Simultaneous Latent-Class Analysis Across Groups 5969 S

Clogg and Goodman (1985). This multigroup LC

Simultaneous Latent-Class Analysis model contains three types of categorical vari-
Across Groups ables: a set of observed indicators (e.g., the item
responses from a QoL questionnaire – ▶ Quality
Miloš Kankaraš1 and Jeroen K. Vermunt2 of Life Questionnaire), one or more unobserved
1
Tilburg University, Tilburg, The Netherlands (latent) variables that account for the relation-
2
Department of Methodology and Statistics, ships between the observed indicators, and
Tilburg University, Tilburg, The Netherlands a grouping variable G, which is a categorical
observed variable that can be associated with
both the latent variable(s) and the indicators.
Synonyms The main model assumption is that the indicators
are mutually independent conditionally on the
Multigroup latent-class analysis latent variable(s) and the grouping variable.
This is usually referred to as the assumption of
local independence (Lazarsfeld & Henry, 1968).
Definition
Three Parameterizations
Simultaneous latent-class analysis across groups SLCAG is used to study measurement equiva-
(SLCAG) is an extension of the standard latent- lence, that is, to determine to which extent
class (LC) model (▶ Latent Class Model) for the the item responses are similar across groups
examination of measurement equivalence/invari- conditional on the individuals’ latent-class
ance (▶ Measurement Invariance). It can be used memberships. For this purpose, one can use
to compare the latent structure derived from a set one of the three types of parameterizations
of discrete item responses between multiple of the multigroup LC model, that is, a parameter-
groups, for example, between males and females; ization in terms of conditional probabilities,
between Japanese, Americans, and Dutch; log-linear parameters, or linear-logistic item
between young and old; and between ill and parameters (Kankaraš, Moors, & Vermunt,
healthy. Whereas the more commonly used 2010). The probabilistic parameterization can be
multigroup confirmatory factor analysis used when both the indicators and the latent
(MCFA) (▶ Factorial Invariance) assumes that variable(s) are nominal variables. Measurement
the underlying latent variables are continuous, equivalence is studied by equating the
SLCAG treats the latent variables (▶ Latent Var- class-specific item response probabilities across
iables) either as nominal, for example, to identify groups (Clogg & Goodman, 1985; Hagenaars &
a typological classification from a given set of McCutcheon, 2002).
categorical indicators, or as ordinal, for example, In the log-linear formulation (Haberman, S
to investigate the scalability of a set of categorical 1979), the item response probabilities are param-
indicators. These two specifications are some- eterized using a log-linear model (Log-linear
times referred as LC cluster and LC factor models) with a main effect for the item, an asso-
models, respectively. ciation between latent variable(s) and item, an
association between group and item (called direct
effect), and a three-variable term between item,
Description latent variable(s), and grouping variable (called
interaction effect). The latter two terms capture
LC analysis was first introduced by Lazarsfeld for possible differences between groups. A direct
dichotomous items (Lazarsfeld & Henry, 1968) effect quantifies group differences in item
and formalized and extended to nominal vari- responses that cannot fully be explained by
ables by Goodman (1974). The multigroup exten- group differences in the latent variable(s) – that
sion of the standard LC model was proposed by is, when the group variable influences the
S 5970 Simultaneous Latent-Class Analysis Across Groups

indicator concerned independently of the latent parameters (Kankaraš et al., 2010). More specif-
variable(s). The presence of an interaction effect ically, these parameters are restricted to be pro-
indicates that the association between the item portional to the item category scores.
and the latent variable is modified by the group Similar restrictions can be used to define
membership. The third parameterization is the models in which the latent variables have ordinal
linear-logistic specification (Kankaraš et al., instead of nominal categories. This is achieved by
2010) in which the item response probabilities assigning scores to (which is referred to as locat-
are parameterized in terms of item intercepts ing) the latent classes. Such LC models with
and latent variable regression slopes. These inter- discrete-ordinal latent variables are sometimes
cepts and slopes may vary across groups, which is called LC factor models since they resemble lin-
similar to the formulation used in multigroup ear factor analysis (▶ Factor Analysis) (Vermunt
factor analysis. The meaning of unequal slopes & Magidson, 2005). In most aspects, multigroup
is the same as the presence of log-linear interac- LC factor analysis is equivalent to standard
tion effects, and the interpretation of unequal multigroup LC analysis, with the main difference
intercepts is the same as the presence of log- being that instead of comparing typologies, it
linear direct effects. compares latent dimensions of observed discrete
In an unrestricted SLCAG, in which all param- variables across groups.
