Pith. sign in

REVIEW 4 major objections 6 minor 46 references

Equitable Discrimination in Survival Prediction: The Maximum Expected C-Index

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A survival model's ranking power is capped by the expected C-Index of the true hazard model, not by 1.

desk verdict The subpopulation C-index is a useful descriptive tool, but the oracle-based upper bound is computed from observed survival times, so the headline 0.75 ECI and the discrimination ratios are not valid. read the letter →

arxiv 2506.05592 v1 pith:G22HOEQE submitted 2025-06-05 stat.AP

Felipe Simon , Francisco Perez-Galarce , Joris van de Klundert This is my paper
classification stat.AP MSC 62N0162N0262P10
keywords C-Indexexpectedsubpopulationsurvivalanalysisproportionalhazardsdiscriminationratiopredictionfairnesskidneygraft
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

Under the standard proportional-hazards assumption of survival analysis, this paper derives a ceiling on how well any prediction model can rank patients by survival time. The ceiling—the expected C-Index of the true hazard model—replaces the trivial upper bound of 1 with a number that is often much lower: around 0.75 for ten-year death-censored graft survival after kidney transplant in the UNOS/OPTN registry. At that ceiling, a commonly reported C-Index around 0.61 is not a mediocre score but roughly halfway between random ranking (0.5) and the best rank ordering the data allow. The paper also defines a subpopulation version of the C-Index and shows that, because subpopulation ceilings are nearly identical while achieved scores differ significantly, a standard Cox model ranks some ethnic groups less accurately than others—a model bias with implications for allocation and treatment decisions.

What carries the argument

The key object is the pairwise win probability $p_{ij} = P(T_i > T_j) = h_j/(h_i+h_j)$, where $h_i$ is the time-independent hazard multiplier of member $i$ under proportional hazards. This identity carries the argument because it lets the expected C-Index be written as a sum of terms that the true ranking maximizes termwise, independently of the baseline survival function. In practice the bound is computed by recovering empirical hazard multipliers $h^*_i$ from an estimated baseline survival function and observed event times, then forming the model $M^*$ whose expected C-Index is the upper bound.

What would settle it

Simulate survival times from a known proportional-hazards model with a Weibull baseline and two strong predictors, compute the ECI ceiling from the true hazard multipliers, then fit a Cox model that omits one predictor; if its observed C-Index exceeds the ECI ceiling across repeated random 50/50 splits, the empirical version of the bound fails as a universal ceiling.

Watch

Extended reading notes

Core claim

Proposition 2 is the load-bearing result: for any prediction model $\hat M$, $E[CI(\hat M)] \le E[CI(M)]$, where $M$ is the true model generated by the proportional-hazards survival functions; and in uncensored complete-case data $E[CI(M)] \le E[CI^U(M)]$. The proof rests on Proposition 1, which gives the pairwise probability that member $i$ outlives member $j$ as $p_{ij} = h_j/(h_i+h_j)$, a quantity independent of the baseline hazard. Because the true model maximizes concordance in expectation, scores that look low against 1 can be near the attainable ceiling when judged against ECI. In the UNOS/OPTN case study the ceiling for ten-year death-censored graft survival is about 0.75, and a Cox model scoring 0.61-0.62 has a discrimination ratio of roughly 37-52 percent—closing about half the gap between chance and the boundary. For subpopulations, the SUBECI ceilings barely differ across Asian, Black, Hispanic/Latino, and White patients, while the SUBCIs differ significantly, which the paper reads as evidence that the Cox model is ethnically biased.

Load-bearing premise

The load-bearing premise is that the hazard multipliers $h^*_i$ used to compute the ceiling are unbiased estimates of the true hazards, which requires the proportional-hazards model to be correctly specified with no important predictors omitted; if the model is misspecified, the computed ECI is not a ceiling every other model must respect.

Editorial extensions

If this is right

  • Reported C-Index values should be read against ECI rather than against 1: a 0.61 score on ten-year death-censored graft survival lies about halfway between 0.5 and the 0.75 ceiling, not in the lower-middle range toward 1.
  • The discrimination ratio $DR = (CI - 0.5)/(ECI - 0.5)$ gives a normalized performance measure; values around 0.37-0.52 in this case study quantify how much of the attainable ranking power the Cox model captures.
  • Because subpopulation ceilings are similar while observed SUB C-Indices differ significantly, the gap between SUBCI and SUBECI can flag biased models even when the overall C-Index looks acceptable.
  • Undersampling to balance ethnic subpopulations raises most subpopulation C-Indices but does not remove the differences, so data imbalance alone does not explain the bias.
  • The upper bound applies to any model whose implied survival functions satisfy proportional hazards, so machine-learning survival models can be compared with Cox models against the same ECI benchmark.

