Tie-Augmented Bradley–Terry Model

• Tie-augmented Bradley–Terry models are probabilistic frameworks that extend classical paired comparisons by incorporating tie outcomes to better capture ambiguous decisions.
• Multiple parameterizations, such as the Rao–Kupper and Davidson models, provide distinct mechanisms to model tie propensities and discrimination in pairwise data.
• Efficient estimation methods—including MM, EM, and Bayesian approaches—enable scalable inference and robust performance in applications like sports analytics and preference learning.

The tie-augmented Bradley–Terry model generalizes the classical paired-comparison framework to account for the empirically observed phenomenon that some comparisons are legitimately ambiguous or indistinguishable, resulting in ties. This framework extends the two-outcome (win/loss) Bradley–Terry model to three (win, loss, tie), providing a more accurate and information-preserving probabilistic model for pairwise data arising in sports, psychometrics, policy evaluation, machine learning preference aggregation, and other domains. Multiple parameterizations exist—including the Rao–Kupper and Davidson models—each specifying different mechanisms for tie-propensity, and diverse estimation strategies ranging from MM/EM algorithms and Bayesian inference to quasi-likelihood U-statistic approaches.

1. Mathematical Formulation and Tie-Probability Parameterizations

The classical Bradley–Terry model posits for items $i$ and $j$ with strengths $\lambda_i$, $\lambda_j > 0$:

$$P(i \succ j) = \frac{\lambda_i}{\lambda_i + \lambda_j}$$

Admitting ties, several parameterizations are used:

Rao–Kupper model: Introduces a tie-parameter $\theta > 1$ governing discrimination: $$\begin{align*} P(i \succ j) &= \frac{\lambda_i}{\lambda_i + \theta\lambda_j} \ P(j \succ i) &= \frac{\lambda_j}{\theta\lambda_i + \lambda_j} \ P(i \equiv j) &= \frac{(\theta^2-1) \lambda_i\lambda_j}{(\lambda_i+\theta\lambda_j)(\theta\lambda_i+\lambda_j)} \end{align*}$$ (Seymour et al., 2024, Yan, 2014, Caron et al., 2010, Vojnovic et al., 2019)

Davidson model: Uses a geometric mean for tie propensity, with tie-parameter $\nu \geq 0$: $$\begin{align*} P(i \succ j) &= \frac{\lambda_i}{\lambda_i + \lambda_j + \nu\sqrt{\lambda_i\lambda_j}} \ P(j \succ i) &= \frac{\lambda_j}{\lambda_i + \lambda_j + \nu\sqrt{\lambda_i\lambda_j}} \ P(i \equiv j) &= \frac{\nu\sqrt{\lambda_i\lambda_j}}{\lambda_i+\lambda_j+\nu\sqrt{\lambda_i\lambda_j}} \end{align*}$$ (Yan, 2014, Chen et al., 2024)

BTT “symmetric tie” form (used in RLHF): Three unnormalized potentials, where tie propensity is parameterized via $2e^{(\theta_i+\theta_j)/2}$: $j$0 (Liu et al., 2024)

Further generalizations include strength-dependent tie-propensities and explicit modeling of order effects (such as home-field advantage), as in extensions for chess or sports (Glickman, 30 May 2025, Whelan et al., 2021).

2. Likelihoods, Identifiability, and Existence of Estimates

Given counts of wins, losses, and ties $j$1 for unordered pairs, the log-likelihood (Rao–Kupper) is

$j$2

Identifiability is ensured by constraints such as $j$3 or fixing one $j$4. For Davidson-type models, $j$5 must be determined from at least one observed tie; otherwise, the tie parameter remains unconstrained (Yan, 2014).

Problems of nonuniqueness or nonexistence in sparse data are mitigated by $j$6-perturbation (adding small positive counts to existing pairwise outcomes), replacing the strong connectivity requirement of the win-loss digraph by weak connectivity of the undirected comparison graph. Under this, and with at least one tie, strict log-likelihood concavity ensures unique MLEs for both principal tie models (Yan, 2014).

3. Estimation Algorithms: MM, EM, and Bayesian Methods

Minorization-Maximization (MM) and Expectation-Maximization (EM): For Rao–Kupper and Davidson models, MM and EM coincide; the surrogate likelihood is constructed by exploiting convexity, allowing closed-form or one-dimensional updates for strengths and tie parameters (Caron et al., 2010, Vojnovic et al., 2019).

Accelerated MM: Parameter rescaling ensures uniform PL (Polyak–Łojasiewicz) condition even when the tie parameter or prior becomes small, resulting in provably linear convergence, outperforming basic MM in ill-conditioned cases (Vojnovic et al., 2019).

