Design and data sources
We conducted an interrupted time series analysis using data from the Veteran Affairs’ Corporate Data Warehouse, a national repository of clinical and administrative data, available through the VA Informatics and Computing Infrastructure. We used data from the VA Diabetes Risk (VADR) cohort; a large, established national cohort of 6,082,246 veterans seen for at least two primary care visits in any VA Medical Centers or VA Community Based Outpatient Clinics prior to January 2008, and with at least two additional visits between January 2008 and December 2016, and free of diabetes at cohort entry. The VADR cohort has been described in detail elsewhere [18]. Briefly, the cohort was created as a part of the Diabetes Location, Environmental Attributes, and Disparities (Diabetes LEAD) network, a collaboration between multiple academic institutions aiming to study the role of community level factors on diabetes incidence, funded by the Centers for Disease Control and Prevention (CDC) [19].
We longitudinally examined healthcare service utilization outcomes among veterans who developed a new diagnosis of diabetes during the VADR follow-up time period of January 1, 2008 through December 31, 2018. We defined incident diabetes as any of three criteria: (1) at least two encounters (inpatient or outpatient) with documentation of a diabetes ICD-9/10 code; or (2) a prescription for a diabetes medication other than metformin or acarbose alone; or (3) at least one encounter with a diabetes ICD-9/10 code and two elevated (≥6.5%) HbA1C test results. For the current analysis, in addition to these criteria, we required that the veterans in the analytic cohort have at least one encounter with the VA health system since March 1, 2018 to ensure adequate follow-up. As of December 31, 2018, 936,627 (15.6%) veterans were newly diagnosed with diabetes during a median cohort follow-up of 5.5 years. Of this population, the analytic sample for this study was 733,006 veterans with incident diabetes and at least 1 VA encounter between March 2018 and March 2020. The analytic sample was followed until March 2021 to assess the study outcomes.
Study variables
Exposure. We defined the COVID-19 pandemic onset as March 2020, when stay-at-home orders were issued nationwide [20], and compared outcomes before and after this month. As a secondary analysis, we examined the patterns of utilization across four phases in the first year of the pandemic; March 2018 through February 2020; March through June 2020; July through December 2020; and January through March 2021 [21].
Utilization Outcomes. We identified HbA1C test for all individuals in the cohort based on laboratory codes and laboratory testing data and determined monthly counts of HbA1C tests. HbA1C values outside a plausible range (< 3.1 or > 19.5) were excluded, as were observations in which an individual had multiple HbA1C tests recorded in a single day with values that were more than one point apart [22]. We assessed changes in routine care by capturing monthly rates of HbA1C tests performed per 1000 veterans in the analytic cohort. Denominators for computing monthly rates were fixed at the total number of veterans in the analytic cohort or within each subgroup for the subgroup analyses.
We extracted monthly counts of all outpatient visits to any VA facility and categorized these visits as in-person or telehealth (telephone or video) according to decision support identifiers (primary and secondary stop codes; Table S1). In a secondary analysis, we limited these stop codes to identify primary care visits only. Changes in visits were assessed by capturing monthly in-person and telehealth visits per 1000 veterans.
Finally, we extracted monthly counts of prescriptions that were filled or refilled in VA pharmacies for diabetes and hypertension medications based on annual VA national formulary [23] (Table S2). We calculated monthly rates of prescription fills per 1000 veterans separately for diabetes and hypertension medications. For computing monthly prescription fill rates, the denominator included veterans who filled their prescriptions, for each medications separately, in the VA pharmacy at least once between March 2018–2021.
Other measures
Pre-pandemic glycemia: Veterans were stratified based on their most recent HbA1c measurements prior to March 2020 to characterize their pre-pandemic glycemia in four strata: HbA1c < 5.7, 5.7–6.49, 6.5–8.9, and > 9.0 [24].
Community type: We defined four community types, measured at the census tract-level, using strata developed by the authors and others in the LEAD Network, described elsewhere [18]. Briefly, these community types are based on a modification of the Rural-Urban Commuting Area (RUCA) codes from the US Department of Agriculture. After collapsing the original 10 RUCA categories into three, we divided census tracts within urbanized areas into two categories based on land area. This resulted in four community type categories along the rural-urban continuum: high density urban, lower density urban, suburban/small town, and rural [19, 25]. Veterans were assigned community types based on the census tracts associated with their addresses when they entered the VADR cohort. Only those addresses that we were able to successfully geocode were assigned community types and included in the subgroup analysis.
Race/ethnicity: We used self-identified individual race/ethnicity to assess variations in utilization outcomes trends across racial/ethnic groups. Four race/ethnicity categories with sufficient data were considered for comparison: Non-Hispanic White, Non-Hispanic Black, Hispanic, and Non-Hispanic Asian-American Pacific-Islander & Native-Indian American (AAPI). Those with missing race/ethnicity information were dropped from this subgroup analysis.
Low income and disability: We created a low income/disability variable as a proxy for socioeconomic status using VA established priority groups, based on veterans’ military service history, disability, income, and eligibility for Medicaid or other VA benefits [26]. The low income/disability variable, categorized as, disabled, low income but non-disabled, and neither, was used to characterize veterans with and without medication fills.
Statistical analysis
We described the demographic characteristics of the veterans with diabetes in our analytic cohort. We used interrupted time series design [27], a quasi-experimental approach, to compare outcome trends at different phases of the pandemic. The interrupted time series design is an alternative to the randomized controlled designs, and leverages observational data and natural experiments in the event of an external shock, such as the COVID-19 pandemic, to assess impact of the shock. We visually assessed the assumptions of stationarity and lack of outliers, required by the design, by generating time series plots and visualizing the longitudinal trends.
Log-linear generalized additive regression model was used to compare the outcomes, rates of HbA1C tests, in-person visits (all outpatients and primary care only), telehealth visits (all outpatients and primary care only) and prescription fills (for diabetes and hypertension medications) with a binary variable indicating pre- and post-pandemic as the main exposure. Secular trends were modeled with a penalized cubic spline with the smoothing term selected using restricted maximum likelihood [28, 29]. To further ensure that there is no residual autocorrelation in the fitted model, we plotted the autocorrelation function of the residuals along with 95% confidence interval. In the subsequent analyses, we also fitted the regression models with a categorical exposure variable to indicate the three different phases in the first year of the pandemic, March–June 2020, July–December 2020, and January–March 2021, with pre-pandemic phase (March 2018–February 2020) as the reference category. Incidence rate ratios along with 95% confidence intervals were reported. Predicted monthly rates and 95% confidence intervals were also generated for each comparison phase.
Sub-group analyses
To assess whether disruptions in care due to COVID-19 had a differential impact among those with poor control compared to those with good control, we conducted analyses by stratifying the veterans based on their pre-pandemic glycemic control. To compare whether the associations differed by geography, we compared monthly rates of utilization outcomes stratified by community type among veterans with geocoded addresses. Finally, we also conducted the analyses stratified by self-identified race/ethnicity. All analyses were conducted using statistical software R (packages ‘mgcv’ and ‘ggeffects’). Statistical significance was two tailed at significance level of 0.05.