## Introduction

Rankings data arise in a range of applications, such as sport tournaments and consumer studies. In rankings data, each observation is an ordering of a set of items. A classic model for such data is the Plackett–Luce model, which stems from Luce’s axiom of choice (Luce 1959, 1977). This axiom states that the odds of choosing one item, $$i_1$$, over another, $$i_2$$, do not depend on the other items available for selection in that choice.

Suppose we have a set of $$J$$ items

\begin{aligned} S = \{i_1, i_2, \ldots , i_J\}. \end{aligned}

Under Luce’s axiom, the probability of selecting some item $$i_j$$ from $$S$$ is given by

\begin{aligned} P(i_j | S) = \frac{\alpha _{i_j}}{\sum _{i \in S} \alpha _i}, \end{aligned}

where $$\alpha _i \ge 0$$ represents the worth of item $$i$$. The worth is a latent characteristic, with higher values indicating that an item is more likely to be selected. In practice, it could represent the ability of an athlete or the attractiveness of a product, for example.

A ranking of $$J$$ items can be viewed as a sequence of independent choices: first choosing the top-ranked item from all items, then choosing the second-ranked item from the remaining items and so on. Let $$a \succ b$$ denote that $$a$$ is ranked higher than $$b$$. It follows from Luce’s axiom that the probability of the ranking $${i_1 \succ \cdots \succ i_J}$$ is

\begin{aligned} \prod _{j=1}^J \frac{\alpha _{i_j}}{\sum _{i \in A_j} \alpha _i}, \end{aligned}
(1)

where $$A_j$$ is the set of alternatives $$\{i_j, i_{j + 1}, \ldots , i_J\}$$ from which item $$i_j$$ is chosen. The above model is also derived in Plackett (1975), hence the name Plackett–Luce model.

In this paper, we present the R package PlackettLuce (Turner et al. 2019). This package implements an extension of the Plackett–Luce model that allows for ties in the rankings. The model can be applied to either complete or partial rankings (complete rankings of subsets of items). PlackettLuce offers a choice of algorithms to fit the model via maximum likelihood. Pseudo-rankings, i.e. pairwise comparisons with a hypothetical item, are used to ensure that the item worths always have finite maximum likelihood estimates (MLEs) with finite standard error. Methods are provided to obtain different parameterizations with corresponding standard errors or quasi-standard errors (that do not change with the identifiability constraint). There is also a method to fit Plackett–Luce trees, which partition the rankings by covariate values to identify subgroups with distinct Plackett–Luce models.

In the next section, we review the available software for modelling rankings and make comparisons with PlackettLuce. Then in Sect. 3 we describe the Plackett–Luce model with ties and the methods implemented in the package for model-fitting and inference. Plackett–Luce trees are introduced in Sect. 4, before we conclude the paper with a discussion in Sect. 5.

## Software for modelling rankings

First, in Sect. 2.1, we consider software to fit the standard Plackett–Luce model. Then in Sect. 2.2, we review the key packages in terms of their support for features beyond fitting the standard model. Finally in Sect. 2.3 we review software implementing alternative models for rankings data and discuss how these approaches compare with the Plackett–Luce model.

### Fitting the Plackett–Luce model

Sections 2.1.1 and 2.1.2 describe two ways the Plackett–Luce model can be fitted with standard statistical software. Section 2.1.3 reviews packages that have been specifically developed to fit the Plackett–Luce model or a related genralization. The performance of the key packages for R are compared in Sect. 2.1.4.

#### Fitting as a log-linear model

Each choice in the ranking can be considered as a multinomial observation, with one item observed out of a possible set. The “Poisson trick” (see, for example, Baker 1994) expresses these multinomial observations as counts (one or zero) of each possible outcome within each choice. For example, the ranking {item 3} $$\succ$$ {item 1} $$\succ$$ {item 2} is equivalent to two multinomial observations: {item 3} from {item 1, item 2, item 3} and {item 1} from {item 2, item 3}. These observations are expressed as counts as follows:

The Plackett–Luce model is then equivalent to a log-linear model of count, with choice included as a nuisance factor to ensure the multinomial probabilities sum to one, and the dummy variables for each item (item 1, item 2, item 3) included to estimate the item worth parameters.

As a result, the Plackett–Luce model can be fitted using standard software for generalized linear models (GLMs). However, expanding the rankings to counts of each possible choice is a non-standard task. Also, the choice factor will have many levels: greater than the number of rankings, for rankings of more than two objects. A standard GLM fitting function such as glm in R will be slow to fit the model, or may even fail as the corresponding model matrix will be too large to fit in memory. This issue can be circumvented by “eliminating” the nuisance choice factor, by taking advantage of the diagonal submatrix that would appear in the Fisher Information matrix to compute updates without requiring the full design matrix. This feature is available, for example, in the gnm function from the R package gnm (Turner and Firth 2018). Even then, the model-fitting may be relatively slow, given the expansion in the number of observations when rankings are converted to counts. This issue is exacerbated when ties are allowed, since the number of possible outcomes increases quickly with the number of items.

