Data covering the period from 1 January 2011 through 21 December 2015 were obtained from the Truven Health MarketScan® Research Databases. Both the Commercial Claims and Encounters (CCAE) and Medicare Supplemental and Coordination of Benefits (MDCR) Databases were used in this study. The CCAE database consists of health insurance claims and patient enrollment data from large employers and health plans across the USA. The MDCR dataset represents health services received by employees, dependents, and retirees in the USA with primary or Medicare supplemental coverage. The CCAE and MDCR databases are generally representative of the population in the USA [18], and these databases provide detailed data on costs, medical resource use, and outcomes for healthcare services performed in both inpatient and outpatient settings. In addition to providing information about inpatient and outpatient services, medical claims are linked to outpatient prescription drug clams and person-level enrollment information. The data are fully de-identified and compliant with the Health Insurance Portability and Accountability Act (HIPAA). This article does not contain any new studies with human or animal subjects performed by any of the authors.
For entry into this study, patients were required to be identified as having T2D during the calendar year 2012. Such patients were identified if they received more diagnoses for T2D than for type 1 diabetes (T1D) and/or if they received a diagnosis for T2D and filled a prescription for an oral glucose-lowering agent (GLA) other than metformin or a sodium-glucose linked transport 2 inhibitor, since the latter agents may be used in the treatment of T1D [4, 19, 20]. Patients were furthermore required to have filled a prescription for basal insulin in 2012, with the date of the first fill identified as the index date. For each patient, data were included from 1 year before the index date (e.g., the pre-period) through 3 years after the index date (e.g., the post-period). Patients were excluded from the study if they: (1) filled a prescription for basal insulin during the pre-period, (2) were identified as pregnant at any time from the start of the pre-period through the end of the post-period, (3) were less than 18 years old as of the index date, (4) did not have continuous insurance coverage from the start of the pre-period through the end of the post-period, (5) did not have valid demographic data, or (6) received their index basal insulin prescription via mail order. The final sample consisted of 21,363 individuals. Figure 1 illustrates how each of the inclusion and exclusion criteria affected sample size.
The aim of the study was to examine the relationship between patient adherence to basal insulin as a class of therapeutics and outcomes, where adherence was proxied by the proportion of days covered (PDC). PDC was constructed as the percentage of days during the post-period that an individual had a supply of basal insulin, with adjustment for the possibility that insulin may not be used in a method consistent with the days’ supply field in a claims database [21]. A patient was considered to be adherent if the PDC reached the 80% threshold, and a patient was considered to be nonadherent if the PDC was < 80% [12, 15]. One advantage of using PDC as the measure of adherence is that it is used by both the Pharmacy Quality Alliance (PQA) and the Centers for Medicare and Medicaid Services (CMS) as a measure for examining the treatment of patients with diabetes [22, 23].
The patient outcomes of interest included medical costs, resource utilization, and acute complications. Costs were constructed using gross payments to a provider for a service, where payments were equal to the amount eligible for payment under the medical plan terms after applying rules such as discounts, but before applying coordination of benefits, copayments, and deductibles. All costs were converted to 2015 amounts using the medical component of the consumer price index. Both all-cause and diabetes-related costs were examined, and each of these cost categories was subcategorized to examine acute care (costs associated with hospitalizations or ER visits), outpatient costs, and prescription drug costs. All-cause costs consisted of all medical costs associated with inpatient, outpatient, and prescription drugs, and diabetes-related costs were constructed as the sum of all costs where (1) there was an accompanying primary or secondary diagnosis of diabetes; (2) there was receipt of a prescription for a GLA or diabetic supplies; or (3) the patient was identified as having hypoglycemia based upon a previously published and validated algorithm [24].
