The new potential for Long-Term Survival (LTS) in glioblastoma (GBM) demonstrated by the EF-14 trial, which showed significant survival benefit from adding Tumor-Treating Fields (TTFields) to maintenance temozolomide (TMZ) in newly diagnosed GBM (Stupp 2017), has made predictive modeling of LTS highly relevant for clinical and policy decision making. Reliable models must account for the evolution in prognosis over time in GBM. Previous studies (Bernard-Arnoux 2016, Connock 2019) have parametrically modeled the survival impact of TTFields with EF-14 data, but their results suffer from poor in-sample fit, clinical inconsistency, and poor fit to long-term epidemiological data. Here we evaluate LTS predictions from a novel, Responder Analysis-based method of parametric modeling with EF-14 data. Patients were classified as responders (R) or non-responders (NR) according to their Progression-Free Survival (PFS). If a patient’s PFS met or exceeded a given threshold time (TT), then they were classified as R. Otherwise they were classified as NR. Differences between R’s and NR’s Kaplan-Meier curves and resultant significance levels were studied for a range of TT’s. These significance levels are the result of the impact of longer PFS on OS and the opposing effect of smaller R-subgroup sample size with larger TT's. In the TTFields+TMZ arm, 10 months was the most significant TT, while in the TMZ-only arm, the most significant TT was 6.5 months. TT’s with high significance in both arms were selected for further study, including 10, 12, 18, and 20 months. Parametric survival distributions were separately calibrated for each R and NR subgroup in each trial arm using Maximum Likelihood Estimation. We studied results from exponential, Weibull, log-normal, log-logistic, Gamma, Gompertz, and Generalized Gamma and F distributions. The long-term conditional survival probabilities from each model were then calculated or estimated with Monte Carlo simulation and compared to epidemiological results from Porter et al. (2011) and from separate SEER database searches. For every TT and parametrization studied, the models for the R-subgroups clearly underestimated conditional LTS probabilities compared to epidemiological results beyond 5 years. Errors in these conditional probability estimates ranged from approximately -20% to -50% with models for TTFields+TMZ responders, while models for TMZ-only responders performed even worse, outputting values near 0% for conditional survival probabilities. This analysis provides further evidence that parametric models trained exclusively on relatively short-term EF-14 trial data, even with responder analysis-based enhancements, suffer from biases that limit their usefulness for LTS modeling in GBM. Long-term epidemiology data and modeling approaches that incorporate this data (Guzauskas 2018) are therefore necessary for informed decision making regarding novel treatments in glioblastoma.