marp: Model-Averaged Inference for Parametric Renewal Processes in R

arXiv:2611.07519 · stat.CO, stat.ME · Submitted 2026-11-05 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Statistics. The day in statistics: methods, theory and the inference under them.

Tom: I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "marp: Model-Averaged Inference for Parametric Renewal Processes in R".

Jane: Here's the gist of the paper. Renewal processes model inter-event times between recurrent events using a parametric distribution, but selecting a single model can ignore model uncertainty and affect inference, particularly in small samples marp introduces an R package for model-averaged inference in parametric renewal processes.

Tom: Renewal processes are widely used to model recurrent events through the distribution of inter-event times (Smith, 1958; Ross, 1996; Cook and Lawless, 2007).

Jane: Common choices include the exponential, gamma, Weibull, log-normal, log-logistic and Brownian passagetime (BPT) distributions.

Tom: When the data do not clearly favour any particular model, inference based on a single selected model ignores model uncertainty and can affect point and interval estimation, particularly in small samples (Chatfield, 1995; Draper, 1995; Buckland et al., 1997).

Title and authors: Tom: So we're looking at this paper called "marp: Model-Averaged Inference for Parametric Renewal Processes in R." Basically, it tackles a problem where you use different mathematical models to describe how often something happens, like a recurring event. The authors say picking just one model can totally ignore the uncertainty about which model is actually right, especially when your sample size is small.

Jane: That sounds really practical for people working in fields where things are hard to measure precisely. So, what exactly does this marp package do? Does it just pick the best model for you automatically?

Tom: Not quite. It fits six different types of waiting-time distributions—exponential, gamma, Weibull, log-normal, log-logistic and Brownian passage time—and then it uses something called Akaike and Bayesian information criteria to weigh those models. It gives you both a single best guess based on the favorite model and a model averaged estimate that combines all the estimates.

Lu: What I find interesting is how they handle that uncertainty; they don't just pick one thing and call it done. They use bootstrap inference, refitting each candidate model inside every resample to get confidence intervals for those averaged estimates, including percentile and Studentised ones.

Meng: From a production standpoint, that sounds heavy on computation if you have a massive stream of data coming in constantly. How scalable is this unified framework when we're dealing with high-throughput systems?

Lalam: It's designed to be a unified framework, which means the core idea is combining the model-specific estimates using weights derived from how well each candidate model fits the data, which they call relative support >

Jane: So it sounds like they are giving you options instead of just one answer when you're trying to figure out something about those recurrent events. How does this compare to what we usually do in statistics?

Tom: Usually, if you don't know which distribution is right, you pick the one that looks best on a graph or maybe just use the one with the lowest error metric, but marp tries to quantify that model uncertainty directly.

Lu: The paper specifically focuses on common estimands like the mean inter-event time and event probability at a threshold y, and they also look at how you define things like hazard functions under different scales.

Meng: Are we talking about just averaging the final numbers, or does this method change how we think about the underlying process itself? Like, is it just a statistical tool or something that influences the engineering decisions?

Tom: It's both. The results in their simulation study showed that while model averaging didn't necessarily improve the point estimate of things like the mean inter-event time, it often improved other types of estimates and provided more stable confidence intervals >

Jane: That’s a subtle but important distinction—it doesn't always change the main number you are trying to predict, but it makes the uncertainty around that prediction much clearer. What about the specific improvements they suggest for future work?

Lu: They suggest distinguishing between natural and transformed scales because averaging hazard functions directly isn't always equivalent to averaging log-hazards and transforming back >

Meng: That sounds like a lot of extra bookkeeping, which is something I can relate to. If we have to track different scales, how does that affect deployment complexity?

Tom: Well, they are trying to be super thorough. They also suggest using a double bootstrap procedure in their confidence intervals, which uses inner bootstrap samples within the outer resample to get those Studentised confidence intervals >

Jane: That sounds like they're really trying to account for both how much variability is in the data and how uncertain we are about which model we chose. That’s a big step for uncertainty quantification.

Lalam: And they also return bootstrap model-selection frequencies, which gives you a way to check the stability of the selection itself across different resamples >

Tom: It sounds like they're giving us a much richer set of diagnostic information than just one final result, which is what I need when shipping models.

