Distinguishing Between Latent Classes and Continuous Factors: Resolution by Maximum Likelihood?

Abstract

Latent variable models exist with continuous, categorical, or both types of latent variables. The role of latent variables is to account for systematic patterns in the observed responses. This article has two goals: (a) to establish whether, based on observed responses, it can be decided that an underlying latent variable is continu- ous or categorical, and (b) to quantify the effect of sample size and class propor- tions on making this distinction. Latent variable models with categorical, continu- ous, or both types of latent variables are fitted to simulated data generated under different types of latent variable models. If an analysis is restricted to fitting con- tinuous latent variable models assuming a homogeneous population and data stem from a heterogeneous population, overextraction of factors may occur. Similarly, if an analysis is restricted to fitting latent class models, overextraction of classes may occur if covariation between observed variables is due to continuous factors. For the data-generating models used in this study, comparing the fit of different explor- atory factor mixture models usually allows one to distinguish correctly between categorical and/or continuous latent variables. Correct model choice depends on class separation and within-class sample size. Starting with the introduction of factor analysis by Spearman (1904), different types of latent variable models have been developed in various areas of the social sciences. Apart from proposed estimation methods, the most obvious differences between these early latent variable models concern the assumed distribution of the

Full text

Distinguishing Between Latent Classes
and Continuous Factors: Resolution
by Maximum Likelihood?
Gitta Lubke
University of Notre Dame
Michael C. Neale
Virginia Commonwealth University
Latent variable models exist with continuous, categorical, or both types of latent
variables. The role of latent variables is to account for systematic patterns in the
observed responses. This article has two goals: (a) to establish whether, based on
observed responses, it can be decided that an underlying latent variable is continu-
ous or categorical, and (b) to quantify the effect of sample size and class propor-
tions on making this distinction. Latent variable models with categorical, continu-
ous, or both types of latent variables are fitted to simulated data generated under
different types of latent variable models. If an analysis is restricted to fitting con
-
tinuous latent variable models assuming a homogeneous population and data stem
from a heterogeneous population, overextraction of factors may occur. Similarly, if
an analysis is restricted to fitting latent class models, overextraction of classes may
occur if covariation between observed variables is due to continuous factors. For
the data-generating models used in this study, comparing the fit of different explor
-
atory factor mixture models usually allows one to distinguish correctly between
categorical and/or continuous latent variables. Correct model choice depends on
class separation and within-class sample size.
Starting with the introduction of factor analysis by Spearman (1904), different
types of latent variable models have been developed in various areas of the social
sciences. Apart from proposed estimation methods, the most obvious differences
between these early latent variable models concern the assumed distribution of the
MULTIVARIATE BEHAVIORAL RESEARCH, 41(4), 499–532
Copyright © 2006, Lawrence Erlbaum Associates, Inc.
Correspondence concerning this article should be addressed to Gitta H. Lubke, Department of Psy
-
chology, 118 Haggar Hall, University of Notre Dame, Notre Dame IN 46556. E-mail: glubke@nd.edu

observed and latent variables. Bartholomew and Knott (1999) and Heinen (1993)
classified models ina2×2table according to whether the observed and latent vari
-
ables are categorical or continuous. Models with categorical latent variables en
-
compass the classic latent class model, which has categorical observed variables,
and the latent profile model (LPM), which has continuous observed variables
(Heinen, 1993; Lazarsfeld & Henry, 1968). Models with continuous latent vari
-
ables include latent trait models, factor models, and structural equation models.
Attempts have been made to show similarities between these models, and to pro
-
pose general frameworks that encompass the different models as submodels
(Bartholomew & Knott, 1999; Heinen, 1996; Langenheine & Rost, 1988;
Lazarsfeld & Henry, 1968).
A common characteristic of categorical and continuous latent variable models
is that both are designed to explain the covariances between observed variables.
More precisely, latent variable models are constructed such that the observed vari
-
ables are independent conditional on the latent variables. In the context of latent
trait and latent class analysis, this characteristic has been called “local independ-
ence” (Lazarsfeld & Henry, 1968; Mellenbergh, 1994). In exploratory and confir-
matory factor models, local independence corresponds to the assumption of
uncorrelated residuals. Consequently, in both continuous and categorical latent
variable models, nonzero covariances between observed variables are due to dif-
ferences between participants with respect to the latent variables. The general
question addressed in this article is whether, based on observed data, it is possible
to distinguish between categorical and continuous latent variables.
The distinction between categorical and continuous observed variables refers to
the response format of the data, whereas the distinction between categorical and
continuous latent variables can be of considerable importance on a theoretical
level. For example, whether individuals with clinical diagnoses of psychiatric dis
-
orders differ qualitatively or quantitatively from those without is a topic of current
debate (Hicks, Krueger, Iacono, McGue, & Patrick, 2004; Krueger et al., 2004;
Pickles & Angold, 2003). Therefore, a model comparison approach that can reli
-
ably discriminate between latent class or LPMs, on one hand, and factor models,
on the other hand, would have great practical value.
Although the difference between categorical and continuous latent variables
can be significant on a conceptual level, the distinction is less clear on a statistical
level. As shown in Bartholomew (1987), the model implied covariances of a
K-class latent class model for continuous observed variables are structurally equiv
-
alent to the model implied covariances of a K – 1 factor model for continuous ob
-
served variables (see also Molenaar & van Eye, 1994, and more recently, Bauer &
Curran, 2004; Meredith & Horn, 2001). Consequently, in an analysis of
covariances (or correlations), a K-class model and a K – 1 factor model fit equally
well, regardless of whether the true data-generating model is a K-class model or a
K – 1 factor model. This fact has serious consequences when making assumptions
regarding the distribution of the data. First, assuming population heterogeneity and
500
LUBKE AND NEALE

restricting an analysis to latent class models may result in an overextraction of
classes if the sample is homogeneous and covariances of observed variables are
due to underlying continuous factors (see Bauer & Curran, 2004, for some interest
-
ing examples). Second, assuming homogeneity of the sample, and conducting an
exploratory factor analysis might result in overextraction of factors if the sample is
heterogenous. In fact, if a population consists of, say, six classes, an exploratory
factor analysis of the pooled covariance (or correlation) matrix derived from a rea
-
sonably large sample will likely result in a nicely fitting five-factor solution (see
Molenaar & van Eye, 1994, for an illustration). Given that exploratory factor anal
-
ysis is more popular than latent class analysis, overextraction of factors may be a
more common mistake than overextraction of classes.
Exploratory and confirmatory factor analyses can be carried out using
covariance matrices of the observed variables because these are sufficient statistics
given the usual assumptions, namely, homogeneity of the sample, multivariate nor
-
mality of the underlying factors, normality of residuals, independence of factors
and residuals, zero autocorrelations of residuals, and linear relations between ob-
served variables and factors. Under these assumptions, the resulting distribution of
the observed variables is multivariate normal, and has therefore zero skewness and
kurtosis. Latent class models, on the other hand, are fitted to raw data, because
class membership is not observed, and therefore within-class covariance matrices
and mean vectors are not available. The distribution of observed variables in a
heterogenous sample consisting of several latent classes is a mixture distribution.
Skewness, kurtosis, and other higher order moments deviate from zero as differ-
ences between classes (e.g., differences in means, variances, and covariances) in-
crease. Raw data contain information concerning all higher order moments of the
joint distribution of the observed data. Therefore, a comparison of latent class
models and exploratory factor models based on maximum likelihood analysis of
the raw data should in principle reveal the correct model. The likelihood of the data
should be larger under the correctly specified distribution than under an incorrectly
specified distribution. However, in an actual analysis of empirical data, it is not
clear whether the distinction can be made with acceptable accuracy.
Recently, general models have been introduced that include both continuous
and categorical latent variables and that can handle both categorical and continu
-
ous observed variables (Arminger, Stein, & Wittenberg, 1999; Dolan & van der
Maas, 1998; Heinen, 1996; Jedidi, Jagpal, & DeSarbo, 1997; Muthén & Shedden,
1999; Vermunt & Magidson, 2003; Yung, 1997). These general models have been
introduced under various names and are referred to as factor mixture models in the
remainder of this article. Factor mixture models allow the specification of two or
more classes, and within-class structures that can range from local independence
to complex structural relations between latent continuous variables. These models
have the potential to estimate structural relations in the presence of population het
-
erogeneity, and to compare different subpopulations with respect to the parameters
of the within-class distribution. Using factor mixture models in an analysis of em
-
LATENT CLASSES AND FACTORS 501

