Title: Comparing discrimination and calibration performance of two flexible link functions in discrete survival models

Authors: Susan Maposa; Alphonce Bere; Caston Sigauke; Charles Chimedza

Addresses: Department of Mathematical and Computational Sciences, University of Venda, Thohoyandou, 0950, South Africa ' Department of Mathematical and Computational Sciences, University of Venda, Thohoyandou, 0950, South Africa ' Department of Mathematical and Computational Sciences, University of Venda, Thohoyandou, 0950, South Africa ' School of Statistics and Actuarial Science, University of the Witwatersrand, Wits 2050, Johannesburg, South Africa

Abstract: This study provides the first direct comparison between the Pareto and Logit-power link functions within discrete survival models, evaluated alongside three commonly used links. We assess their discrimination and calibration using simulated and real-life datasets with varying skewness. Simulations included 100 datasets with symmetric, right-skewed, and left-skewed distributions, and bootstrapping was applied for robust evaluation. The results show that cloglog excels in discrimination, while logit offers superior calibration. The Pareto family demonstrates robust performance, making it a reliable secondary option. However, Logit-power performs poorly in calibration and is unsuitable for discrete survival models. The study offers practical recommendations for implementing the Logitp link, addressing its complex estimation process, and suggests a grid search approach using information criteria for parameter optimisation. These findings highlight the importance of carefully selecting link functions in discrete survival modelling.

Keywords: calibration; bootstrapping; discrimination; discrete survival models; link functions; Pareto family; logit-power family; simulation.

DOI: 10.1504/IJDATS.2025.150914

International Journal of Data Analysis Techniques and Strategies, 2025 Vol.17 No.4, pp.302 - 327

Received: 07 Apr 2024
Accepted: 18 Aug 2024
Published online: 05 Jan 2026 *

Full-text access for editors Full-text access for subscribers Purchase this article Comment on this article