As we begin the last stages of the 2010 Census data collection, our thoughts are turning more and more to evaluating how the country did in this census.

I thought it might be good to talk a little about how censuses are usually assessed. There are some basic ideas that are crucial to any notion of statistical quality.

First, the basic quality criteria are simple. An oft-mentioned goal of the 2010 Census is “enumerating each resident of the US once, only once, and in the right place.” Thus, there are three basic threats to quality in a census: 1) failure to enumerate a person, 2) enumerating someone more than once, and 3) assigning an enumerated person to the wrong physical location. We also seek accurate reports of attributes of the persons enumerated (e.g., age, race/ethnicity).

Second, the quality of a census depends on its use. For example, the first use of the census is distribution of members of the House of Representatives based on population of states. The key quality of that use is whether we assigned counted persons to their appropriate state of residence. The next important use of the census is redistricting within states, based on counts of persons within blocks. For that use, we need accurate assignment to the correct block of their residence. (This is a much stricter quality hurdle than assignment to the correct state of residence.)

Third, uses based on count totals face tougher quality challenges than those for distributional purposes. For example, if the census count is used by a program to provide $2.17 for every person in an area, its allocations will be affected by every person not counted and every person counted twice. If there is an undercount, fewer dollars are allocated than intended. On the other hand, if a program allocates a fixed sum of funds, say $50 million to the states based on the proportion of the total population in each state, then getting the relative counts right is necessary. For example, if we fail to count exactly 1% of every state’s population, the relative share for each state in the census counts is the same as the relative share truly. But if one state is undercounted by 2% and another by 0.5%, the first is relatively harmed versus the second. (There are many uses of censuses that are distributional, like the apportionment of the House.) It’s easy to see, therefore, why there is great interest in the differential undercount for distributional purposes; that is, the extent to which different groups experience different levels of undercount or overcount in the census.

These are merely conceptual notions of quality, not measures of quality. I’ll talk about those in a later post.

Please submit any questions pertaining to this post to ask.census.gov