ST10CH24_Raftery ARjats.

cls February 14, 2023 14:5

Annual Review of Statistics and Its Application

Model-Based Clustering
Isobel Claire Gormley,1 Thomas Brendan Murphy,1,2
and Adrian E. Raftery3
1
School of Mathematics and Statistics, University College Dublin, Dublin, Ireland;
email: [email protected], [email protected]
2
Collegium de Lyon Institut d’Études Avancées and ERIC, Université de Lyon, Lyon, France
3
Department of Statistics and Department of Sociology, University of Washington, Seattle,
Washington; email: [email protected]

Annu. Rev. Stat. Appl. 2023. 10:573–95 Keywords

First published as a Review in Advance on
clustering, mixture models, expectation–maximization algorithm, Bayesian
October 21, 2022
inference, software
The Annual Review of Statistics and Its Application is
online at statistics.annualreviews.org Abstract
https://doi.org/10.1146/annurev-statistics-033121-
Clustering is the task of automatically gathering observations into ho-
115326
mogeneous groups, where the number of groups is unknown. Through
Copyright © 2023 by the author(s). This work is
its basis in a statistical modeling framework, model-based clustering pro-
licensed under a Creative Commons Attribution 4.0
International License, which permits unrestricted vides a principled and reproducible approach to clustering. In contrast to
use, distribution, and reproduction in any medium, heuristic approaches, model-based clustering allows for robust approaches
provided the original author and source are credited.
to parameter estimation and objective inference on the number of clus-
See credit lines of images or other third-party
material in this article for license information. ters, while providing a clustering solution that accounts for uncertainty in
cluster membership. The aim of this article is to provide a review of the
theory underpinning model-based clustering, to outline associated inferen-
tial approaches, and to highlight recent methodological developments that
facilitate the use of model-based clustering for a broad array of data types.
Since its emergence six decades ago, the literature on model-based cluster-
ing has grown rapidly, and as such, this review provides only a selection of
the bibliography in this dynamic and impactful field.

573
ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

1. INTRODUCTION
Clustering methods address the challenge of collecting observations into a number of homoge-
neous groups based on quantitative measurements on the observations. While humans seemingly
perform such a task effortlessly when the number of observations, measurements, and groups is
small, clustering is a difficult problem: The number of clusters is unknown, as is the clustering
solution, i.e., the cluster membership of each observation. Many approaches to inferring a clus-
tering solution exist and, in general, can themselves be grouped into heuristic and model-based
approaches. Heuristic, algorithmic approaches often rely on dissimilarity measures between ob-
servations and clusters to provide a clustering solution. Such approaches are typically intuitive and
computationally efficient but require many subjective decisions (e.g., which dissimilarity measure
to use, how dissimilar two clusters are, and how many clusters are present), meaning repro-
ducibility and robustness are often lacking. Model-based approaches infer the clustering solution
by employing a statistical modeling framework and by applying standard statistical inferential
methods, allowing for objective and robust inference.

1.1. A Brief History

Grouping objects and individuals with similar characteristics is inherent to language. Perhaps the
first to formalize it was Plato, with his theory of forms, defining a form as an abstract idea, of which
there may be many instances in practice. Aristotle, in his History of Animals, classified animals into
groups based on their characteristics, for the first time drawing heavily on empirical observations.
More systematically, Linnaeus (1753) classified biological species based on measured empirical
characteristics.
Cluster analysis is something more: the search for groups in quantitative data using system-
atic numerical methods, first proposed by Czekanowski (1909). Cluster analysis developed rapidly
in the 1950s, driven by the needs of biological taxonomy and marketing, yielding heuristic algo-
rithmic methods such as single link hierarchical agglomerative clustering (Sneath 1957), and the
average link and complete link methods (Sokal & Michener 1958). These methods were heuristic,
and developed largely in isolation from the statistical methods based on probability models that
were emerging rapidly at the same time. The heuristic methods worked well overall but still did
not answer questions that arose frequently, such as which method should be used in a particular
setting, how many clusters there are, and how outliers should be handled.
Answers to questions like these were to come from expressing cluster analysis in terms of a
probability model, leading to model-based clustering. This was first done for multivariate discrete
data, in the form of the latent class model, in which several discrete characteristics are measured
on each object, the objects are assumed to be grouped, and within each group the characteristics
are statistically independent (Lazarsfeld 1950a,b).
The dominant model for clustering continuous-valued data is the mixture of multivariate nor-
mal distributions, proposed for this purpose by Wolfe (1965, 1967, 1970) and independently by
Day (1969). Binder (1978) developed a Bayesian estimation method for the model, while Murtagh
& Raftery (1984) developed a model-based clustering method based on the eigenvalue decomposi-
tion of the component covariance matrices. This was built on by Banfield & Raftery (1989, 1993),
who also coined the term “model-based clustering.” Celeux & Govaert (1995) introduced max-
imum likelihood estimation for this family of models using the expectation–maximization (EM)
algorithm (Dempster et al. 1977).
Articles reviewing the literature on model-based clustering include those of Fraley & Raftery
(1998, 2002), Ahlquist & Breunig (2012), and McNicholas (2016b). More extensive reviews in

574 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Figure 1
Pairwise scatterplots of the flow cytometry data used to illustrate model-based clustering. The fluorescence
intensity at each of four cell surface markers, SLP76, ZAP70, CD4, and CD45RA, is illustrated. Figure
adapted with permission from O’Hagan et al. (2016).

book form include those of McLachlan & Basford (1988), McLachlan & Peel (2000), McNicholas
(2016a), and Bouveyron et al. (2019).

1.2. A Motivating Example

To motivate and illustrate model-based clustering, we draw on a problem from the field of flow
cytometry that allows for rapid single cell analysis by measuring fluorescence intensity across a
number of cell surface markers (Maier et al. 2007). Figure 1 illustrates 4,669 observations with
four cell surface markers (SLP76, ZAP70, CD4, and CD45RA) from the T cell phosphorylation
data outlined by Pyne et al. (2009). The challenge here is to identify discrete cell populations using
the four cell surface markers only, where the number of cell populations (i.e., clusters) is unknown,
as is the clustering solution. Figure 1 illuminates many of the difficulties that clustering poses,
even in this relatively low-dimensional example: The number of clusters present is unclear, and
different cell markers are suggestive of different numbers of clusters. There are different degrees

www.annualreviews.org • Model-Based Clustering 575

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

of separation between some clusters, and the distribution of data within clusters varies. A model-
based approach to clustering these data elegantly addresses each of these challenges in a unified,
principled framework that allows for reproducibility and uncertainty quantification.
The clustering solution resulting from the application of model-based clustering to these flow
cytometry data is given in Section 2, which reviews the statistical framework underpinning model-
based clustering. Section 3 discusses maximum likelihood and Bayesian inferential approaches,
and model-selection tools used to identify the number of clusters present and the form of the
selected model. Section 4 reviews some recent methodological advances in model-based clustering
across a range of data modes, highlighting the preeminent software available to facilitate practical
implementation. The article concludes in Section 5 with a discussion, focusing on pending issues
in the field of model-based clustering.

2. THE MODEL-BASED CLUSTERING FRAMEWORK

2.1. A Finite Mixture Model
Model-based clustering assumes a probability distribution for the data, typically a finite mixture of
G multivariate distributions. Suppose that for each of n observations we have data on d variables,
denoted yi = (yi,1 , . . . , yi,d ) for observation i. Thus, the probability model is a weighted average of
G probability density functions, i.e.,
∑
G
p(yi ) = τg fg (yi |θ g ), 1.
g=1

where the gth mixing proportion τ g denotes the probability that observation i’s data were gen-
∑
erated by the gth density, where τ g ≥ 0 for g = 1, 2, . . . , G and Gg=1 τg = 1. The density of the
gth mixture component is fg (·|θ g ), where its parameters are collected in θ g . The “model-based
clustering” terminology is often synonymous with the assumption that fg (·|θ g ) is a multivariate
Gaussian distribution, meaning that Equation 1 is then a Gaussian mixture model. In such a
setting, θ g = {µg , 6g }, where µg and 6g denote the mean and covariance of the gth component,
respectively (Banfield & Raftery 1993, Celeux & Govaert 1995). Typically each mixture com-
ponent is taken to correspond to a different cluster, but in cases where a cluster may be better
represented by a mixture of Gaussian distributions rather than by a single Gaussian distribution,
components can be combined to provide more substantively refined clustering solutions (Baudry
et al. 2010, Hennig 2010).
For data with large d, employing a full covariance matrix for each mixture component requires
estimation of a large number of parameters, which can lead to a lack of precision, generalizability,
and interpretability. Thus it is common to employ more parsimonious component covariance
matrices by considering their geometric interpretation. Under the Gaussian density assumption,
clusters are ellipsoidal, with their volume, shape, and orientation determined by the component
covariance matrices. The widely used software for the implementation of model-based clustering,
mclust (Scrucca et al. 2016, 2022), considers an eigen-decomposition of the covariance matrix 6g
of the form
6g = λg Dg Ag D⊤
g ,

where Dg is the matrix of eigenvectors of 6g , Ag is a diagonal matrix with elements that are pro-
portional to the eigenvalues of 6g in descending order, and λg is the associated proportionality
constant. Geometrically, λg controls the volume of the ellipsoid, Ag specifies the shape of the
density contours, and Dg determines the orientation of the corresponding ellipsoid (Banfield &
Raftery 1993, Celeux & Govaert 1995). The volume, shape, and orientation of the cluster densities
can be constrained to be equal (E) or variable (V) across clusters.
576 Gormley • Murphy • Raftery
ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Table 1 Parameterizations of the covariance matrix 6g available in mclust and the associated
number of covariance parameters ν when G = 9 for d = 4 and d = 40
ν
Model name Description d=4 d = 40
EII Spherical, equal volume 1 1
VII Spherical, varying volume 9 9
EEI Diagonal, equal volume and shape 4 40
VEI Diagonal, equal shape 12 48
EVI Diagonal, equal volume, varying shape 28 352
VVI Diagonal, varying volume and shape 36 360
EEE Ellipsoidal, equal volume, shape and orientation 10 820
VEE Ellipsoidal, equal shape and orientation 18 828
EVE Ellipsoidal, equal volume and orientation 34 1132
VVE Ellipsoidal, equal orientation 42 1140
EEV Ellipsoidal, equal volume and shape 58 7060
VEV Ellipsoidal, equal shape 66 7068
EVV Ellipsoidal, equal volume 82 7372
VVV Ellipsoidal, varying volume, shape, and 90 7380
orientation