Reading between the lines

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

  • Testable extension: apply the same ECI and SUBECI machinery to other allocation-driven survival problems, such as liver or heart transplant, credit-default, or customer-churn survival, to see whether subpopulation ceilings stay nearly flat while model scores diverge; the kidney-transplant equity conclusion would be strengthened if the pattern generalizes.
  • Implicit consequence: if ECI is a tight ceiling, then the room for improvement in a Cox model is exactly the gap between its discrimination ratio and 1; a ranking-oriented machine-learning model that surpasses the Cox DR without exceeding ECI would demonstrate that the ceiling is reachable in practice.
  • Neighbouring problem: the same pairwise win-probability identity suggests closed-form expected-concordance bounds for other C-Index variants that the paper lists as future work; deriving those would turn each variant into a comparable benchmark rather than an unnormalized score.
  • Design implication: because SUBECI ceilings are nearly equal while SUBCIs differ, a fairness intervention could aim to equalize subpopulation discrimination ratios rather than raw C-Indices; this is a target the paper gestures at but does not itself implement.
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.

Referee Report

4 major / 6 minor

Summary. The paper proposes a new upper bound, the expected C-Index (ECI), for the C-Index of survival prediction models under proportional hazards, together with a subpopulation C-Index (SUBCI) and a discrimination ratio (DR). The theoretical claims are given as Proposition 1 (pairwise concordance probability for proportional hazards) and Proposition 2 (the expected C-Index of the true model is an upper bound for any prediction model, with a tighter bound in the uncensored case). The methods are applied to UNOS kidney-transplant data to estimate ECI for death-censored graft survival, to compare overall and subpopulation C-Indices, and to argue that a standard Cox model is ethnically biased. The paper claims ECI around 0.75, making a C-Index of ~0.61 close to half the gap to the upper bound.

Significance. If the central theorem were correct, the ECI would provide a practically useful benchmark for survival prediction, and the subpopulation extension would be a valuable contribution to fairness evaluation in medical prediction. The paper has concrete strengths: Proposition 1 is correct under the stated assumptions; the definition of SUBCI is a meaningful and apparently novel way to measure discrimination within subpopulations; the case study uses a large registry dataset and the code is publicly available. However, the central upper-bound claim (Proposition 2) is not valid for censored data, and the empirical ECI is computed from a data-dependent oracle construction rather than from the true hazards. These flaws are load-bearing: they affect the headline ECI value, the discrimination ratios, and the ethnic-bias conclusions. The contribution is therefore not yet established.

