Seth A. Seabury, PhD

 To assess the extent to which medication adherence in congestive heart failure (CHF) and diabetes may serve as a measure of physician-level quality.

 A retrospective analysis of Medicare data from 2007 to 2009, including parts A (inpatient), B (outpatient), and D (pharmacy).

 For each disease, we assessed the correlation between medication adherence and health outcomes at the physician level. We controlled for selection bias by first regressing patient-level outcomes on a set of covariates including comorbid conditions, demographic attributes, and physician fixed effects. We then classified physicians into 3 levels of average patient medication adherence—low, medium, and high&mdash;and compared health outcomes across these groups.

 There is a clear relationship between average medication adherence and patient health outcomes as measured at the physician level. Within the diabetes sample, among physicians with high average adherence and controlling for patient characteristics, 26.3 per 1000 patients had uncontrolled diabetes compared with 45.9 per 1000 patients among physicians with low average adherence. Within the CHF sample, also controlling for patient characteristics, the average rate of CHF emergency care usage among patients seen by physicians with low average adherence was 16.3% compared with 13.5% for doctors with high average adherence.

 This study’s results establish a physician-level correlation between improved medication adherence and improved health outcomes in the Medicare population. Our findings suggest that medication adherence could be a useful measure of physician quality, at least for chronic conditions for which prescription medications are an important component of treatment.

In both the diabetes and congestive heart failure disease spaces, average levels of patient medication adherence (calculated at the physician level) are significant predictors of patient health.

Average rates of hospitalization, emergency care, and comorbidities are lower among patients treated by physicians with high average adherence rates; this correlation persists after controlling for individual patient characteristics.

The utility of average medication adherence as a potential measure of physician quality should be examined.

Measuring the quality of care offered by healthcare providers is an important tool for effectively promoting value in the US healthcare system. New quality metrics are constantly being developed and refined, and they are increasingly used in performance-based contracting. From the standpoint of providing incentives to increase the value of care, ideally these measures would directly observe the incremental value of care—in economic terms, the “marginal product of care”&mdash;and link reimbursement to the difference between incremental value and incremental cost. However, this direct effect is not easy to measure. Quality metrics sometimes attempt to do so by capturing various elements of healthcare delivery, including process, outcomes, and patient satisfaction.

More recently, quality measures have been developed to evaluate the prescribing patterns, use, and consequences of prescription medications. Some of this growth reflects the increasing availability of medications to treat and prevent chronic illnesses and the growing consensus that adherence is an important part of disease management. Study results routinely show that proper adherence can lead to lower net healthcare costs, particularly for patients with cardiovascular disease.

 Recent work also demonstrates that some health plans are systematically associated with higher medication adherence and better patient outcomes.

 This evidence coincides with the growing use of adherence measures in pay-for-performance systems such as the Medicare Advantage Star Rating System.

There is also a growing effort to increase quality monitoring for individual physicians. As part of the 2010 Affordable Care Act (ACA), CMS reports information on individual physician quality through the Physician Compare Initiative.

 More recently, the Medicare Access and CHIP Reauthorization Act of 2015 created the Quality Payment Program, through which some participating physicians receive reimbursement modifications based on reported quality metrics, including some based on adherence. However, despite the use of these metrics, it is not well established whether adherence varies systematically across physicians or correlates with the quality of medical care provided. Although the association between adherence and outcomes has been established generally, it is possible that at the individual physician level the signal to noise ratio is too weak to make it a useful reimbursement tool. More work is needed to understand if some physicians are better than others in terms of promoting medication adherence and, if so, what this means for patient outcomes.

This study compares average patient medication adherence with select outcomes-based quality measures at the physician level using Medicare claims data. We do this for all patients in a nationally representative sample of Medicare data who have diabetes or congestive heart failure (CHF), 2 expensive chronic conditions that require extensive medication management to treat. We examine whether there are systematic differences in patient adherence across physicians and how those differences are associated with improved patient outcomes. Although this analysis cannot indicate a causal relationship between physician behavior and medication adherence and patient outcomes, it can evaluate whether any association exists between systematic physician-level differences in medication adherence and patient outcomes. Establishing that such a relationship exists is a necessary first step toward validating the use of physician-level measures of medication adherence in incentive-based reimbursement schemes.

This study uses linked Medicare parts A (inpatient), B (outpatient), and D (pharmacy) claims data from a nationally representative, randomly selected 20% sample of Medicare beneficiaries 65 years and older with fee-for-service coverage of Medicare parts A, B, and D. Additionally, beneficiary enrollment, demographic information, and vital status come from the Medicare Denominator File. The inpatient data provide information on all hospital stays, including length of stay, the diagnosis-related group associated with the stay, and up to 10 individual procedure codes and diagnostic codes. The outpatient data include information on outpatient hospital visits, home hospice care, and the use of durable medical equipment. It also includes all claims submitted by physicians and other health providers, including physician office visits. The Part D prescription drug information provides information on prescription drug events, including the National Drug Code, number of days supplied, and date of service. These Part D data are linked with individual beneficiaries’ vital status from the Denominator File, which contains geographic identifiers, date of birth, date of death, gender, race, and age. Because Part D did not begin until 2006, we restrict our analysis to patients from 2007 through 2009. Although these data are older, they provide a useful perspective because they are based on a period prior to the passage of the ACA, when physician prescribing patterns were likely unaffected by the recent drive toward improved quality metrics.

We provide a brief summary of our analytical methods in this paragraph. Complete details are provided in the technical 

). We assess the relationship between adherence and several measures of patient health (including rates of hospitalization, emergency department [ED] care, and comorbidities), all calculated at the physician level. The full set of outcomes considered is listed in 

. To accomplish this, we first used claims data to generate a measure of patient-level adherence, the proportion of days covered (PDC). We next executed a series of patient-level regressions, in which 2 classes of outcomes (medication adherence and quality measures based on patient health) were modeled as a function of patient characteristics (including age, gender, and comorbid conditions) and physician fixed effects. We then predicted physician-level adherence and quality values from these regressions, calculated at the population mean of those patients. This produced measures of physician-level adherence and quality that controlled for observable differences in physician patient populations. Next, we regressed the generated physician-level quality measures on the generated physician-level measures of medication adherence. Because we employed a 2-stage model, we used bootstrap methods to estimate standard errors. Finally, we examined the relative extent to which physician fixed effects and individual patient comorbidity profiles predict patient health outcomes, using analysis of variance.

, we report differences in patient characteristics, aggregated to the physician-year level. The table shows average medication adherence for patients with diabetes and patients with CHF, patient outcomes, number of patients per physician-year, and mean total annual medical expenditures. Our measure of medication adherence is higher in the diabetes sample than the CHF sample, with a mean of 40.2% PDC (interquartile range [IQR], 30.8%-50.0%) for patients with diabetes compared with 36.2% for patients with CHF (IQR, 27.3%-44.7%). The private-sector patients considered by Seabury et al

 showed similar results in that adherence was higher for patients with diabetes than those with CHF, but the average adherence levels were much higher in the private-sector data, perhaps reflecting the sicker nature of patients in Medicare. The average medical expenditures per patient-year are high: $39,954 for patients with diabetes and $54,723 for patients with CHF. Complication rates and spending levels in this analysis are higher than those seen in many other studies because this sample is composed of patients who saw a physician at least once in the year, restricted to physicians who see a relatively large number of patients with diabetes or CHF.

