Presentation (CTI)

OBJECTIVES: In health economic (HE) modelling for oncology, probabilistic sensitivity analysis (PSA) is critical for addressing uncertainty in survival estimates. Bootstrap resampling and Cholesky decomposition are widely used approaches for incorporating uncertainty in parametric survival models. While bootstrap is robust, it can be computationally demanding. In contrast, Cholesky decomposition is computationally efficient alternative leveraging the variance-covariance matrix of survival parameters. However, these methods have not been compared within the same modelling framework. This research addresses this gap by evaluating PSA estimates generated from these two approaches.This analysis aims to compare the PSA estimates from Cholesky decomposition and bootstrap method and assessing the efficiency of these options for use in Excel-based HE models.
METHODS: Parametric survival models were fitted to patient-level data for invasive disease-free survival (IDFS) as part of a semi-Markov model in early breast cancer. Two methods generated 1,000 samples of the survival parameters: (i) bootstrap resampling via repeated sampling and model refitting, and (ii) Cholesky decomposition via sampling from a multivariate normal distribution defined by the variance-covariance matrix. Comparisons were based on descriptive statistics, visual assessments (scatterplots and density plots), statistical tests and run time.
RESULTS: The comparison revealed high overlap in sampled parameters, ranging from 65.94% (log-normal distribution) to 99.44% (Gompertz distribution), with similar correlation. While most estimates were similar, density plots and descriptive statistics indicated potential differences. T-tests showed statistically significant differences (p-value <0.05) in mean parameter estimates for 3 out of 8 survival parameters between the two methods.
CONCLUSIONS: The similarity in PSA estimates between the two approaches highlights Cholesky decomposition as a practical and efficient method for capturing uncertainty, especially when computational resources are limited. Choleskey decomposition may potentially also lead to savings in model run time and file size. Simplifying mathematically complex models benefits HTA bodies and payers, making them easier to use in decision-making.