The three letters in the model name denote, in order, the characteristics of the volume, shape, and orientation across
clusters. Abbreviations: E, equal; I, spherical; V, varying.

Table 1 details the constraints on the volume, shape, and orientation of clusters, and associated
model names, for all models currently available in mclust. Each model is denoted by a three-letter
code, the volume-shape-orientation representation: The first letter denotes whether the volume
is constrained to be equal (E) or varies (V) across clusters; the second letter denotes whether the
shape is constrained to be equal (E) or varies (V) across clusters or if the clusters are spherical (I);
and the final letter, which refers to the clusters’ orientation, is subject to a similar interpretation.
To demonstrate the parsimony gained by such a representation of 6g , the number of covariance
parameters ν in each model is detailed for d = 4, as in the flow cytometry data in Figure 1, under
each mclust model with G = 9. For additional context, the number of covariance parameters in
each model for a higher-dimensional setting where d = 40 and G = 9 is also detailed.
Figure 2 illustrates the clustering solution resulting from the application of model-based clus-
tering (via mclust) to the flow cytometry data. A finite mixture of G = 9 ellipsoidal multivariate
Gaussian distributions with equal shape (i.e., the VEV mclust model) was deemed optimal by
the Bayesian information criterion (BIC) (see Section 3.3). However, combining these compo-
nents hierarchically according to an entropy criterion (Baudry et al. 2010) strongly indicates the
six-cluster solution illustrated in Figure 2.

2.2. Beyond the Multivariate Gaussian Distribution

Model-based clustering’s use of a probability model for the data easily allows for clustering of
different data modes by employing an apposite density (or mass function) fg (·|θ g ) for the data
under study. For example, for multivariate continuous data with heavier tails than allowed for
by the Gaussian, a mixture model in which fg (·|θ g ) is a multivariate t-distribution may be em-
ployed for clustering. For the nonelliptically shaped flow cytometry data in Figure 1, the FLAME
(flow analysis with automated multivariate estimation) method of Pyne et al. (2009) suggested a

www.annualreviews.org • Model-Based Clustering 577

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Figure 2
The six-cluster solution resulting from the application of model-based clustering to the flow cytometry data, with subsequent cluster
merging. Observations within a cluster are represented by points with the same color and shape. The observation with the lowest
clustering uncertainty (see Section 3.1) is illustrated by the magenta circle; it is clustered with observations in cluster 1, illustrated by
black circles. The observation with the highest clustering uncertainty is the yellow square; it is clustered with observations in cluster 2,
illustrated by light blue crosses. SLP76, ZAP70, CD4, and CD45RA are four cell surface markers at which the fluorescence intensity is
measured.

mixture of five skew-t distributions as optimal, while O’Hagan et al. (2016) proffer a mixture of six
multivariate normal inverse Gaussian distributions. McLachlan et al. (2019) provide a review of
finite mixture models with nonnormal but continuous components, including robust mixtures of
Gaussian distributions such as Banfield & Raftery’s (1993) mixture of Gaussian distributions with
a noise component and the trimming approaches of García-Escudero et al. (2008, 2018).
The finite mixture structure can additionally accommodate different forms of data: high-
dimensional data (i.e., n j p) can be clustered using a mixture of latent factor models (McNicholas
& Murphy 2008, Montanari & Viroli 2010, Viroli & McLachlan 2019), and mixtures of la-
tent space models permit clustering of nodes in a network (Handcock et al. 2007, Gormley &
Murphy 2010). Model-based clustering methods for clustering functional data (Bouveyron &
Jacques 2011, Jacques & Preda 2014) or text data (Viroli & Anderlucci 2021) provide further ex-
amples. This adaptability of the model-based clustering framework to appositely model a variety

578 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

of data types and forms ensures its widespread applicability; Section 4 provides a deeper overview
of model-based clustering beyond the Gaussian setting.

2.3. Beyond the Finite Mixture Model

While the finite mixture model is the prevalent basis for model-based approaches to clustering,
many modifications or extensions have been developed. Within the Bayesian paradigm, inference
for the finite mixture model described in Section 2.1 is straightforward (see Section 3.2) but ex-
tensions including sparse finite and infinite mixture models are naturally accommodated. Careful
specification of the hyperparameters of a Dirichlet prior on a finite mixture model’s mixing pro-
portions gives rise to the sparse finite mixture model (Rousseau & Mengersen 2011, Van Havre
et al. 2015, Malsiner-Walli et al. 2016) while the use of a Dirichlet process prior, for example, pro-
vides an infinite mixture model (Papastamoulis 2018, Hejblum et al. 2019, Murphy et al. 2020).
Frühwirth-Schnatter & Malsiner-Walli (2019) discuss links between sparse finite and Dirichlet
process mixtures.
This array of mixture models typically assume the observations are independent, but in cer-
tain settings such an assumption may be invalid. Modifying the basic finite mixture model to
account for longitudinal measures or panel-type data (Frühwirth-Schnatter 2006) is feasible given
its probabilistic basis. In cases where concomitant information is available, mixture-of-experts
models can be employed; in these models the mixing proportions and/or the component den-
sity parameters are modeled as a function of the concomitant variables ( Jordan & Jacobs 1994,
Gormley & Frühwirth-Schnatter 2019). Again, Section 4 provides a deeper overview of
model-based clustering beyond the basic finite mixture model setting.

3. STATISTICAL INFERENCE FOR MODEL-BASED CLUSTERING

While early approaches to inferring the parameters of the mixture model used the method of
moments (Pearson 1894), modern approaches, particularly in the context of model-based cluster-
ing, predominantly rely on maximum likelihood estimation or Bayesian inference. The clustering
problem can be viewed as an incomplete data problem or a latent variable problem, where the
cluster membership of each observation is unobserved or latent. This view lends the clustering
problem well to inference via the well-known EM algorithm (McLachlan & Krishnan 2008) in a
maximum likelihood setting, or to Bayesian inference (Bensmail et al. 1997, Frühwirth-Schnatter
2006) given its elegant handling of latent variables.

3.1. Maximum Likelihood Inference

Since the formalization of the EM algorithm by Dempster et al. (1977), the majority of model-
based clustering applications use the EM algorithm for inference (McLachlan et al. 2019). In this
setting the data are viewed as (yi , zi ) for i = 1, . . . , n, where yi denotes the observed data on d
variables as before, and zi = (zi,1 , . . . , zi,G ) is the unobserved portion of the data. Specifically, we
define {
1 if i belongs to cluster g
zi,g =
0 otherwise.

Thus, each of z1 , . . . zn , assumed to be independent and identically distributed, has a multinomial

distribution with G categories, with probabilities τ 1 , . . . , τ G , respectively. Rather than directly
maximizing the observed data likelihood function
∏
n ∑
G
L(θ, τ|y) = p(y|θ, τ) = τg fg (yi |θ g ),
i=1 g=1

www.annualreviews.org • Model-Based Clustering 579

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

where y = (y1 , . . . , yn ), θ = (θ 1 , . . . , θ G ), and τ = (τ1 , . . . , τG ), the EM algorithm works with the

complete data likelihood function
∏
n ∏
G
[ ]z
LC (θ, τ, z|y) = τg fg (yi |θ g ) i,g , 2.
i=1 g=1

