Handling missing voting preferences with multiple imputation. A new method and possible applications

The handling of missing values in surveys is controversial, in particular when they recur within sensitive questions, political preferences among them. Listwise deletion is one of the most commonly used methods, though it can cause a massive loss of information and sample biases. Here we contributed to the literature which uses multiple imputation for imputing individuals’ voting preferences implementing a cross-country analysis for six European countries. By relying on information about individuals’ socio-economic characteristics and values drawn from Round 9 of European Social Survey, we imputed abstainers’ and non-respondents’ potential electoral preferences. Differently from previous studies, we used a progressive imputation model based on supervised machine learning. We showed two possible applications of the imputed dataset. Firstly, we predicted how the party landscape would have changed under full turnout, looking at winners and losers. Unlike previous works, we built party weights to correct ESS sample selection biases and ensure consistency with the actual political landscape. Secondly, we showed the potentialities of using the dataset for analysing the differences between parties’ actual and imputed voters. We chose variables considered as crucial in the literature on turnout and voting behaviour, focusing on objective and subjective wellbeing in particular. Looking at correlations, we found that while some characteristics predictably recur across individuals, others can be counterintuitive. We also showed that imputed voters of right-wing populist parties tend to belong to lower-income deciles, feel more economically insecure and be less happy than actual and imputed voters from other parties. However, there are differences across the parties. Our findings suggest that analyses on voting behaviour could benefit from this method to obtain specific information on unstated individual party preferences. The method has great potential for predicting future electoral preferences using panel data. However, it can also be used for imputing other variables of interest.