The public opinion polls conducted before the 1992 general election were a disaster. The errors were far in excess of expected sampling variation and can only be explained by biases specific to the context of the 1992 election since previous election results had been forecast accurately. The replication in the polls in the months before the election provides a natural experiment from which the sampling variance of quota samples can be estimated. Using a components-of-variance model, company-specific variances are estimated and are found to be consistent with those which would have been obtained from equivalent random sample designs.

Fitting time series models to polling data prior to elections suggests that there is little change in the opinions about voting in the month prior to the election. If this is true then the need for polls to provide instant results is less pressing. A case is made for a well designed random panel survey to be conducted alongside the standard quota samples used by most polls.