major comments (4)
  1. [Section 2.3, Proposition 2] The proof of the first inequality in Proposition 2 is incorrect, and the statement is false for censored data. The problem appears in the line: 'By definition of M, Mij = 1 implies hj/(\tilde h_i + h_j) > 1/2 > \tilde h_i/(\tilde h_i + h_j).' This implication is not valid: from Mij=1 we have hi < hj, but \tilde h_i can exceed h_j when i is censored, so hj/(\tilde h_i+h_j) may be less than 1/2. Moreover, the comparison with the \hat M_j i term is wrong: when \hat M predicts j>i, the relevant probability is E[A_ji] = hi/(\tilde h_j + hi), not \tilde h_i/(\tilde h_i+h_j). A concrete counterexample: two members, i censored at time 0 (so \tilde T_i=0) and j uncensored, with hi=1, hj=2. The true model M predicts i>j (since E[Ti]>E[Tj]), but E[A_ij]=0 and E[A_ji]=1, so E[CI(M)]=0 while any model that predicts j>i has E[CI]=1. This violates Eq. (14). Since all subsequent results (DR, SUBECI, and the case-study conclusions) rely on Proposition 2 as an upper bound, the censored-data bound is not established.
  2. [Section 3, construction of M*] The empirical ECI is not a valid upper bound because it is computed from oracle hazard rates. The text states: 'For Dret, we can calculate unbiased hazard rates h*_i satisfying E[T*_i] = Ti' and then forms M* with survival functions S*_i(t)=[S0(t)]^{h*_i}. These h*_i are functions of the realized outcome Ti, not the population hazard parameters hi used in Proposition 2; accordingly the paper observes that CI(M*)=1 in Dret. Proposition 2 applies to the true model M whose hazards are fixed population parameters, not to a model that sees the outcomes. Replacing hi by h*_i = g(Ti) can inflate the bound: in a homogeneous population the true PH bound is 0.5, but the construction yields approximately 0.75 for exponential survival. Therefore the ECI values of about 0.75 reported in Tables 1 and 2, the discrimination ratios in Tables 3 and 4, and the claim that a C-Index of 0.61 'closes almost half the gap' are not supported by the theory.
  3. [Section 2.2 and Definition 3, notational inconsistency] The definition of Aij is inconsistent across the manuscript. In Section 2.2, Aij is defined for censored data as Aij = 1 if \tilde T_i > T_j (with \tilde T_i the observed censored/event time), and this is the form used in the proof of Proposition 2 and in the ECI expressions. Definition 3, however, restates Aij = 1 if Ti > Tj, using the unobserved event time, and the surrounding text returns to uncensored notation without cleaning up. This ambiguity makes it difficult to verify which expression is used in the case study; it also affects the interpretation of K and the weights in Section 2.4. The authors should standardize notation and make explicit whether all Aij in the censored setting are based on \tilde T_i.
  4. [Section 3, transfer of the bound from Dret to Dpro] The statement 'if both Dret and Dpro are large enough random samples from the same original data set, any maximum expected C-Index calculated for Dret also applies to Dpro' is an additional assumption that is not justified. Even if the oracle-hazard issue were resolved, the h*_i are defined on the realized event times in Dret, and their distribution in Dpro is not shown to be the same. A formal argument for the transferability of the upper bound is needed; without it, the case-study numbers compare a Dret-derived bound to a Dpro-derived C-Index in an unsupported way. This issue is separate from the misspecification concern raised in the text.
minor comments (6)
  1. [Section 1, Introduction] The abstract and introduction treat 'survival' and 'time to event' interchangeably; the paper should state explicitly that longer survival corresponds to later event time, to avoid confusion with the proportional-hazards direction.
  2. [Section 2.3, Definition 4] Definition 4 uses E[Aij] without specifying the probability space; it would be clearer to write E[Aij] as a function of the hazards and censoring distribution, and to define the expectation over the joint distribution of (Ti, Di).
  3. [Section 4, Tables 1-4] The tables report 95% confidence intervals over 30 random splits. The text should state whether these intervals account for within-split correlation between the model C-Index and the ECI bound; the current description is insufficient for a reader to judge the significance of the differences.
  4. [Section 3, Definition 9] The discrimination ratio DR subtracts 0.5 in both numerator and denominator; this is a valid rescaling only if E[CI(M*)] > 0.5, which is not proven for the censored case. Please add a condition or justification.
  5. [Appendix A.2, Predictor table] Table 6 lists 'Donor's Age' twice; one entry should be removed or replaced by the intended variable (e.g., donor age squared).
  6. [General] The paper contains several typographical errors (e.g., 'C-index' vs 'C-Index' inconsistently; 'shows present an upperbound' in the Conclusion; 'the narrow confidence interval show' in Section 4). A careful proofreading pass is needed.

Circularity Check

2 steps flagged · score 6.0 of 10

Empirical ECI bound is computed from an outcome-calibrated oracle model, so the headline 0.75 bound and the discrimination ratios are fitted quantities rather than independent PH upper bounds.

  1. fitted input called prediction [Section 3, paragraph beginning 'After the transfer of an unbiased baseline survival function...']
    "For Dret, we can calculate unbiased hazard rates h∗ i satisfying E[T ∗ i ] = Ti, i= 1, . . . , m. Being unbiased, the h∗ i can be used to form a prediction model M ∗ with survival functions S∗ i (t) = [S0(t)]h∗ i . Now, because M ∗ ij = 1 if Aij = 1 for every pair i, jin P ret, CI (M ∗) = 1 in Dret. Moreover, E[CI (M ∗)] is an upper bound for the E[CI ( ˆM )] attainable by any prediction model ˆM on Dret and Dpro."