#### Fitting as a Cox proportional hazards model

For complete rankings, the probability in Eq. (1) is equivalent to that for a special case of the rank-ordered logit model or exploded logit model (Allison and Christakis 1994). This model assumes a certain utility $$U_{ri}$$ for each item $$i$$ and respondent $$r$$, which is modelled as

\begin{aligned} U_{ri} = \mu _{ri} + \epsilon _{ri}. \end{aligned}

Here $$\mu _{ri}$$ may be decomposed into a linear function of explanatory variables, while the $$\{\epsilon _{ri}\}$$ are independent and identically distributed with an extreme value distribution. The standard Plackett–Luce model is then given by a rank-ordered logit model with $$U_{ri} = \log (\alpha _i) + \epsilon _{ri}$$.

Allison and Christakis (1994) demonstrated that the rank-ordered logit model can be estimated by fitting a Cox proportional hazards model with equivalent partial likelihood. Therefore, specialist software is not required to fit the model. Software such as the rologit function in Stata or the rologit function from the R package ROlogit (Tan et al. 2017) provide wrappers to methods for the Cox proportional hazards model, so that users do not have to specify the equivalent survival model.

Compared to the log-linear approach, the rank-ordered logit model has the advantage that the ranks can be modelled directly without expanding into counts. Nonetheless, the model fitting requires a dummy variable for each item and a stratification factor for each unique ranking. Therefore, the estimation becomes slow with a high number of items and/or unique rankings.

#### Specialized software

Several software packages have been developed to model rankings data, which provide specialized functions to fit the Plackett–Luce model. Rather than re-expressing the model so that it can be estimated with a standard modelling function, these specialized functions take a direct approach to fitting the Plackett–Luce model.

One such approach is Hunter’s (2004) minorization–maximization (MM) algorithm to maximize the likelihood of the observed data under the Plackett–Luce model. This algorithm is used by the R package StatRank and is one of the algorithms available in the Python package choix (Maystre 2018). Alternatively the likelihood can be maximised directly using a generic optimisation method such as the Broyden–Fletcher–Goldfarb–Shanno (BFGS) algorithm. This is another option offered by choix and is the method used in the R packages pmr (Lee and Yu 2015) and hyper2 (Hankin 2019). A third option offered by choix is the “Luce spectral algorithm”, which in a single iteration (LSR) gives a close approximation to the MLEs and may be used iteratively (I-LSR) to give a maximum likelihood solution (Maystre and Grossglauser 2015).

In addition, a number of algorithms have been proposed to fit Plackett–Luce mixture models. These models assume the rankings are observations from $$G$$ subgroups, each with a distinct Plackett–Luce model. The group membership probabilities are estimated as part of the model-fitting. Therefore, software to fit Plackett–Luce mixture models can be used fit the standard Plackett–Luce model by fixing the number of groups, $$G$$, to be one. The R package mixedMem (Wang and Erosheva 2015) implements a variational Expectation-Maximization (EM) algorithm, which returns MLEs in the case of a single group. The R package PLMIX (Mollica and Tardella 2014, 2017, 2020) uses a Bayesian framework and offers the choice to maximize the posterior distribution via an EM algorithm, which reduces to the MM algorithm with noninformative hyperparameters, or to simulate the posterior distribution using Markov-chain Monte-Carlo (MCMC) techniques.

PlackettLuce offers both iterative scaling, which is equivalent to the MM algorithm for the standard Plackett–Luce model, and generic optimization using either BFGS or a limited memory variant (L-BFGS) via the R package lbfgs package (Coppola et al. 2014).

#### Performance of software to fit the Plackett–Luce model

Even using specialized software, the time taken to fit a Plackett–Luce model can scale poorly with the number of items and/or the number of rankings. Table 1 shows some example run times for fitting the Plackett–Luce model with specialized software available in R, including ROlogit for comparison. The run times are presented for three different data sets taken from the PrefLib library of preference data (Mattei and Walsh 2013). The data sets are selected to represent a range of features, as summarised in Table 2.

As expected, ROlogit, based on R’s coxph function, shows a marked drop in performance when the number of unique rankings is high. A similar drop off is seen for hyper2 where the time is spent specifying the likelihood function for each unique ranking. Although StatRank uses an MM algorithm, the same as used by PlackettLuce and PLMIX in these benchmarks, it shows poor performance for a moderate number of items ($$\ge$$ 10) and/or a high number of unique rankings. Similarly pmr which uses the same optimization method as hyper2 becomes impractical to use with a moderate number of items.

Clearly, the details of the implementation are as important as the estimation method used. PlackettLuce copes well with these moderately-sized data sets and is comparable to mixedMem. PlackettLuce is faster than mixedMem on the Netflix data as it can work with aggregated unique rankings. PLMIX is consistently fast; the superior performance over PlackettLuce when both the number of items and the number of unique rankings is high can be explained by its core functionality being implemented in C++ rather than pure R code.

