Using posterior analysis to predict missing information in passively collected data

Abstract

Trip diaries are increasingly being collected using passive manner for instance via smartphone surveys, where travel information (e.g. travel modes, purposes) is inferred and the participants are asked to validate or correct the inferred information. However, many people neglect to validate their trip resulting in large shares of unvalidated or missing data. To better predict missing data for individuals who have validated some of their trips, we propose the use of posterior analysis. Posterior analysis makes use of the Bayes rule to find the likely location of an individual on a population distribution by conditioning on the individual’s previously observed choices. Using a two-week long trip diary dataset collected in the UK, we find that by making use of posteriors, the average probability of inferring the chosen alternative in a trip purpose model on a testing dataset substantially increases from 0.47 to 0.57 compared to without using posteriors.