OBJECTIVES:

This study investigated how the cured proportion and data maturity impact MCM performance.

METHODS:

A simulation plan was developed to assess four scenarios composed of 1,000 simulations, varying the proportion of cure (25% or 50%) and length of follow-up (13 or 33 months). Survival of uncured patients was simulated according to a Weibull distribution, and survival of cured patients was simulated according to the Guyot’s algorithm based on US life tables. Time to censoring was drawn from an exponential distribution. Standard parametric models and MCMs were then estimated on the obtained datasets. Performance was assessed based on mean survival, estimated cure rate and the squared error terms. In addition, MCMs misestimating the proportion of cured by at least 20% were identified as “implausible models”.

RESULTS:

In the simulated data, the average mean survival was of 77 months. Parametric models always underestimated long-term survival. The MCM using the Weibull function performed best and estimated the mean survival at 70.9 months and at 74.3 months with short and longer follow-up respectively. On average, the mean cure rate was estimated at 26% and 27% for the short and long-term follow-up respectively in the scenario including 25% of cured patients. These proportions were estimated at 50% and 53% for the scenario including 50% of cured. The risk of selecting implausible models (in our case the log-normal and log-logistic distributions) for the MCM was higher with short term follow-ups (15% for the 25% cured scenario and 26% for the 50%) than with longer-term data (2% and 1% respectively).

CONCLUSIONS:

On average, the performance of MCM was not influenced by the proportion of cured nor the length of follow-up. However, the proportion of models associated with “implausible” estimates in terms of proportion of cured and mean survival was higher for data with a shorter follow-up.