Does the diabetes health plan have a differential impact on medication adherence among beneficiaries with fewer financial resources?

BACKGROUND: The Diabetes Health Plan (DHP), a value-based insurance plan that reduces cost sharing, was previously shown to modestly increase employer-level medication adherence. It is unclear how the DHP might impact individuals with different incomes. OBJECTIVE: To examine the impact of the DHP on individual-level medication adherence, by income level. METHODS: This is a retrospective, quasiexperimental study. An employer-level propensity score match was done to identify suitable control employers, followed by individual-level propensity score weighing. These weights were applied to difference-in-difference models examining the effect of the DHP and the effect of income on changes in adherence to metformin, statins, and angiotensin-converting enzymes/angiotensin receptor blockers. The weights were then applied to a differences-in-differences-in-differences model to estimate the differential impact of DHP status on changes in adherence by income group. RESULTS: The study population included 2,065 beneficiaries with DHP and 17,704 matched controls. There were no significant differences in changes to adherence for any medications between beneficiaries enrolled in the DHP vs standard plans. However, adherence to all medications was higher among those with incomes greater than $75,000 (year 1: metformin: +7.3 percentage points; statin +4.3 percentage points; angiotensin-converting enzymes/angiotensin receptor blockers: +6.2 percentage points; P < 0.01) compared with those with incomes less than $50,000. The differences-in-differences-in-differences term examining the impact of income on the DHP effect was not significant for any comparisons. CONCLUSIONS: We did not find significant associations between the DHP and changes in individual-level medication adherence, even for low-income beneficiaries. New strategies to improve consumer engagement may be needed to translate value-based insurance designs into changes in patient behavior.


Plain language summary
The Diabetes Health Plan (DHP) lowers the patient cost of medications, such as metformin and statins. We compared medication adherence among patients with and without the DHP, and for patients in each group who were lower income or higher income. We did not find adherence differences between patients with or without the DHP, even among patients with a lower income. New strategies to increase education and engagement may increase the clinical benefits of insurance plans that reduce patient medication costs.

Implications for managed care pharmacy
This study of a diabetes-specific health plan incorporating value-based insurance design shows that this approach alone may not improve adherence. We used rigorous statistical methods, including a 2-level propensity score and differencein-difference-in-difference models, but did not demonstrate an adherence benefit for patients with a lower income, a group that in theory would benefit most from cost savings. Future value-based insurance design interventions for patients with diabetes should consider additional components to increase the effectiveness of reduced or eliminated copayments.
The relationship between socioeconomic status and health outcomes is well documented. [1][2][3][4] Considerable research has shown that individuals with higher incomes are more likely to live longer and are less likely to report poor health status or health-related activity limitations. 4 Moreover, socioeconomic disparities have been reported for multiple different diseases, including diabetes. 2,4 A major cause of mortality and morbidity, diabetes has been shown to disproportionately impact the poor. Not only are there higher rates of diabetes among those with lower income, but they also face a 2-fold increase in diabetes-related mortality. 5 There is evidence that these disparities may be related to differences in health care access, health care utilization, and differences in patterns of health behavior. 6 The prevalence of diabetes is highest among patients with lower socioeconomic status, who are less able to access reliable and comprehensive health care. 7,8 For example, individuals with lower income report lower rates of important preventive screening measures. [9][10][11][12] Among individuals with diabetes, those with lower wages are less likely to receive glycated hemoglobin (A1C) testing, eye examinations, and foot examinations, measures that are important for decreasing diabetes-related morbidity and mortality. 9 Moreover, because of financial barriers, individuals with lower income are more likely to delay or forego care. 10,11 In one study, lower-wage earners were almost 2.5 times more likely than those with higher wages to report cost-related problems with medication adherence. 11 This cost-related nonadherence is likely clinically significant as higher levels of medication adherence have been shown to be associated with lower rates of diabetes-related complications and emergency department visits. 12,13 Value-based insurance designs (VBIDs) have the potential to help mitigate some of these financial barriers to health care services, with the goal of improving patient outcomes. 14 With reduced out-of-pocket payments for important medications and services, the hope is that providers will work with patients to increase their use of these high-value services, thus leading to improved clinical outcomes. [15][16][17] Evidence thus far has found that the implementation of a VBID can lead to modest improvements in medication adherence, especially when targeting specific patient populations. [16][17][18] The Diabetes Health Plan (DHP) is the first such VBID that specifically targets individuals with diabetes and prediabetes; these plans are not yet well established in the insurance marketplace. Hallmarks of this innovative plan include modestly lower cost sharing for office visits and medications that reduce the incidence of and complications for diabetes (Supplementary Table 1, available in online article). 19 Additionally, this insurance plan offers enrollees free or low-cost resources for diabetes management. Thus far, our group has found that the DHP has led to modest improvements in mean adherence to evidence-based medications for each participating employer and lower rates of incident diabetes among beneficiaries with prediabetes. 20,21 How patients with lower income who bear a disproportionate disease burden might respond to the DHP and other VBIDs is less clear. Given the higher rates of cost-related nonadherence among individuals with lower incomes, it is possible that these patients may be more likely than patients with a high income to respond to interventions that decrease their out-of-pocket payments, leading to a relatively greater uptake of high-value services. 6 Indeed, a recent rapid review of literature evaluating the impact of removing cost sharing for preventive services, including cancer screening, found some evidence that those with lower income or higher previous levels of cost sharing may benefit more from eliminating cost sharing. 22 This is similar to a previous study assessing the impact of eliminating cost sharing for cardiovascular medications, which found a greater increase in medication adherence among non-White patients, a group with historically lower rates of adherence. 23 Thus, the primary objective of this study is to examine if income influences an individual's response to reduced cost sharing under the DHP, by comparing changes in medication adherence over time among individuals of different incomes who were and were not offered this VBID.

