Title: A censored generalized linear finite mixture model for the columbia card task
Authors: Nienke Dijkstra - Erasmus University Rotterdam (Netherlands) [presenting]
Patrick Groenen - Erasmus University Rotterdam (Netherlands)
Henning Tiemeier - Harvard University (United States)
Abstract: The Columbia Card Task (CCT) is a card game that measures risk behavior. Participants gain points by turning over win cards, but turning a loss card ends the round and costs points. At any time the participant can stop playing a round. The aim is to estimate the number of cards a participant intends to turn over and relate this outcome to participant characteristics. When the participant faces a loss card, the observation is censored. The purpose is to build a statistical model that appropriately addresses the features of the CCT, which is performed in 16 rounds each by 3326 children aged 8-10 years. The new model should accommodate censoring and take into account the unobserved heterogeneity across individuals. We propose a censored generalized linear finite mixture model. Several distributions and link functions are investigated. They are compared based on their predictive performance, computational speed, convenience for usage, and their interpretablility. In particular, we study the Poisson distribution with an identity link that gives convenient linear effects. Alternatively, to weaken the equidispersion assumption of the Poisson distribution, we consider the negative binomial distribution.