Conroy, Elizabeth 
ORCID: 0000-0003-4858-727X
(2022)
Learning and clustering: statistical adjustment for the learning curve and clustering effects in randomised surgical trials.
PhD thesis, University of Liverpool.
|
Text
200447797_Dec2021.pdf - Unspecified Download (5MB) | Preview |
Abstract
The need for more, and better, randomised surgical trials is well recognised, and recently the number of surgical trials has grown. Rigorous design and analysis of such studies is important to support clinical decision making. Two associated methodological challenges are clustering effects, by centre and surgeon, and the surgical learning curve. This thesis aims to improve the design and analysis of surgical trials by investigating existing guidance, establishing current practice and demonstrating how design and analysis can incorporate clustering and learning. Existing guidance for managing these challenges exist, but this work identifies that they focus more on design than analysis. As there is no single document, triallists must access multiple documents to gain full understanding, ultimately leading to inconsistencies in practice. Two novel reviews of surgical trials are undertaken covering a time period of remarkable growth of surgical trials. The first, of 247 published trials, demonstrates that clustering and learning considerations are underreported, methods used to do so vary and reporting guidelines are poorly adhered to. It is recommended that triallists report these methods, or justify where not, to support results interpretation. Early consideration of these effects is vital. The second, comprising fifty funded grant applications by a leading UK funder, identifies early consideration of these effects and the funder as a potential driver of better practice. Recommendations are provided about when and how to address surgical learning and clustering in the design and analysis. To complete understanding of current practice, forty-seven statisticians from UK clinical trials Units were surveyed. Widespread awareness of challenges in design and analysis are identified. Approaches used to manage clustering and learning varies both across and within Units, suggesting that agreed principles, across a range of trial scenarios, are needed. A number of real surgical trials, varying by intervention and setting, are presented as a practical demonstration of approaches to design and analysis. Statistical methods for exploring the presence of clustering and learning, by centre and surgeon, and any impact on trial conclusions are demonstrated using real trial datasets. For clustering, simulated data were used to explore the impact of clustering under different scenarios. Clustering became a greater concern as the intraclass correlation and true treatment difference increased. For learning, a curve was identified but it did not impact trial outcomes. Developing better measures of learning for use within such explorations is recommended. Good design can minimise the impact of clustering and learning, but statistical methods that fail to account for these effects, if present, can lead to biased treatment estimates and reduced power. Clustering and learning should be managed using a design and analysis approach. Considerations should be made early, and on a trial-by-trial basis, to ensure that the trial conclusions are valid.
| Item Type: | Thesis (PhD) |
|---|---|
| Divisions: | Faculty of Health and Life Sciences Faculty of Health and Life Sciences > Institute of Population Health |
| Depositing User: | Symplectic Admin |
| Date Deposited: | 04 Aug 2022 09:24 |
| Last Modified: | 18 Jan 2023 21:00 |
| DOI: | 10.17638/03155529 |
| Supervisors: |
|
| URI: | https://livrepository.liverpool.ac.uk/id/eprint/3155529 |
Dimensions
Dimensions
