Methods of Adjusting for Exposure-affected Time-varying Confounding

Download files
Access & Terms of Use
open access
Copyright: Clare, Philip
Altmetric
Abstract
There is growing understanding that estimates of causal effects can be obtained from non-randomised studies if appropriate statistical techniques are employed, provided certain assumptions are met. However, doing so can introduce additional sources of bias that must be addressed. One such issue is that of exposure-affected time-varying confounding, where confounders of the relationship between an exposure and an outcome are themselves effected by past exposure. While methods capable of adjusting for such confounding exist, they remain underused in the literature. The thesis had two main aims: firstly, to examine and compare methods for handling exposure-affected time-varying confounding; and secondly to provide guidance on the implementation of these methods, to make them more accessible to applied researchers. The thesis presents: a systematic review of the literature of the use of the methods over a 16-year period; two simulation studies comparing different possible methods for adjusting for exposure-affected time-varying confounding, and handling missing data in analysis when such confounding is present; a tutorial on the use and implementation of targeted maximum likelihood estimation; and two applications of TMLE in the area of substance use research, an area of research which could benefit from greater use of these methods. The findings of the thesis support greater use of ‘doubly robust’ methods for causal inference in observational data. In order to promote this, it provides specific guidance on the use of TMLE, which performed well in simulations, and is also relatively easy to implement in R, using machine learning to avoid the need to manually specify models. The thesis also provides further support for the use of the models based on the findings of the applied research, which found that conclusions of TMLE models differed substantially from naïve analysis that did not account for exposure-affected time-varying confounding. Causal inference using observational data is possible, provided appropriate analysis methods are employed. Despite this, such methods have seen relatively little use in the literature, in part due to lack of familiarity and difficulty in implementation. This thesis provides evidence-based guidance on the use of the methods, and TMLE in particular, in order to make robust causal inference more accessible to applied researchers.
Persistent link to this record
Link to Publisher Version
Link to Open Access Version
Additional Link
Author(s)
Clare, Philip
Supervisor(s)
Dobbins, Timothy
Mattick, Richard
Bruno, Raimondo
Creator(s)
Editor(s)
Translator(s)
Curator(s)
Designer(s)
Arranger(s)
Composer(s)
Recordist(s)
Conference Proceedings Editor(s)
Other Contributor(s)
Corporate/Industry Contributor(s)
Publication Year
2020
Resource Type
Thesis
Degree Type
PhD Doctorate
UNSW Faculty
Files
download public version.pdf 2.58 MB Adobe Portable Document Format
Related dataset(s)