The average physician-year in the data covered 52 patients in the diabetes sample and 34 patients in the CHF sample. These figures are skewed by a subset of physicians who see a large number of patients; the physician in the 90th percentile saw 103 patients with diabetes and 65 with CHF in a year.

 compares the differences in health outcomes with average adherence across physician-years in the diabetes sample. The top panel reports the unadjusted association between outcomes and adherence, whereas the bottom panel reports the association between outcomes and adherence adjusted for patient characteristics, including age, sex, and comorbidities. For each outcome, the complication rate is averaged across physicians with low, moderate, or high average medication adherence among their patients. The results show a clear relationship between the average adherence and outcomes of a physician’s patients with diabetes. In the top “unadjusted” panel, physicians whose patients have low adherence have 49.6 patients per 1000 with uncontrolled diabetes compared with 22.1 per 1000 among physicians with high medication adherence. A similar relationship, in terms of both direction and magnitude, holds for ED visits and short-term and long-term complications. The magnitude of the effect diminishes following adjustment for patient characteristics: 45.9 per 1000 patients of physicians with low average adherence have uncontrolled diabetes compared with 26.3 per 1000 patients of physicians with high average adherence. In all cases, the bootstrap tests found that groupwise differences were statistically significant (

 reports analogous results for the CHF sample. As with the diabetes sample, patients of physicians with better average adherence also experienced better average outcomes. The unadjusted CHF hospitalization rate of patients of the low-adherence physicians was 51.8% compared with 45.7% for patients of physicians with high average adherence, a difference of 6.1 percentage points or 11.8%. In terms of CHF-related ED visits, the rate was 16.4% for patients of physicians with low adherence compared with 13.5% for physicians with high average adherence, a 17.8% difference. These results are consistent in terms of direction and magnitude for hospitalizations and ED visits for the other conditions. Comparing the top and bottom panels indicates that the regression adjustment for age, gender, and comorbidities has very little effect on the magnitude of the difference across physicians with low-adherence or high-adherence patients. As in the diabetes sample, in all cases the differences are statistically significant at the 1% level (

 compares the proportion of the variance in the regression models of each medication adherence measure and outcome that is explained by physician fixed effects or patients’ own comorbidity profiles. Physician fixed effects consistently account for a much larger portion of the explained variation than do individual comorbidities. In the diabetes sample, physician fixed effects account for about 15% to 20% of the explained variation in medication adherence and 3% to 5% in outcomes. This is considerably more than patient comorbidities, which generally account for less than 1% of the explained variation in outcomes. For the CHF sample, physician fixed effects account for somewhat less of the explained variation in adherence (about 15%-18%) but more of the explained variation in outcomes (about 8%-11%). However, patient comorbidities still account for very little of the explained variation, typically 2% or less. One possible explanation is that the patient comorbidity profile is a relatively weak predictor of patient health, although some comorbidities (eg, depression) are strongly indicative of worse outcomes. In any case, the analyses from Tables 3 through 5 suggest that there are unobserved characteristics of physicians that systematically predict medication adherence and, further, that unobserved characteristics are predictive of patient outcomes. Of course, which unobserved characteristics led to better patient adherence and which led to better patient outcomes remains unknown.

Our analysis finds a strong correlation between medication adherence and patient outcomes at the physician level. This suggests that measuring the medication adherence of a physician’s patients can provide a signal of provider performance as measured by patient outcomes. More effort should be given to developing physician-level and/or disease-specific quality indications based on adherence and to evaluating whether the use of these measures could improve care and outcomes for patients with chronic diseases.

Although the relationship between improved adherence and better outcomes is well established at the patient level, it has not been established that the same relationship holds for individual physicians.

 Recent work shows that adherence can be a similarly useful measure of health plan performance.

 Prior work also shows that physicians with different measured quality also vary systematically in their prescribing patterns. For example, patients with type 2 diabetes of physicians who routinely prescribe metformin as frontline therapy have better outcomes than those of physicians who prescribe sulfonylureas.

 However, in this study we explicitly documented a correlation between improved patient outcomes and improved adherence to medication in the case of both diabetes and CHF.

Unfortunately, our data do not cast light on 

 some physicians have patients with better medication adherence. Evidence suggests that many physicians feel untrained or ill equipped to promote medication adherence.

 If that is true, it is possible that the systematic differences in medication adherence across physicians could reflect something intrinsic about the physician—patient relationship. Past work has found that good physician–patient relationships were important for promoting medication adherence.

 Other work has found that physician characteristics, such as job satisfaction and propensity to schedule follow-up visits, were associated with improved adherence.

 At the very least, the existing evidence is clear that there is an association between a physician’s engagement with patients and better medication adherence by patients. Future work should better establish the activities or qualities of physicians that promote medication adherence and use these to develop physician guidelines to help promote these more broadly. Additionally, there is likely correlation between average adherence at the physician level and average adherence at the pharmacy level; future work should investigate which of these levels is better able to influence adherence.

 Our findings suggest that medication adherence measures have the potential to be a useful quality measurement tool, at least for chronic conditions for which prescription medications are an important component of treatment. There are many potential advantages to using medication adherence as a quality metric. For one, although medication use is imperfect, there are commonly accepted and objective metrics that can be constructed using pharmacy claims data. Also, medication use is much more common than many types of complications, so it can be more accurately measured and is potentially more informative than a performance metric based on outcomes.

Although more needs to be done to establish a causal result, the cost implications could be large. Seabury et al

 demonstrated that in addition to predicting outcomes, plan-level medication adherence is a strong signal of costs. In other words, plans whose patients had better adherence also had lower total average expenditures among their patients with diabetes and CHF. Although it is unclear how much of the difference between high- and low-adherence physicians is actually due to factors within a physician’s control, our findings suggest the potential for large savings if we can sufficiently train and incentivize physicians to help their patients better adhere to their medication regimens. In the United States in 2014, there were more than 330,000 ED admissions for hyperglycemia or hypoglycemia and about 1 million for CHF.

 Prior studies have estimated the cost of these admissions at around $2000 each.

 If improved adherence reduced the incidence of these events by even a small amount, the cost savings would be large. For example, a 5% reduction in each type of admission would yield an estimated cost savings of almost $130 million. This suggests a need for careful, rigorous evaluation of new initiatives such as the Quality Payment Program to better understand what impact, if any, incentivizing physicians to improve quality has on patient outcomes.

This study has several important limitations. First, our sample was large, but it was limited to Medicare patients. Although we expect the relationship between adherence and outcomes to generalize to commercial patients, this should be demonstrated in future work. Also, the quality measures were limited by the fact that our data are claims based and lack information about disease severity. Some outcomes that we considered are relatively rare, and many physicians had no patients with these complications. Thus, although the relationship between adherence and outcomes is informative in the aggregate, patient outcomes may provide limited information about the performance of any single provider. Further studies might consider how closely average medication adherence predicts more clinically relevant measures of provider performance (eg, average glycated hemoglobin levels). Also, although we defined good adherence as having a PDC above 80%, the level chosen by CMS, it would be useful to evaluate the predictive ability of different threshold values.

