Bayesian variable and hazard structure selection in the General Hazard model
summary
The gist
The authors introduce a Bayesian variable selection approach within a general hazard (GH) model framework…
In short
The authors developed a Bayesian variable selection method for General Hazard (GH) models to simultaneously select relevant covariates and determine the underlying hazard structure (like PH or AFT). They introduced two specific priors for regression coefficients and a hierarchical prior on the model space to manage complexity. The work shows these methods are consistent, effectively identifying important variables and suggesting that the AFT model is often most probable.
Key concepts
- General Hazard (GH) Model
- This is a flexible survival model structure that allows for different ways risk changes over time. It combines a baseline hazard function with effects from covariates ($oldsymbol{eta}$), enabling it to represent various failure patterns, including proportional hazards and accelerated failure time.
- g-priors (Likelihood-Curvature-Matching and Product)
- These are specific prior distributions used for the regression coefficients. The authors propose two types: one that matches the likelihood curvature and another product prior. These priors are designed to work well with the GH model's two-component structure, ensuring that when they select a model, it is consistent with the data.
- Model Space Prior
- This is a prior placed on the set of possible models themselves. It goes beyond simple models by allowing covariates to potentially play multiple roles—such as affecting both the hazard rate and the time scale. This helps penalize overly complex models that might arise from this flexibility.
- Extended Add–Delete–Swap Algorithm
- This is a computational tool used within an MCMC strategy. It allows the sampling process to efficiently move between different models within the GH framework by jointly updating both which variables are included and what the specific hazard structure is, improving exploration of the parameter space.
Terminology used across episodes
This episode discusses
The paper
Bayesian variable and hazard structure selection in the General Hazard model · Read on arXiv
University College London
Transcript
Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.
Tom: We're looking at the paper: "Bayesian variable and hazard structure selection in the General Hazard model".
Jane: Here's the gist of the paper. The authors introduce a Bayesian variable selection approach within a general hazard (GH) model framework, designed to simultaneously select relevant covariates and determine the underlying hazard structure, which can encompass proportional hazards (PH), accelerated failure time (AFT), and other structures as special cases.
Tom: To manage complexity and account for multiplicity, they propose two types of g-priors for regression coefficients—a likelihood-curvature-matching prior and a product g-prior—which are shown to lead to consistent model selection under suitable regularity conditions.
Jane: Furthermore, a hierarchical prior on the model space is introduced to penalize model complexity arising from multiplicity.
Tom: The GH model is defined by the hazard function h(t alpha, beta, xi) = h 0(t xi) (n x tilde o xi (n x tilde o), where h 0(times xi) is a parametric baseline hazard function.
Title and authors: Tom: So, we're talking about this paper today, "Bayesian variable and hazard structure selection in the General Hazard model." It looks like it tackles a big problem: how to pick both which variables matter and what kind of hazard structure the data actually follows.
Jane: That's right, Tom. It takes models like proportional hazards or accelerated failure time and expands them into this General Hazard model, which is more flexible because it lets variables affect the hazard rate, the time scale, or both at once.
Lu: The authors are trying to solve that simultaneously by using a Bayesian variable selection approach that can determine both the covariates and the underlying hazard structure.
Meng: From an engineering standpoint, what I’m hearing is they’re moving away from assuming you pick one structure first and then selecting variables within it, which is where a lot of complexity comes from.
Lalam: The paper introduces two specific prior setups for the regression coefficients, the likelihood-curvature-matching prior and a product g-prior.
Tom: Right, so those priors are designed to match the two different components of this General Hazard model structure and they show that both ways lead to consistent model selection under certain conditions.
Jane: So, it sounds like they've made a specific choice there because they need something that handles the flexibility of this GH model properly.
Lu: And on top of that, they introduce a hierarchical prior over the entire model space to handle the issue of multiplicity, which is when you have many variables competing for attention.
Meng: Multiplicity is huge in production; if you're trying to select features and structures at the same time, you can get way too much noise unless you have a good way to penalize complexity.
Lalam: They also developed some computational tools, specifically an extended Add-Delete-Swap algorithm for joint sampling of variable inclusion indicators and hazard structures within a custom Markov chain Monte Carlo strategy.
Tom: That’s the mechanics part, right? How they actually get these complex models to sample correctly when you have both variables and structures floating around.
Jane: And to make those samples more useful, they implemented fast approximations for the marginal likelihood using Laplace and integrated Laplace approximations, which is pretty key for figuring out posterior model probabilities.