### Features beyond fitting the standard model

So far, we have compared packages in terms of their ability to fit the standard Plackett–Luce model. In practice, rankings data can have features that require a generalization of the standard model, or that can cause issues with maximum likelihood estimation. In addition, not all packages provide facilities for inference on the estimated item worths.

In this section, we compare the key packages for fitting the Plackett–Luce model in terms of their ability to meet these additional requirements.

#### Partial rankings

As the number of items increases, it is typically more common to observe partial rankings than complete rankings. Partial rankings can be of two types: subset rankings, where subsets of items are completely ranked in each observation, and top $$n$$rankings, where the top $$n$$ items are selected and the remaining items are unranked, but implicitly ranked lower than the top $$n$$. Both types can be accommodated by the Plackett–Luce model but imply a different form for the likelihood. For subset rankings, the set of alternatives in the denominator of (1) will only comprise remaining items from the subset of items in the current ranking. For top $$n$$ rankings the set of alternatives is always the full set of items minus the items that have been selected so far in the current ranking.

Given this difference in the likelihood, implementations tend to support only one type of partial ranking. PlackettLuce and choix handle subset rankings only, while PLMIX, ROlogit and mixedMem handle top $$n$$ rankings only. The documentation for StatRank suggests that it supports subset rankings, but we were unable to validate such support. hyper2 can handle both types of partial ranking, since it requires the likelihood to be specified for each observation. However, hyper2 provides a utility function for subset rankings making these easier to fit with the package.

Table 3 shows run times for fitting the Plackett–Luce model to the NASCAR subset rankings from Hunter (2004). These results illustrate that PlackettLuce is more efficient than hyper2 for modelling subset rankings of a relatively large number of items.

#### Tied ranks

ROlogit can handle tied ranks as tied survival times in the underlying Cox proportional hazards model. The ties are accommodated in the likelihood by summing over all possible strict rankings that could have resulted in the observed tie. This assumes that one of these strict rankings is correct, but a lack of precision means we observe a tie.

However in some contexts, a tie is a valid outcome, for example, an equal number of goals in a sports match. As far as we are aware, PlackettLuce is the only software to model ties as an explicit outcome based on item worth, through an extension of the Plackett–Luce model described in Sect. 3.

#### Inference

In order to conduct frequentist inference on the item parameters, it is necessary to compute standard errors. The only packages based on maximum likelihood estimation with this functionality are PlackettLuce and ROlogit. The computation of the variance-covariance matrix in PlackettLuce is currently based on the methodology for generalized linear models, so the rankings are internally expanded as counts as described in Sect. 2.1.1.

PLMIX allows for a fully Bayesian inference based on the approximation of the posterior distribution.

#### Disconnected or weakly connected networks

In some cases, the MLE does not exist, or has infinite standard error. This happens when the network of wins and losses implied by the rankings is not strongly connected—an issue explored further in Sect. 2.2.4. PlackettLuce ensures that the item worths can always be estimated by using pseudo-rankings. This approach can be viewed as incorporating a “shrinkage” prior as in a Bayesian analysis (but with estimation based on maximization of the posterior density). The choix package implements methods of regularization that have a similar effect.

Since PLMIX uses a Bayesian framework that incorporates a prior, the package will always produce finite worth estimates.

#### Heterogeneity

As noted in Sect. 2.1.3PLMIX and mixedMem offer the facility to model heterogeneous rankings via mixture models, which allow for latent groups with different item worth. This is similar in spirit to the tree models offered by PlackettLuce, which define such groups through recursive binary splits on attributes of the judges or other variables that covary with the rankings. This partitioning approach does not rely on parametric assumptions, whereas the mixture models assume a heterogeneous sampling distribution.

#### Summary of software features

A summary of the features of the main software for Plackett–Luce models is given in Table 4. pmr and StatRank are not included as they do not support any of the features mentioned (with the possible exception of partial rankings in the case of StatRank, where the documented support could not be validated).

Some of the software for Plackett–Luce models have features beyond the scope of PlackettLuce and these features are also included in Table 4. ROlogit can model the item worth as a linear function of covariates. hyper2 can handle rankings of combinations of items, for example, team rankings in sports. This feature is implemented by extending the Plackett–Luce model to describe the selection of a combination $$T_j$$ of items from the alternatives, that is

\begin{aligned} \prod _{j=1}^J \frac{\sum _{i \in T_j} \alpha _{i}}{\sum _{i \in A_j} \alpha _i}. \end{aligned}

### Other software for modelling rankings

Although we are focusing on the Plackett–Luce model in this paper, there are other approaches to modelling rankings data. This section reviews the main alternatives and the available implementations.