Bayesian inference: Gibbs samplers are constructed via latent variable augmentation, e.g., Gamma or Pólya–Gamma variables, allowing efficient updates for strengths and ties. The full posterior (with proper priors) facilitates credible intervals and uncertainty quantification. For high-dimensional or structured problems (e.g., spatial priors), block Gibbs sampling and latent variable augmentation remain scalable (Seymour et al., 2024, Caron et al., 2010).

Sampling Details Table: (Algorithms for Rao–Kupper/Davidson models)

Step	Method	Notes
Strength update	Closed-form/Gibbs	EM/MM (analytic), Bayesian (Gaussian/Gamma)
Tie-parameter update	1D root-finding, MH/Gibbs	Surrogate maximization or latent variable
Scaling/centering	Affine normalization	Ensures identifiability

4. Model Generalizations and Extensions

Order effects and strength-dependent ties: In chess and some sports, tie rates and first-mover or home advantage are empirically found to be higher among stronger players. Models thus generalize tie propensities as $j$7, and order-effects with $j$8 (Glickman, 30 May 2025).

Multiple outcomes: Extension to more than three outcomes (e.g., regulation, overtime, shootout in hockey) is accomplished by modeling outcome probabilities as multinomial softmaxes, with fractional points as exponents, calibrated to reproduce league standings and observed margins (Whelan et al., 2021).

Preference learning (RLHF/DPO): In RLHF and Direct Preference Optimization, integrating tie-aware models (Rao–Kupper, Davidson, BTT) into loss functions allows the learning signal from tied (ambiguous) pairs to regularize model behavior, improve calibration, and reduce bias in reward estimates (Liu et al., 2024, Chen et al., 2024).

5. Theoretical Properties and Implications

Omitting ties and treating them as random results biases preference-strength estimates toward zero; this attenuation grows with the tie-parameter (Liu et al., 2024). Analytic expressions quantify this shrinkage. Tie-augmented models are theoretically guaranteed to assign KL-regularized, less-extreme margins to ambiguous pairs, preventing overfitting and improving out-of-sample consistency—critical for preference learning where ambiguous or adversarial comparisons are common (Chen et al., 2024).

Quasi-likelihood frameworks embed ties within unbiased U-statistics for the Kemeny correlation, offering strictly unbiased, analytic estimators, Edgeworth corrections, and concentration bounds independent of parametric assumptions (Hurley, 30 Dec 2025).

6. Computational Scalability and Practical Implementation

Pólya–Gamma augmentation yields block-conjugate updates for strength vectors, allowing $j$9 (or less, with sparsity/exploiting structure) solves per MCMC iteration (Seymour et al., 2024). EM/MM and Gibbs schemes for moderate $\lambda_i$0 (dozens to thousands) are tractable in practice, with acceleration stratagems applicable for slow-mixing or high-dimensional settings (Caron et al., 2010, Vojnovic et al., 2019).

Out-of-the-box Hamiltonian Monte Carlo (e.g., via Stan) is effective for joint strength-tie inference, given the twice-differentiable, strictly concave posterior (Whelan et al., 2021). In practical use cases (e.g., hundreds of entities, thousands of comparisons), convergence diagnostics and effective sample size per second favor Pólya–Gamma over naive random walk Metropolis samplers (Seymour et al., 2024).

7. Empirical Performance, Application Domains, and Guidelines

Tie-augmented Bradley–Terry models have shown improved fit and calibration in competitive sports (NFL, chess, hockey), psychometrics, and RLHF reward modeling. For instance, in community-level abuse mapping, tie probabilities around 0.14–0.19 (for equal-strength units) were inferred, with efficient, scalable MCMC sampling demonstrated in real field data (Seymour et al., 2024).

In preference optimization, inclusion of ties with proper modeling yields stronger regularization and improved generalization; empirical metrics (summarization, translation) confirm that vanilla preference models degrade with ties, while tie-augmented variants regain or exceed baseline performance, matching real-world ambiguity (Chen et al., 2024, Liu et al., 2024).

Practical recommendations:

• Do not discard tied comparisons; use tie-parameterized models and optimize both strength and tie parameters.
• Set the tie-sensitivity parameter (e.g., $\lambda_i$1 so that tie-probability at zero margin is 50%) for robust, interpretable results (Chen et al., 2024).
• Enforce identifiability via centering or affine constraints at each iteration.

Conclusion:

The tie-augmented Bradley–Terry model, via principled parameterizations (Rao–Kupper, Davidson, BTT, and their generalizations), robust estimation strategies, and scalable inference procedures, provides a statistically rigorous and practically indispensable framework for modern applications in ranking, preference learning, and paired-comparison studies where ties are an intrinsic phenomenon.