Title and authors: Lu: The simulation studies suggest that the performance of variable selection is stable across different scenarios in terms of sensitivity and specificity when you use these methods.
Tom: The simulations also showed that the likelihood-curvature-matching prior has a tendency toward greater sparsity when the true model is actually a General Hazard model.
Meng: Real data testing on datasets like FLC confirmed that this methodology can correctly identify important variables and what their roles are, which is good news for practical application.
Jane: And they even tested it on the NKI70 dataset, which was pretty challenging because it had a moderately large number of predictors and small sample sizes with high censoring.
Lu: In that challenging scenario, both methods agreed that the accelerated failure time model was the most probable hazard structure they found.
Tom: That’s interesting because it suggests that even though this is a General Hazard model, sometimes the data points strongly toward a simpler structure like AFT being correct.
Jane: And looking at how these methods perform asymptotically, the analysis of Bayes factor rates shows that for both priors, in overfitted models, the rates only depend on how many spurious variables are at each level and their hyperparameters.
Lu: For non-overfitted models, those rates depend on terms involving censoring and potential model misspecification through expected log-likelihood.
Tom: So for a clean model, the probability of being correct approaches one in the limit using these priors, which is a strong theoretical result for consistency.
Jane: Corollary one and Corollary two indicate that asymptotically, the proposed methods recover the model closest to the true data-generating mechanism when measured by a generalized Kullback-Leibler divergence that accounts for censoring.
Meng: But they also admit that in finite samples, performance can still be affected by sample size, censoring rate, and follow-up length.
Lu: And one limitation they flag is that misspecification of the parametric General Hazard model itself—like getting the baseline hazard shape wrong—can affect both structure selection and the power to detect small effects in finite samples.
Tom: So for a listener just looking at what this means for survival analysis, it's that these methods are robust, but you still have to be careful about how well you model the baseline hazard.
Title and authors: Jane: And they conclude that both proposed methods point toward the AFT model being most probable in their test cases, which suggests a proportional hazards model might be misspecified in those situations.
Lu: So while they propose this flexible framework for variable and structure selection, the practical implication is that you can now formally test not just if a structure fits, but also whether variables act on the hazard or the time scale.
Meng: For someone deploying this in production, it means you have a more rigorous way to determine which features are useful and what kind of effect they have before you commit to a model structure.
Tom: It’s about moving from just picking one structure and then adding variables on top to having a joint selection process that handles both at once.
Jane: So, in wrapping up this discussion on the paper "Bayesian variable and hazard structure selection in the General Hazard model," we see a system that allows us to look at variables and hazard structures together.
Lu: The core of their work is introducing those two g-prior formulations to ensure consistent selection under suitable regularity conditions, plus the hierarchical prior for model space complexity.
Meng: For me, the computational tools like the extended Add-Delete-Swap algorithm are important because they actually make this kind of joint sampling feasible rather than just theoretical.
Tom: And those fast Laplace approximations help keep the MCMC moves efficient so we don't spend forever just calculating model probabilities during sampling.
Jane: The overall implication is that we can now formally quantify our uncertainty about both the variables and the structure simultaneously, which is a step up from what was previously possible with single-structure methods.
Lu: And asymptotically, they recover the best model because of those generalized Kullback-Leibler divergence terms mentioned in Corollary one and two even accounting for censoring.
Meng: However, they are honest about the finite sample issues—the results can still be sensitive to things like the censoring rate or how long you follow up.
Tom: So for the listener, this paper gives us a more principled way to select both inputs and outputs of our survival models within that flexible General Hazard framework.
Jane: That’s it for today on "Bayesian variable and hazard structure selection in the General Hazard model." Next time we'll look at some work on sequence-derived feature distributions from deep sequencing data.
The paper's summary: Tom: So, we're looking at this paper about Bayesian variable and hazard structure selection in the General Hazard model today.
Jane: To start off, they've put together a framework that lets you pick both which variables are important and what kind of hazard structure the data actually follows simultaneously.
Lu: Exactly. It moves past picking one thing first, like just choosing proportional hazards or just picking variables for a simple model, and lets you figure out both at the same time within this General Hazard model.
Meng: I mean, so instead of saying "this variable affects the hazard rate," it lets you test if it acts on the hazard rate or maybe on how fast time moves.
Tom: Right. And they use these two specific priors for the regression coefficients—one is a likelihood-curvature-matching prior, and the other is a product g-prior.
Jane: They argue that these two setups are designed to match up with the two parts of this General Hazard model structure, and they claim that both lead to consistent model selection if you have decent regularity conditions.
