The use of an 80% threshold or other binary cut point may be insufficient for characterizing the relationship between medication adherence and Medicare spending.
Patients are commonly considered to be adherent when they have medication on hand at least 80% of the time. We find evidence that a binary adherence measure
may be insufficient to characterize the association between adherent behavior and Medicare spending due to nonlinearities in the adherence/spending relationship among Part D enrollees with diabetes.
Prescription medicines are an important component in the treatment of diabetes and diabetes-related complications. When taken as directed, those medicines play a key role in slowing disease progression, preventing hospitalizations, and reducing premature mortality. Despite considerable evidence for the efficacy of medications in managing diabetes and other chronic diseases, it is estimated that patients never fill 20% to 30% of newly written prescriptions, and more than 50% of medications are not taken as prescribed.1,2 Improving adherence is critically important for reducing avoidable hospitalizations and preventing the use of costly acute care, but gaps persist in our knowledge about which patients should be targeted for intervention. In particular, we lack data about the specific patient populations in whom increasing adherence could have the greatest impact to improve outcomes and lower costs.3,4
A substantial body of research demonstrates that adherence to medications used to prevent or manage diabetes and diabetes-related complications is associated with improved health outcomes and lower healthcare spending.5-13 To classify patients as adherent, researchers have typically used a cutoff of 80% for measurements of the medication possession ratio or proportion of days covered (PDC).12,14-16 This approach makes strong assumptions about the nature of the relationship between adherence and spending. These assumptions are typically treated as maintained rather than testable assumptions, and it remains unclear whether the functional form of this relationship is best characterized by a binary measure of adherence. A better understanding of the relationship between adherence and healthcare spending would help to inform stakeholders’ expectations about the magnitude of spending reductions that are achievable through improvements in adherence. Additionally, it would assist payers and providers in identifying and targeting patients who are likely to benefit from intervention.
Data and Sample Selection
Our study used survey and claims data from the 2007 to 2010 Medicare Current Beneficiary Survey (MCBS), which is a nationally representative sample of Medicare beneficiaries. We limited our analysis to community-dwelling beneficiaries with diabetes who filled Part D—covered prescriptions for angiotensin-converting enzyme (ACE) inhibitors or angiotensin receptor blockers (ARBs), and were continuously enrolled in Medicare Parts A and B and a stand-alone Part D prescription drug plan for a full calendar year between 2007 and 2010. We classified beneficiaries as having diabetes if they self-reported having ever been given a diagnosis of diabetes by a physician. ACE inhibitors/ARBs were chosen for this analysis because they have been shown to be effective for the prevention of vascular complications in patients with diabetes and are recommended for virtually all elderly diabetics.17 Beneficiaries who received drug coverage through capitated Medicare Advantage plans were excluded because these plans do not generate the Parts A and B claims needed to evaluate medical spending.
Our dependent variable was total Medicare Parts A and B expenditures, adjusted to 2010 dollars using the all urban consumers component of the Consumer Price Index (CPI-U). To calculate adherence to ACE inhibitor/ARB therapy, we used the Part D prescription event data to determine the PDC, measured as a continuous variable ranging from 0 to 1. PDC is the adherence metric recommended by the Pharmacy Quality Alliance and is defined as the number of days a full supply of medication is on hand each month divided by the number of days in the month.18 Days spent in the hospital or a skilled nursing facility were excluded from both the numerator and de nominator, as medications were presumably administered from facility supplies during inpatient stays.
Covariates in our regression analyses included sociodemographic characteristics, measures of health status, and controls for healthy adherer bias. Sociodemographic characteristics included age, gender, race and ethnicity, US Census region, and whether or not the beneficiary received the low-income subsidy. Health status was measured by: 1) a count of self-reported comorbid chronic conditions, including hypertension, hyperlipidemia, asthma/emphysema/chronic obstructive pulmonary disease, congestive heart failure, heart disease, osteoarthritis, and depression; 2) self-reported diabetes-related complications related to the eyes, feet, or kidneys; and 3) an indicator for whether the beneficiary originally qualified for Medicare on the basis of a disability.
