Answers are grounded in this paper and cite their sources.
Tumour size is an important prognostic factor in breast cancer, but its association with survival may be non-linear. Fractional polynomials provide a parsimonious approach for modelling such effects; however, the conventional Royston–Altman procedure selects powers from a prespecified set before estimating regression coefficients. We compared continuous joint maximum-likelihood (ML) estimation with conventional sequential fractional polynomial estimation for modelling tumour size in breast cancer survival. A total of 329,013 women with invasive breast cancer, diagnosed from 2010–2015, were analysed. A jointly estimated FP2 Royston–Parmar flexible parametric survival model was utilised. Predictive performance was evaluated using discrimination and calibration and integrated Brier score, and joint ML stability was explored using profile likelihood and Hessian diagnostics and multi-start optimisation. Covariate effect shapes, sample sizes, and degrees of censoring were assessed using Monte Carlo simulation. The two approaches showed comparable predictive performance, with C-indices of 0.8431 for both models and integrated Brier scores of 0.05936 and 0.05934, respectively. Joint ML FP power estimates showed weak identification and numerical sensitivity. Joint ML produced greater bias and RMSE, with a 95% Wald coverage from 48% to 76%. Sequential fractional polynomial estimation was more stable for point estimation, whereas joint estimation and profile-likelihood diagnostics might assist with uncertainty evaluation when the functional form is non-linear.
No takes yet. Share an insight, caveat, or question.
Atoyebi et al. (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: