Stage-wise semiparametric analysis of multivariate event times under informative censoring via vine copula

arXiv:2502.20608 · stat.ME, stat.AP, stat.CO · Submitted 2025-02-28 · 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: "Stage-wise semiparametric analysis of multivariate event times under informative censoring via vine copula".

Jane: Here's the gist of the paper. The study introduces a novel analysis framework leveraging vine copulas to directly estimate the joint density of multivariate times to nonterminal and terminal events This approach excels in capturing the heterogeneous dependence between each pair of event times (nonterminal-terminal and between-nonterminal) in terms of strength and structure Problem Statement The primary challenge in analyzing such multivariate event times is the presence of informative censoring by the terminal event Existing literature primarily focuses on data involving a single nonterminal event, commonly known as semi-competing risks Two recent studies have explored the analysis of the METIC data However, these methods rely on relatively strong structural assumptions and do not offer convenient approaches to conducting regressions on individual nonterminal event times Specifically, the nested copula structure in Li et al.

Tom: (2023) assumes that each nonterminal event time has identical associations (in terms of strength and functional form) with the terminal, and that each pairwise between-nonterminal dependence is homogeneous Furthermore, the multivariate copula in Li et al.

Jane: (2020) and the bivariate copula in Li et al.

Tom: (2023) are restricted to the Archimedean copula family Vine Copula Modeling The approach leverages vine copulas (Joe, 2014) to decompose the complex multivariate dependence among all event times into a catalog of pairwise dependencies organized through a sequence of trees (connected acyclic graphs) Each pairwise dependence is represented by a bivariate copula, which can be from a different family The first tree assumes a star shape, with the root node representing the terminal event time and connecting all nonterminal event times This configuration enables direct regression modeling for each nonterminal event time The joint copula density is given in Proposition 1 Stage-wise Estimation The authors develop a computationally efficient stage-wise pseudo maximum likelihood estimation (PMLE) procedure to estimate the parameters associated with each tree sequentially The process involves three stages: 1.

Title and authors: Tom: So we're diving into this paper now, "Stage-wise semiparametric analysis of multivariate event times under informative censoring via vine copula". It looks like the core idea is tackling that tricky problem where you have multiple events happening, and one event—the terminal one—is censoring the others in a way that's not random.

Jane: Exactly. The naive question here is how do we even start modeling this when we have these multiple event times, each potentially having different relationships with each other? It seems like existing tools are pretty limited, mostly focusing on just one nonterminal event at a time.

Lu: Yeah, and what this paper aims to do is introduce a new framework that uses vine copulas to directly estimate the joint density of all those multivariate times to both nonterminal and terminal events in one go. It’s trying to handle the fact that different pairs of events might have completely different dependence structures.

Meng: From an engineer's standpoint, I’m wondering how much computational overhead this vine copula approach adds compared to just running a standard regression model on each event time separately? Because if we have thousands of data points, that sequential estimation sounds slow.

Tom: That's a fair point, Meng. But the paper suggests a stage-wise pseudo maximum likelihood estimation procedure that breaks the problem down into smaller steps for each tree in the vine copula structure. It claims this is more efficient than trying to maximize the full likelihood all at once on page two of THIS PAPER <ref:2502.20608#pg1>.

Jane: That sounds like they are trying to make it manageable without giving up on capturing that heterogeneity we talked about earlier, Tom. So, what’s the main payoff from using this vine copula structure instead of those simpler models?

Lu: The main thing is that vine copulas decompose the complex multivariate dependence among all event times into a catalog of pairwise dependencies organized through a sequence of trees, and each tree uses a different bivariate copula family. This allows them to capture the heterogeneity in association between every pair of events, whether it's nonterminal-terminal or between-nonterminal.

Meng: So they are not assuming everything is uniform across the board? Because that’s what those nested copula models did, and we know that assumption is often where things break down in real data.