    Proposition 2 bounds E[CI(ˆM)] by the expected C-Index of the true model M, whose hazards h_i are fixed population parameters. The empirical implementation instead defines h*_i by solving E[T*_i] = T_i, i.e., by calibrating each hazard to the observed outcome. Because E[T] decreases in the hazard, this forces M*'s predicted ordering to match the observed event order, which is why CI(M*)=1 on Dret. E[CI(M*)] is therefore a function of the realized T_i values, not an upper bound derived from the true PH model. In a homogeneous population the true bound is 0.5, yet this construction can produce values near 0.75, so the reported ECI and SUBECI bounds are fitted, outcome-dependent quantities, not independent upper limits.

  2. fitted input called prediction [Definition 9 and Abstract]
    "DR( ˆM , M∗) = CI ( ˆM ) − 0.5 / E[CI (M ∗)] − 0.5 . (23) ... With an ECI of 0.75 for 10-year survival, the new upperbound is well below 1. A C-Index performance around 0.61 or slightly above as commonly reported in literature and replicated in this study therefore closes almost half of the gap between the ECI and the 0.5 threshold."

    The discrimination ratio and the 'closes almost half the gap' claim normalize the observed C-Index by E[CI(M*)]-0.5. Since E[CI(M*)] is inflated by the outcome-calibrated construction of M*, the denominator is not the true-model bound from Proposition 2 but a fitted benchmark. The ratio CI(ˆM)-0.5 over an outcome-fit denominator does not measure how much of the PH-achievable discrimination is captured; it measures performance against an oracle that has seen the event times. Thus the headline numerical conclusion depends on the same fitted values used to define the benchmark.

full rationale

The theoretical contribution in Proposition 1 and Proposition 2 is not circular: the proof derives P(T_i > T_j) = h_j/(h_i + h_j) from the PH assumption and then shows that the true model's expected C-Index dominates any other model's expected C-Index purely from the ordering of hazards. That argument is self-contained and does not depend on the case study. The circularity enters in the empirical application. Section 3 replaces the true hazards h_i of Proposition 2 with h*_i chosen so that E[T*_i] equals each observed event time T_i. This is an outcome-calibrated oracle model, and the paper itself notes that CI(M*)=1 on Dret. Calling these hazards 'unbiased' does not make M* the true PH model; it makes M* a function of the realized outcomes over which the C-Index and its bound are computed. Consequently, the reported ECI of about 0.75, the SUBECI bounds, and the discrimination ratios are not upper bounds for models under the PH assumption but quantities forced by the h* construction. The ethnic-bias comparison of SUBCIs against these inflated SUBECI bounds inherits the same problem. The paper contains self-citations, but they are not the load-bearing source of circularity; the reduction is by construction in the empirical ECI. Score 6 reflects that the central empirical claim reduces to a fit, while the theorem itself remains independent.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The paper introduces new metrics (ECI, SUBCI, SUBECI, DR) but no physical entities. The key free parameters are the Cox model coefficients and baseline survival, which directly determine the reported upper bound. The correctness of the central claim depends on the proportional hazards assumption, independent censoring, correct model specification, and the asserted unbiasedness of the estimated hazard rates.

free parameters (4)
  • Cox model regression coefficients = Not reported individually
    The ECI and all C-Indices are computed from a Cox model with predictors KDPI, EPTS, HLA mismatch, and patient/donor characteristics (Table 6). These coefficients determine the fitted hazard rates h*_i used to compute the upper bound.
  • Baseline survival function S0(t) = Estimated nonparametrically via Breslow/Kaplan-Meier
    The baseline survival is estimated from Dret and is used to relate hazards to expected survival times and concordance probabilities.
  • Split proportion (0.5) = 0.5
    The 50/50 split is chosen to minimize the variance of the ECI across 30 simulations (Appendix A.3); it is a design choice that affects the estimated bounds.
  • Perturbation epsilon for ties = Small, unspecified
    Used to break ties in discretized survival times (Section 4); the paper states effects are negligible, but the value is not given.
assumptions (5)
  • domain assumption True hazard functions are proportional: h_i(t) = h_i * h_0(t)
    Equation (1) in Section 2.1; this is the core model assumption for all theoretical results.
  • domain assumption Independent right censoring
    Assumed so that Kaplan-Meier/Breslow estimators are unbiased (Section 3); required for the censor-adjusted expected C-Index.
  • ad hoc to paper Correct model specification: the Cox covariate set captures the true hazards
    The paper's estimated h*_i are treated as the true hazards when computing the upper bound, but the covariates KDPI, EPTS, and HLA mismatch are a standard but incomplete set; omitted predictors would make the bound invalid.
  • ad hoc to paper Unbiasedness of h*_i: E[T*_i] = T_i
    Section 3 asserts this without proof; if h*_i are calibrated to observed event times, the bound is overfit, and if they come from a misspecified Cox model, they are not the true hazards.
  • domain assumption Distinct event times
    Assumed for pairwise comparisons (Section 2.3); resolved in practice by small random perturbations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Equitable Discrimination in Survival Prediction: The Maximum Expected C-Index." pith.science (2026). https://pith.science/paper/G22HOEQE

