OBECTIVE: A multi-state model (MSM) of a standard three health state model requires modelling of three transitions: 1) progression-free survival (PFS) to progression (P), 2) PFS to death (D) and 3) P to D. Early censoring in PFS might be problematic, as the MSM needs to assign the post-PFS censoring time to a certain transition. The aim is to assess the impact of early censoring of PFS on the predicted PFS by an MSM.

METHODS: Hypothetical patient level data (PLD) were generated for an intervention arm (n=400) and a placebo arm (n=400). The PFS curve had 330 events and 258 events for the placebo and active arm respectively. In PFS 19 patients and 22 patients had an early censoring event in the placebo arm and active arm followed by a death event later onwards respectively. The MSM package in R was used.

RESULTS: The MSM assumes that patients do not progress between PFS censoring time and the OS event. It assigns the time of the OS event to both the 1) PFS to P (censor) and the 2) PFS to D (event) transition. Consequently, the PFS curve resulting from PFS to P and PFS to D in the MSM contains 349 events in the placebo arm and 280 events in the active arm. The MSM considers more events for estimated PFS than in the original KM PFS curve. Thus, the MSM underestimates PFS.

DISCUSSION: In case of early censoring of PFS but not OS, the MSM is likely to underestimate PFS compared to the KM, which impacts the modelled quality of life. In case early censoring is problematic, potentially models for the third transition that allow right and left censoring should be considered or endpoints less sensitive to censoring, such as time to treatment discontinuation.