A network model for multiple selection questions in opinion surveys

Opinion surveys can contain closed questions to which respondents can give multiple answers. We propose to model these data as networks in which vertices are the eligible items and arcs are the respondents. This representation opens up the possibility of using complex networks methodologies to retrieve information and most prominently, the possibility of using clustering/community detection techniques to reduce data complexity. We will take advantage of the implicit null hypothesis of the modularity function, namely, that items are chosen without any preferential pairing, to show how the hypothesis can be tested through the usual calculation of p-values. We illustrate the methodology with an application to Eurobarometer data. There, a question about national concerns can receive up to two selections. We will show that community clustering groups together concerns that can be interpreted in a consistent way and in general terms, such as Economy, or Security or Welfare issues. Moreover, we will show how different society groups are worried by different class of items.