Another limitation is that we only measured an association between average adherence and patient outcomes, and this relationship may not be causal. If healthier patients systematically select to certain physicians, it could overstate the correlation between adherence and outcomes due to what is known as the “healthy adherer” effect. For example, Dormuth et al

 found that patients with better adherence also display many other positive health habits, such as safer driving. Although it is not clear that patients systematically select to physicians according to unobserved health factors, if they do, it would significantly impact the interpretation of our findings. If the association that we estimated represents a true causal relationship, then we would expect that tying reimbursement to adherence could improve outcomes for patients by providing physicians with enhanced incentives for better performance. If, on the other hand, high-adherence patients self-select toward some physicians more than others, reimbursing physicians based on adherence would simply reward physicians for something they have limited impact on. This could even generate adverse incentives, if it promoted “cream skimming” and gave physicians incentives to avoid sick patients. We find some evidence to suggest that the selection across providers is relatively minor, at least according to noncardiovascular comorbid conditions. Nevertheless, this accentuates the importance of understanding the mechanisms through which physicians influence adherence, as this will help policy makers devise quality metrics and reimbursement schemes to promote the right incentives to providers and unlock value for patients.

An additional limitation is that it is difficult to precisely measure adherence to injectable medications in claims data, due to the lack of information on titration instructions. This adds some uncertainty to our findings for the diabetes population, but that uncertainty is partially offset by the parallel conclusions in the CHF population. Also, some patients are seen by more than 1 physician in a given year and thus may contribute to the adherence and outcome fixed effects for multiple physicians. We assume that no single patient could create significant correlation across physicians, based on the fact that we required a minimum of 10 patients per physician.

Lastly, we were unable to adjust for the effect of social determinants of health (SDH). Several studies have demonstrated that SDH and patient outcomes are closely related. However, the observational claim-based Medicare database provides very limited information on SDH, which may result in unmeasured confounding.

We sought to assess the potential for utilization of physician-level measures of medication adherence as quality indicators. For both CHF and diabetes and controlling for other factors, we found that the patients of physicians with higher average rates of adherence have better health outcomes than do patients of physicians with low average rates of adherence. This suggests that medication adherence may be a useful indicator of quality of care at the physician level.

 Keck School of Medicine, University of Southern California (SAS), Los Angeles, CA; Pharmaceutical Research and Manufacturers of America (PhRMA) (JSD), Washington, DC; Precision Health Economics (JS), Boston, MA.

 Dr Seabury is a consultant for Precision Health Economics, a life-sciences consulting company. Mr Sullivan is an employee of and holds stock options in Precision Medicine Group, which consults for pharmaceutical companies. The remaining author reports 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 (SAS, JSD); analysis and interpretation of data (JSD, JS); drafting of the manuscript (SAS); critical revision of the manuscript for important intellectual content (SAS, JSD, JS); and statistical analysis (SAS, JS).

 Seth A. Seabury, PhD, USC Schaeffer Center, VPD Suite 414, Los Angeles, CA 90087. Email: seabury@usc.edu.

Crossing the Quality Chasm: A New Health System for the 21st Century

. Washington, DC; National Academies Press; 2001.

2. Tamblyn R, Laprise R, Hanley JA, et al. Adverse events associated with prescription drug cost-sharing among poor and elderly persons. 

. 2001;285(4):421-429. doi: 10.1001/jama.285.4.421.

3. Heisler M, Langa KM, Eby EL, Fendrick AM, Kabeto MU, Piette JD. The health effects of restricting prescription medication use because of cost. 

. 2004;42(7):626-634. doi: 10.1097/01.mlr.0000129352.36733.cc.

4. Goldman DP, Joyce GF, Lawless G, Crown WH, Willey V. Benefit design and specialty drug use. 

. 2006;25(5):1319-1331. doi: 10.1377/hlthaff.25.5.1319.

5. Hsu J, Price M, Huang J, et al. Unintended consequences of caps on Medicare drug benefits. 

. 2006;354(22):2349-2359. doi: 10.1056/NEJMsa054436.

6. Goldman DP, Joyce GF, Zheng Y. Prescription drug cost sharing: associations with medication and medical utilization and spending and health. 

. 2007;298(1):61-69. doi: 10.1001/jama.298.1.61.

7. Thornton Snider J, Seabury S, Lopez J, McKenzie S, Goldman DP. Impact of type 2 diabetes medication cost sharing on patient outcomes and health plan costs. 

8. Seabury SA, Lakdawalla DN, Dougherty JS, Sullivan J, Goldman DP. Medication adherence and measures of health plan quality. 

9. Physician Compare. Medicare.gov website. medicare.gov/physiciancompare. Accessed January 10, 2016.

10. Ghosh A, Simon K, Sommers BD. The effect of state Medicaid expansions on prescription drug use: evidence from the Affordable Care Act [NBER working paper no. 23044]. National Bureau of Economic Research website. nber.org/papers/w23044. Published January 2017. Accessed June 1, 2017.

11. Mulcahy AW, Hollands S, Duffy EL, Strong A, Wynn BO. Modeling the economic impact of a California workers’ compensation formulary. RAND Corporation website. rand.org/pubs/research_reports/RR1971.html. Published 2017. Accessed June 1, 2017.

12. Bonakdar Tehrani A, Cunningham PJ. Closing the Medicare doughnut hole: changes in prescription drug utilization and out-of-pocket spending among Medicare beneficiaries with Part D coverage after the Affordable Care Act. 

. 2017;55(1):43-49. doi: 10.1097/MLR.0000000000000613.

13. Bognar K, Bell KF, Lakdawalla D, et al. Clinical outcomes associated with rates of sulfonylurea use among physicians.

. 2013;19(spec no. 5):SP182-SP186. ajmc.com/journals/evidence-based-diabetes-management/2013/2013-1-vol19-sp5/clinical-outcomes-associated-with-rates-of-sulfonylurea-use-among-physicians. Accessed January 28, 2019.

Articles

Physician-level medication adherence is a strong predictor of patient health and should be considered as a measure of physician quality.

Medication Adherence as a Measure of the Quality of Care Provided by Physicians

State Medicaid programs are under increasing pressure to contain pharmaceutical spending. Many states have attempted to limit spending through greater Medicaid managed care penetration, which rose nationally from 54.5% in 1999 to 74.9% in 2011. It is not clear how this expansion has affected beneficiaries with serious mental illness (SMI)—a vulnerable population that often has their drug spending “carved out” from their managed care benefit. We sought to assess the association between managed care penetration and pharmaceutical spending on drugs for SMIs in these states.

State-year observations were constructed to study the relationship between managed care penetration and pharmaceutical spending on drugs for SMIs over the period 1999 to 2011. We analyzed the relationship using both cross-sectional and panel-data methods.

Our cross-sectional analyses suggested that carve-out states with greater managed care penetration spend significantly less per enrollee on pharmaceuticals for the treatment of mental disorders: our panel data analyses did not generate statistically meaningful results.

Future studies should address whether any effects of managed care on mental health prescription utilization and spending reflect improved care coordination or worsening access to valuable care for the population with SMI.

	Our findings suggest that states with higher Medicaid managed care penetration have lower spending on serious mental illness (SMI) drugs. We cannot conclude that the relationship is causal, but the finding does raise a number of interesting issues for health policy. 

		If managed care is associated with less use of ineffective drugs, then this lower utilization can be viewed as cost-effective, with state Medicaid programs achieving similar outcomes at a lower cost. 

