Thesis
Statistical confounding and related issues in causal modeling
Honours, Murdoch University
2011
Abstract
In this thesis, we explore causal inference in observational studies with particular emphasis on the impacts of confounding in causal models. Failure to include confounding variables can result in significant bias when assessing the causal effect of an explanatory variables on a response, and the typical approach is to control for all potential confounders. While this is appropriate in most cases, there are instances where controlling for such variables does not reduce bias in estimates of association but actually increases bias. vVe examine a range of possible causal relationships among an explanatory variable, response. and potential confounders, demonstrating how bias can be avoided or induced by controlling for potential confounders. Additionally, we show how this bias can be quantified. Through a simulation study and application, we demonstrate how Monte Carlo approximation can be used to provide probabilistic evidence for whether controlling for a confounding variable will actually induce greater bias than not controlling for it, providing a mechanism by which to determine when to control for a potential confounder or not.
Details
- Title
- Statistical confounding and related issues in causal modeling
- Authors/Creators
- Elyse Corless
- Contributors
- Ryan Admiraal (Supervisor)
- Awarding Institution
- Murdoch University; Honours
- Identifiers
- 991005543205907891
- Murdoch Affiliation
- School of Mathematical and Physical Sciences
- Language
- English
- Resource Type
- Thesis
Metrics
10 File views/ downloads
64 Record Views