Pith. sign in

REVIEW 4 major objections 4 minor 20 references

A joint model of tumour growth and event hazards forecasts PFS and OS from immature trial data

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-01 16:37 UTC pith:OUCA2SIU

load-bearing objection Solid integration of existing joint-modelling components, but the month-4 'calibrated PFS forecast' claim is undercut by an LFO comparison that never generates the unenrolled patients. the 4 major comments →

arxiv 2607.17908 v1 pith:OUCA2SIU submitted 2026-07-20 stat.AP math.STstat.MEstat.OTstat.TH

PIONEER: Bayesian Joint Modelling of Mechanistic Tumour Growth and Time-to-Event Endpoints for Dynamic Prediction of Ongoing Oncology Trials

Karim Naguib , Roger Berch\'e , Lu Li , Antonia Bevan , Sajan Khosla , Jessica Davies , Paul Metcalfe This is my paper
classification stat.AP math.STstat.MEstat.OTstat.TH MSC 62P1062F1592C50
keywords Bayesian joint modellingtumour growth dynamicsmultistate survival modelleave-future-out validationdynamic predictionsmall-cell lung cancerPFS and OS forecastingGompertz growth
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

PIONEER aims to show that an ongoing oncology trial's future survival curves can be forecast reliably from immature data—sparse longitudinal tumour measurements and only a few events—by coupling a mechanistic two-component tumour-growth model to a multistate event-history model in a single Bayesian fit. The claim is demonstrated in extensive-stage small-cell lung cancer: at month 4 of enrolment, with only 9 target-trial patients, the model produces calibrated PFS forecasts that cover the mature month-19 Kaplan–Meier curve, and by month 11 (39 patients) the OS forecast converges—roughly 8 months of advance information with quantified uncertainty. If this holds generally, it would let trialists make go/no-go decisions earlier, without waiting for survival events to mature, and with honest uncertainty estimates. The paper is framed as a methodological demonstration, not yet a validated clinical tool, for individual-level prediction.

Core claim

PIONEER claims that immature longitudinal tumour-size data, together with whatever events have accumulated, are enough to produce calibrated forecasts of the mature PFS and OS distributions in an ongoing oncology trial—provided the tumour dynamics are constrained by a mechanistic two-component model (a treatment-responsive shrinking compartment plus a refractory growing compartment with Gompertz attenuation) and the hazards are learned jointly with it. In leave-future-out validation on two extensive-stage small-cell lung cancer trials (497 patients), the model's month-4 forecast (9 target-trial patients) already covers the observed month-19 PFS Kaplan–Meier curve, and the month-11 forecast (

What carries the argument

The load-bearing object is the latent two-component burden trajectory B_i(t) = D_i(t) + G_i(t), where D_i decays exponentially at rate r_dec and G_i grows at rate r_gro under a Gompertz decay that gives the growth arm a finite carrying capacity. This latent trajectory, inferred from noisy sum-of-longest-diameter (SLD) observations, feeds transition-specific proportional hazards through the bridge W_i(t) = (standardised log-burden, log decrease rate, log growth rate). Target-lesion progression is read off deterministically where the trajectory crosses the RECIST +20%-from-nadir threshold, while a visit-gated hazard handles non-target progression, and a full multistate routing (progression, di

Load-bearing premise

The extrapolated latent tumour-growth trajectory—especially the population growth rate and the Gompertz carrying capacity learned from a historical trial with no anti-tumour effect—represents the target population, because early PFS forecasts depend on when that trajectory crosses the RECIST progression threshold.

What would settle it

Refit the model at the month-4 cut-off with the population growth-rate prior replaced by one centred on the slow-growing half of the historical trial, and check whether the Arm A PFS forecast stops sitting systematically below the cutoff-censored Kaplan–Meier curve; if the pessimism persists unchanged, the bias is not driven by the growth-rate identifiability mechanism described in the paper. A second check is to apply the same leave-future-out protocol to a target trial whose kinetics differ from the historical trial and see whether the early PFS forecast remains calibrated.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • At month 4 of enrolment (9 patients), the model's PFS forecast is calibrated enough to cover the mature month-19 Kaplan–Meier curve, and at month 11 (39 patients) OS converges, implying at least 8 months of advance forecasting with quantified uncertainty.
  • Because all endpoints are read off the joint posterior, median PFS, median OS, and objective response rate inherit full parameter uncertainty without two-stage plug-in; ORR, derived entirely from the latent trajectory, is available even when hazard-only models could not produce it.
  • Transition-specific structure lets the same biomarker act in opposite directions on different transitions—e.g., haemoglobin lowers the progression hazard but raises post-progression death hazard—which a single-hazard joint model would collapse.
  • The framework is intended to accelerate go/no-go decisions; the paper argues that the observed conservative bias in early PFS forecasts (Arm A) aligns with the asymmetric cost of continuing a failing programme versus terminating a viable one.
  • Population-level calibration is necessary but not sufficient for individual-level predictions; the authors explicitly flag that patient-level clinical use requires further held-out validation.

