Geographic access to United States SARS-CoV-2 testing sites highlights healthcare disparities and may bias transmission estimates

Uniform access to SARS-CoV-2 testing is crucial for controlling the COVID-19 epidemic.¹ Lack of testing can result in the epidemic spreading undetected² and increase the risk of extensive local transmission. The USA has been slow to develop reliable diagnostic tests and, while there has been recent improvement in testing capabilities,³ large-scale testing remains a serious concern.

Inequalities in geographic accessibility to healthcare in the USA have been documented to cause negative health outcomes for seasonal influenza transmission and other diseases.⁴ Further, travel time negatively impacts healthcare-seeking behaviour.⁵ The deployment of SARS-CoV-2 testing within existing medical infrastructure, while logistically efficient, may exacerbate this disparity in health outcomes⁶ and underestimate disease burden in disadvantaged populations.

Geographic accessibility to SARS-CoV-2 testing sites, to our knowledge, has not been systematically quantified. Therefore, we evaluated whether testing sites were equally accessible to populations across the USA, leveraging two public SARS-CoV-2 testing site datasets and a high-resolution map of travel times.

American Community Survey (2014–2018) data for contiguous US states were used to tabulate county-level covariates including population, population density (), median income, percent uninsured and percent minority ().

A national database of SARS-CoV-2 testing sites was curated using the Carbon Health (N = 5376) and CodersAgainstCovid (N = 1547) datasets (accessed 7th April 2020). Carbon Health (carbonhealth.com/covid-19-testing-centers) prospectively called urgent care centers and hospitals on publicly listed telephone numbers starting 17th March 2020 to ask whether SARS-CoV-19 testing was being offered. Additionally, a verified, non-exhaustive collection of publicly documented and user-entered testing sites were included. CodersAgainstCovid identified urgent care centers, hospitals, drive-throughs, health departments and other facility types prospectively starting 15 March 2020, through volunteer-verified ‘webscraping’ and crowdsourcing (https://codersagainstcovid.org/).

We identified and geocoded (R v.3.6.2 ggmap v3.0.0) 6236 unique sites (687 excluded following manual de-duplication and cleaning). Related site ontologies were collapsed into meta-ontologies (e.g. Urgent with Immediate Care). To date, this is the largest database of US testing sites known to the authors. To evaluate completeness (as of 20 April 2020), we identified public testing sites listed in sample areas: 34 in Illinois (https://www.dph.illinois.gov/covid19/covid-19-testing-sites), five in Colorado (https://covid19.colorado.gov/testing-covid-19) and 104 in West Virginia (https://www.wvhealthconnection.com/covid-19). Our database included 169, 85 and 60 sites in each area, respectively. We confirmed our database identified at least one site in every city in Texas operating a drive-through (https://www.dshs.state.tx.us/coronavirus/testing.aspx).

We used published friction-based travel times⁷ between ~1 km² gridded cells in the USA, accounting for topography and the most efficient non-air travel method. Median travel times for the shortest path to testing sites across all grid cells in each county (N = 3108) were calculated using the Dijkstra’s algorithm.⁸

Generalized linear models (R stats v3.6.2) were used to estimate the correlation of population density, percent minority, percent uninsured and median income on median travel time, by county. We also tested for potential interactions between population density and percent minority or percent uninsured. Influential counties with a Cook’s distance measure over 4/N were excluded (up to N = 175).

We collated 6236 SARS-Cov-2 testing sites in the contiguous US states. Testing sites (Supplementary Table 1) were often affiliated with medical centers (43%) and urgent care (47%) and were infrequently drive-through (3%). Testing sites were spatially clustered (Moran’s I = 0.037; z = 61.4; P < 10⁻⁵), around US urban centers (Supplementary Figure 1).

The travel time from each 1 km² grid cell to the nearest US testing site is spatially heterogeneous at the national and state level (Figure 1A–C). Thirty percent of the population live in a county (N = 1920) with a median travel time over 20 minutes, though with pronounced regional differences (Figure 1D) ranging from 5 to 86%.

Figure 1.

Distribution of SARS-CoV-2 testing sites. (A) Travel time to the nearest testing site per 1 km² area (shorter travel time in darker blue) in the 48 contiguous US states plus DC. (B) Travel time as in Panel A enlarged to show detail in the state of Texas. (C) Percent minority () by county in Texas. (D) Median travel time by county versus the cumulative population for each geographic region (excluding two outlier counties). Vertical dashed line at 20 minute median travel time. Horizontal dotted lines indicate cumulative population percentage in that region (in parenthesis) residing in counties with <20 minutes median travel time.

