Using physician’s prescribing preference as an instrumental variable in comparative effectiveness research

Zhang, Lisong (2023) Using physician’s prescribing preference as an instrumental variable in comparative effectiveness research. PhD thesis, University of Glasgow.

Abstract

Background:
Comparative effectiveness research (CER) studies using non-randomised study designs sometimes employ instrumental variables (IVs) to address the problem of unmeasured confounding. Physician’s prescribing preference (PPP) is a commonly used IV in this context and had been shown to have utility in many CERs. However, these IVs are generally used as a supplementary method rather than the main analytical strategy. In this thesis, I aim to test the validity of PPP IVs, including an evaluation of the different ways they can be constructed to help promote their more widespread use in CER.

Methods:
This thesis consists of a range of underpinning methodological approaches, including a literature review summarising applied and simulation studies between 2005 and 2020 that use PPP IV in CER; applied CERs using PPP IV in studies utilising routinely-collected health datasets; target trial emulation approaches based on benchmarking from a randomised clinical trial; and simulation studies to test the performance of PPP IV in multiple CER settings.

Results:
My literature review provides guidance on the further use of physician’s prescribing preference as instrumental variables in comparative effectiveness research. It highlighted that practical use of PPP needs to consider the findings from simulation studies in the area. In my empirical chapters, I provide strong evidence that PPP is a valid IV approach for conducting CERs using non-randomised study designs. I found that constructing PPP using longer prescription histories generally produces stronger instruments, which in turn leads to greater precision in estimation of treatment effects. In practice, validation of assumptions is crucial for the utility of IVs in CER. In my applied research, I found strong real-world evidence that supports diazepam is associated with lower risk of rehospitalisation and mortality due to the alcohol intoxication and harmful than chlordiazepoxide; that disulfiram is superior to acamprosate in terms of preventing alcohol dependence-related hospitalisations; and that sulfonylureas (SU) performs better than dipeptidyl peptidase-4 inhibitor (DPP-4 inhibitor) in reducing HbA1c levels as the second-line treatment for Type-2 diabetes patients. In my simulation studies, I found PPP IV, when unmeasured confounding exists, can produce less biased estimates of treatment effects than conventional multivariable regressions that only adjust for measured confounding variables, albeit with lower statistical power. The simulations also show PPP IV has potential in alleviating noncollapsibility in non-linear IV approaches.

Implications:
Findings from this thesis indicate that PPP IVs can be valid IVs and reduce unmeasured confounding in observational CER studies. However, I have found that there is room for improvement in the application of PPP IV in CER studies; researchers need to pay more attention on validating IV assumptions and carefully consider how different formulations of PPP IVs can be applied in order to improve the quality of statistical inference. Future applied PPP IV research should consider findings from relevant simulation studies to inform study designs and analysis plans. Conversely, one also needs information on PPP IVs from empirical studies to inform future simulation study design and to gain further knowledge from triangulation between applied and simulation findings. Many of my thesis findings can be generalised to the use of non-PPP IV approaches in CER.

• Item Type: Thesis (PhD)
• Qualification Level: Doctoral
• Subjects: H Social Sciences > HA Statistics; R Medicine > RA Public aspects of medicine
• Colleges/Schools: College of Medical Veterinary and Life Sciences > School of Health & Wellbeing > Health Economics and Health Technology Assessment
• Supervisor's Name: Lewsey, Professor Jim and McAllister, Professor David
• Date of Award: 2023
• Depositing User: Theses Team
• Unique ID: glathesis:2023-84016
• Copyright: Copyright of this thesis is held by the author.
• Date Deposited: 09 Jan 2024 09:07
• Last Modified: 09 Jan 2024 10:41
• Thesis DOI: 10.5525/gla.thesis.84016
• URI: https://theses.gla.ac.uk/id/eprint/84016