Markov Renewal Proportional Hazards is All You Need

Elvis Han Cui

REVIEW 4 major objections 5 minor 74 references

Markov Renewal Proportional Hazards is All You Need

T0 review · 4 major / 5 minor · reviewed 2026-08-10 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read Sojourn-aware multi-state models give smoother, later transition curves for transplant patients than memoryless Markov models.

desk verdict A readable tutorial re-derivation of the DSH estimator whose advertised smoothness claim rests only on eyeballing four plots; send it to review only if the claim is quantified or the paper is repositioned as a tutorial. read the letter →

arxiv 2502.03479 v6 pith:3JVAOPLM submitted 2025-01-27 stat.AP stat.CO

Elvis Han Cui This is my paper · ORCID

classification stat.APstat.CO MSC 60K15

keywords semi-MarkovmodelsMarkovrenewalprocessestransitionprobabilitiesDabrowska-Sun-HorowitzestimatorAalen-Johansenmulti-statesurvivalanalysisstem-celltransplantdata

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

This paper argues that in multi-state clinical modeling, recording and using the time spent in each state (the sojourn time) makes patient trajectories more realistic than a memoryless Markov model, and that the Dabrowska-Sun-Horowitz (DSH) estimator of transition probabilities, which is derived from the Markov renewal equation, produces smoother probability curves than the classical Aalen-Johansen (AJ) product-integral estimator. The demonstration is on a stem-cell transplant registry with six clinical states, where the DSH estimator smooths transitions into prolonged states such as relapse, and the semi-Markov model delays the accumulation of probability in absorbing states such as relapse or death. The paper also reorganizes the asymptotic theory for the semi-parametric Markov renewal Cox regression model, proving weak convergence of the regression coefficients and the baseline hazards with empirical process tools and the Burkholder-Davis-Gundy inequality. If the claims hold, clinical modelers get a more temporally sensitive framework and a concrete formula—$P(t) = R * G(t)$—for computing transition probabilities.

What carries the argument

The central object is the Markov renewal function and the renewal equation it solves. With the semi-Markov kernel $Q_{ij}(x)$ (probability that the next transition from $i$ goes to $j$ within sojourn $x$), the diagonal state-survival matrix $G$, and the matrix convolution $Q * P$, the transition probability matrix satisfies $P(t) = G(t) + Q * P(t)$; the unique solution is $P(t) = R * G(t)$, where $R = \sum_{p=0}^{\infty} Q^{(p)}$ is the Markov renewal function. Plugging in estimators for $Q$ and $G$ yields the DSH estimator, and $R$ is truncated at a few convolution terms in practice. A second load-bearing component is the Cox-type intensity model $\alpha_{ij}(x) = \alpha_{0ij}(x) e^{\beta^T Z_{ij}}$ with its profile likelihood; the asymptotic results for $\beta$ and the baseline hazard are carried by empirical process theory and the Burkholder-Davis-Gundy inequality.

What would settle it

Compute the DSH transition probabilities for the same transplant cohort with the renewal series $R = \sum_p Q^{(p)}$ extended to ten or twenty convolutions and with bootstrap standard errors; if the probability curves move by more than the width of their confidence bands, the reported smoothness and delayed absorption are truncation artifacts.

Watch

Extended reading notes

Core claim

The paper's central claim is that transition probabilities in multi-state clinical processes should be modeled as a Markov renewal process, because the distribution of sojourn times carries information that memoryless Markov models discard. On the transplant data, the DSH estimator—obtained by plugging nonparametric estimates of the semi-Markov kernel (the probability that the next transition from state $i$ goes to $j$ within sojourn time $x$) into the renewal solution—is consistently smoother than the AJ estimator, with the difference most visible in states with long dwell times, and the semi-Markov models shift probability out of absorbing states such as relapse and death toward later times. The paper also states and proves weak convergence of the Cox-regression parameter and baseline hazard in this Markov renewal model, and gives a bootstrap scheme for confidence bands.

Load-bearing premise

The comparison depends on assuming that truncating the repeated-convolution series $R = \sum_{p} Q^{(p)}$ after three or four terms leaves an error too small to change the curves; the paper says this error is still to be quantified.

