survey research

In 1936 the Literary Digest forecast that Franklin D. Roosevelt would lose that year's US presidential election. The forecast was based on some 2 million telephone calls. Simultaneously some of the first Gallup polls, having sampled only around 2,000 people, were correctly predicting that Roosevelt would win by a landslide. The Literary Digest poll failed because telephones, in 1936, mostly belonged to the rich, who mostly opposed Roosevelt: thus it had an incorrect ‘sampling frame’. In statistical theory a sample is representative of a population if, and only if, each member of the population had an equal probability of being selected for the sample. The rigorous way to achieve a representative sample is to get a list of the eligible population (such as the electoral register, for voting samples), and select every (P/n)th member of the population, where P is the size of the population and n is the desired sample size. In practice most surveys use the rough-and-ready ‘quota’ method, in which the interviewer is instructed to interview the correct proportion of each principal social and demographic group in the population.

If a sample is correctly drawn, the laws of statistics enable us to predict how close its distribution of the trait being examined (such as voting intention) is to the unknown distribution of that trait in the population from which it is drawn. The form of this prediction is always that there is a high probability (usually 95 or 99 per cent) that the sample distribution varies by not more than a small proportion (typically 1 or 3 per cent) on either side of the true distribution. Headline writers, who typically overinterpret small shifts in voting intentions revealed by successive polls, insufficiently understand this form. One remedy is to take a ‘poll of polls’ including all the reputable polls taken at roughly the same time, and pool their results for the best available forecast of voting intentions.

The surveys of most interest to political science are not those of voting intention (‘If there were a general election tomorrow, how would you vote?’; but every respondent knows there will not be), but rather those which tap attitudes and behaviour at deeper levels. Nationwide election surveys which have run continuously since 1952 in the United States and since 1963 in Britain have built up a full picture of why people vote in the ways that they do. The earlier surveys were held to justify the Michigan school's picture of the electorate as ill-informed and responding more to their inherited ‘party identification’ than to the issues. More recent work suggests that voters are closer to the rational choice school's image of people who choose the party which is closest to offering them what they want; current analysis accepts that both schools of thought have a valid contribution to make towards understanding what makes voters tick.

Survey research underpins almost all good empirical work on large populations in political science and sociology, as there is no other way of making reliable generalizations about them (which does not deter many self-confident people from making unreliable ones).