Date of Graduation
Doctor of Philosophy in Business Administration (PhD)
Second Committee Member
Causal inferences, confounder adjustment, economic significance, expected utility, model averaging, model uncertainty
This study examines the role of knowledge about underlying causal relationships in classifying controls in order to mitigate omitted- and included-variable biases. Using simulations and accounting examples, the study shows that the researcher may not distinguish good and bad controls because the underlying causal relationships are unobservable, and remedying strategies (such as the relative timing of measurement) may not remove the uncertainty in the classification. Because of the uncertainty about which controls to use, two or more models will be credible as will the distinct estimates derived from them. Next, the study shows that the current standard practice of singling out a preferred model (e.g., one about which the researcher is most confident) masks model uncertainty and is generally not supported by theories of rational decision making under uncertainty. Specifically, relying on the estimate from a preferred model will often involve higher risks in statements made from the study than relying on another estimate or combination of estimates. Accordingly, the study recommends that the researcher disclose all the relevant credible estimates without signalling which estimate is more important or choose an applicable theory that recommends which estimate to rely on in making conclusions and other important statements from their study.
Dusenge, D. (2022). Full Disclosure: Model Uncertainty in Adjusting for Confounders. Graduate Theses and Dissertations Retrieved from https://scholarworks.uark.edu/etd/4646