Self-Response Data Q&A

The 2020 rates are published daily by the Census Bureau at 3pm eastern (some days the data is available a bit before 3pm, other days soon after 3pm). The daily rates are available either as a CSV file or via the Census Bureau’s API. Information about the API specific to the response rates is available here for both the 2020 and 2010 rates.

One option with the API is that you can simply enter the following URL in a browser to download the rates:
https://api.census.gov/data/2020/dec/responserate?get=GEO_ID,RESP_DATE,DRRINT,DRRALL,CRRINT,CRRALL

You need a bit of patience, because this method downloads about 123,000 rows of data directly into your browser. But some days the API is updated before the CSV file is uploaded, so the URL can provide access to the latest rates sooner than later.

If you use the API, you can request data for specific geographies as well as specific fields in the table. The Bureau provides examples here.

If you use the downloadable CSV file, it contains the response rates for all levels of geography in a single file. You can use these codes to distinguish the rates for different geographic types:

• 0100000 - National level
• 0200000 - Census Bureau regions
• 0400000 - States
• 0500000 - Counties
• 0600000 - County subdivision
• 1400000 - Tracts
• 1600000 - Incorporated places (cities)
• 1700000 - Consolidated city
• 2500000 - American Indian Area/Native Hawaiian Area
• 2560000 - Tribal census tracts
• 5001600 - Congressional districts

Remember that the 2010 rates are “final” rates from the end of the self-response operation during the 2010 Census. These can be considered “goal post” rates; each community, county, city, and state could be working to boost 2020 rates to meet and surpass the 2010 final rate.

We have already downloaded the final 2010 rates using the Census Bureau’s API. These files do not change. They are available at this shared Dropbox directory, in CSV format.

Our team will be hosting an archive of the daily CSV files published by the Census Bureau, containing the self-response rates for each day that the Bureau publishes the data. The Bureau began publishing the data on March 20 and plans to continue for the duration of the self-response operation. We plan to upload the daily files to our archive for the same time period.

The archive is for informational purposes only. The official source of the 2020 Census self-response rates is the U.S. Census Bureau itself. You should plan to access the Census Bureau's data files directly, and only access our archive if you cannot obtain the response rates from the Census Bureau.

The archive is available as a shared Dropbox folder.

In early March 2020 we discovered that the Census Bureau planned to use “new” census tract boundaries in order to publish the 2020 (and 2010) self-response rates. To many (all?) groups outside the Census Bureau, this was unexpected. The Bureau apparently decided to use these new tract boundaries because these are eventually what the Bureau will use to publish the final 2020 population counts.

After several organizations explained to the Bureau that no one outside the Bureau had access to these new tracts, the Bureau created an FTP site from which you can download the data.

There are two zipped files at that link with the tract boundaries: “sr20_500k.zip” and “sr20_detailed.zip”. The “500k” file presents the tracts with generalized boundaries and omits tracts wholly in the water. The “detailed” file has more spatially precise boundaries and includes the water tracts.

The “rr_tract_rel.zip” file is a crosswalk between the “new” 2020 tracts and the current TIGER tracts. More on this below.

One thing to keep in mind with these new tract boundaries is that they may change at some point. The Bureau tells us that the tracts are currently being verified, and could change in the future. We have asked for clarification on whether they might change during the 2020 Census operation, but we have not heard back.

The differences between current TIGER tract boundaries and the new 2020 boundaries in some places are substantial. Nationwide the new 2020 file has about 84,000 tracts, while the current tract file everyone else is using has about 73,000 tracts. In other words, many current tracts have been split into multiple new tracts.

For example, in Co-Op City in the Bronx, this large, private housing development is currently included in one tract; see the image on the left from an early version of our Census 2020 HTC map. But the new tract file splits Co-Op City into 5 tracts, and none of them have the same ID as the current tract; see the image on the right from the Census Bureau’s Response Rate map:

NB: In order to present the 2020 and 2010 response rates on our HTC map consistently with how the Census Bureau is reporting the rates, we have updated the geography at our map to use these new 2020 tracts.

Not directly. Although most current tracts are unchanged in the new 2020 tract database, many have changed substantially (such as the Co-op City example above). That means that demographic data for the current tracts that have been changed will be very different for the new tracts.

If you wanted to produce an analysis that said “In tracts where 30% or more householders have no internet access, the 2020 response rates in Week 1 were X, compared to other tracts where the rates were Y”, your analysis in many parts of the country would be incomplete because it would omit any new tracts.

In order to help address this problem, the Census Bureau has produced a “crosswalk” file.

The idea behind the crosswalk is that it lists all the new tracts and provides the share of housing units from the current tracts that are in the overlapping parts of the new tracts. This enables you to convert current ACS estimates, for example, to the new 2020 tracts.

