When the census question changes

A number can change because the world changed. It can also change because the question used to measure the world changed. Before asking an AI to explain a population trend, find out whether the two counts were made under the same rules.

The U.S. Census Bureau’s Questions and Answers for Census 2000 Data on Race, dated March 14, 2001, gives a precise historical example. In 1990, respondents reported one race. Census 2000 allowed respondents to report more than one race. The Bureau says the two sets of data are not directly comparable, and it names other differences that may affect comparison as well, including the order of the race and Hispanic-origin questions and changes to wording and format.

That does not mean every change between the counts came from the questionnaire. It means the question is part of the evidence. A change in the measured total may combine population change with a change in how responses could be recorded and tabulated. Treating the later number as a clean continuation of the earlier category can make the apparent trend look more definite than the measurement supports.

The tabulations add a second trap. Census 2000 could report race groups alone, or alone or in combination with one or more other races. In the combination-inclusive version, a person who reported Asian and Black or African American appears in both corresponding group totals. The person is counted once in the census total, while the overlapping category counts can add up to more than the total number of people. That is not double-counting people in the census. It is overlap in the way the categories are reported.

Here is a fully fictional arithmetic example. Imagine 100 people, including 10 who report both Cedar and Flint. A Cedar-alone-or-in-combination count could include those 10, and a Flint-alone-or-in-combination count could include the same 10. Adding the two category totals would count those people in both category sums, while the population total would still contain each person once. The names and numbers are invented; the overlap pattern illustrates the Census Bureau’s explanation.

The practical lesson is a question to ask before asking for a trend: are the categories, response options, wording, order, and tabulation method comparable? If they are not, an answer should describe the break in measurement before it offers a story about the population. Later researchers may harmonize categories for a particular purpose, and that work can be useful. It does not erase the fact that the original questions produced different kinds of data.

This is a historical example from the 2001 Census Bureau page, not a statement about the current 2026 census questionnaire. The habit is the part I want to keep: compare the measuring instrument before interpreting the measured change.