Pith. sign in

REVIEW 2 major objections 6 minor 47 references

Isotonic Bradley-Terry Model for Paired Comparison Data

T0 review · 2 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read The isotonic Bradley-Terry model learns the link function from data instead of fixing it, and by alternating that with player-strength updates it guarantees monotone training-error improvement and more accurate win predictions and rankings

desk verdict New alternating scheme for learning the Bradley-Terry link deserves a referee, but the monotonicity guarantee is asserted rather than proven and the ranking gains lean on tie-abstention. read the letter →

arxiv 2608.02081 v1 pith:JEBK333L submitted 2026-08-03 cs.LG

Ryoya Yamasaki This is my paper
classification cs.LG
keywords isotonicregressionBradley-Terrymodelpairedcomparisondatawinprobabilitypredictionrankinginverselinkfunctionmisspecificationpool-adjacent-violatorsalgorithm
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

The paper proposes the isotonic Bradley-Terry model, which estimates the win-probability link function by isotonic regression while updating player strength parameters by subgradient descent, and alternates the two. The central claim is that this procedure never increases training error, can output an exact tie when data are too sparse to separate two players, and improves test win-probability prediction and ranking over the standard fixed-logistic Bradley-Terry model. The sympathetic reader cares because the approach removes a common model-misspecification assumption while keeping the interpretable real-valued strength parameters. Synthetic experiments and real-world data from football, baseball, and tennis support the improvement, especially when training data are scarce or the true link function differs from the logistic.

What carries the argument

The key machinery is the alternating optimization: rate parameters are updated by a subgradient method with line search (the paper notes this step may be non-convex), and the inverse link function is updated by isotonic regression solved with the pool-adjacent-violators (PAV) algorithm. The PAV solution is a nondecreasing, skew-symmetric polyline that satisfies the order constraints on win probabilities implied by the current rate differences. This learned link, written as σ̂(u), then feeds back into the rate update, and the cycle repeats.

What would settle it

Construct a dataset where the alternating procedure's training error increases between two consecutive iterations, or show a sequence of rate parameters and link functions that oscillates without converging. A simple check is to run the algorithm on any standard paired-comparison benchmark and print the training loss after each alternation; any observed increase contradicts the monotone-improvement guarantee.

Watch

Extended reading notes

Core claim

On the paper's own terms, the inverse link function in a Bradley-Terry model does not have to be fixed in advance. By replacing it with a monotone polyline learned from data via the pool-adjacent-violators algorithm and alternating that step with (sub-)gradient updates of the rate parameters, the model attains a monotonic decrease in training error and tends to declare a tie (predicted win probability 0.5) when the data cannot establish a strict ranking. This directly addresses the misspecification risk of a pre-specified link function while retaining the real-valued representation of player strength. The paper also recommends ranking players by Borda count rather than raw rate parameters, s

Load-bearing premise

The load-bearing premise is that after the first isotonic update, the objective becomes non-convex and non-smooth, yet a subgradient method with line search can always be steered to a point that does not increase training error—a claim stated without a formal proof or explicit condition.

Editorial extensions

If this is right

  • Training error is guaranteed to be non-increasing across alternations, as long as the subgradient line-search step succeeds.
  • When training data for a pair are insufficient, the model is likely to output a predicted win probability of exactly 0.5, effectively withholding a strict ranking decision.
  • Ranking should be done through Borda count rather than the rate parameters themselves, because the learned link may be flat on some intervals.
  • Experiments on English Premier League, MLB, and ATP data show improved win-probability prediction (squared loss) and higher Kendall's Tau in most settings, particularly with sparse training data.
  • With negative log-likelihood loss, test win-probability evaluation often produces NaN values, so the model is unsuitable for that loss when predicting win probabilities directly, though ranking still improves.