eters vary freely across groups, the three param- Although LC factor models are typically used
eterizations of the multigroup LC model are in combination with ordinal indicators,
essentially equivalent. This does, however, not multigroup LC factor models with nominally
apply to restricted SLCAG because the three defined indicators can be very useful in cross-
parameterizations allow for slightly different cultural research as it allows for simultaneous
types of model restrictions which have important analysis of measurement equivalence and various
implications when testing for measurement types of response styles biases (Response Style
equivalence. First, in the probabilistic parameter- Bias) that may affect survey responses (Moors,
ization, equivalence is studied by restricting 2004).
probabilities to be group invariant, in the log-
linear parameterization by eliminating interac- Analysis of Measurement Equivalence
tion effects and direct effects, and in logistic In an unrestricted SLCAG, all model parameters
formulation by restricting intercepts and slopes are allowed to differ across groups. This may
to be invariant across groups. Second, the latter yield completely different definitions of the latent
two parameterizations are needed to formulate classes in the investigated groups, making it dif-
models in which indicators and/or latent variables ficult to perform comparisons between the
are treated as ordinal variables. groups. However, this is not the type of model
a researcher is typically aiming at since he or she
Multigroup LC Models for Ordinal Indicators wants to be able to compare the latent classes
and with Ordinal Latent Variables across groups, that is, across medical centers,
When using either the log-linear or the logistic age groups, regions, and different types of
parameterization of the (multigroup) LC model, patients. To determine whether this comparison
it is possible to define restricted models for ordi- is possible, the researcher has to check whether
nal indicators. This is important because in many latent classes have the same meaning in all
areas of social sciences, including quality of life groups, that is, whether measurement equiva-
research (▶ Quality of Life Research), question- lence can be established. In the context of LC
naire items are often of an ordinal nature (e.g., analysis measurement, equivalence is established
rating scales). The constrained LC models for when the class-specific conditional response
ordinal items impose linear restrictions on the probabilities are equal across groups. This
log-linear indicator-latent variable association implies that it is necessary to impose across-
parameters or on the latent variable slope group equality restrictions on these conditional
Simultaneous Latent-Class Analysis Across Groups 5971 S
X Y X Y X Y

G G G

Simultaneous Latent-Class Analysis Across Groups, multigroup LC models. (a) Heterogeneity (inequivalence),
Fig. 1 Relationships between latent variable (X), indicator (b) partial homogeneity, (c) structural homogeneity
variables (Y), and group variable (G) in three different

probabilities in order to test for measurement restricted to be equal across groups, the model is
equivalence. As is shown below, using an called partially homogeneous (Clogg & Good-
SLCAG approach, various levels of homogeneity man, 1985).