		However, if managed care is also associated with less use of effective drugs, then this decreased utilization may not be cost-effective due to the increased likelihood of negative health outcomes, such emergency department visits, hospitalizations, homelessness, incarcerations, and death. 

		Most states carve out some or all behavioral health services from managed care benefits. A potential downside of such arrangements is fragmentation of care, particularly for vulnerable beneficiaries with SMI, who are likely consumers of other medical services.

Medicaid is the nation’s largest source of financing for behavioral health services.

 It serves many beneficiaries with serious mental illness (SMI), including schizophrenia, bipolar disorder, and major depressive disorder.

 Although many dually eligible beneficiaries with SMI receive their drug coverage through Medicare under the Part D program, drugs treating SMI still account for almost 20% of Medicaid pharmacy spending.

 In 2009, approximately 5 million individuals (11.9% of all Medicaid beneficiaries) received antipsychotics or antidepressant medications through Medicaid.

One of the major provisions of the Affordable Care Act (ACA) is a sweeping increase in Medicaid eligibility for those states choosing to opt in. The Congressional Budget Office estimates that Medicaid expansion will increase the number of nonelderly enrollees in Medicaid and the Children’s Health Insurance Program (CHIP) by 10 million in 2016.

 The costs of this expansion initially will be fully financed by the federal government and, by 2020, the federal share of Medicaid expansion funding will decline to 90%, which will require states to pay for the balance of new enrollees’ costs.

 Increased enrollment in Medicaid, coupled with states’ growing obligation to finance the program’s costs, will likely impose new pressures on states to curb Medicaid spending—for both newly and previously eligible populations. 

Historically, states have attempted to control Medicaid fee-for-service (FFS) prescription drug spending with tools such as formulary restrictions, step therapies, and prior authorization; however, these have sometimes produced unintended consequences in the FFS setting. In the case of drug treatments for SMI, prior research has demonstrated that the use of formulary restrictions may have significant adverse effects for patients—especially those who may benefit from atypical antipsychotics or other novel therapies.

 Furthermore, in many cases, these poor outcomes were significant enough to eliminate any potential cost savings from the policies.

 Similarly, the use of prior authorization for atypical antipsychotics by the Maine Medicaid program was found to increase medication discontinuations without lowering spending.

Increasingly, states have turned to contracts with Medicaid managed care plans in order to better control costs and reduce budgetary uncertainty. However, in many states, prescription drug spending is “carved out” (ie, not included) in the managed care benefit. Under a carve-out arrangement, prescription drug benefits are managed on an FFS basis, which excludes them from the set of services for which a managed care plan has oversight and direct financial liability. Conceptually, this suggests that carve-out reduces both the ability of and incentives for managed care plans to coordinate pharmaceutical use with spending on other health services, potentially leading to “cost spillovers” elsewhere in the system.

This issue is particularly relevant in the treatment of SMI, where poor adherence and discontinuation may have serious consequences, including emergency department (ED) visits, inpatient hospitalization, homelessness, incarceration, or even suicide.

 Indeed, a simple change in drug coverage may result in unintended consequences. For example, following the implementation of Medicare Part D, many dually eligible beneficiaries had difficulty accessing psychiatric medications, leading to increased psychiatric ED visits.

 On the other hand, in light of growing evidence that prescription drug adherence might result in lower spending on services that are carved into most Medicaid managed care programs (eg, nondrug inpatient and urgent care), these plans might still have an incentive to promote pharmaceutical adherence among carved-out beneficiaries with SMI.

New research is needed to understand the implications of Medicaid managed care for the treatment of SMI. In this study, we examined the relationship between the expansion of managed care in Medicaid programs and Medicaid spending on pharmaceuticals for the SMI population in carve-out states.

We collected historical Medicaid managed care penetration rates, SMI prescription utilization data, and additional state-level information for all states and the District of Columbia from 1999 to 2011 (inclusive). These data were used to analyze whether states’ levels of Medicaid managed care penetration were associated with their level of prescription use for the SMI population.

Medicaid managed care penetration rates were derived from annual enrollment reports published by CMS and its predecessor, the Health Care Financing Administration.

 These reports document enrollment in both comprehensive and limited benefit managed care plans and include the proportion of each state’s Medicaid beneficiaries enrolled in managed care plans (ie, the penetration rate). Several states administered Medicaid programs on an exclusively FFS basis: New Hampshire in 2004, 2010, and 2011; Mississippi in 2002, 2003, and 2007; and Alaska and Wyoming from 1999 to 2011. We retained these observations in our analyses, coding these states as having no Medicaid managed care penetration in the applicable years.

Drug utilization data were collected from CMS annual Medicaid State Drug Utilization Data files. Each file contains information for covered outpatient drugs that were paid for by a state Medicaid agency. Specifically, the files include information on the number of prescriptions and the amount of reimbursement for those prescriptions by National Drug Code (NDC). Notably, the files do not include prescriptions paid by Medicare Part D plans on behalf of dual eligible beneficiaries. The CMS database of State Drug Utilization files is relatively complete, missing files for only 6 of the 323 (1.9%) state-year observations in which states were eligible to report prescription drug spending data (ie, states that carved out the managed care prescription drug benefit, or for which the managed care penetration rate was 0, in a particular year).

Because the CMS annual Medicaid State Drug Utilization Data only record prescriptions that were paid for via FFS (either due to a carve-out of the drug benefit or through a traditional FFS model), we used the Medicaid Analytic Extract (MAX) Prescription Drug Tables, published by CMS for the years 1999 to 2011, to identify each state’s carve-out status by year.

 We treated a state as operating under a carve-out arrangement in a given year if it carved out the prescription drug benefit from all of its managed care plans—or prepaid health plans&mdash;for at least three-fourths of that year. The MAX Prescription Drug Tables did not provide information about carve-out arrangements for 2000 and 2010, so we imputed data for these years by conservatively coding a state as carving out the drug benefit only if it did so in both adjacent years (ie, 1999 and 2001 for the missing 2000 data, and 2009 and 2011 for the missing 2010 data). States whose carve-out status could not be determined (Louisiana and Vermont) were dropped from our analyses, as were the 2 states (Maryland and Minnesota) that did not carve out the prescription drug benefit from their Medicaid managed care plans in any year of our study period.

Other state-year—level data were derived from 2 main sources. First, various resident sociodemographic characteristics and insurance coverage rates were calculated from the Annual Social and Economic Supplement to the Current Population Survey (CPS), a large national survey of households by the Census Bureau.

 Second, information on states’ healthcare provider counts were collected from the Area Health Resource File (AHRF) Access System, a compendium of geocoded data compiled by the Health Resources and Services Administration. Rates of providers per capita were calculated using AHRF state provider totals and CPS state population estimates. Because provider counts are not available for every year, missing data were imputed by modeling state-specific quadratic time trends.