SETTING
The DHP, developed by UnitedHealthcare, is an innovative, disease-specific, opt-out health plan for beneficiaries with diabetes and prediabetes. 24 Key features include reduced or eliminated copayments for office visits to primary care providers and endocrinologists and for recommended antiglycemic, antihypertensive, and statin medications. The DHP also provides access to diabetes-specific telephone case management and other online resources. In addition to these benefits, the DHP provides scorecards with reminders to complete health maintenance activities, such as biannual A1C and cholesterol screening, an annual retinal eye examination, and age-appropriate cancer screening. Employers who purchase the DHP can modify the standard benefit design to meet their needs. Altogether, the DHP provides $150-$500 in annual out-of-pocket savings for enrollees. 20

STUDY DESIGN AND POPULATION
For this study, we used a pre-post quasi-experimental "intent-to-treat" design, with a concurrent control group and statins in the first year and the second year after DHP implementation, calculated as the mean proportion of days covered. We calculated adherence in the first year after implementation using the last 9 months of the year, using the first 3 months to account for medication carryforward. We modified adherence in the second year after implementation such that the full 12 months was used to calculate adherence (and the last 3 months of the first year after DHP implementation was used to calculate the prescription carry-forward). We did not control for multiple prescriptions within a medication class, but if 2 or more prescriptions were filled on the same day, we included the prescription with the greater number of days supply in calculating adherence.

INCOME
Household income for each individual at the beginning of the study was estimated by UnitedHealthcare using an algorithm powered by the AmeriLINK Consumer Database, which includes information such as net worth, home value, bankruptcy suppression, credit card information, buying behavior, investment interests, and occupation. [25][26][27] These data were combined with census-level data to identify income distribution and the Internal Revenue Service data on zip code level income to predict household income. We used the following income categories: less than $50,000, $50,000-$74,999, and greater than or equal to $75,000. Notably, small sample sizes limited our ability to further stratify individuals with annual household incomes less than $50,000.