Lu: And they don't stop there; they introduce a hierarchical prior over the entire model space itself.
Meng: So it’s like they’re penalizing complexity when multiple variables might be playing different roles or affecting different aspects of time and hazard.
Tom: And computationally, they built this extended Add-Delete-Swap algorithm for joint sampling. This lets them sample variable inclusion indicators and hazard structures together within a custom Markov chain Monte Carlo strategy.
Jane: That’s the engine part, right? They have these fast approximations for the marginal likelihood using Laplace and integrated Laplace approximations to make those MCMC moves efficient.
Lu: The simulation results suggest that variable selection performance is pretty stable across different scenarios in terms of sensitivity and specificity when you use these methods.
Meng: I saw some real data tests on the FLC dataset, and they confirmed that this method can actually find important variables and correctly identify their roles.
Tom: They even tested it on the NKI70 dataset, which was pretty tough because it had a lot of predictors but a small sample size with high censoring.
Jane: Even there, both methods agreed that the accelerated failure time model was the most probable hazard structure they found.
Lu: That’s an interesting finding because it suggests that even in this flexible General Hazard setup, sometimes the data strongly points toward a simpler structure like Accelerated Failure Time being correct.
Tom: Looking at the math on Bayes factor rates, they show that for both priors, in overfitted models, those rates only depend on how many spurious variables are at each level and their hyperparameters.
Jane: For non-overfitted models, those rates depend on terms involving censoring and potential model misspecification through expected log likelihood.
Lu: Corollary one and corollary two show that asymptotically, the proposed methods recover the model closest to the true data-generating mechanism when measured by a generalized Kullback-Leibler divergence that accounts for censoring.
Meng: But they also flag that in finite samples, performance can still be affected by sample size, censoring rate, and follow-up length.
Tom: And one thing they warn about is that if you misspecify the parametric General Hazard model itself—like getting the baseline hazard shape wrong—that can affect both structure selection and the power to find small effects in finite samples.
Jane: So while they conclude that both proposed methods point toward the Accelerated Failure Time model being most probable, it suggests a proportional hazards model might be misspecified in those specific situations.
Lu: The overall implication is that you can now formally quantify your uncertainty about both the variables and the structure at the same time, which is a step up from what was possible with single-structure methods.
Tom: It’s about moving from just picking one structure and then adding variables on top to having a joint selection process that handles both at once.
Jane: So while they give us these tools for simultaneous selection, they also remind us that the quality of the underlying hazard function model is still critical for getting those numbers right in practice.
Lu: The next thing we're looking at is how this flexible framework compares to other ways of modeling sequence data features, like those deep sequencing distributions.
The paper's improvements: Tom: So, we're talking about how they improved their approach to variable and hazard structure selection in this General Hazard model today.
Jane: Basically, they’re suggesting two main ways to get better results: one is using those two specific g-priors we talked about earlier—the likelihood-curvature-matching and the product g-prior.
Lu: They argue that these priors are specifically tailored to match the General Hazard model's two component structure, which is what makes them consistent for model selection under certain conditions.
Meng: So if you’re building a system, it sounds like you can’t just pick any random prior; it has to be designed with this specific knowledge of the model’s components in mind.
Tom: And they also introduced a hierarchical prior over the entire model space to handle that multiplicity issue we mentioned before.
Jane: That means instead of just picking one variable or one structure, you penalize yourself if your model gets too complex because multiple variables are competing for attention at different levels.
Lu: That moves the process from just selecting inputs to selecting a whole hierarchy of possibilities, which should make the selection process more robust against noise.
Meng: From an engineering standpoint, that complexity sounds good on paper but how much computational overhead does this hierarchical prior add when we’re running models on real traffic data?
Tom: Well, they did develop computational tools for this joint sampling—that extended Add-Delete-Swap algorithm—which is supposed to make exploring this space feasible instead of just theoretical.
Jane: And they also refined how we estimate the model probabilities by using those fast Laplace and integrated Laplace approximations, which makes those MCMC moves much faster.
Lu: These improvements aim to provide a more stable way to determine which variables are relevant and whether they act on the hazard or the time scale, even in complex scenarios.
Meng: If it works well at scale, that’s a big deal for production deployment because it means we can trust which features we’re using.
Tom: And asymptotically, they showed that for non-overfitted models, those Bayes factor rates approach one in the limit.
Jane: That means if your model is actually right and you have enough data, this method will correctly identify the most probable structure without getting lost in noise.
