Bayesian estimation of hazard models of airline passengers' cancellation behavior

This study explores the use of Bayesian methods to estimate hazard models of airline passengers' cancellation behavior. We show how the discrete time proportional odds (DTPO) cancellation model can be rewritten as an equivalent fixed parameter discrete choice model that can be easily estimated using Bayesian methods and extended to random parameters that account for unobserved heterogeneity. The use of Bayesian methods allows us to address several limitations of existing airline cancellation models. First, because of the random parameter reformulation, it is possible to calculate individual-specific cancellation probabilities. Second, unlike existing DTPO models that forecast average cancellation probabilities only, our model can be used to forecast both means and a measure of variance (credible intervals) associated with an individual's cancellation probability. We apply the Bayesian estimation method to a dataset of tickets purchased over a two-year period by employees of a university in Atlanta, Georgia. During this time period, the major carrier in Atlanta terminated an agreement in which it allowed employees to purchase discounted fares that could be refunded or exchanged without a fee. The data allow us to investigate how passenger cancellation behavior changed when these fares were discontinued. Cancellations are reduced on average 3.3% when customers must pay to exchange their tickets. For a simulated hypothetical flight the coefficient of variation of cancellation is 43% when the state rate was offered, and 83% without state rates.