Editorial extensions

If this is right

If sojourn time matters as claimed, Markov-based analyses of transplant data will tend to overstate how quickly patients reach relapse or death; semi-Markov models move those probabilities later.
The DSH estimator gives smoother curves for states with prolonged sojourns, such as recovery with adverse events and relapse, which is useful for sparse low-frequency transitions.
The stated weak-convergence results justify confidence intervals and bootstrap confidence bands for the Markov renewal Cox model, following the paper's bootstrap scheme.
Lemma 3.2 provides a practical formula for prediction probabilities $P(s,t)$ from an arbitrary calendar time $s$, supporting dynamic prediction of patient trajectories.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

Editorial inference: The smoothness of DSH may be partly a convolution artifact rather than evidence for semi-Markov structure; one could test by simulating a true memoryless Markov process and checking whether DSH still appears smoother.
Editorial inference: A formal bound on the truncation error of the renewal series $R = \sum_p Q^{(p)}$ is missing; a total-variation or spectral-radius argument could turn the practical stopping rule into a provable approximation.
Editorial inference: The progressive-state assumption excludes back transitions such as relapse followed by recovery; extending the method to recurrent states would require another truncation strategy and is not covered by the current comparison.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Here's my take: this is a tutorial-style application of the Dabrowska-Sun-Horowitz estimator to the EBMT data. The genuinely new parts are a simulation, an application, and a reorganized proof of asymptotic normality; the estimator, the asymptotics, and the prediction formula come from prior work (Dabrowska, Sun and Horowitz; Spitoni et al.). The paper is honest about that, which I respect.

What it does well: the proof reorganization in Appendix 8.5 is genuinely more readable than the original sources; Lemma 3.2 gives a clean reformulation of the prediction-probability formula; Table 1 is a useful comparison of semi-Markov versus Markov machinery. The paper also states plainly, in Section 5.2, that the smoothness claim 'lacks a formal theoretical justification.' That is the right kind of honesty.

The soft spots are where the abstract overreaches. The central claim—DSH consistently yields smoother curves—is supported only by visual inspection of Figures 6-9. There is no quantitative smoothness metric, no confidence bands, no bootstrap intervals, and no code. For a paper whose stated aim is to address the lack of computational tools, releasing no code is a real gap. The stress-test note about infinite convolution truncation actually does not land: the EBMT model is progressive with maximum path length 3, so the renewal sum is finite and 'no change after three or four convolutions' is an exact zero tail, not a truncation error. The load-bearing issue is upstream: the smoothness comparison is qualitative. There is also an unresolved ambiguity: 'Markov model with the DSH estimator' is not a standard estimator of a calendar-time Markov process unless the Markov model is assumed time-homogeneous, and the paper never states that. So the four-way comparison is not as clean as presented.

Who is this for? A reader who wants a self-contained walkthrough of Markov renewal estimation and a compact restatement of the DSH asymptotics will get value. A reader looking for a new method or a validated empirical finding will not. It is honest, coherent, and not wrong on its own terms, but the advertised conclusion is unsupported.

Recommendation: I would send it to a serious referee rather than desk-reject—it is a coherent tutorial with a useful proof reorganization, and an editor can reasonably ask for the smoothness claim to be quantified and the code released. But I would not accept the empirical claim as stated, and I would cite it only as a tutorial, not as a source of new results.

Referee Report

4 major / 5 minor

Summary. The paper develops and applies a Markov renewal proportional hazards model with the Dabrowska-Sun-Horowitz (DSH) estimator for transition probabilities, and compares it with Aalen-Johansen (AJ) estimation under Markov and semi-Markov assumptions. The methodological core is a review and reorganization of existing asymptotic theory, followed by a simulation study and an application to the EBMT hematopoietic stem cell transplantation dataset. The central empirical claim is that the DSH estimator consistently yields smoother transition probability curves than AJ, and that semi-Markov models delay absorption into absorbing states relative to Markov models. The paper also presents an appendix reorganizing the weak convergence proof for the regression coefficients and baseline cumulative hazards.

Significance. If the empirical claim were quantitatively supported, the paper would provide useful practical guidance for clinical multi-state modeling and would make a case for preferring semi-Markov models with the DSH estimator. The paper has clear strengths: it gives a comprehensive and readable review of Markov renewal theory, a transparent plug-in estimation framework, an explicit simulation algorithm, and a real-data application in a clinically important setting. It also candidly acknowledges limitations, including the lack of formal justification for the smoothing effect and the unquantified truncation error. However, the paper ships no code or machine-checked proofs, and its main advertised finding rests on visual inspection. The asymptotic results are largely borrowed from prior work by Dabrowska, Sun and Horowitz, Spitoni et al., and others, so the novel contribution is primarily expository and empirical rather than theoretical.

major comments (4)

