Content area
Abstract
This thesis motivates, develops, and applies a technique for automatically searching for alternative causal explanations of statistical data. The search technique is embodied in a computer program, TETRAD, and it is based upon a new and provably correct algorithm for calculating testable features of any linear model expressable as a directed graph. "Causal modeling", a set of quantitative techniques for modeling the causal structure of a social system in nonexperimental settings, is commonly employed by researchers in psychology, sociology, political science, econometrics, epidemiology, biology, and other disciplines. The thesis begins by considering the philosophical foundations of the causal modeling formalism. It then attempts to address the most serious problem facing causal modelers, that of searching for a good model in a space that routinely numbers in the billions and up. There are two kinds of information one can use in performing such a search, substantive information about the domain under study and structural relations between the models and the data. The technique I develop combines both. It allows a researcher to specify a simple initial model, and then uses structural criteria, taken partly from the history and philosophy of science, to automatically search for elaborations to this initial model. The program that embodies this technique, TETRAD, is then applied to simulated and empirical cases. In simulated cases, data is generated from a known model. A fraction of the model is given to a TETRAD user along with the data; the task is to recover the model that generated the data. The TETRAD user is successful in each case. In every empirical case attempted, TETRAD helps to find an alternative to the one published that is as plausible and that fares as well or better on standard statistical tests. Finally, I suggest a framework for evaluating the capacity of any search procedure (based on the features of linear models TETRAD computes) to distinguish among different causal models. Given a certain class of initial models that are subgraphs of the model that generated the data, one can detect which pair of variables ought to receive extra causal connection. One can also, for certain pairs of variables, uniquely determine the causal order of extra causal connection.
