Edson: Bureau of Census devises advanced sampling scheme
By Peter Edson
1944 being one of those years when the polltakers will be abroad in the land, sampling public opinion and thereby predicting who is going to be elected dogcatcher where and by how much, it is worthy of note that the U.S. Bureau of Census has worked out some new wrinkles on straw balloting which have produced amazingly accurate results.
It should be made clear at the start that the Bureau of Census wasn’t and still isn’t interested in political prophesying, the work of the bureau being limited to strictly economic fact-finding. But the bureau used its new sampling technique so effectively in the recent survey of consumer requirements that it has set the private, non-governmental polltakers like Crossley, Gallup and Roper to studying the results to see if their own methods may not need some revision.
The possibility that changes in scientific sampling methods would have to be made from time to time of course has been admitted by the commercial poll-taking organizations. The old Literary Digest poll was accurate up to 1936, when it missed completely. Gallup’s organization was close in 1936 and 1940, but it was off in 1942.
The new Bureau of Census technique may not represent as much of a refinement over the Gallup method as Gallup was an improvement over the Literary Digest, but it is hailed as an improvement.
The polltakers all get their results by gathering the opinions of only a limited number of people – from 3,000 to 60,000. In the first instance, that’s approximately one out of every 43,000 people in the country. In the latter, it’s one for every 2,150. In neither case is it a big sample and that may explain why you never knew anyone who was asked for his opinion in a poll.
The trick, of course, is to pick the right 3,000 to 60,000 people for the poll so as to get a representative cross-section of the population, correctly divided as to geographic areas, income levels, sex, occupation, age groups and other pre-determined classification. This is known as “purposive selection,” to get exactly a true percentage of each classification in the entire population. Where the Literary Digest went wrong in 1936, of course, was that all of the two million straw votes it received were cast by people whose names were in phonebooks or owned autos.
‘Cell’ system used
The Bureau of Census, for its recent poll of consumer requirements, went after the problem on a different basis from both these others. First, by running through its 1940 census returns, as corrected by all the intra-census studies it has made of population shifts since then, the bureau was able to select 68 areas or “cells” which were statistically representative of other similar areas in the United States. The determination was made as to location, population, division of rural and urban population, type of farming, and other occupation. Two counties representative of each of these areas were chosen. Then from the census listing of households in all of these counties, providing perhaps 10 times as many households as it was desired to survey, a random selection was made of every tenth household. The census enumerators were told to go to those specific addresses to ask their questions.
What the census poll came up with in the end was a list of roughly 4,900 households out of 36.5 million households in the country, or approximately one for every 7,400 families.
Unfortunately, they weren’t permitted to ask their 4,900 sample families who was going to be the next President.