Approximation Issues in Simulation-Based Estimation of Random Coefficient Models

This paper discusses issues of approximation in the estimation of mixed logit models that contain variation in tastes across respondents and possible additional variation across choices for each respondent. The authors identify several simplifications to the true log-likelihood function that have been utilized in past work and implemented in existing software. The authors examine the accuracy of these simplifications through Monte Carlo methods. For models with inter-respondent variation in tastes but no intra-respondent variation, the authors find that treating the multiple choices of each respondent as if they were choices by different individuals (i.e., treating the panel data as if they were cross-sectional) results in only a small loss of efficiency when sample sizes are sufficiently large. However, for models with intra- as well as inter-variation in tastes, simplifications of the log-likelihood function result in a large loss of efficiency, even with numerous observations. This latter result implies that the computationally intensive formula for the log-likelihood function, which fully accounts for the intra- and inter-respondent variation, needs to be used, at least until an accurate approximation is developed.


  • English

Media Info

  • Media Type: DVD
  • Features: Figures; References; Tables;
  • Pagination: 19p
  • Monograph Title: TRB 89th Annual Meeting Compendium of Papers DVD

Subject/Index Terms

Filing Info

  • Accession Number: 01155527
  • Record Type: Publication
  • Report/Paper Numbers: 10-2986
  • Files: TRIS, TRB
  • Created Date: Jan 25 2010 11:28AM