatient, outpatient, and physician services as well as drug costs.

To control for differences in beneficiary characteristics between groups, all outcomes (hospitalization, mortality, and costs) were adjusted for age, sex, race, number of Elixhauser comorbidities, and the Centers for Medicare & Medicaid Services Hierarchical Condition Category (CMS-HCC) scores; for pharmaceutical spending, we used the analogous prescription drug hierarchical condition category (CMS-RxHCC) scores.

In Medicare Part D, generic drug coverage was cost saving compared with no coverage in bipolar disorder and schizophrenia while improving health outcomes.

Yuting Zhang, PhD

Articles

Using a Modified Next Generation ACO Benchmark Can Improve the MSSP

Medication Adherence and Readmission After Myocardial Infarction in the Medicare Population

Effects of Medicare Part D Coverage Gap on Medication Adherence

Cost-Effectiveness of Medicare Drug Plans in Schizophrenia and Bipolar Disorder