Reading between the lines

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

  • The success of the method suggests that a fully nonparametric link function can absorb some of the misspecification that normally requires heavier models, so extending the same alternating idea to Elo-style time-dependent strengths or to covariate-adjusted rates is a natural next step.
  • The tie behavior could be reinterpreted as a data-driven confidence signal: a flat portion of the learned link flags pairs whose ordering is not supported by the data, which may be useful for active learning or match scheduling.
  • Because the paper leaves the number of alternations to cross-validation, a stopping rule based on validation error or on a change-point in the link shape could be investigated empirically.
  • A concrete test of the paper's central guarantee would be to monitor the training-error trajectory on a benchmark dataset and check that no iteration increases it; if an increase is observed, the subgradient line-search assumption is violated.
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

2 major / 6 minor

Summary. The paper proposes an isotonic Bradley-Terry (IBT) model for paired comparison data, which alternates between (sub)gradient updates of player rate parameters and isotonic-regression (PAV) updates of the inverse link function. The authors argue that conventional fixed-link BT models suffer from misspecification, and that learning the link nonparametrically can improve win-probability prediction and ranking. The main theoretical claims are that the isotonic link update is globally optimal for a given rate vector (Theorems 1 and 2) and that the alternating procedure guarantees monotonic improvement in training error. Numerical experiments with synthetic data and real-world football, baseball, and tennis data compare the IBT model with a logistic-link BT baseline, using squared loss and NLL loss (in the appendix), with the number of alternating updates selected by 10-fold cross-validation.

Significance. If the monotonicity guarantee is rigorously established and the ranking gains are not an artifact of tie-abstention, the IBT model is a useful contribution: it provides a flexible, low-dimensional nonparametric link for paired comparisons while retaining interpretable player strengths and an automatic mechanism for declaring ties when data are scarce. The paper is commendable for releasing the experimental code, for running extensive synthetic and real-world experiments, and for candidly disclosing the NLL-loss NaN issue in the appendix. However, two load-bearing points need attention: the unproven claim that a subgradient line search can always avoid increasing the training error under a non-convex, non-smooth objective, and the mismatch between the recommended Borda-count ranking and the Kendall-tau evaluation actually used in the experiments.

major comments (2)
  1. [Section 3 (after Theorem 2)] The statement that 'we can employ a sub-gradient method with line search of the step size to update the rate parameters without increasing the training error' is load-bearing for the abstract's guarantee of monotonic improvement in training error, but no proof or formal condition is supplied. After the first isotonic update, the objective (2) with σ=σ^[t] is generally non-convex and non-smooth; a subgradient is not necessarily a descent direction, and a line search may fail to find a non-increasing step without additional assumptions. Please provide a theorem with explicit conditions under which such a step exists, or weaken the guarantee to 'the isotonic update does not increase the training error; the rate-parameter update is heuristic.'
  2. [Section 5.1, Eq. (4); Section 3 (Borda count)] Ranking performance is evaluated only via Kendall's Tau on predicted pairwise win probabilities, where tied predictions (σ=0.5) are counted as neither concordant nor discordant and therefore do not penalize the score. This metric can reward tie-abstention on difficult pairs. Section 3, however, recommends ranking by Borda count for the IBT model, yet no experiment evaluates Borda-count ordering. The empirical claim of improved ranking is therefore not established as improvement in the recommended ranking; please add Borda-count-based evaluation or justify the metric choice.
minor comments (6)
  1. [Section 3 (after Eq. (5))] Typographical errors: 'relay on' should be 'rely on'; 'For the the player ranking' has a doubled article.
  2. [Abstract / Conclusion] The claim that the model 'improves win probability prediction' should be qualified as demonstrated for the squared loss. With NLL loss, the test error frequently becomes NaN, as the authors themselves acknowledge in Section 3 and Appendix B.
  3. [Sections 5.1 and 5.2] The 10-fold CV procedure for selecting the number of alternating updates t is not specified in detail (e.g., folds over player pairs or over individual comparisons). The step-size line search for the (sub)gradient updates is also not described; please provide enough detail for reproducibility.
  4. [Section 3, Borda-count discussion] The implication notation '⇐' appears to be a typo for '⇏' in the discussion of non-strictly increasing σ; please clarify the direction of the implication.
  5. [Theorem 1] The assumption that φ is strictly convex in u at every v is not satisfied by φ_nll when v=0, although the paired objective in Algorithm 1, φ(u,v)+φ(1-u,1-v), is strictly convex for all v∈[0,1]. Please restate the assumption for the paired objective to avoid a technical inaccuracy.
  6. [Tables 1, B1, and B2] The 'meanstd' column headers are typeset without separation, making the numbers difficult to parse. Please separate mean and standard deviation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: training-loss monotonicity is a design property, and test improvements are evaluated on held-out pairs; the only self-citation is not load-bearing.