pirical data represents an additional challenge regarding the distinction between
latent categorical and latent continuous variables. Specifically, the task is to deter
-
mine simultaneously the number of latent classes and the dimensionality of the la
-
tent factor space within each class.
Factor mixture models are based on several assumptions including multivariate
normality within class. The joint distribution of observed variables is a mixture
distributions where the mixture components represent the latent classes. It should
be noted that mixture distributions have two main areas of applications: One is to
use mixture components to model clusters in a population, whereas the other is to
use mixture components to approximate non-normal distributions (McLachlan &
Peel, 2000). When using factor mixture models, one should be aware of the fact
that deviations from the assumed multivariate normality within class can lead to
overextraction of classes. This fact has been demonstrated by Bauer and Curran
(2003a) in the context of factor mixture models for longitudinal data (see also
McLachlan & Peel, 2000, for some classic examples).
In this article, we focus on the recovery of the correct within-class factor struc-
ture in case of possible population heterogeneity. Mixture components are thought
to represent clusters of subjects. Our approach is to compare models in an analysis
that is exploratory with respect to both the within-class structure and the number of
latent classes. This article addresses the question whether model comparisons lead
to a correct model choice regarding the nature of the latent variables (continuous,
categorical, or both), and their dimensionality (number of classes and number of
factors) when model assumptions are not violated. Artificial data are generated un-
der different models including (a) latent class models with two and three classes,
(b) single- and two-factor models where the population is homogenous (as in con-
ventional factor models), and (c) single- and two-factor models where the popula
-
tion consists of two latent classes. Correct and incorrect models are fitted to the dif
-
ferent types of data and compared with respect to model fit. The accuracy of model
selection is likely to depend on the separation between classes when the population
is heterogeneous (Lubke & Muthén, 2007; Yung, 1997). Therefore, data are drawn
from heterogeneous populations with different degrees of separation. Model pa
-
rameters such as intercepts and factor loadings are class invariant or class specific.
Class proportions are equal or unequal. In addition, sample sizes are varied for a
subset of the generated data to investigate minimum sample size requirements for
accurate model selection.
MODELS
General models that encompass continuous and categorical latent variables can be
conceptualized as generalizations of classic latent class models. In classic latent
class models, observed variables within class are assumed to be independent. The
502
LUBKE AND NEALE

generalization consists of the possibility to impose more complicated structures on
observed variables within class, such as factor models. Various types of these gen
-
eralized latent class models have been described elsewhere (Arminger et al., 1999;
Dolan & van der Maas, 1998; Heinen, 1996; Jedidi et al., 1997; Muthén &
Shedden, 1999; Vermunt & Magidson, 2005; Yung, 1997).
1
Here, we follow an ap
-
proach comparable to Bartholomew and Knott (1999) and start with the joint dis
-
tribution of observed variables Y in the total population. The joint distribution is a
mixture distribution where each mixture component corresponds to the distribu
-
tion of observed variables within a particular latent class. The joint probability dis
-
tribution of the data is a weighted sum of the probability distributions of the com
-
ponents. The weights are the proportions of the latent classes in the total
population, and sum to unity. The mixture with k = 1,…,K components can be
denoted as
where p
k
is the proportion of class k and f(?) is a probability distribution. The
within-class distribution f(y
k
) can be parameterized in different ways to derive spe-
cific submodels.
Latent Profile Model
In the LPM, observed variables are assumed to be continuous, and independent
conditional on class membership. Specifically, where the
within-class mean vectors m
k
equal the observed within-class means, and where
the within-class covariance matrices S
k
are diagonal. That is, within each class, the
observed variables are uncorrelated.
Exploratory Factor Mixture Models
The exploratory factor model for a single homogeneous population is derived by
letting K = 1. Hence, there is a single within-class distribution f(y
k
) and the sub
-
script k can be omitted. The observed variables Y are normally distributed with
. In the exploratory factor model the mean vector m is unstructured
such that each observed variable has a single unique free parameter for its mean,
and the model implied covariance matrix is structured as
LATENT CLASSES AND FACTORS 503
1
In a less technical article, Lubke and Muthén (2005) illlustrated how the latent class model can be
extended stepwise to derive a factor mixture model.
=1
() = ( ), (1)
K
kk
k
ffp
å
yy
(,)
kkk
N SY : m
(,)N Sy : m

The factor loading matrix L has dimension J × L where J is the number of observed
variables and L is the number of factors. We assume uncorrelated normally distrib
-
utedfactorswith unit variancesuchthatthe factorcovariance matrix C is anidentity
matrix.Somerestrictionson L arenecessaryfor reasons of identification(Lawley&
Maxwell, 1971). The residuals are assumed to be uncorrelated with the factors, and
normally distributed with zero means and diagonal covariance matrix Q.
By letting k = 1,…, K, exploratory factor mixture models can be specified for K
classes. As before, the within-class mean vectors m
k
equal the observed within-
class means, and the within-class covariance matrices S
k
are structured as in Equa
-
tion 2. Factor loadings and residual variances may differ across classes.
Confirmatory factor models for a homogeneous population with K = 1 classes,
or for a heterogeneous population with k = 1,…, K classes can be derived by im
-
posing restrictions on the within-class factor loading matrices. Confirmatory fac-
tor models are not further considered in this article.
SIMULATION STUDY DESIGN
The simulation study presented in this article is designed to investigate the merits
of exploratory factor mixture analysis. Exploratory factor mixture analysis in-
volves fitting a series of different mixture models, and has the potential to recover
the within-class factor structure in the presence of possible population heterogene-
ity. In this study, mixture models are fitted with an increasing number of classes.
Within-class, exploratory factor models are specified with an increasing number of
factors. Note that the within-class model of the LPM can be conceptualized as a
zero-factor model, or as an L-factor model where the factors have zero variance.
LPMs impose the most restrictive structure within class in this study (i.e., local in
-
dependence within class). Increasing the number of factors within class relaxes
this restriction. In this study, the dimension of the within-class factor space, and the
number of classes of the true models used for data generation is kept low.
Exploratory factor models are lenient because all observed variables have load
-
ings on all factors, but they are not entirely unstructured. Because restrictions on
the within-class distribution can result in overextraction of latent classes (Bauer &
Curran, 2004), it might be preferable to start an analysis by fitting mixtures with
unstructured within-class mean vectors and covariance matrices. However, such an
approach would be impractical for questionnaire data with a large number of ob
-
served variables. Estimation may be problematic due to the fact that the number of
parameters to be estimated increases rapidly with increasing numbers of classes.
Furthermore, the likelihood surface of a (normal) mixture with class-specific
504
LUBKE AND NEALE
=. (2)+
t
SLCLQ

covariance matrices often has many singularities (McLachlan & Peel, 2000). In
-
deed, in an initial pilot simulation we found that fitting unrestricted within-class
models led to high nonconvergence rates (e.g., more than 70%) for data with more
than eight observed variables.
The simulation study has two parts, which are organized in a similar way. The
general approach consists of fitting a set of exploratory factor mixture models to
artificial data to investigate whether comparing the fit of different models results in
the correct model choice. In Part 1, detection of the correct model is investigated
for different types of data-generating models, for different class separations in the
true data, and for class-specific versus class-invariant intercepts and loadings. In
Part 2, detection of the correct model is investigated for varying sample sizes and
for unequal class proportions.
Throughout, data are generated without violating assumptions regarding the
within-class factor models (viz., within-class homogeneity, multivariate normality
of the underlying factors, normality of residuals, independence of factors and re
-
siduals, zero autocorrelations of residuals, and linear relations between observed
variables and factors). It should be noted that, for example, deviations from nor-
mality within-class can lead to overextraction of classes. This fact has been dem-
onstrated in a simulation study concerning confirmatory factor mixture models
(e.g., growth mixture models) by Bauer and Curran (2003a) and has been dis-
cussed by Bauer and Curran (2003b), Cudeck and Henly (2003), Muthén (2003),
and Rindskopf (2003). Our article represents a first step to evaluate the potential of
exploratory factor mixture models and focuses on the possibility to discriminate
between relatively simple models that differ with respect to the distribution of la-
tent variables and some other aspects such as class-specific or invariant model pa-
rameters, class proportions, and sample size. Although beyond the scope of this ar
-
ticle, extending our study presented here to include different possible violations of
underlying model assumptions is of clear interest for future research.
Data Generation
For both parts of the simulation, 100 data sets are generated, which differ only with
respect to factor scores and/or residual scores drawn from (multivariate) normal
distributions and with respect to a score drawn from the uniform distribution,
which is used together with the prior class probabilities to assign participants to a
given class. The number of observed variables is 10 throughout. The observed
variables are continuous, and multivariate normally distributed conditional on
class membership. A detailed overview of all within-class parameter values used
for the data generation can be found in the Appendix.
For the first part of the study, data are generated under nine different models.
The first two models are LPMs with two and three classes, abbreviated LPMc2 and
LPMc3, respectively. Models 3, 4, and 5 are factor models for a single population,
LATENT CLASSES AND FACTORS 505

namely, a single-factor/single-class model, denoted as F1c1; a two-factor/sin
-
gle-class model with simple structure (F2c1SS); and a two-factor/single-class
model with cross-loadings (F2c1CL). Models 6, 7, and 8 are single-fac
-
tor/two-class models. Specifically, F1c2MI is measurement invariant across
classes (i.e, intercepts, factor loadings, and residual variances of the within-class
factor model are invariant across classes), F1c2nMI1 has class-specific intercepts
(i.e., vk in Equation 3), and F1c2nMI2 has noninvariant intercepts and factor load
-
ings. The latter two models are not measurement invariant (Lubke, Dolan,
Kelderman, & Mellenbergh, 2003; Meredith, 1993; Widaman & Reise, 1997).
Finally, Model 9 is a two-factor/two-class model. The factor structure is equal to
the two-factor/single-class model.
In this part of the study, the number of participants within class is approxi
-
mately 200 for all models. Consequently, models with more than one class have a
higher total number of participants. This choice is made because the emphasis of
the simulation is on detecting the correct within-class parameterization. The prior
class probabilities in the first part of the study are .5 for two-class models and .333
for the three-class model. The distance between the classes is varied such that the
impact of class separation on model selection can be evaluated. The separation be-
tween classes for the two-class models is a separation of either 1.5 or 3 as mea-
sured by the multivariate Mahalanobis distance.
2
The only three-class model in the
set of data-generating models is the three-class LPM. The Mahalanobis distance
between consecutive classes equals 1.5 for the small class separation and 3 for the
large separation. The difference between Classes 1 and 3 equals 3.0 and 6.0 for
small and large separation, respectively.
The second part of the study focuses on different sample sizes and unequal class
proportions. Data are generated under the single-factor/two-class model with inter
-
ceptdifferences(i.e.,F1c2nMI).Whilekeepingthepriorclass probabilities at .5,the
within-class sample sizes are 25, 50, 75, and 1000. While keeping the within-class
sample size at 200, prior class probabilities are set to .9 and .1, respectively.
Model Fitting
A set of models were fitted to each type of data. For each fitted model, 50 sets of
random starting values are provided, and 10 initial iterations are computed for
these 50 sets. The 10 sets of starting values that result in the highest log likelihood
are then iterated until convergence, and the best solution is chosen as the final re
-
sult. Each set of fitted models comprises models with an increasing number of
within-class factors and increasing number of classes, including LPMs with an in
-
creasing number of classes. The number of factors within class and the number of
classes is increased until the higher dimensional model is rejected in favor of the
lower dimensional model. Factor variances of fitted exploratory factor models are
506
LUBKE AND NEALE
2
The Mahalanobis distance between two classes equals M = (m
1
– m
2
)
t
S
–1
(m
1
– m
2
).