(i.e., measurement equivalence) can be tested, Among the various possible partially homoge-
each of which involves restricting specific sets neous models, the one presented in Fig. 1b with
of model parameters to be equal across groups. no “group–latent variable” interaction terms is
The ideal situation for an applied researcher especially important. This model still allows for
who wishes to compare groups occurs when all “direct effects” or group-specific intercept
measurement model parameters can be set equal parameters, which means that the values of the
across groups. From this perspective, the objec- conditional response probabilities (i.e., their “dif-
tive of researching measurement equivalence is ficulties”) are different across populations. How-
to find the model with the highest level of equiv- ever, as there are no group–latent variable
alence possible that fits the data well. The model “interaction effects” in the model (as slope
selection procedure usually starts by determining parameters are assumed to be equal across
the required number of latent classes or discrete groups), relationships between the latent variable
latent factors for each group. If the number of and the responses are the same across groups,
classes is the same across groups, then the het- which make it possible to compare group differ-
erogeneous model is fitted to the data, followed ences in latent-class membership (McCutcheon
by a series of nested, increasingly restricted & Hagenaars, 1997). This is conceptually similar
models which are evaluated in terms of model with the “metric equivalence” model in MCFA in
fit (Eid, Langeheine, & Diener, 2003; Hagenaars, which factor loadings are equal across groups,
1990). Graphical representations of the three pro- but item intercepts may be unequal. Likewise, it
totypical models that differ in the assumed level resembles the situation of “uniform” differential S
of measurement equivalence are provided in item functioning (DIF) (▶ Differential Item
Fig. 1. Functioning (DIF)) in item response theory
The heterogeneous, unrestricted multigroup (IRT) (▶ Item Response Theory [IRT]) frame-
LC model, graphically presented in Fig. 1a, work. It should be noted that the partially homo-
depicts the situation of complete lack of compa- geneous model presented in Fig. 1b can only be
rability of results across groups as all measure- specified using the log-linear and logistic param-
ment model parameters are group eterizations – distinguishing direct and interac-
specific. Comparability is only established if we tion effects or intercepts and slope parameters,
can impose across-group restrictions on the respectively – and, thus, not with the probabilistic
model parameters without deteriorating the fit of parameterization. The partially homogeneous
the model with the data. Imposing restrictions model can be tested against the unrestricted het-
create various nested homogeneous models. If erogeneous model. If the difference in fit between
some, but not all, of the model parameters are the two models is not significant, a researcher can
S 5972 Simultaneous Latent-Class Analysis Across Groups

conclude that interaction effects are not needed in direct effects) at the item level, we need to assume
the model and can proceed with the next step in equivalence in the slope parameters. Therefore,
the analysis. testing equivalence of the intercept parameters of
In comparative social research, researchers are item Y is based on the comparison of the partially
typically interested in establishing full compara- homogeneous model with equal slope parameters
bility of the measurement across groups – that is, for all items (Fig. 1b) with the model which in
they want to attain complete measurement equiv- addition assumes equal intercept parameters for
alence. In order to do so in the context of LC item Y. This procedure is very similar to the one
analysis, it is necessary to establish structural used in MCFA where it is referred to as “partial
equivalence (Hagenaars & McCutcheon, 2002). equivalence.” It should be noted that SLCAG
In a structurally equivalent (homogeneous) differs from MCFA in that it does not require the
model (Fig. 1c), both direct and interaction use of an invariant anchor item for identification
effects are excluded from the log-linear model purposes.
(set to zero), or in the alternative logistic formu- Analysis of measurement equivalence with
lation, both intercept and slope parameters are set LC analysis is not by definition restricted to com-
to be equal across groups. This means that the paring the three models drawn in Fig. 1. Various
conditional item response probabilities are combinations of within- and across-group restric-
restricted to be equal across groups, making the tions and different parameterizations are possi-
item responses independent of the group variable, ble. For instance, it is possible to impose
when controlled for the latent-class membership. equalities constraints across items within groups
Structural equivalence is established if this model or to define models with different numbers of
does not fit the data significantly worse than the latent classes across groups while still assuming
partially homogenous and heterogeneous models. measurement equivalence for particular classes
The homogeneous model is comparable with the (Kankaraš et al., 2010).
“scalar equivalent” model in MCFA that defines
both factor loadings and item intercepts to be the Parameter Estimation and Assessment of
same across groups. It is also similar to an IRT Model Fit
model with both item “difficulty” and “discrimi- LC models are usually estimated by means of
nation” parameters invariant across groups. maximum likelihood (ML) under the assumption
The SLCAG procedure explained so far con- of a multinomial distribution for the indicator
cerns an analysis at the scale level; that is, it uses variables used in model. Maximization of the
models in which a particular restriction is log-likelihood function is typically performed
imposed or relaxed for all indicators simulta- using an expectation–maximization (EM) algo-
neously. However, SLCAG for testing measure- rithm, a Newton–Raphson algorithm, or
ment equivalence can also be conducted at the a combination of these two.