full rationale

The derivation chain is self-contained. Section 3 alternates two block-minimization steps: rate parameters are updated for a fixed inverse link by (sub-)gradient with line search, and the inverse link is updated by isotonic regression for fixed rates. Each step minimizes the same empirical training loss, so monotonic non-increase of training error is a property of the algorithm's design, not a prediction used to validate the model. All claimed predictive and ranking improvements are evaluated on held-out test pairs D_tes using test criteria (3) and (4), with the number of alternating updates selected by 10-fold cross-validation; no fitted constant is renamed as a prediction. The Borda-count recommendation in Section 3 cites external work [36] and is not used in the experiments, so it does not smuggle in the result. The comparison with the nonparametric Bradley-Terry model in Section 4 is a substantive model distinction, not a renaming. The only self-citation ([27], for robust losses) is auxiliary and does not support any central premise. Unproved assertions such as the existence of a non-increasing sub-gradient line-search step in the non-convex regime and the possible inflation of Kendall's Tau by tie-abstention are correctness or validity concerns, not circularity. Accordingly, no circular step can be exhibited with the required reduction to inputs by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claim relies on standard isotonic regression theory and on the stochastic transitivity assumption that win probabilities are monotone in latent strength differences. The only tunable hyperparameter is the number of alternating updates, chosen by cross-validation. No new physical or conceptual entities are introduced.

free parameters (1)
  • number of alternating updates t = selected by 10-fold cross-validation in Procedure 2; fixed over 1..10 in Procedure 1
    Model behavior depends on t; excessive updates cause overfitting. This is a standard tuning hyperparameter, not a fitted physical constant.
assumptions (3)
  • standard math The isotonic regression problem (6) is convex and separable under strictly convex loss, so PAV finds the global optimum.
    Theorem 1 relies on convexity and separability, citing Best and Chakravarti and Barlow et al.; this is standard isotonic regression theory.
  • domain assumption Stochastic transitivity: true win probabilities are a non-decreasing function of the difference in latent player strengths.
    The entire model class Sigma and the isotonic constraints in (6) rest on this assumption, introduced in Sections 2 and 3. If false, the model is misspecified.
  • ad hoc to paper At each rate-parameter update, a line search can find a step that does not increase the training error under the non-convex and non-smooth objective.
    The monotone improvement guarantee is asserted in the abstract and Section 3 without a formal proof; this premise is what makes the alternating procedure monotone.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Isotonic Bradley-Terry Model for Paired Comparison Data." pith.science (2026). https://pith.science/paper/JEBK333L

@misc{pith2026260802081,
  author       = {Pith},
  title        = {Pith review of: Isotonic Bradley-Terry Model for Paired Comparison Data},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JEBK333L}},
  note         = {Machine review of arXiv:2608.02081}
}
read the original abstract

In this paper, we study prediction problems for paired comparison data, for example, predicting the win probability between two unmatched players and ranking all the players according to the order of their strengths by using win probability data between two matched players. Paired comparison data are typically analyzed using Bradley-Terry and Thurstone-Mosteller models. These models predict the win probability by transforming the difference between learned rate parameters, which represent players'\;strengths, with a pre-specified inverse link function, and employ the order of learned rate parameters for player ranking. However, these models may suffer from model misspecification owing to the selection of a fixed inverse link function. Therefore, in this study, we propose to learn the rate parameters by a (sub-)gradient method and the inverse link function by an isotonic regression technique alternately. The proposed model guarantees monotonic improvement in training error, and is likely to yield an exact tie when the available data is insufficient to establish a strict ranking. We also verified that the proposed model could improve the win probability prediction and ranking performance through numerical experiments with synthetic data and real-world data of football Premier League, baseball MLB, and tennis ATP tour.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 14 canonical work pages

