Census and Sensibility
by Howard Fienberg
August 12, 2002
This spring, the U.S. Census Bureau announced plans to
effectively replace the "long form" of the Census with the annual
(and relatively new) American Communities Survey (ACS). Despite the potentially
serious ramifications, the debate about this policy has not exactly been
front-page news. When the Washington Post examined the debate on April 8, it
remarked, "Most in the statistical community support the concept,
although some have concerns." Yet nowhere in the article could be found
what such concerns might be.
You don't have to be a technocrat to care. The numbers
from the Census Bureau guide funding and effort across government at all
levels. They tell us where new schools should be built to accommodate an
influx of children and where new roads and mass transit systems should be
added to deal with dreadful commutes. They also drive the wrenching political
disputes over apportioning legislative districts and seats.
During the decennial census, every American household is
supposed to receive a short form questionnaire, asking basic questions like
how many people live in their domicile. One in six also gets a long form of
53 questions, which go into much greater demographic depth and detail (like income,
housing, language, etc). The Bureau proposes to replace the long form with
the ACS, an annual survey in pilot-testing since 1996 which covers ground
similar to the long form, querying about 3 million people.
There are pluses and minuses to supplanting the long form
with the ACS. On the plus side:
would give the Bureau a permanent professional apparatus dedicated to
running the ACS, which would make running the decennial census and
training the decennial census-takers much easier and cheaper. The Bureau
estimates that the long form accounts for about sixty percent of its
work. Science magazine reported that in 2000, the Bureau spent about a
million dollars to train temporary workers for follow-up interviewing
with households that received the long form but did not respond. By
contrast, the ACS would have several thousand dedicated staff,
well-versed in survey-taking.
would require the Bureau to compile a permanent master address file.
Instead of creating one from scratch every ten years, a very time-consuming
and costly production, this constantly updated file would make the
decennial census much more accurate and save a fair bit of money over
would yield more timely data. Not only could data be collected annually
instead of every ten years, new and timely questions could be added to
the questionnaire for faster results.
the ACS has a significantly smaller sample, is run over the telephone
and is conducted by professionals, response rates are dramatically
better than for the Census long form. While the ACS manages only 52
percent response on the initial run, dedicated follow-up attempts have
raised the response rate as high as 96 percent.
However, on the con side:
ACS enterprise is expensive. Officials expect to need about $124 million
for the ACS in fiscal year 2003, and even more the year after. In the
long run, the ACS might save money. But the long form, at the moment, is
a bargain - getting this extra info at every decennial Census, aside
from the hassles in follow-up interviews, costs very little.
ACS is more timely, and reasonably rapid
response questions can be added to the questionnaire, but the Bureau
doesn't do "rapid" very well. Its forays into experimental
questions have achieved mixed results.
base, the smaller sample of the ACS translates to a larger margin of
error. Off the record, some statisticians are concerned that the Bureau
has not adequately tested the accuracy of the ACS sampling mechanism.
ACS might end up missing the same people claimed to be missed even in
the huge effort of the decennial survey. A test of the ACS method showed
that response rates differ dramatically by race. While whites answered
about 60 percent of the mailed surveys, blacks and Latinos answered only
about 35 percent. The Bureau does follow-up interviews, of course, but
only does in-person follow-up for 1 of every 3 households that do not
return calls or cannot be reached by phone.
larger question with the ACS is what to do with the data. While the
Census data gets specific about regions, states, counties, cities, even
street blocks, the ACS is a national sample. Because of the national
sampling method, specific local data can not be inferred from the ACS.
Statisticians have yet to decide how to solve this "statistical
inference" problem. Any attempt to influence legislative
apportionment using ACS data could prompt a political fight greater than
any previous battle over statistical sampling in the Census.
Most data-hungry people, whether in government or
business, are eager to give the ACS a try. They desperately crave data that
is not two to three years out-of-date by the time it reaches them. But once
the long-form goes the way of the do-do, it probably can't be resurrected.
This major change in how we collect official statistics ought to inspire at
least a little more debate before any final decisions are made.
See the original: http://www.tcsdaily.com/article.aspx?id=081202B
return to Howard Fienberg's page