Lu: But remember what they said about finite samples—the actual performance still depends on factors like sample size and how much censoring there is.
Meng: So, while the theory looks great for large datasets, we still have to be careful about our follow-up time and data quality when we deploy this in the real world.
Tom: The paper confirms that structure selection usually doesn't drastically change variable selection performance, but they did find a few things interesting regarding how misspecification affects power to detect small effects.
Jane: It’s a good reminder that even with this sophisticated framework, you still have to check if your baseline hazard shape is modeled correctly in finite samples.
Lu: So the core idea here is taking a flexible General Hazard model and giving it a rigorous, multi-layered Bayesian prior structure to handle both variables and structures at once.
Tom: It’s about getting the right tools for the job so we can make more principled decisions about which features are useful and what kind of effect they have before we commit to a final structure.
Jane: And while it’s a solid theoretical step, the next big thing will be testing how these ideas translate when we apply them directly to sequence-derived feature distributions from deep sequencing data.
Conclusion: Tom: So we’ve covered the main points of this paper on "Bayesian variable and hazard structure selection in the General Hazard model."
Jane: To recap, they’ve shown how to simultaneously select variables and determine the underlying hazard structure within this flexible General Hazard framework.
Lu: The core idea is using specific g-priors for the regression coefficients to ensure consistent model selection across different structures like proportional hazards or accelerated failure time.
Meng: It’s about giving the AI a way to test not just if a variable matters, but also whether it’s affecting the hazard rate or maybe how fast time moves.
Tom: And they built tools like the extended Add-Delete-Swap algorithm for joint sampling, which makes exploring this model space much more practical than just theory.
Jane: They even put in fast approximations for those likelihood calculations so the MCMC sampling doesn't take forever to run.
Lu: The results show that variable selection performance is stable across different scenarios in terms of sensitivity and specificity when you use these methods.
Meng: I saw some real data tests on the FLC dataset, and they confirmed this methodology can actually find important variables and correctly identify their roles.
Tom: They even tested it on the NKI70 dataset, which was pretty tough with its many predictors but small sample sizes and high censoring rates.
Jane: Even there, both methods agreed that the accelerated failure time model was the most probable hazard structure they found.
Lu: That suggests that even in this flexible General Hazard setup, sometimes the data strongly points toward a simpler structure like Accelerated Failure Time being correct.
Tom: Looking at how they analyze things asymptotically, for non-overfitted models, those Bayes factor rates approach one in the limit.
Jane: That means if your model is actually right and you have enough data, this method will correctly identify the most probable structure without getting lost in noise.
Lu: But they also flagged that in finite samples, performance still depends on things like sample size and how much censoring there is.
Meng: So while the theory looks great for large datasets, we still have to be careful about our follow-up time and data quality when we deploy this in the real world.
Tom: The paper confirms that structure selection usually doesn't drastically change variable selection performance, but they did find a few things interesting regarding how misspecification affects power to detect small effects.
Jane: It’s a good reminder that even with this sophisticated framework, you still have to check if your baseline hazard shape is modeled correctly in finite samples.
Lu: The core idea here is taking a flexible General Hazard model and giving it a rigorous, multi-layered Bayesian prior structure to handle both variables and structures at once.
Meng: It’s about getting the right tools for the job so we can make more principled decisions about which features are useful and what kind of effect they have before we commit to a final structure.
Tom: And Lalam is pretty sure this approach is very solid, connecting it directly to what it changes for survival analysis by making variable selection and structure determination one unified task.
Jane: Exactly, so while the framework itself is complex, the goal is giving practitioners a more rigorous way to select both inputs and outputs of their survival models at the same time.
Lu: The next big thing we're looking at will be testing how these ideas translate when we apply them directly to sequence-derived feature distributions from deep sequencing data.
More episodes
- Daily Summary for 2026-11-19
- 2310.07973-Statistical Performance Guarantee for Subgroup Identification with Generic Machine Learning
- 2408.09155-Learning Robust Treatment Rules for Censored Data
- Daily Summary for 2026-11-19
- 2603.23397-Kinetic Langevin Splitting Schemes for Constrained Sampling
- 2606.09906-An information-geometric framework for mapping maximum potential biodiversity
- 2603.27718-Induced replication and the assessment of models
- 1307.0898-Entropy of a Zipfian Distributed Lexicon
- 2406.19974-Generalizing self-normalized importance sampling with couplings
- 2412.10213-Collaborative Optimal Design of Controlled Experiments with Subject Covariates