Predicting Times to Event Based on Vine Copula Models

TL;DR

This paper introduces a vine copula-based method for predicting event times in right-censored data, offering flexibility and improved accuracy over traditional models when assumptions are violated.

Contribution

The paper presents a novel vine copula approach for time-to-event prediction that handles mixed data types and outperforms classical models under certain conditions.

Findings

Vine copula models approximate traditional methods well on standard datasets.

They outperform Cox and AFT models when their assumptions are violated.

The approach is flexible for various time-to-event data shapes.

Abstract

In statistics, time-to-event analysis methods traditionally focus on the estimation of hazards. In recent years, machine learning methods have been proposed to directly predict the event times. We propose a method based on vine copula models to make point and interval predictions for a right-censored response variable given mixed discrete-continuous explanatory variables. Extensive experiments on simulated and real datasets show that our proposed vine copula approach provides a decent approximation to other time-to-event analysis models including Cox proportional hazards and Accelerate Failure Time models. When the Cox proportional hazards or Accelerate Failure Time assumptions do not hold, predictions based on vine copulas can significantly outperform other models, depending on the shape of the conditional quantile functions. This shows the flexibility of our proposed vine copula approach for general time-to-event datasets.