Robustness of RRM modelling outcomes towards omitted attributes
As discrete choice models may be misspecified, it is crucial for choice modellers to have knowledge on the robustness of their modelling outcomes towards misspecification. One type of model misspecification occurs when not all attributes that are relevant for the choice are included in the choice model. Such omission of attributes may result in larger finite sample bias, possibly leading to erroneous modelling outcomes such as poor market share forecasts and inaccurate estimates of demand elasticities.
The robustness of RUM-based choice modelling outcomes towards the omission of a relevant attribute has frequently been studied. Contrary to RUM-based choice modelling outcomes however, the robustness of RRM-based choice modelling outcomes towards the omission of a relevant attribute has not been studied. Questions such as “are RRM modelling outcomes relatively less robust towards omitted attributes as compared to RUM modelling outcomes, e.g. due to the fact that RRM models account for context effects?”, or “are RRM relatively robust towards one sort of omitted attribute, but little robust towards another?” are yet unanswered. This lack of understanding currently hampers adequate interpretation of RRM modelling outcomes.
To investigate the robustness of RRM modelling outcomes towards the omission of a relevant attribute we conducted several Monte Carlo experiments. Acknowledging that in real life the ‘true’ decision rule is inherently unknown to the choice modeller, choices are generated based on RUM and RRM. Furthermore, to enhance the interpretation of the results, we investigated the effects of the omitted attribute on the robustness of the RRM modelling results alongside with the effects of the omitted attribute on the robustness of the RUM modelling results
In particular, the impacts of the omitted attribute on implied elastcities are explored.
RRM models are found to be fairly robust towards the presence of an orthogonal omitted attribute, though not as robust as the linear-additive RUM model. The differences in robustness between the RUM and RRM models are however quite subtle, and occur only in the quite stylized situation in which the omitted attribute is orthogonal. Considering the fact that regret is conceived to emerge in a more complex way than utilities, this results seem intuitive. See Van Cranenburgh, S. & Prato, C.G. (2016) for more details on this study.
Van Cranenburgh, S. & Prato, C.G. (2016) “On the robustness of Random Regret Minimization modelling outcomes towards omitted attributes.” Journal of Choice modelling(18) 51-70.