Tom: Right. And the authors suggest a way to fix that by using a sequence of bivariate copulas, which means each pair can have its own specific mathematical structure, not just one Archimedean family like before.

Jane: That flexibility is key because the paper points out that existing methods assume too much homogeneity in dependence between pairs of events, which isn't realistic for data like crowdfunding or medical survivorship studies.

Title and authors: Lu: And they specifically address the issue where previous work assumed identical associations between each nonterminal event time and the terminal event, which page three of THIS PAPER points out as a problem.

Tom: So instead of forcing everything into one symmetric structure, they are building a flexible catalog using the vine copula to model those different pairwise relationships directly.

Jane: And what about the actual estimation part? How do they actually get those joint density estimates without getting stuck in an infinite loop or something similar?

Lu: They use this stage-wise pseudo maximum likelihood estimation procedure, which is broken into three stages: first estimating parameters for the first tree, then separately estimating parameters for each nonterminal event time using a semi-competing risk method, and finally estimating the bivariate copula parameter for each edge.

Meng: That sounds like a lot of separate optimization steps. How robust are these pseudo maximum likelihood estimates when we move this to a production environment where we need fast, reliable inference?

Tom: The simulation studies on page five show that the finite sample performance is satisfactory, and the relative bias falls as the sample size increases, which gives some confidence in the consistency of their estimation procedure.

Jane: So they’re showing that even with this complex setup, their estimators don't totally drift away from the true values when you get more data. That's reassuring for anyone trying to apply this to real-world data.

Lu: Simulation study two shows that the vine copula method actually results in a much smaller relative bias for Kendall’s tau parameters compared to that nested modeling approach they compared it against, which is a pretty strong result.

Meng: That comparison is interesting because it shows the practical advantage of the vine copula over those highly symmetric structures they were trying to replace. What about running this on actual data?

Tom: The application to crowdfunding data showed that none of the covariate effects turned out to be significant under any copula family they tested, which is interesting because it suggests the underlying dependence structure might not be as easily influenced by those covariates.

Jane: Wait, so even though they found no significant covariate effects in one dataset, they did find a significant negative effect from email verification status on all four marginals. That seems like a concrete finding we can use.

Lu: That’s because the analysis revealed that the email verification status has a significant negative effect on all four marginals, even though the copula structure itself didn't pick up those covariate effects as significant.

Title and authors: Meng: So it means the model is sensitive to some things, but maybe not in the way we expected from a simple linear adjustment. What’s next for this framework?

Tom: The paper does suggest that to make this more robust, you could use non-parametric bivariate copulas for the pairs and sieve approaches for modeling those marginals, which would help avoid some of the misspecification issues they encountered.

Jane: It sounds like the authors are being honest about where their current model might need strengthening if we want to push it further into new applications.

Lu: They also flag a limitation: because they achieved a closed-form expression for the joint density function of T to J, they still require numerical integrations to get the joint survival and log likelihood functions.

Meng: So, even with all this machinery, you still need numerical integration at some point for those final calculations? That adds complexity back into the deployment pipeline.

Tom: Yeah, that’s a real caveat for production use; it means it’s not a fully closed analytical solution that you can just plug numbers into and get an answer instantly.

Jane: So to wrap up on this paper, the work "Stage-wise semiparametric analysis of multivariate event times under informative censoring via vine copula" gives us a way to model complex relationships between multiple event times when one is censored by the terminal event, using flexible vine copulas instead of rigid nested structures.

Lu: It allows for direct regression modeling of all marginals, which is a big deal because it lets you look at individual nonterminal event times without first having to detour through the survival functions of the first event time.

Meng: From an engineering view, it’s a powerful tool for understanding complex dependencies, but we need those numerical integration steps and the flexibility in specifying the copulas to be handled carefully when we ship this out.

Tom: It’s a solid framework for dissecting these kinds of multivariate event times, and it shows how vine copulas can handle heterogeneous dependence structures better than some previous approaches.

