Why are opinion polls sometimes far from reality? An analysis with scenarios in Python ¿Por qué en ocasiones los resultados de las encuestas de opinión se alejan de la realidad? Un análisis con escenarios en Python

Abstract

Before the runoff election that took place in Costa Rica in 2018, several opinion public polls were executed in order to survey the electorate tendencies. The poll companies that made these surveys showed a technical draw between both candidates or, in other cases, a clear advantage for the ``Restauración Nacional'' candidate over the ``Acción Ciudadana'' candidate. Opposed to this, the final election result gave the ``Acción Ciudadana'' candidate a clear win, with an advantage that was considerably larger than the error percentage that all polls estimated. Considering this, a natural question can be asked: why do polls fail? One of the answers provided by the poll's companies was that they were not allowed to make surveys unit one week before the election, and the electorate opinion changes with time.

In the present work we give a series of hypothetical scenarios in which polls may fails, and a methodology to detect these biased results. Particularly, we analyze the electorate's polls answer rate and its influence in the polls estimates. In fact, if the supporters of a given candidate don't want to answer a poll survey, the poll's results will be biased, independently of the time in which it was made. In this work, we create these scenarios using Python and we also propose alternate weighing methods to compensate these biases. We pretend to open a discussion in poll's results that take into account technical, mathematical and computer simulation aspects that are not generally taken into account in Costa Rican's public opinion surveys and avoid myths such as the sample size is unable to predict the entire population preferences.