Election Forecasting Techniques - Part II

This is the second of two special issues devoted to current topics and innovative approaches in the field of election forecasting techniques. The articles included in these special issues were submitted to the journal after a call for papers was circulated in mid-2013, soliciting contributions that advance the current state of the literature and/or promote novel approaches to political opinion polling, with special emphasis on uses of forecasting techniques of election results. The articles hosted in the two issues cover topics ranging from exit polls, explanatory statistical models based on structural variables (economic trends, government approval ratings, etc.), prediction markets, social media-based election forecasting, the web as a means to collect data on voting preferences, and measures of forecast accuracy. The first contribution in this second issue – “Vote miscount or poll response bias? What causes discrepancy between polls and election results?” by Kathy Dopp and Ron Baiman – develops a statistical method for estimating the rates of vote miscount and relative contributions of random sampling error, exit poll response bias, and vote miscount causing the discrepancies between weighted and unadjusted exit poll results and reported election results prior to official certification of election outcomes. Simulations provided by the authors show the method recovers both direction and magnitude of vote shift and relative response bias parameters. Moreover, through an analysis of the 2004 U.S. Presidential election in Ohio, the authors demonstrate that the estimated discrepancies are consistent with vote miscount sufficient to have altered the outcome of the Presidential election. The following article is “Using social media to forecast electoral results. A review of the state of the art”, by Andrea Ceron, Luigi Curini, and Stefano M. Iacus. The authors discuss the advantages of performing a Supervised Aggregated Sentiment Analysis (SASA), and in particular the two variants of the SASA methodology, ReadMe and iSA, as an attempt to forecast electoral results. Authors provide estimates of voting intention of Internet users – obtained by analyzing millions of comments related to elections held between 2011 and 2013 in France, Italy and the United States – based on both ReadMe and iSA. Then they compare these estimates with the actual voting results and show that, on average, their forecasts of electoral results are accurate with a Mean Absolute Error (MAE) of approximately 2.5%. To further explore the determinants of the accuracy of social media predictions, they compare the MAE of 80 social-media-based forecasts, highlighting the fact that ReadMe increases the accuracy of estimates when compared to forecasts based on traditional techniques, and iSA appears to perform better than does ReadMe. As regards the accuracy of published pre-election polls on voting intentions, the issue hosts two papers. In “Polling and multi-party accuracy measures: evidence from the Italian general elections”, Graziella Castro and Venera Tomaselli attempt to ascertain the statistical advantages and disadvantages in employing two measures of accuracy of polls in a multiparty system, that is, a revised version of the Martin, Traugott, and Kennedy accuracy measure (the “A’ measure”) and the Polling Accuracy measure of Arzheimer and Evans (the “B measure”). The data refer to the Italian general elections held in 2006, 2008, and 2013 and polls published during the last 6 months of those election campaigns. The authors are primarily interested in testing three hypotheses: the centre-right coalition has been underestimated across time; a lower degree of error is estimated when getting closer to election day; the two statistical tests of the B measure provide more reliable estimates of the number of biased polls in the Italian polling than the confidence interval of the A’ measure. They find strong evidence in favor of the first hypothesis, weak evidence for the second, and mixed evidence as regards the third hypothesis. The authors conclude that, on the whole, the use of the two measures is appropriate under certain conditions. Rinaldo Vignati and Giancarlo Gasperoni (“The predictive ability of preelection polls in Italy: a regional focus”) also use the A and the B measures in their exploration of published pre-election polls in Italy, but with a specific emphasis on regional elections and region-level surveys for the national Senate election. In 2010 regional elections the winning candidates’ level of support was generally underestimated by pre-election polls, a result in sharp contrast with previous findings regarding national elections in 2001, 2006 and 2013. In the 2013 region-level Senate elections, poll results tend to severely underestimate the success of a new actor on the Italian political stage: the Five-Star Movement. One of the submitted contributions concerns explanatory statistical models based on structural variables, in that media content is a variable indicator much like government approval ratings or economic trends. More specifically, in “News coverage and candidate preferences: using media content to predict election poll movement”, Dan Cassino and Christian Kolmer study the effect of media coverage on candidates’ support by focusing on the problem of how a “middle-man” changes his/her opinion in the real world. The point they make is how to exclude the middleman so as to account for news coverage at the macro media level only in order to predict election results. The authors test three hypotheses about the effect of news coverage: greater positivity in statements will increase support for a candidate; greater positivity in the difference between the number of positive and negative statements will decrease support for a candidate’s opponent; media coverage should have a greater impact on support for a candidate during the second half of the campaign. To test the three hypotheses, different techniques, like the Wald test on causality and the ARIMA model, are used. Findings reveal strong evidence in favor of the first two hypotheses and mixed support for the third. Finally, “Estimating representatives from election poll proportions: The Spanish case”, by Jose M. Pavía, Belén García-Cárceles and Elena Badal, deals with the issue of translating votes into seats in the context of the Spanish system. The authors state that no universal solutions exist for mapping vote shares onto representatives, as each electoral system has a specific electoral formula. The article attempts to fill this gap by offering some initial exploratory answers to the issue. The authors develop three groups of models. As to the first group, the authors fit independent univariate linear equations for each party, where the dependent variable is the proportion of seats in the 2011 Spanish general election and the independent variable is the proportion of votes. For the second group, authors fit univariate linear models per each party and constituency between the total proportion of votes and the corresponding party-constituency proportion of votes, and then use this proportion estimates obtained for each party in each constituency to allocate seats in the corresponding constituency. In the third group of models, authors investigate the performance of the second model with an alternative sampling design in which an extra polling effort is made in those constituencies with a historically strong presence of regional parties. In order to assess the performances of the proposed models, the authors simulate a number of polls, and compare the Parliaments that would be obtained applying the d’Hondt rule (Direct Poll Forecasts system) to those that would be obtained by applying their models. They conclude that their models express a better perfomances compared to the Direct Poll Forecasts, as the use of the former improves overall forecasts and reduces the total mean squared error