Impacts of survey design and model specification on willingness-to-pay estimates from discrete choice models

Discrete choice models infer individuals’ preferences from observed choices. Analysts can thereby contribute to producing more reliable demand forecasts and assess welfare impacts of policy/scenario changes. However, the risk of model misspecification errors may bias parameter estimates and lead to incorrect demand forecasts and policy recommendations. This thesis examines three types of model misspecifications: i) ignoring travel time constraints, ii) measurement error in the income variable, and iii) ignoring the behavioural phenomenon of the zero-price (ZP) effect. We are particularly interested in understanding the policy implications of these misspecifications on the marginal valuation of qualitative variables. Our analyses are relevant to policy makers as these specification errors prevail in some ‘state-of-the-practice’ model representations commonly used in support of cost-benefit analyses. This thesis first examines the issue of ignoring travel time constraints for simple time-cost trade-offs. Analysts may ignore that some alternatives are not available to individuals as the travel times presented could exceed their time allowances for such journey. We find via simulation that the value of travel time (VTT) can be significantly over-estimated when travel time constraints are not accounted for in estimation. More importantly, we identify the confounding issue between travel time constraints and taste heterogeneity. This thesis then turns to the issue of the measurement error in the income variable. We investigate the extent to which the income measure used in the estimation of choice models contributes to the disparity between the cross-sectional and inter-temporal income elasticity of the VTT. We compile various income measures that are varied in terms of the income re-distribution measures and the intra-household budget allocation based on secondary expenditure data. We empirically test the new income measures based on the modelling framework developed for the 2014/15 UK VTT study. Our results indicate that by additionally accounting for social benefits, the cross-sectional income elasticity of VTT approaches unity. This closes the gap between the cross-sectional and inter-temporal income elasticity. We find the behavioural VTTs, which represent the averages of the VTTs estimated from behavioural models across respondents, to be consistent despite the income variations. However, we find that when moving from the stated choice (SC) to the national travel survey to obtain a nationally representative figure for appraisal, appraisal values diverge as per the income variations due to the sampling bias in the income variable in behavioural model. We highlight the requirement for the sampling of the income to be consistent between the estimation and implementation tool. We finally explore the issue of ignoring the ZP effect in choice modelling. ZP effect is a well-established notion in behavioural economics which explains the tendency for individuals to over-react to free alternatives. The lack of attention to the ZP effect in the choice modelling literature is particularly worrying since ‘free’ status quo (SQ) alternatives are at the heart of many SC surveys, especially outside of transport, and form the basis of contrasting the (policy) ‘interventions’. We develop alternative stated survey designs to identify the ZP effect. We find that the observed preference for remaining at the SQ is largely attributed to the ZP effect within our data. We also present experimental design features that allow separation of the ZP effect from the non-linear cost sensitivity. We stress that the prevalence of the ZP effect in observed choice behaviour may introduce bias to the prediction of welfare when the perfect confounding between the ZP and SQ effects is broken. Overall, this thesis highlights the significant bias on WTP estimates that may be caused by ignoring some basic and fundamental misspecification issues. This thesis closes by suggesting some future improvements required to avoid model misspecification issues identified.