Medication Adherence and Medicare Expenditure Among Beneficiaries With Heart Failure

Our multivariate models included an extensive set of covariates to control for confounding due to health status, predilection for prescription drug use, access to care, and healthy adherer bias. Health status measures were captured from 3 sources: (1) self-reported measures (a standard 5-item scale of health status from excellent to poor, body mass index computed from self-reported height and weight, and limitations in activities of daily living); (2) Medicare administrative data (current and former recipients of Social Security Disability Insurance); and (3) select claims-based comorbidities common in the elderly, including hypertension, ischemic heart disease, diabetes, chronic obstructive pulmonary disease, chronic renal disease, hyperlipidemia, and osteoarthritis, plus an index of coexisting conditions. Demographic factors potentially influencing both prescription use and outcomes included age, sex, race, marital status, and education. Access to care was captured by data on income and the possession of supplemental medical or drug coverage. Dummy variables were used to capture censoring (loss to follow-up, admission to a long-term care facility, death), as well as year of induction into the MCBS (a proxy for temporal trends in treatment).

We controlled for healthy adherer bias by including additional measures of prescription drug utilization for each of the 4 drug classes in each drug-specific equation. This approach enabled us to estimate the effect of an incremental increase in adherence with each medication conditional upon utilization of other medications in the regimen. Except for the drug utilization and censoring variables, all covariates were measured during the baseline year.

Statistical Analysis

Regressions for each of the drug groups were used to estimate the relationship between medication adherence and 3-year Medicare spending. Models were estimated using a generalized linear model with a gamma distribution and log link to approximate the right skewed distribution of Medicare costs.

Three sets of 4 models were estimated. The models within sets were estimated by restriction to users of the 4 distinct drug groups. The first set modeled the bivariate relationship between Medicare spending and average daily pill counts within the specified drug group without control variables.

The other 2 sets of models included an extensive set of covariates (Table 1) as well as additional drug measures to address confounding due to unobserved healthy adherer effects. The drug measures in the second set included daily pill counts for the specified drug group in addition to dichotomous variables indicating use of the other 3 groups of drugs. The third set of models included measures of daily pill counts for all 4 drug groups as well as indicators of use of each.

We contend that by including measures of utilization for several drugs in the same equation, any shared variance due to a healthy adherer effect is removed. However, this does give rise to larger standard errors and introduces the possibility of underestimating the true impact of drugs with similar effects on health outcomes due to multicollinearity between the drug measures.

To facilitate interpretation, all model coefficients were converted to marginal effects, holding all covariates at their sample means. All models were estimated using Stata 11 (Statacorp, College Station, Texas).

RESULTS

Descriptive Findings

The characteristics of the study sample and medication user groups are presented in Table 1. Despite a CHF diagnosis confirmed by claims data, almost 17% received no drug therapy, with more than half the cohort taking an ACE inhibitor or ARB (57.6%), 71.8% taking a diuretic of any class, 37.1% taking a beta-blocker, and a third (33.6%) on cardiac glycoside therapy (Table 2). As anticipated, higher proportions of each drug user group were generally concentrated in the later years, possibly reflecting worsening disease as well as evolution in recommended treatment regimens and the diffusion of new products in some drug classes over the study period. While beta-blocker and cardiac glycoside groups had slightly lower percentages of minorities than ACE inhibitor/ARB and diuretics users, generally the groups were demographically similar. Despite their heart failure diagnoses, more than half the subjects in each group described their health as “good” or “excellent.” Table 2 also presents data on medication adherence and Medicare expenditure. Cumulative mean 3-year spending on Medicare Part A and Part B services exceeded $50,000 for each user group, measured in constant 2006 dollars. The median pill count ranged from 0.63 to 0.82 per day over the 36-month follow-up period for all drug groups examined.

Table 3 presents the bivariate regression results under model set 1 and the multivariate regression results under model set 2. The models quantify the estimated impact of changes in prescription drug adherence, as measured in daily pill counts, on Medicare expenditures, with a negative association reflecting an overall cost saving to the program. For each drug group, the first row shows the unadjusted impact on Medicare expenditure of a 10% increase in daily pill counts estimated in the simple regressions of model set 1. The results show that increased utilization of ACE inhibitors or ARBs, beta-blockers, and cardiac glycosides was generally associated with lower Medicare spending, with only the effect of cardiac glycosides reaching statistical significance (P <.05). Results from model set 2 show that a 10% increase in daily pill counts of ACE inhibitor/ ARB, beta-blockers, diuretics, and cardiac glycosides resulted in savings to Medicare of $390 (P = .06), $510 (P = .04), $13 (P = 0.92), and $923 (P = .01), respectively.

Table 4 presents results from the fully controlled models of the third set. A 10% increase in daily beta-blocker and cardiac glycoside pill counts resulted in savings to Medicare of $561 (P <.05) and $750 (P <.05), respectively. As expected, the effect size and statistical significance generally decreased in the multivariate models. We attribute this decrease to the shared variance associated with the healthy adherer effect. The direction of effect was consistent across all 3 sets of models. Among