Multiple time scales in multi-state models

In multi‐state models, it has been the tradition to model all transition intensities on one time scale, usually the time since entry into the study (‘clock‐forward’ approach). The effect of time since an intermediate event has been accommodated either by changing the time scale to time since entry t...

Full description

Saved in:
Bibliographic Details
Published in:Statistics in medicine 2013-12, Vol.32 (30), p.5315-5327
Main Authors: Iacobelli, Simona, Carstensen, Bendix
Format: Article
Language:English
Subjects:
Data collection
Failure
flexible parametric models
Humans
Leukemia
Leukemia, Myelogenous, Chronic, BCR-ABL Positive - therapy
Markov analysis
Markov Chains
Mesenchymal Stem Cell Transplantation - standards
Models, Statistical
multi-state
Poisson distribution
Poisson model
Recurrence
Time
Time Factors
time scales
time-dependent covariate
Transplants & implants
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:In multi‐state models, it has been the tradition to model all transition intensities on one time scale, usually the time since entry into the study (‘clock‐forward’ approach). The effect of time since an intermediate event has been accommodated either by changing the time scale to time since entry to the new state (‘clock‐back’ approach) or by including the time at entry to the new state as a covariate. In this paper, we argue that the choice of time scale for the various transitions in a multi‐state model should be dealt with as an empirical question, as also the question of whether a single time scale is sufficient. We illustrate that these questions are best addressed by using parametric models for the transition rates, as opposed to the traditional Cox‐model‐based approaches. Specific advantages are that dependence of failure rates on multiple time scales can be made explicit and described in informative graphical displays.Using a single common time scale for all transitions greatly facilitates computations of probabilities of being in a particular state at a given time, because the machinery from the theory of Markov chains can be applied. However, a realistic model for transition rates is preferable, especially when the focus is not on prediction of final outcomes from start but on the analysis of instantaneous risk or on dynamic prediction.We illustrate the various approaches using a data set from stem cell transplant in leukemia and provide supplementary online material in R. Copyright © 2013 John Wiley & Sons, Ltd.
ISSN:0277-6715
1097-0258
DOI:10.1002/sim.5976