Medication is an essential aspect of tertiary prevention, as it often addresses symptoms, may restore function, and minimizes adverse consequences associated with chronic conditions. Medication adherence is most often studied in the context of Medicare Part D. In a newly published Medical Care article, Drs. Roebuck, Kaestner, and Dougherty, instead measure the associations between medication adherence and health services utilization among Medicaid beneficiaries in 10 states with seven conditions.
The authors report relationships between the proportion of days covered (PDC) [pdf] and lower health services utilization, particularly for patients with congestive heart failure, hypertension, diabetes, and schizophrenia or bipolar disorder. They also provide evidence that the threshold of 0.80 PDC commonly used to dichotomize patients into “adherent” or “non-adherent” might be missing the benefits potentially accruing to patients who are below the threshold but still receiving benefits from their medication.
This idea echoes earlier results from a paper by Yang and colleagues that found lower adherence to anti-hypertensive medication, measured using the medication possession ratio, was associated with higher CVD risk in a population of fee-for-service Medicaid beneficiaries. They also estimated that costs from increased medication use were offset by lower costs from having fewer CVD events.
As with so many issues, the details matter. I guess my recent thinking about gardening lately (trying to turn my thumb greener) might have translated into blogging. The rest of this post gets into the weeds a bit about how medication adherence is measured and estimated. I’ve been thinking a lot about these issues lately as I watch family members try to keep up with varied medication regimes.
The article by Roebuck and colleagues used the Medicaid Analytic eXtract (MAX) files, which included fee-for-service and managed care claims for Alabama, California, Florida, Illinois, Indiana, Louisiana, New Hampshire, New Mexico, New York, and Virginia from 2008 to 2010. The authors included individuals aged 18-64, with Medicaid coverage only, without coverage gaps, and who did not reside in long-term care facilities. All patients who were diagnosed during an inpatient stay or at 2 outpatient visits with one of the target conditions were classified as requiring pharmacotherapy. If the patient didn’t have any prescription drug fills, their PDC was zero.
As pointed out by the authors, this approach assumes that all patients meeting the diagnostic criteria laid out in the study were prescribed medication for the condition. This method catches people who are prescribed medication but never initiate pharmacotherapy (they never have any fills). Another standard approach in the literature is to assume only patients with at least one prescription drug refill during the study period were prescribed medication. Both options have strengths and weaknesses, and Roebuck and colleagues report results for this alternative measure in their online Appendix. Estimates of the percent of patients on medication for each condition falls within previous estimates for the percentage of patients who do not initiate pharmacotherapy even after being prescribed medication.
The authors use both continuous and categorical versions of PDC (lagged one year). For most of the conditions, there is very little evidence about the thresholds of PDC that might impact health service use. Numbers of inpatient hospitalizations, emergency department visits, and outpatient visits are modeled using Poisson regression with patient-level fixed effects, year dummy variables, and several covariates, including the Charlson comorbidity index. Separate regressions were run for each outcome and each condition. The samples were not mutually exclusive: someone with both diabetes and depression would have been included in the sample for each of those regression models. The sample was divided based on Medicaid eligibility, with individuals qualifying as blind or disabled separated from other adults; as expected, sample sizes were much bigger for the blind or disabled cohort.
In the blind or disabled cohort, there were statistically significant negative associations between the continuous measure of PDC and inpatient hospitalizations for 5 of the 7 conditions–translating to 0.17-0.56 fewer annual hospitalizations per patient for PDC values moving from zero to one. Several of the conditions’ regressions showed qualitatively similar results in the cohort of other adults, but with some differences in which conditions showed strong associations.
Medication adherence was also negatively related to emergency department use for 5/7 conditions for the blind/disabled cohort and 4/7 for the cohort of other adults. Percent decreases in emergency visits ranged from 3%-10% for the blind/disabled group (except depression was positively associated,+3%) and from 8%-12% for the other adult cohort (PDC from zero to one). PDC was negatively related to outpatient/clinic visits for nearly every condition in both groups (reductions of 1-15% for change in PDC of zero to one). These percent changes are potentially significant, but of course, they represent a completed change in PDC, from no adherence to perfect adherence.
The next set of models described in the paper estimate the same regressions but with categorically entered PDC: 0-0.40 (reference), 0.40-0.59, 0.60-0.79, and 0.80-1.0. For several of the conditions, there are robust associations with decreased health service use at categories below the typical 0.80 thresholds. For some of the conditions, such as depression in the cohort of other adults, each category is with fewer outpatient/clinic visits, with marginal effects ranging from -0.19 to -0.30. In summary, the higher PDC categories were more likely to be significantly related to reduced health services utilization compared to the reference category for more of the conditions (as we might expect). Interestingly, even 0.40-0.59 PDC was negatively associated with health services use compared to <0.40 PDC for some conditions, such as hypertension in the blind/disabled cohort and dyslipidemia in the other adult cohort.
Most likely because of sample size, utilization was measured as total utilization over a year period and was not condition specific. While it would be nice to assume the estimated associations would be strictly additive, we just don’t have enough information to make that claim. Roebuck, Kaestner, and Dougherty have made a good start on this complicated issue and have furthered the discussion of how adherence levels are linked with health services utilization for different health conditions.
The article makes a good case for prescription benefits to remain included with Medicaid and suggests which conditions might be fruitful intervention targets from a cost perspective. However, additional research about the benefits of medication adherence across multiple conditions for individuals with complex health needs is necessary to make a case for specific interventions. As the authors point out, just because a prescription was refilled, doesn’t mean that the medication was taken as indicated.
Individuals may use a variety of strategies to make medicines “last longer.” This is particularly true when cost is a barrier. Most medications are provided by state Medicaid plan with little to no co-pay. However, even a small copay may be a meaningful difference for patients. Also, co-pays only represent the direct out-of-pocket costs — they do not include other costs such time and travel to a pharmacy.
Additionally, the data used in the paper were before the expansion of Medicaid under the Affordable Care Act (ACA). Given the range of eligibility criteria for the states in the analysis, a detailed understanding of the populations would inform generalizability and future work using data from the rollout of the ACA would provide a very interesting extension.
Jess Williams