The crosswalk also provides housing unit percentages that enable you to allocate the current 2020 rates (or 2010 rates) that use the new 2020 tracts, back to current “2010-vintage” tract boundaries.

Remember, though, that either approach involves tradeoffs. If you allocate the ACS estimates to the new 2020 tracts, remember that the ACS data are estimates, and you may want to account for margins of error when doing this allocation.

If you allocate the 2020 rates back to the current tract boundaries, you will therefore be working with different rates than what the Census Bureau has published, which might be confusing to your audience.

Neither approach is ideal. But we are stuck with using one or the other due to the Bureau’s decision to use “new” tracts for the self-response rates.

You can’t reliably convert the LRS values, which are based on current tract geography, to the new 2020 tracts. The LRS documentation is clear that the LRS values are specific to the tracts for which they were calculated (partly because the LRS model uses tract-level demographic data from the 2010-vintage tracts).

You could use the crosswalk file to convert the 2020 rates back to the current tracts in order to compare with the Low Response Scores. But then you would have the problem of creating a different set of response rates than what the Bureau is publishing.

As it turns out, despite the name of the Bureau’s Low “Response” Scores, the LRS model does not predict self-response. According to the Bureau staff who created the LRS, it predicts “mail return rates”, which are different enough from response rates that it presents a problem to compare them.

The issue is how “return” and “response” rates are calculated. Both rates reflect the number of households that fill out the census questionnaire on their own. That number is then divided into a denominator to calculate the rate. But the denominators are different for “return” and “response”, and that’s the issue.

Response rates are the number of self-responding households divided by the universe of housing units in the Bureau’s “master address file” (MAF). Return rates are different, because the denominator is calculated by subtracting from the MAF any unit that is vacant, deleted after the count is done because they didn’t exist (demolished, etc), or considered by the Postal Service to be “undeliverable as addressed."

The denominator for response rates is therefore larger than return rates, which means that response rates are typically smaller than return rates. Also, response rates are calculated during the census. Return rates can only be calculated after the census (because at this point we don’t know yet how many units are vacant, etc).

The Low Response Score (despite its name) tries to predict the equivalent of the 2010 “mail return rate”. The predicted LRS values are higher than what we’ll likely be seeing with the 2020 response rates.

The 2010 self-response rate, which we now show at our HTC map, is the only metric that is directly comparable to the 2020 response rate.

Therefore, if your community is predicted to have a 73% return rate based on the LRS, you may want to also look at your community’s response rate from 2010 which is likely lower, and modify your expectations accordingly.

There are two components to this answer: one having to do with how self-response is measured, and the other having to do with how the census is conducted.

Regarding the self-response metric itself, the Census Bureau in 2010 used two different ways of measuring self-response during the census: response rates, and participation rates. These are different; they represent an apples-to-oranges comparison.

At the national level, the Bureau published day-to-day response rates, which can be compared directly with the daily self-response rates from 2020. You can find those in Appendix H in this Census Bureau memorandum [PDF].

At the local level, however, the Bureau did not publish response rates. Instead, they published participation rates.

Participation rates measure the number of households responding on their own to the census, but that number is divided into a different denominator than response rates. The denominator for response rates is the universe of all housing units. The denominator for participation rates removes from the denominator addresses considered “undeliverable” by the Postal Service. Therefore, participation rates are almost always higher than response rates. It would be misleading to compare the day-to-day participation rate trend for a given area in 2010 with the current day-to-day response rate trend.

The other reason you need to use caution in comparing where we are now with self-response rates compared with 2010 is because the difference in census operations between then and now, and of course because the nation was not in the midst of a pandemic in 2010 like we are now.

During a recent webinar about our HTC map, we were asked “What’s the typical census self-response rate at this point?” Our response was there is no “typical” response rate in general, and we are certainly not in a “typical” situation now.

Remember that census operations were different in 2010. which may impact the difference in the national trend. For example, there was no online option, there were only two mailings (rather than staggered waves of initial mailings and multiple follow up mailings in 2020), and of course there was no pandemic in 2010 that caused key census operations to be delayed and to completely upend Get Out the Count (GOTC) campaigns.

Nonetheless, stakeholders will be following the trajectory of the 2020 national self-response rate to monitor whether America will be on track to meet and beat its 2010 rate.

On a local level, perhaps a better way to evaluate how you’re doing today and on an ongoing basis, is to look back at the 2010 self-response rate for your area and strive to meet and beat that rate. And to compare how you area is doing compared with others nearby, or other similar areas across the state, and for larger regions such as counties and the state and nation as a whole.