where z = (z1 , . . . , zn ).
The EM algorithm is an iterative algorithm where each iteration consists of an expectation step
(E-step) and a maximization step (M-step). In the E-step, the conditional expectation of the com-
plete data log-likelihood function, given the observed data and the current parameter estimates, is
computed. In the M-step, the expected complete data log-likelihood function from the E-step is
maximized with respect to the model parameters. Iterating these E- and M-steps until convergence
achieves at least a local maximum of the observed data likelihood function, under mild regularity
conditions (Dempster et al. 1977). As with any iterative algorithm, convergence of the EM al-
gorithm can be assessed in several ways, predominantly by tracking the change in log-likelihood
and/or parameter estimates between successive iterations, or using Aitken’s acceleration-based
stopping criterion (McLachlan & Peel 2000, McLachlan & Krishnan 2008). Starting values for
the EM algorithm also require careful consideration; O’Hagan et al. (2012) propose some useful
initialization strategies.
In practice, for the complete data log-likelihood function (Equation 2), where fg (·|θ g ) is the
multivariate Gaussian distribution, the E-step involves computing the conditional expectation of
zi, g for i = 1, . . . , n and g = 1, . . . , G given the current parameter estimates and the data y. Given
the expected values ẑ, the M-step involves maximizing the expected complete data log-likelihood
function with respect to the mixing proportions and mean parameters for which solutions are
available in closed form; closed form solutions for the covariance matrices are available for some
covariance parameterizations. On convergence, the value of ẑi,g , the conditional expectation of zi, g ,
is the estimated conditional probability that observation i belongs to cluster g. Thus, a hard classifi-
cation of cluster membership for each observation is available through allocating each observation
to the cluster g′ for which {g′ |ẑi,g′ = maxg ẑi,g }, and the uncertainty in that cluster membership is
quantified by (1 − maxg ẑi,g ) for observation i (Bensmail et al. 1997).
Figure 3 illustrates the clustering uncertainty for each of the observations in the flow cytom-
etry data set under the BIC-optimal mixture of G = 9 Gaussian distributions. The maximum
possible uncertainty, i.e., 1 − 1/G, is highlighted; in general, the clustering uncertainty is low, and
particularly so for some observations. In Figure 2, the magenta circle illustrates the observation
with the lowest clustering uncertainty (<10−14 ) under the optimal mclust G = 9 VEV solution;
this observation is clustered with observations in cluster 1, illustrated by black circles. The yel-
low square is the observation with the highest clustering uncertainty (of 0.68) under the optimal
mclust G = 9 VEV solution; this observation is clustered with observations in cluster 2, illustrated
by light blue crosses. The yellow square lies on the boundary of several clusters, and thus the clus-
ter membership of this observation is highly uncertain. The model-based approach to clustering
facilitates the quantification of this uncertainty within a principled probabilistic framework.

3.2. Bayesian Inference

In the Bayesian framework, it is intuitive to view the finite mixture model as a hierarchical la-
tent variable model where the distribution of each observation’s data depends on a discrete latent
variable that indicates component membership. These latent variables, s = (s1 , . . . , sn ), where si 
{1, . . . , G}, are typically assumed to be independent and have a multinomial distribution with one

580 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Figure 3
The clustering uncertainty for each of the observations in the flow cytometry data. Each observation is
colored according to its cluster membership under the optimal G = 9 VEV mclust model. VEV indicates
that the volume varies across clusters, the shape is constrained to be equal, and the orientation varies.

trial and category probabilities τ = (τ1 , . . . , τG ). This gives the complete data likelihood
∏
n ∏
G
p(y, s|θ, τ) = {p(yi |θ g )τg }1{si =g} , 3.
i=1 g=1

where 1{·} denotes the indicator function. In this setting, Bayesian inference on (s, θ, τ) typically
proceeds by sampling from the complete data posterior distribution p(s, θ, τ|y). This posterior
distribution is available by combining the complete data likelihood (Equation 3) with prior
distributions on θ and τ through Bayes’ theorem, i.e.,

p(s, θ, τ|y) ∝ p(y|s, θ)p(s|τ)p(θ)p(τ ),

where independence between θ and τ is assumed a priori. Through this data augmentation ap-
proach (Tanner & Wong 1987), it is straightforward to sample from the posterior distribution
using Markov chain Monte Carlo (MCMC) methods, in particular through the use of Gibbs
sampling. The importance of Gibbs sampling for Bayesian estimation of mixture models is well
established; early work includes that of Diebolt & Robert (1994), Mengersen & Robert (1996), and
Raftery (1996), with Frühwirth-Schnatter (2006) and Frühwirth-Schnatter et al. (2019) providing
comprehensive resources.
Specification of the prior distributions on the component parameters depends on the data
mode. For a finite mixture of Gaussian distributions, Frühwirth-Schnatter & Malsiner-Walli

www.annualreviews.org • Model-Based Clustering 581

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

(2019) recommend putting a normal-gamma prior on component means and a conjugate hierar-
chical Wishart prior on the component precision matrices. For the mixing proportions, a Dirichlet
prior D(α) is typically used, and much work has been devoted to discussion of the choice of α. Based
on results of Rousseau & Mengersen (2011) and Frühwirth-Schnatter (2011), Malsiner-Walli et al.
(2016) encourage small values of α in a sparse finite mixture (i.e., a mixture in which G is most
likely larger than the number of clusters) to allow emptying of superfluous components.
The number of components G need not be finite, and the Bayesian framework naturally han-
dles infinite mixture models, such as Dirichlet process mixture models. Using a Dirichlet process
prior (Ferguson 1973), Dirichlet process mixtures have been well utilized as model-based cluster-
ing tools (e.g., Quintana & Iglesias 2003). Bayesian inference for the Dirichlet process mixture
model proceeds via full conditional MCMC sampling (Ishwaran & James 2001). More flexible al-
ternatives, such as the Pitman-Yor process prior (of which the Dirichlet process is a special case),
have also been employed in a model-based clustering setting, e.g., by Murphy et al. (2020), where
slice sampling (Kalli et al. 2011) is invoked for inference.
Sparse finite mixture models (Malsiner-Walli et al. 2016, Frühwirth-Schnatter & Malsiner-
Walli 2019), which are also termed overfitted mixture models in the literature (Rousseau &
Mengersen 2011), offer a bridge between finite mixture and Dirichlet process mixture models.
Frühwirth-Schnatter & Malsiner-Walli (2019) develop sparse finite mixture models to cluster a
broad range of non-Gaussian data; Bayesian inference here is a straightforward extension of in-
ference for a standard finite mixture (Frühwirth-Schnatter 2006), requiring only one additional
step. Furthermore, Frühwirth-Schnatter & Malsiner-Walli (2019) unify and compare sparse finite
mixtures and Dirichlet process mixtures, highlighting the importance of placing hyperpriors on
the Dirichlet parameters and thereby linking the two approaches.
The identifiability of finite mixture distributions (Teicher 1963) requires special attention when
performing inference in the Bayesian paradigm. Frühwirth-Schnatter (2006) outlines three types
of nonidentifiability: invariance to relabeling the components of the mixture model, nonidenti-
fiability due to potential overfitting, and a generic nonidentifiability that occurs only for certain
classes of mixture distributions. To draw valid posterior inference, the invariance to relabeling
characteristic requires particular attention. An array of approaches have been developed to ad-
dress this so-called label-switching problem: Stephens (2000b) proposes a loss function based
approach, Celeux et al. (2000) outline a clustering approach, and Murphy et al. (2020) employ
a cost-minimizing permutation given by the square assignment algorithm (Carpaneto & Toth
1980). Jasra et al. (2005) provide an encompassing review.

3.3. Model Selection

A central challenge in many clustering problems is inferring the number of clusters G that are
present. While the adaptability of model-based clustering to different data modes and forms is
advantageous, it can lead to the additional need to choose between competing component den-
sities. The probabilistic model underpinning model-based clustering allows these choices to be
jointly addressed using objective and statistically principled model-selection tools.
A typical application of model-based clustering involves fitting a number of models,
M1 , . . . , MK , to the data y, where each model has a different value of G and a different form
of component densities. In mclust, for example, models with different values of G and different
constraints on the covariance matrix (see Table 1) may be fitted. A natural approach to select-
ing the optimal model is then to consider the model that is most likely a posteriori. Assuming all
fitted models are equally probable a priori, this amounts to choosing the model with the highest
marginal likelihood. Evaluating the integral that defines the marginal likelihood of model Mk can

582 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

be difficult, but for regular models it can be approximated by the BIC:

BIC = 2 log p(y|Mk ) ≈ 2 log p(y|θ̂, τ̂) − νMk log(n),

where θ̂ and τ̂ denote the maximum likelihood estimates of θ and τ, respectively, and νMk is the
number of independent parameters to be estimated in model Mk . While the necessary regularity
conditions do not hold for mixture models in general (Aitkin & Rubin 1985), Keribin (2000)
showed that the BIC is consistent for the number of groups, subject to the assumption of a bounded
likelihood. While a bounded likelihood is not the case in general for Gaussian mixtures, mild
restrictions, such as ensuring a small lower bound on the smallest eigenvalue of the covariance
matrices, can ensure this. Moreover, the BIC has been widely used in clustering practice with
good results (e.g., Fraley & Raftery 2002, McNicholas & Murphy 2008, McParland et al. 2017,
Murphy et al. 2021).
Figure 4 illustrates the BIC values resulting from fitting 210 different mclust models to the
flow cytometry data. Each model had G  {1, . . . , 15} and used one of the 14 covariance con-
strained Gaussian densities detailed in Table 1. The optimal model was the G = 9 VEV model,
which requires estimation of 110 parameters, but there were many closely competing models.
The next best model, according to the BIC, was the G = 12 VVI model requiring 107 parameters,
while the third best was the G = 11 VVI model with 103 parameters. This illustrates the model

Figure 4
The BIC values for all 14 mclust models fitted to the flow cytometry data, with the number of components
G = 1, . . . , 15. The optimal model, as deemed by the BIC, is the G = 9 VEV model. The three letters in the
model name denote, in order, the volume, shape, and orientation across clusters. Abbreviations: BIC,
Bayesian information criterion; E, equal; I, spherical; V, varying.

