(186) Flexible parametric survival models for simultaneous estimation on multiple timescales: an application using nationwide Swedish register data

Background: In many survival analysis applications, time-to-event outcomes can be modelled on different timescales, such as time since diagnosis or attained age. However, models estimating effects simultaneously across multiple timescales are rare, and their implementation in standard software can be computationally intensive.

Objectives: The main objective of this study was to demonstrate how flexible parametric survival models can easily and efficiently model thrombosis rates on multiple timescales simultaneously in patients with myeloproliferative neoplasms (MPN). A secondary aim was to assess bias in different scenarios where simultaneous modelling on multiple timescales is omitted.

Methods: Swedish MPN patients diagnosed during 1987-2009 were selected from national registers and matched with population controls. Using flexible parametric models, thrombosis rates and hazard ratios (MPN vs non-MPN) with 95% confidence intervals were estimated over two timescales, time since diagnosis and attained age, simultaneously. Furthermore, we performed a simulation study to compare models using one timescale and adjusting for variables on the second timescale to models with two timescales, to assess the magnitude of bias in different scenarios.

Results: Thrombosis rates were significantly higher for MPN patients of all ages, particularly in the first two years after diagnosis, after which the rates remained higher than for the comparators of the same age, and increased with attained age. However, the duration of the disease did not significantly impact thrombosis rates after the first couple of years following diagnosis. Simulations demonstrated that the omission of modelling simultaneously on two timescales results in large bias if the covariate of interest has an interaction with one or both timescales, i.e. non-proportional hazards. Otherwise, models with one timescale performed well.

Conclusions: Flexible parametric survival models can be used to fit multiple timescales simultaneously, using standard software. Using Swedish register data, these models helped to better understand the relationship between age, time since diagnosis, and thrombosis rates, guiding clinicians for optimized management. Based on the simulation results, these models are crucial in situations where non-proportional hazards are present.