This graph shows the PES net coverage error estimates for the total state household population from 1990 to 2020.

The PES state estimates for 1990, 2000 and 2010 were all synthetic estimates. This means there were no state-specific effects in the models used to estimate net coverage error. In other words, the state net coverage error estimates were determined by:

• The characteristics of the people in the states.
• The national-level estimates of census coverage for these characteristics.

An impact of the synthetic estimation is that the state estimates were driven by national and regional trends. In 2020, the estimation models included state effects that reduced the synthetic bias but also increased the variability in the net coverage error estimates.

Let’s focus on Washington, D.C., represented in the graph by the dashed line. From 1990 to 2010, Washington, D.C., had a point estimate of negative net coverage error, though these estimated net undercounts were not necessarily statistically significant. This makes sense because of its large Black or African American and renter populations – two groups historically undercounted at the national level. The synthetic estimates in these years reflected those national-level coverage patterns.

These national coverage patterns persisted in 2020, but the models for 2020 included a state effect. Using these models, the net coverage error estimate for Washington, D.C., shows an overcount (which was not statistically significantly different from zero) in 2020. The state effect for Washington, D.C., clearly overpowered the national-level effects for race and homeownership. Perhaps the Black or African American population and renters interacted with the census differently here than in the rest of the United States.

While we believe this reduction in synthetic bias in the state estimates is an improvement in the methodology, the tradeoff is that some state estimates of net coverage error had large sampling errors. Thus, even at the state level, there is much uncertainty in the PES estimates. Using the PES results to adjust the population estimates base at the state level would infuse this uncertainty into the population estimates at a magnitude that could overtake the adjustment being made.

Furthermore, this change in methodology for the state estimates shows how different the synthetic estimates of census coverage and estimates that include state effects can be. Because the 2020 PES sample size was too small to support models with effects for substate geographies or subnational demographic groups, any estimates of census coverage for counties or subnational demographic groups would be subject to synthetic bias in addition to substantial sampling variability. This would make the resulting population estimates unacceptably imprecise.

Instead, we recommend continuing to use the PES to help inform the strengths and limitations of the 2020 Census counts. This also provides insight on the strengths and limitations of the population estimates to the extent that they are derived from the counts.

The Census Bureau should start planning early for a coverage measurement program that can adjust the population estimates base if there are large coverage errors in the 2030 Census.

The 2020 PES was never designed to produce results that could be used to adjust the official population estimates for coverage errors in the 2020 Census.

Since the 1990 Census, various technical limitations have led the Census Bureau to determine that the PES results – while informative at the national and state levels – were insufficiently robust for improving the population estimates base for lower levels of geography.

• After the 1990 Census, the Committee to Adjust the Postcensal Estimates (CAPE), a senior level group of Census Bureau statisticians and demographers, decided that an adjustment using the PES would improve the accuracy of the population estimates at the national level. But it recommended against using the PES to adjust the population estimates base because it was unclear if it would also improve the accuracy of the state and county estimates. CAPE’s findings are discussed in a 1993 Federal Register notice (download the pdf on the left of the Federal Register page; the discussion begins on p. 69 of the printed document, which is p. 77 of the pdf). The 1990 PES had a sample size of approximately 170,000 addresses and included both housing units and noninstitutional group quarters.
• The initial plan for the 2000 Census was to conduct a PES with a sample size of 750,000 housing units and use the results to adjust the census and the base for the population estimates for coverage errors. However, a U.S. Supreme Court decision in 1999 prevented the Census Bureau from using a sample statistic to adjust the census results for the purpose of apportionment. The plans for the 2000 PES were ultimately revised and the sample size was reduced to 300,000 housing units. The Supreme Court decision did not preclude the Census Bureau from adjusting the base for the intercensal population estimates. However, the Census Bureau decided to use the unadjusted census counts for the intercensal population estimates base, citing various technical concerns.
• The 2010 PES was originally planned with a sample size of 300,000 housing units, but this was reduced in 2009 to 170,000 housing units as part of an effort to lower nonsampling errors through additional quality control operations. The 2010 PES produced net coverage error estimates for the 50 states and Washington, D.C., as well as counties and places with populations of a least 100,000 people. The estimates for states, counties and places were all synthetic estimates.
• The 2020 PES kept the sample size at 170,000 housing units, but changed the methodology for state estimates and did not release county or place estimates. The 2020 PES faced many of the same COVID-19 pandemic-related challenges and delays as the 2020 Census. For example, the primary interview operation was delayed from June to September 2020. Also, the data collection period ran longer than expected and was even reopened to give people more time to respond. Despite these challenges, the 2020 PES provided measures of census coverage for demographic groups and census operations. However, the 2020 PES was not designed to adjust the population estimates, especially at lower levels of geography where the sample size was too small to support high-quality coverage estimates.

BERT, based on its findings from PES staff and coverage experts, recommends the Census Bureau start planning now for a coverage measurement program that could be used to adjust the population estimates base for potential coverage errors in the 2030 Census.

This could include changes to the sample size and overall design of the estimation methodology — if the PES program is planning to field a traditional post-enumeration survey — or including the population estimates base as a use-case if a different approach to coverage estimation is being considered.

Another possibility is for the Census Bureau to explore whether we can use administrative records to produce estimates of net coverage error for the 2030 Census.

We often get questions about whether a specific city or neighborhood was undercounted in the 2020 Census. We even get questions about undercounts for specific population groups in a city or county.

Again, the coverage programs for the 2020 Census were not able to produce reliable estimates at these lower levels of geography. In fact, if we wanted to produce accurate net coverage error estimates for every city or neighborhood using a post-enumeration survey, we would need to essentially conduct a second census, which is not practical.

The key is for the Census Bureau to start planning now so the coverage measurement program is ready in time for the 2030 Census, and so the early stages of research are underway.