pith. sign in

arxiv: 2607.01476 · v1 · pith:Y2VSXC64new · submitted 2026-07-01 · 📊 stat.AP · stat.ME

Sensitivity Analysis and Optimization of Stochastic Epidemic Models under Parameter Uncertainty

Nicholas R. Wu , Michael C. Fu This is my paper

Pith reviewed 2026-07-03 00:57 UTC · model grok-4.3

classification 📊 stat.AP stat.ME
keywords stochastic epidemic modelsensitivity analysisgradient estimatorsparameter uncertaintyvaccination optimizationherd immunityBayesian calibrationdiscrete-time simulation
0
0 comments X

The pith

Unbiased gradient estimators show that parameter uncertainty reduces sensitivity in stochastic epidemic models and produces more conservative optimization policies.

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

The paper derives unbiased gradient estimators for analyzing how total infections respond to changes in the proportion immunized and the contact rate when those parameters are uncertain and represented as distributions. These estimators are applied to a discrete-time stochastic epidemic model and compared against deterministic large-population approximations that ignore finite-size effects. Simulation results establish that incorporating parameter uncertainty lowers estimated sensitivities, weakens indirect effects such as herd immunity, and shifts cost-balancing optimization toward more cautious intervention levels than when parameters are treated as known.

Core claim

The authors derive unbiased gradient estimators that accommodate uncertainties represented as distributions over the parameters of interest, such as those arising from Bayesian calibration. Specifically, they estimate the sensitivity of total infections over a finite time horizon with respect to the proportion immunized (v) and the contact rate (β). Comparing the proposed estimators with deterministic limit approximations based on large populations reveals differences due to the finite population and time horizon. The estimators exhibit lower variance than finite-difference estimators for the derivative with respect to β, but higher variance for the derivative with respect to v. Simulation e

What carries the argument

Unbiased gradient estimators that accommodate distributions over model parameters for sensitivity of total infections to immunization proportion and contact rate.

If this is right

  • Sensitivity estimates differ from those obtained via deterministic large-population limits because of finite population size and finite time horizon.
  • The estimators have lower variance than finite-difference methods when differentiating with respect to the contact rate but higher variance when differentiating with respect to the immunization proportion.
  • Indirect effects such as herd immunity appear less pronounced once parameter uncertainty is included.
  • Optimization that balances intervention costs against infection costs yields more conservative policies under parameter uncertainty than under point estimates.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Real-world vaccination planning that ignores parameter uncertainty may overestimate the strength of indirect protection and therefore adopt insufficiently cautious targets.
  • The same estimator construction could be tested on continuous-time epidemic models or on networks with heterogeneous contact patterns to check whether the reduction in sensitivity persists.
  • If Bayesian calibration produces wider parameter distributions than those used in the simulations, the shift toward conservative policies would be even stronger.

Load-bearing premise

The discrete-time stochastic epidemic model is correctly specified and the parameter distributions accurately capture the uncertainty arising from Bayesian calibration.

What would settle it

A controlled simulation on a known stochastic epidemic process with specified parameter distributions in which the proposed gradient estimators exhibit bias, or empirical epidemic data in which incorporating parameter distributions fails to produce more conservative optimization outcomes than fixed-parameter analysis.

Figures

Figures reproduced from arXiv: 2607.01476 by Michael C. Fu, Nicholas R. Wu.

Figure 1
Figure 1. Figure 1: Contour plots of the derivative-based sensitivity from different meth￾ods vs. the parameters β and v. The top row shows the estimated sensitivities of the total infections with respect to v, and the bottom row shows the estimated sensitivities of the total infections with respect to β. From left to right, we show the sensitivities from the final size equations; the derivatives Jv and Jβ; and the expected s… view at source ↗
Figure 2
Figure 2. Figure 2: Solutions to optimization problems as functions of intervention costs, with known vs. uncertain parameters, estimated via stochastic approximation. The left panel shows the solution v ∗ to the vaccination problem (10) as a function of the cost parameter cv. The right panel shows the solution β ∗ to the contact problem (11) as a function of the cost parameter cβ. In all plots, the shaded region is the 95% C… view at source ↗
Figure 3
Figure 3. Figure 3: Critical unit costs C ∗ v and C ∗ β , as a function of β and v, respectively. The left plot shows estimates of C ∗ v , corresponding to the vaccination problem. The right plot shows estimates of C ∗ β , corresponding to the contact problem. The blue lines are the critical unit costs without parameter uncertainty, and the orange lines are the critical unit costs under parameter uncertainty, both estimated f… view at source ↗
read the original abstract

