Assessment and correction of endogeneity problems in discrete choice models

Abstract

The term endogeneity is used when there is a correlation between one or more observed explanatory variables (independent variables) and the error term of an econometric model. Endogeneity is considered a practically inevitable phenomenon in econometric modelling, as there are many potential causes behind it: omitted variables, measurement or specification errors, simultaneous estimation and self-selection. The problem is that it may give rise to inconsistent parameter estimates, and if its effects are not considered when estimating a model, the analyst may come to wrong forecasts and conclusions. Correcting for endogeneity has been widely addressed in the linear models (LM) literature, but LM have a limited scope in certain areas. This is particularly the case in planning and social evaluation of transport projects, where Discrete Choice Models (DCM), which are highly non-linear, play a fundamental role. Unfortunately, DCM are not often corrected for endogeneity, so a gap has been identified in the state of knowledge that this thesis intends to close. Thus, the general aim of this Ph.D. dissertation is to develop a set of guidelines that allow for the assessment and correction of endogeneity problems in DCM. We establish conclusions of a theoretical, empirical and methodological nature. In the first instance, it is desired to determine adequate instrumental variables for endogeneity correction in transport modelling and measure the impact of this correction on strategic modal split models. We can reduce the errors associated with the estimation of DCM, improve its forecasting capabilities, and achieve consistent parameters resulting in corrected estimates of model valuation measures, such as the subjective value of time (SVT). Furthermore, we formulate an empirical methodology, supported by Monte Carlo simulation, to predict using DCM corrected for endogeneity with a new and more adequate version of the CF method. We also define guidelines to clarify under what conditions discrete indicators work (or not) when DCM are corrected for endogeneity using the MIS method. Finally, we structure a methodology to detect weak DCM instruments based on what has been proposed for linear models