www.annualreviews.org • Model-Based Clustering 583

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

complexity and number of clusters trade-off in model selection in the context of model-based clus-
tering: Where the more parsimonious VVI model is fitted, more clusters are required to provide
a good representation of the data. Where the more complex VEV model is fitted, fewer clusters
are sufficient.
Many other approaches to model selection have been proposed; Steele & Raftery (2010) pro-
vide a review. While the BIC seeks out the number of components in a mixture model, if interest
is primarily in the clustering solution, the integrated completed likelihood (Biernacki et al. 2000)
provides an alternative to the BIC. The well-known Akaike information criterion (Schwarz 1978)
tends to overestimate the number of components in a mixture model; the deviance informa-
tion criterion (Spiegelhalter et al. 2002), which is an Akaike information criterion–like likelihood
penalization criterion, has shown varied performance.
In the Bayesian framework, inferring the number of clusters is naturally available through the
posterior mode, or maximum a posteriori estimate of G. For finite mixtures this can be computed
by reversible jump MCMC (Richardson & Green 1997) or by birth-death processes (Stephens
2000a), among other methods. For sparse finite and infinite mixture models, inferring the number
of clusters is somewhat automatic, for example, by examining the posterior modal number of
nonempty clusters. Frühwirth-Schnatter (2006) covers many approaches to inferring the number
of components in Bayesian mixture models.
Model-selection criteria have also been used in the Bayesian clustering context. For example,
for finite mixtures, Murphy et al. (2020) use the BIC-MCMC of Frühwirth-Schnatter (2011).
Raftery et al. (2007) propose the BIC–Monte Carlo (BICM), where BICM = 2 log L̃ − 2sl2 log(n).
Here L̃ and sl2 denote, respectively, the largest value and the sample variance of the log-likelihoods
in the posterior sample. Murphy et al. (2020) employ the BICM for a finite mixture of infinite
latent factor models; the BICM is particularly useful in the context of such nonparametric models
where the number of free parameters can be difficult to quantify.
Including all available variables in a cluster analysis can result in a poor clustering solution
if some of the variables cloud the clustering structure. In the model-based framework, selecting
the informative clustering variables is naturally framed as a model selection problem; Section 4.9
outlines some approaches to address this issue.

4. MODEL-BASED CLUSTERING IN PRACTICE

The model-based clustering approach has been developed to account for a wide range of data
structures beyond the finite Gaussian mixture model described in Section 2.1. Thus, the structure
of the data being clustered can be appropriately accounted for using the model-based approach.
In this section we outline a selection of such extensions.
A wide range of software packages are available for implementing model-based clustering for
different data types. The availability of open-source software has been a particular strong point of
the model-based clustering research effort, dating back at least as far as Wolfe (1967). In addition
to the software highlighted in this section, the sidebar titled Software describes aspects of five
model-based clustering software packages that are particularly widely used.

4.1. Continuous Data

The Gaussian mixture model (Section 2.1) is the most common approach for model-based clus-
tering of multivariate continuous data. The model has been used since the origins of model-based
clustering (see also Section 1.1).
The mclust software package (Banfield & Raftery 1993, Fraley & Raftery 2002, Scrucca
et al. 2016) is one of the most commonly used software packages for model-based clustering of
continuous data. The model underlying mclust assumes that the data arise from a finite Gaussian

584 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

SOFTWARE
While a range of software packages is available to implement model-based clustering, here we describe five widely
used packages.
■ The preeminent software package for model-based clustering is mclust (Scrucca et al. 2016, 2022), which is
used to cluster continuous data. The software uses Gaussian mixture models with parsimonious covariance
structures and facilitates model-based clustering, classification, density estimation, dimension reduction, and
visualization.
■ The poLCA software package (Linzer & Lewis 2011) is a widely used package for clustering categorical data.
Based on the latent class analysis (LCA) model, poLCA uses the EM algorithm and Newton-Raphson for model
fitting. The package includes latent class regression, which is an extension of LCA to include covariates.
■ The clustMD software package (McParland & Gormley 2017) clusters mixed data. The software can model
data consisting of continuous, binary, ordinal and/or nominal variables using a parsimonious mixture of latent
Gaussian variable models.
■ The flexmix software package (Leisch 2004; Grün & Leisch 2007, 2008) implements model-based clustering
for a range of component distributions. The package can handle a wide range of data types (continuous,
categorical, count, etc.) and allows for clustering with covariates.
■ The mixtools software package (Benaglia et al. 2009) contains approaches for clustering different data types
and for model-based clustering with covariates.

mixture model where the component covariance matrices are constrained to have a parsimonious
structure. Celeux & Govaert (1995) developed a similar model-based clustering approach called
Gaussian parsimonious clustering models (GPCM); the Rmixmod software (Lebret et al. 2015)
implements this approach.
Alternative Gaussian model–based clustering methods have been developed by assuming dif-
ferent covariance structures. Examples include pgmm (McNicholas & Murphy 2008, McNicholas
et al. 2021), which is based on the mixture of factor analyzers model (Ghahramani & Hinton 1996,
McLachlan et al. 2003); longclust (McNicholas & Murphy 2010, McNicholas et al. 2019), which
uses a constrained Cholesky decomposition for the component covariances; and mixggm (Fop et al.
2019), which assumes a sparse covariance structure based on Gaussian graphical models.
Many extensions of model-based clustering for continuous data have been developed by re-
laxing the Gaussian assumption for the component densities. The multivariate t-distribution has
been proposed by McLachlan & Peel (1998) and Stephens (2000a) to accommodate heavy tailed
data clusters. The EMMIX software (McLachlan et al. 1999) was developed for fitting mixtures of
multivariate t-distributions for clustering heavy tailed continuous data.
A wide range of more flexible component densities that accommodate heavy tails and skewness
have also been proposed. Examples include model-based clustering methods based on the skew-
normal and t-distributions (e.g., Lee & McLachlan 2018) and mixtures of generalized hyperbolic
distributions (e.g., Morris & McNicholas 2016, Tang et al. 2018).
Lee & McLachlan (2013), McNicholas (2016a), McLachlan et al. (2019), Bouveyron et al.
(2019, chapter 9), and Lee & McLachlan (2022) give an overview of many non-Gaussian mixture
models proposed for clustering continuous data.

4.2. Categorical Data

Multivariate categorical data arise in a wide range of application domains. The latent class anal-
ysis (LCA) model (Lazarsfeld 1950a,b) (see also Section 1.1) is the primary model that is used

www.annualreviews.org • Model-Based Clustering 585

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

for clustering multivariate categorical data. LCA is based on assuming that the data arise from a
finite mixture model, where the component distribution assumes that the categorical variables are
conditionally independent given the cluster memberships; this is known as the local independence
assumption.
Extensions of the latent class model have been proposed to relax the local independence as-
sumption by allowing for dependence within clusters using latent variables. Examples of this
approach include mixtures of Rasch models (Rost 1990), multi-level mixture item response models
(Vermunt 2007), and mixtures of latent trait analyzers (Cagnone & Viroli 2012, Gollini & Murphy
2014).
The poLCA (Linzer & Lewis 2011) software is widely used for LCA, the BayesLCA software
(White & Murphy 2014) implements Bayesian LCA, and the lvm4net software (Gollini 2019)
includes the mixture of latent trait analyzers model.

4.3. Mixed Data

Mixed data arise when observations consist of different data types, including continuous, nomi-
nal categorical, and ordinal categorical data, for example. These data arise widely in a range of
applications, in particular in survey data.
Everitt (1984, chapter 4) proposed a mixture model for mixed data that assumes local inde-
pendence between the variables, as used in LCA. The location model (Hunt & Jorgensen 1999,
2003) was developed for model-based clustering of mixed data, but where the local independence
assumption is relaxed. The model can accommodate arbitrary dependence structures for the cat-
egorical variables, and the continuous variables can be dependent on the categorical variables.
The model is structured to accommodate data with a small number of categorical variables and a
potentially large number of continuous variables.
McParland et al. (2014, 2017) and McParland & Gormley (2016) developed the clustMD
framework for clustering mixed data. The clustMD model assumes that the observed variables
(nominal, ordinal, and continuous) are manifestations of underlying continuous latent variables.
This approach facilitates clustering a large number of categorical variables compared with the
location model. The clustMD (McParland & Gormley 2017) software package implements the
methodology. A review of many of the approaches for model-based clustering for mixed data is
given by McParland & Murphy (2019).

4.4. Count Data

A number of model-based clustering approaches have been developed for clustering multivariate
count data that arise in application domains including ecology, epidemiology, bioinformatics, and
text analytics. One of the simplest model-based clustering approaches for multivariate count data
uses mixtures with a locally independent Poisson assumption, similar to the local independence
assumption used in LCA (Section 4.2). A number of approaches have been proposed to cluster
multivariate count data while allowing for dependence and overdispersion in the counts. Karlis &
Meligkotsidou (2007) and Karlis (2019) developed model-based clustering methods based on the
multivariate Poisson distribution. Subedi & Browne (2020) developed a model-based clustering
approach for count data using multivariate Poisson-log normal component distributions. Roick
et al. (2021) developed a model-based clustering approach based on integer-valued autoregressive
(INAR) models. Ng & Murphy (2021) developed a model-based clustering approach based on the
Gaussian Cox process to cluster count process data. Karlis (2019) gave an overview of model-based
clustering of discrete data, including count-valued data.

