Extracting and Modeling Survival Data

We used GetData Graph Digitizer (v2.26, http://getdata.sourceforge.net/download.html) to extract survival data from published PFS and OS Kaplan-Meier curves. To reconstruct individual patient data, we used the Guyot's method, which is the most accurate data reproduction method currently known for cases where individual patient data are not available (35 , 36 (link)). Log cumulative hazards and schoenfeld residual test plots (Supplementary material 2) showed proportional hazard (PH) or piecewise models were not suitable in this analysis. In accordance with the shapes of the survival curves, the non-PH NMA models considered in this study were first- and second-order fractional polynomial (FP) models (37 (link)). We fitted first- and second-order FP models with power parameters −2, −1, −0.5, 0, 0.5, 1, 2, and 3, with three parallel Markov chains consisting of 10,000 samples after a 10,000 samples burn-in. To reconstruct and extrapolate the PFS curve of the standard chemotherapy, and the OS and PFS curves of the second-line docetaxel, we considered parametric functions including Exponential, Weibull, Gompertz, Gamma, Log-logistic, Log-normal, Generalized Gamma, GenF, FP, Restricted Cubic Spline, and Royston and Parmar (RP) models. Goodness-of-fit was evaluated by visual inspection of survival curves, Akaike information criterion (AIC) and deviance information criterion (DIC). Lower AIC and DIC combined with reasonable visual effects indicated a better performance of the selected model (38 ). Survival modeling was conducted in R (v4.1.2) and Winbugs (v1.4.3) (39 , 40 ). R codes for relative methods can be found on Github (https://github.com/TaihangShao/NMA_methodology).