pith. sign in

arxiv: 2607.00980 · v1 · pith:O5V6UC6Bnew · submitted 2026-07-01 · 📊 stat.ME

An Instrumental Variable Approach to Account for Informative Treatment Switching in Real-world Evidence

Yang Liu , Andrew Ying , Zongqi Xia , Jue Hou This is my paper

Pith reviewed 2026-07-02 07:37 UTC · model grok-4.3

classification 📊 stat.ME
keywords instrumental variabletreatment switchingsurvival analysisdoubly robuststructural cumulative survival modelreal-world evidencemultiple sclerosisadditive hazards model
0
0 comments X

The pith

Baseline treatment serves as an instrumental variable to identify treatment effects under informative switching in survival data.

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

The paper develops an instrumental variable method to address treatment switching that depends on expected outcomes in observational survival studies. Baseline treatment is established as a valid instrument for the switching decision. An estimating equation is formed by linking the centered instrument to a martingale-style residual process, which identifies the treatment effect parameter in a structural cumulative survival model. The resulting estimator is doubly robust, remaining consistent when either the baseline propensity model or the no-switching outcome model is correctly specified. A co-training procedure and baseline-survival-corrected cross-fitting allow implementation with semi-parametric additive hazards models and general machine learning for nuisance functions, as demonstrated in simulations and an application to multiple sclerosis treatment comparison.

Core claim

Treating baseline treatment as an instrumental variable for the switching decision, the authors construct an estimating equation that associates the centered instrumental variable with a martingale residual process. This identifies the treatment effect under the structural cumulative survival model. The procedure is doubly robust and can be implemented by jointly estimating the treatment effect parameter and the survival outcome regression under semi-parametric additive hazards models, without requiring observation of a no-switching subset. Baseline-survival-corrected cross-fitting further permits flexible machine learning estimation of nuisance models.

What carries the argument

The estimating equation that associates the centered baseline-treatment instrumental variable with the martingale-style residual process under the structural cumulative survival model.

If this is right

  • The estimator remains consistent whenever either the baseline propensity score model or the no-switching outcome model is correctly specified.
  • The method recovers the treatment effect without requiring a subset of patients who never switch treatment.
  • Simulations show the estimator is valid in settings where standard approaches yield biased or contradictory results.
  • The method produces a comparison of high-efficacy versus standard-efficacy disease-modifying treatments as second-line therapy in multiple sclerosis.

Where Pith is reading between the lines

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

  • The approach may extend naturally to other longitudinal outcomes where switching decisions depend on unobserved prognosis.
  • Choosing baseline covariates that satisfy the instrumental variable conditions could reduce bias in other real-world evidence settings with time-varying treatment.
  • Mild violations of the instrumental variable assumption could be assessed by sensitivity analyses that vary the strength of the direct effect of baseline treatment on the outcome.

Load-bearing premise

Baseline treatment is a valid instrumental variable, meaning it influences the switching decision but has no direct effect on the outcome except through the treatment actually received.

What would settle it

A dataset generated from a known data-generating process with informative switching, where the baseline treatment is unrelated to switching or directly affects the outcome, would produce inconsistent estimates if the identification strategy fails.

read the original abstract

Reproducible and generalizable assessment of treatment decisions requires principled handling of subsequent treatment switching that may inform expected outcomes and shift across cohorts and over time. To effectively account for informative treatment switching, we propose an instrumental variable approach that characterizes the poorly documented expected outcomes at switching as unmeasured confounding. After establishing the baseline treatment as a viable instrumental variable, we constructed an estimating equation based on the association between the centered instrumental variable and a martingale style residual process that identifies the treatment effect under structural cumulative survival model. Our proposed method is doubly robust, i.e., valid whenever either of baseline propensity model or no-switching outcome model is consistently estimated. A co-training of treatment effect parameter and survival outcome regression model eliminated the requirement of observing a no-switching subset under semi-parametric additive hazards models. We further developed an baseline-survival-corrected cross-fitting approach to incorporate general machine learning models for estimating nuisance models. Numerical results demonstrated the validity of our method in various settings when a basket of benchmark solutions produced biased or contradictory results. We applied our method to comparison of high-efficacy vs standard efficacy disease modifying treatments as the second line therapy of multiple sclerosis.