STATISTICAL ANALYSIS
To minimize potential selection bias, particularly at the employer level, we used an employer-level propensity score match and a beneficiary-level propensity score weighting approach, similar to the study by Wharam et al. 28 First, we sought to identify control employers that are most likely to be similar to DHP employers. This propensity score model predicted the likelihood that an employer would purchase the DHP. The model included the following variables: employer size; geographic region; an estimated measure of the overall generosity of benefits for each health plan; proportion of beneficiaries with a high-deductible health plan; proportion of beneficiaries in age strata; proportion of beneficiaries in income strata; race and ethnicity; and sex, proportion of beneficiaries with diabetes, hypertension, coronary artery disease, congestive heart failure, dementia, schizophrenia, anxiety, depression, osteoarthritis, rheumatoid arthritis, nonskin cancer, chronic obstructive pulmonary disease, atrial fibrillation, end-stage renal disease, stroke, peripheral vascular disease, or hyperlipidemia. Propensity score of employers that did not offer the DHP for comparison. Administrative data provided by the plan, including enrollment, prescription claims, medical claims, and laboratory results, were used to identify the study sample and create the analytic measures. Our study population included commercially insured employees and their covered dependents (together considered "beneficiaries"), classified according to whether their employer offered the DHP (ie, DHP beneficiaries) vs a standard medical insurance plan (ie, controls). We began with 43 DHP employers (n = 1,224,890 individuals). We excluded 21 employers (n = 964,746 individuals) who did not have at least 3 years of continuous enrollment data defining the study period (1 year pre-DHP implementation and 2 years postimplementation) and 12 employers (n = 165,609 individuals) that did not have complete claims data. We then limited our study sample to individuals from the remaining DHP employers who had 3 years of continuous enrollment, who were aged between 18 and 63 years, who did not receive Medicare coverage during the entire study period, who were not pregnant during the study period, and who had a diagnosis of diabetes (Supplementary Figure 1). A diabetes diagnosis was defined as having any of the following prior to the implementation of the DHP: (1) at least one 250. xx International Classification of Diseases, Ninth Revision, Clinical Modification diagnosis code from an inpatient, outpatient, or emergency department claim; (2) laboratory A1C value of 6.5% or greater, last fasting plasma glucose level of 125 mg/dl or greater, or a 2-hour value on an oral glucose tolerance test of greater than 200 mg/dL; or (3) at least one prescription fill for insulin, or antiglycemic medication other than metformin. Altogether, our study sample included 2,397 individuals with diabetes employed by 10 employers, prior to propensity score matching (Supplementary Figure 1).
We applied the same exclusion criteria to our initial 658 control employers. In addition, we also excluded 39 employers (n = 79,202 individuals) located in the Mid-Atlantic region, where no DHP employers were located, and 22 employers (n = 67,685 individuals) in which more than 90% of beneficiaries had a high-deductible health plan. This left 472 employers (n = 38,934 individuals) as potential control employers who offered standard health plans (Supplementary Figure 2). For these control employers, we defined the "preperiod" as calendar year 2010 and the postperiod as January 2011 to December 2012, to match the most common implementation date for employers who purchased the DHP.

OUTCOME MEASURES
The primary outcome variable was individual-level medication adherence for metformin, angiotensin-converting enzymes (ACEs)/angiotensin receptor blockers (ARBs), and ethnicity, education, income, baseline medication adherence to the medication of interest, the number of other chronic conditions, baseline utilization (defined as the presence of 2 or more outpatient claims vs 1 or fewer in the last year), and diabetes severity (defined as diet and lifestyle-controlled vs any antiglycemic medications). We used these estimated propensity scores to assign a weight to each individual that was inversely proportional to the probability of each beneficiary being in the treatment group in which they were actually included.
These weights were applied to our full analytical model. We first constructed a difference-in-difference model examining the effect of the DHP on change in adherence to metformin, statins, and ACEs/ ARBs. We then constructed a second difference-in-difference model to examine how adherence to medications varies by income within our study population, irrespective of DHP status. Finally, we used a linear differences-in-differences-in-differences regression model to examine pre-post changes in medication adherence in individuals with a DHP vs a standard insurance plan who fall into 1 of 3 income groups. The differences-in-differences-in-differences term of interest was a 3-way interaction between indicators of DHP vs standard insurance, prestudy vs poststudy period, and indicators for each income category. This reflects the impact of DHP on medication adherence among different income groups. We also conducted a sensitivity analysis for each medication, examining whether the DHP was associated with differences in the proportion of the sample achieving 80% adherence. All relevant 2-way interaction terms were also included in the model. Analyses were conducted using SAS 9.3 and Stata 14.2.
an individual-level propensity score for every individual for each medication of interest within each matched sample among the complete cases. These propensity scores reflected the probability of an individual having a DHP employer instead of a control employer, based on age, sex, race matching yielded 190 control employers in the region of common support.
We then used inverse propensity score weighting to adjust for individual-level differences between beneficiaries of companies that offered the DHP and those who did not offer the DHP. To do so, we calculated  TABLE 1 Does the diabetes health plan have a differential impact on medication adherence among beneficiaries with fewer financial resources?
differences in demographic characteristics between DHP beneficiaries and control beneficiaries who were taking one of the medications under study.

MEDICATION ADHERENCE AND DHP UPTAKE (DIFFERENCE-IN-DIFFERENCES)
In adjusted results (Table 2), the changes in mean predicted adherence rates over time among beneficiaries offered the DHP were similar to those among beneficiaries offered standard insurance plans, irrespective of income (P > 0.05 for all 6 comparisons). The difference-in-differences interaction effects between DHP classification and time were not statistically significant for any of the 3 medication classes among all income groups combined.