fixed to unity, and the matrix of factor loadings is constrained such that factor inde
-
terminacy within class is avoided. Note that all fitted models are exploratory in the
sense that model parameters such as loadings, intercepts, and residual variances (if
applicable) are class specific. The criteria used for model comparisons are the
Akaike information criterion (AIC), the Bayesian information criterion (BIC), the
sample size adjusted BIC (saBIC), the Consistent AIC (CAIC), and the adjusted
likelihood ratio test (aLRT) statistic (Akaike, 1974, 1987; Bozdogan, 1987; Lo,
Mendell, & Rubin, 2001; Schwarz, 1978). The formulae for the information crite
-
ria can be found in the Appendix. Information criteria differ in the way they correct
for sample size and the number of free parameters of the fitted model. The number
of free parameters of the models fitted in this study are shown in Table 1.
For each of the data-generating models, we provide the percentage of Monte
Carlo (MC) replications for which the correct model is the first or second choice
based on the information criteria AIC, BIC, CAIC, and saBIC, and the aLRT. The
percentages of correct model selection, however, do not reveal which models are
competing models, or how “close” competing models are with respect to their fit
indices. For this reason, for each of the fitted models, we also compute fit indices
averaged over the 100 MC replications. Regarding the average fit indices, the BIC
and CAIC never lead to diverging results. The CAIC is therefore omitted from the
tables reporting averages, but can be easily calculated by adding the number of free
parameters (see Table 1) to the average BIC.
Inaddition,parameterestimatesareevaluatedthatare informative with respectto
model choice. For instance, factor correlations approaching unity may be regarded
as an indication of a single underlying dimension. Similarly, very small factor load-
ingsmayindicateabsenceofacommonunderlyingfactor.Modelsarefittedusingan
extended version of the RUNALL utility designed for MC simulations with Mplus.
The original Runall utility is available at www.statmodel.com/ runutil.html.
RESULTS
Part 1: Identifying the Correct Model
Convergence rates were above 95% if not otherwise mentioned. Results of the first
part of the study are presented separately for each of the nine data-generating mod
-
LATENT CLASSES AND FACTORS 507
TABLE 1
Number of Free Parameters for All Fitted Models
F1C1 F2C1 F3C1 F1C2 F2C2 LPAc2 LPAc3 LPAc4
30 39 47 51 69 41 62 83
Note. For the two-class factor models, intercept differences between classes are estimated. Resid
-
ual variances, factor loadings, and factor correlations (if part of the model) are estimated for each class.

els. The results of the nine data-generating models are summarized each in two ta
-
bles. The first of the two tables shows the proportion of times the correct model is
the first or second choice with respect to the likelihood value and different infor
-
mation criteria. In this table we also show the proportion of MC replications for
which the aLRT would lead to a correct test result. Note that more restrictive mod
-
els may emerge as first choice with respect to the likelihood value even though less
restrictive, nested models were part of the set of fitted model. This is due to
nonconvergence of the less restrictive models (see, e.g., Tables 2 and 6). The sec
-
ond, larger table shows the average fit measures corresponding to each of the fitted
models. Informative average parameter estimates are discussed in the accompany
-
ing text. Results of the second part of the study are presented in a similar fashion.
Latent Profile Model Two-Class Data (Tables 2 and 3)
Seven different models were fitted to the LPM two-class data. These are LPMs
with one, two, or three classes; exploratory factor models for a single class with
one, two or three factors; and an exploratory factor model for two classes with a
single factor. The results are shown in Tables 2 and 3.
Small class separation.
Table 3 shows that if the analysis had been re-
stricted to a latent profile analysis (LPA), the correct model would have been cho-
sen based on most indices. Only the AIC favors the three-class LPM, although the
necessity of the third class is rejected by the aLRT. If the analysis had been re-
stricted to an exploratory factor analysis for a homogeneous population, the results
are less clear. When comparing the corresponding models (i.e., exploratory factor
models for a single class), it can be seen that the AIC favors the three-factor/sin
-
gle-class model (F3C1). BIC and adjusted BIC (aBIC) favor the single-factor/sin
-
gle-class model (F1C1). Given the fact that the true model has zero factors, at least
one artificial factor would be extracted.
The fact that the aBIC and the BIC favor the single-factor/single-class model
(F1C1) when comparing the full set of fitted models highlights the difficulty in dis
-
criminating between LPMs and factor models. Table 2 shows that in an overall
comparison of all fitted models, the proportion of correct model choice is very low.
Even as the second choice, the proportions for the different fit indices do not rise to
acceptable levels. Recall that the first- and second-order moments of the F1C1 are
structurally equivalent to the true two-class LPM (Bartholomew, 1987). Raw data
contain information concerning higher order moments. However, in the case of
small class separation, the higher order moments of the data may not deviate suffi
-
ciently from the zero values that are expected if data were generated under the sin
-
gle-factor/single-class model.
Compelling evidence against the factor models, however, is provided by the pa
-
rameter estimates. The factor loadings of all factor models were small, and the fac
-
508
LUBKE AND NEALE

509
TABLE 2
Data Generated Under the Two-Class Latent Profile Model: Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
Using Information Criteria, and Proportion the aLRT Indicates
the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.06 0.11 0.00 0.14 0.00 0.97
Second choice 0.52 0.34 0.60 0.50 0.33 0.00
Larger class separation
First choice 0.06 0.12 0.97 0.77 0.97 1.00
Second choice 0.56 0.73 0.00 0.2 0.00 0.00
Note. A proportion of .11 indicates that when fitting the correct model the corresponding infor
-
mation criterion had a minimum value 11% of the time when comparing all fitted models. The last col
-
umn shows the proportion of Monte Carlo replications for which the aLRT provided the correct test re
-
sult. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike information
criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC = Consistent
AIC.
TABLE 3
Averaged Fit Statistics for 200 Data Sets Generated by a Two-Class Latent
Profile Model (LPM), Fitting Latent Profile Models LPMc
j
With
J
Classes
and Latent Factor Models F
i
c
j
With
I
Factors and
J
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPMc1 –4559.22 9158.45 9238.27 9174.81 NA
LPMc2 –4484.82 9051.63 9215.28 9085.19 0.01
LPMc3 –4458.66 9041.32 9288.79 9092.06 0.47
F1c1 –4496.85 9053.70 9173.45 9078.25 NA
F2c1 –4493.23 9064.46 9220.13 9096.38 NA
F3c1 –4476.24 9046.48 9234.07 9084.94 NA
F1c2 –4452.22 9026.45 9269.93 9076.37 0.60
Larger class separation
LPMc1 –5100.32 10240.63 10320.46 10257.00 NA
LPMc2 –4599.16 9280.33 9443.98 9313.88 0.00
LPMc3 –4573.00 9270.00 9517.47 9320.74 0.43
F1c1 –4692.42 9444.84 9564.59 9469.40 NA
F2c1 –4682.15 9442.30 9597.96 9474.21 NA
F3c1 –4685.49 9464.99 9652.59 9503.45 NA
F1c2 –4577.46 9276.91 9520.39 9326.83 0.06
Note. Lower information criteria correspond to better fitting models. A significant aLRT indicates
that the model with one class less fits significantly worse; presented are the average p values. Results for
the true model are in bold, as are the values of fit indices that favor misspecified models. logL-val = log
likelihood value; AIC = Akaike information criterion; BIC = Bayesian information criterion; saBIC =
sample size adjusted BIC; aLRT = adjusted likelihood ratio test.

