Capturing heterogeneous time-variation in covariate effects in non-proportional hazard regression models
Niklas Hagemann, Thomas Kneib, Kathrin Möllenhoff · 2025
Read Capturing heterogeneous time-variation in covariate effects in non-proportional hazard regression models by Niklas Hagemann, Thomas Kneib, Kathrin Möllenhoff (2025) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2501-13525.txt.
Introduction
Survival analysis asks how factors such as age or treatment change how long patients live. Niklas Hagemann, Thomas Kneib, and Kathrin Möllenhoff note that a factor's effect can change over time and differ between groups. The way it changes over time can itself differ between groups. The standard Cox model allows none of this, and no single framework has covered all three features at once. The authors build one by adding flexible effect curves to a piecewise exponential mixed model, so each group's effect follows its own smooth path over time.
The curves are fitted with penalized splines, which keep them from bending too wildly when the data show no real pattern. The setup also mostly removes the need to choose time intervals by hand. In simulations the method beat competing approaches. A brain tumor case study shows how it works in practice.
Catalog record
| CREATOR | Niklas Hagemann, Thomas Kneib, Kathrin Möllenhoff |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2501-13525
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
Capturing heterogeneous time-variation in covariate effects in non-proportional hazard regression models Niklas Hagemann, Thomas Kneib, Kathrin Möllenhoff Abstract A central focus in survival analysis is examining how covariates influence survival time. These covariate effects are often found to be either time-varying, heterogeneous – such as being specific to patients, treatments, or subgroups – or exhibit both characteristics simultaneously. While the standard model, the Cox proportional hazards model, allows neither time-varying nor heterogeneous effects, several extensions to the Cox model as well as alternative modeling frameworks have been introduced. However, no unified framework for incorporating heterogeneously time-varying effects of covariates has been proposed. Such effects occur when a covariate influences survival not only in a heterogeneous and time-varying manner, but when the time-variation is also heterogeneous. We propose to model such effects by introducing heterogeneously time-varying coefficients to piecewise exponential additive mixed models. We deploy functional random effects, also known as factor smooths, to model such coefficients as the interaction effect of heterogeneity and time-variation. Our approach allows for non-linear time-effects due to being based on penalized splines and uses an efficient random effects basis to model the heterogeneity. Using a penalized basis prevents overfitting in case of absence of such effects. In addition, the penalization mostly solves the problem of choosing the number of intervals which is usually present in unregularized piecewise exponential approaches. We demonstrate the superiority of our approach in comparison to competitors by means of a simulation study. Finally, the practical application and relevance are outlined by presenting a brain tumor case study. 1 Introduction One of the major topics in survival analysis is analyzing the effect covariates have on the survival time. Frequently, the effects of these covariates can be observed to be either time-varying, heterogeneous, i.e. patient-, treatment- or subgroup-specific, or even both. If the goal is to analyze the effects of covariates on the survival time, hazard regression models play a critical role. They estimate the hazard function, which represents the instantaneous rate of occurrence of the event at a given time, conditional on survival up to that time. By incorporating covariates, hazard regression models allow to assess the impact of various factors on the hazard rate and, hence, to identify significant effects. The widely used standard model is the Cox proportional hazards model ( Cox,, 1972 ) where the hazard rate of an observation i\in\{1,...,n\} with corresponding covariate vector \bm{x}_{i} is given by \lambda_{i}(t)=\lambda_{0}(t)\exp(\bm{x}_{i}^{\top}\bm{\beta}), where \bm{\beta} is the vector of regression coefficients. The assumption of proportionality of the hazards results from the model being strictly split into the time-dependent baseline hazard \lambda_{0}(t) and the time-constant covariate effects \exp(x_{i}^{\top}\beta) . In addition, the Cox model assumes the covariate effects to be (exp-transformed) linear effects. These strict assumptions are often not fulfilled in practice ( Li et al.,, 2015 ; Jachno et al.,, 2019 ) . Frequently, this is caused by non-proportional hazards, i.e. the covariates or their effects are time-dependent. In addition, in many cases the