Application and development of statistical methods for prediction modelling in healthcare research

Archer, Lucinda

Abstract

This thesis explores methods and applications for prediction modelling in a healthcare setting, focusing on continuous outcomes, sample size, and external validation of model performance.

It begins by discussing current practices around the dichotomisation of birthweight, along with the issues associated with the dichotomisation of continuous outcomes prior to modelling. Methods are proposed to retain model development on the continuous outcome scale and subsequently generate predicted probabilities for the dichotomised outcome, if needed.

Models for continuous and dichotomised pain score are externally validated, demonstrating large uncertainty in statistical measures of predictive performance due to the small sample size available for validation. This motivates development of new sample size calculations to target precise estimation of performance when externally validating a clinical prediction model with a continuous outcome.

A further external validation shows precision in performance estimates when using individual participant data, combining data from multiple sources to boost the sample size for validating a prediction model for continuous birthweight. Each included cohort surpassed the minimum recommended sample size, based on the newly proposed methods, thus high precision could be expected. However, accounting for heterogeneity in performance across included populations through meta-analysis led to wider confidence intervals for pooled performance statistics than in any individual cohort.

Heterogeneity in model performance is further demonstrated in the external validation of a prediction model for serious falls. This validation involved utilising electronic health records to assess model performance across a range of general practice populations. Pooled performance estimates, though precise on average, hid the large variation in model performance across practices, giving an unrealistic summary of how the model might perform in practice.

In summary, the thesis demonstrates the importance of methodological rigour within clinical prediction model research, to ensure efficient and rigorous models are produced and evaluated.

Citation

Archer, L. (2025). Application and development of statistical methods for prediction modelling in healthcare research. (Thesis). Keele University. Retrieved from https://keele-repository.worktribe.com/output/1109632

• Thesis Type: Thesis
• Deposit Date: Mar 21, 2025
• Publicly Available Date: Mar 21, 2025
• Public URL: https://keele-repository.worktribe.com/output/1109632
• Award Date: 2025-03