Impact of multiple imputation of missing socio-economic data in discrete choice analysis

Abstract

This paper aims to assess the impact of missing income data on estimates of the accuracy of marginal costs of aviation noise externalities, namely missing socio-economic attributes which are normally used to segment the sample for policy purposes. The experimental approach comprised a comparison between the coefficient parameter estimates obtained by fitting multinomial logit models using econometric software to both the complete data set and to data sets where several amounts of income data were replaced by missing values (5%, 10%, 25% and 50%). The latter data sets were analysed employing the two most common methods for missingness: the LD and the Multiple Imputation (MI) method using the expectation-maximization Bayesian bootstrapping algorithm within the Amelia II software program. The most appropriate number of imputations and the set of variables used to obtain the imputations with the MI approach are further investigated. The comparisons will be made through measures of accuracy of parameter estimates and of the combined model.