Where Pith is reading between the lines

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

  • The advance-forecasting claim rests on borrowing tumour kinetics from a historical trial whose drug had no anti-tumour effect; a natural extension is to test whether borrowing from a truly anti-tumour comparator or from earlier windows of the same target trial changes the direction or magnitude of the acknowledged Arm A pessimism.
  • Because ORR is derived entirely from the latent trajectory, the architecture suggests a route to early response prediction before radiological confirmation, which could be tested against independent biomarkers such as ctDNA or serial LDH.
  • A stress test that perturbs the Gompertz carrying-capacity prior—for instance, centring it on the slow-growing half of the historical population—would show whether the systematic early PFS pessimism is an artefact of the prior or a structural feature of the growth-rate identifiability mechanism.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper introduces PIONEER, a Bayesian joint model that couples a mechanistic two-component state-space model of tumour size dynamics (decreasing and growing compartments, with Gompertz growth attenuation) to a multistate proportional-hazards model for progression, death, and dropout. Both submodels are fitted in a single posterior, and clinical endpoints (PFS, OS, ORR) are computed by forward simulation from the joint posterior, either conditionally on each patient's observed history or unconditionally from treatment start. The method is applied to two ES-SCLC trials (N=497), using the Amgen darbepoetin trial as historical data and the Lilly CXCR4 trial as the target. The headline claim is that leave-future-out cross-validation shows calibrated PFS forecasts at month 4 of enrolment (9 patients) and OS forecast convergence at month 11 (39 patients), giving 8 months of advance forecasting with quantified uncertainty.

Significance. If the central forecasting claim were supported, PIONEER would be a substantial methodological contribution: it integrates mechanistic tumour kinetics, multistate competing risks, and full Bayesian uncertainty propagation, and could enable earlier go/no-go decisions in oncology. The paper also has real strengths: the model is specified in detail, inference is performed with modern HMC in Stan, the joint-likelihood architecture avoids two-stage plug-in, and the LFO idea is appropriate in principle. However, the validation as reported does not establish the headline claim, and the paper's own discussion acknowledges a directional miscalibration for Arm A. The framework may still be valuable, but the evidence presented here is not yet sufficient for the abstract's 'calibrated forecasts' statement.