tor(s) explained only a small percentage of the variance of the observed variables.
For instance, for the fitted single-factor/single-class model, the observed variable
R-square averaged over the 100 MC replications ranged between .02 and .23 for
the 10 items, with an average standard deviation of .04. Therefore, when compar
-
ing the entire set of fitted models, and integrating information provided by fit indi
-
ces and parameter estimates, the correct model without underlying continuous fac
-
tors within class would be the most likely choice. Furthermore, on average, the
factor mixture model with two classes is not supported by the aLRT.
Larger class separation.
For larger separation, the correct model is favored
by most indices, and the proportion of correct model choice are greater than .95 for
BIC and CAIC. The AIC prefers the LPM with three classes, but the necessity of
an additional class is again rejected by the aLRT. For large separation, comparing
only the exploratory factor models for a single population based on information
criteria is inconclusive. The single-factor/two-class model can be regarded as a
competing model based on the AIC and the fact that the aLRT does not reject the
second class. However, the amount of variation in the observed variables explained
by the factors for the single-factor/two-class model is low. The observed variable
R-square averaged over the 100 replications for the 10 items for the two classes
range between 0.01 and 0.16, with an average standard deviation of 0.03. As for
the small separation, the correct model without underlying continuous factors
within class would be the logical choice.
When comparing the full set of models, there is no convincing evidence indicat-
ing an overextraction of factors. Apparently, with larger separation, higher order
moments deviate sufficiently from zero for a clearer distinction between latent
class models and factor models.
Latent Profile Model Three-Class Data (Tables 4 and 5)
The fitted models were LPMs with two, three, and four classes; exploratory fac
-
tor models for a single class with one, two, and three factors; exploratory factor
models for two classes with a single and two factors; and an exploratory factor
model for three classes with a single factor (see Tables 4 and 5). The four-class
LPMs had high nonconvergence (e.g., 43% and 39% for the two distances).
Small class separation.
Restricting the analysis to fitting LPMs would re
-
sult in a correct model choice. As in the case of the LPM two-class data, the AIC
points to a model with one additional class, which is rejected by the aLRT. Table 4
showsthataddingtheproportionsoffirstandsecondchoiceoftheAIC leadstoapro
-
portion similar to the BICs and CAIC. Note that restricting the analysis to explor
-
atory factor models for a single population would result in choosing the single-fac
-
tor/single-class model (F1C1) based on the information criteria. The observed
510
LUBKE AND NEALE

511
TABLE 4
Data Generated Under the Three-Class Latent Profile Model: Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
Using Information Criteria, and Proportion the aLRT Indicates the
Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.43 0.46 0.87 0.86 0.87 1.00
Second choice 0.44 0.41 0.00 0.01 0.00 0.00
Larger class separation
First choice 0.39 0.40 0.71 0.68 0.71 1.00
Second choice 0.32 0.31 0.00 0.03 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 5
Averaged Fit Statistics for 200 Data Sets Generated by a Three-Class
Latent Profile Model (LPM), Fitting Latent Profile Models LPMc
j
With
j
Classes and Latent Factor Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPMc2 –7026.45 14134.91 14315.18 14185.02 0.00
LPMc3 –6715.59 13555.18 13827.79 13630.95 0.00
LPMc4 –6693.15 13552.30 13917.25 13653.75 0.42
F1c1 –7127.11 14314.21 14446.12 14350.88 NA
F2c1 –7119.31 14316.61 14488.09 14364.28 NA
F3c1 –7116.47 14326.94 14533.59 14384.38 NA
F1c2 –6832.88 13787.76 14055.97 13862.32 0.03
Larger class separation
LPMc2 –7413.06 14908.12 15088.39 14958.23 0.00
LPMc3 –6776.34 13676.67 13949.28 13752.45 0.00
LPMc4 –6756.27 13678.54 14043.49 13779.98 0.42
F1C1 –7308.00 14676.00 14807.91 14712.67 NA
F2C1 –7302.81 14683.61 14855.09 14731.28 NA
F3C1 –7300.56 14695.11 14901.77 14752.56 NA
F1C2 –6981.01 14084.02 14352.23 14158.58 0.02
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

R-square for the 10 items averaged overthe 100 replications ranges between .26 and
.49, with an average standard deviation of 0.03. The explained factor variance is
higher than for the two-class LPM data. This difference is most likely due to the fol
-
lowing.Althoughconsecutiveclassesinthethree-classdatahaveaseparationequiv
-
alent to the two-class data, Classes 1 and 3 have twice the distance. Hence, multi
-
variate kurtosis is clearly more pronounced in the three-class data. As classes in the
data-generating model are more separated, overextraction of factors becomes a
more serious threat if an analysis is restricted to exploratoryfactor models assuming
a homogeneous population. Tables 4 and 5 show that when comparing the whole set
of fitted models, the correct three-class LPM would be chosen. Note that the conver
-
gencerateofthe two-factor/two-classmodelandthesingle-factor/three-classmodel
were below 40%, and average results are not presented.
Large class separation.
All indices identify the correct model as the best-
fitting model. The proportions of correct model choice are slightly lower than for
the smaller separation. This is due to the fact that the four-class model is a compet-
ing model. However, the fourth class is rejected by the aLRT. Restricting the analy-
sis to exploratory factor model for a homogeneous population would again result
in choosing the single-factor/single-class model. The average observed variable
R-square corresponding to this fitted model is higher for large separation than for
small separation and ranges between .51 and .67, with an average standard devia-
tion of .03. This shows that the probability of overextracting factors may increase
with larger class separation.
Single-Factor/Single-Class Data (Tables 6 and 7)
Six different models were fitted to the single-factor/single-class data: LPMs
with two and three classes; exploratory single-class models with one, two, or three
factors; and an exploratory single-factor/two-class model. The results are shown in
Tables 6 and 7.
Comparing only the results for the LPMs, the LPA model with three classes
would have been chosen. A closer look at individual output files from the three-
class LPA model does not reveal any obvious signs of a possible model mis
-
specification such as inadmissible parameter estimates or extremely small classes.
Therefore, the results provide evidence for an overextraction of classes to account
for covariances that are due to an underlying continuous dimension if the analyses
are restricted to an LPA.
When comparing results of all fitted models, the three information criteria con
-
sistently reject the LPMs in favor of a factor model. The BIC and CAIC select the
correct model 93% of the time. Based on AIC and saBIC, the correct model would
be the first choice 33% and 42% of the time, respectively, and the second choice
39% of the time. The reason for the poorer performance of the AIC and the saBIC
512
LUBKE AND NEALE

is that they favor the overparameterized F1C2 and F2C1 models as can be seen in
Table 7 showing the averaged fit indices. However, the two-class model is not sup
-
ported by a nonsignificant aLTR, which indicates that the second class is not nec
-
essary. Individual output files show that the factor correlation in the F2C1 model is
close to unity. More specifically, the factor correlation averaged over the 100 repli
-
cations equals .99 with a standard deviation of .17. Taken together, the correct
model would most likely have been chosen.
Two-Factor/Single-Class Data With Simple Structure
(Tables 8 and 9)
The same models were fitted as for the single-factor/single-class data (see Table
9). As shown previously, if an analysis had been restricted to an LPA, too many
classes would have been extracted. A comparison of LPMs based on information
LATENT CLASSES AND FACTORS 513
TABLE 6
Data Generated Under the Single-Factor/Single-Class Model: Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
Using Information Criteria, and Proportion the aLRT Indicates the
Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
First choice 0.03 0.33 0.93 0.42 0.93 1.00
Second choice 0.17 0.39 0.00 0.39 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 7
Averaged Fit Statistics for 200 Data Sets Generated by a
Single-Class/Single-Factor Model, Fitting Latent Profile Models LPMc
j
With
j
Classes and Latent Factor Models F
i
c
j
with
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
LPAc2 –2691.47 5464.94 5600.17 5470.27 0.02
LPAc3 –2527.07 5178.15 5382.64 5186.22 0.08
F1C1 –2445.65 4951.31 5050.26 4955.21 NA
F2C1 –2433.13 4944.25 5072.89 4949.33 NA
F3C1 –2432.44 4958.87 5113.89 4964.99 NA
F1C2 –2421.10 4944.19 5112.41 4950.83 0.76
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

criteria would have resulted in choosing the three-class LPM when the population
in fact consists of only a single class. The aLRT favored the two-class LPM. How
-
ever, if exploratory factor mixture models are part of the set of fitted models, the
correct model would have been chosen. On average, all indices are in agreement
and point to the two-factor/single-class model. Note that, again, the AIC and
saBIC only select the correct model 51% and 66% of the time. Adding the sec
-
ond-choice proportion leads to acceptable values (i.e., > .95). As can be seen in
Table 9 containing the average values, this is due to the competing three-factor sin
-
gle-class model. That model would, however, be rejected due to high factor corre
-
lations pertaining to the third factor.
514
LUBKE AND NEALE
TABLE 8
Data Generated Under the Two-Factor/Single-Class Model With Simple
Structure: Proportions of Monte Carlo Replications the Correct Model Is First
and Second Choice Using Information Criteria, and Proportion the aLRT
Indicates the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
First choice 0.00 0.51 1.00 0.66 1.00 1.00
Second choice 0.17 0.44 0.00 0.30 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 9
Averaged Fit Statistics for 200 Data Sets Generated by a
Two-Factor/Single-Class Model With Simple Structure, Fitting Latent
Profile Models LPMc
j
With
j
Classes and Latent Factor Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
LPAc2 –2888.88 5859.76 5995.00 5865.10 0.02
LPAc3 –2780.10 5684.19 5888.69 5692.26 0.14
F1C1 –2788.44 5636.89 5735.84 5640.79 NA
F2C1 –2589.50 5257.01 5385.64 5262.08 NA
F3C1 –2584.58 5263.17 5418.19 5269.29 NA
F1C2 –2623.65 5369.31 5570.51 5377.25 0.13
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

