A Comment on Electoral Forecasting from Poll Data

Abstract

I heartily approve of research based upon poll data, in which I include specifically the contribution by Paul Whiteley, ‘Electoral Forecasting from Poll Data: The British Case’ (this Journal, IX (1979), 219–36). I note with pleasure the fact that the author has relied heavily upon Gallup data for the testing of his time-series model. There is one specific objection that I have to the detail of the contents of this paper. I refer to the sentence bridging pages 231–2, in which he states: ‘the large inaccuracy of the 1951 forecast is not, in fact, due to the model so much as Gallup data, which were extremely inaccurate in that year’. There is a footnote which states that this was inferred from the residuals of the forecasting model. The error quoted is 8·3 per cent. This does refer to the last pre-campaign poll, admittedly. The last campaign poll, the only one that could be compared with the result itself, shows, according to Gallup records, a deviation of 2·2 per cent. The bland assertion of Mr Whiteley that our results were inaccurate in that year is totally unsupported by any evidence of a practical kind. Instead we have the statement that the fit of the model was rather bad on this occasion. Remembering that public opinion, including support for a specific political party, is known to change radically between elections, between local elections and by-elections, and so on (I confine myself here to actual elections – I do not depend upon poll data to substantiate my case) and knowing that among the causes of this are political events, and events of a social and economic nature which affect the mood and attitude of the electorate, it does not surprise me that a purely mathematical model does not necessarily fit from time to time. Indeed elsewhere in the article the author mentions the possibility of shocks affecting public opinion and the degree to which these shocks are effected in the auto-regressive scheme that he has put forward.