Our primary outcomes were annual rates of SMI prescription use and spending for the Medicaid SMI population in states and years in which the Medicaid prescription drug benefit was carved out of managed care or that ran a traditional FFS Medicaid program. We designated drugs as SMI prescriptions on the basis of their pharmacologic class (eg, atypical antipsychotics)

 or specific SMI indications. In order to check the robustness of our results, we chose to construct 2 alternate definitions of SMI prescriptions: a narrow definition including only antipsychotic medications and a broader definition including additional drugs with primary psychiatric indications, like antidepressants. Although all our psychiatric prescriptions consist of psychiatric drugs without major nonpsychiatric uses, these 2 measures allow us to isolate our findings for antipsychotics relative to the broader class of SMI drugs. A full list of classes included in each definition is provided in the 

The construction of the outcome variables involved the following steps. For each state and year, we tabulated the total annual Medicaid spending and number of prescriptions for each NDC falling within our SMI definitions. We excluded a small number of drug utilization outliers (0.1% of annual drug observations) because their prescription totals were larger than 100 times the state median for that drug. We then calculated the total state SMI prescription utilization and spending among all drugs within our SMI definitions. We further excluded 7 state-year observations that exhibited very high SMI prescription spending relative to the state trend: Washington (2006), South Dakota (2007), and Tennessee (2003-2007). After applying this restriction, we obtained a final analytic sample of 310 observations from 46 managed care carve-out and Medicaid FFS states that had valid SMI prescription spending and utilization data (

). Total spending and utilization amounts were divided by the size of each state’s Medicaid SMI population, which was estimated using the annual count of Medicaid beneficiaries in the CPS and 2011-2012 survey estimates of the rates of adult SMI by state.

We employed 2 alternate modeling approaches to assess the association between managed care penetration and SMI prescription outcomes. The first, a pooled cross-sectional approach, tested for whether the levels of a state’s prescription outcomes in a given year were associated with the level of its managed care penetration rate, controlling for time-varying state characteristics and shared time trends in prescription outcomes across all states. This analysis consisted of linear regression models of state prescription outcomes on the state Medicaid managed care penetration rate (the variable of interest), year indicators, and state characteristics.

The second modeling approach, a panel data method, tested for whether changes in a state’s managed care penetration rate over time were associated with changes in prescription outcomes. These regressions included 2 additional sets of terms: state indicators and interactions between these indicators and a linear time trend (ie, state-specific time trends). These additional terms control for unobserved but fixed state characteristics that contribute to the level and linear growth rate of the outcome variable. In other words, one can interpret the panel models as controlling for states’ baseline differences in outcomes as well as their outcomes’ linear trend. All models were run on state-year observations for which a prescription drug carve-out arrangement was in effect or where a state only operated a traditional FFS Medicaid program.

The regression models also included controls for the following time-varying state covariates: the proportion of residents by age group (aged 0 to 25, 26 to 50, and 51 to 75 years), the proportion of white residents, the labor force participation rate, the unemployment rate, the proportion of residents below 100% of the federal poverty level, the proportion of residents who are uninsured, Medicare beneficiaries, Medicaid beneficiaries, dual eligibles, the number of active nonfederal medical doctors per resident, the number of psychiatrists per resident, and the number of short-term general hospital beds per resident.

All models also included an additional term to account for the effect of Medicare Part D on Medicaid drug claims. Although year indicators account for the average effect of the policy, we also included a term allowing that effect to vary according to the size of a state’s dual eligible population. This term is an interaction between the proportions of Medicaid beneficiaries who are also Medicare beneficiaries, calculated from the CPS, and an indicator for years 2006 and later, when dual eligibles no longer received Medicaid pharmacy benefits for drugs covered by Part D.

All regression analyses were weighted by state population and standard errors were clustered at the state level.

Our sample consisted of 310 state-year observations for prescription drug utilization and spending from 46 states that provided valid spending and utilization data, and that either carved out the prescription drug benefit for Medicaid managed care enrollees or operated FFS Medicaid programs (Figure 1). Across these states and years, the average Medicaid managed care penetration rate was 63.2%. This varied substantially across states: for example, Alaska and Wyoming had no managed care penetration, whereas all of Tennessee’s Medicaid beneficiaries were enrolled in managed care (

Over the study period, Medicaid managed care penetration increased from 54.5% nationally in 1999 to 74.9% nationally in 2011 (

). Considerable heterogeneity was found in managed care growth across states. For example, managed care penetration increased by 56 percentage points in Illinois and by 51 percentage points in Texas, from 1999 to 2011. In Mississippi and New Hampshire, managed care penetration varied from 0% to over 70% when these states alternately expanded and reduced the use of managed care contracts in their Medicaid programs. Managed care penetration was more stable in states such as Massachusetts, Wisconsin, and Iowa over our study period.

Medicaid SMI prescription spending also varied substantially across carve-out states (

). When using the broad definition of SMI drugs, annual prescription spending averaged $3730 per beneficiary with SMI, ranging from an average of $111 in New Mexico to $7701 in New Hampshire, from our sample of 310 state-year observations. The corresponding average for our narrow definition of drugs for SMI was $2344, ranging from $68 in New Mexico to $4938 in New Hampshire. Annual counts of prescriptions per beneficiary averaged 25.3 and 8.8 for our broad and narrow definitions of SMI prescriptions, respectively. Summary statistics for additional state characteristics are present in the eAppendix.

We found cross-sectional, negative associations between the managed care penetration rate and each measure of SMI prescription utilization, although not all estimates were statistically significant (

). In particular, a 10 percentage point increase in Medicaid beneficiaries enrolled in managed care was associated with 0.87 fewer SMI prescriptions per beneficiary with SMI (using the broad definition of SMI prescriptions) and with 0.34 fewer prescriptions (using the narrow definition of prescriptions). Similarly, a 10 percentage point increase in managed care penetration was associated with $103 lower SMI drug spending per beneficiary using the broad definition. These estimates grew closer to 0, and were no longer statistically significant, when panel data models were employed.

Managed care has become the dominant form of care delivery in Medicaid, and it continues to grow in use in the current era of Medicaid expansion. Our findings suggest that even when prescriptions are carved out of managed care, states with higher Medicaid managed care penetration have lower spending on SMI drugs. We cannot conclude that the relationship is causal, but the finding does raise a number of interesting issues for health policy.

A key question is why we observed different findings across the cross-sectional and panel data models. Each analysis has some advantages and disadvantages: the cross-sectional analysis generates more precise estimates (ie, smaller standard errors), but the results cannot be interpreted as causal because states with more managed care penetration may differ from other states in ways we are unable to model. The panel data estimates have the advantage of allowing for more valid causal inference, but the estimates are imprecise. Although there was a substantial expansion of Medicaid managed care across the nation during our study period, not all states experienced large changes in managed care penetration over time.

Some states, like New Hampshire and Mississippi, did have wide variation in managed care penetration over time, whereas others had little or no changes in managed care penetration over our study period, which effectively reduced the number of states contributing to the panel data analysis. Moreover, our study period was only from 1999 to 2011, which did not allow us to harness some states’ early introduction of managed care programs—covering more rapid ramp-ups of managed care penetration&mdash;for statistical identification. Ultimately, the large standard errors for our panel data results suggest that increasing Medicaid managed care penetration was not strongly associated with changes in SMI use and spending. Thus, the 2 sets of findings are consistent with one another.

A key question for policy makers will be identifying the underlying mechanism by which managed care is associated with lower utilization. Many of the levers used by managed care to control utilization—prior authorization, limits on medication number/dosing, step therapy, and utilization review&mdash;are also used in Medicaid FFS programs. Although restrictive provider networks might play a role in limiting utilization, this mechanism is limited by federal requirements regarding network adequacy. Our results cannot be explained by favorable selection of healthier patients by managed care organizations (MCOs), because we measured state-level use of SMI drugs for both FFS and managed care enrollees.

