Prediction of long-term cumulative incidences based on short-term parametric model for competing risks: application...

Use of parametric statistical models can be a solution to reduce the follow-up period time required to estimate long-term survival. Mould and Boag were the first to use the lognormal model. Competing risks methodology seems more suitable when a particular event type is of interest than classical survival analysis. The objective was to evaluate the ability of the Jeong and Fine model to predict long-term cumulative incidence. Survival data recorded by Institut Curie (Paris) from 4761 breast cancer patients treated and followed between 1981 and 2013 were used. Long-term cumulative incidence rates predicted by the model using short-term follow-up data were compared to non-parametric estimation using complete follow-up data. 20- or 25-year cumulative incidence rates for loco-regional recurrence and distant metastasis predicted by the model using a maximum of 10 years of follow-up data had a maximum difference of around 6 % compared to non-parametric estimation. Prediction rates were underestimated for the third and composite event (contralateral or second cancer or death). Predictive ability of Jeong and Fine model on breast cancer data was generally good considering the short follow-up period time used for the estimation especially when a proportion of patient did not experience loco-regional recurrence or distant metastasis.