The error that results from using a probability sample as opposed to using the population of all observations relevant to the given problem. It is possible to quantify this error, which is often referred to as the standard error of the estimate. Nonprobability sampling (see convenience sample) introduces error because the sample is likely to be unrepresentative. Traditional measures of probability sampling error do not account for nonresponse bias and response errors; in many practical situations, these errors are often much larger than sampling errors. Consider political polling, in which the situation is well-known to the respondents. Lau (1994) examined the errors in 56 national surveys concerning the 1992 U.S. presidential election. The sample sizes varied from 575 to 2,086. Although the errors varied substantially across the surveys, they were only weakly related to sample size. Perry (1979) estimated that the total error for U.S. political election polls was 30% larger than the sampling error. Buchanan (1986) studied 155 elections from nine countries from1949 to 1985 and estimated that sampling error, given the typical sample size of 1,500, would yield a 95% prediction interval of
plus or minus 2.5%. However, the actual 95% prediction interval was plus or
minus 5.1%. One would expect that the size of other errors would be even larger relative to sampling errors if the analyst were forecasting for an unusual new product rather than for a political candidate.
- Perry, P. (1979), “Certain problems with election survey
methodology,”
*Public Opinion Quarterly*, 43, 312-325.
- Buchanan, W. (1986), “Election predictions: An empirical
assessment,”
*Public Opinion Quarterly*, 50, 222-227.