Comparison Of Standard Parametric Survival Methods Versus More Flexible Approaches

OBJECTIVES: Several standard parametric methods for extrapolating overall survival (OS) exist. However, more flexible methods such as spline-based models and the Generalized Gamma or Generalized F models are less often applied, despite being recommended by the NICE Decision Support Unit (UK). The objective of this study is to compare standard with these more flexible models in simulated datasets. METHODS: Six datasets with active and comparator arms were simulated. The comparators arms were based on three published datasets in which the baseline hazard over time was (1) decreasing, (2) increasing, and (3) fluctuating, respectively. The corresponding active arms were simulated with (1) a constant hazard ratio (HR) and (2) improving HR over time.The following models were tested; standard parametric models (Weibull, Exponential, Lognormal, Loglogistic, Gamma and Gompertz), spline models with one knot (Weibull, Lognormal and Loglogistic), Generalized Gamma models, and Generalized F models. The tested models were fitted (1) with treatment as constant covariate, (2) with treatment as time varying covariate, and (3) as two individual curves over the separate arms. The models were compared based on visual fit of the Kaplan Meier curve, log cumulative hazard, Akaike's information criterion (AIC) and the Bayesian information criterion (BIG). RESULTS: For the decreasing hazards dataset, the Lognormal and spline models had the lowest AIC/BIC for constant and improving HRs, respectively. In case of increasing hazards, the Generalized Gamma and Gompertz had the lowest AIC/BIC for constant and improving HRs, respectively. Finally, for the fluctuating hazards dataset, the Generalized Gamma and spline models had the lowest AIC/BIC for constant and improving HRs, respectively. Visual fits confirmed these results. CONCLUSIONS: Flexible models had a better fit compared to standard parametric models in four out of six datasets. Thus, we recommend the use of these models as key alternative among standard options for extrapolating OS from clinical trials.