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 paper proposes an instrumental variable estimator for treatment effects in the presence of informative treatment switching, using baseline treatment as the IV. It constructs an estimating equation that correlates the centered baseline-treatment IV with a martingale residual under a structural cumulative survival model, claims double robustness with respect to the baseline propensity score or the no-switching outcome model, introduces co-training of the treatment effect parameter with the survival regression and a baseline-survival-corrected cross-fitting procedure to allow machine-learning nuisances, demonstrates performance in simulations against biased benchmarks, and applies the method to compare high- versus standard-efficacy disease-modifying therapies as second-line treatment in multiple sclerosis.

Significance. If the identification strategy is valid, the method supplies a doubly robust route to estimating causal effects on survival outcomes when switching is informative and no pure no-switching subpopulation is observed. The combination of structural cumulative survival modeling, co-training, and cross-fitting for flexible nuisances is a potentially useful extension of existing IV and survival techniques for real-world evidence studies.

major comments (2)
  1. [Abstract and §3 (estimating equation)] The central identification result rests on baseline treatment satisfying the IV conditions (relevance for switching, exclusion restriction, and independence from unmeasured confounders). The abstract and methods assert that baseline treatment is a viable IV, but the double-robustness property protects only against misspecification of the propensity or outcome nuisance models and does not relax these IV conditions. A dedicated subsection should state the precise IV assumptions required for consistency of the estimating equation and discuss their plausibility in the MS application.
  2. [§3 (identification and estimating equation)] The claim that the estimator identifies the treatment effect under the structural cumulative survival model via the martingale residual process requires explicit verification that the estimating equation is unbiased under the stated model when either nuisance is correct. The paper should provide the key steps showing how the co-training step eliminates the need to observe a no-switching subset while preserving the double-robustness property.
minor comments (2)
  1. [§2–3] Notation for the structural cumulative survival model and the martingale residual should be introduced with a short table or explicit definitions before the estimating equation is presented.
  2. [Simulation study] In the simulation section, report the exact data-generating process parameters used to enforce the IV conditions so that readers can reproduce the relevance and exclusion settings.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help clarify the identification strategy and strengthen the presentation. We address each major comment below and will revise the manuscript to incorporate the requested details.

read point-by-point responses

  1. Referee: [Abstract and §3 (estimating equation)] The central identification result rests on baseline treatment satisfying the IV conditions (relevance for switching, exclusion restriction, and independence from unmeasured confounders). The abstract and methods assert that baseline treatment is a viable IV, but the double-robustness property protects only against misspecification of the propensity or outcome nuisance models and does not relax these IV conditions. A dedicated subsection should state the precise IV assumptions required for consistency of the estimating equation and discuss their plausibility in the MS application.

    Authors: We agree that the IV assumptions are foundational and that double robustness applies only to the nuisance models, not the core IV conditions. In the revised manuscript we will add a dedicated subsection in §3 that explicitly lists the three IV assumptions (relevance, exclusion restriction, and independence from unmeasured confounders) required for the estimating equation to identify the treatment effect. We will also provide a brief discussion of their plausibility in the multiple sclerosis application, supported by references to clinical literature on baseline treatment assignment practices. revision: yes

  2. Referee: [§3 (identification and estimating equation)] The claim that the estimator identifies the treatment effect under the structural cumulative survival model via the martingale residual process requires explicit verification that the estimating equation is unbiased under the stated model when either nuisance is correct. The paper should provide the key steps showing how the co-training step eliminates the need to observe a no-switching subset while preserving the double-robustness property.