Jane: It gives us a more nuanced tool for analyzing data where censoring is informative, pushing us closer to understanding the full picture of what’s happening across all those related events.

Lu: For future work, exploring non-parametric bivariate copulas or sieve approaches for marginals would be the next logical step to make this framework even more adaptable and less restrictive.

Meng: I hope they look into making the numerical parts faster so it doesn't slow down our inference times when we actually try to deploy these models at scale.

Tom: That’s the engineering reality for most complex statistical methods, but the conceptual advancement here in how they decomposed that dependence is definitely worth paying attention to.

The paper's summary: Tom: so we’ve seen how they use those vine copulas to break down the complex web of dependencies between all these different event times, and now we’re looking at what that whole framework is actually trying to do in practice.

Jane: it seems like the main goal is getting a direct picture of the joint density for every single event time—both the nonterminal ones and that final terminal one—even when the terminal event is censoring us.

Tom: yeah, that’s right, they’re tackling informative censoring head-on by using this vine structure to organize all those pairwise dependencies into trees.

Jane: what does it mean for someone who just listens to the show? It means we can finally look at how one specific nonterminal event time relates to another one directly, without first having to go through the survival function of the terminal event.

Tom: exactly, they’re giving us direct regression on all those marginals, which is a big deal because it cuts out that detour through the first event time survival functions.

Jane: and they do that by using this stage-wise estimation procedure, which basically means they don't try to solve everything at once but tackle the trees one by one—first the root node connecting everything, then each nonterminal event time individually, and finally each pair of events.

Tom: it’s a computational trade-off; they claim this sequence is more convenient for implementing in practice with lower computational cost than trying to do a full simultaneous maximum likelihood estimation on all the data at once.

Jane: and the results from their simulations show that even with this stage-wise setup, the estimators are actually consistent as the sample size grows, meaning they don't drift away from being correct.

Tom: and simulation two specifically showed that this vine copula method gave a much smaller relative bias for those Kendall tau parameters compared to those nested models they were comparing it against earlier.

Jane: so what does this mean for real data? it suggests that when you have complex dependence structures, like the ones in crowdfunding or medical records, you can use this flexible modeling instead of sticking to the simpler models where they assume every pair of events is exactly the same way.

Tom: and they’re also being pretty pragmatic about limitations; because they found a closed-form expression for the joint density function, they still need to do numerical integration to get those final survival functions and log likelihoods, which adds some complexity to the deployment pipeline.

Jane: so it’s powerful but not a plug-and-play analytical solution yet, and what they suggest for future work is using non-parametric copulas or sieve approaches for the marginals to make the model even more flexible.

Tom: yeah, so it’s a solid framework for dissecting these multivariate event times, and we should keep an eye on how those numerical integration steps can be optimized to make this framework faster when we actually try to deploy these models at scale.

The paper's improvements: Tom: so we’ve looked at how they set up that stage-wise estimation using vine copulas to handle those different pairwise dependencies, and now we’re looking at what they suggest for making that whole process even better.

Jane: it seems like the authors are being realistic about where the current model could be stronger, and they point out a couple of ways to beef up their framework.

Tom: right, they say that instead of just using standard bivariate copulas for every pair, you could specify them non-parametrically, which means you don't have to assume a fixed mathematical shape for the dependence between any two event times.

Jane: that’s a big deal because it helps avoid misspecification issues; if the true relationship isn't perfectly Archimedean or whatever they were using, then being non-parametric gives you more room to fit the data structure.

Tom: and they also suggest using sieve approaches for modeling those marginals—the individual event times themselves—which is something they already referenced from Zhao et al. in a different paper.

Jane: that sounds like a good way to make the marginals more flexible, especially when you’re dealing with things like crowdfunding data where the underlying distribution might be messy.

