On the Likelihood of Improving the Accuracy of the Census Through Statistical Adjustment

In this article, we sketch procedures for taking the census, making adjustments, and evaluating the results. Despite what you read in the newspapers, the census is remarkably accurate. Statistical adjustment is unlikely to improve on the census, because adjustment can easily put in more error than it takes out. Indeed, error rates in the adjustment turn out to be comparable to—if not larger than—errors in the census. The data suggest a strong geographical pattern to such errors, even after controlling for demography—which contradicts a basic premise of adjustment. The complex demographic controls built into the adjustment mechanism turn out to be counter-productive. Proponents of adjustment have cited “loss function analysis” to compare the accuracy of the census and adjustment, generally to the advantage of the latter. However, these analyses make assumptions that are highly stylized, and quite favorable to adjustment. With more realistic assumptions, loss function analysis is neutral or favors the census. At the heart of the adjustment mechanism, there is a large sample survey—the post enumeration survey. The size of the survey cannot be justified. The adjustment process now consumes too large a share of the Census Bureau’s scarce resources, which should be reallocated to other Bureau programs.