[Abstract; §6.3, Figs. 6–9] The central claim that "the DSH estimator consistently yields smoother probability curves" is not supported by any quantitative evidence. Section 6.3 reports only qualitative observations about Figures 6–9, and Section 5.2 explicitly concedes that the smoothing effect "lacks a formal theoretical justification." To make the claim reproducible, the authors should define a smoothness metric (e.g., total variation, number of monotonicity changes, integrated squared second difference), report its value for each estimator, and provide pointwise confidence bands or bootstrap intervals for the transition probabilities. Without this, the apparent smoothness could be an artifact of plotting choices, different time grids, or visual perception.
[§6.3, Eqs. (10)–(15)] The four modeling approaches listed in Section 6.3 include "a Markov model with the DSH estimator" and "a Semi-Markov model with the AJ estimator," but these combinations are not well defined. The DSH estimator in Eq. (15) targets transition probabilities of a Markov renewal/semi-Markov process, while the AJ estimator targets a Markov process. Applying DSH under a Markov assumption, or AJ under a semi-Markov assumption, is a model misspecification unless the intended data-generating model and the target estimand are explicitly stated. As written, the comparison confounds the choice of estimator with the choice of model, so the observed differences in smoothness and absorption timing cannot be attributed to the estimator alone.
[§8.2, Eq. (15)] The paper admits that the infinite renewal convolution bR = sum_p bQ^(p) is truncated after three or four convolutions and that "theoretically, the error is still to be estimated." For the EBMT application, the state graph is progressive and the maximum path length is at most 3, so the tail is exactly zero there; the reader's truncation concern does not land for that dataset. However, the paper presents the truncation rule as a general computational strategy for Markov renewal processes without restricting to progressive graphs. Since recurrent states make the convolution sum genuinely infinite, the authors should either prove a finite-support condition, provide an explicit error bound, or clearly limit the method to progressive processes.
[§8.5, Lemma 8.5 and Theorem 4.3] The proof of Theorem 4.2 relies on Lemma 8.5, but the fourth-moment bound is not established. The function f(n,a)=4n^3-(6+12a)n^2+(4-12a+12a^2)n is claimed to be nonnegative after minimizing at a=(n+1)/2, yielding 4n^3-12n^2+n; this quantity is negative for small n (e.g., n=1 gives -7). The argument covers only n≥12 and leaves small jump counts untreated. In addition, the proof of Theorem 4.3 is only a statement that Hadamard differentiability "can be verified," with no detailed verification. Since the abstract claims to show weak convergence of the estimator, the manuscript should either complete the missing cases and verification, or explicitly identify Theorem 4.3 as a cited result from Dabrowska (1995) and Spitoni et al. (2012).

minor comments (5)

[§8.3.1] The expression bP(s,t)=bP(0,t)/bP(0,s) is mathematically imprecise because transition probability matrices do not commute; the correct Markov expression is P(0,t)P(0,s)^{-1} or, better, the product integral over (s,t]. Please revise the notation to avoid ambiguity.
[Table 3, §6.3] Only the semi-Markov DSH model's coefficient estimates are reported. For a fair comparison of the four approaches, the authors should report coefficient estimates and standard errors for all fitted models, or at least state explicitly why only one model's coefficients are shown.
[Title and abstract] The title "Markov Renewal Proportional Hazards is All You Need" and the abstract's language "we demonstrate" overstate the evidence, given the paper's own admissions in §5.2 and §8.2 that the smoothing effect has no formal justification and the truncation error is unquantified. More cautious wording would better match the manuscript's contributions.
[§8.2] Since the EBMT state graph is progressive and has maximum path length 3, the statement "after three or four convolutions there is no change in the estimate" is exact for that dataset, not a truncation approximation. The authors should say this explicitly to prevent readers from misreading the statement as a general convergence claim.
[References and Figure 5] Several bibliographic entries contain formatting issues (e.g., "V oelkel" in [70], incomplete package references in Table 5), and the Sankey diagram in Figure 5 has small, hard-to-read labels. Please proofread the reference list and provide a higher-resolution vector figure.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity: the DSH estimator, its asymptotic theory, and the EBMT comparison are all imported from external prior work or standard estimating equations; the paper's empirical 'smoother curves' claim is weakly supported but is not a fitted parameter renamed as a prediction.