major comments (4)
  1. [§5.2 and Fig. 9] The LFO design does not support the abstract's claim. The conditional forecasting mode of §2.3.6 retains each observed patient's history and simulates only the unresolved future; it has no mechanism for generating the 69 target-trial patients not yet enrolled at the month-4 cutoff. The dashed 'Overall' reference is the full-cohort mature KM. Comparing a 9-patient conditional forecast to that KM is not a like-for-like evaluation; the apparent 'coverage' can simply reflect small-sample width. The paper must either compare conditional forecasts with the same patients' mature KM or switch to an unconditional (spop) simulation that includes future enrollees, and report coverage/interval scores.
  2. [§6.1 and Fig. 9 vs. abstract] The claim of 'calibrated PFS forecasts' is contradicted by the paper's own text. Section 6.1 states that for Arm A the model is 'systematically pessimistic' and 'overconfident in a pessimistic direction,' with the predictive interval 'consistently below' the observed PFS KM. No quantitative calibration metric (empirical coverage, Brier/IPA, interval score) is reported anywhere in §5.2. Visual 'covering' in Fig. 9 is not a calibration assessment. Please provide numerical calibration by arm and cutoff, and revise the abstract and conclusions accordingly.
  3. [§2.2.3 and §6.1] The long-horizon PFS forecast depends critically on the Gompertz saturating-growth parameter κ and the population growth-rate prior. The paper acknowledges that the population growth rate is biased upward by fast-regrowing patients and that the prior on ακ is chosen so that the carrying capacity is 'finite and reachable within the forecast window.' Because the central claim is calibrated early forecasting, the absence of a prior-sensitivity analysis for (ακ, population r_gro) is load-bearing. A re-fit with alternative priors, or with the historical kinetics excluded, is needed to establish that the early forecasts are not an artifact of these choices.
  4. [§4.1 and §5.2] The historical trial (Amgen) has no anti-tumour effect from the investigational agent, whereas the target Arm A includes a CXCR4 inhibitor. Borrowing population kinetics from a no-effect trial may systematically distort early forecasts for the experimental arm, especially at the month-4 cutoff when only 9 target-trial patients are available. Please report a sensitivity analysis that does not borrow kinetics from the Amgen trial or that allows treatment-arm-specific population kinetics, to assess the impact of this transportability assumption.
minor comments (4)
  1. [§5.3, Fig. 11] The text says the observed mature median PFS is ~6.2 months, but Table 1 reports arm-specific medians of 4.8 and 5.5 months. Please clarify which value is used and why the figure's dashed line differs from the tabulated arm-level values.
  2. [§2.2.2] Equation (3) is written as a differential equation but is then referred to as 'Substituting (Equation 5)' when deriving Eq. (4). The cross-referencing between equations and the text is confusing; please correct the equation numbering and references.
  3. [§5.2, Fig. 9] The 'At cutoff' observed KM for n=9 will be heavily censored and may contain very few discrete drops. Plotting it alongside a 9-patient forecast and a full-cohort mature KM makes visual comparison difficult. Please also show the number at risk and censoring indicators for each cut-off, and consider plotting the same-patient mature KM as the reference.
  4. [§2.3.5] The likelihood section is dense and would benefit from a small table summarizing the six terminal-pattern contributions. As written, the distinction between the gated λ01 and the deterministic target-lesion progression is easy to miss, despite being essential for interpreting the model.

Circularity Check

0 steps flagged

No significant circularity: PIONEER's forecasts are genuine out-of-sample posterior functionals; endpoints are not fitted constants relabeled as predictions.

full rationale

The derivation chain is self-contained. The mechanistic submodel (Eqs. 2–15) defines latent SLD dynamics; the multistate submodel (Eqs. 16–23) defines transition hazards; all endpoints are posterior functionals generated by forward simulation from the joint posterior (Section 2.3.6). The leave-future-out evaluation refits the model at each cut-off and compares posterior predictive KM curves with mature observed KM not available at training time. No fitted parameter is renamed as a prediction: PFS, OS, and ORR are deterministic or stochastic transforms of the posterior latent trajectories and hazards, not re-statements of fitted values. The Gompertz prior choice is a modeling assumption, not an output-derived constraint. The paper invokes no author self-citation chain and no uniqueness theorem; its citations to Stein, Claret, Bruno, and others are external prior work. The acknowledged Arm A pessimism and the ambiguity in the month-4 comparison (conditional forecasts for enrolled patients versus a full-cohort reference KM) are calibration/design concerns, not circularity.

Axiom & Free-Parameter Ledger

8 free parameters · 8 axioms · 0 invented entities

The model's central forecast rests on the parametric tumor-growth decomposition and the population hierarchy; the most consequential choices are fitting patient kinetics from sparse data and sharing them across trials. No physically new entities are introduced; the latent compartments are modeling constructs.