586 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

4.5. Sequence Data

Sequence data arise when observations consist of strings of categorical values selected from a
finite set. Sequence data arise in many social science applications, in particular, where life-course
trajectories are collected, as well as in many other application domains, including computer science
and biology.
Erosheva et al. (2014) provided a review of a number of finite mixture models for life-
course sequence data, including the group-based trajectory and growth mixture models. More
recently, Murphy et al. (2021) developed a distance-based mixture model for clustering life-course
trajectory data; this approach was implemented in the MEDseq software (Murphy et al. 2022).
A number of approaches for model-based clustering of sequences have been developed which
are based on Markov chain models (Melnykov 2016b, Zhang et al. 2021). In this approach, the
clustering of observations is focused on the transition between states in the sequences rather than
the specific trajectories. The ClickClust software (Melnykov 2016a) provides an implementation
of Markov chain model-based clustering for sequence data.

4.6. Rank Data

Rank data arise when a set of judges are asked to list a set of objects in order of preference. The
resulting data are ordered lists of objects. Rank data arise in a number of contexts, including voting,
education, and marketing. A number of model-based clustering methods have been developed for
ranking data.
Gormley & Murphy (2006, 2008) developed finite mixtures of Plackett–Luce (Plackett 1975)
and Benter (Benter 1994) models for clustering rank data; model inference was implemented in
a maximum likelihood framework. Mollica & Tardella (2014, 2017, 2021) further developed the
finite mixture of Plackett–Luce modeling framework by considering an extension of the Plackett–
Luce model and Bayesian methods for inference. Furthermore, the PLMIX software (Mollica &
Tardella 2020) facilitates fitting the mixture of Plackett–Luce models (and extensions) to rank
data.
An alternative approach to clustering rank data was proposed by Murphy & Martin (2003)
using a finite mixture of distance-based models, known as Mallows models (Mallows 1957). This
modeling approach was extended by Busse et al. (2007), who developed computationally efficient
algorithms for inference. More recently, Bayesian mixtures of Mallows models have been pro-
posed for clustering rank data by Vitelli et al. (2017) and Liu et al. (2019). The BayesMallows
software (Sørensen et al. 2020) facilitates fitting mixtures of Mallows models, with a number of
rank distances. Related to the Mallows modeling approach, Biernacki & Jacques (2013) proposed
a model-based clustering approach for rank data based on insertion and sort algorithms.

4.7. Network Data

Network data arise when the connections (or relations) between entities or nodes are recorded.
Often the connections are binary in nature, but it is also possible to have weighted connections
(showing the strength of the connection); these data arise widely in biology and the social sciences.
The emphasis when clustering social network data is on clustering the entities into meaningful
groups.
A number of models have been proposed for model-based clustering of network data. For
example, Snijders & Nowicki (1997) and Nowicki & Snijders (2001) proposed the stochastic block-
model for clustering network data. The stochastic blockmodel assumes a latent clustering variable

www.annualreviews.org • Model-Based Clustering 587

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

and that connections are formed independently, conditional on the latent clustering variable. The
stochastic blockmodel has been extended in a number of directions, including the mixed mem-
bership stochastic blockmodel (Airoldi et al. 2008) and the overlapping stochastic blockmodel
(Latouche et al. 2011).
The latent position cluster model (LPCM), developed by Handcock et al. (2007) for clustering
network data, is based on a latent Gaussian mixture model. The LPCM was extended by Krivitsky
et al. (2009) to account for complex structures that arise in network data. The latentnet soft-
ware (Krivitsky & Handcock 2008, 2020) has been developed to cluster network data using the
LPCM. Gormley & Murphy (2010) include node covariates within the LPCM, with implementa-
tion available through the MEclustnet software package (Gormley & Murphy 2019). Wasserman
et al. (2007), Salter-Townshend et al. (2012), and Bouveyron et al. (2019, chapter 10) provide
reviews of a number of model-based clustering approaches for network data.

4.8. Clustering with Covariates

Jacobs et al. (1991) developed the mixture-of-experts modeling framework, which provides a sys-
tematic way to include covariates in the finite mixture model–based clustering approach. The
mixture-of-experts framework has been used by a number of authors to extend model-based clus-
tering methods to the situation where covariate information is also available. Examples include
Murphy & Murphy (2020), who developed the MoEClust family of models that extend the mclust
modeling framework to include covariates, and Gormley & Murphy (2008), who extended the
Plackett–Luce mixture model (Section 4.6) to include covariates. Yuksel et al. (2012), Gormley
& Frühwirth-Schnatter (2019), and Bouveyron et al. (2019, chapter 11) provide reviews of the
application of mixture-of-experts models.

4.9. Variable Selection

Many applications of model-based clustering involve high-dimensional data, which are difficult
to model. To extend the scope of model-based clustering to high-dimensional settings, a number
of approaches have been developed for simultaneously clustering data and performing variable
selection.
Raftery & Dean (2006) proposed a model-based clustering approach that includes variable
selection; the approach is based on the same Gaussian mixture model as mclust (Section 2.1).
This approach was further developed and extended by Maugis et al. (2009a,b) and Celeux et al.
(2011). The clustvarsel software package (Scrucca & Raftery 2018) implements this variable
selection approach.
Dean & Raftery (2010) also developed a model-based clustering approach with variable selec-
tion for categorical data, using the LCA model (Section 4.2). This approach was extended by Fop
et al. (2017) to partition data variables into clustering, redundant, and irrelevant variables; both
approaches are implemented in the LCAvarsel software package (Fop & Murphy 2017). Detailed
reviews of variable selection methods and clustering high-dimensional data are given by Celeux
et al. (2014), Fop & Murphy (2018) and McParland & Murphy (2019).

5. DISCUSSION
In this article we have reviewed the well-established yet rapidly developing area of model-based
clustering. From its beginnings in the 1960s, model-based clustering has contributed much to
the field of statistics, not only through its own developments but also through the impact those
developments have had on other areas of statistics. As such, this article does not provide a review of

588 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

the total existing literature on model-based clustering, nor does it cover all the emerging themes,
of which there are many.
Perhaps the most urgent aspect of model-based clustering that requires development is its
scalability to be able to handle the volume, velocity, and variety of data currently being generated.
While Section 4 highlights literature that has begun to address these aspects (in particular variety),
there is significant scope for advancement. For example, DNA methylation data are collected at a
scale of circa 1 million variables being recorded per observation, of which there are typically few.
While mixtures of latent variable models have potential given their dimension reduction charac-
teristics (e.g., McLachlan et al. 2002), they break down at such scale. Similarly, for data sets with
large n, the reliance on the likelihood function often means the model-based clustering approach is
intractable. While Antonazzo et al. (2021) proposed a binned technique for scalable model-based
clustering on huge data sets, there is much scope for further novel developments here.
Section 3.3 alluded to the trade-off between the use of complex component densities and the
number of clusters. The idea of cluster merging (Baudry et al. 2010), as used to arrive at Figure 2,
suggests one way of alleviating this tension. Recent work has used mixtures of complex component
densities to cluster non-Gaussian continuous data (e.g, Murray et al. 2020, Lee & McLachlan
2022), and this area is ripe for further advancements to be made. Similarly, model-based clustering
approaches for object-oriented data analysis through the use of complex component densities
require attention.
From an applications viewpoint, the utility of a clustering solution may be low if the resulting
clusters are not associated with (e.g., clinical) outcomes of interest. Developing a guided model-
based clustering approach that can infer clustering structure with specific outcomes in mind would
provide an approach that delivers deeper impact in a range of application domains.
From a software viewpoint, Section 4 and the sidebar titled Software highlight the broad ar-
ray of software available to facilitate the widespread use of model-based clustering approaches.
However, predominantly these software packages employ the maximum likelihood approach to
model-based clustering, and there is little on offer from the Bayesian paradigm given the as-
sociated technical and computational challenges. However, as approaches such as sparse finite
mixture models gain further traction, there is scope for the Bayesian approach to become more
accessible to the nonstatistical community through the development of associated open-source
software.
As unsupervised learning methods gain traction and popularity in a broad range of domains
and application areas, the model-based clustering approach remains strongly competitive given its
inherent statistical rigor and flexibility. There is significant scope for the development of bespoke,
apposite model-based clustering approaches for data types and problems that we have not yet
seen.

DISCLOSURE STATEMENT
The authors are not aware of any affiliations, memberships, funding, or financial holdings that
might be perceived as affecting the objectivity of this review.