Tom: and then there's this interesting plug-in estimate they mention—for the nonterminal event times and their copula parameters—which they get because under certain Archimedean copulas, they can explicitly write that nonterminal event time as a function of the terminal event probability, that specific nonterminal survival function, and some other parameters.

Jane: so it’s a way to get a more direct estimate for those individual marginals without having to rely entirely on the complex vine structure for every single calculation.

Tom: but here’s the caveat they put out there; because they got that closed-form expression for the joint density function of T to J, they still have to use numerical integration when you want the actual joint survival functions and log likelihood functions.

Jane: so while it’s a huge step in decomposing dependence, it doesn't magically give us a perfect analytical answer for every calculation; we still need those numerical steps.

Tom: that’s the reality of their current development; they aren't claiming a fully closed solution for everything yet, which keeps them grounded on what they can actually run efficiently.

Jane: so while the framework is conceptually very sound for handling heterogeneous dependence, it needs those non-parametric tools and maybe faster numerical methods to be truly ready for widespread use.

Tom: it’s a solid path forward, moving from a strong theoretical decomposition to a more flexible modeling approach that can handle the messy reality of real-world data better.

Conclusion: Tom: so to wrap up, we’re talking about "Stage-wise semiparametric analysis of multivariate event times under informative censoring via vine copula." basically, this paper shows how you can use these vine copulas to model the complex web of dependencies between multiple event times when one is being censored by the terminal event.

Jane: it means we get a much more direct handle on those relationships without having to rely on overly restrictive assumptions about how all those events are linked together.

Tom: and they show that this stage-wise approach, while computationally staged, gives us consistent results when you run the simulation studies, especially compared to those nested models they looked at before.

Jane: it’s a pretty solid way to move beyond just looking at one event time in isolation and start seeing the whole picture of what’s happening across all those related events.

Tom: and for someone listening who only cares about the practical side, it means you can run direct regressions on individual nonterminal event times, which is much cleaner than having to go through the first event time survival functions first.

Jane: that’s a huge win for anyone trying to build a model that actually has some real predictive power in its deployment pipeline.

Tom: Lu, what’s your take on the overall ambition of this framework? Does it really change how we think about these multivariate problems?

Lu: i think the idea of using vine copulas to decompose dependence into those trees is powerful because it lets you move past assuming everything is perfectly symmetric across all pairs.

Meng: from a production standpoint, i’m still watching those numerical integration requirements; if that integration doesn't become faster, it adds too much overhead for our systems.

Lalam: the literature confirms this direction is promising; the way they handle heterogeneity in dependence structures is something we see a lot in complex time series analysis.

Tom: and Lalam, you’re saying this method has enough confidence to be considered reliable for analyzing these types of data?

Lalam: i'm pretty sure it does; based on the evidence they presented, i'm confident in their estimation procedure for this framework.

Jane: so we’ve seen the mechanics, the simulation results, and the limitations—it really shows how flexible vine copulas can be when dealing with informative censoring.

Tom: yeah, it’s a strong contribution to handling these kinds of complex survival data problems in a way that acknowledges real-world dependency structures.

Jane: so keep an eye on those suggestions for non-parametric copulas; that’s where the next big refinement for this whole approach is going to be.

Xinyuan Chen, Yiwei Li, Qian M. Zhou

Department of Mathematics and Statistics, Mississippi State University · Department of Marketing and International Business, Lingnan University

stat.ME, stat.AP, stat.CO

Submitted: 2025-02-28

Updated: 2026-11-02

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

Importance score: 80/100

The gist: The study introduces a novel analysis framework leveraging vine copulas to directly estimate the joint density of multivariate times to nonterminal and terminal events This approach excels in…

