The Census Bureau has begun processing the data collected for the 2020 Census. Data collection for the decennial census is always a herculean task and 2020 was no exception. To the usual list of challenges were added the COVID-19 pandemic, hurricanes, wildfires, civil unrest, and a condensed schedule.  Through the hard work of thousands of dedicated employees, both temporary and permanent, the Census Bureau was able to overcome these challenges and count nationally 99.98% of addresses included in the 2020 Census. Fundamental to this operational success was the hugely successful deployment of new technologies to aid the count that yielded a more accurate address list from which to work, and tools that dramatically increased the productivity of our enumerators in the field. Some have suggested that these challenges would lead to an inaccurate or even a failed census. Our analysis of the data is just underway, and as with all prior censuses we’re seeing and working through data quality issues as we prepare the data for tabulation. As I discuss further below, some issues appear to be pandemic related, but most are what we experience with every decennial census and other Census Bureau surveys. Importantly, we’ve not uncovered anything so far that would suggest that the 2020 Census will not be fit for its constitutional and statutory purposes.

The decennial census is comprised of many operations that require years of planning and coordination. Data collection commenced as planned in January 2020 with the Remote Alaska component of our Update Enumerate operations. This was also the time we were learning of the first U.S. case of COVID-19. We began collecting responses via the internet from most U.S. households on March 12, the day after the World Health Organization declared a pandemic and just a few days before the president issued a two-week stay-at-home order. Twice in March, we announced two-week delays in 2020 Census field operations including the onboarding and training of thousands of temporary enumerators. In April, we announced a plan to recommence field operations on June 1. This plan essentially shifted the original schedule by three months with field data collection ending on Oct. 31 rather than July 31. Because of this, we requested an extension to the deadlines for the delivery of statutorily mandated 2020 Census data products.

However, as the summer progressed it became clear that an extension would not be granted, forcing us to crash the schedule to find a way to comply with the Dec. 31, 2020, deadline for delivery of the apportionment counts. To accomplish this, we needed to adjust the plans for both field data collection and for post-collection data processing. The bulk of work for field data collection consists of sending enumerators to households that have not already self-responded. For the 2020 Census, over a third of the approximately 152 million residential addresses in the country required one or more visits by our enumerators. Put simply, the time needed to complete the Nonresponse Followup (NRFU) workload is a function of how many cases enumerators can complete each hour (i.e., enumerator productivity) and the total number of hours worked. In 2010, enumerators recorded responses on paper forms. In 2020, we deployed iPhones on which enumerators received optimally routed daily case assignments, entered responses, and tracked their hours and mileage expenses. From our 2018 Census Test in Providence, R.I., we knew that this technology upgrade substantially improved productivity such that we expected enumerators to complete 1.55 cases per hour versus 1.05 in 2010. Next, we deployed pay and bonus incentives designed to both get more enumerators hired and trained and to incentivize them to work more than the average of 19 hours per week. While we were somewhat successful in hiring enumerators to replace the usual attrition, we were less successful in encouraging folks to work additional hours per week. What did help, however, was that realized productivity was much higher than predicted. Our iPhone-enabled enumerators finished data collection by completing an average of 1.92 cases per hour — nearly double the productivity of their predecessors in 2010. 

Another key innovation area for the 2020 Census were improvements to our Master Address File (MAF) that featured strong partnerships with tribal, federal, state and local agencies and the use of modern Geographic Information Systems (GIS) to ensure we have a complete and accurate list of addresses to enumerate. Using GIS tools, we were able to reduce our expensive in-field Address Canvassing operation from 100% of addresses in 2010 to only 35% of addresses in 2020. Our accurate and updated MAF enabled us to make it easier for residents to respond through what we call “NonID response,” where a housing unit ID is not required to respond. We’ve received over 22 million such responses and have been able to match the address the respondents provided to the MAF in over 90% of the cases.