Two-Factor/Single-Class Data With Cross Loadings
(Tables 10 and 11)
The same set of models were fitted as for the preceding simple structure case.
The results show no real difference with the models fitted to the two-factor sin
-
gle-class data with simple structure (cf. Tables 8 and 9 to Tables 10 and 11). The
proportions of correct model choice are slightly improved for the AIC and
saBIC.
LATENT CLASSES AND FACTORS 515
TABLE 10
Data Generated Under the Two-Factor/Single-Class Model With Cross
Loadings: Proportions of MC Replications the Correct Model Is First
and Second Choice Using Information Criteria, and Proportion the aLRT
Indicates the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
First choice 0.032 0.78 1.00 0.84 1.00 1.00
Second choice 0.168 0.22 0.00 0.16 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 11
Averaged Fit Statistics for 200 Data Sets Generated
by a Two-Factor/Single-Class Model Wth Cross Loadings, Fitting Latent
Profile Models LPMc
j
With
j
Classes and Latent Factor Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
LPMc2 –2790.98 5663.96 5799.19 5669.30 0.01
LPMc3C3 –2695.18 5514.36 5718.86 5522.43 0.13
F1C1 –2676.90 5413.90 5512.90 5417.80 NA
F2C1 –2558.30 5194.60 5323.30 5199.70 NA
F3C1 –2552.10 5198.20 5353.30 5204.40 NA
F1C2 –2603.37 5308.73 5476.95 5315.38 0.24
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

Measurement Invariant Single-Factor/Two-Class Data
(Tables 12 and 13)
Here, the classes in the true data only differ with respect to the factor mean. The
set of fitted models includes LPMs with two, three, and four classes; exploratory
factor models with a single, two, and three factors; and a two-class exploratory fac
-
tor model with a single factor. Tables 12 and 13 depict the results for this data type.
Small separation.
Restricting the analysis to fitting only LPMs would lead
to an overextraction of classes. Among LPMs, information criteria favor the
four-class model. Similar to the previous results, a model with one class less would
be chosen if the decision is based on the aLRT, that is, a three-class model would be
chosen where the true data consist of two classes. Exploratory factor analysis for a
single population lead to the following average results. The AIC favors the
three-factor/single-class model (F3C1), but this would in practice be rejected due
to very high factor correlations. Averaged over the 100 replications, all factor cor-
relations are .99 with standard deviations of .13, .09, and .14 for the three different
factor correlations. The BIC and saBIC are smallest for the single-factor/sin-
gle-class model (F1C1). The aLRT for the correct single-factor/two-class model
(F1C2) versus the incorrect F1C1 is only marginally significant, hence the neces-
sity of a second class would be debatable in practice. Apparently, with small sepa-
ration between classes and a within-class sample size of 200, for the measurement
invariant F1C2 model there is a possibility of underestimating the number of
classes. Table 13 shows that there are a number of competing models with small
differences in average information criteria. This is consistent with the very low
proportions of correct model choice presented in Table 12.
516
LUBKE AND NEALE
TABLE 12
Data Generated Under the Measurement Invariant
Single-Factor/Two-Class Model: Proportions of Monte Carlo Replications
the Correct Model Is First and Second Choice Using Information Criteria,
and Proportion the aLRT Indicates the Necessity of the Second Class
When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.85 0.18 0.00 0.02 0.00 0.63
Second choice 0.09 0.26 0.18 0.26 0.17 0.00
Larger class separation
First choice 1.00 0.99 0.02 0.90 0.00 0.99
Second choice 0.00 0.00 0.47 0.09 0.37 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.

Larger separation.
With larger class separation, the risk of overextracting
classes when fitting only latent class models is higher than for small separation. All
indices point to the LPM with four classes. Exploratory factor analysis for a single
population would be inconclusive as the information criteria do not consistently fa
-
vor the same model. However, comparing the entire set of fitted models would, on
average, lead to a correct model choice. The proportions of correct model choice
are good for AIC and saBIC (i.e., > .95) but low for BIC and CAIC. As we see, the
situation is much improved when the data-generating model deviates from mea
-
surement invariance (e.g., class-specific intercepts and/or loadings).
Noninvariant Single-Factor/Two-Class Data:
Class-Specific Intercepts (Tables 14 and 15)
Small separation.
The same models were fitted as to the measurement in
-
variant single-factor model/two-class data. In the noninvariant case, with small
class separation, the AIC favors the correct model in a much higher proportion
of MC replications than in the measurement invariance (MI) case (cf. Tables 14
LATENT CLASSES AND FACTORS 517
TABLE 13
Averaged Fit Statistics for 200 Data Sets Generated
by a Measurement Invariant Single-Factor/Two-Class Model,
Fitting Latent Profile Models LPMc
j
With
j
Classes and Latent Factor
Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPMc2 –5791.85 11665.70 11829.35 11699.25 0.01
LPMc3 –5352.87 10829.74 11077.21 10880.48 0.04
LPMc4 –5152.52 10471.03 10802.33 10538.96 0.11
F1C1 –5002.60 10065.20 10185.00 10089.80 NA
F2C1 –4994.50 10066.90 10222.60 10098.90 NA
F3C1 –4984.60 10063.30 10250.90 10101.80 NA
F1C2 –4985.56 10073.12 10276.68 10114.86 0.08
Larger class separation
LPMc2 –6265.20 12612.39 12776.04 12645.95 0.00
LPMc3 –5759.67 11643.35 11890.82 11694.09 0.09
LPMc4 –5445.28 11056.56 11387.85 11124.48 0.05
F1C1 –5153.00 10366.00 10485.70 10390.50 NA
F2C1 –5144.50 10367.00 10522.70 10398.90 NA
F3C1 –5133.80 10361.50 10549.20 10400.00 NA
F1C2 –5113.49 10328.97 10532.54 10370.71 0.00
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

518
TABLE 14
Data Generated Under the Single-Factor/Two-Class Model
With Class-Specific Intercepts: Proportions of Monte Carlo Replications
the Correct Model Is First and Second Choice Using Information Criteria,
and Proportion the aLRT Indicates the Necessity of the Second Class
When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.94 0.52 0.00 0.19 0.00 1.00
Second choice 0.05 0.29 0.41 0.55 0.19 0.00
Larger class separation
First choice 1.00 1.00 1.00 1.00 1.00 1.00
Second choice 0.00 0.00 0.00 0.00 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 15
Averaged Fit Statistics for 200 Data Sets Generated
by a Single-Factor/Two-Class Model With Intercept Differences,
Fitting Latent Profile Models LPMc
j
With
j
Classes and Latent Factor
Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPMc2 –5560.68 11203.36 11367.01 11236.92 0.01
LPMc3 –5280.87 10685.73 10933.20 10736.47 0.07
LPMc4 –5161.74 10489.48 10820.77 10557.40 0.13
F1C1 –5114.00 10288.10 10408.00 10313.01 NA
F2C1 –5056.90 10191.80 10347.50 10223.70 NA
F3C1 –5050.40 10194.80 10382.40 10233.30 NA
F1C2 –5043.93 10189.86 10393.42 10231.60 0.00
Larger class separation
LPMc2 –5946.46 11974.92 12138.57 12008.47 0.01
LPMc3 –5715.29 11554.58 11802.05 11605.32 0.06
LPMc4 –5506.48 11178.96 11510.25 11246.88 0.06
F1C1 –5639.00 11338.01 11458.01 11363.03 NA
F2C1 –5252.50 10583.00 10738.60 10614.90 NA
F3C1 –5249.10 10592.21 10779.80 10630.60 NA
F1C2 –5158.86 10419.73 10623.29 10461.46 0.00
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

and 15). The BIC and the aBIC do not perform well, and point to the two-fac
-
tor/single-class model (see Tables 14 and 15) . However, the average factor cor
-
relation is again high, namely, .95 with a standard deviation of .04. Restricting
the analysis to exploratory factor analysis for a single population would there
-
fore be unlikely to result in an overextraction of factors. Comparison of the en
-
tire set of fitted models would likely result in a correct model choice because the
aLRT clearly indicated the necessity of a second class.
Larger separation.
For the large class separation, all indices are in accor
-
dance and would lead to a correct model choice. The proportion correct model se
-
lection equals unity for all information criteria. Overall, the results are clearly
better than for the measurement invariant model.
Noninvariant Single-Factor/Two-Class Data:
Class-Specific Intercepts and Factor Loadings
(Tables 16 and 17)
Small separation.
Class-specific factor loadings in addition to class-spe-
cific intercepts further increase the proportion correct model selection when com-
pared to the measurement invariant model and the model with class-specific inter-
cepts. Note that the class separation had been kept unchanged when introducing
class specific parameters during data generation. The AIC and the sample size
aBIC outperform the BIC and CAIC.
Larger separation.
As for the model with class-specific intercepts, model
choice is unproblematic for larger class separation.
LATENT CLASSES AND FACTORS 519
TABLE 16
Data Generated Under the Single-Factor/Two-Class Model
With Class-Specific Intercepts and Factor Loadings: Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
Using Information Criteria, and Proportion the aLRT Indicates
the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 1.00 0.99 0.35 0.98 0.16 1.00
Second choice 0.00 0.01 0.62 0.01 0.66 0.00
Larger class separation
First choice 1.00 1.00 1.00 1.00 1.00 1.00
Second choice 0.00 0.00 0.00 0.00 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.