Population density, a determinant of population distribution, was associated with a shorter median county-level travel time (Table 1). While controlling for population density as a potential confounder, percent minority was associated with an increase in travel time, as was percent uninsured. These associations remained when also adjusting for median income. We found a significant negative interaction between percent uninsured and population density (P < 0.01) suggesting that the disparity of longer rural travel times is greater in counties where a higher proportion of the population is uninsured. Percent minority and population density did not interact statistically.

Table 1.

Generalized linear regression models. Associations between covariates and median travel time in minutes by county in the 48 contiguous US states and DC

	Model 1	Model 2	Model 3	Model 4
Intercept	61.45***	59.99***	56.29***	51.36***
	[59.82, 63.08]	[58.34, 61.64]	[54.17, 58.41]	[47.70, 55.03]
Log of population density	−13.41***	−14.14***	−12.94***	−14.13***
	[−14.02, −12.79]	[−14.76, −13.52]	[−13.56, −12.32]	[−14.78, −13.47]
Percent minority (%)		0.15***		0.13***
		[0.12, 0.18]		[0.10, 0.17]
Percent uninsured (%)			0.41***	0.23**
			[0.30, 0.53]	[0.09, 0.38]
Median income ($10 000)				2.52***
				[1.46, 3.59]
N	2942	2934	2942	2931
AIC	24 321.32	24 192.59	24 291.55	24 097.61
Pseudo R2	0.38	0.41	0.39	0.41

^*** P < 0.001; ^**P < 0.01; ^*P < 0.05.

Using two large, national datasets of SARS-CoV-2 testing sites paired with estimates of travel times, we demonstrate an uneven distribution of critical public health resources. The testing site distribution recapitulates structural disparities, including inequities among minority, uninsured and rural groups, which may further perpetuate disparities as the pandemic progresses. Differential accessibility to testing may lead to biases in estimation of disease incidence and potentially delay identification of COVID-19 hotspots. In the absence of representative testing, syndromic surveillance tools may provide early warning signals, and augment targeted-testing and other public health interventions.

Despite efforts to ensure comprehensiveness, in some regions, our dataset may be missing testing sites (e.g. West Virginia). While some additional testing sites have been created, given recent difficulties scaling up, we believe our database remains representative.⁹ There remains potential for differential missingness of sites in areas with reduced ‘webscraping’ visibility or sites specifically placed to address inaccessibility. Nevertheless, this work highlights the need for comprehensive resources and the utility of data sharing during a pandemic.

The travel-time metric used here accounts for the presence of public transportation and routine traffic. Early evidence shows widespread variability in mobility reductions during the epidemic.¹⁰ Our estimates of differential access present a conservative picture of inequality in the USA, which may be worse if public transit closures and private transportation were also modelled. Additionally, our models do not examine other, non-geographic barriers to SARS-CoV-2 testing access (e.g. economic), nor geography for residents in Alaska and Hawaii. Travel time, for example, is shorter for urban uninsured minority groups, and therefore does not explain the below average testing rates in disadvantaged urban areas (e.g. Philadelphia).

In summary, reduced geographic access to SARS-CoV-2 testing sites is associated with sociodemographic factors that, in turn, are linked to poor structural access to care and health outcomes. The location of future testing sites should explicitly account for travel time and sociodemographic predictors, in addition to other public health testing requirements.

Author Contributions

B.R., C.M.A., J.S.B. and M.U.G.K. contributed to conceptualization. B.R., K.T.L.S. and K.S. contributed to data acquisition. B.R., C.M.A. and K.T.L.S. contributed to the data analysis. All authors contributed to the interpretation of results and manuscript writing.

Data and Code Availability

Median travel time to testing center by county, sociodemographic variables, code and raster for main analysis: https://figshare.com/s/4b2af17d00e4751685c5. For access to the Carbon Health dataset, please email: coviddata@carbonhealth.com. For access to the CodersAgainstCOVID dataset, visit: http://github.com/codersagainstcovid.org.

Funding

This work was supported by Google.org and the Tides Foundation [TF2003–089662], the National Institutes of Health [K23 DK120899 to C.M.A.], Boston Children’s Hospital Office of Faculty Development Career Development Award [to C.M.A.], the Monitoring Outbreak events for Disease surveillance in a data science context project [to M.U.G.K.] and a Branco Weiss Fellowship [to M.U.G.K.]. The funding bodies had no role in the design and conduct of the study; collection, management, analysis and interpretation of the data; preparation, review or approval of the manuscript; and decision to submit the manuscript for publication. All authors have seen and approved the manuscript.

Conflict of interest

None declared.