The Census Bureau has long used various statistical safeguards to help us avoid disclosing individuals’ information when we release statistics. As we’ve described in other materials, including the handbook Disclosure Avoidance for the 2020 Census: An Introduction, we are using a stronger system to protect your 2020 Census information against today’s cyberthreats.

The Detailed DHC-A used an algorithm called SafeTab-P to protect respondent confidentiality. This algorithm infuses noise – small, random additions or subtractions – in the total population counts and the sex-by-age statistics for detailed race and ethnicity groups and American Indian and Alaska Native tribes and villages.

In the previous blog, we explained how the level of detail available for a group in the Detailed DHC-A is based on population size (using the noise-infused counts) and geographic level. The blog also explained that we set accuracy targets – in the form of margins of error (MOE) – for how much disclosure avoidance-related uncertainty would be in the published statistics. Table 1 recaps the thresholds and MOEs for the detailed and regional groups.

The MOE is how close we can expect the noise-infused counts to be to the enumerated counts approximately 95% of the time. This means that for a noise-infused count of 20 with an MOE of three, we can be 95% confident that the enumerated or “true” count is somewhere between 17 and 23.

Our internal metrics that compare final noise-infused counts to the original, enumerated counts show that these expected MOEs are consistently met. Fewer than 5% of counts are outside of the target MOE.

What does this look like for the typical detailed group?

Table 2 shows how close, on average, the noise-infused detailed counts are to the enumerated counts. For example, the cell in the top-right corner shows that the average noise-infused national sex-by-age count is within about one person of the enumerated count. For counties, the average noise-infused sex-by-age count is within about four people of the enumerated count – so, for example, the noise-infused count of 45- to 64-year-old women who are Hungarian alone in Cook County, Illinois, is 667, and the enumerated count is expected to be within about four or five people of that. 

The sex-by-age counts have similar expected levels of noise because, for a given level of geography, each count has the same MOE – no matter whether the population receives four, nine or 23 age categories.

• National and state counts have MOEs of ±3 and, as shown in the top two rows of the sex-by-age counts column, the noise-infused counts are within about one person of their enumerated counts.
•  Sub-state and American Indian, Alaska Native and Native Hawaiian (AIANNH) area counts have MOEs of ±11, and their sex-by-age counts are within about four or five people on average.

The total population counts are a mix of counts that receive only a total count and those that are aggregations of sex by four, nine or 23 age categories. (When groups receive sex-by-age data, we add up the sex-by-age counts to produce the total population count for the group.) These aggregations have slightly higher MOEs than the total-only counts due to the way that noise adds together. (For details, refer to the Detailed DHC-A technical documentation.)

Even with this additional noise due to aggregation, the average difference between the noise-infused counts and the enumerated counts remains reasonable:

• On average, national total population counts are within three or four people of the enumerated count.
• State total population counts are within one or two people.
• Sub-state and AIANNH total population counts are within five to seven people.

Table 3 repeats this analysis for the regional groups’ counts. The noise-infused sex-by-age counts are within 19 or 20 people of the enumerated counts, on average, while the noise-infused total population counts are slightly farther off because of the effects of aggregation, but they are well within the expected average based on the amount of noise infused.

As we mentioned earlier, the Detailed DHC-A provides a substantial increase in the number of population counts being published relative to the 2010 Census. However, we have implemented some data suppression, processed in the following order, to ensure demographic reasonableness.

• We implemented a minimum population threshold for sub-state and AIANNH geographies (Table 1) to prevent the infused noise from disclosure avoidance making it look like populations exist in areas where they do not.
• We suppressed cases where a noise-infused count fell below zero.
• We suppressed the alone count in areas where the group’s noise-infused alone population count was greater than its noise-infused alone or in any combination population count for that area.

Suppression is specific to a particular geographic area and does not apply to an entire geographic level; a group may be below the minimum population thresholds in a county (and be suppressed) but may exceed the threshold in other counties and receive totals or sex-by-age counts for those counties.

Table 4 shows the proportion of detailed counts for each geographic level that are suppressed. At first glance, many of these numbers appear quite high. For example, 95.5% of county total population counts are suppressed. This is likely due to a combination of geographic concentration – many groups live in only a few counties nationwide – and small county total population sizes throughout much of the country. Despite these high proportions, nearly all these suppressions occur for groups in areas where there were no enumerated members of that group. Counts are still produced, but in a targeted way, for groups in areas where there is sufficient population.

Similarly, Table 5 shows the proportions of counts for regional groups that are suppressed. Because the regional groups are typically larger than the detailed groups, there is not as much suppression of the total population counts across the different geographic types.

Regional groups that get sex-by-age counts are larger than regional groups that get only a total population count, which is why fewer sex-by-age counts are suppressed. Suppression can still happen for sex-by-age counts if a group has a large enough total population to qualify for the breakdown and meets one of the following criteria:

• The noisy count for an individual sex-by-age category fell below zero.
• The noisy alone count for that specific sex-by-age category was less than its corresponding count for the alone or in any combination population.

Suppression happens more frequently at the tract level, as shown in Table 5, because many tracts have unique sex or age concentrations – that is, they are made up of primarily college students, for instance, or older people or single-sex group quarters residents. The thresholds for these tables mean that a tract may qualify for more granular sex-by-age categories based on its total population but, because of its age composition, many of those sex-by-age counts have few or no people in them and are suppressed.

In previously released 2020 Census data products (i.e., the Redistricting Data Summary File, the Demographic and Housing Characteristics File and Demographic Profile), data generally become more accurate as you aggregate them. However, this isn’t the case when aggregating data in the Detailed DHC-A because of the way noise was infused in the data.

Data users should exercise caution when creating custom aggregations, adding or subtracting as little as possible to minimize the accumulation of noise across counts. This is because the expected amount of noise in a sum (or a difference) gets larger with each additional count being added or subtracted.

• To create counts for a custom geography, remove or add as few geographies as possible. For example:
  • To create a count of the Samoan alone population for the Pacific West states, add together the Samoan alone counts for Alaska, California, Hawaii, Oregon and Washington, rather than subtracting the Samoan alone counts of the 46 other states and state equivalents from the national Samoan alone count.
  • To create a count of the Mexican population for the 10 Arizona counties that are majority urban, remove Apache, Graham, Greenlee, Navajo and Santa Cruz counties from the Arizona state total Mexican population rather than combining the Mexican population for the 10 majority urban counties.
•  To create custom counts using age data, collapse as few categories as possible. For example:
  • To compare groups that have nine and 23 age categories, collapse the 23 age categories into the nine rather than into four age categories or other custom categories when possible.

As another resource for understanding this guidance, we also discussed these examples during the pre-release webinar on September 13.

Data users who want to calculate a variety of percentages as they work with the data may find that not all denominators necessary to calculate the percentages are included in the Detailed DHC-A. Table 6 shows which data source to use when calculating a percentage.