Jane: So to wrap up on this paper called "marp: Model-Averaged Inference for Parametric Renewal Processes in R," the main value is that it makes distributional uncertainty easier to assess and report in applied renewal-process modelling >

Tom: It gives us a unified framework where we can compare selected-model summaries against model-averaged ones, inspect those model weights, and quantify uncertainty using bootstrap intervals all at once >

Lu: The simulation study shows that this approach is useful when you have several candidate parametric models that provide plausible descriptions of the data >

Meng: I just need to see if this complexity actually translates into a manageable runtime on real-world traffic, but the idea of quantifying selection stability is interesting for debugging model drift >

Lalam: It’s a very organized way to handle model comparison and averaging in parametric renewal processes using R >

The paper's summary: Tom: So, to recap, this paper is about taking all those different mathematical models we use to describe when things happen again—like waiting times—and instead of just picking one winner, they’re building a system that weighs all those possible models together using information criteria >

Jane: Exactly. It’s not just about picking the best graph; it’s about acknowledging that there's uncertainty in which model is actually correct for your data and then giving you a combined estimate that takes that uncertainty into account >

Tom: The core idea is this package, called marp, which lets you run these things together: fitting the models, doing the model averaging based on those weights, and then building confidence intervals using bootstrap methods >

Lu: What I find really interesting is how they handle the scales. They make a clear distinction between natural probabilities and logit probabilities because averaging one scale doesn't automatically give you the same result as transforming everything else back again >

Jane: That’s a big caveat for anyone trying to use these things for real-world engineering or biomedical stuff, because if you mix the scales up wrong, your final number could be totally misleading >

Tom: And then they get into some pretty heavy-duty uncertainty quantification. They use a double bootstrap procedure that does an inner sample inside every outer sample to get those Studentised confidence intervals >

Meng: From my side, I’m worried about the computational cost of running that kind of resampling for every single candidate model in a massive data stream, but if it stabilizes the estimate enough it might be worth the overhead >

Lu: The authors show that this method is useful when you have several different parametric models that all seem plausible descriptions of your data, especially when your sample size is modest >

Jane: So what does this mean for someone just looking at their data? It means instead of getting one single answer and saying "this is the mean waiting time," you get a range that reflects both how much the data points vary and how unsure we are about which underlying process we're actually observing >

Tom: It changes how we report results. You can compare a single best model estimate directly against this model averaged estimate, which gives you a clearer picture of where the uncertainty is coming from >

Lu: The authors also provide these diagnostic tools, like bootstrap model-selection frequencies, which tell you how often each candidate model gets picked across different resamples >

Jane: That’s important because it lets you check if your selection process is stable or if it keeps jumping around randomly depending on the data noise >

Tom: So the big value here is making distributional uncertainty much easier to assess and report in applied renewal process modeling, letting you see both what the best guess is and how much you should trust that guess >

Jane: It’s a solid framework for when you’re dealing with those recurrent events where you can't be totally sure which mathematical shape the data actually follows, so next time we look at something like this, we'll see if they can apply this to more complex, non-independent event processes.

The paper's improvements: Tom: So, we’re talking about what the authors think marp needs to get even better for production use, right? It seems like they’re focused on making sure that when you run this thing in a real system, it doesn't just give you a pretty number but also gives you reliable diagnostics >

Jane: Right. They point out that their current method is good for getting the main numbers, but they need stronger tools to show *how* stable those estimates are when things change or when the data quality dips >

Lu: The improvements they suggest focus on making the workflow totally integrated. Instead of you having to do all these steps—fitting, averaging, then bootstrapping—it wants one unified framework where those model-averaged inferences happen smoothly >

Tom: So less manual work for the user. They want the whole process from picking a model to getting your confidence intervals wrapped up in one go, which is way more efficient when you’re dealing with a lot of different models to test >

Jane: And they’re also being specific about how they define things. They insist on clearly separating the natural probability scale from the transformed scales because mixing those up can completely mess with the math >

Lu: That distinction is important because averaging hazard functions directly isn't always equivalent to averaging log-hazards and transforming everything back, so you have to be careful about that equivalence >

Meng: From an engineering standpoint, I agree. If the underlying math is confusing, it just means more surface area for bugs when we try to scale it up >

Tom: And they’re pushing for better diagnostics on model selection itself. They want the package to tell you how often each model gets picked during your bootstrap resampling >

Jane: That’s smart. Because if you see that one model is selected almost all the time, you know your data strongly favors that structure, which gives you confidence in your result >

