|
Abstract:
|
Individualized risk prediction is an important goal in health research. Towards this, most established methods for risk prediction solely consider a univariate outcome, and thus are unsuited to the `semi-competing risks' setting in which both a non-terminal event and subsequent terminal event are of interest. While several so-called `illness-death' multistate models have been proposed for jointly modeling such outcomes, no frequentist methods exist for the important task of variable selection in such joint prediction models. To resolve this gap, we propose a penalized estimation framework for variable selection in a large class of illness-death models that: (i) permits flexible parametric and semi-parametric specification of baseline hazards; (ii) admits a random shared frailty; and, (iii) enables use of a broad class of penalties including (adaptive) Lasso, SCAD, and structured fusion Lasso. A unified MM algorithm for estimation is proposed, and performance is shown by asymptotic results and by simulation. An example is presented of developing a joint risk prediction model for preeclampsia, a non-terminal pregnancy complication for which delivery of the baby is the terminal event.
|
