Publication Detail
Causal Modeling of Travel Behavior Using Simultaneous Equations Systems: A Critical Examination
|
UCD-ITS-RR-92-19 Research Report Alumni Theses and Dissertations |
Suggested Citation:
Pendyala, Ram M. (1993) Causal Modeling of Travel Behavior Using Simultaneous Equations Systems: A Critical Examination. Institute of Transportation Studies, University of California, Davis, Research Report UCD-ITS-RR-92-19
Over the past two decades, increasing attention in travel behavior research has been directed toward the development and estimation of simultaneous equations systems which allow one to model relationships among multiple dependent variables. It has been conjectured that such systems are able to better replicate causal relationships underlying behavioral phenomena.
In economics and the biological sciences, model specification is driven by well-established theories. In the behavioral sciences where well-established theories are scarce, researchers have been estimating simultaneous equations models to identify causal theories by statistically testing causal hypotheses. This dissertation addresses several issues regarding the adoption of such an exploratory approach.
The first is concerned with structural heterogeneity, where different behavioral units may have different causal structures driving their decision making processes. Most modeling efforts assign the same causal structure to all behavioral units and use resulting model parameter estimates to obtain forecasts. This dissertation determines the validity of such an approach by examining the sensitivity of model parameters to the presence of structural heterogeneity. It is found that parameter estimates are highly biased and inefficient in the presence of structural heterogeneity.
Second, this dissertation examines whether a causal structure can be uniquely identified on a given data set. Most simultaneous equations systems can be algebraically manipulated to represent several entirely different causal structures. As these structures are mathematically identical, they would provide identical statistical indications when applied to the same data set, making it impossible to uniquely identify the true causal theory. This dissertation corroborates this argument and shows that tests of competing causal hypotheses can never be conclusive, if based on statistical indications.
Finally, the adequacy of method-of-moment estimation software packages, such as LISCOMP, developed for simultaneous equations systems is examined. These packages are being used to perform exploratory causal analysis and test competing hypotheses. However, finite sample properties of parameter estimates obtained using these packages, under varying conditions of exogenous variable distributions and error correlations, are unknown. This study finds that LISCOMP provides the worst results when it is most needed; in the presence of irrelevant parameters, categorical exogenous variables, and small sample sizes.
In economics and the biological sciences, model specification is driven by well-established theories. In the behavioral sciences where well-established theories are scarce, researchers have been estimating simultaneous equations models to identify causal theories by statistically testing causal hypotheses. This dissertation addresses several issues regarding the adoption of such an exploratory approach.
The first is concerned with structural heterogeneity, where different behavioral units may have different causal structures driving their decision making processes. Most modeling efforts assign the same causal structure to all behavioral units and use resulting model parameter estimates to obtain forecasts. This dissertation determines the validity of such an approach by examining the sensitivity of model parameters to the presence of structural heterogeneity. It is found that parameter estimates are highly biased and inefficient in the presence of structural heterogeneity.
Second, this dissertation examines whether a causal structure can be uniquely identified on a given data set. Most simultaneous equations systems can be algebraically manipulated to represent several entirely different causal structures. As these structures are mathematically identical, they would provide identical statistical indications when applied to the same data set, making it impossible to uniquely identify the true causal theory. This dissertation corroborates this argument and shows that tests of competing causal hypotheses can never be conclusive, if based on statistical indications.
Finally, the adequacy of method-of-moment estimation software packages, such as LISCOMP, developed for simultaneous equations systems is examined. These packages are being used to perform exploratory causal analysis and test competing hypotheses. However, finite sample properties of parameter estimates obtained using these packages, under varying conditions of exogenous variable distributions and error correlations, are unknown. This study finds that LISCOMP provides the worst results when it is most needed; in the presence of irrelevant parameters, categorical exogenous variables, and small sample sizes.
Ph.D. Dissertation