@misc{pith2026250605592,
  author       = {Pith},
  title        = {Pith review of: Equitable Discrimination in Survival Prediction: The Maximum Expected C-Index},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/G22HOEQE}},
  note         = {Machine review of arXiv:2506.05592}
}
read the original abstract

The C-Index measures the discrimination performance of survival prediction models. C-Index scores are often well below the upperbound of 1 that represents perfect prediction and closer to 0.5 as achieved by random prediction. Our first objective is to provide a tighter C-Index upperbound for proportional hazards models. Our second research objective is to measure discrimination performance for subpopulations, also relative to subpopulation specific upperbounds. We present the expected C-Index (ECI) as a tight upperbound for proportional hazards models. Moreover, we define the subpopulation C-Index (SUBCI) and a sub-population specific expected C-Index (SUBECI). The metrics are applied to predict death censored graft survival (DCGF) after deceased donor kidney transplant in the US with a Cox model using standard donor (KDPI), patient (EPTS), and (Class 1) mismatch predictors. With an ECI of 0.75 for 10-year survival, the new upperbound is well below 1. A C-Index performance around 0.61 or slightly above as commonly reported in literature and replicated in this study therefore closes almost half of the gap between the ECI and the 0.5 threshold. SUBECIs don't vary significantly from the overall ECI but there are substantial and significant differences among the SUBCIs. Extending this upperbound and C-Index to subpopulations enables to identify differences in discrimination upperbounds across subpopulations and in prediction model biases. A standard Cox model for DCGF in the US can be ethnically biased.

Figures

Figures reproduced from arXiv: 2506.05592 by the authors.

Figure 1
Figure 1. The standard deviation of the expected C-index over 30 simulations using different [PITH_FULL_IMAGE:figures/full_fig_p032_1.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

46 extracted references · 44 canonical work pages

  1. [1]

    Regression models and life-tables

    David R Cox. Regression models and life-tables. Journal of the Royal Statistical Society: Series B (Methodological), 34(2):187–202, 1972

    1972

  2. [2]

    Regression modelling strategies for improved prognostic prediction

    Frank E Harrell Jr, Kerry L Lee, Robert M Califf, David B Pryor, and Robert A Rosati. Regression modelling strategies for improved prognostic prediction. Statistics in medicine, 3(2):143–152, 1984

    work page 1984

  3. [3]

    Time to default in credit scoring using survival analysis: a benchmark study

    Lore Dirick, Gerda Claeskens, and Bart Baesens. Time to default in credit scoring using survival analysis: a benchmark study. Journal of the Operational Research Society , 68(6):652–665, 2017

    work page 2017

  4. [4]

    Prognostic modeling of predictive maintenance with survival analysis for mobile work equipment

    Zhen Yang, Juho Kanniainen, Tomi Krogerus, and Frank Emmert-Streib. Prognostic modeling of predictive maintenance with survival analysis for mobile work equipment. Scientific Reports, 12(1):8529, 2022

    work page 2022

  5. [5]

    Trans- parent reporting of a multivariable prediction model for individual prognosis or diagnosis (tripod) the tripod statement

    Gary S Collins, Johannes B Reitsma, Douglas G Altman, and Karel GM Moons. Trans- parent reporting of a multivariable prediction model for individual prognosis or diagnosis (tripod) the tripod statement. Circulation, 131(2):211–219, 2015

    work page 2015

  6. [6]

    Probast: a tool to assess the risk of bias and applicability of prediction model studies

    Robert F Wolff, Karel GM Moons, Richard D Riley, Penny F Whiting, Marie Westwood, Gary S Collins, Johannes B Reitsma, Jos Kleijnen, Sue Mallett, and PROBAST Group†. Probast: a tool to assess the risk of bias and applicability of prediction model studies. Annals of internal medicine , 170(1):51–58, 2019

    work page 2019

  7. [7]