    Authors: We agree that the unbiasedness of the estimating equation and the role of co-training merit explicit derivation. In the revision we will insert the key algebraic steps showing that the estimating equation is unbiased under the structural cumulative survival model whenever either the baseline propensity score or the no-switching outcome model is correctly specified. We will also detail how the co-training procedure—jointly solving for the treatment-effect parameter and the survival regression coefficients—removes the requirement of observing a no-switching subpopulation while preserving double robustness. revision: yes

Circularity Check

0 steps flagged

No significant circularity; derivation relies on explicit IV assumptions and standard doubly robust construction

full rationale

The paper's core identification uses an estimating equation linking the centered baseline-treatment IV to a martingale residual under the structural cumulative survival model, with double robustness protecting nuisance models but not relaxing the stated IV conditions (relevance, exclusion, independence). No quoted step reduces the target parameter to a fitted input by construction, renames a known result, or chains to a self-citation that itself assumes the result. The baseline-treatment IV validity is presented as an assumption to be established from domain knowledge rather than derived internally. The method is self-contained against external benchmarks once the IV conditions hold.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the validity of the instrumental variable assumption for baseline treatment and the structural cumulative survival model; these are domain assumptions rather than new inventions or fitted parameters.

axioms (1)
  • domain assumption Baseline treatment is a valid instrumental variable for the switching process.
    Explicitly stated in abstract as the foundation after 'establishing the baseline treatment as a viable instrumental variable'.

pith-pipeline@v0.9.1-grok · 5731 in / 1043 out tokens · 24506 ms · 2026-07-02T07:37:00.329284+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

51 extracted references · 1 canonical work pages

  1. [1]

    Robinson, P. M. , title =. Econometrica , year =

  2. [2]

    Andersen, P. K. and Gill, R. D. , title =. The Annals of Statistics , year =

  3. [3]

    2020 , volume =

    Examining the use of real-world evidence in the regulatory process , journal =. 2020 , volume =

    2020

  4. [4]

    and Bonafede, Machaon , title =

    Bowen, James and Mehta, Rina and Pelletier, Corey and Tian, Marc and Noxon, Virginia and Johnson, Barbara H. and Bonafede, Machaon , title =. Advances in Therapy , year =

  5. [5]

    The Annals of Family Medicine , year =

    Boyd, Cynthia and Paolino, Valerie and Goodrich, Glenn and Norton, Jonathan and Kraus, Courtney and Sheehan, Orla and Reeve, Emily and Bayliss, Elizabeth and Barrow, Jennifer and Palen, Ted , title =. The Annals of Family Medicine , year =

  6. [6]

    Econometrics , year =

    Chen, Jau-er and Huang, Chien-Hsun and Tien, Jia-Jyun , title =. Econometrics , year =

  7. [7]

    The Econometrics Journal , year =

    Chernozhukov, Victor and Chetverikov, Denis and Demirer, Mert and Duflo, Esther and Hansen, Christian and Newey, Whitney and Robins, James , title =. The Econometrics Journal , year =

  8. [8]

    Clephas, Pascal R. D. and Malgie, Jishnu and Schaap, Jeroen and Koudstaal, Stefan and Emans, Mireille and Linssen, Gerard C. M. and. Guideline implementation, drug sequencing, and quality of care in heart failure: Design and rationale of. ESC Heart Failure , year =

  9. [9]

    Statistics in Medicine , year =

    Cuzick, Jack and Edwards, Robert and Segnan, Nereo , title =. Statistics in Medicine , year =

  10. [10]

    ClinicoEconomics and Outcomes Research , year =

    Degli Esposti, Luca and Favalli, Ennio Giulio and Sangiorgi, Diego and Di Turi, Roberta and Farina, Giuseppina and Gambera, Marco and Ravasio, Roberto , title =. ClinicoEconomics and Outcomes Research , year =

  11. [11]

    Statistical Science , year =

    Ding, Peng and Fan, Jianqing , title =. Statistical Science , year =

  12. [12]

    and Vansteelandt, Stijn , title =

    Dukes, Oliver and Martinussen, Torben and Tchetgen Tchetgen, Eric J. and Vansteelandt, Stijn , title =. Biometrics , year =