Lu: Plus, they are looking at ways to make the optimization process more robust. They want stronger tools for optimizing those candidate models to ensure they find the best fit reliably >

Tom: So it’s about making sure the package is not just a nice statistical tool but something that's actually sturdy enough to ship in production environments where things get messy >

Jane: The big picture here is that they are trying to move this from a neat research paper into something you can actually trust when you’re making decisions based on these recurrent event patterns >

Lu: Their future work also looks at extending this beyond simple renewal processes, moving toward more complex, covariate-dependent models or even processes where the events aren't independent >

Tom: Exactly. They want to see if this model averaging technique can handle those more complicated scenarios where the events actually depend on other factors in your dataset >

Jane: It’s an interesting direction because that’s where most real-world data lives, so seeing how this framework handles those kinds of dependencies is really what we need to watch for next >

Conclusion: Tom: So, to wrap up, this paper on "marp: Model-Averaged Inference for Parametric Renewal Processes in R" is essentially about giving us a unified way to handle uncertainty when we use multiple mathematical models for recurring events >

Jane: Right. It shows you how to combine the results from different models using weights derived from how well each one fits your specific data, and then quantify the uncertainty around that combined estimate >

Lu: The authors really drive home the idea that this isn't just a single-model fit; it’s about building a more robust system that acknowledges and weighs model uncertainty directly in the inference process >

Meng: It sounds like a solid way to handle those tricky situations where you have several plausible distributions, but I still worry about how computationally heavy the bootstrap procedures are when you try to run them across many models >

Lalam: I’m confident that this framework provides a much more transparent way to report results than just picking one model blindly; it shows the selection process itself >

Jane: So for someone just listening, it means when you see a result from this paper, you can look at the model weights and see exactly how much confidence is attached to each candidate distribution >

Tom: It gives us that comparison between the selected-model estimate and that averaged estimate, which makes it much clearer where the uncertainty really resides >

Lu: The authors are also looking ahead to making this even more powerful, exploring how it can handle more complicated structures like covariate-dependent models or processes where events aren't independent >

Meng: If they can actually build stable tools for those more complex structures, then this has real potential for production modeling instead of just theoretical exercises >

Lalam: I think the paper lays a strong foundation because it gives us these clear diagnostic tools to check the stability of model selection across different resamples, which is crucial when you’re deploying models >

Jane: That sounds like a good direction to look at next, especially since we’ve seen how other papers are tackling similar issues with missing data or complex statistical theory >

Tom: We’ll definitely keep an eye on how they tackle those more complicated scenarios because that's where the real application lies >

School of Mathematics and Statistics, The University of Sydney · Charles Perkins Centre, The University of Sydney · Sydney Precision Data Science Centre, The University of Sydney

stat.CO, stat.ME

Submitted: 2026-11-05

Updated: 2026-11-05

Comments: 32 pages, 8 figures; includes supplementary material

Code: https://github.com/kanji709/marp

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 77/100

The gist: Renewal processes model inter-event times between recurrent events using a parametric distribution, but selecting a single model can ignore model uncertainty and affect inference…

Key concepts

Renewal Processes
These are statistical models used to describe events that occur repeatedly over time, such as customer arrivals or equipment failures. The key feature is that the time between one event and the next (the inter-event time) follows a specific probability distribution.
Model Averaging
When data don't strongly favor one specific mathematical model among several possibilities, selecting just one can be misleading. Model averaging combines estimates from all candidate models using weights derived from how well each model fits the data, resulting in a more stable and less biased final estimate.
Double Bootstrap Procedure
This method is used to calculate confidence intervals for the model-averaged results. It involves running an outer bootstrap resampling process and then performing an inner bootstrap within each resample to generate Studentized confidence intervals, which accounts for both sampling variability and uncertainty arising from the choice of candidate models.

Terminology

Summary

Renewal processes model inter-event times between recurrent events using a parametric distribution, but selecting a single model can ignore model uncertainty and affect inference, particularly in small samples marp introduces an R package for model-averaged inference in parametric renewal processes. p. 0

Where the paper says it

au) Abstract Renewal processes model inter-event times between recurrent events using a parametric distribution, but selecting a single model can ignore model uncertainty and affect inference, particularly in small samples. This paper introduces marp, an R package for model-averaged inference in parametric renewal processes.