MEDICATION ADHERENCE BY INCOME GROUP (DIFFERENCE-IN-DIFFERENCES)
In the adjusted results, the mean predicted rates of adherence to all medication categories examined were higher among those with higher incomes, during all years examined (Table 3). At baseline, medication adherence to metformin, statins, and ACEs/ARBs was 6.9 percentage points, 5.8 percentage points, and 5.5 percentage points higher among those in the highest-income strata (> $75K) compared with

SAMPLE CHARACTERISTICS
The final analytic sample included 2,065 DHP and 17,704 control beneficiaries (Table 1). At baseline, DHP beneficiaries were older (53.9 years vs 52.6 years; P < 0.001) and a greater percentage of DHP beneficiaries were female (44.9% vs 41.9%; P = 0.008). Additionally, there were racial and ethnic differences between DHP vs control beneficiaries, with a greater proportion of DHP beneficiaries who were African American and a greater proportion of control employers who were Hispanic (P < 0.001; Table 1). There were no statistically significant differences in the baseline number of comorbidities, health care utilization (ie, 0-1 vs 2 or more outpatient visits in the preceding year, P = 0.792), and diabetes severity (ie, the percentage taking antiglycemic medications vs those with diet and lifestyle-controlled diabetes). Baseline medication adherence to metformin was similar between groups, but baseline adherence to statins and ACEs/ARBs was lower in the DHP group (71.5% vs 74.2%, P < 0.001, and 76.0% vs 78.4%, P = 0.001, respectively). After adjusting with individual weights that include adherence for each medication of interest, there were no statistically significant

Discussion
In summary, we examined the association between the DHP, a disease-specific value-based insurance design product, and individual-level adherence to evidence-based medications among individuals with diabetes and prediabetes, across different income groups. Our study has 2 notable findings: (1) there was no difference in how individuals with different incomes responded to the DHP and (2) we did not observe an association between DHP implementation and medication adherence at the individual level. Consistent with what has been previously reported, in our sample, medication adherence increased with income. 29 However, we did not observe a differential effect in how individuals with different incomes responded to the DHP. There are many reasons why individuals with a lower income may not preferentially respond to a VBID. First, VBIDs often involve complex cost-sharing structures. Multiple studies have demonstrated that as cost-sharing structures become more complex, patients' understanding of their insurance benefits decreases. 12,30,31 Moreover, understanding appears to be associated with income. Compared with those with lower incomes, those with higher incomes had improved knowledge of their benefits in a VBID and may report those in the lowest-income strata (< $50K), respectively. The difference-in-differences interaction effects between income and time were not statistically significant for any of the 3 medication classes, without regard to DHP classification. That is, the differences between the lowest-and highest-income strata in changes in medication adherence over time were not statistically significant (P > 0.05 for all 6 comparisons).

CHANGE IN MEDICATION ADHERENCE WITH DHP, BY INCOME (DIFFERENCE-IN-DIFFERENCE-IN-DIFFERENCES)
The 3-way interaction term examining the impact of income on the DHP effect on adherence over time was not significant for any of the 3 medication classes (Table 4). That is, the absolute differences in predicted changes in adherence between baseline and year 1 or year 2 with DHP vs a standard insurance plan were not significantly different (all P > 0.05) for the highest-income (> $75K) and lowest-income groups (< $50K).
Our sensitivity analyses examining the impact of the DHP on the 80% adherence threshold were not statistically significant. Predicted Medication Adherence Among Beneficiaries in the Lowest-Income Strata (< $50K) vs the Highest-Income Strata ($75 + K) Irrespective of DHP DID   TABLE 3 Does the diabetes health plan have a differential impact on medication adherence among beneficiaries with fewer financial resources?
size and could not generate a measure of federal poverty level (FPL) for each family. This has important implications as different incomes translate into different levels of poverty depending on family size. For example, an annual income of $50,000 is 400% of FPL for a household of 1 but only 194% of the FPL for a household of 4. 40 Nevertheless, we set $50,000 as the upper threshold for our lowest-income strata, because this annual income would qualify even the smallest household for financial assistance in the Affordable Care Act insurance marketplaces. 41 Finally, we did not have access to use of third-party platforms, such as GoodRx, which may have influenced medication use. 42 Prior studies evaluating the impact of VBIDs on medication adherence have found modest, but significant, improvements in adherence, ranging from 0.1% to 14%. 15,16 Similarly, our group previously demonstrated an improvement in medication adherence after DHP implementation, at the employer level. 20 The lack of an association between DHP implementation and medication adherence in the current study was somewhat surprising, but there are important differences in study design that must be considered. First, the sample populations differed somewhat across the 2 papers. We updated the dataset for the current analysis, and only 8 of the 10 DHP employers included in this study overlapped with those included in the previous study. Importantly, baseline adherence to the medications of interest was higher in our current study compared with the previous study; as such, a substantial proportion of previously nonadherent would need to become adherent to achieve statistical significance. Third, there were some differences between the 2 papers in the included covariates and in the structure of the propensity score. Finally, this was an individual-level study, compared with the prior employer-level aggregate analysis. Although the greater use of low-cost preventive services. 32,33 As a result of lower levels of health insurance literacy as well as limited health care access, individuals with a lower income may not be able to fully take advantage of benefits offered by VBIDs, like the DHP, and may avoid preventive services because of perceived cost-related concerns. 11,12,32,[34][35][36] We did not have information on strategies that employers used to discuss benefits of the DHP and could not account for this in our analyses. Finally, the cost savings associated with the DHP were relatively modest for the generic medications included in this analysis, and similar studies with more expensive brand medications may have been more likely to show differences in adherence. 37 Others have hypothesized about the role of scarcity because individuals with lower incomes have more financial stressors, may not have the bandwidth to prioritize health, and may not be influenced by a VBID structure. 38 Consistent with this hypothesis, studies have found that colorectal screening rates and adherence to statins improved upon cost-sharing reduction in those with middle and high incomes, but there was not a detectable improvement among those with lower incomes. 28,39 Further studies should continue to include income as a variable in their analyses to clarify its relationship with response to VBID offerings.
There are methodological limitations that may explain why response to the DHP did not vary by income. First, our sample was limited to individuals with employer-based insurance. Owing to sample size, we were also limited to analyzing those with incomes less than $50,000 as the lowest-income group. In other words, we could not disaggregate those with smaller incomes. As a result, our study is not generalizable to those with very low incomes, who may have responded differently to a VBID such as the DHP. Furthermore, we did not have information on household