In addition to examining costs, the analyses also examined medical resource utilization and acute complications. The resource utilization outcomes examined included the probability of a hospitalization, the probability of an ER visit, the number of hospitalizations, the number of ER visits, and hospital length of stay (LOS). As with costs, resource utilization was categorized as either all-cause or diabetes-related. The study also examined the probability of being diagnosed with an acute complication over the 3-year post-period. An acute complication was identified based upon receipt of a primary or secondary diagnosis of hyperglycemia or diabetic coma or identification of hypoglycemia based upon the same algorithm used to identify the condition when constructing disease-specific costs [24].
When examining the relationship between adherence to basal insulin therapy and outcomes, the multivariable analyses controlled for factors that may potentially influence patient outcomes. Given the factors available in the database, the analyses controlled for patient demographic characteristics, general health and comorbidities, type of provider visited, medication use, and A1c tests. Patient demographic characteristics that were measured at the index date consisted of age, sex, region of residence, and insurance plan type. Patient overall general health was measured over the pre-period and assessed using the Charlson comorbidity index (CCI) [25], while, the Diabetes Complications Severity Index (DCSI) was utilized to proxy the severity of diabetes complications [26]. In addition to these index scores, the models also included pre-period comorbid diagnoses of anxiety, depression, and hyperlipidemia, since these diagnoses have been shown to be common in patients with diabetes or linked to patient nonadherence but were not captured in either of the two index scores [27, 28]. The analyses also controlled for visits to specialists in the pre-period with indicator variables for patient visits to a cardiologist, endocrinologist, ophthalmologist, or nephrologist, and they controlled for the type of basal insulin prescribed at the index date, the number of GLAs, and overall medication use. These factors were captured by an indicator variable if the index prescription was for U-500 basal insulin. As well, indicator variables captured the number of classes of GLAs prescribed in the pre-period and the number of non-GLA medications prescribed in the pre-period. Finally, the analyses controlled for the number of A1c tests the patient received in the pre-period. Consistent with previous research, this measure was used as a proxy for glycemic control [21], since laboratory results were unavailable in the database.
All multivariable models used the covariates discussed above, while the specific functional form depended upon the dependent variable being examined. In all cases, the estimated outcome was constructed from instrumental variables models, which are designed to allow for the role of unmeasured confounding. The use of retrospective data does not allow for patients to be randomized to treatment, potentially leading to sample selection bias. Instrumental variables are used to adjust for unmeasured confounding and the sample selection bias [29]. Consistent with previous research [30], the following variables were constructed over the first 30 days of the post-period and used as instruments: (1) the percentage of total payments for basal insulin prescriptions that were paid by coinsurance; and (2) the dollar (US) amount of copayments that were paid for basal insulin prescriptions. It is hypothesized that these variables will directly affect patient adherence while not being directly related to patient outcomes.
General linear models with gamma distribution and log link were used to estimate all costs models except those for acute care. Such general linear models have been shown to account for the skewed nature of cost data and to behave well in the estimation of population means of healthcare costs [31]. Acute care costs were estimated using a two-part model where, in the first part, a binary regression model was used to estimate the probability of having an acute care visit. In the second stage, a general linear model with gamma distribution and log link was used to estimate acute care costs for patients who had at least one acute care visit. The predicted probability of an acute care visit, estimated from the first part of the model, was then multiplied by costs estimates from the second part of the model to obtain unconditional average costs estimates.
Negative binomial models were used to examine the number of hospitalizations, the number of ER visits, and hospital LOS. Logistic models were used to examine the probability of being hospitalized or visiting the ER and the probability of being diagnosed with an acute complication.
Descriptive statistics were examined using Chi-square statistics for categorical variables and Kruskal–Wallis tests for continuous variables. For logistic models the odds ratios (ORs) and 95% confidence intervals (CIs) associated with being adherent (vs. nonadherent) were reported. For all other outcomes, the multivariable analyses were used to predict adjusted mean outcomes. Robust standard errors for each of these adjusted means were used to construct 95% confidence intervals for the adjusted means. Complete results of all multivariable analyses are presented in the supplementary material. All analyses were conducted using the SAS version 9.4 statistical software (SAS Institute, Cary, NC, USA). P values of < 0.05 were considered, a priori, to be statistically significant.