Estimation of Underrepresented Strata in Preelection Polls: A Comparative Study

Abstract

In this work we aim at increasing the utility of a preelection poll, by improving the quality of the vote share estimates, both at macro and micro level. Three different methodologies are applied with that purpose: (1) polls aggregation, using existing auxiliary polling; (2) application of multilevel regression methods, using the multilevel structure of the data; and (3) methods of small area estimation, making use of auxiliary information through the application of the Empirical Best Linear Unbiased Prediction (EBLUP). These methods are applied to real data collected from a survey with the aim of estimating the vote share in the Portuguese legislative elections. When auxiliary information is required, we concluded that polls aggregations and EBLUP have to be applied with caution, since this information is extremely important for a good application of these models to the data set and to obtain good reliable forecasts. On the other hand, if auxiliary information is not available or if it is not of good quality, then multilevel regression can and should be seen as a safe alternative to obtain more precise estimates, either at the micro or macro level. Besides, this is the method which further improves the precision of the estimates. In the presence of good auxiliary information, EBLUP proved to be the method with greater proximity with real values.