ACKNOWLEDGMENTS
The authors would like to thank the members of the Working Group in Model-Based Clustering
for discussions that strongly informed this work. This work was supported by Science Foundation
Ireland grants (12/RC/2289_P2, 16/RC/3835), a stay at Collegium de Lyon, and National Insti-
tutes of Health (NIH) grant R01 HD-70936 from the Eunice Kennedy Shriver National Institute
of Child Health and Human Development (NICHD).

www.annualreviews.org • Model-Based Clustering 589

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

LITERATURE CITED
Ahlquist JS, Breunig C. 2012. Model-based clustering and typologies in the social sciences. Political Anal.
20:92–112
Airoldi EM, Blei DM, Fienberg SE, Xing EP. 2008. Mixed-membership stochastic blockmodels. J. Mach. Learn.
Res. 9(65):1981–2014
Aitkin M, Rubin DB. 1985. Estimation and hypothesis testing in finite mixture models. J. R. Stat. Soc. Ser. B
47(1):67–75
Antonazzo F, Biernacki C, Keribin C. 2021. A binned technique for scalable model-based clustering on huge
datasets. In Book of Short Papers of the 5th International Workshop on Models and Learning for Clustering and
Classification (MBC2 2020), Catania, Italy, ed. S Ingrassia, A Punzo, R Rocci, pp. 11–16. Milan: Ledizioni
Banfield JD, Raftery AE. 1989. Model-based Gaussian and non-Gaussian clustering. Tech. Rep. 186, Dep. Stat.,
Univ. Washington, Seattle, WA
Banfield JD, Raftery AE. 1993. Model-based Gaussian and non-Gaussian clustering. Biometrics 49(3):803–21
Baudry JP, Raftery AE, Celeux G, Lo K, Gottardo R. 2010. Combining mixture components for clustering.
J. Comput. Graph. Stat. 19:332–53
Benaglia T, Chauveau D, Hunter DR, Young D. 2009. mixtools: An R package for analyzing finite mixture
models. J. Stat. Softw. 32(6):1–29
Bensmail H, Celeux G, Raftery AE, Robert CP. 1997. Inference in model-based cluster analysis. Stat. Comput.
7(1):1–10
Benter W. 1994. Computer based horse race handicapping and wagering systems: a report. In Efficiency of
Racetrack Betting Markets, ed. WT Ziemba, VS Lo, DB Haush, pp. 183–98. Singapore: World Sci.
Biernacki C, Celeux G, Govaert G. 2000. Assessing a mixture model for clustering with the integrated
completed likelihood. IEEE Trans. Pattern Anal. 22(7):719–25
Biernacki C, Jacques J. 2013. A generative model for rank data based on insertion sort algorithm. Comput. Stat.
Data Anal. 58:162–76
Binder DA. 1978. Bayesian cluster analysis. Biometrika 65(1):31–38
Bouveyron C, Celeux G, Murphy TB, Raftery AE. 2019. Model-Based Clustering and Classification for Data
Science: With Applications in R. Cambridge, UK: Cambridge Univ. Press
Bouveyron C, Jacques J. 2011. Model-based clustering of time series in group-specific functional subspaces.
Adv. Data Anal. Classif. 5(4):281–300
Busse LM, Orbanz P, Buhmann JM. 2007. Cluster analysis of heterogeneous rank data. In Proceedings of the
24th International Conference on Machine Learning, ICML ’07, pp. 113–20. New York: ACM
Cagnone S, Viroli C. 2012. A factor mixture analysis model for multivariate binary data. Stat. Model. 12(3):257–
77
Carpaneto G, Toth P. 1980. Algorithm 548: Solution of the assignment problem [H]. ACM Trans. Math. Softw.
6(1):104–11
Celeux G, Govaert G. 1995. Gaussian parsimonious clustering models. Pattern Recognit. 28(5):781–93
Celeux G, Hurn M, Robert CP. 2000. Computational and inferential difficulties with mixture posterior
distributions. J. Am. Stat. Assoc. 95(451):957–70
Celeux G, Martin-Magniette ML, Maugis C, Raftery AE. 2011. Letter to the editor: “A framework for feature
selection in clustering.” J. Am. Stat. Assoc. 106(493):383
Celeux G, Martin-Magniette ML, Maugis-Rabusseau C, Raftery AE. 2014. Comparing model selection and
regularization approaches to variable selection in model-based clustering. J. Soc. Fr. Stat. 155(2):57–71
Czekanowski J. 1909. Zur differential-diagnose der Neadertalgruppe. Korresp. Bl. Dtsch. Ges. Anthropol. Ethnol.
Urgesch. 40:44–47
Day NE. 1969. Estimating the components of a mixture of two normal distributions. Biometrika 56(3):463–74
Dean N, Raftery AE. 2010. Latent class analysis variable selection. Ann. Inst. Stat. Math. 62(1):11–35
Dempster AP, Laird NM, Rubin DB. 1977. Maximum likelihood from incomplete data via the EM algorithm.
With discussion. J. R. Stat. Soc. Ser. B 39(1):1–38
Diebolt J, Robert CP. 1994. Estimation of finite mixture distributions through Bayesian sampling. J. R. Stat.
Soc. Ser. B 56(2):363–75

590 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Erosheva EA, Matsueda RL, Telesca D. 2014. Breaking bad: two decades of life-course data analysis in
criminology, developmental psychology, and beyond. Annu. Rev. Stat. Appl. 1:301–32
Everitt B. 1984. An Introduction to Latent Variable Models. London: Chapman and Hall
Ferguson TS. 1973. A Bayesian analysis of some nonparametric problems. Ann. Stat. 1:209–30
Fop M, Murphy TB. 2017. LCAvarsel: variable selection for latent class analysis. R Package, version 1.1
Fop M, Murphy TB. 2018. Variable selection methods for model-based clustering. Stat. Surv. 12:18–65
Fop M, Murphy TB, Scrucca L. 2019. Model-based clustering with sparse covariance matrices. Stat. Comput.
29(4):791–819
Fop M, Smart K, Murphy TB. 2017. Variable selection for latent class analysis with application to low back
pain diagnosis. Ann. Appl. Stat. 11:2085–115
Fraley C, Raftery AE. 1998. How many clusters? Which clustering method? Answers via model-based cluster
analysis. Comput. J. 41(8):578–88
Fraley C, Raftery AE. 2002. Model-based clustering, discriminant analysis and density estimation. J. Am. Stat.
Assoc. 97(458):611–31
Frühwirth-Schnatter S. 2006. Finite Mixture and Markov Switching Models. New York: Springer
Frühwirth-Schnatter S. 2011. Dealing with label switching under model uncertainty. In Mixtures: Estimation
and Application, ed. K Mengersen, CP Robert, DM Titterington, pp. 213–39. New York: Wiley
Frühwirth-Schnatter S, Celeux G, Robert CP. 2019. Handbook of Mixture Analysis. Boca Raton, FL: Chapman
and Hall/CRC
Frühwirth-Schnatter S, Malsiner-Walli G. 2019. From here to infinity: sparse finite versus Dirichlet process
mixtures in model-based clustering. Adv. Data Anal. Classif. 13(1):33–64
García-Escudero LA, Gordaliza A, Greselin F, Ingrassia S, Mayo-Iscar A. 2018. Eigenvalues and constraints
in mixture modeling: geometric and computational issues. Adv. Data Anal. Classif. 12(2):203–33
García-Escudero LA, Gordaliza A, Matrán C, Mayo-Iscar A. 2008. A general trimming approach to robust
cluster analysis. Ann. Stat. 36(3):1324–45
Ghahramani Z, Hinton GE. 1996. The EM algorithm for mixtures of factor analyzers. Tech. Rep. CRG-TR-96-1,
Dep. Comput. Sci., Univ. Toronto, Toronto, Can.
Gollini I. 2019. lvm4net: Latent variable models for networks. R Package, version 0.3
Gollini I, Murphy TB. 2014. Mixture of latent trait analyzers for model-based clustering of categorical data.
Stat. Comput. 24(4):569–88
Gormley IC, Frühwirth-Schnatter S. 2019. Mixture of experts models. In Handbook of Mixture Analysis, ed. S
Frühwirth-Schnatter, G Celeux, CP Robert, pp. 271–307. Boca Raton, FL: Chapman and Hall/CRC
Gormley IC, Murphy TB. 2006. Analysis of Irish third-level college applications data. J. R. Stat. Soc. Ser. A
169(2):361–79
Gormley IC, Murphy TB. 2008. Exploring voting blocs within the Irish electorate: a mixture modeling
approach. J. Am. Stat. Assoc. 103(483):1014–27
Gormley IC, Murphy TB. 2010. A mixture of experts latent position cluster model for social network data.
Stat. Methodol. 7(3):385–405
Gormley IC, Murphy TB. 2019. MEclustnet: the mixture of experts latent position cluster model for network
data. R Package, version 1.2.2
Grün B, Leisch F. 2007. Fitting finite mixtures of generalized linear regressions in R. Comput. Stat. Data Anal.
51(11):5247–52
Grün B, Leisch F. 2008. FlexMix version 2: Finite mixtures with concomitant variables and varying and
constant parameters. J. Stat. Softw. 28(4):1–35
Handcock MS, Raftery AE, Tantrum JM. 2007. Model-based clustering for social networks. J. R. Stat. Soc. Ser.
A 170(2):1–22
Hejblum BP, Alkhassim C, Gottardo R, Caron F, Thiébaut R. 2019. Sequential Dirichlet process mixtures
of multivariate skew t-distributions for model-based clustering of flow cytometry data. Ann. Appl. Stat.
13(1):638–60
Hennig C. 2010. Methods for merging Gaussian mixture components. Adv. Data Anal. Classif. 4(1):3–34
Hunt L, Jorgensen M. 1999. Theory & methods: mixture model clustering using the MULTIMIX program.
Aust. N. Z. J. Stat. 41(2):154–71