Key concepts

Vine Copulas
Vine copulas are a tool used to model complex multivariate dependencies among multiple variables by breaking them down into a series of simpler pairwise dependencies organized in a tree structure. This allows the model to capture how different pairs of event times relate to each other, whether they are nonterminal-terminal or between-nonterminal.
Stage-wise Pseudo Maximum Likelihood Estimation (PMLE)
This is a computational method used to estimate the parameters of the vine copula model sequentially, one tree at a time. Instead of solving for all parameters at once, it estimates them in stages. This makes the process more manageable computationally and allows for direct regression modeling of individual event times.
Informative Censoring
This occurs when the terminal event time is not observed directly but is influenced by the occurrence of other events. In this context, it means that knowing when a nonterminal event happens provides extra information about when the final terminal event might occur, complicating standard survival analysis.
Marginal Regression Modeling
This refers to the ability of the framework to perform direct regression analyses on every individual event time (both nonterminal and terminal) using flexible semiparametric transformation models. This avoids the need to first analyze or 'detour through' the survival functions of other event times.

Terminology

Summary

The study introduces a novel analysis framework leveraging vine copulas to directly estimate the joint density of multivariate times to nonterminal and terminal events This approach excels in capturing the heterogeneous dependence between each pair of event times (nonterminal-terminal and between-nonterminal) in terms of strength and structure

Problem Statement

The primary challenge in analyzing such multivariate event times is the presence of informative censoring by the terminal event Existing literature primarily focuses on data involving a single nonterminal event, commonly known as semi-competing risks Two recent studies have explored the analysis of the METIC data However, these methods rely on relatively strong structural assumptions and do not offer convenient approaches to conducting regressions on individual nonterminal event times Specifically, the nested copula structure in Li et al. (2023) assumes that each nonterminal event time has identical associations (in terms of strength and functional form) with the terminal, and that each pairwise between-nonterminal dependence is homogeneous Furthermore, the multivariate copula in Li et al. p. 2

Where the paper says it

Specifically, the nested copula structure in Li et al. (2023) assumes that each nonterminal event time has identical associations (in terms of strength and functional form) with the terminal, and that each pairwise between-nonterminal dependence is homogeneous.

(2020) and the bivariate copula in Li et al. p. 2
Where the paper says it

In addition, the multivariate copula in Li et al. (2020) and the bivariate copula in Li et al.

(2023) are restricted to the Archimedean copula family

Vine Copula Modeling

The approach leverages vine copulas (Joe, 2014) to decompose the complex multivariate dependence among all event times into a catalog of pairwise dependencies organized through a sequence of trees (connected acyclic graphs) Each pairwise dependence is represented by a bivariate copula, which can be from a different family The first tree assumes a star shape, with the root node representing the terminal event time and connecting all nonterminal event times This configuration enables direct regression modeling for each nonterminal event time The joint copula density is given in Proposition 1

Stage-wise Estimation

The authors develop a computationally efficient stage-wise pseudo maximum likelihood estimation (PMLE) procedure to estimate the parameters associated with each tree sequentially The process involves three stages:

  1. Stage 1 estimates the parameters associated with the first tree, including model parameters in SJ (tJ Z) This step obtains the estimate θbJ by maximizing a specific log-likelihood function

  2. Stage 2 separately estimates the parameters for each nonterminal event time j ∈ [J − 1] via the PMLE method proposed for semi-competing risk data This is done by maximizing a pseudo log-likelihood function

  3. Stage k estimates the bivariate copula parameter γek separately for each edge ek ∈ Ek This involves maximizing a pseudo-log-likelihood function

Performance and Application

Simulation studies demonstrate the satisfactory finite-sample performance of the proposed stage-wise estimators and analytical variance estimators In Simulation Study I, relative bias (rBIAS) diminishes as sample size increases, demonstrating the consistency of the estimation procedure In Simulation Study II, the vine copula method shows a much smaller rBIAS for Kendall’s τ parameters compared to the nested modeling approach The application to crowdfunding data showed that none of the covariate effects are significant under any copula family The analysis revealed that the email verification status has a significant negative effect on all four marginals

