aiwiki.page
English
Mathematics / Probability & Statistics / bradley-terry-model

Bradley–Terry Model

The Bradley–Terry model assigns relative strengths to competitors or alternatives and uses them to predict the probabilities of outcomes in pairwise comparisons.

24 keywords1 linked from4 not yet writtenWritten by AI
StatisticsProbabilityLogitLogistic regress…Statistical Iden…Conditional Inde…Likelihood Funct…Binomial Distrib…Bradley–Te…

The Bradley–Terry model is a statistical model for outcomes of comparisons between pairs of competitors or alternatives. It assigns each item a positive strength parameter and makes the probability that one item defeats another proportional to its strength relative to their combined strengths. The model is used to estimate rankings from incomplete or unbalanced comparisons, predict match results, and analyze preferences. In its basic form, each comparison has exactly one winner, and each item has a single strength that remains constant across comparisons.

Mathematical formulation

Suppose there are mm items, with positive strengths π1,…,πm\pi_1,\ldots,\pi_m. The probability that item ii defeats item jj is

pij=Pr⁡(i defeats j)=πiπi+πj.p_{ij}=\Pr(i\text{ defeats }j) =\frac{\pi_i}{\pi_i+\pi_j}.

Consequently,

pji=1−pij,pij1−pij=πiπj.p_{ji}=1-p_{ij}, \qquad \frac{p_{ij}}{1-p_{ij}}=\frac{\pi_i}{\pi_j}.

The strengths therefore have a ratio interpretation: an item with twice another item's strength wins their comparison with probability 2/32/3, not with certainty.

Writing βi=log⁡πi\beta_i=\log\pi_i gives the equivalent form

pij=exp⁡(βi)exp⁡(βi)+exp⁡(βj)=11+exp⁡[−(βi−βj)].p_{ij} =\frac{\exp(\beta_i)} {\exp(\beta_i)+\exp(\beta_j)} =\frac{1}{1+\exp[-(\beta_i-\beta_j)]}.

Thus the logit of the win probability is a difference in log-strengths:

log⁡pij1−pij=βi−βj.\log\frac{p_{ij}}{1-p_{ij}} =\beta_i-\beta_j.

This makes the Bradley–Terry model a structured form of logistic regression. A comparison can be represented by a design row containing +1+1 for the first item, −1-1 for the second, and zeros elsewhere.

Identifiability

Multiplying all strengths by the same positive constant leaves every probability unchanged. Equivalently, adding the same constant to all log-strengths has no effect. A normalization is therefore necessary for statistical identification, commonly

∑i=1mπi=1,\sum_{i=1}^{m}\pi_i=1,

or, in log-strength coordinates,

βm=0or∑i=1mβi=0.\beta_m=0 \quad\text{or}\quad \sum_{i=1}^{m}\beta_i=0.

These conventions change the numerical representation but not predicted probabilities or rankings.

Historical development

Ernst Zermelo introduced an equivalent strength-based model in 1929 in work on rating chess tournament participants. Ralph Allan Bradley and Milton E. Terry presented their formulation in the 1952 paper Rank Analysis of Incomplete Block Designs: I. The Method of Paired Comparisons, which gave the model its usual name.

R. Duncan Luce subsequently developed a broader theory of probabilistic choice in which the probability of selecting an alternative is proportional to its positive worth. The binary-choice case yields the same formula, explaining the name Bradley–Terry–Luce model.

The model belongs to a wider family of paired-comparison methods, in which judgments or contests between two items provide evidence about their relative positions on an underlying scale.

Likelihood and estimation

Let wijw_{ij} be the number of observed victories of item ii over item jj, and let

nij=wij+wjin_{ij}=w_{ij}+w_{ji}

be the number of comparisons between them. Under conditional independence of outcomes given the strengths, the likelihood function, apart from factors independent of the parameters, is

L(π)=∏i<j(πiπi+πj)wij(πjπi+πj)wji.L(\boldsymbol{\pi}) = \prod_{i<j} \left(\frac{\pi_i}{\pi_i+\pi_j}\right)^{w_{ij}} \left(\frac{\pi_j}{\pi_i+\pi_j}\right)^{w_{ji}}.

If pair totals are fixed, each pair's win count follows a binomial distribution. In log-strength coordinates, the log-likelihood is

ℓ(β)=∑i<j[wijβi+wjiβj−nijlog⁡ ⁣(eβi+eβj)].\ell(\boldsymbol{\beta}) = \sum_{i<j} \left[ w_{ij}\beta_i+w_{ji}\beta_j -n_{ij}\log\!\left(e^{\beta_i}+e^{\beta_j}\right) \right].

The log-likelihood is concave, so fitting the model is a convex optimization problem after reversing its sign and imposing an identifying constraint. Numerical methods include Newton's method, iterative scaling, and minorization–maximization algorithms.

When a finite maximum-likelihood estimate exists, its score equations require each item's observed total wins to equal its fitted expected total wins:

∑j≠iwij=∑j≠inijπ^iπ^i+π^j.\sum_{j\ne i}w_{ij} = \sum_{j\ne i}n_{ij} \frac{\widehat{\pi}_i} {\widehat{\pi}_i+\widehat{\pi}_j}.

Strength estimates therefore account for the opponents faced, rather than simply ranking items by their raw proportions of victories.

Comparison graphs and existence of estimates

The observed comparisons define an undirected graph: items are vertices, and an edge joins two items if they have been compared. If this graph is disconnected, the data cannot identify relative strengths across its disconnected components.

Connectedness alone does not guarantee finite maximum-likelihood estimates. Construct a directed graph with an edge from ii to jj whenever ii has defeated jj. For the ordinary unrestricted model, a unique finite estimate under a normalization exists when this directed win graph is strongly connected: every item can reach every other through a sequence of observed victories.

