1. The business question?
When the pandemic forced healthcare online in 2020, telehealth stopped being a convenience and became the only way many Medicare patients could see a doctor. Five years later, the emergency has passed, and usage has settled at a new level.
The interesting question is not how far adoption fell. It is who stayed.
If telehealth removed barriers, the patients who face the most barriers should have gained the most and held on the longest. If it mostly served people who already had good access, then a technology sold as an equalizer quietly did the opposite. A national average cannot answer this, because an average is exactly the place where a disparity goes to hide.
2. The data
The Centers for Medicare and Medicaid Services publishes a Medicare Telehealth Trends dataset covering utilization from January 2020 through December 2025. The extract used here contains 36,120 rows across 13 columns, spanning all 50 states and breaking utilization down by race and ethnicity, dual eligibility status, rural and urban residence, age, and sex.
It is administrative claims data rather than survey data, which matters. These are visits that actually happened and were billed, not visits people reported in an interview. The tradeoff is that the dataset shows behaviour without explaining it. It can establish that rural adoption is lower. It cannot establish why. Every causal claim below is therefore argued from the published literature rather than from these claims alone, and I have flagged where the evidence is strong and where it is contested.
The method
Extraction and shaping in MySQL, statistical analysis in Python, visualization in Tableau.
The SQL layer handled cleaning, type casting, and the aggregations feeding each view: national totals by year, state rankings, and utilization by demographic segment. Python handled year-over-year change calculations, segment gap measurements, and checks on whether observed gaps held consistently across the period or moved. Tableau carried the six views making up the final dashboard.
The analytical choice worth naming is that every finding compares segments against each other within the same year, not against the national line. Comparing a group to the national average tells you how far it sits from the middle. Comparing groups to each other tells you the size of the gap, which is the number that matters if you are trying to close one.
Finding 1: a new floor, not a failure
Adoption peaked at 47.9% in 2020 and settled at 23.2% by 2025.
The steepest fall came immediately, a 13.71 percentage-point drop between 2020 and 2021, after which the decline flattened. This is the shape of a temporary substitution unwinding, not a technology being rejected. Roughly half of peak usage stuck.
That residual is the real finding, and it is consistent with the wider picture. KFF's analysis of Medicare telehealth coverage reports that although use has fallen from its early-pandemic peak, it remains close to double the pre-pandemic level (KFF, 2026). The correct read of 23.2% is not a decline. It is a new permanent floor, roughly twice the pre-pandemic baseline.
Finding 2: dual eligible patients outperform, and the literature agrees
Dual-eligible patients used telehealth at a rate 13 percentage points higher than Medicare-only patients in 2025.
This runs against intuition. Patients eligible for both Medicare and Medicaid are by definition lower income and face more of the barriers usually blamed for low technology adoption. They used telehealth more, and they kept using it.
The finding replicates. KFF reports 2024 telehealth use of 35% among dual eligible beneficiaries against 23% among those not Medicaid eligible, a 12 point gap sitting almost exactly on top of the 13 points measured here (KFF, 2026). A national Medicare claims analysis covering 2019 and 2020 found dually eligible beneficiaries more likely to use telehealth than non-dual beneficiaries in both years (Bogulski et al., 2024, PMID 38314738). Work on Medicare Advantage populations reached the same conclusion, finding higher primary care telehealth use among beneficiaries with low-income status, disability, or frailty than among those without those access challenges (Boudreau et al., 2024).
The mechanism most consistent with this is that dual eligible patients are not a low-access group in every sense. They carry higher clinical need, have more frequent contact with the system, and more often have care coordination attached to them. Where a care coordinator exists, telehealth has someone to schedule it. Connection to the system predicts adoption better than income does.
One caution. The evidence is not unanimous. A difference-in-differences study of Medicare beneficiaries inside a single Accountable Care Organization found the opposite, with lower telehealth use among dually eligible patients (Cao et al., 2021, PMC8679021). That study looked at one ACO rather than a national population, which likely explains the divergence, but it is worth stating that the direction of this effect is not settled at every level of analysis.
Finding 3: the rural gap is infrastructure, not willingness
Rural patients ran roughly 7 percentage points below urban patients for the entire period. The gap did not open during the surge and did not close during the decline. It was there at the peak, and it was there at the plateau.
A gap that stable is structural, and the literature identifies the structure with unusual clarity.
The strongest evidence comes from a county-level ecological study merging broadband capacity data with Medicare Fee-for-Service telehealth utilization across 3,107 US counties. Counties in the highest quintile of broadband availability showed 47% higher telehealth utilization than counties in the lowest quintile. In the adjusted model, a one standard deviation increase in broadband access was associated with a 1.54 percentage point increase in utilization, while rural county designation was independently associated with a 1.96 percentage point decrease (Pandit et al., 2025).
That last detail matters. Rurality remained a negative predictor even after controlling for broadband, meaning connectivity explains part of the rural gap but not all of it. Provider supply is the likely remainder. Fewer than 10% of US physicians practise in rural communities where roughly a quarter of Americans live, and researchers have argued that inadequate rural broadband prevents telemedicine from compensating for exactly that shortage (National Rural Health Association, 2024).
The willingness question has been tested directly and answered. A nationally representative survey study concluded that disparities in telehealth access, rather than differences in willingness to use the services, most plausibly explain rural telehealth gaps (Ko et al., 2023;39(3):617-624).