Conclusion and Limitations

The proposed framework addresses heterogeneous pairwise dependencies and allows for direct regression modeling of all marginals Limitations include the requirement for numerical integrations to obtain the joint survival and log-likelihood functions The paper suggests that more robust specifications of model components could be adopted, such as non-parametric bivariate copulas or sieve approaches for marginals Additionally, the terminal event time is subject to administrative censoring in many applications The vine copula-based method requires numerical integrations to obtain the joint survival and log-likelihood functions because it is developed upon attaining a closed-form expression of the joint density function of T[J]

References

Acar, O. A., Dahl, D. W., Fuchs, C., and Schreier, M. (2021), “The signal value of crowdfunded products,” Journal of Marketing Research, 58, 644–661

Arachchige, S. J., Chen, X., and Zhou, Q. M. (2025), “Two-stage pseudo maximum likelihood estimation of semiparametric copula-based regression models for semi-competing risks data,” Lifetime Data Analysis, 31, 52–75

Beck, B. B., Wuyts, S., and Jap, S. (2024), “Guardians of trust: How review platforms can fight fakery and build consumer trust,” Journal of Marketing Research, 61, 682–699

Bedford, T. and Cooke, R. p. 5

Where the paper says it

Each edge is associated with a bivariate copula. The vine graph satisfies the following conditions (Bedford and Cooke, 2002): (i) N1 = [J]; (ii) Nk = Ek−1, k = 2, .

M. (2002), “Vines–a new graphical model for dependent random variables,” The Annals of Statistics, 30, 1031–1068

Camacho, N., Nam, H., Kannan, P., and Stremersch, S. (2019), “Tournaments to crowdsource innovation: The role of moderator feedback and participation intensity,” Journal of Marketing, 83, 138–157

Chen, X., Fan, Y., Pouzo, D., and Ying, Z. (2010), “Estimation and model selection of semiparametric multivariate survival functions under general censorship,” Journal of Econometrics, 157, 129–142

Chen, Y.-H., “Maximum likelihood analysis of semicompeting risks data with semiparametric regression models,” Lifetime Data Analysis, 18, 36–57

Cornelius, P. B. and Gokpinar, B., “The role of customer investor involvement in crowdfunding success,” Management Science, 66, 452–472

Czado, C. and Nagler, T. p. 5

Where the paper says it

, two edges in Ek−1, share a common node. We denote edge ek from tree Tk by (j1, j2|Dek ), where Dek is the conditioning set of this edge (Czado and Nagler, 2022); for T1, (j1, j2|De1 ) = (j1, j2) since De1 ≡ ∅.

, “Vine copula based modeling,” Annual Review of Statistics and Its Application, 9, 453–477

Dai, H. and Zhang, D. p. 1

Where the paper says it

1 Introduction Multivariate Event Times under Informative Censoring (METIC) data arise when a terminal event precludes multiple nonterminal events, where censoring by the terminal event is informative for the nonterminal event times due to their possible dependence. For instance, in a business study on crowdfunding (Dai and Zhang, 2019; Su et al.

J., “Prosocial goal pursuit in crowdfunding: Evidence from Kickstarter,” Journal of Marketing Research, 56, 498–517

Davis, M. K., Li, D., Wai, E., Tyldesley, S., Simmons, C., Baliski, C., and McBride, M. L. (2014), “Hospital-related cardiac morbidity among survivors of breast cancer: Long-term risks and predictors,” Journal of Cardiac Failure, 20, S44–S45

Fan, T., Gao, L., and Steinhart, Y. (2020), “The small predicts large effect in crowdfunding,” Journal of Consumer Research, 47, 544–565

Fine, J. P., Jiang, H., and Chappell, R. (2001), “On semi-competing risks data,” Biometrika, 88, 907–919

Gao, H., Wang, X. S., Li, X., and Cotte, J. (2025), “Crowdfunding success for female versus male entrepreneurs depends on whether a consumer versus investor decision frame is salient,” Journal of Marketing Research, 0, 1–20

Herd, K. B., Mallapragada, G., and Narayan, V. (2022), “Do backer affiliations help or hurt crowdfunding success? p. 0

Where the paper says it

Through extensive simulation studies, we demonstrate the satisfactory finite-sample performance of our proposed stage-wise estimators and analytical variance estimators, as well as their advantages over existing methods. We apply the developed approach to data from a crowdfunding platform to investigate the relationship between various types of creator-backer interactions and a creator’s lifetime on the platform.