Two-Factor/Two-Class Data (Tables 18 and 19)
The data are generated under a model with moderate factor correlations in both
classes (i.e., .5), and class-specific intercepts. The fitted models include LPMs
with two to four classes, exploratory single-class factor models with one to three
factors, and exploratory two-class factor mixture models with one and two factors.
The results are shown in Tables 18 and 19.
Small separation.
An LPA would lead to overextraction of classes. All in
-
formation criteria favor the four-class LPM when considering only LPMs. The
aLRT points to the three-class LPM. An exploratory factor analysis would lead to
the correct number of factors. Although the fit measures indicate three factors, one
of the factors has correlations with the other two factors approaching unity. When
considering all fitted models, the correct model would be chosen. Although the
percentage of correct model choice is low for the AIC, and close to zero for the
BIC and CAIC, Table 19 shows that the competing models are two- and three-fac
-
tor single-class models. The aLRT is significant in 83% of the MC replications (see
520
LUBKE AND NEALE
TABLE 17
Averaged Fit Statistics for 200 Data Sets Generated
by a Single-Factor/Two-Class Model With Intercept and Loading
Differences, Fitting Latent Profile Models LPMc
j
With
j
Classes
and Latent Factor Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPAc2 –5568.18 11218.36 11382.01 11251.91 0.01
LPAc3 –5295.87 10715.74 10963.21 10766.48 0.07
LPAc4 –5180.56 10527.13 10858.42 10595.06 0.11
F1C1 –5142.26 10344.53 10464.27 10369.08 NA
F2C1 –5087.63 10253.27 10408.94 10285.19 NA
F3C1 –5079.22 10252.45 10440.05 10290.91 NA
F1C2 –5054.18 10210.36 10413.92 10252.09 0.00
Larger class separation
LPAc2 –5949.27 11980.54 12144.19 12014.09 0.00
LPAc3 –5718.41 11560.82 11808.29 11611.56 0.06
LPAc4 –5511.03 11188.05 11519.35 11255.98 0.10
F1C1 –5654.92 11369.83 11489.58 11394.39 NA
F2C1 –5281.25 10640.49 10796.16 10672.41 NA
F3C1 –5271.35 10636.69 10824.29 10675.16 NA
F1C2 –5159.15 10420.3 10623.86 10462.04 0.00
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

521
TABLE 18
Data Generated Under the Two-Factor/Two-Class Model: Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
Using Information Criteria, and Proportion the aLRT Indicates
the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.99 0.57 0.00 0.17 0.00 0.83
Second choice 0.01 0.43 0.53 0.83 0.22 0.00
Larger class separation
First choice 1.00 1.00 1.00 1.00 1.00 1.00
Second choice 0.00 0.00 0.00 0.00 0.00 0.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.
TABLE 19
Averaged Fit Statistics for 200 Data Sets Generated
by a Two Factor/two Class Model, Fitting Latent Profile Models LPMc
j
With
j
Classes and Latent Factor Models F
i
c
j
With
i
Factors and
j
Classes
logL-val AIC BIC saBIC aLRT
Small class separation
LPAc2 –5925.73 11933.47 12097.12 11967.02 0.01
LPAc3 –5753.79 11631.58 11879.05 11682.32 0.09
LPAc4 –5623.29 11412.59 11743.88 11480.52 0.13
F1C1 –5745.32 11550.63 11670.38 11575.18 NA
F2C1 –5400.01 10878.02 11033.68 10909.93 NA
F3C1 –5350.53 10795.06 10982.66 10833.52 NA
F1C2 –5556.19 11214.37 11417.94 11256.11 0.05
F2C2 –5338.01 10794.02 11029.52 10842.31 0.04
Larger class separation
LPAc2 –6250.65 12583.31 12746.96 12616.86 0.02
LPAc3 –6083.58 12291.15 12538.62 12341.89 0.11
LPAc4 –5953.88 12073.76 12405.05 12141.68 0.16
F1C1 –6116.62 12293.23 12412.98 12317.79 NA
F2C1 –5759.40 11596.79 11752.46 11628.71 NA
F3C1 –5542.90 11179.80 11367.40 11218.27 NA
F1C2 –5812.71 11727.43 11930.99 11769.17 0.01
F2C2 –5453.08 11024.17 11259.66 11072.45 0.00
Note. Results for the true model are in bold, as are the values of fit indices that favor misspecified
models. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian informa
-
tion criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likelihood
ratio test.

Table 18), indicating the need for a second class. The three-factor model is rejected
due to high factor correlations.
Larger separation.
Correct model choice is unproblematic for larger class
separation.
Part 2: Effects of Sample Size and Class Proportions
Varying Sample Size (Table 20)
The data-generating model used to investigate the effect of sample size on the
detection of the correct model is the single-factor/two-class model with class-spe
-
cific intercepts. The within-class sample size is denoted as N
wc
. Results for N
wc
=
200 for this model were also presented in Part 1. Here, results are presented for
within-class sample sizes of N
wc
=25, 50, 75, 150, 200, 300, and 1000. Table 20
shows the proportion correct model choice (i.e., “first choice” in the previous ta-
bles) as a function of increasing sample size. The averaged results are discussed in
the text.
522
LUBKE AND NEALE
TABLE 20
Data Generated Under the Single-Factor/Two-Class Model With Intercept
Differences: Proportions of Correct Model Choice as a Function
of Increasing Sample Size, and Proportion the aLRT Indicates
the Necessity of the Second Class When Fitting the Correct Model
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
N
wc
= 25 0.25 0.30 0.00 0.48 0.00 0.06
N
wc
= 50 0.62 0.25 0.00 0.52 0.00 0.37
N
wc
= 75 0.78 0.34 0.00 0.42 0.00 0.77
N
wc
= 150 0.87 0.40 0.00 0.20 0.00 0.98
N
wc
= 200 0.94 0.52 0.00 0.19 0.00 1.00
N
wc
= 300 0.94 0.68 0.00 0.19 0.00 1.00
N
wc
= 1,000 1.00 1.00 0.10 0.85 0.01 1.00
Larger class separation
N
wc
= 25 0.86 0.78 0.18 0.88 0.04 0.90
N
wc
= 50 1.00 1.00 0.57 1.00 0.34 0.99
N
wc
= 75 1.00 1.00 0.88 1.00 0.67 1.00
N
wc
= 150 1.00 1.00 1.00 1.00 1.00 1.00
N
wc
= 200 1.00 1.00 1.00 1.00 1.00 1.00
N
wc
= 300 1.00 1.00 1.00 1.00 1.00 1.00
N
wc
= 1,000 1.00 1.00 1.00 1.00 1.00 1.00
Note. aLRT = adjusted likelihood ratio test; logL-val = log likelihood value; AIC = Akaike infor
-
mation criterion; BIC = Bayesian information criterion; saBIC = sample size adjusted BIC; CAIC =
Consistent AIC.

Small separation.
For very small within-class sample sizes (i.e., N
wc
= 25),
the only distinction that can be made is between LPMs on one hand and factor
models on the other hand. LPMs with two and three classes have clearly higher in
-
formation criteria and would therefore be rejected in favor of the factor models.
However, a comparison of the information criteria corresponding to the different
factor models does not allow for a clear decision. Average information criteria
show that single-class models with one or two factors are competing models. The
BIC and CAIC favor the single-factor/one-class model, whereas the AIC and
saBIC favor the two-factor model. Factor correlations are high on average (i.e.,
.88). The aLRT does not support the necessity of the second class. In only 6 of the
100 MC replications was the aLRT less than or equal to .05. In sum, most likely the
population heterogeneity would not be detected, and a single-factor or two-factor
model for a single class would be chosen.
Increasing the within-class sample size results in an increasing proportion cor
-
rect model selection when considering AIC and aLRT. Interestingly, the saBIC
proportion correct drops to .2 for sample sizes between 150 and 300 before picking
up again for large sample sizes. The two-factor/one-class model remains the com-
peting model for the BIC and CAIC. As sample size increases, the factor correla-
tion approaches unity (i.e., at N
wc
= 200, it equals .95). When considering informa-
tion criteria, aLRT, and parameter estimates jointly, a within-class sample size of
N
wc
= 200 is a conservative choice to achieve correct model selection, and reason-
able results can be expected with 75 participants within class for this type of data.
Larger separation.
For larger class separation, the results in Table 20 show
that AIC, saBIC, and aLRT perform well at a sample size as small as 25 partici-
pants within class. For the BIC and CAIC, the competing model is again the
two-factor/single-class model. When considering the BIC or CAIC jointly with the
factor correlations, a within-class sample size of 50 would be sufficient.
It is important to note that if model comparisons are restricted to LPMs, larger
sample sizes lead to accepting models with a greater number of classes. For the
largest sample size in this study, the four-class LPM would be accepted although
the true data only have two classes. This indicates that the risk of overestimating
the number of classes when using only latent class or LPMs is pronounced in stud
-
ies with large sample sizes. This result emphasizes the importance of fitting both
latent class models and models with continuous latent variables to a given data set.
Unequal class proportions (Table 21)
The data-generating model is again the single-factor/two-class model with in
-
tercept differences. Here, class proportions are .9 and .1, respectively. The results
in Table 21 can therefore be compared to Table 14 showing the results of the same
data-generating model with equal class proportions.
LATENT CLASSES AND FACTORS 523