full rationale

The paper's derivation chain is not circular. The transition probability formula bP(t) = bR * bG(t) (Equation 15) is a plug-in version of the standard Markov renewal equation P(t) = G(t) + Q * P(t) (Equation 7), whose unique solution P(t) = R * G(t) is cited to Cinlar's external textbook (Theorem 2.10, 'The proof is omitted and can be found in page 324 in Cinlar [20]'). The DSH estimator is explicitly introduced as the plug-in estimator of that known renewal solution, not as a quantity defined in terms of the paper's own target claim. The regression coefficient estimator is obtained from the standard profile log-likelihood estimating equation of Dabrowska, Sun and Horowitz [31] and Sun [62], and the weak convergence results in Theorem 4.2 and Proposition 1 are re-statements or reorganizations of external results by Sun, Dabrowska, and Gill; the paper's own Appendix proof is an exposition of those prior arguments rather than a self-referential deduction. The only self-citations are Cui [22,23], which appear in Section 2 solely as tutorial references for readers and are not used in any proof or estimation step. No uniqueness claim is imported from the present authors: the uniqueness of the renewal solution is Cinlar's external theorem. The abstract's headline claim that 'the DSH estimator consistently yields smoother probability curves' is an empirical, visually based comparison of Figures 6-9; this is weakly supported and lacks a formal smoothness metric, but it is not circular because the DSH curves are not constructed from the AJ curves or from a smoothness objective. Section 8.2's remark that after three or four convolutions 'there is no change in the estimate' and that 'theoretically, the error is still to be estimated' is a conservative numerical remark, not a circular step; for the progressive EBMT graph the maximum path length is at most three transitions, so higher-order convolutions Q^(p) are exactly zero. In summary, the paper's central estimator and asymptotics reduce to standard external semi-Markov theory, and the applied comparison is an estimation exercise on a public dataset, so no load-bearing step reduces to its own inputs.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The paper relies on standard Markov renewal assumptions and nonparametric estimation; no new entities are introduced. The unquantified truncation of the renewal convolution is the most fragile computational assumption, and the fitted baseline hazards are never shown.

free parameters (3)

Regression coefficients beta_ij = Table 3: agecl20-40 0.108, agecl>40 0.244, proph -0.177, gender_mismatch -0.018
Estimated from the EBMT data via the profile likelihood in Section 3.2; these fitted values drive the transition probability curves and the reported comparison.
Baseline cumulative hazards A0ij = not reported for EBMT fits
Estimated nonparametrically by the Breslow-type estimator; they enter the semi-Markov kernel and the DSH transition probabilities. The fitted functions are never displayed, so the reader cannot assess their role.
Truncation order for infinite renewal convolution = 3 or 4
Section 8.2 fixes the number of convolution terms as the point where estimates stop changing, but no criterion or error bound is given. This choice affects the reported transition probability curves.

assumptions (6)

domain assumption Each subject's observed trajectory is generated by a Markov renewal process on the finite state space (Definition 2.1).
The EBMT analysis and all estimators are justified only under this process assumption.
domain assumption Censoring is governed by a predictable process K(t) with C_n in [T_n, T_{n+1}] (Definition 2.5).
This random censorship setup is needed for the counting process estimators and asymptotic claims.
standard math Transition intensities A_ij are absolutely continuous (Definition 2.6).
Used to relate survival G_i to product integrals and to define the likelihood; assumes no atom in the transition measures.
domain assumption The process is progressive (states not revisited), so the renewal convolution has finite terms (Section 3.3).
The paper notes non-progressive processes are computationally intractable; the EBMT analysis silently relies on this.
domain assumption Regularity conditions in Appendix 8.5.1: bounded covariates, positive definite information matrix, finite baseline hazard, Glivenko-Cantelli conditions.
Required for the weak convergence results in Theorem 4.2 and Lemma 8.3; not checked on EBMT data.
ad hoc to paper The infinite renewal sum can be truncated after three or four convolutions without meaningful error.
Section 8.2 asserts this based on the observation that estimates stop changing, but explicitly says the error remains to be estimated. The reported curves depend on this untested approximation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Markov Renewal Proportional Hazards is All You Need." pith.science (2026). https://pith.science/paper/3JVAOPLM

@misc{pith2026250203479,
  author       = {Pith},
  title        = {Pith review of: Markov Renewal Proportional Hazards is All You Need},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3JVAOPLM}},
  note         = {Machine review of arXiv:2502.03479}
}

read the original abstract

Transition probability estimation plays a critical role in multi-state modeling, especially in clinical research. This paper investigates the application of semi-Markov and Markov renewal frameworks to the EBMT dataset, focusing on six clinical states encountered during hematopoietic stem cell transplantation. By comparing Aalen-Johansen (AJ) and Dabrowska-Sun-Horowitz (DSH) estimators, we demonstrate that semi-Markov models, which incorporate sojourn times, provide a more nuanced and temporally sensitive depiction of patient trajectories compared to memoryless Markov models. The DSH estimator consistently yields smoother probability curves, particularly for transitions involving prolonged states. We use empirical process theory and Burkholder-Davis-Gundy inequality to show weak convergence of the estimator. Future work includes extending the framework to accommodate advanced covariate structures and non-Markovian dynamics.