Variables to control for potential healthy adherer bias included self-reported health status, the number of limitations in activities of daily living, any use of statins or oral antidiabetic medications, any insulin use, education level, marital status, and diabetes management behaviors including checking one’s own blood sugar, checking for sores on feet, and having attended a diabetes management class. We hypothesized that beneficiaries who had more years of education, had spousal support, and took steps to actively manage their diabetes would be more likely to be engaged in their healthcare. If engaged beneficiaries were more likely to have both higher adherence and better overall health outcomes, this healthy adherer effect could confound the relationship between adherence and spending, thereby overestimating the impact of medication adherence.5,19-22
To characterize the bivariate association between level of medication adherence and Medicare Parts A and B spending, we stratified the sample into quintiles based on the proportion of days covered. Due to the panel design of the MCBS, beneficiaries could contribute multiple years of data to our pooled sample. This design violates a key assumption of many standard statistical analytic approaches—namely that the error terms across observations are independent and identically distributed. To account for correlation in the error terms, as well as the nonnormal and heavily right-skewed distribution of Medicare spending, we used cluster-robust generalized linear models with a gamma distribution and log link to estimate the relationship between adherence and total Medicare Parts A and B spending.23 Robust standard errors were computed using an exchangeable correlation matrix.24 Since virtually all beneficiaries in the sample (97%) had positive spending, we replaced the small share of observations with zero spending with a value of $1. To more precisely investigate the functional form of the relationship between adherence and spending, we also performed multivariate regression using Part A spending and Part B spending as dependent variables in separate regression models.
We investigated nonlinearity in the relationship between adherence and spending using polynomial transformations of the continuous adherence measure. Tests for statistical significance on models including more than 1 polynomial function (eg, squared and cubed values) are subject to type II error due to the high correlation between the transformed variables. Orthogonal polynomials reduce the possibility of type II error by creating independent nonlinear functions of a given continuous variable in order to ensure that each transformed variable is evaluated individually without risk of unfounded correlation with other transformed variables. We used Hermite polynomials—part of the family of classical orthogonal polynomials—to develop nonlinear transformations of the PDC variable for inclusion in the regression models.25 All analyses were conducted using SAS 9.2 (SAS Institute, Cary, North Carolina). The study was approved by the Institutional Review Board of the University of Maryland - Baltimore.
Application of our inclusion criteria resulted in a sample of 1881 ACE inhibitor/ARB users, 61.1% of whom met the conventional 80% threshold for adherence. The distribution of PDC was left-skewed, with an average of 77.2%, a median of 85.8%, and an interquartile range of 68.5% to 92.3%. After stratifying the sample by quintile of PDC, each quintile contained 376 observations, except for the middle quintile, which contained 377 observations.
The unadjusted bivariate association between adherence and Medicare Parts A and B spending clearly suggests a nonlinear relationship (). For total Parts A and B spending, increased adherence was associated with consistently lower spending across quintiles 1 through 4, reaching a turning point between quintiles 4 and 5. In quintile 5—which represents the most highly adherent beneficiaries—average total spending was considerably higher than for the rest of the sample, except for the least adherent beneficiaries in quintile 1. Although beneficiaries in quintiles 4 and 5 both meet the conventional definition of high adherence, we observed large spending differences between these groups. For instance, the average total spending for beneficiaries in quintile 4 was $9277 versus $15,501 for those in quintile 5. Part A spending was also highest overall for quintile 5 ($8489)—more than double the amount for quintiles 3 ($4237) and 4 ($3485). We observed less overall variation in Part B spending across adherence quintiles, with no consistent pattern in the bivariate relationship between adherence and spending.
The summarizes the statistical significance and direction of the relationship between medication adherence and Medicare Parts A and B spending, adjusting for sociodemographic characteristics, health status, and healthy adherer effects. Our interest is in the functional form of the relationship, and therefore, we focus here on the significance of the coefficients, rather than their magnitudes, and report the full regression results in an online (available at www.ajmc.com). The statistical significance of the second- and third-order Hermite polynomials provides evidence for nonlinearities in the relationship between medication adherence and spending, but suggests a more complex nonlinear relationship than is reflected in the Figure. Specifically, the statistical significance of the third order term suggests that the relationship is characterized by multiple inflection points. The results presented in the Table indicate that this relationship between adherence and Parts A and B spending appears to be entirely driven by Part A, as we did not observe a statistically significant association between adherence and Part B spending.
Consistent with the literature, we found evidence of a statistically significant association between Medicare spending and adherence to ACE inhibitor/ARB therapy among Part D enrollees with diabetes.5,9,12,13 Unlike prior studies, which have typically used a binary threshold to classify patients as adherent, we sought to better understand the functional form of the relationship between adherence and spending, and whether this relationship varied across the distribution of adherence. Our findings suggest that the relationship between medication adherence and Medicare spending is indeed nonlinear, and that binary measures of adherence are insufficient to fully characterize the relationship between adherent behavior and healthcare spending.
We observed considerable heterogeneity in average total spending for beneficiaries above the conventional 80% adherence threshold, including substantially higher total costs and Part A costs for beneficiaries with the highest levels of medication adherence. This heterogeneity suggests that there are important differences among beneficiaries commonly considered to be adherent. We also observed that the statistical significance of the relationship varied by type of Medicare spending, and that adherence had a much stronger association with Part A hospital spending than with outpatient and physician spending incurred under Part B.
