Regret minimisation and utility maximisation in a freight transport context

Abstract

This study is the first to compare random regret minimisation (RRM) and random utility maximisation (RUM) in freight transport application. This paper aims to compare RRM and RUM in a freight transport scenario involving negative shock in the reference alternative. Based on data from two stated choice experiments conducted among Swiss logistics managers, this study contributes to related literature by exploring for the first time the use of mixed logit models in the most recent version of the RRM approach. We further investigate two paradigm choices by computing elasticities and forecasting choice probability. We find that regret is important in describing the managers’ choices. Regret increases in the shock scenario, supporting the idea that a shift in reference point can cause a shift towards regret minimisation. Differences in elasticities and forecast probability are identified and discussed appropriately.

Keywords:

• random regret minimisation
• random utility maximisation
• freight transport
• stated preference
• shift of reference point

Notes

1. The idea that regret minimisation is an important choice behaviour is well established in many fields, namely, marketing (Simonson 1992; Zeelenberg and Pieters 2007), microeconomics (Loomes and Sugden 1982; Sarver 2008), psychology (Zeelenberg 1999; Connolly and Reb 2005), management sciences (Savage 1954; Bell 1982), and transportation (Chorus et al. 2006, 2009). The random regret minimisation approach to discrete choice models translates this conceptual notion into an operational, easily estimable logit model for the analysis of risky (Chorus 2012a) and riskless choices (Chorus 2010).

2. For a more exhaustive overview of comparisons between RUM and RRM, see Chorus (2012c).

3. We acknowledge that Hess, Stathopoulos, and Daly (2012) allowed for heterogeneity within the RRM model but did not observe such. We further note that Hess, Stathopoulos, and Daly (2012) referred to the version of RRM proposed in 2008 (alternative specific regret; Chorus, Arentze, and Timmermans 2008). This current paper refers to the version of RRM developed in 2010 (attribute specific regret; Chorus 2010).

4. The use of constrained triangular distribution ensures that the estimated coefficient is consistent with the micro-economic perspective (for further details, see Hensher and Greene 2003). As noted by an anonymous reviewer, distributions with estimated bounds can better identify mean and heterogeneity, but they require significantly more computational time.

5. Negative random error ω is distributed in extreme value type I.

6. Considering the error component nesting specifications result in correctly identified models as shown in Walker, Ben-Akiva, and Bolduc (2007, 1107–1109 and 1112 for panel data). The models in the present study were further tested through estimation based on different random starting values (following the procedure for Python biogeme proposed by Boeri 2011).

7. For more details on the study, see Masiero and Maggi (2012).

8. See Masiero and Hensher (2011) for application of similar data on shift of reference point in a reference-dependent specification.

9. The criteria for setting transitional transport were derived from an in-depth phone survey of six of the most important shippers in the region. The criteria reflect the consequences of detour via the second best road alternative, namely, the San Bernardino road corridor.

10. The statistic population comprised 101 medium and 19 large firms operating in the manufacturing sector in Ticino (Swiss Federal Office, Neuchatel).

11. We adopted the test for non-nested models explained by Ben-Akiva and Swait (1986).

12. See Hensher, Greene, and Chorus (2011) for the formal derivation of elasticities in an estimated RRM model. A routine is available in NLOGIT to compute the RRM-based elasticities. We acknowledge the limitations of computing elasticity values in a stated preference experiment, which are computed in this paper to provide appropriate empirical evidence on the differences in the two choice paradigm assumptions (for further examples, see Hensher, Greene, and Chorus 2011; Thiene, Boeri, and Chorus 2012).

13. The elasticities based on MXL estimates can be linked to assumptions on distribution and are difficult to obtain and interpret. As such, the following comparison is based exclusively on estimates obtained from the RU-MNL and RR-MNL specifications.