To address sensitivity analysis and optimization for a discrete-time stochastic epidemic model, we derive unbiased gradient estimators that accommodate uncertainties represented as distributions over the parameters of interest, such as those arising from Bayesian calibration. Specifically, we estimate the sensitivity of total infections over a finite time horizon with respect to the proportion immunized ($v$) and the contact rate ($\beta$). Comparing the proposed estimators with deterministic limit approximations based on large populations reveals differences due to the finite population and time horizon. The estimators exhibit lower variance than finite-difference estimators for the derivative with respect to $\beta$, but higher variance for the derivative with respect to $v$. Simulation experiments indicate parameter uncertainty reduces sensitivity to the parameters of interest. In particular, indirect effects of vaccination, such as herd immunity, are less pronounced compared to when parameters are known. For optimization problems balancing intervention and infection costs, incorporating parametric uncertainty leads to more conservative policies.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript claims to derive unbiased gradient estimators for sensitivity analysis of total infections over a finite horizon in a discrete-time stochastic epidemic model, with respect to vaccination proportion v and contact rate β, while accommodating parameter uncertainty represented as distributions (e.g., from Bayesian calibration). It compares these estimators to deterministic large-population approximations, reports variance comparisons versus finite-difference methods, and uses simulations to conclude that parameter uncertainty reduces sensitivity (e.g., weaker indirect vaccination effects like herd immunity) and yields more conservative optimal policies when balancing intervention and infection costs.

Significance. If the unbiasedness claims hold under verifiable regularity conditions and the simulation results prove robust, the approach could enable more reliable sensitivity and optimization analyses for epidemic models that incorporate epistemic uncertainty from parameter estimation, addressing a practical gap between deterministic approximations and stochastic models with finite populations.

major comments (2)
  1. [Abstract and gradient estimator derivation] Abstract (first paragraph) and the section deriving the gradient estimators: the central claim of unbiased estimators for sensitivities under random parameters implicitly requires interchanging differentiation and expectation, i.e., d/dθ E_θ[f(X)] = E_θ[d/dθ f(X)] for the epidemic trajectory functional f. No regularity conditions (e.g., dominated convergence, Lipschitz continuity of the discrete-time transition kernel, or parameter-independent support) are stated to justify this interchange; if violated for the chosen parameter distributions, the estimators are biased and the downstream claims on reduced sensitivity and conservative policies do not follow.
  2. [Simulation experiments] The simulation experiments section: the reported reductions in sensitivity and shifts to conservative policies rest on the assumption that the discrete-time stochastic model is correctly specified and that the parameter distributions accurately capture uncertainty from Bayesian calibration, but no details on the model equations, transition probabilities, or experimental setup (e.g., population size, time horizon, specific distributions) are provided to allow verification of these effects.
minor comments (2)
  1. [Abstract] The abstract states that the estimators exhibit lower variance than finite-difference methods for the derivative w.r.t. β but higher for v, yet provides no quantitative values, sample sizes, or variance reduction factors to support this comparison.
  2. [Model description] Notation for the epidemic process (e.g., definition of total infections, the functional f, and how v and β enter the transition kernel) should be introduced explicitly early in the manuscript rather than assumed from context.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful comments, which help strengthen the presentation of our results on unbiased gradient estimators for stochastic epidemic models. We respond to each major comment below.

read point-by-point responses

  1. Referee: [Abstract and gradient estimator derivation] Abstract (first paragraph) and the section deriving the gradient estimators: the central claim of unbiased estimators for sensitivities under random parameters implicitly requires interchanging differentiation and expectation, i.e., d/dθ E_θ[f(X)] = E_θ[d/dθ f(X)] for the epidemic trajectory functional f. No regularity conditions (e.g., dominated convergence, Lipschitz continuity of the discrete-time transition kernel, or parameter-independent support) are stated to justify this interchange; if violated for the chosen parameter distributions, the estimators are biased and the downstream claims on reduced sensitivity and conservative policies do not follow.