Figures

Figures reproduced from arXiv: 2502.03479 by the authors.

**Figure 2.** Prediction probabilities under a Markov model with the Aalen-Johansen estimator [PITH_FULL_IMAGE:figures/full_fig_p016_2.png] view at source ↗

**Figure 4.** State transition diagram illustrating the possible transitions between clinical states in [PITH_FULL_IMAGE:figures/full_fig_p017_4.png] view at source ↗

Figures from the paper (3 more)

**Figure 5.** Figure 5: Enhanced Sankey diagram illustrating patient transitions across clinical states in the [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

**Figure 6.** Figure 6: Transition probabilities under a Markov model with the Aalen-Johansen estimator. 0% 20% 40% 0 2000 4000 6000 Calendar Time (Days) Probability State AE Death PLT Rec & AE Relapse TX Cox Regression in A Markov Model with Dabrowska−Sun−Horowitz Estimator Transition Probab…

**Figure 8.** Figure 8: Transition probabilities under a Semi-Markov model with the AalenJohansen estimator. 0% 20% 40% 0 2000 4000 6000 Calendar Time (Days) Probability State AE Death PLT Rec & AE Relapse TX Cox Regression in A Semi−Markov Model with Dabrowska−Sun−Horowitz Estimator Transit…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

74 extracted references · 74 canonical work pages

[1]

A ALEN , O. O. (1975). Statistical inference for a family of counting processes. University of California, Berkeley

work page 1975
[2]

A ALEN , O. (1978). Nonparametric inference for a family of counting processes. The Annals of Statistics 701–726

work page 1978
[3]

and S CHUMACHER , M

A LLIGNOL , A., B EYERSMANN , J. and S CHUMACHER , M. (2008). Mvna: an R package for the Nelson– Aalen estimator in multistate models. The Newsletter of the R Project Volume 8/2, October 2008 8 48

work page 2008
[4]

K., E SBJERG , S

A NDERSEN , P. K., E SBJERG , S. and S ØRENSEN , T. I. (2000). Multi-state models for bleeding episodes and mortality in liver cirrhosis. Statistics in medicine 19 587–599

work page 2000
[5]

A NDERSEN , P. K. and G ILL , R. D. (1982). Cox’s regression model for counting processes: a large sample study. The annals of statistics 1100–1120

work page 1982
[6]

A NDERSEN , P. K. and K EIDING , N. (2002). Multi-state models for event history analysis. Statistical meth- ods in medical research 11 91–115

work page 2002
[7]

A NDERSEN , P. K. and R AVN, H. (2023). Models for multi-state survival data: rates, risks, and pseudo- values. CRC Press

work page 2023
[8]

K., B ORGAN , O., G ILL , R

A NDERSEN , P. K., B ORGAN , O., G ILL , R. D. and K EIDING , N. (2012). Statistical models based on counting processes. Springer Science & Business Media

work page 2012

Show all 74 references

[9]

A RALIS , H. J. (2016). Modeling multistate processes with back transitions: Statistical challenges and ap- plications. University of California, Los Angeles

2016
[10]

and B ROOKMEYER , R

A RALIS , H. and B ROOKMEYER , R. (2019). A stochastic estimation procedure for intermittently-observed semi-Markov multistate models with back transitions.Statistical methods in medical research28 770– 787

2019
[11]

and R OCA -PARDIÑAS , J

A RAÚJO , A., M EIRA -MACHADO , L. and R OCA -PARDIÑAS , J. (2015). TPmsm: Estimation of the transi- tion probabilities in 3-state models. Journal of Statistical Software 62 1–29

2015
[12]

and N AZARATHY , Y

A SANJARANI , A., L IQUET , B. and N AZARATHY , Y. (2022). Estimation of semi-Markov multi-state mod- els: a comparison of the sojourn times and transition intensities approaches. The international journal of biostatistics 18 243–262

2022
[13]

B ILLINGSLEY , P. (2013). Convergence of probability measures. John Wiley & Sons

2013
[14]

and F URRER , C

B LADT , M. and F URRER , C. (2023). Conditional Aalen–Johansen estimation. arXiv preprint arXiv:2303.02119