  13. [13]

    Journal of Machine Learning Research , year =

    Dukes, Oliver and Vansteelandt, Stijn and Whitney, David , title =. Journal of Machine Learning Research , year =

  14. [14]

    and Rubin, Donald B

    Frangakis, Constantine E. and Rubin, Donald B. , title =. Biometrics , year =

  15. [15]

    and Pawar, Ajinkya and Martin, David and Glynn, Robert J

    Franklin, Jessica M. and Pawar, Ajinkya and Martin, David and Glynn, Robert J. and Levenson, Mark and Temple, Robert and Schneeweiss, Sebastian , title =. Clinical Pharmacology & Therapeutics , year =

  16. [16]

    and Liaw, Kai-Li and Iyasu, Solomon and Critchlow, Cathy W

    Franklin, Jessica M. and Liaw, Kai-Li and Iyasu, Solomon and Critchlow, Cathy W. and Dreyer, Nancy A. , title =. Pharmacoepidemiology and Drug Safety , year =

  17. [17]

    Scheike , title =

    Anders Gorst-Rasmussen and Thomas H. Scheike , title =. Journal of Statistical Software , year =

  18. [18]

    Beyond the intention to treat in comparative effectiveness research , journal =

    Hern. Beyond the intention to treat in comparative effectiveness research , journal =. 2012 , volume =

    2012

  19. [19]

    Instruments for causal inference: An epidemiologist's dream? , journal =

    Hern. Instruments for causal inference: An epidemiologist's dream? , journal =. 2006 , volume =

    2006

  20. [20]

    Per-protocol analyses of pragmatic trials , journal =

    Hern. Per-protocol analyses of pragmatic trials , journal =. 2017 , volume =

    2017

  21. [21]

    Eric C Chi and Tamara G Kolda

    Hou, Jue and Bradic, Jelena and Xu, Ronghui , journal =. Treatment. 2023 , publisher =. doi:10.1080/01621459.2021.1930546 , langid =

    work page doi:10.1080/01621459.2021.1930546 2023

  22. [22]

    JAMA Network Open , year =

    Hou, Jue and Kim, Nicole and Cai, Tianrun and Dahal, Kumar and Weiner, Howard and Chitnis, Tanuja and Cai, Tianxi and Xia, Zongqi , title =. JAMA Network Open , year =

  23. [23]

    Journal of Medical Internet Research , year =

    Hou, Jue and Zhao, Rachel and Gronsbell, Jessica and others , title =. Journal of Medical Internet Research , year =

  24. [24]

    Journal of Machine Learning Research , year =

    Hou, Jue and Mukherjee, Rajarshi and Cai, Tianxi , title =. Journal of Machine Learning Research , year =

  25. [25]

    , title =

    Ishwaran, Hemant and Kogalur, Udaya B. , title =. R News , year =

  26. [26]

    Multiple Sclerosis and Related Disorders , year =

    Liang, Liang and Kim, Nicole and Hou, Jue and Cai, Tianrun and Dahal, Kumar and Lin, Chen and Finan, Sean and Savova, Guergana and Rosso, Mattia and Polgar-Turcsanyi, Mariann and Weiner, Howard and Chitnis, Tanuja and Cai, Tianxi and Xia, Zongqi , title =. Multiple Sclerosis and Related Disorders , year =

  27. [27]

    Lin, D. Y. and Ying, Zhiliang , title =. Biometrika , year =

  28. [28]

    , title =

    Martinussen, Torben and Scheike, Thomas H. , title =. 2006 , doi =

    2006

  29. [29]

    and Zucker, David M

    Martinussen, Torben and Vansteelandt, Stijn and Tchetgen Tchetgen, Eric J. and Zucker, David M. , title =. Biometrics , year =

  30. [30]

    , title =

    Michael, James and Cui, Ziyi and Lorch, Scott and Tchetgen Tchetgen, Eric J. , title =. Journal of the American Statistical Association , year =