This reframes the problem. If the binding constraint is infrastructure rather than attitude, then outreach campaigns aimed at persuading rural patients are spending effort on a variable that is not binding.
Finding 4: American Indian and Alaska Native patients ended lowest
This group finished at 21.1% in 2025, the lowest of any racial or ethnic category in the dataset.
The same county-level broadband study found that a one standard deviation increase in the proportion of Native American and Pacific Islander residents was independently associated with a 0.59 percentage point decrease in telehealth utilization, after adjusting for broadband and rurality (Pandit et al., 2025). Separate analysis of adoption trends found telehealth uptake remained low in Alaska and South Dakota, both states with large Native populations, and argued that meaningful expansion would require substantial infrastructure investment rather than programme effort alone (Chartis, 2022).
Research on American Indian and Alaska Native veterans found significant rural and urban differences in video telehealth use within that population during rapid mental health care virtualization, pointing to geographic and infrastructural barriers rather than demand (Kusters et al., 2023, JAMA Psychiatry;80(10):1055-1060).
The combination is the concerning part. This population carries elevated rates of the chronic conditions telehealth manages well, and it has the lowest access to the modality.
Finding 5: Washington looks average and is not.
Washington sits at 23.4%, effectively at the national average, ranked 18th of 54.
Broken down, the state is not average at all. Hispanic patients in Washington used telehealth at 17.9%. Black and African American patients used it at 25.8%. That is a spread of nearly 8 percentage points inside a state whose headline number suggests nothing is happening.
The Hispanic figure aligns with a well-documented national pattern. Among Medicare beneficiaries, Hispanic and Latino individuals have been found roughly 35% less likely than White beneficiaries to use telehealth, and language is a leading candidate mechanism, with about 28% of the US Hispanic population having limited English proficiency (Yoo et al., 2025).
Recent work isolates the mechanism further. Using 2023 and 2024 National Health Interview Survey data, researchers found that digital health literacy mediated close to half of the difference in telehealth use between Latino adults preferring a non-English language and those preferring English, identifying digital health literacy as a modifiable point of intervention rather than a fixed characteristic (Linares et al., 2026).
The Washington Black and African American figure is the more interesting half. National studies generally find lower telehealth use among Black patients than White patients, including roughly 29% lower odds among Medicare beneficiaries. But the same literature notes that when digital access barriers are controlled for, Black patients have shown comparable or higher telehealth engagement in primary care settings (Yoo et al., 2025). Washington's relatively high figure is consistent with an environment where access barriers are lower, which is a testable proposition rather than a conclusion.
Statistical Tests
Three statistical tests were conducted to validate the key findings. A two-sample t-test confirmed the rural-urban adoption gap is statistically significant (t = -8.01, p < 0.001). The gap between dual eligible and Medicare Only patients is even more pronounced (t = 27.67, p < 0.001), the strongest signal in the entire analysis. A Pearson correlation between state-level adoption and Medicare Part B enrollment size returned r = 0.43 (p = 0.001), confirming a moderate positive relationship, larger states tend toward higher adoption, though state size alone explains less than half the variation, suggesting broadband access, demographics, and state policy play equally important roles.
What follows from this?
1. Fund rural telehealth as broadband policy.
The association between county broadband capacity and telehealth utilization is strong enough that connectivity investment should be treated as a health access intervention and evaluated on health access terms. However, because rurality remained significant even after controlling for broadband, connectivity alone will not close the gap. Provider supply must be addressed alongside infrastructure investment.
2. Study what dual eligibility does differently.
Dual eligible patients are the only structurally disadvantaged group in this dataset that consistently outperforms their counterparts, and the result replicates across independent national datasets. The most plausible mechanism is care coordination. That mechanism is worth isolating and extending to Medicare Only patients with comparable clinical need.
3. Treat digital health literacy as the primary lever for language-based gaps.
Evidence that digital health literacy mediates roughly half the disparity for non-English-preferring Latino adults makes it a more tractable intervention target than language access alone. Literacy programmes should be prioritised ahead of translation services as the first line of response.
4. Report below the state line.
Washington State is the clearest argument for sub-state reporting. State-level data showed a system performing at the national average while an 8 percentage point internal racial gap went unrecorded. Equity monitoring that stops at the state boundary will continue missing the disparities it exists to detect.
5. Anchor planning to 23.2%, not to the peak.
Capacity and reimbursement models built on 2020 adoption figures will systematically overestimate demand. Models built on the current baseline of 23.2% will not. The post-pandemic floor is the correct planning anchor for any forward-looking telehealth investment decision.
Limitations
Claims data shows utilization, not need, so a low figure cannot be separated into unwillingness and inability. The broadband and care coordination explanations are drawn from the published literature and are associations rather than demonstrated causes in this dataset. The county-level broadband study is ecological, meaning its findings apply to counties rather than to individuals, and it covers 2020 specifically. Demographic categories follow CMS definitions and carry the known reporting inconsistencies of administrative race and ethnicity fields, which likely understate some groups, particularly American Indian and Alaska Native populations. The 2025 figures reflect data available at time of analysis and may be revised.
The full interactive dashboard and analysis will be available on my Tableau Public profile and GitHub repository soon.
Let's connect
Contacts
1-206-476-0218
sineke@dennyanddata.com
