A propensity score weighted difference-in-difference analysis was used to estimate the effect of telehealth use on healthcare utilization among patients with type 2 diabetes using de-identified Medicaid claims data from March 2019 to August 2021. This model mitigates selection bias and controls for observed factors related to telehealth use. Zip code level data from the 2019 American Community Survey (ACS) were linked to Medicaid claims based on residential 5-digit zip codes to acquire environmental characteristics.

We selected the sample using the de-identified Medicaid claims data. The sample was restricted to those who continuously enrolled in Louisiana Medicaid over the study period (March 2019 to August 2021) to avoid issues with compositional changes resulting from increased enrollment due to the COVID-19 pandemic. We excluded patients dual-eligible for both Medicare and Medicaid as we lack access to Medicare claims. Only beneficiaries with type 2 diabetes were included in this study. Type 1 diabetes and gestational diabetes were excluded using International Classification of Diseases (ICD) 10 codes.

We identified patients with type 2 diabetes according to a modified Surveillance, Prevention and Management of Diabetes Mellitus (SUPREME-DM) diabetes definition [19] because of the absence of laboratory results in Medicaid claims data: (1) one or more of the International Classification of Disease, Tenth Revision, Clinical Modification (ICD-10-CM) codes (E11.xx) for type 2 diabetes associated with inpatient encounters; (2) two or more ICD codes associated with outpatient encounters on different days within 2 years; (3) combination of any ICD codes and antidiabetic medications with outpatient encounters on different days within 2 years.

For comparison purposes, beneficiaries were segmented into two groups. The first group, termed as telehealth users or the treatment group, consisted of those with any telehealth-related claims since March 2020. Identification of these services utilized specific procedure codes appended with modifiers such as GT, GQ, 95 and a particular place of service code (02) from Louisiana Medicaid claims. All other Medicaid beneficiaries, with no telehealth claims during the stipulated period, were categorized as the control group or non-telehealth users. This binary classification was chosen to distinctly evaluate the broad impact of any telehealth usage against exclusive non-telehealth utilization. The dates of the first telehealth encountered were coded as the initiation dates (index dates). We then assigned index dates randomly for non-telehealth beneficiaries (control group) based on the distribution of initiation dates in the treated population. The baseline period was 12 months before the index date, and the selected cohort also required at least 6 months of follow-up after the index date to ensure a sufficient evaluation period. For example, patients who had their first telehealth service after February 2021 were not included in this study because they had < 6 months of follow-up data. We also required all beneficiaries to have had at least one outpatient visit at the baseline. Beneficiaries without available zip codes to be linked with ACS data were finally excluded. The sample selection is shown in Figure S1 (supplementary material).

In assessing healthcare utilization, our primary focus was on the visit frequency of various healthcare services. This encompassed outpatient visits, which included in-person visits with claim types of outpatient hospitals, clinics and physician services. Additionally, we considered inpatient visits, Emergency Department (ED) visits—which were identified in accordance with the Healthcare Effectiveness Data and Information Set measures [20]—and hemoglobin A1C (HbA1C) tests. For each of these categories, the frequency was represented as an average number of visits per month, with separate calculations for the periods before and after the designated index date during the COVID-19 pandemic.

Our study also assessed secondary outcomes. Specifically, we considered ED and inpatient visits associated with Major Adverse Cardiovascular Events (MACE) and Ambulatory Care Sensitive Conditions (ACSC), given their potential preventability with improved primary care access.

To ensure a comprehensive assessment, we incorporated various control variables in our analysis. This included demographic factors like age, sex, and race/ethnicity. We also considered baseline healthcare utilization, represented by monthly utilization of ED visits, hospital stays, outpatient visits and HbA1C tests. Furthermore, we considered any existing comorbidities and evaluated technological accessibility at the zip code level and use rates of computer, internet and telephone in beneficiaries' residential area.

We first used the propensity score weighting method to obtain a successful balance between treatment and control groups. To succeed in balancing, we used group-based trajectory modeling to categorize individuals into latent groups with similar patterns of outpatient visits over 12 months before the index dates. Once the best group-based trajectory model was chosen, measures of group membership were then incorporated as control variables in propensity score weighting protocols and regression models [21]. Detailed explanations of using group-based trajectory models can be found in prior work of other studies [13, 13,22,23,24,25,26,27,28,29]. We then estimated the propensity scores for receiving telehealth during the pandemic using a probit regression model, controlling for baseline characteristics and a binary indicator of each trajectory group. For every weighting process, the standardized mean difference was checked between treatment and control groups before and after weighting to ensure successful weighting as defined as standardized mean differences within 10% for all baseline characteristics.

The outcomes of healthcare utilizations were estimated by the DID model as follows.

$$ Y_} = \alpha + \beta *Effect_} + Post_ + Telehealth_ + X_ + \delta _ + \partial _ + \varepsilon _},~w_ $$

The variable,\(_\), was an interaction between the time indicator (\(_\): 0 for pre and 1 for post) and the indicator of telehealth users (\(_\): 0 for non-telehealth users and 1 for telehealth users). The time indicator was a dichotomous variable, 0 for the pre-period (12 months before the index date) and 1 for the post-period (> = 6 months after the index date). \(\beta \) was the coefficient of interest and captured the change in outcomes in the pre- and post-period between Medicaid beneficiaries with and without telehealth use. The model also included zip code level fixed effect (\(_\)) and time fixed effect (\(_\)) of index to control unobservable zip code level differences and secular trends, respectively. Xi was a set of factors used in the propensity score weighting model, with the exception of the baseline outcome variables, used as control variables in our regressions to help control for additional variation that may remain after matching [30]. Weights \(,_\) were odds of treatment calculated based on propensity scores: wi = 1 for treated units and wi = propensity score/(1- propensity score) for untreated units. The standard error was clustered at the zip code level to account for common variances in these observations.

We identified multiple subgroups of beneficiaries to determine how the impact of telehealth uptake during the COVID-19 pandemic differed by zip code-level environmental characteristics related to telehealth. These characteristics included computer use rate, internet rate and telephone rate. The sample was divided into two subgroups separated by the median of each characteristic. Since telehealth is used as an approach to improve the healthcare access of rural residents, we then examined whether telehealth impacted differently across the rural status of the county where beneficiaries lived. Other subgroup analyses were performed in different age groups (≥ 50 versus < 50 years) and racial groups (Black versus non-Black).

With the lack of healthcare access due to the pandemic, patients with critically poor health could receive telehealth services for follow-up care after they had inpatient visits or emergency department visits. Therefore, we further performed the analysis after excluding those who had ED visits or hospitalizations within 30 days before the telehealth visit.

Telehealth was not exclusively expanded for diabetes care during the pandemic. A proportion of patients used telehealth services solely for non-diabetes care, such as mental health, which could attenuate any detectable impacts on diabetes-related outcomes in this study. We then repeated our main analysis only including patients who used telehealth services for any diabetes care in the treatment group.

The treatment sample may change with different treatment definitions and different outcome assessments; therefore, we repeated our matching process and regenerated propensity scores for non-telehealth beneficiaries to approximate the corresponding counterfactuals in each scenario. All data analyses were performed using SAS version 9.4 and Stata version 15.1 (StataCorp). Mean values were reported with standard deviations and regression coefficients were reported with 95% confidence intervals. Statistical significance was set at P < 0.05, and all tests were two-tailed.

The study was approved by the Tulane University Institutional Review Board (2021–1707). All data used in this study were deidentified; therefore, patient consent was not warranted. This study was conducted in accordance with the Helsinki Declaration of 1964 and its later amendments.