2023 arXiv
[15]

and A BDALLA , N

B ROOKMEYER , R. and A BDALLA , N. (2019). Multistate models and lifetime risk estimation: Application to Alzheimer’s disease. Statistics in medicine 38 1558–1565

2019
[16]

B ROSTRÖM , G. (2024). eha: Event History Analysis R package version 2.10.2

2024
[17]

B ÜHLER , A., C OOK , R. J. and L AWLESS , J. F. (2023). Multistate models as a framework for estimand specification in clinical trials of complex processes. Statistics in Medicine 42 1368–1397

2023
[18]

Ç INLAR , E. (1969). On semi-Markov processes on arbitrary spaces. In Mathematical Proceedings of the Cambridge Philosophical Society 66 381–392. Cambridge University Press

1969
[19]

C INLAR , E. (1969). Markov renewal theory. Advances in Applied Probability 1 123–187

1969
[20]

C INLAR , E. (2013). Introduction to stochastic processes. Courier Corporation

2013
[21]

C OOK , R. J. and L AWLESS , J. F. (2018). Multistate models for the analysis of life history data . Chapman and Hall/CRC. 32

2018
[22]

C UI, E. H. (2022). A Tutorial on Statistical Models Based on Counting Processes. arXiv preprint arXiv:2210.07114

2022 arXiv
[23]

C UI, E. H. (2024). A Tutorial on Brownian Motion for Biostatisticians. arXiv preprint arXiv:2408.16011

2024 arXiv
[24]

D ABROWSKA , D. (1995). Estimation of transition probabilities and bootstrap in a semiparametric Markov renewal model. Journal of Nonparametric Statistics 5 237–259

1995
[25]

D ABROWSKA , D. M. (2006). Estimation in a class of semiparametric transformation models.Lecture Notes- Monograph Series 131–169

2006
[26]

D ABROWSKA , D. M. (2009). Estimation in a semi-parametric two-stage renewal regression model. Statis- tica Sinica 19 981

2009
[27]

D ABROWSKA , D. M. (2012). Estimation in a semi-Markov transformation model.The International Journal of Biostatistics 8

2012
[28]

D ABROWSKA , D. M. (2014). Multivariate Survival Analysis. Wiley StatsRef: Statistics Reference Online 1–12

2014
[29]

D ABROWSKA , D. M. and H O, W. T. (2006). Estimation in a semiparametric modulated renewal process. Statistica Sinica 93–119

2006
[30]

D ABROWSKA , D. M. and L EE, W. (1996). Nonparametric estimation of transition probabilities in a two- stage duration model. Journal of Nonparametric Statistics 7 75–103

1996
[31]

M., S UN, G.- W

D ABROWSKA , D. M., S UN, G.- W. and H OROWITZ , M. M. (1994). Cox regression in a Markov renewal model: an application to the analysis of bone marrow transplant data. Journal of the American Statis- tical Association 89 867–877

1994
[32]

and S ATTEN , G

D ATTA, S. and S ATTEN , G. A. (2002). Estimation of integrated transition hazards and stage occupation probabilities for non-Markov systems under dependent censoring. Biometrics 58 792–802

2002
[33]

C., F IOCCO , M

DE WREEDE , L. C., F IOCCO , M. and P UTTER , H. (2011). mstate: an R package for the analysis of com- peting risks and multi-state models. Journal of statistical software 38 1–30

2011
[34]

G ILL , R. D. (1980). Nonparametric estimation based on censored observations of a Markov renewal pro- cess. Zeitschrift für Wahrscheinlichkeitstheorie und verwandte Gebiete53 97–116

1980
[35]

G ILL , R. (1981). Testing with replacement and the product limit estimator. The Annals of Statistics 9 853– 860

1981
[36]

G ILL , R. D. and J OHANSEN , S. (1990). A survey of product-integration with a view toward application in survival analysis. The annals of statistics 1501–1555

1990
[37]

D., W ELLNER , J

G ILL , R. D., W ELLNER , J. A. and P RÆSTGAARD , J. (1989). Non-and semi-parametric maximum likeli- hood estimators and the von mises method (part 1)[with discussion and reply]. Scandinavian Journal of Statistics 97–128

1989
[38]

and G RAY, M

G RAY, B. and G RAY, M. B. (2014). Package ‘cmprsk’. Subdistribution analysis of competing risks. R package version 2 2–7

2014
[39]