  1. [1]

    Biometrics32(2), 213– 239 (1976) https://doi.org/10.2307/2529494

    Bradley, R.A.: Science, statistics, and paired comparisons. Biometrics32(2), 213– 239 (1976) https://doi.org/10.2307/2529494

    work page doi:10.2307/2529494 1976

  2. [2]

    Statistical Science27(3), 412–433 (2012) https://doi.org/10

    Cattelan, M.: Models for Paired Comparison Data: A Review with Emphasis on Dependent Data. Statistical Science27(3), 412–433 (2012) https://doi.org/10. 1214/12-sts396

    2012

  3. [3]

    Acta Psychologica104(2), 145–166 (2000) https://doi.org/ 10.1016/s0001-6918(00)00018-4

    Kissler, J., B¨ auml, K.-H.: Effects of the beholder’s age on the perception of facial attractiveness. Acta Psychologica104(2), 145–166 (2000) https://doi.org/ 10.1016/s0001-6918(00)00018-4

    work page doi:10.1016/s0001-6918(00)00018-4 2000

  4. [4]

    Medical Care46(4), 346–348 (2008) https: //doi.org/10.1097/mlr.0b013e31816dd8d9

    Maydeu-Olivares, A., B¨ ockenholt, U.: Modeling subjective health outcomes: Top 10 reasons to use Thurstone’s method. Medical Care46(4), 346–348 (2008) https: //doi.org/10.1097/mlr.0b013e31816dd8d9

    work page doi:10.1097/mlr.0b013e31816dd8d9 2008

  5. [5]

    Reliability Engineering & System Safety93(5), 722–731 (2008) https://doi.org/10.1016/j

    Mazzuchi, T.A., Linzey, W.G., Bruning, A.: A paired comparison experiment for gathering expert judgment for an aircraft wiring risk assessment. Reliability Engineering & System Safety93(5), 722–731 (2008) https://doi.org/10.1016/j. ress.2007.03.011

    doi:10.1016/j 2008

  6. [6]

    In: Advances in Neural Information Processing Systems, vol

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C.D., Ermon, S., Finn, C.: Direct preference optimization: Your language model is secretly a reward model. In: Advances in Neural Information Processing Systems, vol. 36, pp. 53728–53741 (2023). https://doi.org/10.52202/075280-2338

    doi:10.52202/075280-2338 2023

  7. [7]

    In: Proceedings of the 41st International Conference on Machine Learning (2024)

    Chiang, W.-L., Zheng, L., Sheng, Y., Angelopoulos, A.N., Li, T., Li, D., Zhu, B., Zhang, H., Jordan, M.I., Gonzalez, J.E., Stoica, I.: Chatbot arena: An open platform for evaluating llms by human preference. In: Proceedings of the 41st International Conference on Machine Learning (2024)

    2024

  8. [8]

    Zermelo, E.: Die Berechnung der Turnier-Ergebnisse als ein Maximumproblem der Wahrscheinlichkeitsrechnung. Mathematische Zeitschrift29, 436–460 (1929) https://doi.org/10.1007/bf01180541 17 𝑁=1,|𝐷 tra|:|𝐷 tes|= 𝑁=5,|𝐷 tra|:|𝐷 tes|= 𝑁=25,|𝐷 tra|:|𝐷 tes|= 1 : 9 5 : 5 9 : 1 1 : 9 5 : 5 9 : 1 1 : 9 5 : 5 9 : 1 (3),𝑛= 25 100 400 (4),𝑛= 25 100 400 (9),𝑛= 25 100 ...

    arXiv 1929

Show all 47 references
  1. [9]

