A pivotal component of prognosis research is the prediction of future outcome risk. This thesis applies, develops and evaluates novel statistical methods for development and validation of risk prediction (prognostic) models. In the first part, a literature review of published prediction models shows that the Cox model remains the most common approach for developing a model using survival data; however, this avoids modelling the baseline hazard and therefore restricts individualised predictions. Flexible parametric survival models are shown to address this by flexibly modelling the baseline hazard, thereby enabling individualised risk predictions over time. Clinical application reveals discrepant mortality rates for different hip replacement procedures, and identifies common issues when developing models using clinical trial data. In the second part, univariate and multivariate random-effects meta-analyses are proposed to summarise a model’s performance across multiple validation studies. The multivariate approach accounts for correlation in multiple statistics (e.g. C-statistic and calibration slope), and allows joint predictions about expected model performance in applied settings. This allows competing implementation strategies (e.g. regarding baseline hazard choice) to be compared and ranked. A simulation study also provides recommendations for the scales on which to combine performance statistics to best satisfy the between-study normality assumption in random-effects meta-analysis.
Identifer | oai:union.ndltd.org:bl.uk/oai:ethos.bl.uk:669063 |
Date | January 2015 |
Creators | Snell, Kym Iris Erika |
Publisher | University of Birmingham |
Source Sets | Ethos UK |
Detected Language | English |
Type | Electronic Thesis or Dissertation |
Source | http://etheses.bham.ac.uk//id/eprint/6259/ |
Page generated in 0.0017 seconds