This successful deployment of technology was critical in enabling us to complete the count during the pandemic. Not only did technology make our enumerators more productive, it allowed us to adjust to conditions in the field much more efficiently than in the past. Being able to manage cases daily and message enumerators in real time allowed managers in our area census offices to adjust operations efficiently when confronted with many challenges mother nature threw at us during this year. It also aided our efforts to redeploy enumerators to areas lagging in the count. We redeployed over 26,000 enumerators outside their home areas and in many cases to different states. This included getting an extra 1,530 enumerators to Louisiana — hard hit by hurricanes Laura and Delta. This and the hard work of our incredibly dedicated field staff enabled us to get every state, the District of Columbia and Puerto Rico above the 99% completion rate and a national rate of 99.98%.

The processing of the data collected from a census is just as important to ensuring quality as is data collection itself and is a large and complex task on its own. It is during post-collection data processing that we ensure responses are coded to the correct location, that duplicate responses are removed, and that a final, complete universe of the U.S. population is created to which we can apply a variety of quality metrics. Prior to beginning post-collection data processing, we conducted an early review of the data as they were being collected, which was considerably more review than in past censuses. These early, real-time examinations of the data were not able to provide a definitive picture of the overall U.S. population, but did allow us to assess initial indicators of data quality and to identify and fix several data processing errors. This work is just starting and, as in past, will continue through the release of the results from our most rigorous and thorough quality assessment — the Post Enumeration Survey (PES). 

We’re seeing increased interest in these metrics from the media, stakeholders and professional organizations. The American Statistical Association issued a report discussing the metrics they would like to see, and I’ve reached out to them to discuss priorities and how they might assist the Census Bureau further. We are committed to publishing quality metrics as they become available, with the first major release coming in December with the release of our Demographic Analysis estimates that provide an independent assessment of the 2020 population counts.

It’s premature to definitively describe the quality of the 2020 Census or assess its fitness for use. But it is possible to look at some preliminary metrics to get a general sense of where we stand. I’ve already discussed one such metric — completion rates — which measures whether we can resolve the status of an address in our MAF and determine the number of residents living there. With a 99.98% completion rate, we know that we were able to get at least some basic information for nearly every known residential address in the country. Another metric is the self-response rate, which measures the percentage of addresses where a householder completed the census online, over the phone, or by returning a paper questionnaire. At 67%, the 2020 Census self-response rate is just higher than what we achieved in 2010. This is important as self-responses yield the most complete and accurate information.

Those households that do not self-respond are visited at least once by an enumerator. If our enumerators are unable to speak to a resident after one or more visits, we attempt to resolve that address with high quality administrative records that have accurate data for that address. This is new to the 2020 Census and has allowed us to accurately enumerate approximately 5.6% of the nation’s addresses. If we fail to contact a resident for an address that does not have high quality administrative data, after repeated visits, our enumerators will attempt to get basic information from a knowledgeable proxy such as a neighbor. The proxy rate is the share of all NRFU interviews conducted with proxy respondents. While still preliminary and subject to change as we continue to process and eliminate duplication of the data, the 2020 proxy rate of approximately 24% appears close to the 2010 rate. There is some evidence of pandemic impacts, however, as preliminary results show proxy rates in college towns are higher. Many residents of off-campus housing left town during the self-response phase that coincided with the first wave of the pandemic and were unlikely to return by the period when enumerators visited to follow up. To reiterate, the proxy rates are based upon early tabulations and a clearer, more definitive picture of the overall extent of proxy usage will emerge once all post-collection collection data processing has been completed.

A final early metric of quality is the completeness of responses. A challenge for all surveys is getting respondents to answer every question fully and accurately, and the decennial census is no exception.  Preliminary indications are that item nonresponse for questions on date of birth, sex, race and Hispanic origin are higher relative to 2010. Some observers were concerned that the compressed schedule would lead enumerators to accept more incomplete responses. However, we are seeing elevated item nonresponse rates for all response modes suggesting something else is at work. This will require additional analysis as processing progresses. But note the Census Bureau has well-established procedures for coping with missing items on the decennial census and its other surveys.

In summary, while there’s much more work to be done to assess the quality of the 2020 Census and its fitness for its constitutional and statutory uses, early indications are that, despite the many challenges that confronted in conducting the 2020 Census, we do not have any evidence yet of any unusual issues.  No census is perfect. The 2020 Census won’t be perfect either, but the imperfections seen thus far can be addressed in standard ways as in prior censuses.