    The American Mathematical Monthly64(8, Part 2), 28–33 (1957) https://doi.org/10

    Ford Jr., L.R.: Solution of a ranking problem from binary comparisons. The American Mathematical Monthly64(8, Part 2), 28–33 (1957) https://doi.org/10. 2307/2308513

    1957

  2. [10]

    The method of paired comparisons

    Bradley, R.A., Terry, M.E.: Rank analysis of incomplete block designs: I. The method of paired comparisons. Biometrika39(3/4), 324–345 (1952) https://doi. org/10.2307/2527550

    1952 doi

  3. [11]

    Luce, R.D.: Individual Choice Behavior vol. 4. John Wiley & Sons, New York (1959)

    1959

  4. [12]

    Psychological Review34(4), 273–286 (1927) https://doi.org/10.1037/h0070288

    Thurstone, L.L.: A law of comparative judgment. Psychological Review34(4), 273–286 (1927) https://doi.org/10.1037/h0070288

    1927 doi

  5. [13]

    The least squares solution assuming equal standard deviations and equal correlations

    Mosteller, F.: Remarks on the method of paired comparisons: I. The least squares solution assuming equal standard deviations and equal correlations. Psychometrika16(1), 3–9 (1951) https://doi.org/10.1007/978-0-387-44956-2 8

    1951 doi

  6. [14]

    Biometrika77(2), 265–273 (1990) https://doi.org/10.2307/2336804

    Stern, H.: A continuum of paired comparisons models. Biometrika77(2), 265–273 (1990) https://doi.org/10.2307/2336804

    1990 doi

  7. [16]

    The Annals of Mathematical Statistics26(4), 607–616 (1955) https://doi.org/10.1214/aoms/ 1177728420

    Brunk, H.D.: Maximum likelihood estimates of monotone parameters. The Annals of Mathematical Statistics26(4), 607–616 (1955) https://doi.org/10.1214/aoms/ 1177728420

    1955 doi

  8. [17]

    Pro- ceedings of the Koninklijke Nederlandse Akademie van Wetenschappen, Seriese A59(4), 444–455 (1956) https://doi.org/10.1016/s1385-7258(56)50060-1

    Van Eeden, C.: Maximum likelihood estimation of ordered probabilities. Pro- ceedings of the Koninklijke Nederlandse Akademie van Wetenschappen, Seriese A59(4), 444–455 (1956) https://doi.org/10.1016/s1385-7258(56)50060-1

    1956 doi

  9. [18]

    John Wiley & Sons, London and New York (1972)

    Barlow, R.E., Bartholomew, D.J., Bremner, J.M., Brunk, H.D.: Statistical Inference Under Order Restrictions: The Theory and Application of Isotonic Regression. John Wiley & Sons, London and New York (1972)

    1972

  10. [19]

    John Wiley & Sons, Chichester (1988)

    Robertson, T., Wright, F.T., Dykstra, R.L.: Order Restricted Statistical Infer- ence. John Wiley & Sons, Chichester (1988)

    1988

  11. [20]

    The Annals of Statistics43(1), 177–214 (2015) https://doi.org/10.1214/14-aos1272

    Chatterjee, S.: Matrix estimation by universal singular value thresholding. The Annals of Statistics43(1), 177–214 (2015) https://doi.org/10.1214/14-aos1272

    2015 doi

  12. [21]

    In: Proceedings of the 33rd International Conference on Machine Learning, pp

    Shah, N., Balakrishnan, S., Guntuboyina, A., Wainwright, M.: Stochastically transitive models for pairwise comparisons: Statistical and computational issues. In: Proceedings of the 33rd International Conference on Machine Learning, pp. 26 11–20 (2016). https://doi.org/10.1109/...