From a policy perspective, the implications of lower Medicaid drug utilization and spending for SMI beneficiaries are unclear without knowing the corresponding impact on healthcare outcomes. On the one hand, if managed care is associated with less use of ineffective drugs, then this lower utilization can be viewed as cost-effective, with state Medicaid programs achieving similar outcomes at a lower cost. However, if managed care is associated with less use of effective drugs, then this decreased utilization may not be cost-effective due to the increased likelihood of negative health outcomes, such as ED visits, hospitalizations, homelessness, incarcerations, and death.

A large amount of the health economics literature has worried about cost spillovers across different services; that is, reducing costs in one service area may be like squeezing a balloon, causing costs to rise in other areas.

 In the case of Medicaid coverage of SMI drugs, the concern is that lower Medicaid spending on drugs will be offset by higher Medicaid spending for other services, such as hospitalizations. The incentive for payers to avoid cost-spillovers depends on the financing structure, which can involve carve-outs for mental health, pharmaceuticals, or both. Most states carve out some or all behavioral health services and pharmaceuticals from managed care benefits and finance them through separate reimbursement contracts.

In addition to the potential for increased cost shifting across services that are carved in and out, another potential downside of such arrangements is fragmentation of care, particularly for vulnerable beneficiaries with SMI who are likely consumers of other medical services. Some evidence suggests that the behavioral health carve-outs have restricted necessary care for patients with SMI.

This study examines the relationship between Medicaid managed care penetration within a state and spending on pharmaceuticals for patients with serious mental illnesses.

Medicaid Managed Care Penetration and Drug Utilization for Patients With Serious Mental Illness

 Medication adherence is increasingly being considered as a measure for performance-based reimbursement contracts in healthcare systems. However, the association between health outcomes and adherence at the plan level is unknown.

 Retrospective analysis of medical and pharmacy claims from a large private sector claims database from 2000 to 2009.

 We compared plan-level measures of medication adherence and health outcomes for patients with diabetes and congestive heart failure (CHF). Plan performance was based on average rates of disease complications. Medication adherence was calculated as the percent of patients having 80% of days covered for medications treating diabetes or CHF. Both adherence and outcomes were adjusted for patient differences using multivariate regression. Plans were stratified into low, moderate, and high adherence, based on adherence in the bottom quartile, middle 2 quartiles, and top quartile, respectively.

 Average adherence varied significantly across plans. Plans with low adherence to diabetes medications had adjusted rates of uncontrolled diabetes admissions of 13.2 per 1000 patients, compared with 11.2 in moderate adherence plans and 8.3 in high adherence plans (

 <.001). The adjusted rate of CHF-related hospitalization was 15.3% in low adherence plans, compared with 12.4% in moderate adherence plans and 12.2% in high adherence plans (

<.001). These patterns were consistent across different types of complications for both diabetes and CHF.

 Private health plans vary considerably in average adherence to medications treating chronic diseases. Plans with higher average adherence had lower rates of disease complications, suggesting that medication adherence measures are potentially useful tools for improving the performance of health plans.

This study examines the association between plan-level measures of health outcomes and medication adherence to assess the viability of adherence as a measure of plan performance.

We found that plan-level averages of medication adherence were associated with significantly lower rates of disease-related complications for both diabetes and congestive heart failure.

These findings suggest that medication adherence measures are potentially useful tools for improving the performance of health plans.

More needs to be done to understand how and why providers and plans can improve medication adherence.

Measuring the quality of care provided to patients by health plans and providers is an increasingly common element of evaluating and improving healthcare delivery. Efforts to increase value in healthcare often include incorporating measures of quality into provider compensation or plan ratings.

 This is reshaping how health plans evaluate themselves, how employers choose plans, and how consumers decide who provides their care. At present, many healthcare organizations participate in quality reporting, whether for internal quality improvement or for public reporting. New payment models are emerging that directly incorporate these quality data into coverage and reimbursement decisions.

Many quality measures have been developed to evaluate the appropriateness of the prescribing, monitoring, and use of prescription medicines in the treatment of both acute and chronic conditions. Medication adherence is a natural candidate for quality measurement, as individuals with higher adherence consistently demonstrate better clinical outcomes as well as lower use of and spending on other medical services.

 For example, a recent study found that poor adherence can lead to higher costs of up to $702 per month for Medicare beneficiaries with diabetes and up to $840 for those with heart failure.

 This persistent empirical relationship was reflected recently in the decision by the Congressional Budget Office to incorporate a reduction in spending on medical services to reflect the health benefits associated with medication use among Medicare beneficiaries.

However, there is little understanding of the ways in which explicitly measuring and tracking medication adherence and outcomes at the plan level will incentivize health plans to implement measures that improve upon those process and outcome metrics. In particular, it is not known whether plans with higher average adherence exhibit better health outcomes or lower health spending. There is considerable research demonstrating that medication adherence is associated with improved clinical outcomes and reduced spending for individuals,

 although this does not necessarily mean that average medication adherence rates can provide useful signals of health plan performance. There are many drivers of patient outcomes and plan performance, and it is possible that medication adherence does not vary enough across plans, or that the patient populations differ too much, to isolate a clear relationship between plan-level adherence and outcomes. To better understand the implications and possible value of using medication adherence as a measure of health plan performance, we need to understand the link between adherence and outcomes at the plan level after controlling for patient differences across plans.

We used a large database of claims from private sector health plans to compute commonly used measures of plan performance on outcomes and plan-level rates of medication adherence for patients with diabetes and congestive heart failure (CHF). We selected these 2 conditions because they are relatively common and lead to significant social costs, and also because there is ample evidence supporting a relationship between adherence to guideline-recommended medications and improved clinical outcomes for these patients on an individual level.

 We tested the relationship between plan-level quality measures of medication adherence and health outcomes to assess whether better adherence was associated with lower disease complication rates and spending.

We estimated the relationship between plan-level adherence and outcomes according to a 2-step process. We first computed average rates of adherence and health complications related to diabetes and CHF at the plan level. We computed unadjusted averages, and we also used multivariate regression to adjust for patient characteristics and to predict the adherence and complication rates that would result if every plan had patients with the same observable characteristics. In the second stage, we estimated the association between adherence and disease complication rates in order to test whether higher adherence was associated with better outcomes. Here we describe these procedures and the data used in greater detail.

We used a database of medical and pharmacy claims data for individuals enrolled in health plans of large private employers from 2000 to 2009. These data have been used extensively to study medication adherence and patient outcomes in past work.

 The database contained medical claims or encounter data from all available healthcare sites, including both inpatient and outpatient care. The billing data were broken down by payer (eg, out-of-pocket or health plan) and provided information on up to 4 associated diagnosis codes and 1 associated procedure code. The database also included pharmacy claims for all outpatient pharmaceutical purchases; each claim included unique drug identifiers, the number of days of medication supplied, and total payments by all payers. Enrollment records provided information on basic demographics, including age, gender, and 3-digit zip code of residence. These claims data were rolled up to a patient level to identify individual diagnoses, characteristics, adherence, and hospitalizations, and then aggregated at the plan level for analysis.