item level. This is particularly relevant when Various model fit statistics are available for
the scale level analysis indicates inequivalence the evaluation in a multigroup LC model. The
in the form of either interactions or direct effects. likelihood-ratio and the Pearson’s chi-square
In such a case, the analysis continues with item goodness-of-fit statistics (Chi-Squared Statistic)
level comparisons in order to check which of the are used as standard measures of discrepancy
items in a scale are the source of inequivalence. between the observed frequencies and the esti-
More specifically, equivalence in the slope mated frequencies according to the model. How-
parameter (presence of interaction effect) for ever, these chi-square tests have a number of
a particular item Y is assessed by comparing the important limitations, the major ones being their
unrestricted, heterogeneous model (Fig. 1a) with limited use when dealing with sparse tables and
a model in which this parameter is equated across their oversensitivity for misfit with large samples.
groups only for this item Y. In order to test for In addition, the goodness-of-fit statistics do not
equivalence in intercept parameters (presence of provide enough control for the number of
Simultaneous Latent-Class Analysis Across Groups 5973 S
parameters in a model, which can sometimes be equivalence in any situation in which the indica-
very large even for models of modest size tors are discrete variables.
(Hagenaars & McCutcheon, 2002).
These limitations prompted the recent devel-
opment and use of several information criteria, Cross-References
such as the Akaike information criterion (AIC),
Bayesian information criterion (BIC), modified ▶ Differential Item Functioning (DIF)
AIC (AIC3), and consistent AIC (CAIC), each of ▶ Factor Analysis
which is designed to penalize models with larger ▶ Factorial Invariance
numbers of parameters. Since increasing the ▶ Item Response Theory [IRT]
number of parameters of a model will increase ▶ Latent Class Model
the likelihood, the information criteria penalize ▶ Latent Variables
the likelihood by a certain amount that is ▶ Measurement Invariance
a function of the number of estimated parameters. ▶ Quality of Life Questionnaire
Thus, a model with a lower value for the infor- ▶ Quality of Life Research
mation criteria yields a better fit to a data, for
a given number of parameters. Since they also
control for sample size, BIC and CAIC are pre- References
ferred fit statistics in situations when sample size
is large. For small to medium sample sizes, the Clogg, C. C., & Goodman, L. A. (1985). Simultaneous
AIC statistic is most commonly used. latent structural analysis in several groups. In N. B.
Tuma (Ed.), Sociological methodology (pp. 81–110).
Software packages that can be used to obtain
San Francisco: Jossey-Bass Publishers.
ML estimates and model fit statistics for Eid, M., Langeheine, R., & Diener, E. (2003). Comparing
multigroup LC models are LEM (Vermunt, typological structures across cultures by multigroup
1997), Latent GOLD (Vermunt & Magidson, latent class analysis. A primer. Journal of Cross-Cul-
tural Psychology, 34, 195–210.
2005), MPlus (Muthén & Muthén, 2006), and
Goodman, L. A. (1974). Exploratory latent structure anal-
GLLAMM (Rabe-Hesketh, Skrondal, & Pickles, ysis using both identifiable and unidentifiable models.
2004). Biometrika, 61, 215–231.
Haberman, S. J. (1979). Analysis of qualitative data:
Vol. 2 New Developments. New York: Academic.
Discussion
Hagenaars, J. A. (1990). Categorical longitudinal data –
LC analysis is an obvious choice when Loglinear analysis of panel, trend and cohort data.
a researcher wishes to compare typological struc- Newbury Park, CA: Sage.
tures across groups – that is, when investigating Hagenaars, J. A., & McCutcheon, A. L. (2002). Applied
latent class analysis. Cambridge, UK: Cambridge Uni-
whether groups differ with respect to the preva-
versity Press.