” Journal of Marketing, 86, 117–134

Hofert, M. and Pham, D. p. 2

Where the paper says it

, Clayton, Gumbel, Frank); the nested copula parameters in Li et al. (2023) are required to satisfy sufficient nesting conditions (Hofert and Pham, 2013) for the joint survival function to be well-defined.

, “Densities of nested Archimedean copulas,” Journal of Multivariate Analysis, 118, 37–52

Hsieh, J.-J., Wang, W., and Ding, A. A. (2008), “Regression analysis based on semicompeting risks data,” Journal of the Royal Statistical Society Series B: Statistical Methodology, 70, 3–20

Hsu, C.-H., Taylor, J. M., Murray, S., and Commenges, D. (2006), “Survival analysis using auxiliary variables via non-parametric multiple imputation,” Statistics in Medicine, 25, 3503–3517

Jabr, W. and Rahman, M. S., “Online reviews and information overload: The role of selective, parsimonious, and concordant top reviews,” MIS Quarterly, 46, 1517–1550

Joe, H., Dependence Modeling with Copulas, Chapman & Hall/CRC Monographs on Statistics and Applied Probability, CRC Press

Kim, C., Kannan, P. K., Trusov, M., and Ordanini, A. (2020), “Modeling dynamics in crowdfunding,” Marketing Science, 39, 339–365

Kozinets, R. V., Valck, K. D., Wojnicki, A. C., and Wilner, S. J. (2010), “Networked narratives: Understanding word-of-mouth marketing in online communities,” Journal of Marketing, 74, 71–89

Li, D., Hu, X. J., McBride, M. L., and Spinelli, J. J. (2020), “Multiple event times in the presence of informative censoring: Modeling and analysis by copulas,” Lifetime Data Analysis, 26, 573–602

Li, D., Hu, X. J., and Wang, R. (2023), “Evaluating association between two event times with observations subject to informative censoring,” Journal of the American Statistical Association, 118, 1282–1294

Liu, Z., Gao, Q., and Rao, R. S. (2025), “Self-donations and charitable contributions in online crowdfunding: An empirical analysis,” Journal of Marketing, 89, 117–134

Nagler, T., “Simplified vine copula models: State of science and affairs,” Risk Sciences, 1, 100022

Nelsen, R. B., An introduction to copulas, Springer Science & Business Media

Palmatier, R. W., Jarvis, C. B., Bechkoff, J. R., and Kardes, F. R. (2009), “The role of customer gratitude in relationship marketing,” Journal of Marketing, 73, 1–18

Peng, L. and Fine, J. p. 1

Where the paper says it

The primary interests in analyzing the METIC data are often to evaluate pairwise associations between event times (nonterminal-terminal and between-nonterminal) and to estimate covariate effects on these associations, as well as individual event times, via regressions. Existing literature primarily focuses on data involving a single nonterminal event, commonly known as semi-competing risks (Fine et al.

P., “Regression modeling of semicompeting risks data,” Biometrics, 63, 96–108

Sklar, A., “Fonctions de r´epartition a n dimensions et leurs marges,” Publications de l’Institut Statistique de l’Universit´e de Paris, 8, 229–231

Su, L., Sengupta, J., Li, Y., and Chen, F. p. 1

Where the paper says it

, 2001; Peng and Fine, 2007; Hsieh et al. , 2008; Chen, 2012; Zhu et al.