We restricted the sample to individuals aged 18 to 64 years who were diagnosed with either diabetes or CHF. We excluded individuals 65 years or older to focus on private health plans and to exclude patients covered by Medicare. Patients with diabetes were identified based on an 

International Classification of Diseases, Ninth Revision, Clinical Modification (ICD-9-CM) 

diagnosis code of 250.xx listed as the primary or secondary diagnosis in 1 inpatient claim, or at least 2 outpatient claims within a 30-day window over 1 calendar year. We required at least 2 outpatient codes to eliminate cases where the 

code was listed on a claim as a possible diagnosis but was later ruled out. CHF was identified the same way, using 

Using these inclusion criteria, we created separate cohorts of individuals with diabetes and with CHF. Any individual diagnosed with both diseases in the same year was included separately in both samples. Once identified as being in either or both samples, individuals were grouped into health plans according to their enrollment records. Individuals were included from the year of initial observed diagnosis for as long as they remained enrolled. We required that individuals be continuously enrolled over a full calendar year to ensure that we observed all their utilization throughout the year. When aggregating the information to the plan level, we also required that plans have at least 100 patients to be included in the analysis.

Many quality measures have been developed to evaluate plan performance, and a full accounting of the relationship between all quality measures tracking medication adherence and health outcomes is beyond the scope of this study. We evaluated a set of quality metrics specific to medication adherence and health outcomes that were endorsed by the National Quality Forum (NQF), relevant for diabetes or CHF, and observable in retrospective claims analysis.

 The NQF is a nonprofit organization whose mission is to build consensus around approved quality measures for health plan performance. The NQF measures are used by CMS in quality reporting programs for physicians, plans, and hospitals.

 We measured medication adherence using the percentage of days covered (PDC) over a 1-year period. PDC is an NQF-endorsed quality measure for adherence that is commonly used by researchers to evaluate patient adherence.

 Therapeutic classes were selected based on treatment guidelines for diabetes and CHF. Diabetes medications included beta-blockers, angiotensin-converting enzyme (ACE) inhibitors, angiotensin receptor blockers (ARBs), calcium channel blockers, oral diabetes medications, and statins. For CHF, we included beta-blockers, ACE inhibitors, ARBs, and diuretics. Within each health plan, we identified all patients who were prescribed each type of medication at least once and computed the PDC for the calendar year after the date of initiation. Patients who met the PDC threshold of 80% or greater over the year were considered to have “good” adherence. The 80% PDC threshold is often used as an NQF-endorsed measure of quality; for example, CMS uses it for certain drug classes in its performance measurement system for Medicare Part D plans.

 We measured patient outcomes using quality measures endorsed by NQF that reflect disease complications generally considered preventable for the 2 relevant diseases. For diabetes, this includes the rate of inpatient admissions for uncontrolled diabetes, emergent care for hypoglycemia or hyperglycemia, rates of short-term diabetes complications, and rates of long-term diabetes complications. Given that some outcomes are relatively uncommon, we reported them in terms of rates per 1000 patients. CHF outcome measures included hospitalizations or emergent care visits for CHF or other comorbid conditions, including diabetes, coronary artery disease, or hypertension. As these were comparatively more common than the diabetes complications, they were computed in terms of percent of patients per year. We also computed the average annual total medical expenditures for patients in each sample, normalized to 2009 dollars using the Consumer Price Index.

The goal of our analysis was to describe how variation in medication adherence across plans correlated with variation in plan-level patient outcomes. However, this correlation could be confounded by the presence of differences across plans; for example, plans with sicker patients will on average tend to have patients using more medications and more medical services, regardless of their adherence patterns. To address for potential bias driven by systematic differences in patients across health plans, we adjusted for patient characteristics using multivariate regression. We then aggregated adjusted adherence and outcome values up to the plan level to construct our key analytic plan-level quality measures of disease complications, spending, and adherence. In the 

), we expand on this and provide a more detailed discussion of our empirical strategy.

We estimated separate regression models for each medication class and each health outcome. All regressions included controls for patient age, gender, and number of comorbidities, as well as fixed effects for state of residence and year. Note that while our data identify plans consistently within a year, the coding of plans is not fully consistent within an employer over time. Thus, a plan in our analysis really refers to a plan-year, which is our primary unit of analysis, and we were unable to use plan-fixed effects. In the eAppendix, we discuss the results of a sensitivity analysis in which we included employer fixed effects and found that our results were largely unchanged. Comorbidities were identified from the medical claims and included ulcer, depression, allergic rhinitis, migraine, osteoarthritis, chronic sinusitis, anxiety or tension disorder, epilepsy, gastric acid disorder, glaucoma, irritable bowel syndrome, malignancies, psychotic illness, thyroid disorder, rheumatoid arthritis, tuberculosis, HIV, anemia, and chronic obstructive pulmonary disease. Other common conditions that were related to cardiovascular health (eg, acute myocardial infarction or hypertension) were excluded because the onset of claims could be endogenously related to individuals’ medication use or other healthcare related to their diabetes or CHF.

We calculated adjusted plan-level adherence and health outcomes as follows. For medication adherence, we regressed the percent of individuals with PDC greater than 80% on these covariates using a linear probability model to obtain predicted PDC values for each individual. We then generated predicted values for adjusted adherence holding other factors fixed at their mean values, so that the remaining variation would come from fixed differences across plan-years and not from differences in observable patient characteristics. We replicated this procedure for health outcomes and spending, leaving us with a set of adjusted, plan-level averages for medication adherence and outcomes. The same process was conducted separately for CHF and diabetes.

This regression adjustment eliminates observed heterogeneity in the patient population across plans but does not necessarily address unobserved heterogeneity across plans that could be correlated with both medication adherence and health outcomes. For example, significant changes in the plan population in a given year (if, say, the plan benefits changed in such a way as to attract sicker patients with poorer adherence) could drive a correlation between adherence and outcomes. Thus, the association between plan-level adherence and outcomes are not necessarily causal, and the findings should be interpreted accordingly.

We computed the average level of medication adherence across plans separately for beta-blockers, ACE inhibitors, ARBs, calcium channel blockers, oral diabetes medications, and statins in the diabetes sample, and beta-blockers, ACE inhibitors, ARBs, and diuretics in the CHF sample. We then combined the PDC measures into a single index of medication adherence for each condition. Based on the value of the index, we grouped health plans into categories of high, moderate, and low levels of adherence.

Specifically, to construct the index we applied the following procedure separately in both disease cohorts. First, for each health plan we computed the average share of patients with at least an 80% PDC separately for each medication class. Then, for each plan, we calculated the weighted average of adherence across drugs for each plan-year, where adherence was weighted by the number of patients in the plan taking each drug. Using this weighted average, we identified the upper and lower quartile values of adherence across plans (ie, the values of adherence that 25% of plans were above and below, respectively). Finally, we stratified plans into 3 categories of adherence—high, moderate, or low&mdash;based on whether their weighted average was above the 75th percentile, between the 75th and 25th percentile, or below the 25th percentile across plans, respectively. We focused our analysis on the top and bottom quartiles of plans because we wanted to emphasize the impact of plans that did particularly well or particularly poorly&mdash;the “outlier” plans.

We used the same process to categorize plans into different adherence categories using both the adjusted and unadjusted values of the PDC, and in both disease cohorts. For our main analysis, we regressed the adjusted health outcomes against the medication adherence variables as the independent variables using ordinary least squares regression. Note that in a 2-step estimation procedure, where we use regression to adjust for health plan characteristics in the first step and regress adjusted outcomes against adjusted adherence in the second step, the variance estimates are biased in the second step.