effects might not be linear but of a more complex form. In particular, heterogeneity in terms of treatment-specific, subgroup-specific (e.g. gender-specific) or individual effects might be present which cannot be captured properly by linear effects. Therefore, several flexible extensions to the Cox model have been introduced. On the one hand, such extensions relax the proportional hazards assumption: The most commonly used may be the stratified Cox model in which for each level of a categorical variable a separate baseline hazard is fitted. Alternatively, several studies ( Zucker and Karr,, 1990 ; Murphy and Sen,, 1991 , see, e.g.,) propose to include time-varying coefficients in order to capture time-dependent covariate effects. In addition, Andersen and Gill, (1982) introduced an approach to include time-dependent covariates. On the other hand, extensions of the Cox model have been introduced to allow for more complex effects: Gray, (1992) added non-linear smooth spline-based covariate effects and Hess, (1994) uses such effects to express covariate effects as a function of time. Regarding the heterogeneity, again the stratified Cox model can be mentioned, where the baseline hazard can be group-specific. However, a more natural way to account for heterogeneous effects, which also allows for more general types of heterogeneity, is to introduce random effects leading to frailty models ( Vaupel et al.,, 1979 ; Ripatti and Palmgren,, 2000 ; Therneau et al.,, 2003 ) . In contrast to adding specific extensions to the Cox model, several recent studies have aimed to introduce a new flexible hazard regression framework: Kneib and Fahrmeir, (2007) introduce Cox-type structured hazard regression models \lambda_{i}(t|\bm{x}_{i})=\lambda_{0}(t)\exp\left(\sum_{k=1}^{K}f_{k}(\bm{x}_{i},t)\right)=\exp\left(\tilde{\lambda}_{0}(t)+\sum_{k=1}^{K}f_{k}(\bm{x}_{i},t)\right), (1) where \tilde{\lambda}_{0}(t) is the log-baseline hazard and f_{k} can resemble different types of effects, e.g. linear effects, smooth (spline-based) effects, time-varying effects or random effects/frailty. The inclusion of time-varying effects allows for explicit modeling of non-proportional hazards. This approach was further investigated by Hofner et al., (2011) ; Hofner et al., (2013) . The corresponding inference is conducted by mixed model-based penalized likelihood estimation. However, the log-likelihood involves an integral over the hazard rate. Hence, the estimation relies on numerical integration which is computationally costly and can be subject to impreciseness. Alternatively, Hennerfeind et al., (2006) proposed a Bayesian estimation scheme for such models. However, this approach is based on the same log-likelihood and, therefore, shares these disadvantages. Another approach to introduce models of the form ( 1 ) is given by piecewise exponential additive mixed models ( Bender et al.,, 2018 ; Bender and Scheipl,, 2018 , PAMM;) which generalizes the concept of piecewise exponential models ( Friedman,, 1982 , PEM;) from linear to additive predictor terms. The underlying idea is to divide the time axis into a finite number of intervals and assume the hazard rate to be piecewise constant within these intervals. While manually choosing the interval cut-off points is a challenging task and a frequent source of criticism for PEMs, PAMMs avoid the arbitrary choice of cut-off points by using a penalized approach providing a sufficiently good fit while preventing overfitting. Under the assumption of piecewise constant hazard rates, restructuring the data leads to the likelihood of the survival model being proportional to the one of a Poisson regression model (see Section 2 of Bender et al., (2018) for details). Hence, both models are equivalent with respect to their maximum likelihood estimators and the model parameters are estimated based on the Poisson model. Therefore, estimation can make use of existing methods and implementations for generalized linear models (GLMs) in the case of PEMs and generalized additive (mixed) models (GAMs/GAMMs) in the case of PAMMs. While these approaches propose several flexible effects such as non-linear, time-varying and random effects, none of them include heterogeneously time-varying effects of covariates. Such effects occur if a covariate influences the survival time not only in a heterogeneous and time-varying manner, but the time-varying effect is heterogeneous, too. A typical example would be that the effect of a covariate is treatment specific, time-varying and that its time-variation is also treatment-specific, e.g., decreasing for an intervention but increasing for a placebo. To the best of our knowledge, this study is the first to propose heterogeneously time-varying covariate effects in hazard regression models. Based on the framework of PAMMs, we introduce these heterogeneously time-varying coefficients as f_{kg}(t)\cdot x_{ik}, where g\in\{1,...,G\}