How Sampling Would Change The Census
By Barbara Vobejda
If they don't reach them by mail, the workers knock on doors, again and again, until someone answers, or until they find a neighbor who can describe the family characteristics of those who couldn't be located.
Now, for the first time with the 2000 census, the Clinton administration is proposing something very different. Rather than counting every person, the government would count as many as possible, but then use information from a small, representative group to estimate the characteristics of the missing households.
That process, known as "sampling," is employed routinely, in the public opinion polls that drive political campaigns and the marketing surveys used to develop and sell new consumer products.
While the Census Bureau has used a limited amount of statistical sampling in its work for years, plans for the 2000 census call for the method to be used much more broadly than ever.
Though they disagree about how to change it, both Republicans and Democrats concede that the traditional method has become less and less accurate for a population in which increasingly large numbers of people do not speak English, are not attached to stable families or move frequently from house to house.
In 1990, for example, the Census Bureau missed between 10 million and 15 million Americans, and double-counted or otherwise erred with 6 million to 9 million people, causing a net undercount of 4 million, according to the General Accounting Office, the auditing arm of Congress.
To improve on that performance in the next census, the bureau, with the backing of an expert panel from the National Academy of Sciences, proposed to combine a traditional head count with statistical sampling.
According to the plan, the agency would send out questionnaires to all households on its mailing list, as it has in the past. And as in the past, the government would send workers into every census tract to find people who didn't respond by mail. But before now, that would be the end of the process. This time, the results of the door-knocking would be used to infer information about the missing households.
For example, if in one census tract the government reached 900 households but believed, based on its address lists, that it was missing another 100 households, it would extrapolate, from the people it had found, the demographic characteristics of those it couldn't locate.
As a quality check, the bureau then would conduct a survey of 750,000 households, a representative subset of the entire nation, and use those results to check the accuracy of its initial head count.
© Copyright 1998 The Washington Post Company