    Authors: We agree that an explicit statement of regularity conditions is needed to rigorously support the interchange of differentiation and expectation. The revised manuscript will add a dedicated paragraph in the gradient estimator section listing the assumptions: finite time horizon, bounded transition probabilities, and application of the dominated convergence theorem given the compact parameter support and discrete state space. These conditions hold for the binomial transition kernels used in the model. revision: yes

  2. Referee: [Simulation experiments] The simulation experiments section: the reported reductions in sensitivity and shifts to conservative policies rest on the assumption that the discrete-time stochastic model is correctly specified and that the parameter distributions accurately capture uncertainty from Bayesian calibration, but no details on the model equations, transition probabilities, or experimental setup (e.g., population size, time horizon, specific distributions) are provided to allow verification of these effects.

    Authors: The model equations and binomial transition probabilities are defined in Section 2 of the manuscript, and the parameter distributions are described as posteriors from Bayesian calibration. To improve self-containment and verifiability, the revised simulation section will add an explicit table listing N=1000, T=50, and the exact distributions (e.g., β ~ N(0.4, 0.05)). We will also include pseudocode for the Monte Carlo procedure. revision: partial

Circularity Check

0 steps flagged

No circularity: gradient estimators derived independently of fitted inputs

full rationale

The paper derives unbiased gradient estimators for sensitivities w.r.t. v and β under parameter distributions using standard stochastic-gradient techniques applied to the discrete-time epidemic process. No step reduces a claimed prediction or sensitivity to a quantity already fitted from the same data by construction, nor does any load-bearing premise rest on a self-citation chain. Simulation results on reduced sensitivity and conservative policies are presented as empirical outcomes, not tautological consequences of the estimators themselves. The derivation chain is self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

Review performed from abstract only; full model equations, estimator derivations, and simulation protocols are unavailable, so the ledger is necessarily incomplete.

axioms (2)
  • standard math Interchange of derivative and expectation is valid for the gradient estimators in the discrete-time Markov chain epidemic model
    Required for unbiasedness claim in first paragraph of abstract
  • domain assumption Parameter uncertainty is accurately captured by the supplied distributions from Bayesian calibration
    Stated as the setting the estimators must accommodate

pith-pipeline@v0.9.1-grok · 5678 in / 1432 out tokens · 31109 ms · 2026-07-03T00:57:59.794689+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

57 extracted references · 8 canonical work pages

  1. [1]

    Ecological Monographs , volume=

    Grenfell, Bryan T and Bj. Ecological Monographs , volume=. 2002 , publisher=

    2002

  2. [2]

    Lomeli, Luis DJ and Ngo, Michelle N and Wakefield, Jon and Shahbaba, Babak and Minin, Vladimir N , journal=

  3. [3]

    SIAM Journal on Optimization , volume =

    Wu, Di and Zhu, Helin and Zhou, Enlu , title =. SIAM Journal on Optimization , volume =. 2018 , doi =. https://doi.org/10.1137/16M1101933 , abstract =

    work page doi:10.1137/16m1101933 2018

  4. [4]

    2021 , publisher=

    Cakmak, Sait and Wu, Di and Zhou, Enlu , journal=. 2021 , publisher=

    2021

  5. [5]

    Mathematical Biosciences , volume=

    Capistr. Mathematical Biosciences , volume=. 2012 , publisher=

    2012

  6. [6]

    2015 , publisher=

    Fu, Michael C , booktitle=. 2015 , publisher=

    2015

  7. [7]

    2015 , publisher=

    Chau, Marie and Fu, Michael C , booktitle=. 2015 , publisher=

    2015

  8. [8]

    2007 , publisher=

    Law, Averill M and Kelton, W David , edition=. 2007 , publisher=

    2007

  9. [9]

    2012 , month=sept, pages=

    Bulletin of Mathematical Biology , author=. 2012 , month=sept, pages=. doi:10.1007/s11538-012-9749-6 , abstractNote=

    work page doi:10.1007/s11538-012-9749-6 2012

  10. [10]

    Serfozo, Richard , year=

  11. [11]

    2024 , month=sept, pages=

    Journal of The Royal Society Interface , author=. 2024 , month=sept, pages=. doi:10.1098/rsif.2024.0299 , abstractNote=

    work page doi:10.1098/rsif.2024.0299 2024

  12. [12]

    Metcalf and M

    C.J.E. Metcalf and M. Ferrari and A.L. Graham and B.T. Grenfell , abstract =. Understanding Herd Immunity , journal =. 2015 , issn =. doi:https://doi.org/10.1016/j.it.2015.10.004 , url =

    work page doi:10.1016/j.it.2015.10.004 2015

  13. [13]