www.annualreviews.org • Model-Based Clustering 591

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Hunt L, Jorgensen M. 2003. Mixture model clustering for mixed data with missing information. Comput. Stat.
Data Anal. 41(3–4):429–40
Ishwaran H, James LF. 2001. Gibbs sampling methods for stick-breaking priors. J. Am. Stat. Assoc. 96(453):161–
73
Jacobs RA, Jordan MI, Nowlan SJ, Hinton GE. 1991. Adaptive mixtures of local experts. Neural Comput.
3(1):79–87
Jacques J, Preda C. 2014. Model-based clustering for multivariate functional data. Comput. Stat. Data Anal.
71:92–106
Jasra A, Holmes CC, Stephens DA. 2005. Markov chain Monte Carlo methods and the label switching problem
in Bayesian mixture modeling. Stat. Sci. 20(1):50–67
Jordan MI, Jacobs RA. 1994. Hierarchical mixtures of experts and the EM algorithm. Neural Comput. 6(2):181–
214
Kalli M, Griffin JE, Walker SG. 2011. Slice sampling mixture models. Stat. Comput. 21(1):93–105
Karlis D. 2019. Mixture modelling of discrete data. In Handbook of Mixture Analysis, ed. S Frühwirth-Schnatter,
G Celeux, CP Robert, pp. 193–218. Boca Raton, FL: Chapman and Hall/CRC
Karlis D, Meligkotsidou L. 2007. Finite mixtures of multivariate Poisson distributions with application. J. Stat.
Plan. Inference 137(6):1942–60
Keribin C. 2000. Consistent estimation of the order of mixture models. Sankhya A 62(1):49–66
Krivitsky PN, Handcock MS. 2008. Fitting latent cluster models for networks with latentnet. J. Stat. Softw.
24(5):1–23
Krivitsky PN, Handcock MS. 2020. latentnet: Latent position and cluster models for statistical networks.
R Package, version 2.10.5
Krivitsky PN, Handcock MS, Raftery AE, Hoff PD. 2009. Representing degree distributions, clustering, and
homophily in social networks with latent cluster random effects models. Soc. Netw. 31(3):204–13
Latouche P, Birmelé E, Ambroise C. 2011. Overlapping stochastic block models with application to the French
political blogosphere. Ann. Appl. Stat. 5(1):309–36
Lazarsfeld PF. 1950a. The logical and mathematical foundations of latent structure analysis. In Studies in Social
Psychology in World War II. Vol. IV: Measurement and Prediction, ed. SA Stouffer, L Guttman, EA Suchman,
PF Lazarsfeld, pp. 362–412. Princeton, NJ: Princeton Univ. Press
Lazarsfeld PF. 1950b. Some latent structures. In Studies in Social Psychology in World War II. Vol. IV: Measurement
and Prediction, ed. SA Stouffer, L Guttman, EA Suchman, PF Lazarsfeld, pp. 413–73. Princeton, NJ:
Princeton Univ. Press
Lebret R, Iovleff S, Langrognet F, Biernacki C, Celeux G, Govaert G. 2015. Rmixmod: the R package of the
model-based unsupervised, supervised, and semi-supervised classification mixmod library. J. Stat. Softw.
67(6):1–29
Lee SX, McLachlan GJ. 2013. Model-based clustering and classification with non-normal mixture
distributions. Stat. Methods Appl. 22(4):427–54
Lee SX, McLachlan GJ. 2018. EMMIXcskew: an R package for the fitting of a mixture of canonical fundamental
skew t-distributions. J. Stat. Softw. 83(3):1–32
Lee SX, McLachlan GJ. 2022. An overview of skew distributions in model-based clustering. J. Multivar. Anal.
188:104853
Leisch F. 2004. FlexMix: a general framework for finite mixture models and latent class regression in R. J. Stat.
Softw. 11(8):1–18
Linnaeus C. 1753. Species Plantarum. Stockholm: Laurentii Salvii. 1st ed.
Linzer DA, Lewis JB. 2011. poLCA: An R package for polytomous variable latent class analysis. J. Stat. Softw.
42(10):1–29
Liu Q, Crispino M, Scheel I, Vitelli V, Frigessi A. 2019. Model-based learning from preference data. Annu.
Rev. Stat. Appl. 6:329–54
Maier LM, Anderson DE, De Jager PL, Wicker LS, Hafler DA. 2007. Allelic variant in CTLA4 alters T cell
phosphorylation patterns. PNAS 104(47):18607–12
Mallows CL. 1957. Non-null ranking models. Biometrika 44(1/2):114–30
Malsiner-Walli G, Frühwirth-Schnatter S, Grün B. 2016. Model-based clustering based on sparse finite
Gaussian mixtures. Stat. Comput. 26:303–24

592 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Maugis C, Celeux G, Martin-Magniette ML. 2009a. Variable selection for clustering with Gaussian mixture
models. Biometrics 65(3):701–9
Maugis C, Celeux G, Martin-Magniette ML. 2009b. Variable selection in model-based clustering: a general
variable role modeling. Comput. Stat. Data Anal. 53(11):3872–82
McLachlan G, Peel D. 1998. Robust cluster analysis via mixtures of multivariate t-distributions. In Advances
in Pattern Recognition, ed. A Amin, D Dori, P Pudil, H Freeman, pp. 658–66. Berlin: Springer-Verlag
McLachlan G, Peel D. 2000. Finite Mixture Models. New York: Wiley-Interscience
McLachlan GJ, Basford KE. 1988. Mixture Models: Inference and Applications to Clustering. New York: Marcel
Dekker
McLachlan GJ, Bean RW, Peel D. 2002. A mixture model–based approach to the clustering of microarray
expression data. Bioinformatics 18(3):413–22
McLachlan GJ, Krishnan T. 2008. The EM Algorithm and Extensions. New York: Wiley-Interscience. 2nd ed.
McLachlan GJ, Lee SX, Rathnayake SI. 2019. Finite mixture models. Annu. Rev. Stat. Appl. 6:355–78
McLachlan GJ, Peel D, Basford KE, Adams P. 1999. The EMMIX software for the fitting of mixtures of
normal and t-components. J. Stat. Softw. 4(2):1–14
McLachlan GJ, Peel D, Bean RW. 2003. Modelling high-dimensional data by mixtures of factor analyzers.
Comput. Stat. Data Anal. 41(3–4):379–88
McNicholas PD. 2016a. Mixture Model–Based Classification. Boca Raton, FL: Chapman and Hall/CRC
McNicholas PD. 2016b. Model-based clustering. J. Classif. 33:331–73
McNicholas PD, ElSherbiny A, McDaid AF, Murphy TB. 2021. pgmm: Parsimonious Gaussian mixture models.
R Package, version 1.2.5
McNicholas PD, Jampani KR, Subedi S. 2019. longclust: Model-based clustering and classification for
longitudinal data. R Package, version 1.2.3
McNicholas PD, Murphy TB. 2008. Parsimonious Gaussian mixture models. Stat. Comput. 18(3):285–96
McNicholas PD, Murphy TB. 2010. Model-based clustering of longitudinal data. Can. J. Stat. 38(1):153–68
McParland D, Gormley IC. 2016. Model based clustering for mixed data: clustMD. Adv. Data Anal. Classif.
10(2):155–69
McParland D, Gormley IC. 2017. clustMD: Model based clustering for mixed data. R Package, version 1.2.1
McParland D, Gormley IC, McCormick TH, Clark SJ, Kabudula CW, Collinson MA. 2014. Clustering South
African households based on their asset status using latent variable models. Ann. Appl. Stat. 8(2):747–76
McParland D, Murphy TB. 2019. Mixture modelling of high-dimensional data. In Handbook of Mixture
Analysis, ed. S Frühwirth-Schnatter, G Celeux, CP Robert, pp. 39–70. Boca Raton, FL: Chapman and
Hall/CRC
McParland D, Phillips CM, Brennan L, Roche HM, Gormley IC. 2017. Clustering high-dimensional mixed
data to uncover sub-phenotypes: joint analysis of phenotypic and genotypic data. Stat. Med. 36(28):4548–
69
Melnykov V. 2016a. ClickClust: an R package for model-based clustering of categorical sequences. J. Stat.
Softw. 74(9):1–34
Melnykov V. 2016b. Model-based biclustering of clickstream data. Comput. Stat. Data Anal. 93:31–45
Mengersen KL, Robert CP. 1996. Testing for mixtures: a Bayesian entropic approach. In Bayesian Statistics 5:
Proceedings of the Fifth Valencia International Meeting, ed. JM Bernardo, JO Berger, AP Dawid, AFM Smith,
pp. 255–76. Oxford, UK: Oxford Univ. Press
Mollica C, Tardella L. 2014. Epitope profiling via mixture modeling of ranked data. Stat. Med. 33(21):3738–58
Mollica C, Tardella L. 2017. Bayesian Plackett-Luce mixture models for partially ranked data. Psychometrika
82(2):442–58
Mollica C, Tardella L. 2020. PLMIX: an R package for modelling and clustering partially ranked data. J. Stat.
Comput. Simul. 90(5):925–59
Mollica C, Tardella L. 2021. Bayesian analysis of ranking data with the extended Plackett-Luce model. Stat.
Methods Appl. 30(1):175–94
Montanari A, Viroli C. 2010. Heteroscedastic factor mixture analysis. Stat. Model. 10(4):441–60
Morris K, McNicholas PD. 2016. Clustering, classification, discriminant analysis, and dimension reduction
via generalized hyperbolic mixtures. Comput. Stat. Data Anal. 97:133–50