On average, the most highly adherent beneficiaries incurred substantially higher Part A spending than their less-adherent counterparts. This should not and does not imply, however, that the relationship between good medication-taking behavior and poor health outcomes is a causal one. One possible explanation for this counterintuitive association is that highly adherent beneficiaries are further along in their disease progression. In this case, poor health status would increase beneficiaries’ underlying risk of hospitalization, while simultaneously increasing their likelihood of adhering to therapy. Similarly, it may be the case that highly adherent beneficiaries experienced significant health shocks resulting in hospitalization early in the year, and improved the regularity of their medication-taking behavior following discharge. In this case, the high level of adherence observed for these beneficiaries would be the outcome, rather than the driver, of high Medicare spending. Based on the current cross-sectional design, we cannot determine which case prevails. At the same time, our findings indicate that studies employing a cross-sectional design may miss opportunities to develop new insights into adherence and spending when an 80% cutoff is employed. Future adherence research employing longitudinal designs can build on the findings from this study to investigate the temporal relationship between adherence and spending, while accounting for nonlinearities in the relationship.
A few studies have challenged the use of an 80% adherence threshold, suggesting evidence of a broader dose-response relationship between adherence and outcomes. For example, Esposito and colleagues found evidence of a graded relationship between adherence and healthcare costs for patients with congestive heart failure, with spending rising monotonically as the level of adherence declined.15 Similarly, Briesacher and colleagues reported significant reductions in fracture risk and overall health costs associated with adherence levels as low as 40% for patients with osteoporosis, while Stuart and colleagues observed significant reductions in 3-year Medicare Parts A and B spending associated with adherence rates of 60% or higher for patients with diabetes taking ACE inhibitors/ ARBs.12,16 In contrast, a recent study by Choudhry and colleagues found that patients with adherence levels below 80% did not experience the same significant reductions in adverse clinical events post myocardial infarction as did their more highly adherent counterparts.14
An important limitation of our study was its small sample size, which limited the ability to investigate subgroup effects via interactions or defined subgroups. The reported results may mask differences between patient subgroups defined by demographic or clinical conditions. Future research with larger samples is necessary to confirm our findings and to explore whether the relationship between adherence and spending varies across conditions and therapeutic classes. Indicators of patient participation in transitional care programs or other programs that directly support medication adherence are not available from this data set; however, participation in these programs may be correlated with underlying hospitalization risk.
This study makes an important contribution by highlighting the inadequacy of defining adherent behavior with a single cut point in some instances, a key assumption made in much of the existing adherence and cost offsets literature. Our findings also reinforce the need for additional longitudinal research to determine the level of adherence necessary to achieve better health outcomes. Studies investigating the relationship between adherence and spending should build on a rich, nuanced conceptualization of the patient- and non—patient-level factors contributing to higher observed adherence. A better understanding of these relationships would assist policy makers, payers, and providers in identifying and targeting patients most likely to benefit from intervention.
Author Affiliations: Department of Pharmaceutical Health Services Research, School of Pharmacy, University of Maryland (PNR, EO), Baltimore, MD; Pharmaceutical Research and Manufacturers of America (PNR), Washington, DC.
Source of Funding: This work was conducted as an unfunded student project. The lead author is employed on a part-time basis by the Pharmaceutical Research and Manufacturers of America (PhRMA), which had no role in the funding or conduct of this analysis. Any views expressed here are solely those of the authors and do not reflect the views of PhRMA.
Author Disclosures: Ms Roberto is a part-time employee at PhRMA. Dr Onukwugha reports 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 (EO, PNR); acquisition of data (PNR); analysis and interpretation of data (EO, PNR); drafting of the manuscript (PNR); critical revision of the manuscript for important intellectual content (EO, PNR); statistical analysis (PNR); and supervision (EO).
Address correspondence to: Pamela N. Roberto, MPP, 220 Arch St, Baltimore, MD 21201. E-mail: firstname.lastname@example.org.
1. Fischer MA, Stedman MR, Lii J, et al. Primary medication non-adherence: analysis of 195,930 electronic prescriptions. J Gen Internal Med. 2010;25(4):284-290.
2. Tamblyn R, Eguale T, Huang A, Winslade N, Doran P. The incidents and determinants of primary nonadherence with prescribed medication in primary care: a cohort study. Ann Internal Med. 2014;160(7):441-450.
3. Seabury S, Gupta CN, Philipson TJ, Henkaus LE; PhRMA Medication Adherence Advisory Council. Understanding and overcoming barriers to medication adherence: a review of research priorities. J Manag Care Pharm. 2014;20(8):775-783.
