Bias Correction by Sub-population Weighting for the 2016 United States Presidential Election

摘要

The 2016 Presidential Election was an international surprise, as President Donald Trump came back from a seemingly large deficit in the pre-election opinion polls. As most, if not all, of the major polls missed the election results, the public started to doubt the credibility of pre-election polls. This article proposes that there was a methodological error in the polls. The polls used the census data of American population to weigh their data. However, population may not have a correlation with turnout, meaning that a certain population group may not vote much; not contributing to the electorate. For this reason, the polls based on population might systematically over or underestimate a particular candidate. Thereby, the proposition is that the polling agencies should consider the electorate, not the population for modifying the polling results. The proposition is substantiated with a series of statistical simulations supporting the claim that a poll conducted based on the electorate resembles the actual result more accurately. Conclusively, it argues that, as the polls play a pivotal role in affecting the media and the electorate, the improvement of polls is necessary for well-informed forecasts to be available.