The lay abstract featured today (for Structured learning in time-dependent Cox models by Guanbo Wang, Yi Lian, Archer Y. Yang, Robert W. Platt, Rui Wang, Sylvie Perreault, Marc Dorais, Mireille E. Schnitzer) is from Statistics in Medicine with the full article now available to read here.

Wang G, Lian Y, Yang AY, et al. Structured learning in time-dependent Cox models. Statistics in Medicine. 2024; 1–20. doi: 10.1002/sim.10116

Abstract

Cox models, which are essential for analyzing time-to-event data, often need to manage variables and coefficients that change over time. These models are particularly useful when dealing with a large number of variables, and selecting the most relevant ones is crucial. However, traditional methods lack flexibility and struggle to incorporate the researcher’s prior knowledge about variable relationships. To address this, we have introduced an innovative approach that allows for the integration of complex selection rules in the variable selection process. For instance, if an interaction effect is selected, our method can ensure that the main effects are also chosen. Our approach effectively handles various structural relationships among variables, such as temporal, spatial, and even intricate patterns like those found in trees or graphs. It is highly accurate, minimizing the risk of falsely identifying irrelevant variables as significant. We have implemented this new method in a software package called ‘sox.’ This package uses an advanced algorithm to efficiently manage these complex scenarios. ‘sox’ is user-friendly and fast, making it suitable for practical, real-world applications. We demonstrate its effectiveness through several examples, including a study on heart disease patients that identifies key predictors of mortality. This tool allows researchers to tailor their analyses to specific needs, integrating their existing knowledge and significantly improving both the precision and applicability of their findings.