Abstract:
Non-response or non-participation can introduce selection bias in the analytic results by confounding the behavioral parameters of interest with parameters that determine response. By incorporating a model of selection into the data likelihood function, it is possible to correct potential selection bias. Correctly specified full information maximum likelihood (FIML) methods should be able to correct for selection bias even in discrete choice models with truncated data when no data is available for non-respondents or non-participants. Unfortunately, attempts to implement such methods have not been very successful. This paper carefully examines the FIML approach for estimating discrete choice models with truncated data in order to better understand the difficulties with implementing this estimation technique. Simulation results are used to demonstrate the difficulty of implementing FIML estimates. Given the failure of FIML methods, alternative estimation techniques are proposed.