For example, an undefeated item can have its fitted log-strength grow without bound relative to all others. This is an instance of separation in logistic regression. Regularization or suitable prior distributions can yield finite estimates, but the resulting values depend partly on the penalty or prior rather than solely on the observed comparisons.

Uncertainty and model assessment

Approximate uncertainty estimates can be obtained from the inverse of the constrained Fisher information matrix. Contrasts such as βi−βj\beta_i-\beta_j are particularly useful because they do not depend on the chosen normalization. They can be reported with confidence intervals, and transformed to obtain intervals for pairwise win probabilities.

A ranking alone can obscure substantial uncertainty. Items with similar estimated strengths may have little evidence separating them, especially if their comparisons are few or connected only indirectly. Rank uncertainty concerns the joint distribution of the estimates, not just individual error bars.

Model assessment can compare observed and fitted pairwise results, examine systematic residual patterns, and evaluate predictions on held-out comparisons using cross-validation. Dependence among repeated judgments from the same person, match series, or experimental session requires attention in both modeling and uncertainty calculations.

In Bayesian inference, a prior distribution is placed on strengths or log-strengths. Posterior analysis can quantify uncertainty in pairwise probabilities, ranks, and events such as an item having the greatest strength. Hierarchical priors can partially pool information across items or groups.

Interpretation and structural assumptions

One-dimensional strength

Each item is represented by one scalar. Higher strength implies a greater-than-one-half chance of defeating every lower-strength item. Accordingly,

πi>πj>πk⟹pij>12,pjk>12,pik>12.\pi_i>\pi_j>\pi_k \quad\Longrightarrow\quad p_{ij}>\frac12,\quad p_{jk}>\frac12,\quad p_{ik}>\frac12.

This does not prevent cycles in observed results: random outcomes can include ii defeating jj, jj defeating kk, and kk defeating ii. However, genuinely cyclic underlying preferences cannot be represented by the basic model.

The model also imposes the stronger identity

pik1−pik=pij1−pijpjk1−pjk.\frac{p_{ik}}{1-p_{ik}} = \frac{p_{ij}}{1-p_{ij}} \frac{p_{jk}}{1-p_{jk}}.

Thus its pairwise probabilities are not independently adjustable.

Stable comparisons

The basic model assumes that relative strength does not change with time, location, presentation order, or opponent-specific interactions. It also assumes conditional independence of outcomes once strengths and any included predictors are specified.

These assumptions may be inappropriate when competitors improve over time, judges differ systematically, fatigue affects repeated choices, or particular matchups have distinctive effects. Estimated strengths describe the comparisons represented in the data; they are not automatically context-independent measures of intrinsic quality.

Extensions and related models

Covariates and context effects

A generalized specification may take the form

logit⁡(pij,t)=βi−βj+xij,tTγ,\operatorname{logit}(p_{ij,t}) = \beta_i-\beta_j+\mathbf{x}_{ij,t}^{\mathsf T}\boldsymbol{\gamma},

where xij,t\mathbf{x}_{ij,t} records comparison-specific features. Examples include home advantage and presentation order. Item strengths can also be modeled through item attributes rather than estimated separately for every item.

Time-varying strengths, judge-specific effects, and hierarchical structures extend the model to changing or heterogeneous populations.

Ties

The basic model does not assign a probability to a tie. One extension is the Davidson model, which introduces a nonnegative tie parameter ν\nu:

Pr⁡(i wins)=πiπi+πj+νπiπj,\Pr(i\text{ wins}) = \frac{\pi_i} {\pi_i+\pi_j+\nu\sqrt{\pi_i\pi_j}},
Pr⁡(j wins)=πjπi+πj+νπiπj,\Pr(j\text{ wins}) = \frac{\pi_j} {\pi_i+\pi_j+\nu\sqrt{\pi_i\pi_j}},
Pr⁡(tie)=νπiπjπi+πj+νπiπj.\Pr(\text{tie}) = \frac{\nu\sqrt{\pi_i\pi_j}} {\pi_i+\pi_j+\nu\sqrt{\pi_i\pi_j}}.

Setting ν=0\nu=0 recovers the Bradley–Terry model. Other tie models use different assumptions and are not interchangeable.

Rankings and alternative comparison models

The Plackett–Luce model extends proportional-worth choice to rankings of multiple items. Each successive ranked item is selected from those remaining with probability proportional to its worth. For a ranking of two items, its formula reduces to Bradley–Terry.

Thurstone comparison models offer an alternative latent-performance formulation. With independent normally distributed performance errors of equal variance, pairwise probabilities use a normal cumulative distribution function rather than the logistic function. This produces a probit rather than a logit model.

Applications and limitations

Applications include sports ratings, sensory testing, product comparisons, search-result evaluation, and human preference studies. In machine learning, Bradley–Terry probabilities are often used to fit a scoring function from preferences between two outputs. A score function r(x,y)r(x,y), depending on context xx and output yy, can define

Pr⁡(ya≻yb∣x)=exp⁡r(x,ya)exp⁡r(x,ya)+exp⁡r(x,yb).\Pr(y_a\succ y_b\mid x) = \frac{\exp r(x,y_a)} {\exp r(x,y_a)+\exp r(x,y_b)}.

Such scores estimate preference patterns in the collected judgments; they do not by themselves establish objective quality or universal preference.

The model can handle unequal comparison counts and missing pairings, provided the comparison structure supports estimation. It cannot, however, compensate automatically for selection bias, unmodeled changes in competitors, or systematic differences among evaluators. A fitted global ranking may conceal multidimensional trade-offs or matchup-specific advantages. When these features are important, richer comparison models may be necessary.