#### Mallows’ Bradley–Terry model

For partial rankings of two items, i.e., paired comparisons, the Plackett–Luce model reduces to the Bradley–Terry model (Bradley and Terry 1952). Mallows’ Bradley–Terry model (Mallows 1957) is an alternative way to extend the Bradley–Terry model for complete rankings. Each ranking is considered as a set of paired comparisons and their joint probability is modelled assuming independence, normalizing over all possible rankings. Mallows’ Bradley–Terry model is implemented by the eba and prefmod packages for R (Wickelmaier and Schmid 2004; and Hatzinger and Dittrich 2012, respectively). Although both packages estimate item worths, neither eba nor prefmod provide methods for inference based on these estimates.

The prefmod package implements Mallows’ Bradley–Terry model as a special case of a wider class of pattern models, where a pattern is a set of paired comparisons considered simultaneously. Several extensions of the basic model are implemented in prefmod: allowing for dependencies between paired comparisons that involve the same item; allowing for ties and partial rankings; incorporating item, subject or order effects, and allowing for latent subject effects via mixture models. However, not all of these features can be combined. Specifically, item effects, subject effects due to continuous attributes, order effects and mixture models are not compatible with partial rankings. Partial rankings are implemented by treating unranked items as having missing ranks and all possible rankings including missing values must be considered. This means that the methods do not scale well to a large number of items (the package documentation suggests 8 items as a reasonable maximum).

#### Mallows distance-based model

The Mallows distance-based model (Mallows 1957) does not model the worth of items, but rather a global ranking. Specifically, it models the modal ranking and a dispersion parameter, which describes how quickly the probability of a ranking decreases with increasing distance from the modal ranking. Extensions include the generalized Mallows model that allows a separate dispersion parameter for each free position in the ranking, which can give some indication of the relative worth. Standard and generalized versions of Mallows models are implemented in the pmr and PerMallowsR packages (Irurozki et al. 2016).

The rankdistR package implements mixtures of the standard Mallows model for complete or subset rankings via maximum likelihood (Qian and Yu 2019). The BayesMallowR package (Vitelli et al. 2018) implements mixtures of the standard Mallows model in a Bayesian framework. This has the advantage of facilitating inference on the global ranking (in each sub-population), e.g. 95% highest posterior density intervals on ranking position; probability of an item appearing in the top 10, etc. In addition, subset rankings, top $$n$$ rankings, missing ranks and ties can all be accommodated in the Bayesian model.

#### Insertion sorting rank model

The Rankcluster package (Grimonprez and Jacques 2016) implements an alternative probabilistic model for rankings called the Insertion Sorting Rank model. This is based on a modal ranking and a parameter for the probability of a ranker making an implicit paired comparison consistent with the modal ranking. Heterogeneity is accommodated via a mixture model and the model can be applied to subset or top-$$n$$ rankings. As the package name suggests, the focus is on clustering subjects by their modal ranking and the probability of membership, rather than inference on item worths.

#### Summary

Table 5 provides an overview of software based on alternative models, compared to PlackettLuce.

## Plackett–Luce model with ties

In this section, we describe and demonstrate the core features of the PlackettLuce package: modelling partial rankings with ties, conducting inference on the worth parameters, and handling ranking networks that are not strongly connected.

### Modelling partial rankings with ties

The PlackettLuce package permits rankings of the form

\begin{aligned} R = \{C_1, C_2, \ldots , C_J\}, \end{aligned}

where the items in set $$C_1$$ are ranked higher than (better than) the items in $$C_2$$, and so on. If there are multiple objects in set $$C_j$$ these items are tied in the ranking. To accommodate such ties, PlackettLuce models each choice in the ranking with the Davidson-Luce model as developed recently by Firth et al. (2019).

For a set $$S$$, let

\begin{aligned} f(S) = \delta _{|S|} \left( \prod _{i \in S} \alpha _i \right) ^\frac{1}{|S|}, \end{aligned}

where $$|S|$$ is the cardinality of the set $$S$$, $$\delta _n$$ is a parameter representing the prevalence of ties of order $$n$$ (with $$\delta _1 \equiv 1$$), and $$\alpha _i$$ is the worth of item $$i$$. Then, the probability of a ranking $$R$$ with ties up to order $$D$$ is given by

\begin{aligned} \prod _{j = 1}^J \frac{f(C_j)}{ \sum _{k = 1}^{\min (D_j, D)} \sum _{S \in {\left( {\begin{array}{c}A_j\\ k\end{array}}\right) }} f(S)}. \end{aligned}
(2)

Here $$D_j$$ is the cardinality of $$A_j$$, which is the set of items from which $$C_j$$ is chosen, and $$\left( {\begin{array}{c}A_j\\ k\end{array}}\right)$$ denotes the set of all possible choices of $$k$$ items from $$A_j$$. The value of $$D$$ can be set to the maximum number of tied items observed in the data, so that $$\delta _n = 0$$ for $$n > D$$.

The scale of the worth parameters needs to be constrained for the parameters to be identifiable. PlackettLuce constrains the worth parameters to sum to one, so that they represent the probability that the corresponding item comes first in a ranking of all items, given that first place is not tied.

