Importance of Dwelling, Neighbourhood Attributes in Residential Location Modelling: Best Worst Scaling vs. Discrete Choice

Abstract

Traditional discrete choice experiments do not differentiate between the intrinsic importance of an attribute and that associated with its levels of variation. It has been suggested recently that best-worst (B-W) scaling (Case 2) allows for this differentiation. Here we pool B-W answers with binary stated choice (SC) data to study the importance of dwelling and neighbourhood attributes for apartment seekers in the centre of Santiago, Chile. Previous research had shown how these diverse elicitation methods can be pooled (albeit without including any type of heterogeneity), suggesting that the “best” (as opposed to “worst”) responses are most compatible with the binary SC data. In this paper we extend this work to allow for heterogeneity in preferences through (a) systematic taste variations alone, (b) correct treatment of panel effects alone, and finally, (c) the combined effect of both. In all cases the best resulting model is obtained by pooling the “best” answers with the binary SC, under the assumption of common and specific attributes to each dataset. Nevertheless, when the model included only unobserved heterogeneity through error components (to treat the panel effect), the datasets did not pool as well as when we did not consider it. The joint model had half of the attributes specific to each dataset while in the previous case only two were specific. On the other hand, when considering observed and unobserved heterogeneity, the two datasets pooled better than in the other cases, needing to consider only one attribute as specific to each set (and the remaining seven as common). We also analysed how the inclusion of heterogeneity changed the attribute estimates. In particular, the scales that can be constructed for the attributes and for their levels of variation do not change their relative order, but the magnitudes do, especially in cases where the estimates were not significantly different from zero at the 95% confidence level.