Structural Causal Modeling for Advancement of Information System Research
thesis
posted on 2024-08-01, 00:00authored bySujin Park
A causal structure is a property of the physical world. The concept of "causal" is distinguished from "associational" given the continuity from the cause to the effect. Causation is not about co-occurrence but about the precedence of cause to effect. For decades, substantial methodological efforts have been made to observe and capture such a causal process. Randomized controlled trials (RCT) or experimental designs are perhaps the most rigorous method to make causal inferences. By minimizing potential bias and confounding, RCTs enable us to disentangle true causal links from spurious relationships and explore the underlying causal mechanism. Observational designs such as natural experiments and quasi-experiments, econometric models utilizing instruments and control variables, or propensity score matching, are other methods for causal inference. Even though the manipulation of the treatment (i.e., cause) is not under our control, these approaches mimic the randomization process so that some biases inherent in the observational data can be reduced.
However, some practical considerations often limit the use of these methods. Given the non-negligible monetary and human costs, it can be challenging to conduct randomized experiments. Experiments might sometimes raise ethical and privacy concerns, especially when the research subjects are humans. In addition, the observational designs involve stringent assumptions that are often difficult to justify; for example, to be an instrument, variables need to satisfy the assumptions that the instrument is related to the exposure of interest but not directly to the outcome and is independent of confounders. Causality has great potential to transform the way we solve a variety of real-world problems, but its methodological challenges may limit our efforts to navigate numerous research opportunities. In this dissertation, we explore the recent methodological advance in causal inference and discuss its potential applicability in various areas. In Chapter 1, we introduce a novel methodology called "transportability" and discuss some practical considerations for the application of transportability while providing a detailed procedure for its implementation. In Chapter 2, we extend the applicability of the transportability framework and focus on the potential usefulness of the transportability method in enhancing the replicability of randomized experiments. Chapter 3 demonstrates how applicable the fundamental idea of the transportability framework is by considering the other potential fields for which using this causal notion can be beneficial, i.e., improving machine learning and AI models' generalizability and ethical capability.
History
Advisor
Ali Tafti
Department
Information Decision and Sciences
Degree Grantor
University of Illinois Chicago
Degree Level
Doctoral
Degree name
Doctor of Philosophy
Committee Member
Mary Beth Watson-Manheim
Yingda Lu
Elena Zheleva
Galit Shmueli