Estimating the Actual Effect of the Built Environment on Travel Behavior in the Context of Residential Self-Selection: A Comparison of Method

The influence of the built environment (BE) on travel behavior (TB) is of considerable interest to transportation policy makers and land-use planners because the BE is an obvious limiting factor on whether or not individuals even have opportunities to make certain decisions with respect to their behavior. Although many studies have found that suburban residents tend to drive more and walk less than residents of traditional neighborhoods, it is less clear to what extent the built environment directly influences their travel behavior, compared to what extent residents "self-select" into their built environment based on prior predispositions and attitudes. This bias in the estimated effect of the BE on TB that results from not appropriately separating the "true" influence of the built environment from (often unobserved) attitudes that may affect the choice of residential location in the first place has commonly become referred to as residential self-selection (RSS) in the literature. By now, many studies have identified RSS as an issue, and various techniques have been applied to account for its effects. Empirically, however, the findings from the applications of these varied approaches are far from unani¬mous, and the reasons for this are not clear. The question driving the proposed study is, to what extent is the variability in empirical outcomes due to differences in the approach used to account for self-selection, and the method (formula) used to quantify the effects of self-selection?;Among the different approaches for dealing with self-selection, I have chosen three as being of particular interest for this dissertation: statistical control (SC) modeling, propensity score-based techniques (PS), and sample selection (SS) modeling. Statistical control focuses on self-selection that arises from an omitted variables bias. Both propensity score-based techniques and sample selection modeling focus on the idea that self-selection arises from non-random assignment into treatment and control groups. In propensity score regression (the main focus in this study), the propensity score acts as both a substitute for the socioeconomic and attitudinal variables that enter the propensity score equation, and as a sort of residential choice (built environment) variable. In sample selection modeling (as implemen¬ted in this study), outcomes are modeled separately for each residential choice. In theory, sample selection models control for selection on unobservables (either by including an auxiliary term in the outcome equations, or by estimating both selection and outcome equations simultaneously) as well as observables, whereas propensity score techniques (as well as statistical control) only control for selection on observables.;The principal measure of interest in this study is the proportion of the total apparent effect of the built environment on travel behavior that is due to the built environment itself (as opposed to RSS), which I call the "built environment proportion", or BEP.;There were three main objectives of this study. The first objective was to (i) identify a number of plausible methods for estimating the key quantity of interest to this study, namely, the BEP; and (ii) devise a framework for analyzing and comparing these various BEP calculation methods. With respect to (i), I present and evaluate three categories of methods of estimating this principal measure of interest (each native to one of the approaches): variance-explained, modular effects, and treatment effects. A BEP formula associated with a given method can be applied to its native approach or cross-applied to other approaches, sometimes with subtle changes due to the nuances of the particular approach to which the formula is being cross-applied, leading to many different possible values for the BEP. Ultimately, I identify and enumerate 47 potential BEPs, of which 28 were found to be useful for the empirical analysis (the other 19 were either found to be conceptually flawed in my context, or resulted in BEPs outside the valid range of 0 to 1). With respect to (ii), I develop a systematic "framework" for comparing the estimated values of the BEPs across approach and method, with respect to three dimensions. The first dimension compares BEPs computed on a calibration sample to those obtained by applying the calibrated models to a validation sample. The second dimension compares BEPs obtained from models in which each approach's specification is based only on the best possible specification in terms of statistical significance and model fit, to those obtained from models having as close to the same specification as possible across approaches. The third dimension is related to the first: BEPs computed based on a single (partially-)random split of the data into calibration and validation (training and test) samples are compared to the averages obtained from 1000 such random splits. (Abstract shortened by ProQuest.).