The 2-way tie parameter $$\delta _2$$ is related to the probability that two items of equal worth tie for first place, given that the first place is not a 3-way or higher tie. Specifically, that probability is $$\delta _2/(2 + \delta _2)$$. The 3-way and higher tie parameters are similarly interpretable, in terms of tie probabilities among equal-worth items.

Further description of this general model for handling multi-item ties, including its origins and its most important properties, can be found in Firth et al. (2019).

#### Pudding example

When each ranking contains only two items, the model in Eq. (2) reduces to the extended Bradley–Terry model proposed by Davidson (1970). In Davidson ((1970), Example 2), a case study is presented where respondents were asked to test two brands of chocolate pudding from a total of six brands. The results are provided as the pudding data set in PlackettLuce. For each pair of brands $$i$$ and $$j$$, the data set gives the total number of paired comparisons ($$r_{ij}$$), along with the frequencies that brand $$i$$ was preferred ($$w_{ij}$$), that brand $$j$$ was preferred ($$w_{ji}$$) and that the brands were tied ($$t_{ij}$$).

PlackettLuce(), the model-fitting function in PlackettLuce, requires each ranking to be specified by the numeric rank (1, 2, 3, $$\ldots$$) assigned to each item. For the pudding data, it is easier to specify the ordering (winner, loser) initially. In particular, we can define the winner and loser for the three sets of paired comparions: the wins for brand $$i$$, the wins for brand $$j$$ and the ties:

In the last case, the base R function asplit() is used to split the i and j columns of pudding by row, giving a vector of items that we can specify as the winner, while the loser is missing.

The combined orderings can be converted to numeric rankings using the as.rankings() function from PlackettLuce:

The result is an object of class "rankings". The print method displays the rankings in a readable form, indicating the ordering of the brands. However, the underlying data structure stores the rankings in the form of a matrix:

The six columns represent the pudding brands. In each row, 0 represents an unranked brand (not in the comparison), 1 represents the brand(s) ranked in first place and 2 represents the brand in second place, if applicable. PlackettLuce() will also accept rankings in this matrix form, which it will convert to a rankings object using as.rankings() with input = "rankings".

To specify the full set of rankings, we need the frequency of each ranking, which will be specified to the model-fitting function as a weight vector:

Now we can fit the model with PlackettLuce(), passing the rankings object and the weight vector as arguments. Setting npseudo = 0 means that standard maximum likelihood estimation is performed and maxit = 7 limits the number of iterations to obtain the same worth parameters as Davidson (1970):

Here we have specified log = FALSE to report the estimates in the parameterization of Eq. (2), with the worth parameters constrained to sum to one. In Sect. 2.2.3 we discuss why it is more appropriate to use the log scale for inference.

#### Choice of algorithm

As noted in Sect. 2.1.3, PlackettLuce() offers a selection of three fitting algorithms, which can be specified by setting the argument method to "iterative scaling", "BFGS" or "L-BFGS". Iterative scaling is the default because it is a reliable, albeit slow, method. Convergence is determined by direct comparison of the observed and expected values of the sufficient statistics for the parameters.

The sufficient statistics are: for each item, the number of outright wins plus one $$n$$th of the number of $$n$$-way ties ($$n = 1, \ldots , D$$); and for each tie of order $$n$$, the number of $$n$$-way ties. Assessing convergence by how close the sufficient statistics are to their expected value gives more control over the precision of the parameter estimates, compared to specifying a tolerance on the change in log-likelihood.

Iterative scaling is slow because it does not use derivatives of the log-likelihood. However, PlackettLuce() uses Steffensen’s method (see for example, Nievergelt 1991) to accelerate the algorithm when it gets close to the final solution.

The BFGS and L-BFGS algorithms are implemented in PlackettLuce via the workhorse functions optim() (from base R) or lbfgs() (from lfbgs). Both algorithms use derivatives of the log-likelihood, therefore typically require fewer iterations than iterative scaling. However, BFGS approximates the Hessian matrix (matrix of second derivatives), which is computationally expensive, while L-BFGS uses past estimates of the first derivatives to avoid evaluating the Hessian. Overall then, L-BFGS typically has the quickest run time, followed by iterative scaling, then BFGS.

### Inference

#### Standard errors and Z tests

A standard way to report model parameter estimates is to report them along with their corresponding standard error. This is an indication of the estimate’s precision. However, implicitly this invites comparison with zero. Such comparison is made explicit in many summary methods for models in R, with the addition of partial t or Z tests. These are tests of the null hypothesis that a parameter is equal to zero, provided that the other parameters are in the model. However, this hypothesis is generally not of interest for the worth parameters in a Plackett–Luce model: we expect most items to have some worth, the question is whether the items differ in worth. Besides, when the worth parameters are constrained to sum to one, this constrains each parameter to lie between zero and one. A Z test based on asymptotic normality of the MLE is not appropriate for worth parameters near the limits, since it does not account for the constrained parameter space.