    Spatio-Temporal Analysis of Surveillance Data , booktitle=

    Wakefield, Jon and Dong, Tracy Qi and Minin, Vladimir N , year=. Spatio-Temporal Analysis of Surveillance Data , booktitle=

  14. [14]

    Communications of the ACM , volume=

    Likelihood ratio gradient estimation for stochastic systems , author=. Communications of the ACM , volume=. 1990 , publisher=

    1990

  15. [15]

    1996 , volume=

    Optimization of Stochastic Models: The Interface Between Simulation and Optimization , author=. 1996 , volume=

    1996

  16. [16]

    2020 , publisher=

    Saltelli, Andrea and Bammer, Gabriele and Bruno, Isabelle and Charters, Erica and Di Fiore, Monica and Didier, Emmanuel and Nelson Espeland, Wendy and Kay, John and Lo Piano, Samuele and Mayo, Deborah and others , journal=. 2020 , publisher=

    2020

  17. [17]

    2022 , publisher=

    Ioannidis, John PA and Cripps, Sally and Tanner, Martin A , journal=. 2022 , publisher=

    2022

  18. [18]

    2022 , publisher=

    Taleb, Nassim Nicholas and Bar-Yam, Yaneer and Cirillo, Pasquale , journal=. 2022 , publisher=

    2022

  19. [19]

    Pinson, Pierre and Makridakis, Spyros , journal=

  20. [20]

    2000 , publisher=

    Hethcote, Herbert W , journal=. 2000 , publisher=

    2000

  21. [21]

    2015 , publisher=

    Roberts, Mick and Andreasen, Viggo and Lloyd, Alun and Pellis, Lorenzo , journal=. 2015 , publisher=

    2015

  22. [22]

    Coffeng and Philip Dawid and Daniela

    Ben Swallow and Paul Birrell and Joshua Blake and Mark Burgman and Peter Challenor and Luc E. Coffeng and Philip Dawid and Daniela. Epidemics , volume =. 2022 , issn =. doi:https://doi.org/10.1016/j.epidem.2022.100547 , url =

    work page doi:10.1016/j.epidem.2022.100547 2022

  23. [23]

    2008 , publisher=

    Allen, Linda JS , booktitle=. 2008 , publisher=

    2008

  24. [24]

    doi:10.1007/978-1-4612-1158-7 , publisher=

    Andersson, Håkan and Britton, Tom , year=. doi:10.1007/978-1-4612-1158-7 , publisher=

    work page doi:10.1007/978-1-4612-1158-7

  25. [25]

    2019 , publisher=

    Britton, Tom and Pardoux, Etienne and Ball, Franck and Laredo, Catherine and Sirl, David and Tran, Viet Chi , volume=. 2019 , publisher=

    2019

  26. [26]

    Railsback, Steven F and Grimm, Volker , year=

  27. [27]

    BMC Infectious Diseases , volume=

    Ajelli, Marco and Gon. BMC Infectious Diseases , volume=. 2010 , publisher=

    2010

  28. [28]

    History and Philosophy of the Life Sciences , volume=

    Iranzo, Valeriano and P. History and Philosophy of the Life Sciences , volume=. 2021 , publisher=

    2021

  29. [29]

    PLoS Computational Biology , volume=

    Kerr, Cliff C and Stuart, Robyn M and Mistry, Dina and Abeysuriya, Romesh G and Rosenfeld, Katherine and Hart, Gregory R and N. PLoS Computational Biology , volume=

  30. [30]

    Annual Review of Public Health , volume=

    Tracy, Melissa and Cerd. Annual Review of Public Health , volume=. 2018 , publisher=

    2018

  31. [31]

    Nsoesie, Elaine O and Beckman, Richard J and Marathe, Madhav V , journal=

  32. [32]

    2022 , publisher=

    Borgonovo, Emanuele and Pangallo, Marco and Rivkin, Jan and Rizzo, Leonardo and Siggelkow, Nicolaj , journal=. 2022 , publisher=

    2022

  33. [33]

    Kouye, Henri Mermoz , year=

  34. [34]

    2019 , publisher=

    Berhe, Hailay Weldegiorgis and Makinde, Oluwole Daniel and Theuri, David Mwangi , journal=. 2019 , publisher=

    2019

  35. [35]

    Mester, Rachel and Landeros, Alfonso and Rackauckas, Chris and Lange, Kenneth , journal=