    Machine learning for survival analysis: A survey

    Ping Wang, Yan Li, and Chandan K Reddy. Machine learning for survival analysis: A survey. ACM Computing Surveys (CSUR) , 51(6):1–36, 2019

    work page 2019

  8. [8]

    Deep learning for survival analysis: a review

    Simon Wiegrebe, Philipp Kopper, Raphael Sonabend, Bernd Bischl, and Andreas Ben- der. Deep learning for survival analysis: a review. Artificial Intelligence Review , 57(3):65, 2024

    work page 2024

Show all 46 references
  1. [9]

    The comparative performance of models predicting patient and graft survival after kidney transplantation: A systematic review

    Joris van de Klundert, Francisco Perez-Galarce, Marcelo Olivares, Liset Pengel, and Annelies de Weerd. The comparative performance of models predicting patient and graft survival after kidney transplantation: A systematic review. Transplantation Reviews, page 100934, 2025

    2025

  2. [10]

    Verification of forecasts expressed in terms of probability

    Glenn W Brier. Verification of forecasts expressed in terms of probability. Monthly weather review, 78(1):1–3, 1950. 21

    1950

  3. [11]

    Assessment and comparison of prognostic classification schemes for survival data

    Erika Graf, Claudia Schmoor, Willi Sauerbrei, and Martin Schumacher. Assessment and comparison of prognostic classification schemes for survival data. Statistics in medicine, 18(17-18):2529–2545, 1999

    1999

  4. [12]

    Signal detection theory and psychophysics , volume 1

    David Marvin Green, John A Swets, et al. Signal detection theory and psychophysics , volume 1. Wiley New York, 1966

    1966

  5. [13]

    Receiver-operating characteristic (roc) plots: a fundamental evaluation tool in clinical medicine

    Mark H Zweig and Gregory Campbell. Receiver-operating characteristic (roc) plots: a fundamental evaluation tool in clinical medicine. Clinical chemistry, 39(4):561–577, 1993

    1993

  6. [14]

    Assessing the performance of prediction models: a framework for traditional and novel measures

    Ewout W Steyerberg, Andrew J Vickers, Nancy R Cook, Thomas Gerds, Mithat Go- nen, Nancy Obuchowski, Michael J Pencina, and Michael W Kattan. Assessing the performance of prediction models: a framework for traditional and novel measures. Epidemiology, 21(1):128–138, 2010

    2010

  7. [15]

    A com- prehensive risk quantification score for deceased donor kidneys: the kidney donor risk index

    Panduranga S Rao, Douglas E Schaubel, Mary K Guidinger, Kenneth A Andreoni, Robert A Wolfe, Robert M Merion, Friedrich K Port, and Randall S Sung. A com- prehensive risk quantification score for deceased donor kidneys: the kidney donor risk index. Transplantation, 88(2):231–236, 2009

    2009

  8. [16]

    Nicholas Hartman, Sehee Kim, Kevin He, and John D Kalbfleisch. c. Statistics in medicine, 42(13):2179–2190, 2023

    2023

  9. [17]

    Machine learning to predict transplant outcomes: helpful or hype? a national cohort study

    Sunjae Bae, Allan B Massie, Brian S Caffo, Kyle R Jackson, and Dorry L Segev. Machine learning to predict transplant outcomes: helpful or hype? a national cohort study. Transplant International, 33(11):1472–1480, 2020

    2020

  10. [18]

    Large-scale benchmark study of survival prediction methods using multi- omics data

    Moritz Herrmann, Philipp Probst, Roman Hornung, Vindi Jurinovic, and Anne-Laure Boulesteix. Large-scale benchmark study of survival prediction methods using multi- omics data. Briefings in bioinformatics , 22(3):bbaa167, 2021

    2021

  11. [19]

    Ensuring fairness in machine learning to advance health equity.Annals of internal medicine, 169(12):866–872, 2018

    Alvin Rajkomar, Michaela Hardt, Michael D Howell, Greg Corrado, and Marshall H Chin. Ensuring fairness in machine learning to advance health equity.Annals of internal medicine, 169(12):866–872, 2018

    2018

  12. [20]