On the log scale however, before identifiability constraints are applied, there are no bounds on the parameters. We can constrain the worth parameters by computing simple contrasts with a reference value. By default, the summary method for "PlackettLuce" objects contrasts with the worth of the first item (the first element of colnames(R)), which is equivalent to setting the worth parameter for that item to zero:

The Z tests for the item parameters test the null hypothesis that the difference from the worth of item 1 is zero. None of the Z tests provide significant evidence against this hypothesis, which is consistent with the test for equal preferences presented in Davidson (1970). The tie parameter is also shown on the log scale here, but it is an integral part of the model rather than a parameter of interest for inference, and its scale is not as relevant as that of the worth parameters.

The reference value for the item parameters can be changed via the ref argument, for example, setting to NULL sets the mean worth as the reference:

#### Quasi standard errors

As shown by the two summaries of the model fitted to the pudding data in Sect. 3.2.1, the standard error of the item parameters changes with the reference level. In some cases, there will be a natural reference, for example, when comparing an own brand product with competing brands. When this is not the case, inference can depend on an arbitrary choice. This problem can be handled through the use of quasi standard errors (Firth and De Menezes 2004; see also Firth 2003) that remain constant for a given item regardless of the reference. In particular, the quasi standard error is not changed by setting an item as the reference, so even in cases where there is a natural reference, the uncertainty around the worth of that item can be represented.

Quasi standard errors for the item parameters are implemented via a method for the qvcalc function from the qvcalc package:

Again by default, the first item is taken as the reference, but this may be changed via the ref argument. The plot method for the returned object visualizes the item parameters (log worth parameters) along with comparison intervals:

The output is shown in Fig. 1. There is evidence of a difference between any two item parameters if there is no overlap between the corresponding comparison intervals. The quasi-variances allow comparisons that are approximately correct, for every possible contrast among the parameters. The routine error report in the last two lines printed by summary(pudding_qv) tells us that, in this example, the approximation error has been very small: the approximation error for the standard error of any simple contrast among the parameters is less than 0.8%.

### Disconnected or weakly connected networks

Given a ranking of items, we can infer the wins and losses between pairs of items in the ranking. For example, the ranking {item 1} $$\succ$$ {item 3, item 4} $$\succ$$ {item 2}, implies that item 1 wins against items 2, 3 and 4; and items 3 and 4 win against item 2. These wins and losses can be represented as a directed network.

For example, consider the following artificial set of paired comparisons:

Here we have defined the paired comparisons as a matrix of numeric rankings, then used as.rankings() to create a formal rankings object (input is set to "rankings" by default). The as.rankings() function ensures that the rankings are coded as dense rankings, i.e. consecutive integers with no rank skipped for tied items; sets any rankings with less than two items to NA since these are uninformative, and adds item names if necessary.

The adjacency() function from PlackettLuce can be used to convert these rankings to an adjacency matrix, where element $$(i, j)$$ is the number of times item $$i$$ is ranked higher than item $$j$$:

Using functions from the igraphR package (Csardi and Nepusz 2006) we can visualise the corresponding network (output shown in Fig. 2):

A sufficient condition for the worth parameters (on the log scale) to have finite MLEs and standard errors is that the network is strongly connected (Hunter 2004). This means there is a path of wins and a path of losses between each pair of items. This condition is also necessary when there are no ties in the data.

In our example data, A, B and C are strongly connected. For example, C directly loses against B; C never directly beats B, but C does beat A and A, in turn, beats B, so C indirectly beats B. Similar paths of wins and losses can be found for all pairs of A, B and C. On the other hand, D is only observed to lose, therefore the MLE of its log worth would be $$-\infty$$, with infinite standard error.

If one item always wins, the MLE of its log worth would be $$+\infty$$ with infinite standard error. If the items form strongly connected clusters, that are not connected to each other, or connected only by wins or only by losses (weakly connected), then the maximum likelihood solution cannot be represented simply through a vector of relative worths. A variant of our artificial example illustrates this last point, through a pair of clusters that are not connected to one another:

In this example, there is no information that allows the relative worths of items $$A$$ and $$C$$ to be estimated. While the within-cluster worth ratios $$\alpha _A/\alpha _B$$ and $$\alpha _C/\alpha _D$$ both have MLEs equal to 1 here, between-cluster ratios such as $$\alpha _A/\alpha _C$$ are not estimable. When PlackettLuce() is applied to such a set of rankings, with argument npseudo = 0 to specify the use of maximum likelihood estimation, by design the result is an error message:

The case of weakly connected clusters is similar. If, for example, a single further ranking “$$A$$ beats $$C$$” is added to the example just shown, then the two clusters become weakly connected. The maximum likelihood solution in that case has $$\hat{\alpha }_C/\hat{\alpha }_A = \hat{\alpha }_C/\hat{\alpha }_B = \hat{\alpha }_D/\hat{\alpha }_A = \hat{\alpha }_D/\hat{\alpha }_B = 0$$, and requires the notion of a “direction of recession” (Geyer 2009) to describe it fully. This is beyond the scope of PlackettLuce(), which again by design gives an informative error message in such situations.