free parameters (8)
  • Initial decreasing fraction π_i (logit regression) = posterior estimates, values not tabulated
    Latent trajectory is initialized by π_i via Eq. (11)–(12); fitted from SLD and event data.
  • Decrease/growth rates r_dec_i, r_gro_i (magnitude-balance decomposition) = posterior estimates, values not tabulated
    Patient-level rates determine the entire trajectory (Eqs. 8, 13, 14) and feed the hazard bridge.
  • Gompertz decay intercept ακ (κ_i = e^{ακ}) = posterior estimate, value not tabulated
    Single population-level Gompertz decay controls the carrying capacity of the growth compartment (Eq. 10); prior chosen so carrying capacity is "reachable within the forecast window".
  • Observation noise σ_y = posterior estimate, value not tabulated
    Log-Gaussian measurement error in Eq. (15) is estimated and affects how tightly latent trajectories track SLD.
  • Bridge coefficients β_tv for log-burden, log r_dec, log r_gro = posterior estimates, shown in Fig. 15
    These link the mechanistic submodel to the 0→1 and 0→3 hazards (Eq. 21); the log-burden coefficient is described as "the mechanistic heart" of the model.
  • Transition-specific hazard coefficients β_ti for baseline covariates = posterior estimates, shown in Fig. 15
    Covariate effects on each transition are estimated separately (Eq. 16).
  • Frailty covariance Σγ (σ_01, σ_03, Rγ) = posterior estimates, values not tabulated
    Correlated patient-level frailties for progression and dropout are estimated jointly (Eq. 20).
  • GP hyperparameters for population and trial baselines = posterior estimates, values not tabulated
    Each transition hazard has a GP baseline with intercept, marginal SD, and length scale (Section 2.3.4), fitted on coarse knot grids.
axioms (8)
  • domain assumption Two-component decomposition B_i(t) = D_i(t) + G_i(t)
    Phenomenological split between responding and refractory burden, explicitly not claimed to recover clonal subpopulations (Eq. 2, Section 2.2.1).
  • domain assumption Exponential decay/growth with Gompertz attenuation
    Dynamics in Eqs. (3), (6), (8): D decays exponentially, G grows with exponentially decaying rate; κ_i ≥ 0 gives a finite carrying capacity. This parametric form determines all long-horizon extrapolations.
  • ad hoc to paper Covariates enter balance and initial split but not rate magnitude
    Eq. (14) restricts covariate effects to η_bal and η_init, not η_tot, described as a "deliberate identification choice" to avoid poor separation with short/noisy SLD.
  • domain assumption Gaussian log-SLD observation with LoD left-censoring
    Eq. (15) uses a log-normal observation model with shared σ_y and CDF left-censoring at y_LoD, assumed consistent with RECIST 1.1 measurement error.
  • domain assumption Multistate proportional hazards with GP baselines and frailty
    All five transitions share the log-linear hazard form (Eq. 16), with Gaussian-process baselines and multivariate-normal frailty (Eq. 20); semi-Markov clocks for post-progression and post-dropout death are assumed.
  • domain assumption Visit-gating of non-target progression hazard
    Non-target progression can only be detected at scheduled assessments, so λ01 is multiplied by a visit indicator (Eq. 18); this changes the interpretation of the 0→1 coefficient as a per-visit detection rate.
  • ad hoc to paper Cross-trial transportability of population kinetics and hazards
    A single hierarchical model pools the target Lilly CXCR4 trial with the historical Amgen darbepoetin trial, whose drug has no anti-tumor effect (Section 4.1). The claim that the population kinetics are common across the two trials is asserted, not tested.
  • domain assumption Weakly informative priors mapped through QR rotation
    Prior distributions are described as weakly informative and mapped through QR-rotated design matrices (Sections 2.2.5, 2.3.2), but exact prior forms are not specified, leaving prior sensitivity unquantified.

pith-pipeline@v1.3.0-alltime-deepseek · 25614 in / 11895 out tokens · 107900 ms · 2026-08-01T16:37:46.094905+00:00 · methodology

0 comments
read the original abstract

High-stakes decisions in oncology clinical trials must often be made while survival data remains immature: progression-free survival (PFS) and overall survival (OS) are heavily censored, few events have accumulated, and the primary endpoint may be months or years from reading out. What is available at interim data cut-offs is information-rich longitudinal tumour measurements and baseline covariates. We present PIONEER, a Bayesian joint modelling framework that couples a mechanistic two-component state-space submodel of longitudinal tumour size dynamics to a multistate proportional-hazard submodel for competing clinical events, fitted simultaneously under a single posterior. The mechanistic submodel infers latent per-patient tumour trajectories - decomposed into treatment-responsive and refractory compartments with Gompertz-attenuated growth - from sparse, noisy sum-of-longest-diameter (SLD) observations. These latent trajectories feed the multistate hazard as time-varying covariates, while the event data simultaneously refines the tumour dynamics through the joint likelihood. All clinical endpoints (PFS, OS, objective response rate) are derived from the joint posterior in a single forward simulation pass, propagating full parameter uncertainty without any two-stage plug-in. Applied to a case study in extensive-stage small-cell lung cancer (two trials, N = 497), leave-future-out cross-validation demonstrates that at month 4 of enrolment (9 patients) the model produces calibrated PFS forecasts covering the mature month-19 Kaplan-Meier curve, and at month 11 (39 patients) the OS forecast converges - representing at least 8 months of advance forecasting with properly quantified uncertainty. We hope this work paves the way for broader adoption of Bayesian mechanistic state-space frameworks in clinical development, enabling earlier and more informed decision-making from immature trial data.