(2023), “Want versus Need: How linguistic framing influences responses to crowdfunding appeals,” Journal of Consumer Research, 50, 923–944

Taylor, J. M. G., Murray, S., and Hsu, C.-H. (2002), “Survival estimation and testing via multiple imputation,” Statistics & Probability Letters, 58, 221–232

Wei, Y. M., Hong, J., and Tellis, G. J., “Machine learning for creativity: Using similarity networks to design better crowdfunding projects,” Journal of Marketing, 86, 87–104

Zeng, D., “Estimating marginal survival function by adjusting for dependent censoring using many covariates,” The Annals of Statistics, 32, 1533–1555

Zeng, D. and Lin, D. p. 4

Where the paper says it

, XJ,i, ∆J,i, Zi), i ∈ [n]}. We assume that there exists a maximum follow-up time t < ∞ and a constant p > 0 such that P(Xj > t|Z) > p for all j ∈ [J], and every element of Z is bounded with probability one (Zeng, 2004; Zeng and Lin, 2006; Li et al.

Y., “Efficient estimation of semiparametric transformation models for counting processes,” Biometrika, 93, 627–640

Zhang, J. and Liu, P., “Rational herding in microloan markets,” Management Science, 58, 892–912

Zhang, Y., Hua, L., and Huang, J., “A spline-based semiparametric maximum likelihood estimation method for the Cox model with interval-censored data,” Scandinavian Journal of Statistics, 37, 338–354

Zhao, X., Wu, Y., and Yin, G., “Sieve maximum likelihood estimation for a general class of accelerated hazards models with bundled parameters,” Bernoulli, 23, 3385–3411

Zhu, H., Lan, Y., Ning, J., and Shen, Y.

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. Direct estimation of joint density for multivariate event times under informative censoring can be achieved by leveraging vine copulas to decompose complex multivariate dependence into a catalog of pairwise dependencies organized through a sequence of trees (connected acyclic graphs). This allows the model to capture the heterogeneity in the association between each pair of event times, whether they are nonterminal-terminal or between-nonterminal.

  2. The proposed stage-wise pseudo maximum likelihood estimation procedure reduces computational cost compared to simultaneous maximum likelihood estimation by estimating parameters sequentially for each tree. This is described as being less efficient but more convenient to implement in practice, with lower computational cost than the MLE obtained by maximizing the full likelihood given in (7).

  3. The model allows for direct regression modeling of all marginals, including nonterminal and terminal event times, by specifying them using a flexible semiparametric transformation model (equation 5). This enables direct regression analyses on individual event times without detouring through the first event time survival functions.

  4. The framework can be adapted to avoid misspecification issues by allowing bivariate copulas to be specified non-parametrically and covariates to be modeled using semi-parametric specifications, such as the bundled sieve approach for marginals developed in Zhao et al. (2017).

  5. The system can perform a plug-in estimate for the nonterminal marginals and their corresponding nonterminal-terminal copula parameters when using the vine copula structure, which is achieved because "Only under Archimedean copula, Sj (t) can be explicitly expressed as a function of S∗j (t) = P(T∗j > t), SJ (t), and θ[2]."

Abstract

The analysis of times to multiple nonterminal events arises in many lifetime data settings, where the nonterminal event times are subject to informative censoring by a terminal event, posing substantial challenges. We propose a likelihood-based regression framework that models dependence among multiple nonterminal and terminal event times using a vine copula. This framework allows pairwise associations between nonterminal and terminal events, as well as among nonterminal events, to vary across event types, while enabling valid regression inference for the marginal distributions of individual event times. To facilitate computation, we develop a stage-wise pseudo-likelihood estimation procedure and establish its large-sample properties. We demonstrate satisfactory finite-sample performance of the proposed estimators and associated variance estimators through simulation studies. We also illustrate the proposed methodology using data from a crowdfunding platform to examine the relationship between different types of creator-backer interactions and a creator's lifetime on the platform.

Related papers