    Gianfrancesco, Suzanne Tamang, Jinoos Yazdany, and Gabriela Schmajuk

    Milena A. Gianfrancesco, Suzanne Tamang, Jinoos Yazdany, and Gabriela Schmajuk. Potential Biases in Machine Learning Algorithms Using Electronic Health Record Data. JAMA Internal Medicine , 178(11):1544–1547, 11 2018. 22

    2018

  13. [21]

    Risk of bias in studies on prediction models developed using supervised machine learning techniques: systematic review

    Constanza L Andaur Navarro, Johanna AA Damen, Toshihiko Takada, Steven WJ Nij- man, Paula Dhiman, Jie Ma, Gary S Collins, Ram Bajpai, Richard D Riley, Karel GM Moons, et al. Risk of bias in studies on prediction models developed using supervised machine learning techniques: sy...

    2021

  14. [22]

    Predictably unequal: understanding and address- ing concerns that algorithmic clinical prediction may increase health disparities

    Jessica K Paulus and David M Kent. Predictably unequal: understanding and address- ing concerns that algorithmic clinical prediction may increase health disparities. NPJ digital medicine, 3(1):99, 2020

    2020

  15. [23]

    The effectiveness, equity and explainability of health service resource allocation-with applications in kidney transplantation & family planning

    Joris Van De Klundert, Harwin De Vries, Francisco P´ erez Galarce, Nieves Valdes, and Felipe Simon. The effectiveness, equity and explainability of health service resource allocation-with applications in kidney transplantation & family planning. Frontiers in Health Services, 5...

    2025

  16. [24]

    Statistical modeling: Regression, survival analysis, and time series analysis

    Lawrence Leemis. Statistical modeling: Regression, survival analysis, and time series analysis. N/A, 2023

    2023

  17. [25]

    Empirical likelihood in survival analysis

    Gang Li, Runze Li, and Mai Zhou. Empirical likelihood in survival analysis. In Contem- porary Multivariate Analysis And Design Of Experiments: In Celebration of Professor Kai-Tai Fang’s 65th Birthday , pages 337–349. World Scientific, 2005

    2005

  18. [26]

    Statistical models based on counting processes

    Per K Andersen, Ornulf Borgan, Richard D Gill, and Niels Keiding. Statistical models based on counting processes. Springer Science & Business Media, 2012

    2012

  19. [27]

    Two-sided bias bound of the kaplan-meier estimator

    M Zhou. Two-sided bias bound of the kaplan-meier estimator. Probability theory and related fields, 79(2):165–173, 1988

    1988

  20. [28]

    The bias of kaplan-meier integrals

    Winfried Stute. The bias of kaplan-meier integrals. Scandinavian Journal of Statistics , pages 475–484, 1994

    1994

  21. [29]

    The kidney allocation system

    John J Friedewald, Ciara J Samana, Bertram L Kasiske, Ajay K Israni, Darren Stewart, Wida Cherikh, and Richard N Formica. The kidney allocation system. Surgical Clinics, 93(6):1395–1406, 2013

    2013

  22. [30]

    A guide to calculating and interpreting the estimated post-transplant survival (epts) score used in the kidney allocation system (kas)

    OPTN. A guide to calculating and interpreting the estimated post-transplant survival (epts) score used in the kidney allocation system (kas). Kidney, 2, 2012

    2012

  23. [31]

    A guide to calculating and interpreting the kidney donor profile index (kdpi)

    OPTN. A guide to calculating and interpreting the kidney donor profile index (kdpi). Accesed at https://optnpilot. unos. org/media/1512/guide to calculating interpreting kdpi. pdf on , 8, 2018. 23

    2018

  24. [32]

    Long-term kidney transplant graft survival—making progress when most needed

    Emilio D Poggio, Joshua J Augustine, Susana Arrigain, Daniel C Brennan, and Jesse D Schold. Long-term kidney transplant graft survival—making progress when most needed. American journal of transplantation , 21(8):2824–2832, 2021

    2021

  25. [33]

    The statistical sign test

    Wilfrid J Dixon and Alexander M Mood. The statistical sign test. Journal of the American Statistical Association, 41(236):557–566, 1946

    1946

  26. [34]

    On a test of whether one of two random variables is stochastically larger than the other

    Henry B Mann and Donald R Whitney. On a test of whether one of two random variables is stochastically larger than the other. The annals of mathematical statistics , pages 50–60, 1947

    1947

  27. [35]