  31. [31]

    and Caniglia, Ellen C

    Murray, Eleanor J. and Caniglia, Ellen C. and Petito, Lucia C. , title =. Research Methods in Medicine & Health Sciences , year =

  32. [32]

    and Darsaut, T

    Olijnyk, L. and Darsaut, T. E. and. Understanding intent to treat analyses: An important lesson from the international cooperative study on the timing of aneurysm surgery , journal =. 2022 , volume =

    2022

  33. [33]

    2025 , howpublished =

    Papazoglou, Theodosios and Waddingham, Ed and Young, Alastair , title =. 2025 , howpublished =

    2025

  34. [34]

    , title =

    Prentice, Ross L. , title =. Statistics in Medicine , year =

  35. [35]

    and Finkelstein, Dianne M

    Robins, James M. and Finkelstein, Dianne M. , title =. Biometrics , year =

  36. [36]

    and Rotnitzky, Andrea and Zhao, Lue Ping , title =

    Robins, James M. and Rotnitzky, Andrea and Zhao, Lue Ping , title =. Journal of the American Statistical Association , year =

  37. [37]

    and Tsiatis, Anastasios A

    Robins, James M. and Tsiatis, Anastasios A. , title =. Communications in Statistics - Theory and Methods , year =

  38. [38]

    and Rubin, Donald B

    Rosenbaum, Paul R. and Rubin, Donald B. , title =. Biometrika , year =

  39. [39]

    Annals of General Psychiatry , year =

    Roussidis, Andreas and Kalkavoura, Christina and Dimelis, Dimos and Theodorou, Afroditi and Ioannidou, Ina and Mellos, Eleytherios and Mylonaki, Triantafyllia and Spyropoulou, Areti and Yfantis, Andreas , title =. Annals of General Psychiatry , year =

  40. [40]

    Biometrics , year =

    Seaman, Shaun and Dukes, Oliver and Keogh, Ruth and Vansteelandt, Stijn , title =. Biometrics , year =

  41. [41]

    and Hayward, Rodney A

    Sussman, Jeremy B. and Hayward, Rodney A. , title =. BMJ , year =

  42. [42]

    and Atkinson, Beth , title =

    Therneau, Terry M. and Atkinson, Beth , title =. 2023 , note =

    2023

  43. [43]

    and Dekker, Friedo W

    Tripepi, Giovanni and Chesnaye, Nicholas C. and Dekker, Friedo W. and Zoccali, Carmine and Jager, Kitty J. , title =. Nephrology , year =

  44. [44]

    and Oxnard, Geoffrey R

    Gong, Yutao and Kehl, Kenneth L. and Oxnard, Geoffrey R. and Khozin, Sean and Mishra-Kalyani, Pallavi Shruti and Blumenthal, Gideon Michael , title =. Journal of Clinical Oncology , year =

  45. [45]

    2006 , volume =

    Targeted maximum likelihood learning , journal =. 2006 , volume =

    2006

  46. [46]

    1996 , series =

    Weak Convergence and Empirical Processes , publisher =. 1996 , series =

    1996

  47. [47]

    Statistical Science , year =

    Vansteelandt, Stijn and Joffe, Marshall , title =. Statistical Science , year =

  48. [48]

    , title =

    Wang, Yan and Lee, Mihye and Liu, Pengfei and Shi, Liuhua and Yu, Zhi and Awad, Yara Abu and Zanobetti, Antonella and Schwartz, Joel D. , title =. Epidemiology , year =

  49. [49]

    , title =

    Ying, Andrew and Tchetgen Tchetgen, Eric J. , title =. Biometrics , year =

  50. [50]

    Information and Inference: A Journal of the IMA , year =

    Zhang, Yuqian and Chakrabortty, Abhishek and Bradic, Jelena , title =. Information and Inference: A Journal of the IMA , year =

  51. [51]

    Biometrics , year =

    Almirall, Daniel and Ten Have, Thomas and Murphy, Susan A , title =. Biometrics , year =