Automated causal modeling, latent constructs, and abductive inference

In the traditional scientific framework, a scientist conjectures a hypothesis and then conducts an experiment to test it. However, a new paradigm advocated by Clark Glymour and his associates propose a method of causal discovery. In their methodology that is in a similar vein to data mining, large data sets are collected and automated algorithms are employed to exhaust virtually all possible combinations of the relationships among variables. Hypotheses about how variables are related are generated and tested along the way. Glymour proposes that the model identified as the best fit to the data by the automated search method is to be considered the causal conclusion. To substantiate this assertion, Glymour and colleagues developed Tetra Difference (TETRAD), a plug-in module for standard Structural Equating Modeling (SEM) software applications. This dissertation challenges the claim that automated induction, especially data mining by TETRAD, constitutes a paradigm shift in causal discovery, because TETRAD focuses on the structural aspect of causal modeling but overlooks how conceptualization affects projection and causal inferences. On the contrary, the abductive approach recognizes the role of conceptualization in every step of inquiry. A thoughtful researcher should always keep an open mind about whether these constructs are subject to revision based on new information. In conclusion, a thorough inquiry with respect to causal inferences should include abduction, deduction, and induction. Abduction aims to suggest new constructs or plausible causal hypotheses; deduction builds a logical and testable model based upon plausible premises; and induction assesses the adequacy of the hypothesized model with empirical data. Tangible applications of this integration were demonstrated in Evidence-Centered Design (ECD) and Construct-map item banking.