  36. [36]

    2023 , publisher=

    Lu, Xuefei and Borgonovo, Emanuele , journal=. 2023 , publisher=

    2023

  37. [37]

    Saltelli, Andrea and Ratto, Marco and Andres, Terry and Campolongo, Francesca and Cariboni, Jessica and Gatelli, Debora and Saisana, Michaela and Tarantola, Stefano , year=

  38. [38]

    2016 , publisher=

    Kucherenko, Sergei and Song, Shugfang , booktitle=. 2016 , publisher=

    2016

  39. [39]

    2016 , publisher=

    De Lozzo, Matthias and Marrel, Amandine , journal=. 2016 , publisher=

    2016

  40. [40]

    2026 , publisher=

    Mirsaeedi, Fatemeh and Sheikhalishahi, Mohammad and Mohammadi, Mehrdad and Pirayesh, Amir and Ivanov, Dmitry , journal=. 2026 , publisher=

    2026

  41. [41]

    Nsoesie, Elaine O and Beckman, Richard J and Shashaani, Sara and Nagaraj, Kalyani S and Marathe, Madhav V , journal=

  42. [42]

    2021 , publisher=

    Gillis, Melissa and Urban, Ryley and Saif, Ahmed and Kamal, Noreen and Murphy, Matthew , journal=. 2021 , publisher=

    2021

  43. [43]

    2017 , publisher=

    Paleshi, Arsalan and Bae, Ki-Hwan and Evans, Gerald and Heragu, Sunderesh , journal=. 2017 , publisher=

    2017

  44. [44]

    2025 , publisher=

    Ye, Yang and Pandey, Abhishek and Bawden, Carolyn and Sumsuzzman, Dewan Md and Rajput, Rimpi and Shoukat, Affan and Singer, Burton H and Moghadas, Seyed M and Galvani, Alison P , journal=. 2025 , publisher=

    2025

  45. [45]

    2008 , publisher=

    Tanner, Matthew W and Sattenspiel, Lisa and Ntaimo, Lewis , journal=. 2008 , publisher=

    2008

  46. [46]

    IISE Transactions on Healthcare Systems Engineering , volume=

    Yin, Xuecheng and B. IISE Transactions on Healthcare Systems Engineering , volume=. 2022 , publisher=

    2022

  47. [47]

    and La, Richard J

    Park, Shinkyu and Certorio, Jair and Martins, Nuno C. and La, Richard J. , journal=. Epidemic Population Games and Perturbed Best Response Dynamics , year=

  48. [48]

    Gao, Zhijian and Li, Shuxin and An, Bo , journal=

  49. [49]

    Chopra, Ayush and Rodr

  50. [50]

    PLoS One , volume=

    Navascu. PLoS One , volume=. 2021 , publisher=

    2021

  51. [51]

    2024 , publisher=

    Trostle, Parker and Guinness, Joseph and Reich, Brian J , journal=. 2024 , publisher=

    2024

  52. [52]

    2020 , publisher=

    Peng, Yijie and Fu, Michael C and Heidergott, Bernd and Lam, Henry , journal=. 2020 , publisher=

    2020

  53. [53]

    2013 , publisher=

    Wu, Jianyong and Dhingra, Radhika and Gambhir, Manoj and Remais, Justin V , journal=. 2013 , publisher=

    2013

  54. [54]

    2017 , organization=

    Alaeddini, Atiye and Klein, Daniel J , booktitle=. 2017 , organization=

    2017

  55. [55]

    Computational Statistics , volume=

    Rupp, Kevin and Schill, Rudolf and S. Computational Statistics , volume=. 2024 , publisher=

    2024

  56. [56]

    Robertson and Ben Swallow and Cerian R

    Liza Hadley and Peter Challenor and Chris Dent and Valerie Isham and Denis Mollison and Duncan A. Robertson and Ben Swallow and Cerian R. Webb , keywords =. Epidemics , volume =. 2021 , issn =. doi:https://doi.org/10.1016/j.epidem.2021.100499 , url =

    work page doi:10.1016/j.epidem.2021.100499 2021

  57. [57]

    Mathematics and Computers in Simulation , volume =

    I.M Sobol' , keywords =. Mathematics and Computers in Simulation , volume =. 2001 , issn =. doi:https://doi.org/10.1016/S0378-4754(00)00270-6 , url =

    work page doi:10.1016/s0378-4754(00)00270-6 2001