4. Redberg RF. Medication regimen adherence and patient outcomes. JAMA Intern Med. 2014;174(2):193.
5. Stuart B, Davidoff A, Lopert A, Shaffer T, Samantha Shoemaker JS. Lloyd J. Does medication adherence lower Medicare spending among beneficiaries with diabetes? Health Serv Res. 2011;46(4):1180-1199.
6. Balkrishnan R, Rajagopalan R, Camacho FT, Huston SA, Murray FT, Anderson RT. Predictors of medication adherence and associated health care costs in an older population with type 2 diabetes mellitus: a longitudinal cohort study. Clin Ther. 2003;25(11):2958-2971.
7. Encinosa WE, Bernard D, Dor A. Does prescription drug adherence reduce hospitalizations and costs? the case of diabetes. Adv Health Econ Health Serv Res. 2010;22:151-173.
8. Jha AK, Aubert RE, Yao J, Teagarden JR, Epstein RS. Greater adherence to diabetes drugs is linked to less hospital use and could save nearly $5 billion annually. Health Aff (Millwood). 2012;31(8):1836-1846.
9. Roebuck MC, Liberman JN, Gemmill-Toyama M, Brennan TA. Medication adherence leads to lower health care use and costs despite increased drug spending. Health Aff (Millwood). 2011;30(1):91-99.
10. Lau DT, Nau DP. Oral antihyperglycemic medication nonadherence and subsequent hospitalization among individuals with type 2 diabetes. Diabetes Care. 2004;27(9):2149-2153.
11. Sokol MC, McGuigan KA, Verbrugge RR, Epstein RS. Impact of medication adherence on hospitalization risk and healthcare cost. Med Care. 2005;43(6):521-530.
12. Stuart B, Loh FE, Roberto P, Miller L. Incident user cohorts for assessing medication cost-offsets [published online March 15, 2014]. Health Serv Res. 2014;49(4):1364-1386. doi:10.1111/1475-6773.12170.
13. Stuart B, Loh FE, Roberto P, Miller LM. Increasing Medicare part D enrollment in medication therapy management could improve health and lower costs. Health Aff (Millwood). 2013;32(7):1212-1220.
14. Choudhry NK, Glynn RJ, Avorn J, et al. Untangling the relationship between medication adherence and post-myocardial infarction outcomes: medication adherence and clinical outcomes. Am Heart J. 2014;167(1):51-58.e5.
15. Esposito D, Bagchi AD, Verdier JM, Bencio DS, Kim MS. Medicaid beneficiaries with congestive heart failure: association of medication adherence with healthcare use and costs. Am J Manag Care. 2009;15(7):437-445.
16. Briesacher BA, Andrade SE, Yood RA, Kahler KH. Consequences of poor compliance with bisphosphonates. Bone. 2007;41(5):882-887.
17. Rosen AB. Indications for and utilization of ACE inhibitors in older individuals with diabetes: findings from the National Health and Nutrition Examination Survey 1999 to 2002. J Gen Intern Med. 2006;21(4):315-319.
18. Nau DP. Proportion of days covered (PDC) as a preferred method of measuring medication adherence. Pharmacy Quality Alliance website.
http://www.pqaalliance.org/images/uploads/files/PQA%20PDC%20vs%20%20MPR.pdf. Accessed October 14, 2013.
19. Fairman KA, Curtiss FR. Still looking for health outcomes in all the wrong places? misinterpreted observational evidence, medication adherence promotion, and value-based insurance design. J Manag Care Pharm. 2009;15(6):501-507.
20. Salas M, Hughes D, Zuluaga A, Vardeva K, Lebmeier M. Costs of medication nonadherence in patients with diabetes mellitus: a systematic review and critical analysis of the literature. Value Health. 2009;12(6):915-922.
21. Brookhart MA, Patrick AR, Dormuth C, et al. Adherence to lipid-lowering therapy and the use of preventive health services: an investigation of the healthy user effect. Am J Epidemiol. 2007;166(3):348-354.
22. Chewning B. The healthy adherer and the placebo effect. BMJ. 2006;333(7557):18-19.
23. Glick HA, Doshi JA, Sonnad SS, Polsky D. Economic Evaluation in Clinical Trials. New York, NY: Oxford University Press; 2007.
24. Fitzmaurice GM, Laird NM, Ware JH. Applied Longitudinal Analysis. 2nd ed. Hoboken, NJ: John Wiley & Sons; 2011.
25. Greene WH. Econometric Analysis. 6th ed. Upper Saddle River, NJ: Prentice Hall; 2007.