www.annualreviews.org • Model-Based Clustering 593

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Murphy K, Murphy TB. 2020. Gaussian parsimonious clustering models with covariates and a noise
component. Adv. Data Anal. Classif. 14(2):293–325
Murphy K, Murphy TB, Piccarreta R, Gormley IC. 2021. Clustering longitudinal life-course sequences using
mixtures of exponential-distance models. J. R. Stat. Soc. Ser. A 184(4):1414–51
Murphy K, Murphy TB, Piccarreta R, Gormley IC. 2022. MEDseq: Mixtures of exponential-distance models
with covariates. R Package, version 1.3.3
Murphy K, Viroli C, Gormley IC. 2020. Infinite mixtures of infinite factor analysers. Bayesian Anal. 15(3):937–
63
Murphy TB, Martin D. 2003. Mixtures of distance-based models for ranking data. Comput. Stat. Data Anal.
41(3–4):645–55
Murray PM, Browne RP, McNicholas PD. 2020. Mixtures of hidden truncation hyperbolic factor analyzers.
J. Classif. 37(2):366–79
Murtagh F, Raftery AE. 1984. Fitting straight lines to point patterns. Pattern Recognit. 17(5):479–83
Ng TLJ, Murphy TB. 2021. Model-based clustering of count processes. J. Classif. 38(2):188–211
Nowicki K, Snijders TAB. 2001. Estimation and prediction of stochastic blockstructures. J. Am. Stat. Assoc.
96(455):1077–87
O’Hagan A, Murphy TB, Gormley IC. 2012. Computational aspects of fitting mixture models via the
expectation–maximization algorithm. Comput. Stat. Data Anal. 56(12):3843–64
O’Hagan A, Murphy TB, Gormley IC, McNicholas PD, Karlis D. 2016. Clustering with the multivariate
normal inverse Gaussian distribution. Comput. Stat. Data Anal. 93:18–30
Papastamoulis P. 2018. Overfitting Bayesian mixtures of factor analyzers with an unknown number of
components. Comput. Stat. Data Anal. 124:220–34
Pearson K. 1894. Contributions to the mathematical theory of evolution. Philos. Trans. R. Soc. A 185:71–110
Plackett RL. 1975. The analysis of permutations. J. R. Stat. Soc. Ser. C 24(2):193–202
Pyne S, Hu X, Wang K, Rossin E, Lin TI, et al. 2009. Automated high-dimensional flow cytometric data
analysis. PNAS 106(21):8519–24
Quintana FA, Iglesias PL. 2003. Bayesian clustering and product partition models. J. R. Stat. Soc. Ser. B
65(2):557–74
Raftery AE. 1996. Hypothesis testing and model selection. In Markov Chain Monte Carlo in Practice, ed. WR
Gilks, S Richardson, DJ Spiegelhalter, pp. 163–88. London: Chapman and Hall
Raftery AE, Dean N. 2006. Variable selection for model-based clustering. J. Am. Stat. Assoc. 101(473):168–78
Raftery AE, Newton M, Satagopan J, Krivitsky P. 2007. Estimating the integrated likelihood via posterior
simulation using the harmonic mean identity. In Bayesian Statistics 8, ed. JM Bernardo, MJ Bayarri, JO
Berger, AP Dawid, D Heckerman, et al., pp. 1–45. Oxford, UK: Oxford Univ. Press
Richardson S, Green PJ. 1997. On Bayesian analysis of mixtures with an unknown number of components
(with discussion). J. R. Stat. Soc. Ser. B 59(4):731–92
Roick T, Karlis D, McNicholas PD. 2021. Clustering discrete-valued time series. Adv. Data Anal. Classif.
15(1):209–29
Rost J. 1990. Rasch models in latent classes: an integration of two approaches to item analysis. Appl. Psychol.
Meas. 14(3):271–82
Rousseau J, Mengersen K. 2011. Asymptotic behaviour of the posterior distribution in overfitted mixture
models. J. R. Stat. Soc. Ser. B 73(5):689–710
Salter-Townshend M, White A, Gollini I, Murphy TB. 2012. Review of statistical network analysis: models,
algorithms, and software. Stat. Anal. Data Min. 5(4):243–64
Schwarz G. 1978. Estimating the dimension of a model. Ann. Stat. 6(2):461–64
Scrucca L, Fop M, Murphy TB, Raftery AE. 2016. mclust 5: Clustering, classification and density estimation
using Gaussian finite mixture models. R J. 8(1):289–317
Scrucca L, Fraley C, Murphy TB, Raftery AE. 2022. Model-Based Clustering, Classification and Density Estimation
Using mclust in R. Boca Raton, FL: Chapman and Hall/CRC
Scrucca L, Raftery AE. 2018. clustvarsel: A package implementing variable selection for Gaussian model-
based clustering in R. J. Stat. Softw. 84(1):1–28
Sneath PHA. 1957. The application of computers to taxonomy. J. Gen. Microbiol. 17:201–6

594 Gormley • Murphy • Raftery

ST10CH24_Raftery ARjats.cls February 14, 2023 14:5

Snijders TAB, Nowicki K. 1997. Estimation and prediction for stochastic blockmodels for graphs with latent
block structure. J. Classif. 14(1):75–100
Sokal RR, Michener CD. 1958. A statistical method for evaluating systematic relationships. Univ. Kans. Sci.
Bull. 38:1409–38
Sørensen Ø, Crispino M, Liu Q, Vitelli V. 2020. BayesMallows: an R package for the Bayesian Mallows model.
R J. 12(1):324–42
Spiegelhalter DJ, Best NG, Carlin BP, Van Der Linde A. 2002. Bayesian measures of model complexity and
fit. J. R. Stat. Soc. Ser. B 64(4):583–639
Steele RJ, Raftery AE. 2010. Performance of Bayesian model selection criteria for Gaussian mixture models.
In Frontiers of Statistical Decision Making and Bayesian Analysis, ed. M-H Chen, P Müller, D Sun, K Ye, DK
Dey, pp. 113–30. New York: Springer
Stephens M. 2000a. Bayesian analysis of mixture models with an unknown number of components—an
alternative to reversible jump methods. Ann. Stat. 28(1):40–74
Stephens M. 2000b. Dealing with label switching in mixture models. J. R. Stat. Soc. Ser. B 62(4):795–809
Subedi S, Browne RP. 2020. A family of parsimonious mixtures of multivariate Poisson-lognormal distributions
for clustering multivariate count data. Stat 9(1):e310
Tang Y, Browne RP, McNicholas PD. 2018. Flexible clustering of high-dimensional data via mixtures of joint
generalized hyperbolic distributions. Stat 7(1):e177
Tanner MA, Wong WH. 1987. The calculation of posterior distributions by data augmentation. J. Am. Stat.
Assoc. 82(398):528–40
Teicher H. 1963. Identifiability of finite mixtures. Ann. Math. Stat. 34(4):1265–69
Van Havre Z, White N, Rousseau J, Mengersen K. 2015. Overfitting Bayesian mixture models with an
unknown number of components. PLOS ONE 10(7):e0131739
Vermunt J. 2007. Multilevel mixture item response theory models: an application in education testing. In
Proceedings of the 56th Session of the International Statistical Institute, Lisbon, Portugal. Voorburg, Neth.: Int.
Stat. Inst.
Viroli C, Anderlucci L. 2021. Deep mixtures of unigrams for uncovering topics in textual data. Stat. Comput.
31(3):1–10
Viroli C, McLachlan GJ. 2019. Deep Gaussian mixture models. Stat. Comput. 29(1):43–51
Vitelli V, Sørensen Ø, Crispino M, Frigessi A, Arjas E. 2017. Probabilistic preference learning with the
Mallows rank model. J. Mach. Learn. Res. 18:1–49
Wasserman S, Robins G, Steinley D. 2007. Statistical models for networks: a brief review of some recent
research. In Statistical Network Analysis: Models, Issues, and New Directions, ed. EM Airoldi, DM Blei, SE
Fienberg, A Goldenberg, EP Xing, AX Zheng, pp. 45–56. New York: Springer
White A, Murphy TB. 2014. BayesLCA: an R package for Bayesian latent class analysis. J. Stat. Softw. 61(13):1–
28
Wolfe JH. 1965. A computer program for the maximum-likelihood analysis of types. USNPRA Tech. Bull. 65-15,
US Naval Pers. Res. Act., San Diego, CA
Wolfe JH. 1967. NORMIX: computational methods for estimating the parameters of multivariate normal mixture
distributions of types. USNPRA Tech. Bull. 68-2, US Naval Pers. Res. Act., San Diego, CA
Wolfe JH. 1970. Pattern clustering by multivariate mixture analysis. Multivar. Behav. Res. 5(3):329–50
Yuksel SE, Wilson JN, Gader PD. 2012. Twenty years of mixture of experts. IEEE Trans. Neural Netw. Learn.
Syst. 23(8):1177–93
Zhang Y, Melnykov V, Zhu X. 2021. Model-based clustering of time-dependent categorical sequences with
application to the analysis of major life event patterns. Stat. Anal. Data Min. 14(3):230–40

www.annualreviews.org • Model-Based Clustering 595