International Choice Modelling Conference, International Choice Modelling Conference 2017

Font Size: 
Reference-dependent choice set formation for modelling of long-term and medium-term residential location choices
Md Bashirul Haque, Charisma Choudhury, Stephane Hess

Last modified: 28 March 2017


Modelling of residential location choice is complex due to numerous methodological and empirical challenges, a key one being choice set formation. Large scale residential location choice modelling using revealed preference data needs to consider a huge number of choice alternatives (several hundred in zone level models and millions in dwelling level models). The estimation of a model with such a massive choice set is computationally difficult or in some cases intractable. More importantly however, doubts can also be expressed about the behavioural realism of considering all alternatives when it is clear that a decision maker will likely consider just a limited subset of areas/properties. However, in many of the residential location choice models, the full choice set has been considered for each respondent [1]. Random sampling techniques have also been used in several applications to estimate the model with a subset of alternatives [2]. Sampling is helpful in reducing computational complexity and will give consistent parameter estimates if the universal choice set is the true choice set of individual decision makers.

The universal choice set consisting all zones (in zone level model) or dwellings (in dwelling level model) in the study area is however unlikely to be the true choice set of households. In practice, households are neither aware of the full choice set of alternatives nor consider all alternatives they are aware of. Different households might thus have different consideration sets based on household preferences, sociodemographic characteristics and their knowledge of available alternatives. Therefore, it is expected that better treatment of the choice set will make the models computationally easier and behaviorally more representative.


Probabilistic choice set formation approaches are infeasible when the universal choice set is very large because the number of possible choice set increases exponentially with the number of alternatives (2M-1). Therefore,  deterministic constraint approaches have been used widely in residential location choice modelling to eliminate unrealistic alternatives from the choice set [3].  Zolfaghari [3]  compared the performance of alternative exogenous choice set formation approaches and he found poor performance of deterministic constraint based choice set generation techniques. From the evidence in the published literature, the development of appropriate approaches for choice set formation in the context of residential location choice modelling with a large set of alternatives still remains a gap in the literature. In this paper, we propose an alternative approach for choice set generation.

The dataset used in this research is generated by combining London household survey data, ward atlas data set and data from the London transport studies model. In the data set, 62% households are found to choose their new dwelling in the same borough as their previous dwelling and most of the households from the rest are found to choose their new dwelling in nearby boroughs. This finding gives a clear indication of household preference to the residential location alternatives in their current location or in nearby zones. However, this preference is not absolute.

On the basis of the above, a reference-dependent (current location) choice set formation technique is adopted in this research to increase the propensity of the alternatives close to current location to be considered in the choice set. A distance threshold from current location is applied for sampling of a higher portion of alternatives from the area covered by threshold distance and a lower portion of alternatives from the other part of the study area. The proportion of alternatives chosen from inside and outside of the sampling zones is optimized by several trials. Two alternate approaches are applied to capture the optimum threshold distance from the current location.  In the first approach, the distance between the reference location and the new location (chosen alternative) of all households in the data set are measured and a distance which includes the reference location and the chosen alternative of most of the households is selected as a threshold point. In the second approach, unique distance threshold is applied for each individual household based on the distance between the reference location and the new location (chosen) of the respective household.

The decision of residential location is broadly categorized into ownership (considered as long term decision) and renting (considered as medium term decision). Preferences of households to these difference time scale choices are expected to be different, as confirmed by earlier research by the authors[4]. The proposed method of choice set formation is applied for modelling of residential ownership and renting decision to investigate the heterogeneity of underlying preference structure in the consideration set of owners and renters.

Models are estimated using the full choice set (where electoral ward boundaries in Greater London Area are considered as location alternatives), a random subsample and choice set generated using the proposed method. In all cases, substantial heterogeneity has been observed in the choices of owners and renter but the reference-dependent choice set approach demonstrates a better gain in the model fit due to capturing underlying preference structure. We also undertook a validation test to evaluate the performance of the proposed approach. Despite having behavioural credibility, the proposed method of choice set generation can be improved by applying techniques for filtering of alternatives based on household preferences on commute distance and dwelling price along with reference location which remains as a scope for future work. The proposed technique of choice set generation can be used effectively in the context of residential location choice modelling where the probabilistic approach is infeasible due to a large number of alternatives.



1.            Bhat, C.R. and J. Guo, A mixed spatially correlated logit model: formulation and application to residential choice modeling. Transportation Research Part B: Methodological, 2004. 38(2): p. 147-168.

2.            Lee, B.H. and P. Waddell, Residential mobility and location choice: a nested logit model with sampling of alternatives. Transportation, 2010. 37(4): p. 587-601.

3.            Zolfaghari, A., Methodological and empirical challenges in modelling residential location choices. 2013.

4.            Haque, M.B., C. Choudhury, and S. Hess, Investigating the differences between long-term and medium- term residential location choices: a case-study of London. Submitted in Transport Research Board Conference 2017.

Conference registration is required in order to view papers.