H OFF, R., P UTTER , H., M EHLUM , I. S. and G RAN , J. M. (2019). Landmark estimation of transition probabilities in non-Markov multi-state models with covariates. Lifetime data analysis 25 660–680

2019
[40]

H OUGAARD , P. (1999). Multi-state models: a review. Lifetime data analysis 5 239–264

1999
[41]

J ACKSON , C. (2011). Multi-state models for panel data: the msm package for R. Journal of statistical software 38 1–28

2011
[42]

J ACKSON , C. H. (2016). flexsurv: a platform for parametric survival modeling in R. Journal of statistical software 70

2016
[43]

J OHANSEN , S. (1983). An extension of Cox’s regression model. International Statistical Review/Revue Internationale de Statistique 165–174

1983
[44]

K ASSAMBARA , A. et al. (2023). survminer: An R package for survival analysis and visualization. The Comprehensive R Archive Network (CRAN)

2023
[45]

K RUGER , F. et al. (2015). SemiMarkov: An R package for parametric estimation in multi-state semi- Markov models. The R Journal 7 1–10

2015
[46]

L UCAS , A., M ONTEIRO , A. A. and S MIRNOV , G. V. (2006). Nonparametric Estimation for Non- Homogeneous Semi-Markov Processes: An Application to Credit Risk Discussion Paper No. 06-024/2, Tinbergen Institute. 43 pages

2006
[47]

and G RAN , J

M ALTZAHN , N., H OFF, R., A ALEN , O., M EHLUM , I., P UTTER , H. and G RAN , J. (2020). A hybrid land- mark Aalen-Johansen estimator for transition probabilities in partially non-Markov multi-state models. arXiv preprint arXiv:2007.00974

2020 arXiv
[48]

and A NDERSEN , P

M EIRA -M ACHADO , L., DE UÑA-ÁLVAREZ, J., C ADARSO -S UÁREZ , C. and A NDERSEN , P. K. (2009). Multi-state models for the analysis of time-to-event data. Statistical methods in medical research 18 195–222

2009
[49]

M OORE , E. H. and P YKE , R. (1968). Estimation of the transition distributions of a Markov renewal process. Annals of the Institute of Statistical Mathematics 20 411–424. MARKOV RENEW AL PROPORTIONAL HAZARDS MODEL 33

1968
[50]

J., M ATEJA, A., F OLLMANN , D

M ORSOMME , R., L IANG , C. J., M ATEJA, A., F OLLMANN , D. A., O’B RIEN , M. P., W ANG , C. and FINTZI , J. (2025). Assessing treatment efficacy for interval-censored endpoints using multistate semi- Markov models fit to multiple data streams. arXiv preprint arXiv:2501.14097

2025 arXiv
[51]

P UTTER , H. (2011). Tutorial in biostatistics: Competing risks and multi-state models Analyses using the mstate package. Companion file for the mstate package 145

2011
[52]

and G ESKUS , R

P UTTER , H., F IOCCO , M. and G ESKUS , R. B. (2007). Tutorial in biostatistics: competing risks and multi- state models. Statistics in medicine 26 2389–2430

2007
[53]

P YKE , R. (1961a). Markov renewal processes with finitely many states.The Annals of Mathematical Statis- tics 1243–1259

1961
[54]

P YKE , R. (1961b). Markov renewal processes: definitions and preliminary properties. The Annals of Math- ematical Statistics 1231–1242

1961
[55]

P., M CGRATH, E., M IDDELKOOP , B., ADAMS , G., A TLIJA , M., A YUK , F

S ACCARDI , R., P UTTER , H., E IKEMA , D.-J., B USTO , M. P., M CGRATH, E., M IDDELKOOP , B., ADAMS , G., A TLIJA , M., A YUK , F. A., B ALDOMERO , H. et al. (2023). Benchmarking of survival outcomes following Haematopoietic Stem Cell Transplantation (HSCT): an update of the ...