The remainder of this section describes two distinct approaches to the use of PlackettLuce with disconnected networks: first, the separate analysis by maximum likelihood of an identified sub-network that is strongly connected within itself; and second, a general approach that uses pseudo-rankings to fully connect the network of items being compared.

We return to the small “toy” set of rankings depicted in Fig. 2. The connectivity() function from PlackettLuce checks how connected the items are by the wins and losses recorded in the adjacency matrix:

If the network of wins and losses is not strongly connected, information on the clusters within the network is returned. In this example, we might proceed by excluding item D:

Since ABCD_rankings is a rankings object, the subset method checks the rankings are still valid after excluding item D. In this case, the paired comparison with item A only has one item left in the ranking, so it is set to NA.

The handling of any rankings that are NA is determined by the na.action argument of PlackettLuce(). This is set to "na.omit" by default, so that missing rankings are dropped. In our example, the network corresponding to the remaining rankings is strongly connected, so the MLEs in the Plackett–Luce model all have finite worth estimates and standard errors:

However, it is not necessary to exclude item D, since PlackettLuce provides a way to handle disconnected/weakly connected networks through the addition of pseudo-rankings. This works by adding wins and losses between each item and a hypothetical or ghost item with fixed worth, making the network strongly connected. This method has an interpretation as a Bayesian prior, in particular, an exchangeable prior in which all items have equal worth.

The npseudo argument defines the number of wins and loses with the ghost item that are added for each real item. The larger npseudo is, the stronger the influence of the prior; by default npseudo is set to 0.5, so that each pseudo-ranking is weighted by 0.5. This fractional weight is enough to connect the network. Note that this weight is only applied to the pseudo-rankings, the weights argument defines the weights (frequencies) for the observed rankings.

In this toy example, when we fit the model to the full data using pseudo-rankings the parameters for items B and C change quite considerably compared to the previous fit:

This is because there are only 5 rankings, so weighting each pseudo-ranking by 0.5 gives them relatively large weight. In more realistic examples, with many rankings, the default prior will have a weak shrinkage effect, shrinking the items’ worth parameters towards $$1/J$$, where $$J$$ is the number of items.

Although it is only necessary to use pseudo-rankings when the network is not strongly connected, the default behaviour is always to use them (with a weight of 0.5). This is because the shrinkage that the pseudo-rankings deliver typically reduces both the bias and the variance of the estimators of the worth parameters.

#### NASCAR example

For a practical example of modelling rankings where the underlying network is weakly connected, we consider the NASCAR data from Hunter (2004). The data are results of the 36 races in the 2002 NASCAR season in the United States. Each race involves 43 drivers out of a total of 87 drivers. The nascar data provided by PlackettLuce records the results as an ordering of the drivers in each race:

For example, in the first race, driver 83 came first, followed by driver 18 and so on. The names corresponding to the driver IDs are available as an attribute of nascar. We can provide these names when converting the orderings to rankings via the items argument:

Maximum likelihood estimation cannot be used in this example, because four drivers placed last in each race they entered. For this reason, Hunter (2004) dropped these four drivers to fit the Plackett–Luce model, which we can reproduce as follows:

In order to demonstrate the correspondence with the results from Hunter (2004), we order the item parameters by the driver’s average rank:

Now we fit the Plackett–Luce model to the full data, using the default pseudo-rankings method.

For items that were in the previous model, we see that the log worth parameters generally shrink towards zero:

The new items have relatively large negative log worth, but the estimates are nonetheless finite with finite standard error:

We can plot the quasi-variances to visualise the relative ability of all drivers (output shown in Fig. 3):

As in the pudding example, we can use summary(nascar_qv) to see a report on the accuracy of the quasi-variance approximation. In this example the error of approximation, across the standard errors of all 3741 possible simple contrasts (contrasts between pairs of the 87 driver-specific parameters), ranges between -0.7% and +6.7%. This is still remarkably accurate and means that the plot of comparison intervals is a good visual guide to the uncertainty about drivers’ relative abilities. The results of using summary(nascar_qv) are not shown here, as they would occupy too much space.

## Plackett–Luce trees

A Plackett–Luce model that assumes the same worth parameters across all rankings may sometimes be an over-simplification. For example, if rankings are made by different judges, the worth parameters may vary between groups of judges with different characteristics. Alternatively, the worth parameters may vary by conditions in which the ranking took place, such as the weather at the time of a race.

Model-based partitioning provides an automatic way to determine subgroups with significantly different sets of worth parameters. The rankings are recursively split into two groups according to their value on covariates whose values vary between rankings, and a Plackett–Luce model is fitted to each subgroup formed. The method proceeds as follows:

1. 1.

Fit a Plackett–Luce model to the full data.

2. 2.

Assess the stability of the worth parameters with respect to each covariate. That is, test for a structural change in the ranking-level contributions to the first derivatives of the log-likelihood, when these contributions are ordered by covariate values.

3. 3.

If there is significant instability, split the full data by the covariate with the strongest instability, using the cut-point that gives the highest improvement in model fit.

4. 4.

