BackgroundOptimizing cancer screening and surveillance frequency requires accurate information on parameters such as sojourn time and cancer risk from premalignant lesions. These parameters can be estimated using multistate cancer models applied to screening or surveillance data. However, the performance of these models has not been thoroughly investigated in settings in which cancer precursors are treated upon detection, preventing progression to cancer. Our main goal is understanding the performance of available multistate methods in this challenging censoring setting.MethodsWe assumed progression hazards between consecutive health states in a 3-state model (healthy HE, cancer precursor, and cancer) to be either time independent or dependent on time since state entry and compared 6 methods implemented in R software packages with varying assumptions: time-independent hazards (msm), hazards dependent on time since state entry (msm with a phase-type model, cthmm, smms, BayesTSM), and hazards dependent on time since the start of the process (hmm). Risk estimates from each method were compared in simulations and illustrated using colorectal cancer surveillance data from 734 individuals, classified into 3 health states: HE, non-advanced adenoma (nAA), and advanced neoplasia (AN).ResultsAll methods performed well with time-independent hazards in the simulation study. With hazards dependent on time since state entry, only smms and BayesTSM provided unbiased risk estimates. In the application, only msm,hmm, and BayesTSM yielded converged solutions. The nAA risk estimates were similar between hmm and BayesTSM but differed for msm, while AN risk estimates varied across methods.ConclusionsMethods for multistate cancer models, specifically with unobservable precursor-to-cancer transition, are strongly affected by the time dependency of the hazard. With time-dependent hazards since state entry, BayesTSM provided robust estimates, in both the simulation and application.HighlightsThis study presents the first comprehensive comparison of available multistate modeling options for screening and surveillance data, focusing on the specific setting of a 3-state progressive model (healthy, cancer precursor, cancer) in which cancer precursors are treated upon detection so that the transition to cancer is prevented (censoring after intervention). Sample R code and simulated data demonstrating the compared methods, along with documentation (including installation instructions, manual, and/or worked examples) for the corresponding R software packages, are available at https://github.com/EddymurphyAkwiwu/MultiStateMethods.All methods provide unbiased risk estimates for transition times when the true progression hazards are time independent. With more realistic models in which progression hazards are dependent on time since state entry, only BayesTSM and smms yield unbiased risk estimates for transition times.In situations with weakly identifiable likelihoods, the smms package may suffer from numerical and optimization problems. The BayesTSM package overcomes these problems by applying regularized parameter estimation using weakly informative priors.Methods for multistate cancer models, more specifically with unobservable precursor-to-cancer transition, are strongly affected by the time dependency of the hazard. An inappropriate choice can lead to biased parameter estimates.
Akwiwu et al. (Fri,) studied this question.