2023
[56]

S ATTEN , G. A. and S TERNBERG , M. R. (1999). Fitting semi-Markov models to interval-censored data with unknown initiation times. Biometrics 55 507–513

1999
[57]

S CHEIKE , T. H. and Z HANG , M.-J. (2011). Analyzing competing risk data using the R timereg package. Journal of statistical software 38

2011
[58]

S HORACK , G. R. and W ELLNER , J. A. (2009). Empirical processes with applications to statistics. SIAM

2009
[59]

S HU, Y., KLEIN , J. P. and ZHANG , M.-J. (2007). Asymptotic theory for the Cox semi-Markov illness-death model. Lifetime Data Analysis 13 91–117

2007
[60]

S PEZIA , L. et al. (2022). An R package for hidden hybrid Markov/semi-Markov models. The R Journal 14 1–20

2022
[61]

and P UTTER , H

S PITONI , C., V ERDUIJN , M. and P UTTER , H. (2012). Estimation and asymptotic theory for transition probabilities in Markov renewal multi-state models. The International Journal of Biostatistics 8

2012
[62]

S UN, G.-W. (1992). Markov renewal proportional hazards model for longitudinal survival data. University of California, Los Angeles

1992
[63]

T AGA, Y. (1963). On the limiting distributions in Markov renewal processes with finitely many states. Annals of the Institute of Statistical Mathematics 15 1–10

1963
[64]

T HERNEAU , T. (2015). A package for survival analysis in S. R package version 2 2014

2015
[65]

T HERNEAU , T. M. (2020). survival: A package for survival analysis. R Foundation

2020
[66]

T HERNEAU , T. M. and O U, F.-S. (2024). Using multistate models with clinical trial data for a deeper understanding of complex disease processes. Clinical Trials 21 531–540

2024
[67]

T ITMAN , A. C. and S HARPLES , L. D. (2010). Semi-Markov models with phase-type sojourn distributions. Biometrics 66 742–752

2010
[68]

V AART, A. V. D. and W ELLNER , J. A. (2023). Empirical processes. In Weak Convergence and Empirical Processes: With Applications to Statistics 127–384. Springer

2023
[69]

K., M OLINA , K

V ASQUEZ , J. K., M OLINA , K. C., T OMAZELLA , V., DINIZ , C. A. and S UZUKI , A. K. (2024). Multistate models with nested frailty for lifetime analysis: Application to bone marrow transplantation recovery patients. Communications in Statistics-Theory and Methods 1–19

2024
[70]

V OELKEL , J. G. and C ROWLEY, J. (1984). Nonparametric inference for a class of semi-Markov processes with censored observations. The Annals of Statistics 142–160

1984
[71]

W EISS , G. H. and Z ELEN , M. (1965). A semi-Markov model for clinical trials. Journal of Applied Proba- bility 2 269–285

1965
[72]

and E NGELHARDT , B

W ILKINSON , D. and E NGELHARDT , B. E. (2017). An R package for estimating the Markov chain from discrete-time data. Journal of Statistical Software 81 1–17

2017
[73]

and N AIR , V

Y ANG , Y. and N AIR , V. N. (2011). Parametric inference for time-to-failure in multi-state semi-Markov models: A comparison of marginal and process approaches. Canadian Journal of Statistics 39 537– 555

2011
[74]

and I ANNACCHIONE , V

Z HAO, L., Z HANG , Y. and I ANNACCHIONE , V. G. (2012). msSurv: An R package for nonparametric estimation of multistate models. Journal of Statistical Software 50 1–28

2012

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.

[1] [1]

A ALEN , O. O. (1975). Statistical inference for a family of counting processes. University of California, Berkeley

work page 1975

[2] [2]

A ALEN , O. (1978). Nonparametric inference for a family of counting processes. The Annals of Statistics 701–726

work page 1978

[3] [3]

and S CHUMACHER , M

A LLIGNOL , A., B EYERSMANN , J. and S CHUMACHER , M. (2008). Mvna: an R package for the Nelson– Aalen estimator in multistate models. The Newsletter of the R Project Volume 8/2, October 2008 8 48

work page 2008

[4] [4]

K., E SBJERG , S

A NDERSEN , P. K., E SBJERG , S. and S ØRENSEN , T. I. (2000). Multi-state models for bleeding episodes and mortality in liver cirrhosis. Statistics in medicine 19 587–599

work page 2000

[5] [5]

A NDERSEN , P. K. and G ILL , R. D. (1982). Cox’s regression model for counting processes: a large sample study. The annals of statistics 1100–1120

work page 1982

[6] [6]

A NDERSEN , P. K. and K EIDING , N. (2002). Multi-state models for event history analysis. Statistical meth- ods in medical research 11 91–115

work page 2002

[7] [7]

A NDERSEN , P. K. and R AVN, H. (2023). Models for multi-state survival data: rates, risks, and pseudo- values. CRC Press

work page 2023

[8] [8]

K., B ORGAN , O., G ILL , R

A NDERSEN , P. K., B ORGAN , O., G ILL , R. D. and K EIDING , N. (2012). Statistical models based on counting processes. Springer Science & Business Media

work page 2012