Overview
The Dartmouth Atlas Project has long based its analyses on the natural markets where residents of the United States receive their care. The 306 hospital referral regions (HRRs) have become a widely-used standard for health care policy and research because they correspond to local travel patterns, which often cross county or state lines.
We applied these methods to the COVID-19 epidemic, using the county-level data that are being collected, organized, and updated daily by The New York Times. County case and death rates were aggregated to the 306 hospital referral regions. This HRR-mapped COVID-19 dataset covers data from January 2020 through October 2021. This dataset includes the following:
- population-based rates of reported COVID-19 cases
- population-based rates of reported deaths from COVID-19
- average daily growth rates over the past seven days
- new COVID-19 cases for past fourteen days
- new COVID-19 deaths for the past fourteen days
These maps and data provide 4 helpful insights:
- They account for differences in population size across states and regions, revealing the proportion (rather than the absolute number) of the population that has been diagnosed with COVID-19.
- They reveal differences across regions within states, which can be dramatic, particularly in large states such as California and Texas.
- Many counties are very small. A substantial number have no cases, leaving the false impression that the virus is not present in the region. Small numbers provide statistically imprecise estimates of prevalence. HRRs improve statistical precision and reveal that COVID-19 is present in every HRR.
- As HRRs were constructed originally to measure health care catchment areas, they provide a more accurate tool than administrative boundaries for assessing the relative capacity of local health care systems. Data on hospital capacity, easily linked by HRR, can be found at https://globalepidemics.org/our-data/hospital-capacity/ .
This dataset was used to create the maps shown on the Dartmouth Atlas of Health Care website, https://www.dartmouthatlas.org/covid-19/hrr-mapping. Screenshots of these maps are included here.
Methods
Data for these analyses were drawn from two sources, The New York Times daily update of county case and death counts (https://github.com/nytimes/covid-19-data), and the MABLE datafile from the Missouri Census Data Center (geocorr2014.csv), which provides a crosswalk between each U.S. county and the Dartmouth Atlas of Health Care-defined hospital referral regions. This crosswalk file reports the total population of each county and the number of residents of that county who reside within a given HRR, allowing the proportion of residents of a county within a given HRR to be determined.
The New York Times case and death counts were used to calculate the county rates based on population data from the U.S. Census. HRR rates were then calculated as the weighted average of county-level rates, with the weight equal to the proportion of residents of each county within a given HRR. Growth rates were calculated using the following formula:
Average growth rate = ((Rate today / Rate 7 days ago)(1/7)) - 1
Mapping county to HRR and associated calculated rates mentioned were done using an R script. Everything to replicate the HRR-mapped COVID-19 dataset is included here:
- cumulativecountytohrr.R (R script that creates HRR-mapped dataset)
- us-counties.csv (copy of NYT dataset, through October 2021)
- geocorr2014.csv (MABLE datafile)
- CasesandDeathsbyHRR.csv (sample HRR-mapped output; used for our maps)
Limitations
Alignment of Counties and HRRs
Although many counties fall within HRRs, some span HRR boundaries. The HRR rates are thus an estimate and will only be exactly correct when the constituent counties align with the HRR, or when the per-capita rates are similar across counties. We calculated an accuracy index to let readers understand how strongly these align for each HRR. This was calculated as the weighted average of the proportions of the county populations that fall within an HRR, where the weight is the ratio of the “in-HRR” county population to the total HRR population. A value of 1 means perfect alignment; a value closer to 0 implies that there is less confidence in the estimate. 90% of HRRs have accuracy index values over 0.75. The four HRRs with values under 0.25 are Evanston, IL (0.16), Blue Island, IL (0.15), Palm Springs, CA (0.11), and Sun City, AZ (0.08). The download file includes this measure.
Population Differences
Counties vary in the characteristics of their populations, including differences in age, gender, race, and comorbidities. Because we did not know the demographic or health characteristics of the cases and deaths, we could not adjust for these differences. It is likely that these differences had less influence on the spread of the disease and much more influence on case fatality rates, as was made clear in a recent Health Affairs Blog using Dartmouth HRRs.
Limitations of The New York Times Data
The New York Times data are based on the efforts of multiple journalists working to analyze data releases from states and local health departments, seeking to clarify the data, and correcting and updating the data wherever possible. Additional details on their methods can be found in their github repository’s README file. Their files are updated daily, with corrections to earlier time periods with each update. The Atlas calculations were rerun each day, including the corrected data for prior days.
Looking Back vs Looking Forward
The calculation of the growth rates was based on the prior seven days and should not be assumed to predict future rates.