Introduction to the Framework

Renewal processes are widely used to model recurrent events through the distribution of inter-event times (Smith, 1958; Ross, 1996; Cook and Lawless, 2007). Common choices include the exponential, gamma, Weibull, log-normal, log-logistic and Brownian passagetime (BPT) distributions. p. 0

Where the paper says it

1 Introduction Renewal processes are widely used to model recurrent events through the distribution of inter-event times (Smith, 1958; Ross, 1996; Cook and Lawless, 2007). Common choices include the exponential, gamma, Weibull, log-normal, log-logistic and Brownian passagetime (BPT) distributions, which accommodate hazard functions ranging from the constant (Lawless, 2003) to flexible monotone (Lawless, 2003; Klein and Moeschberger, 2003) and non-monotone forms (Lawless, 2003; Bennett, 1983).

When the data do not clearly favour any particular model, inference based on a single selected model ignores model uncertainty and can affect point and interval estimation, particularly in small samples (Chatfield, 1995; Draper, 1995; Buckland et al. p. 0
Where the paper says it

Candidate models are typically fitted by maximum likelihood and selected using the Akaike Information Criterion (AIC) (Akaike, 1973) or Bayesian Information Criterion (BIC) (Schwarz, 1978). When the data do not clearly favour any particular model, inference based on a single selected model ignores model uncertainty and can affect point and interval estimation, particularly in small samples (Chatfield, 1995; Draper, 1995; Buckland et al.

, 1997). Frequentist model averaging addresses this by combining model-specific estimates using weights derived from the relative support for each candidate model (Hjort and Claeskens, 2003; Claeskens and Hjort, 2012). p. 0
Where the paper says it

When the data do not clearly favour any particular model, inference based on a single selected model ignores model uncertainty and can affect point and interval estimation, particularly in small samples (Chatfield, 1995; Draper, 1995; Buckland et al. Frequentist model averaging addresses this by combining model-specific estimates using weights derived from the relative support for each candidate model (Hjort and Claeskens, 2003; Claeskens and Hjort, 2012).

The Statistical Methodology

The framework assumes a renewal process where inter-event times are independent and identically distributed non-negative random variables (Smith, 1958). The candidate model set in marp consists of six parametric waiting-time distributions: exponential, gamma, Weibull, log-normal, log-logistic and Brownian passage time (BPT). p. 1

Where the paper says it

Censoring covariates, subject-level heterogeneity and dependence between successive waiting times are outside the scope of the current implementation. 2 Candidate models The candidate model set in marp consists of six parametric waiting-time distributions: exponential, gamma, Weibull, log-normal, log-logistic and Brownian passage time (BPT).

Model-specific parameter vectors are estimated by maximum likelihood. p. 2
Where the paper says it

The model-specific parameter vector is estimated by maximum likelihood. If ℓk (θk ) denotes the log-likelihood under model Mk , then the maximum likelihood estimator is bθk = arg max θk∈Θk ℓk (θk ).

The analysis focuses on common estimands such as the mean inter-event time, event probability at a threshold y, and the hazard function. p. 2
Where the paper says it

3 Estimands The analysis focuses on estimands that are common to all candidate waiting-time models. The main quantities are the mean inter-event time, µ = E(X), the event probability, or distribution function, at a threshold y, F(y) = Pr(X ≤ y), and the hazard function at selected time points t, h(t) = f(t) S(t) .

Model Comparison and Averaging

The selected-model estimator is obtained by choosing the model with the minimum Akaike Information Criterion (AIC). AIC weights are computed as exp(-∆k/2) summed over all candidate models, where ∆k is the difference between AIC and the minimum AIC. p. 3

Where the paper says it

AIC weights are computed as wk = exp(−∆k/2) ∑ K j=1 exp(−∆j/2) , ∆k = AICk − min j AICj . These weights measure relative empirical support within the specified candidate model set.

The model-averaged estimator of a target functional ψ is defined as ψbMA = K sum wkψbk. p. 3
Where the paper says it

They are used for model-averaging weights but should not be interpreted as posterior probabilities that the models are true. The AIC-weighted model-averaged estimator of ψ is ψbMA = K ∑ k=1 wkψbk .

