Background: Multistate modelling (MSM) is an alternative to partitioned survival modelling. For standard oncology models, MSMs separate individual patient event times into three transitions: (1) time to progression (TTP), (2) TTP to death, and (3) time to death after progression. Williams et al. suggests that MSMs can be created from aggregated data, though assumptions are required when creating individual patient-level data (IPD) with PFS and corresponding OS time per patient. The aim was to assess the challenges of creating an MSM using aggregated data.

Methods: Based on simulated and subsequently aggregated data, OS and PFS Kaplan-Meier curves were used to obtain IPD for OS and PFS using a validated algorithm by Guyot et al. An assumption that the shortest PFS time corresponds to the shortest OS time was used to create IPD with PFS and OS time per patient. Parametric distributions were fitted and combined to estimate PFS and OS.

Results: Due to the assumption of linking shortest PFS to shortest OS, the MSM predicted that all PFS events are progression events, which implied that there are no death events before progression. Also, an artificial correlation was introduced between timing of progression and risk of death. Nevertheless, the MSM predicted PFS accurately, but overestimated OS for the initial months for both treatments, and fitted OS relatively well thereafter.

Conclusions: The challenge of using MSMs with aggregated data is that the time each patient contributes to each transition relies on assumptions which affect the plausibility of extrapolations. OS is likely overestimated due to the artificial correlation not accounted for when estimating the time to death after progression predictions. If the number of PFS deaths are reported, this should be considered when linking OS and PFS time per patient. More research and guidance are needed on estimating MSMs based on aggregated data.