Conclusions
In conclusion, we did not observe a statistically significant change in individual-level medication adherence among beneficiaries with diabetes or prediabetes who were offered the DHP, a disease-specific value-based insurance plan. Moreover, we also did not observe a difference in effect across individuals with different income groups. These results highlight the challenges of translating VBIDs into changes in patient behavior that will ideally improve their health outcomes. Further research should be conducted to clarify additional strategies that can improve consumer engagement, with the goal of designing more targeted, impactful policies. income may impact the effectiveness of any VBID. 38,44,45 LIMITATIONS This study has several limitations, in addition to those previously described. Although we used propensity matching at the employer level and weighting at the individual level to minimize confounding, there may still be unmeasured differences between the populations (eg, employer-level "wellness culture," individual engagement). Patients within the matched control employers may have been more or less similar to each other as compared with patients of DHP employers, in terms of unmeasured individual characteristics related to medication-taking behavior. We also conducted a complete case analysis: this may have biased our results if this sample differs systematically from those with any missing data. However, only 11.8% of the sample was removed because of missing data. Our study also has several notable strengths. For example, we measured medication adherence 1 and 2 years after DHP implementation, allowing us to explore its short-term and long-term impacts. Additionally, unlike prior studies that used area-level income, we used an estimate of individual household-level income, allowing us to more precisely examine the effect of income on VBID effectiveness. 46,47 Altogether, this study may have implications for policies that aim to encourage the uptake of high-value, evidence-based therapies through thoughtful benefit designs that aim to promote equity across income levels, in combination with both patient behavioral interventions and provider interventions (eg, reducing clinical inertia in diabetes care). The lack of an observed adherence effect of the DHP suggests that modest reductions in consumer cost sharing may not be sufficient in the absence employer-level analysis does help mitigate against selection bias, it has been well established that relationships seen at the aggregate level do not necessarily transfer to the individual level and vice versa, resulting in inferences that are specific to the level of analysis used. [43][44][45] As with all analyses using administrative data, we were limited in our ability to measure individuallevel characteristics that are linked to medication-taking behaviors. For example, the DHP employers may have a wide range of medication-taking behaviors among their employees, and preferentially match to similar control employers, such that there may have been a subset of individuals in each group who benefited (or would have benefited) from VBIDs. This effect may have been obscured by the other individuals making up the larger sample.

DISCLOSURES
Differences in study design may also explain the lack of associations between the DHP and medication adherence in our study as compared with evaluations of other VBID programs that demonstrated positive findings. For example, one study reported no change in medication adherence among those with a VBID alone but a significant 6.5-percentage point increase in medication adherence among VBID patients who also participated in a disease management program. 40 We did not have details of disease management or wellness programs among our employers, precluding our ability to control for these variables or conduct subgroup analyses. Additionally, in contrast to prior studies that used area-level income measures, our study used a measure of individual household income that incorporated individual financial behavior. [40][41][42]46 Given the well-established relationship between income and adherence, it is possible that individual household