Except for the aLRT, unequal class proportions do not seem to have a major im-
pact on the results. Interestingly, the AIC and saBIC perform slightly better than
for equal class proportions, which is consistent with the findings concerning small
sample sizes. The aLRT seems to be very sensitive to unequal class proportions
and fails to support the second class 95% of the time when class separation is
small. In general, however, the conclusions are similar to the ones concerning
equal class proportions. The two-factor/single-class model is a competing model.
This model would be rejected when considering factor correlations. For larger
class separation, the correct model is detected without difficulty.
CONCLUSIONS
The results of this study demonstrate the value of exploratory factor mixture analy
-
sis for investigating the within-class factor structure while accounting for possible
population heterogeneity. For the data-generating models and sample sizes used in
this study, comparing the fit of a set of exploratory factor mixture models including
LPMs correctly distinguishes between categorical and/or continuous latent vari
-
ables in most cases.
A consistent finding in this study is that restricting analyses to fitting only latent
class models results in overextraction of classes if continuous factors are part of the
latent structure of the true data-generating model. The finding confirms examples
provided in Bauer and Curran (2004), where fitting overly restricted within-class
models resulted in spurious classes. In our simulation, overextraction of classes
seems to depend neither on the type of within-class factor structure (i.e., simple
structure vs. presence of cross loadings) nor on the restrictiveness of the
524
LUBKE AND NEALE
TABLE 21
Data Generated Under the Single-Factor/Two-Class Model With Intercept
Differences: Effect of Unequal Class Proportions on the Proportions
of Monte Carlo Replications the Correct Model Is First and Second Choice
logL-val AIC BIC saBIC CAIC aLRT
Small class separation
First choice 0.89 0.77 0.00 0.39 0.00 0.05
Second choice 0.06 0.11 0.09 0.24 0.06 0.00
Larger class separation
First choice 1.00 1.00 1.00 1.00 1.00 0.98
Second choice 0.00 0.00 0.00 0.00 0.00 0.00
Note. logL-val = log likelihood value; AIC = Akaike information criterion; BIC = Bayesian infor
-
mation criterion; saBIC = sample size adjusted BIC; CAIC = Consistent AIC; aLRT = adjusted likeli
-
hood ratio test.

within-class model for the means (i.e., measurement invariant vs. intercept differ
-
ences). Overextraction occurred when the number of within-class factors was
misspecified (i.e., specified as zero when in fact it was nonzero). It is noteworthy
that the risk of overextracting classes is higher for large sample sizes.
Evidence for an overextraction of factors if an analysis is restricted to explor
-
atory factor analysis and the population is incorrectly assumed to be homoge
-
neous is less clear. The correct number of factors in a latent class model is zero.
The researcher conducting factor analysis will likely conclude that there is at
least one factor. That is, latent classes can generate data that appear to indicate a
latent factor. This is a minimal level of overextraction. Sometimes two or more
factors provide a better fit than a single-factor model, which is a more serious
level of overextraction. However, although the fit indices can indicate the need
for additional, spurious factors to account for covariation due to population het
-
erogeneity, individual output files in this study show that factor structures result
-
ing from fitting models with too many factors are rather unsatisfactory. High fac
-
tor correlations suggest that the dimension of the factor space in the true model
may be lower. Similarly, very low percentages of variance explained by the fac-
tor can be used as an indication that models without a continuous latent factor
may be more appropriate.
It is a well-known fact in mixture analysis (McLachlan, Peel, & Bean, 2001)
that correct model choice depends heavily on the separation between classes in the
true model. In our study, classes were separated between 1.5 and 3 standard devia-
tions. Note that when classes are separated by a 1.5 standard deviation factor mean
difference and factors within class are normally distributed, the mixture distribu-
tion of factor scores would still look approximately normal, and not bimodal
(McLachlan & Peel, 2000). Smaller class separations are investigated in Lubke
and Muthén (2007). Our study clearly demonstrates that sample size plays an im
-
portant role in addition to class separation. Regarding correct model choice, there
is a clear trade-off between sample size and class separation. Interestingly, for the
data-generating model used in the second part of our simulation, sample sizes as
small as 75 participants within class often result in a correct model choice even if
the distance between classes is so small that the heterogeneity would be hard to de
-
tect when plotting the data. For a larger separation, sample sizes as small as 25 par
-
ticipants within class result on average in correct model detection. Interestingly,
the AIC and saBIC outperform the BIC and CAIC. The aLRT performs well at
very small sample sizes. Further research is needed to investigate in how far results
with respect to sample size can be generalized to other types of mixture models and
other degrees of separation.
A limitation of this study is the fact that the data-generating models did not vio
-
late any of the model assumptions. There is no doubt that model assumptions such
as multivariate normality are frequently unrealistic in practice. Our study repre
-
sents a first step in evaluating exploratory factor mixture models when neither the
LATENT CLASSES AND FACTORS 525

number of factors nor the number of classes is known. Future research will show in
how far our results are robust against violations of model assumptions. In addition,
the emphasis in this study was on mean differences between classes. It remains to
be seen if covariance structure differences can be detected in the absence of sub
-
stantial mean differences.
In summary, comparing the fit of a set of exploratory factor mixture models in
-
cluding classic latent class models as proposed in this article helps to prevent
overextraction of classes or factors. This study is could be extended in several
ways. First, as previously mentioned, the data were generated in accordance with
the assumptions of exploratory factor mixture models. Violations of some of these
assumptions (e.g., normally distributed variables within class) have been shown to
induce an overestimation of the number of classes (Bauer & Curran, 2003a) and
therefore warrant further investigation. Second, the number of latent classes and
the number of factors was kept low in our study. The choice between a four-fac
-
tor/three-class model and a three-factor/four-class model based on information cri
-
teria, the aLRT, and inspection of individual output files may be more difficult than
for one or two factors and/or classes. Higher dimensions of the latent space are cur-
rently under investigation. Third, the set of exploratory factor mixture models will
in practice be unlikely to include the true model. Real data may well stem from
more complex data-generating processes, including, for instance, differences be-
tween classes with respect to the structure of the within-class model. Hence, in
practice, model selection will be between more or less misspecified models, and
decisions are likely to be less clear cut. A broader set of candidate models could
therefore be tested. Finally, it is not clear how reliable distinctions will be when the
observed data consist of ordinal or mixed data instead of the continuous measures
used here. This issue is also currently being investigated.
ACKNOWLEDGMENTS
The research of both authors was supported through grant MH65322 by the
National Institutes of Mental Health and through grant DA018673 by the National
Institute on Drug Abuse.
REFERENCES
Akaike, H. (1974). A new look at statistical model identification. IEEE Transactions on automatic Con
-
trol, AU-19, 719–722.
Akaike, H. (1987). Factor analysis and AIC. Psychometrika, 52, 317–332.
Arminger, G., Stein, P., & Wittenberg, J. (1999). Mixtures of conditional meanand covariance structure
models. Psychometrika, 64, 475–494.
526 LUBKE AND NEALE

Bartholomew, D. J. (1987). Latent variables models and factor analysis. New York: Oxford University
Press.
Bartholomew, D. J., & Knott, M. (1999). Latent variables models and factor analysis (2nd ed.). Lon
-
don: Arnold.
Bauer, D. B., & Curran, P. J. (2003a). Distributional assumptions of growth mixture models: Implica
-
tions for overextraction of latent trajectory classes. Psychological Methods, 8, 338–363.
Bauer, D. B., & Curran, P. J. (2003b). Overextraction of latent trajectory classes: Much ado about noth
-
ing? Reply to Rindskopf (2003), Muthen (2003), and Cudeck and Henly (2003). Psychological
Methods, 8, 384–393.
Bauer, D. B., & Curran, P. J. (2004). The integration of continuous and discrete latent variable models:
Potential problems and promising opportunities. Psychological Methods, 9, 3–29.
Bozdogan, H. (1987). Model selection and Akaike’s Information Criterion (AIC): The general theory
and its analytical extensions. Psychometrika, 52, 345–370.
Cudeck, R., & Henly, S. J. (2003). A realistic perspective on pattern prepresentation in growth data:
Comment on Bauer and Curran. Psychological Methods, 8, 378–383.
Dolan, C. V., & van der Maas, H. L. J. (1998). Fitting multivariate normal finite mixtures subject to
structural equation modeling. Psychometrika, 63, 227–253.
Heinen, T. (1993). Discrete latent variable models. Tilburg, The Netherlands: Tilburg University Press.
Heinen, T. (1996). Latent class and discrete latent trait models: Similarities and differences. Thousand
Oaks, CA: Sage.
Hicks, B. M., Krueger, R. F., Iacono, W. G., McGue, M., & Patrick, C. J. (2004). Family transmission
and heritability of externalizing disorders: A twin-family study. Archives of General Psychiatry, 61,
922–928.
Jedidi, K., Jagpal, H. S., & DeSarbo, W. S. (1997). Finite mixture structural equation models for re-
sponse based segmentation and unobserved heterogeneity. Marketing Science, 16, 39–59.
Krueger, R. F., Nichol, P. E., Hicks, B. M., Markon, K. E., Patrick, C. J., Lacono, W. G., et al. (2004).
Using latent trait modeling to conceptualize an alcohol problems continuum. Psychological Assess-
ment, 16, 107–119.
Langenheine, R., & Rost, J. (1988). Latent trait and latent class models. New York: Plenum.
Lawley, D. N., & Maxwell, A. E. (1971). Factor analysis as a statistical method. London: Butterworth.
Lazarsfeld, P. F., & Henry, N. W. (1968). Latent structure analysis. Boston: Houghton Mifflin.
Lo, Y., Mendell, N., & Rubin, D. B. (2001). Testing the number of components in a normal mixture.
Biometrika, 88, 767–778.
Lubke, G. H., Dolan, C. V., Kelderman, H., & Mellenbergh, G. J. (2003). Absence of measurement bias
with respect to unmeasured variables: An implication of strict factorial invariance. British Journal of
Mathematical and Statistical Psychology, 56, 231–248.
Lubke, G. H., & Muthén, B. O. (2005). Investigating population heterogeneity with factor mixture
models. Psychological Methods, 10, 21–39.
Lubke, G. H., & Muthén, B. O. (2007). Performance of factor mixture models as a function of model
size, covariate effects, and class-specific parameters. Structural Equation Modeling, 14, 26–47.
McLachlan, G. J., & Peel, D. (2000). Finite mixture models. New York: Wiley.
McLachlan,G.J., Peel, D.,& Bean, R.W.(2001). Modelling high-dimensionaldataby mixtures of factor
analyzers(Tech.Rep.).Brisbane,Australia:Department ofMathematics,UniversityofQueensland.
Mellenbergh, G. J. (1994). Generalized linear item response theory. Psychological Bulletin: Quantita
-
tive Methods in Psychology, 115, 300–307.
Meredith, W. (1993). Measurement invariance, factor analysis, and factorial invariance.
Psychometrika, 58, 525–543.
Meredith, W., & Horn, J. (2001). The role of factorial invariance in modeling growth and change. In L.
M. Collins & A. G. Sayer (Eds.), New methods for the analysis of change (pp. 203–240). Washing
-
ton, DC: American Psychological Association.
LATENT CLASSES AND FACTORS
527