Figures

Figures reproduced from arXiv: 2607.17908 by Antonia Bevan, Jessica Davies, Karim Naguib, Lu Li, Paul Metcalfe, Roger Berch\'e, Sajan Khosla.

Figure 4
Figure 4. Figure 4: — multistate state diagram (this section, Figure [PITH_FULL_IMAGE:figures/full_fig_p024_4.png] view at source ↗

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

20 extracted references · 13 canonical work pages

  1. [1]

    Bruno, Rene, Dean Bottino, Dinesh P de Alwis, et al. 2020. ``Progress and Opportunities to Advance Clinical Cancer Therapeutics Using Tumor Dynamic Models.'' Clinical Cancer Research 26 (8): 1787--95. https://doi.org/10.1158/1078-0432.CCR-19-0287

    work page doi:10.1158/1078-0432.ccr-19-0287 2020

  2. [2]

    Bruno, Rene, Mathilde Marchand, Kenta Yoshida, Jing Shao, Laurent Creancier, and Jin Y Jin. 2023. ``Tumor Dynamic Model-Based Decision Support for Phase Ib/II Combination Studies: A Retrospective Assessment Based on Resampling of the Phase III Study IMpower150 .'' Clinical Cancer Research 29 (6): 1047--55. https://doi.org/10.1158/1078-0432.CCR-22-2681

    work page doi:10.1158/1078-0432.ccr-22-2681 2023

  3. [3]

    Bruno, Rene, Francois Mercier, and Laurent Claret. 2014. ``Evaluation of Tumor Size Response Metrics to Predict Survival in Oncology Clinical Trials.'' Clinical Pharmacology & Therapeutics 95 (4): 386--93. https://doi.org/10.1038/clpt.2014.4

    work page doi:10.1038/clpt.2014.4 2014

  4. [4]

    Claret, Laurent, Rene Bruno, Jin-Fang Lu, Yu-Nien Sun, and Chih-Peng Hsu. 2012. ``Simulations Using a Drug-Disease Modeling Framework and Phase II Data Predict Phase III Survival Outcome in First-Line Non-Small-Cell Lung Cancer.'' Clinical Pharmacology & Therapeutics 92 (5): 631--34. https://doi.org/10.1038/clpt.2012.78

    work page doi:10.1038/clpt.2012.78 2012

  5. [5]

    Ruixi Zuideveld Klaas P and

    Claret, Laurent, Pascal Girard, Paulo M Hoff, Eric Van Cutsem, K and"; Ruixi Zuideveld Klaas P and"; Joerg, and Rene Bruno. 2009. ``Model-Based Prediction of Phase III Overall Survival in Colorectal Cancer on the Basis of Phase II Tumor Dynamics.'' Journal of Clinical Oncology 27 (25): 4103--8. https://doi.org/10.1200/JCO.2008.21.0807

    work page doi:10.1200/jco.2008.21.0807 2009

  6. [6]

    Desmée, Solène, France Mentré, Christine Veyrat-Follet, Bénédicte Sébastien, and Jérémie Guedj. 2017. ``Nonlinear Joint Models for Individual Dynamic Prediction of Risk of Death Using H amiltonian M onte C arlo: Application to Metastatic Prostate Cancer.'' BMC Medical Research Methodology 17: 105. https://doi.org/10.1186/s12874-017-0382-9

    doi:10.1186/s12874-017-0382-9 2017

  7. [7]

    Hickey, Graeme L, Pete Philipson, Andrea Jorgensen, and Ruwanthi Kolamunnage-Dona. 2016. ``Joint Modelling of Time-to-Event and Multivariate Longitudinal Outcomes: Recent Developments and Issues.'' BMC Medical Research Methodology 16: 117. https://doi.org/10.1186/s12874-016-0212-5

    work page doi:10.1186/s12874-016-0212-5 2016

  8. [8]

    Kerioui, Marion, Julie Bertrand, Rene Bruno, Francois Mercier, Jérémie Guedj, and Solène Desmée. 2022. ``Modelling the Association Between Biomarkers and Clinical Outcome: An Introduction to Nonlinear Joint Models.'' British Journal of Clinical Pharmacology 88 (4): 1452--63. https://doi.org/10.1111/bcp.15200

    doi:10.1111/bcp.15200 2022

  9. [9]

