calibmsm: An R package for calibration plots of the transition probabilities in a multistate model

Background and objective: Multistate models, which allow the prediction of complex multistate survival processes such as multimorbidity, or recovery, relapse and death following treatment for cancer, are being used for clinical prediction. It is paramount to evaluate the calibration (as well as other metrics) of a risk prediction model before implementation of the model. While there are a number of software applications available for developing multistate models, currently no software exists to aid in assessing the calibration of a multistate model, and as a result evaluation of model performance is uncommon. calibmsm has been developed to fill this gap. Methods: Assessing the calibration of predicted transition probabilities between any two states is made possible through three approaches. The first two utilise calibration techniques for binary and multinomial logistic regression models in combination with inverse probability of censoring weights, whereas the third utilises pseudo-values. All methods are implemented in conjunction with landmarking to allow calibration assessment of predictions made at any time beyond the start of follow up. This study focuses on calibration curves, but the methodological framework also allows estimation of calibration slopes and intercepts. Results: This article serves as a guide on how to use calibmsm to assess the calibration of any multistate model, via a comprehensive example evaluating a model developed to predict recovery, adverse events, relapse and survival in patients with blood cancer after a transplantation. The calibration plots indicate that predictions of relapse made at the time of transplant are poorly calibrated, however predictions of death are well calibrated. The calibration of all predictions made at 100 days post transplant appear to be poor, although a larger validation sample is required to make stronger conclusions. Conclusions: calibmsm is an R package which allows users to assess the calibration of predicted transition probabilities from a multistate model. Evaluation of model performance is a key step in the pathway to model implementation, yet evaluation of the performance of predictions from multistate models is not common. We hope availability of software will help model developers evaluate the calibration of models being developed.