    2016

  13. [22]

    Bernoulli24(2), 1072–1100 (2018) https://doi.org/10.3150/ 16-bej865

    Chatterjee, S., Guntuboyina, A., Sen, B.: On matrix estimation under mono- tonicity constraints. Bernoulli24(2), 1072–1100 (2018) https://doi.org/10.3150/ 16-bej865

    2018

  14. [23]

    Arco Publishing, Inc., New York (1978)

    Elo, A.E., Sloan, S.: The Rating of Chessplayers: Past and Present. Arco Publishing, Inc., New York (1978)

    1978

  15. [24]

    The Annals of Statistics32(1), 384–406 (2004) https://doi.org/10.1214/aos/1079120141

    Hunter, D.R.: MM algorithms for generalized Bradley-Terry models. The Annals of Statistics32(1), 384–406 (2004) https://doi.org/10.1214/aos/1079120141

    2004

  16. [25]

    Journal of the Royal Statistical Society Series A: Statistics in Society135(3), 370–384 (1972) https://doi.org/10.2307/2344614

    Nelder, J.A., Wedderburn, R.W.M.: Generalized linear models. Journal of the Royal Statistical Society Series A: Statistics in Society135(3), 370–384 (1972) https://doi.org/10.2307/2344614

    1972 doi

  17. [26]

    In: Robust Statistics, Data Analysis, and Computer Intensive Methods, pp

    Bianco, A.M., Yohai, V.J.: Robust estimation in the logistic regression model. In: Robust Statistics, Data Analysis, and Computer Intensive Methods, pp. 17–34 (1996). https://doi.org/10.1007/978-1-4612-2380-1 2

    1996 doi

  18. [27]

    arXiv preprint arXiv:2305.08501 (2023)

    Yamasaki, R., Tanaka, T.: Label Smoothing is Robustification against Model Misspecification. arXiv preprint arXiv:2305.08501 (2023)

    2023 arXiv

  19. [28]

    Biometrika30(1-2), 81–93 (1938) https://doi.org/10.2307/2332226

    Kendall, M.G.: A new measure of rank correlation. Biometrika30(1-2), 81–93 (1938) https://doi.org/10.2307/2332226

    1938 doi

  20. [29]

    The American Journal of Psychology15(1), 72–101 (1904) https://doi.org/10

    Spearman, C.: The Proof and Measurement of Association between Two Things. The American Journal of Psychology15(1), 72–101 (1904) https://doi.org/10. 1037/11491-005

    1904

  21. [30]

    Noether, G.E.: Why kendall tau? Teaching Statistics3(2), 41–43 (1981) https: //doi.org/10.1111/j.1467-9639.1981.tb00422.x

    1981

  22. [31]

    Pacific Journal of Mathematics7(1), 833–847 (1957) https: //doi.org/10.2140/pjm.1957.7.833

    Brunk, H.D., Ewing, G.M., Utz, W.R.: Minimizing integrals in certain classes of monotone functions. Pacific Journal of Mathematics7(1), 833–847 (1957) https: //doi.org/10.2140/pjm.1957.7.833

    1957 doi

  23. [32]

    The Annals of Mathematical Statistics33(1), 273–289 (1962) https://doi.org/10

    Thompson Jr., W.A.: The problem of negative estimates of variance components. The Annals of Mathematical Statistics33(1), 273–289 (1962) https://doi.org/10. 1214/aoms/1177704731

    1962

  24. [33]

    The Annals of Statistics11(2), 467–477 (1983) https://doi.org/10.1214/aos/1176346153

    Lee, C.-I.C.: The min-max algorithm and isotonic regression. The Annals of Statistics11(2), 467–477 (1983) https://doi.org/10.1214/aos/1176346153

    1983

  25. [34]

    Mathematical Programming47(1), 425–439 (1990) https: //doi.org/10.1007/bf01580873 27

    Best, M.J., Chakravarti, N.: Active set algorithms for isotonic regression; a unifying framework. Mathematical Programming47(1), 425–439 (1990) https: //doi.org/10.1007/bf01580873 27

    1990 doi

  26. [35]

    Journal of Statistical Software32(5), 1–24 (2010) https://doi.org/10.18637/jss.v032.i05

    Leeuw, J., Hornik, K., Mair, P.: Isotone optimization in R: Pool-adjacent-violators algorithm (PA V A) and active set methods. Journal of Statistical Software32(5), 1–24 (2010) https://doi.org/10.18637/jss.v032.i05