    Kerioui, Marion, Francois Mercier, Julie Bertrand, et al. 2020. ``Bayesian Inference Using H amiltonian M onte- C arlo Algorithm for Nonlinear Joint Modeling in the Context of Cancer Immunotherapy.'' Statistics in Medicine 39 (30): 4853--68. https://doi.org/10.1002/sim.8756

    work page doi:10.1002/sim.8756 2020

  10. [10]

    Meira-Machado, Lu\' s, Jacobo de Uña-Álvarez, Carmen Cadarso-Suárez, and Per Kragh Andersen. 2009. ``Multi-State Models for the Analysis of Time-to-Event Data.'' Statistical Methods in Medical Research 18 (2): 195--222. https://doi.org/10.1177/0962280208092301

    work page doi:10.1177/0962280208092301 2009

  11. [11]

    Putter, Hein, Marta Fiocco, and Ronald B Geskus. 2007. ``Tutorial in Biostatistics: Competing Risks and Multi-State Models.'' Statistics in Medicine 26 (11): 2389--430. https://doi.org/10.1002/sim.2712

    doi:10.1002/sim.2712 2007

  12. [12]

    Ribba, Benjamin, Nicholas HG Holford, Paolo Magni, et al. 2014. ``A Review of Mixed-Effects Models of Tumor Growth and Effects of Anticancer Drug Treatment Used in Population Analysis.'' CPT: Pharmacometrics & Systems Pharmacology 3 (5): e113. https://doi.org/10.1038/psp.2014.12

    work page doi:10.1038/psp.2014.12 2014

  13. [13]

    Rizopoulos, Dimitris. 2012. Joint Models for Longitudinal and Time-to-Event Data: With Applications in R . Chapman; Hall/CRC. https://doi.org/10.1201/b12208

    work page doi:10.1201/b12208 2012

  14. [14]

    Simeoni, Massimiliano, Paolo Magni, Cristiano Cammia, et al. 2004. ``Predictive Pharmacokinetic-Pharmacodynamic Modeling of Tumor Growth Kinetics in Xenograft Models After Administration of Anticancer Agents.'' Cancer Research 64 (3): 1094--101. https://doi.org/10.1158/0008-5472.CAN-03-2524

    work page doi:10.1158/0008-5472.can-03-2524 2004

  15. [15]

    Stein, Wilfred D, William Doug Figg, William Dahut, et al. 2008. ``Tumor Growth Rates Derived from Data for Patients in a Clinical Trial Correlate Strongly with Patient Survival: A Novel Strategy for Evaluation of Clinical Trial Data.'' The Oncologist 13 (10): 1046--54

    2008

  16. [16]

    Stein, Wilfred D, James L Gulley, Jeffrey Schlom, et al. 2011. ``Tumor Regression and Growth Rates Determined in Five Intramural NCI Prostate Cancer Trials: The Growth Rate Constant as an Indicator of Therapeutic Efficacy.'' Clinical Cancer Research 17 (4): 907--17. https://doi.org/10.1158/1078-0432.CCR-10-1762

    work page doi:10.1158/1078-0432.ccr-10-1762 2011

  17. [17]

    Tardivon, Coralie, Solène Desmée, Marion Kerioui, et al. 2019. ``Association Between Tumor Size Kinetics and Survival in Patients with Urothelial Carcinoma Treated with Atezolizumab: Implications for Patient Follow-up.'' Clinical Pharmacology & Therapeutics 106 (4): 810--20. https://doi.org/10.1002/cpt.1450

    doi:10.1002/cpt.1450 2019

  18. [18]

    Food and Drug Administration

    U.S. Food and Drug Administration. 2021. Model-Informed Drug Development: Clarifying Fit for Purpose. Center for Drug Evaluation; Research / Center for Biologics Evaluation; Research. https://www.fda.gov/media/154907/download

    2021

  19. [19]

    Wang, Yaning, Chyi Sung, Celine Dartois, et al. 2009. ``Elucidation of Relationship Between Tumor Size and Survival in Non-Small-Cell Lung Cancer Patients Can Aid Early Decision Making in Clinical Drug Development.'' Clinical Pharmacology & Therapeutics 86 (2): 167--74. https://doi.org/10.1038/clpt.2009.64

    work page doi:10.1038/clpt.2009.64 2009

  20. [20]

    Wooldridge, Jeffrey M. 2010. Econometric Analysis of Cross Section and Panel Data. 2nd ed. MIT Press. CSLReferences document

    2010