Pre-election poll accuracy: meta-analysis as multilevel approach

In pre-election polls, more than in other surveys, the accuracy of predicted outcomes is verifiable. Comparing the expected result of a party with the actual electoral performance, in fact, it is possible to compute a measure of poll predictive accuracy. Some scholars proposed a measure whereby the accuracy of pre-election polls is assessable. Employing such measure, in this paper, we present a meta-analysis of accuracy of pre-election polls aiming to explain the differences among the accuracy measures in each poll. In such as, we intend to identify the potential sources of biases affecting poll results. We carry out a meta-analysis of accuracy measures specifying a ‘special’ multilevel regression model. Unlike the standard multilevel model, in fact, we do not have raw data of interviewed subjects as level-1 units but only summary statistics varying across polls. The accuracy measure is used as dependent variable in a hierarchical model where each outcome is affected by a specific sampling error assumed normally distributed and with a known variance. By the means of intra-class correlation coefficient, then, it is possible estimate the proportion of total variance due to the single poll. Multilevel regression models allow including in the model the poll features as explanatory variables. The paper shows the results of meta-analysis regarding a data set of pre-election polls, carried out in parliamentary elections in Italy and published on website: www.sondaggilettorali.it, from 2001 to 2013.