    2010 doi

  27. [36]

    Journal of Machine Learning Research18(199), 1–38 (2018)

    Shah, N.B., Wainwright, M.J.: Simple, robust and optimal ranking from pairwise comparisons. Journal of Machine Learning Research18(199), 1–38 (2018)

    2018

  28. [37]

    Annals of Applied Probability30(5), 2491–2515 (2020) https://doi.org/ 10.1214/20-aap1564

    Han, R., Ye, R., Tan, C., Chen, K.: Asymptotic theory of sparse Bradley–Terry model. Annals of Applied Probability30(5), 2491–2515 (2020) https://doi.org/ 10.1214/20-aap1564

    2020 doi

  29. [38]

    The Annals of Statistics27(3), 1041–1060 (1999) https://doi.org/10.1214/aos/1018031267

    Simons, G., Yao, Y.-C.: Asymptotics when the number of parameters tends to infinity in the Bradley-Terry model for paired comparisons. The Annals of Statistics27(3), 1041–1060 (1999) https://doi.org/10.1214/aos/1018031267

    1999

  30. [39]

    Journal of Machine Learning Research12, 2825–2830 (2011)

    Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O., Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A., Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, E.: Scikit-learn: Machine learning in Python. Journal of Mach...

    2011

  31. [40]

    Journal of the Royal Statistical Society Series C: Applied Statistics22(1), 59–68 (1973) https://doi.org/10.2307/2346303

    Springall, A.: Response surface fitting using a generalization of the Bradley- Terry paired comparison model. Journal of the Royal Statistical Society Series C: Applied Statistics22(1), 59–68 (1973) https://doi.org/10.2307/2346303

    1973 doi

  32. [41]

    Psychometrika58(2), 233–256 (1993) https://doi.org/10.1007/bf02294575

    De Soete, G., Winsberg, S.: A Thurstonian pairwise choice model with univariate and multivariate spline transformations. Psychometrika58(2), 233–256 (1993) https://doi.org/10.1007/bf02294575

    1993 doi

  33. [42]

    David, H.A.: The Method of Paired Comparisons vol. 12. Charles Griffin and Company Ltd., London (1963)

    1963

  34. [43]

    American Chess Journal 3(1), 59–102 (1995)

    Glickman, M.E.: A comprehensive guide to chess ratings. American Chess Journal 3(1), 59–102 (1995)

    1995

  35. [44]

    Biometrics16(1), 86–109 (1960) https://doi

    Glenn, W.A., David, H.A.: Ties in paired-comparison experiments using a mod- ified Thurstone-Mosteller model. Biometrics16(1), 86–109 (1960) https://doi. org/10.2307/2527957

    1960 doi

  36. [45]

    Journal of the American Statistical Association 62(317), 194–204 (1967) https://doi.org/10.2307/2285921

    Rao, P.V., Kupper, L.L.: Ties in paired-comparison experiments: A generaliza- tion of the Bradley-Terry model. Journal of the American Statistical Association 62(317), 194–204 (1967) https://doi.org/10.2307/2285921

    1967 doi

  37. [46]

    Journal of the American Statistical Association 65(329), 317–328 (1970) https://doi.org/10.1080/01621459.1970.10481082

    Davidson, R.R.: On extending the Bradley-Terry model to accommodate ties in paired comparison experiments. Journal of the American Statistical Association 65(329), 317–328 (1970) https://doi.org/10.1080/01621459.1970.10481082

    1970

  38. [47]

    Journal of the Royal Statistical 28 Society Series C: Applied Statistics24(2), 193–202 (1975) https://doi.org/10

    Plackett, R.L.: The analysis of permutations. Journal of the Royal Statistical 28 Society Series C: Applied Statistics24(2), 193–202 (1975) https://doi.org/10. 2307/2346567

    1975

  39. [48]

    CRC Press, Boca Raton (1996) 29

    Marden, J.I.: Analyzing and Modeling Rank Data. CRC Press, Boca Raton (1996) 29

    1996

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.