Repeat steps 1-3 for each subgroup until there are no more significant instabilities, or a split produces a subgroup below a given size threshold.

The result is a Plackett–Luce tree, since the recursive binary splits may be represented in a tree diagram and a Plackett–Luce model is fitted to the subgroups defined by each branch. This is an extension of Bradley–Terry trees, implemented in the R package psychotree (Zeileis et al. 2018) and described in more detail by Strobl et al. (2011).

### Beans example

To illustrate Plackett–Luce trees, we consider data from a trial in Nicaragua of ten different varieties of bean, run by Bioversity International (Van Etten et al. 2016). Farmers were asked to grow three out of the ten trial varieties in one of the growing seasons, Primera (May–August), Postrera (September–October) or Apante (November–January). At the end of the season, they were asked which variety they thought was best and which variety they thought was worst, to give a ranking of the three varieties. In addition, they were asked to compare each trial variety to the standard local variety and say whether it was better or worse.

The data are provided as the dataset beans in PlackettLuce. The data require some preparation to collate the rankings, which is detailed in “Appendix 2”. The same code is provided in the examples section of the help file of beans

The result is a rankings object R with all rankings of the three trial varieties and the output of their comparison with the local variety.

In order to fit a Plackett–Luce tree, we need to create a "grouped_rankings" object, that defines how the rankings map to the covariate values. In this case we wish to group by each record in the original data set. So we group by an index that identifies the record number for each of the four rankings from each farmer (one ranking of order three plus three pairwise rankings with the local variety):

For each record in the original data, we have three covariates: season the season-year the beans were planted, year the year of planting, and maxTN the maximum temperature at night during the vegetative cycle. The following code fits a Plackett–Luce tree with up to three nodes and at least 5% of the records in each node:

The algorithm identifies three nodes, with the first split defined by high night-time temperatures and the second split separating the single Primera season from the others. So for early planting in regions where the night-time temperatures were not too high, INTA Rojo (variety 7) was most preferred, closely followed by the local variety. During the regular growing seasons (Postrera and Apante) in regions where the night-time temperatures were not too high, the local variety was most preferred, closely followed by INTA Sequia (8). Finally in regions where the maximum night-time temperature was high, INTA Sequia (8) was most preferred, closely followed by BRT 103-182 (2) and INTA Centro Sur (3). A plot method is provided to visualise the tree (output shown in Fig. 4):

Note that when working with high dimensional data in particular, tree models can be quite unstable. For a more robust model, an ensemble method such as bagging or random forests could be employed (Strobl et al. 2009).

## Discussion

PlackettLuce is a feature-rich package for the analysis of ranking data. The package provides methods for importing and handling partial rankings data, and for estimation and inference from a generalization of the Plackett–Luce model that can handle subset rankings and ties of arbitrary order. Disconnected item networks are handled by appropriately augmenting the data with pseudo-rankings for a hypothetical item. The package also facilitates the construction of Plackett–Luce trees, which allow item worth parameters to vary between rankings, according to available covariate information on the rankings.

A disadvantage of modelling ties explicitly is that the normalising constant in Eq. (2), that normalises over all possible choices for a given rank, involves a number of terms that increases rapidly with the tie order $$D$$. This causes a corresponding increase in the run time of PlackettLuce(). Also, as noted in Sect. 2.2.3, our current method for estimating the variance-covariance matrix relies on the technique of expanding rankings into counts. When this expansion is large, estimating the standard errors for the model parameters can take almost as long as fitting the model, or can even become infeasible due to memory limits. Improving the efficiency of both the model-fitting and the variance-covariance estimation is a focus of current work.

Meanwhile, we expect that PlackettLuce() and all its methods work well on data with ties up to order 4, for moderate numbers of items and unique rankings. As a guide, for 5000 sub-rankings of 10 items, out of 100 items in total, with ties up to order 4, a test model fit took 18.1 s and variance-covariance estimation took 14.8 s on the same machine used in the performance evaluation of Sect. 2.1.4. This makes PlackettLuce suitable for applications such as sport, where ties of order greater than 3 are typically extremely rare. For data with higher-order ties, a method that does not explicitly model tied events, such as the rank-ordered logit model or the Mallows distance-based model, may be more suitable.

Our plans for future development of PlackettLuce include support for online estimation from streams of partial rankings and formally accounting for spatio-temporal heterogeneity in worth parameters. Algorithms for online estimation will be more efficient, enabling the package to handle larger data sets. To realise this potential, alternative data structures may be required for partial rankings, which are currently represented as full rankings with some zero ranks.

Other potential directions for future work include implementing some of the features offered by other software. In particular, top $$n$$ rankings often arise in practice and are compatible with the algorithms that are currently implemented. Modelling top $$n$$ rankings is more computationally demanding than modelling subset rankings of the same length, since each choice in the ranking selects from a wider set of alternatives. So supporting top $$n$$ rankings will have wider practical application if it is paired with a more efficient algorithm and/or a more efficient implementation using compiled code.