Flexible Parametric Models Top Survival Extrapolation Test in FDA Oncology Trial Analysis

A new analysis published in BMC Medical Research Methodology examined which of six families of parametric survival models performs this extrapolation best, using reconstructed individual patient data drawn from oncology trials that supported Food and Drug Administration-approved indications between 2014 and 2023. The study’s central message is a cautionary one: the goodness-of-fit statistics most commonly used to choose a model, the Akaike and Bayesian Information Criteria, showed essentially no correlation with how accurately a model extrapolates survival into the long term.

The research team, led by Mingye Zhao, Taihang Shao and Hanqiao Shao with colleagues including Wenxi Tang, included 109 trials that provided 346 updated Kaplan-Meier survival curves. They then compared how six model classes performed: standard parametric models, fractional polynomials, restricted cubic splines, Royston-Parmar models, generalized additive models and parametric mixture models.

The evaluation rested on two complementary yardsticks. Direct goodness-of-fit measured how closely each model’s extrapolated curve matched the updated Kaplan-Meier curve, quantified by the mean absolute error in months per year and by the difference in restricted mean survival time. Relative goodness-of-fit asked a different question: whether a model preserved the treatment effect, the survival difference between regimens, which is often the quantity that matters most for deciding whether a new therapy is worth funding. The authors also defined a combined metric blending within-sample fit and extrapolation, and used Wilcoxon signed-rank tests for formal model comparisons, with subgroup and regression analyses to probe robustness.

The headline result concerns the information criteria. Spearman correlations between within-sample fit and extrapolation accuracy were near zero, with coefficients below 0.1 for both AIC and BIC. In practical terms, a model that fits the observed trial data beautifully may extrapolate poorly, and vice versa. The study suggests that relying solely on these criteria is insufficient, particularly when choosing across different classes of parametric models or among flexible parametric approaches.

Against that backdrop, one model family stood out. In the direct extrapolation comparison against standard parametric models, which had a median mean absolute error of 0.446 months per year, the median error difference for Royston-Parmar models was -0.032 months per year, a 95 percent confidence interval of -0.064 to -0.001 confirming a statistically significant advantage. Every competitor fared worse: parametric mixture models added a median 0.097 months per year of error, fractional polynomials 0.134, generalized additive models 0.192 and restricted cubic splines 0.246, each with confidence intervals excluding zero.

The relative analysis reinforced the ranking. Model performance judged by the combined within-sample and extrapolation metric was consistent with the pure extrapolation results, and sensitivity analyses, including subgroup analyses by clinical characteristics and regression analyses, confirmed the robustness of the findings.

Yet on this large, real-world dataset of oncology trials, the flexible approach delivered the most accurate long-term predictions by both direct and relative measures. The authors conclude that these models offer the best extrapolation accuracy and hold potential for long-term survival predictions in oncology.

The study’s scope and limitations deserve emphasis. The evidence base consists of oncology trials supporting FDA-approved indications with updated Kaplan-Meier curves available in the 2014 to 2023 window, and the individual patient data were reconstructed from published curves rather than observed directly, a step that introduces its own approximation. Extrapolation accuracy was judged against updated Kaplan-Meier curves, which themselves reflect finite follow-up rather than true long-term survival. The findings therefore speak to model selection practice in oncology health technology assessment rather than to other disease areas, and the authors’ statistical comparisons, while formally tested, describe performance across this specific collection of trials.

Even with those caveats, the implications are concrete. Analysts choosing survival extrapolations for cancer models should not treat a low AIC or BIC as a guarantee of reliable long-term projections, and should consider giving flexible parametric Royston-Parmar models a more prominent place in the candidate set.

Subject of Research: Parametric survival model selection and extrapolation accuracy in oncology trials

Article Title: Optimizing parametric survival model selection: an analysis using constructed individual patient data extracted from oncology trials from 2014 to 2023

Article References: Zhao, M., Shao, T., Shao, H., Li, Y., Fang, H., & Tang, W. (2026). Optimizing parametric survival model selection: an analysis using constructed individual patient data extracted from oncology trials from 2014 to 2023. BMC Medical Research Methodology. https://doi.org/10.1186/s12874-026-03051-4

Image Credits: AI Generated

DOI: 10.1186/s12874-026-03051-4

Keywords: parametric survival models, survival extrapolation, oncology trials, individual patient data reconstruction, Akaike Information Criterion, Bayesian Information Criterion, Royston-Parmar models, health technology assessment, Kaplan-Meier curves, flexible parametric models, Optimizing, parametric