Uncertainty Quantification

Bootstrap inference is applied by refitting candidate models within each outer bootstrap resample, recalculating information criteria, and recomputing estimates. Confidence intervals for model-averaged estimates are obtained from the bootstrap distribution, including percentile and Studentised bootstrap intervals. p. 0

Where the paper says it

This paper introduces marp, an R package for model-averaged inference in parametric renewal processes. The package fits six commonly used distributions, compares them using Akaike and Bayesian information criteria, and reports selected-model and model-averaged estimates of the mean inter-event time, event probability and hazard function, with percentile and Studentised bootstrap confidence intervals.

Practical Applications

The package illustrates its use in two real time-to-event datasets: a biomedical recurrence example based on seizure onset times and an engineering reliability example based on airplane glass failure times. These examples demonstrate how marp can be used when the number of observed waiting times is modest and several candidate parametric models provide plausible descriptions of the data. p. 1

Where the paper says it

2 Candidate models The candidate model set in marp consists of six parametric waiting-time distributions: exponential, gamma, Weibull, log-normal, log-logistic and Brownian passage time (BPT). These distributions provide a range of possible inter-event-time behaviours and hazard shapes, from the constant hazard implied by the exponential model to more flexible monotone or non-monotone hazards under the remaining candidate models.

The results show that model averaging can be useful, but should be accompanied by diagnostics, model-weight summaries and bootstrap failure rates. p. 0
Where the paper says it

A simulation study under three data-generating models showed that model averaging did not improve point estimation but often improved interval coverage relative to the AIC-selected model. Applications to seizure-recurrence and engineering failure-time data illustrate the proposed methods.

Conclusion

The main value of marp is that it makes distributional uncertainty easier to assess and report in applied renewal-process modelling. Users can compare selected-model and model-averaged summaries, inspect model weights, and quantify uncertainty using bootstrap intervals within a single framework. p. 3

Where the paper says it

Therefore, the software implementation and simulation study distinguish between natural- and transformed-scale summaries. 5 Bootstrap confidence intervals Bootstrap inference is used to quantify uncertainty in both selected-model and modelaveraged estimators.

The simulation study and applications show that model averaging can be useful, but should be accompanied by diagnostics, model-weight summaries and bootstrap failure rates. p. 0
Where the paper says it

A simulation study under three data-generating models showed that model averaging did not improve point estimation but often improved interval coverage relative to the AIC-selected model. Applications to seizure-recurrence and engineering failure-time data illustrate the proposed methods.

Future development will focus on stronger diagnostic tools, more robust optimisation, additional candidate distributions, covariate-dependent models, and extensions for heterogeneous or dependent recurrent-event processes. p. 0
Where the paper says it

In renewal-process models, model averaging combines estimates of common estimands, such as the mean inter-event time, event probabilities and hazard function, across multiple candidate waiting-time distributions to account for model uncertainty. This motivates the development of a unified framework for model-averaged inference in renewal processes.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Integrated workflow implementation for model-averaged inference This package provides a unified framework for fitting parametric renewal-process models, performing information-criterion model averaging, and constructing bootstrap confidence intervals, which reduces the amount of manual implementation required by combining these steps in a common framework.

  2. Explicit distinction between natural and transformed scales This paper distinguishes between the natural event probability F(y), the logit event probability logit F(y), the hazard h(t), and the log-hazard log h(t) because averaging hazards on the natural scale is not equivalent to averaging log-hazards and transforming back.

  3. Robust uncertainty quantification through double bootstrap procedures This method, implemented in confint, uses a double bootstrap procedure where inner bootstrap samples are drawn within each outer resample to obtain Studentised confidence intervals, which incorporates both sampling variability and model-selection uncertainty.

  4. Enhanced diagnostic reporting for practical application analysis The package returns bootstrap model-selection frequencies as descriptive measures of selection stability, allowing users to inspect how often different models are selected across bootstrap samples.

Abstract

Renewal processes model inter-event times between recurrent events using a parametric distribution, but selecting a single model can ignore model uncertainty and affect inference, particularly in small samples. This paper introduces marp, an R package for model-averaged inference in parametric renewal processes. The package fits six commonly used distributions, compares them using Akaike and Bayesian information criteria, and reports selected-model and model-averaged estimates of the mean inter-event time, event probability and hazard function, with percentile and Studentised bootstrap confidence intervals. A simulation study under three data-generating models showed that model averaging did not improve point estimation but often improved interval coverage relative to the AIC-selected model. Applications to seizure-recurrence and engineering failure-time data illustrate the proposed methods. The marp package provides a reproducible framework for inference that accounts for model uncertainty in parametric renewal processes.

Related papers