The analyses were conducted using de-identified Market Clarity data from Optum®. The data were obtained from administrative health insurance claims for members located across the USA. The approximately 17–19 million lives covered in this database were all insured via large commercial and Medicare Advantage health plans. The data were fully de-identified and Health Insurance Portability and Accountability Act (HIPAA) compliant. For this study, the dataset supplied longitudinal information on patient demographics, coverage eligibility, inpatient and outpatient services, outpatient prescription fills, payments, and laboratory test results. Data for this study covered the period from January 1, 2015 through June 30, 2021, and permission to access the data was obtained from Optum®. The data for this study are available from Optum® but restrictions apply to the availability of these data, which were used under license for the current study. Given the use of retrospective and de-identified data, ethics committee approval was not required.

Patients were required to have at least one recorded HbA1c result at any time from January 1, 2016 through July 1, 2020 (the identification window). For each patient, the date of the first such HbA1c result was identified as the index date. Patients were also required to have had T2D during the 12 months prior to the index date (the pre-period) based upon receipt of two or more diagnoses of T2D and no receipt of any diagnoses of type 1 diabetes. Patients were excluded from the analyses if they were younger than age 18 years on the index date or were diagnosed with gestational diabetes or pregnancy at any time from the start of the pre-period through 1 year after the index date (the post-period). The selection of a 1-year post-period is consistent with economic models which update risks and outcomes annually [9, 10]. Finally, in order to ensure complete records of diagnoses, costs, and resource utilization, all patients were required to be insured continually from the start of the pre-period through the end of the post-period.

The primary outcome of interest was annual all-cause total healthcare costs. Consistent with previous research, costs were calculated as the sum of standard costs, copayments, and deductibles [11, 12]. In addition, costs were subcategorized into outpatient, acute care (inpatient and emergency room), and drug costs. All cost measures were expressed as average per-patient annual costs in 2021 US dollars, as adjusted for inflation by the medical component of the Consumer Price Index [13].

Consistent with ADA guidelines, which suggest that a target of HbA1c < 7% is appropriate for many nonpregnant adults with T2D [7], recommended glycemic control was defined as HbA1c < 7%, and patients were grouped on the basis of whether or not they met that HbA1c target on the index date. The analyses employed propensity score matching and multivariable analyses which controlled for patient demographics and baseline clinical characteristics. Specifically, propensity score matching (PSM) was used to match the group with HbA1c < 7% to the cohort with index HbA1c ≥ 7%, utilizing a greedy nearest neighbor match without replacement and a specified caliper distance of 0.2 [14]. Covariates included in the PSM model were patient age, sex, race, ethnicity, region of residence, and year of index date. The final sample after PSM consisted of 59,830 patients. Figure 1 illustrates how each of the study inclusion and exclusion criteria affected sample size.

Study inclusion–exclusion criteria and sample size. AIdentification window time frame of January 1, 2016 through July 1, 2020 determined by the duration of the data (January 1, 2015–June 30, 2021) and the requirement of 1-year pre- and post-periods. BIdentification of type 2 diabetes (T2D) based upon receipt of two or more diagnoses of T2D and no receipt of any diagnoses of type 1 diabetes

Given the PSM matched cohort, multivariable analyses were used to examine the relationship between index HbA1c and all-cause healthcare costs. Specifically, given the skewed nature of cost data, generalized linear models (GLM) with gamma distribution and log link [15] were used to examine all-cause total costs, outpatient costs, and drug costs, while all-cause acute care costs were examined using a two-part model. In this two-part model, the first step examined the probability of having an acute care visit and the second step estimates acute care costs for individuals with such a visit [16]. In the multivariable models, costs were estimated using the method of recycled predictions, with standard errors calculated from 1000 bootstrap iterations [17].

The multivariable analyses of all-cause costs controlled for patient demographics and pre-period characteristics, including general health, comorbidities, resource utilization, and medication use. Patient demographic information included age, sex, race, ethnicity, and region of residence. Patient general health and comorbidities were measured using the Diabetes Complications Severity Index (DCSI) and the adjusted Charlson Comorbidity Index (CCI). The DCSI is scored on a scale of 0–13, with higher scores assigned to patients with a larger number and/or more severe levels of the following diagnoses: retinopathy, neuropathy, nephropathy, cerebrovascular disease, cardiovascular disease, peripheral vascular disease, and metabolic disease [18]. Meanwhile, the CCI creates a composite morbidity score that reflects mortality risk based upon the presence of any of 19 comorbidities, with individual comorbidities given a score between 1 and 6 [19]. In this study, myocardial infarction, peripheral vascular disease, and diabetes with or without complications were omitted from the calculation of the adjusted CCI either because they applied to all individuals (e.g., diabetes) or because they were included in the DCSI. For example, the CCI condition of myocardial infarction was omitted in our measurement of the CCI since it is captured as a cardiovascular complication in the DCSI. In addition, the comorbidities of anxiety and depression, which are not included in either the CCI or DCSI, were included as covariates. Resource utilization was measured by indicator variables capturing whether the individual visited a cardiologist, endocrinologist, nephrologist, or ophthalmologist in the pre-period, as well as the number of pre-period visits to a family practitioner or internist. Pre-period medication use was used as an additional proxy for disease severity and overall health and was measured by the number of classes of prescriptions filled for insulin, non-insulin glucose-lowering agents (GLAs), and non-GLA medication prescriptions filled. Insulin was subgrouped into basal, bolus and premixed classes, while non-insulin GLAs consisted of alpha-glucosidase inhibitors, amylin analogues, dipeptidyl peptidase IV inhibitors, glucagon-like peptide 1 receptor agonists, meglitinides, metformin, sulfonylureas, sodium-glucose cotransporter 2 inhibitors, and thiazolidinediones.

In addition to the multivariable analyses, unadjusted descriptive statistics were summarized for the cohort across index HbA1c thresholds (< 7% or ≥ 7%) post matching. Differences in continuous variables across groups were examined using t statistics, while differences in categorical variables were examined using chi-square statistics. All analyses were conducted using SAS, version 9.4 (Cary, NC), and a P value < 0.05 was considered, a priori, to be statistically significant.