This paper tests a new method for designing stated preference (SP) experiments. Currently, many SP designs used in practice maintain orthogonality between the attribute levels - or at least seek to minimise deviation from orthogonality - because it is thought this will produce lower standard errors associated with the parameter estimates. The authors however contend that the standard errors of coefficient estimates derived from a logit model are not necessarily minimised when using an orthogonal design, and that variance-minimising parameter estimates are obtained when the binary choices offered to respondents have probabilities of being chosen of 0.917 and 0.083. A previous test of this new design principle suggested that there was the potential to outperform a standard orthogonal design. A further test is reported here comparing two orthogonal designs with three "new" designs. The context is a within-mode train-service choice experiment for commuters into London with the attributes being journey time, ticket cost, service frequency and time spent standing. The "new" designs are a design produced by an unconstrained algorithm, a design which constrains the attributes to be contextually realistic, and a design which both constrains the attributes to be contextually realistic and rounds the attribute levels. The relative merits of the five designs are analysed in terms of the t-statistics on both parameters and on monetary values of attributes. The paper concludes with recommendations for further research to improve the design of SP experiments. For the covering abstract see ITRD E105584.


  • English

Media Info

  • Features: References;
  • Pagination: p. 51-62
  • Serial:
    • Volume: P434

Subject/Index Terms

Filing Info

  • Accession Number: 00796487
  • Record Type: Publication
  • Source Agency: Transport Research Laboratory
  • ISBN: 0-86050-325-9
  • Files: ITRD
  • Created Date: Aug 2 2000 12:00AM