Molenaar, P. C. M., & van Eye, A. (1994). On the arbitrary nature of latent variables. In A. van Eye & C.
C. Clogg (Eds.), Latent variable analysis: Applications for developmental research (pp. 226–242).
Thousand Oaks, CA: Sage.
Muthén, B. O. (2003). Statistical and substantive checking in growth mixture modeling: Comment on
Bauer and Curran (2003). Psychological Methods, 8, 369–377.
Muthén, B. O., & Shedden, K. (1999). Finite mixture modeling with mixture outcomes using the EM
algorithm. Biometrics, 55, 463–469.
Pickles,A., &Angold,A. (2003).Natural categoriesorfundamentaldimensions:Oncarving natureatthe
joints and the rearticulation of psychopathology. Development and Psychopathology, 15, 529–551.
Rindskopf, D. (2003). Mixture or homogeneous? Comment on Bauer and Curran (2003). Psychologi
-
cal Methods, 8, 364–368.
Schwarz, G. (1978). Estimating the dimensions of a model. Annals of Statistics, 6, 461–464.
Sclove, S. (1987). Application of model-selection criteria to some problems in multivariate analysis.
Psychometrika, 52, 333–343.
Spearman. (1904). General intelligence, objectively determined and measured. American Journal of
Psychology, 15, 201–293.
Vermunt, J. K., & Magidson, J. (2003). Latent class models for classification. Computational Statistics
& Data Analysis, 41, 531–537.
Vermunt, J. K., & Magidson, J. (2005). Factor analysis with categorical indicators: A comparison be
-
tween traditional and latent class approaches. In A. K. d. Ark, M. A. Croon, & K. Sijtsma (Eds.), New
developments in categorical data analysis for the social and behavioral sciences (pp. 41–62).
Mahwah, NJ: Lawrence Erlbaum Associates, Inc.
Widaman, K. F., & Reise, S. P. (1997). Exploring the measurement invariance of psychological instru-
ments: Applications in the substance use domain. In K. J. Bryant, M. Windle, & S. G. West (Eds.),
The science of prevention: Methodological advances from alcohol and substance abuse research
(pp. 281–324). Washington, DC: American Psychological Association.
Yung, Y. F. (1997). Finite mixtures in confirmatory factor analysis models. Psychometrika, 62,
297–330.
528 LUBKE AND NEALE

APPENDIX
The parameter values used for the data generation are shown for each model. Only
nonzero parameters are listed.
First Part
Two-class latent profile model
Class-invariant parameters:
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
small separation, means class 2 [0.35 – .2 .6 – .75 0.35 – .2 .6 – .75 0.35 – .2]
t
larger separation, means class 2 [.8 – .8 1.2 – 1.2 .8 – .8 1.2 – 1.2 .8 – .8]
t
Three-Class Latent Profile Model
Class-invariant parameters:
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
means class 1 u = [0 0 0 0 0]
t
small separation, means class 2 [.7 – .8 1.2 – 1.1 .7 – .8 1.2 – 1.1 .7 – .8]
t
small separation, means class 3 [1.3 – 1.2 1.2 – 1.3 1.3 – 1.2 1.3 – 1.2 1.3 – 1.2]
t
larger separation, means class 2 [1.5 – 1.5 1.8 – 1.8 1.5 – 1.5 1.8 – 1.8 1.5 – 1.5]
t
larger separation, means class 3 [2.5 – 2.5 2.3 – 2.3 2.5 – 2.5 2.3 – 2.3 2.5 – 2.5]
t
Single-Factor/Single-Class Model
factor loadings [1 .8 .8 .8 .8 .8 .8 .8 .8 .8]
t
factor variance 1.2
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Two-Factor/Single-Class Model With Simple Structure
factor loadings
factor covariance matrix
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
LATENT CLASSES AND FACTORS 529
1.8.8.8.800000
000001.8.8.8.8
t
éù
êú
êú
ëû
1.2 .5
.5 1.2
éù
êú
êú
ëû

Two-Factor/Single-Class Model With Crossloadings
factor loadings
factor covariance matrix
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Measurement Invariant Single-Factor/Two-Class Model
Class-invariant parameters:
factor loadings [1 .8 .8 .8 .8 .8 .8 .8 .8 .8]
t
factor variance 1.2
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
small class separation factor mean class 2 is 1.8
larger class separation factor mean class 2 is 3.4
Single-Factor/Two-Class Model With Intercept Differences
Class-invariant parameters:
factor loadings [1 .8 .8 .8 .8 .8 .8 .8 .8 .8]
t
factor variance 1.2
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
small class separation intercepts class 2 [0.35 – .2 .6 – .75 0.35 – .2 .6 – .75 0.35 – .2]
larger class separation intercepts class 2 [.8 – .8 1.2 – 1.2 .8 – .8 1.2 – 1.2 .8 – .8]
Single-Factor/Two-Class Model With Intercept and Loading
Differences
Class-invariant parameters:
factor variance 1.2
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
small class separation intercepts class 2 [0.35 – .2 .6 – .75 0.35 – .2 .6 – .75 0.35 – .2]
larger class separation intercepts class 2 [.8 – .8 1.2 – 1.2 .8 – .8 1.2 – 1.2 .8 – .8]
factor loadings class 1 [1 .8 .8 .8 .8 .8 .8 .8 .8 .8]
t
factor loadings class 2 [1 .49 .68 .95 .95 .58 .92 .61 .88 .96]
530
LUBKE AND NEALE
1.8.6.4.80.2 0.40
0 0 .2 .4 0 1 .6 .8 .4 .8
t
éù
êú
êú
ëû
1.2 .5
.5 1.2
éù
êú
êú
ëû

Two-Factor/Two-Class Model With Intercept Differences
Class-invariant parameters:
factor loadings
factor covariance matrix
residual variances [0.7 .5 .5 .5 .5 .5 .5 .5 .5 .5]
t
Class-specific parameters:
small class separation intercepts class 2 [0.35 – .2 .6 – .75 0.35 – .2 .6 – .75 0.35 – .2]
larger class separation intercepts class 2 [.8 – .8 1.2 – 1.2 .8 – .8 1.2 – 1.2 .8 – .8]
Second Part
The parameter values used to generate data with varying sample sizes are the same
as for the single-factor/two-class model with intercept differences.
Prior class probabilities used to generate data with unequal class sizes are .9 and .1.
Information Criteria
All information criteria used in our study are penalized log-likelihood functions
with the general form –2L + f(N)p where L is the loglikelihood of the estimated
model with p free parameters and f(N) is a function that may depend on the total
sample size N (Sclove, 1987). The AIC does not depend on sample size, the pen-
alty is f(N)p =2p (Akaike, 1974, 1987). The BIC, the CAIC, and the saBIC inte
-
grate N in different ways, the respective penalty terms are log(N)p and (log(N)+
1)p for the BIC and the CAIC (Bozdogan, 1987; Schwarz, 1978). The saBIC uses
(N* = (N + 2)/24) instead of N.
Example Mplus Input File: Single-Factor/Two-Class Model
TITLE: F1 C2
DATA: FILE IS yinc.dat;
VARIABLE:
NAMES ARE ycont1-ycont10 eta1 eta2 c1 c2 tc;
USEVARIABLES ARE ycont1-ycont10;
CLASSES = c(2);
ANALYSIS: TYPE = MIXTURE;
ITERATIONS = 1000;
starts=50 10;
stiterations=10;
LATENT CLASSES AND FACTORS 531
1.8.8.8.800000
000001.8.8.8.8
t
éù
êú
êú
ëû
1.2 .5
.5 1.2
éù
êú
êú
ëû

stscale=5;
stseed=198;
MODEL: %OVERALL%
f1 BY ycont1* ycont2-ycont10; f1@1;
%c#2%
f1 BY ycont1* ycont2-ycont10; f1@1; [F1@0];
!estimate class specific factor loadings
[ycont1-ycont10@0]; ycont1-ycont10;
! fix intercepts to zero in this class, estimate intercept differences in the other class
! estimate residual variances
%c#1%
f1 BY ycont1* ycont2-ycont10; f1@1; [F1@0];
[ycont1-ycont10]; ycont1-ycont10;
OUTPUT: tech1 tech8 tech11 standardized;
532
LUBKE AND NEALE