    Multical- ibration: Calibration for the (computationally-identifiable) masses

    Ursula H´ ebert-Johnson, Michael Kim, Omer Reingold, and Guy Rothblum. Multical- ibration: Calibration for the (computationally-identifiable) masses. In International Conference on Machine Learning , pages 1939–1948. PMLR, 2018

    1939

  28. [36]

    Prediction performance and fairness heterogeneity in cardiovascular risk models

    Uri Kartoun, Shaan Khurshid, Bum Chul Kwon, Aniruddh P Patel, Puneet Batra, Anthony Philippakis, Amit V Khera, Patrick T Ellinor, Steven A Lubitz, and Kenney Ng. Prediction performance and fairness heterogeneity in cardiovascular risk models. Scientific Reports, 12(1):12542, 2022

    2022

  29. [37]

    Bias and unfairness in machine learning models: a systematic review on datasets, tools, fairness metrics, and identification and mitigation methods

    Tiago P Pagano, Rafael B Loureiro, Fernanda VN Lisboa, Rodrigo M Peixoto, Guil- herme AS Guimar˜ aes, Gustavo OR Cruz, Maira M Araujo, Lucas L Santos, Marco AS Cruz, Ewerton LS Oliveira, et al. Bias and unfairness in machine learning models: a systematic review on datasets, to...

    2023

  30. [38]

    Fairness in predicting cancer mortality across racial subgroups

    Teja Ganta, Arash Kia, Prathamesh Parchure, Min-heng Wang, Melanie Besculides, Madhu Mazumdar, and Cardinale B Smith. Fairness in predicting cancer mortality across racial subgroups. JAMA Network Open , 7(7):e2421290–e2421290, 2024

    2024

  31. [39]

    Effect of human leukocyte antigen compatibility on kidney graft survival: comparative analysis of two decades

    Gerhard Opelz and Bernd D¨ ohler. Effect of human leukocyte antigen compatibility on kidney graft survival: comparative analysis of two decades. Transplantation, 84(2):137– 143, 2007

    2007

  32. [40]

    Health disparities and health equity: concepts and measurement

    Paula Braveman. Health disparities and health equity: concepts and measurement. Annual review of public health , 27(1):167–194, 2006

    2006

  33. [41]

    What is health equity? Behavioral science & policy , 4(1):1–14, 2018

    Paula Braveman, Elaine Arkin, Tracy Orleans, Dwayne Proctor, Julia Acker, and Alonzo Plough. What is health equity? Behavioral science & policy , 4(1):1–14, 2018. 24

    2018

  34. [42]

    Eliminating transplant waiting time inequities–with an application to kidney allocation in the usa

    Joris Van de Klundert, Liana van der Hagen, and Aniek Markus. Eliminating transplant waiting time inequities–with an application to kidney allocation in the usa. European Journal of Operational Research, 297(3):977–985, 2022

    2022

  35. [43]

    Optn/srtr 2022 annual data report: kidney

    Krista L Lentine, Jodi M Smith, Grace R Lyden, Jonathan M Miller, Thomas G Dolan, Keighly Bradbrook, Lindsay Larkin, Kayla Temple, Dzhuliyana K Handarova, Saman- tha Weiss, et al. Optn/srtr 2022 annual data report: kidney. American Journal of Transplantation, 24(2):S19–S118, 2024

    2022

  36. [44]

    On ranking in survival analysis: Bounds on the concordance index

    Harald Steck, Balaji Krishnapuram, Cary Dehing-Oberije, Philippe Lambin, and Vikas C Raykar. On ranking in survival analysis: Bounds on the concordance index. Advances in neural information processing systems , 20, 2007

    2007

  37. [45]

    On the c-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data

    Hajime Uno, Tianxi Cai, Michael J Pencina, Ralph B D’Agostino, and Lee-Jen Wei. On the c-statistics for evaluating overall adequacy of risk prediction procedures with censored survival data. Statistics in medicine , 30(10):1105–1117, 2011

    2011

  38. [46]

    A time-dependent discrimination index for survival data

    Laura Antolini, Patrizia Boracchi, and Elia Biganzoli. A time-dependent discrimination index for survival data. Statistics in medicine , 24(24):3927–3944, 2005. 25 9 Tables Table 1: Expected C-Index vs actual C-Index for multiple years of transplants. The numbers in parenthese...

    2005

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.