OBJECTIVES: The selection of cut-point(s) is often a point of contention for piecewise survival models. A method to statistically estimate the cut-point(s) has been developed using a piecewise exponential (PWE) model where a constant hazard is assumed within segments.¹ This case study aimed to compare the accuracy of long-term survival estimates of the PWE model with piecewise models that use Kaplan-Meier (KM) data adjoined to a parametric tail (PW-KM).

METHODS: Overall survival KM data from the KEYNOTE-010 31.0-month data cut-off (DCO) were digitized and fitted with the PWE model. The PWE model identified a cut-point at 13 months, and separate exponential models were fitted to each segment. For the PW-KM model, 5 months was selected as a cut-point through visual inspection of hazard plots. 13 months was also selected to facilitate comparison with the PWE model. From the cut-points onward, parametric tails were adjoined to the KM curves. Restricted mean survival time (RMST) at maximum follow-up from the KEYNOTE-010 67.4-month DCO was compared with the estimated RMST from each model at the same time point to evaluate accuracy.

RESULTS: RMST from the KEYNOTE-010 67.4-month DCO was 22.5 months. Estimated RMST from the PWE model was 21.6 months, and average estimated RMST for the 5- and 13-month PW-KM models was 22.3 and 22.4 months, respectively. The three most accurate models were the PW-KM 5-month cut-point generalized gamma, PW-KM 13-month cut-point Gompertz, and PW-KM 5-month cut-point log-logistic models. The PWE was ranked 7/13 models.

CONCLUSIONS: In this case study, the PWE model did not perform better than PW-KM models on average in estimating long-term survival. Although the cut-point was identified objectively using the PWE statistical model, constraining the model to use constant hazards may limit accuracy. Biological plausibility and hazard plots should be assessed to validate this assumption.