 To statistically compare differences across the low, moderate, and high adherence plans, we used a bootstrap procedure. Specifically, we resampled individuals with replacement and computed 

 values for differences across low, moderate, and high adherence plans using bootstrap variance estimates based on 200 draws.

 presents the share of individuals within plans who show evidence of good adherence, as identified by a PDC of 80% or better. Only adherence to medications used in plan quality measures (as specified above) are computed for each sample; for example, we do not report adherence to antidiabetes medications in the CHF sample, even though CHF patients with diabetes may take them. For patients with diabetes, 54% to 56% of patients within a plan exhibit good adherence to therapy, on average, except in the case of statin medications, for which just 45.2% exhibit good adherence. Adherence to ACE inhibitors and ARBs is generally lower for CHF patients within a plan, while adherence to diuretics is particularly low, at 37%. Individuals taking diuretics have been shown in past research to exhibit lower adherence levels than other agents for treating cardiovascular disease.

 As a result, the weighted average of plan-level adherence across therapies is significantly lower for the CHF sample (46.8%, compared with 52.9% for diabetes).

. The table reports the mean values of adherence and other characteristics, including health outcomes, plan size, and average total medical expenditures, for different plan years. The table shows that medication adherence is slightly higher in the diabetes sample (approximately 53% compared with 47%). Also, as we would expect, there are many more diabetes patients per plan, while patients with CHF are sicker on average, demonstrated by their higher rates of complications and higher total healthcare expenditures. Note that because we required at least 100 patients in a plan to be included in the analysis, the number of plans included is much smaller in the CHF sample than the diabetes sample (195 vs 919).

The thresholds below and above which plans are defined as having low or moderate adherence, respectively, were defined using the 25th and 75th percentiles of the adherence across plans. We classified plans with adherence below the 25th percentile as having low adherence, those above the 75th percentile as having high adherence, and those with adherence between these levels as having moderate adherence. We found that there was substantial variability in average medication adherence across plans. In the diabetes sample, the threshold for low plan-level adherence is 48.3%, meaning that less than half of plan enrollees with diabetes are considered to have good adherence. The threshold for high plan-level adherence, the 75th percentile, is 58.6%. The significant differences in adherence across plans are important, because they suggest that there is heterogeneity in adherence across plans that can potentially be targeted for intervention by policy makers. For CHF patients, adherence ranges from 43.5% at the 25th percentile to 50.9% at the 75th percentile. In the eAppendix, we provide some additional detail on the sample and the variation in key variables across plans.

we compare the variation in plan-level health outcomes to plan-level adherence in the diabetes sample. Specifically, we report the average health outcomes for patients in the low, moderate, and high adherence plans. The top panel reports the association between unadjusted outcomes and unadjusted adherence, while the bottom panel reports the association between adjusted outcomes and adjusted adherence. 

 values are reported in parentheses reflecting statistically significant differences between the moderate and high adherence plans and the poor adherence plans.

There is a positive association between poor adherence and adverse outcomes for diabetes patients. In the top “unadjusted” panel, low adherence plans exhibit significantly more uncontrolled diabetes, emergent care for glycemic events, and long-term complications, compared with the high adherence plans. The patterns for diabetes complications are less clear, but adjusting for differences in patient characteristics across plans clarifies the associations.

After adjusting for the characteristics of individual patients, patients in low adherence plans exhibit statistically worse outcomes compared either with those in high adherence plans, or with those in moderate adherence plans. Significance is achieved at the 1% level for all the high adherence comparisons, and at the 5% or 10% level for the moderate adherence comparisons.

 reports the associations between plan adherence measures and outcome measures for CHF patients. As with the previous table, the top panel reports unadjusted values and the bottom panel reports adjusted outcomes, with bootstrap 

For the unadjusted values, low adherence plans have statistically higher (

<.001) rates of hospitalizations and emergency department (ED) visits for CHF itself and all the co-morbidities considered in quality measurement, compared with high adherence plans. The same is true when comparing low adherence plans to moderate adherence plans, where the highest 

 value is .01. These findings are largely unchanged by the regression adjustment, as shown in the bottom panel. Low adherence plans perform significantly worse than those with moderate or high adherence. For example, rates of CHF hospitalization are 3.1 percentage points higher in low versus high adherence plans, a 25% difference. Rates of CHF ED visits for low adherence plans are 3.9 percentage points higher, nearly double those for plans with moderate or high adherence. The adjusted comparison of adherence and coronary artery disease hospitalizations are not significantly different for high adherence versus low adherence plans.

In addition to directly assessing the relationship between quality measures for adherence and health outcomes, we also explored the implications for healthcare spending on patients with diabetes and CHF. Accordingly, we compared adjusted average annual non-drug medical expenditures (defined as the sum of inpatient and outpatient expenditures) by diabetes and CHF patients according to plan adherence (

). Results demonstrated that better adherence is associated with substantially lower average spending per year among both the diabetes and CHF samples. For example, average annual expenditures for CHF patients in plans that had low average adherence were approximately $31,500, compared with $20,407 for patients in high adherence plans (

<.001). Similarly, annual expenditures for patients with diabetes in low adherence plans were $8784, compared with $6766 in high adherence plans (

We used the relative differences in predicted expenditure values to simulate potential cost savings associated with improving performance among plans with low adherence. Our estimates suggest that moving all low adherence plans to the moderate category would, on average, reduce aggregate spending among patients with diabetes by 1.6% and among patients with CHF by 6.3%. Similarly, moving all the low and moderate adherence plans to high adherence would reduce spending among diabetes patients by 14.1% and among CHF patients by 13.7%. Note that because there is overlap between the diabetes and CHF samples, some of the improvement in diabetes patients could be due to improved treatment of CHF and vice versa. Thus, these differences in spending are not additive in terms of the implications for total spending at the plan level.

Medication Adherence and Measures of Health Plan Quality

 To measure the impact of state Medicaid formulary policies on costs for patients with schizophrenia and bipolar disorder.

 Retrospective analysis of medical and pharmacy claims for patients diagnosed with schizophrenia or bipolar disorder in 24 state Medicaid programs.

We combined information on formulary restrictions in Medicaid with the medical and pharmacy claims of 117,908 patients with schizophrenia and 170,596 patients with bipolar disorder in Medicaid who were single-eligible, and newly prescribed a second-generation antipsychotic from 2001 to 2008. We tested the impact of formulary restrictions on the medical costs and utilization of patients in the 12 months after the index prescription. To capture social costs in addition to medical expenditures in Medicaid, we estimated the incremental costs of incarcerating patients with schizophrenia and bipolar disorder associated with formulary restrictions.

Patients with schizophrenia subject to formulary restrictions were more likely to be hospitalized (odds ratio 1.13, 

 <.001), had 23% higher inpatient costs (

 <.001). Similar effects were observed for patients with bipolar disorder. Our estimates suggest restrictive formulary policies in Medicaid increased the number of prisoners by 9920 and incarceration costs by $362 million nationwide in 2008.

 Applying formulary restrictions to atypical antipsychotics is associated with higher total medical expenditures for patients with schizophrenia and bipolar disorder in Medicaid. Combined with the other social costs such as an increase in incarceration rates, these formulary restrictions could increase state costs by $1 billion annually, enough to offset any savings in pharmacy costs.

Applying formulary restrictions to atypical antipsychotics is associated with higher total medical expenditures for patients with schizophrenia and bipolar disorder in Medicaid.