lence of latent classes, while taking into account Kankaraš, M., Moors, G., & Vermunt, J. K. (2010). Test- S
issues of measurement equivalence. With the ing for measurement invariance with latent class anal-
possibility to define the latent variable(s) as dis- ysis. In E. Davidov, P. Schmidt, & J. Billiet (Eds.),
Cross-cultural analysis: Methods and applications
crete ordinal, it is shown that the LC approach
(pp. 359–384). New York: Routledge.
can also be used for cross-cultural comparisons of Lazarsfeld, P. F., & Henry, N. W. (1968). Latent structure
dimensional structures, thus, presenting an alter- analysis. Boston: Houghton Mifflin.
native to the more frequently used MCFA and McCutcheon, A. L., & Hagenaars, J. A. (1997). Compar-
ative social research with multi-sample latent class
IRT approaches (Kankaraš et al., 2010). This is models. In J. Rost & R. Langeheine (Eds.), Applica-
especially true in those situations when some of tions of latent trait and latent class models in the social
the model assumptions of MCFA and IRT do not sciences (pp. 266–277). Münster, Germany: Waxmann.
hold. With its flexible set of tools, combined with Moors, G. (2004). Facts and artefacts in the comparison of
attitudes among ethnic minorities. A multi-group
recent developments in software for multigroup
latent class structure model with adjustment for
LC modeling, the presented approach is a very response style behaviour. European Sociological
attractive option for studying measurement Review, 20, 303–320.
S 5974 SINET

Muthén, L. K., & Muthén, B. O. (2006). Mplus user’s The objective of the Newsletter was to convey
guide (4th ed.). Los Angeles: Muthén and Muthén. information on the activities of the Center for
Rabe-Hesketh, S., Skrondal, A., & Pickles, A. (2004).
GLLAMM Manual. U.C. Berkeley Division of Biosta- Coordination of Research on Social Indicators
tistics Working Paper Series (Working Paper 160). as well as of various activities and organizations
Vermunt, J. K. (1997). LEM 1.0: A general program for involved in the development of social indicators,
the analysis of categorical data. Tilburg, The Nether- including the development of new datasets and
lands: Tilburg University.
Vermunt, J. K., & Magidson, J. (2005). Factor analysis data series, new conceptual frameworks, new
with categorical indicators: A comparison between indicators, and new organizations and groups
traditional and latent class approaches. In A. Van der involved in such activities.
Ark, M. A. Croon, & K. Sijtsma (Eds.), New develop- When the Center closed its doors in 1984,
ments in categorical data analysis for the social and
behavioral sciences (pp. 41–62). Mahwah, NJ: Abbott L. Ferriss of Emory University
Erlbaum. volunteered to carry on the collation, review,
and distribution of news about social indicators
functions of the Newsletter by publishing the
SINET: Social Indicators Network News. Ferriss
SINET edited the quarterly publication of SINET through
the November 1995 issue (number 44). During
Kenneth C. Land his term, SINET became a key source of news
Department of Sociology, Duke University, about social indicators developments. Profes-
Durham, NC, USA sional recognition followed at its 1992 meeting,
the Working Group on Social Indicators and
Quality of Life Measurement of the International
Definition Sociological Association voted to have SINET
publish news of the Working Group commencing
SINET: Social Indicators Network News is with the November 1992 issue. Ferriss also con-
a quarterly review of social reports and research tinued and expanded the review function of
on ▶ social indicators, ▶ social trends, and the SINET in the form of periodic reviews of new
▶ quality of life that also contains news items on social indicators publications and books.
these topics. SINET serves as the Newsletter of Kenneth C. Land of Duke University became
the ▶ International Society for Quality of Life Editor with the Winter (February) 1996 issue
Studies (ISQOLS) and also includes news from number 45 and has served in that capacity
Research Committee 55 on Social Indicators and through the November 2012 issue number 112
Social Reporting of the International Sociologi- and beyond. In his initial issue as Editor
cal Association. It is published three times a year (February 1996), Land (1996) published an
(February, May–August combined issue, and essay entitled “Social Indicators and the Quality
November). of Life: Where Do We Stand in the Mid-1990s?”
In this article, Land reviewed the history and
contemporary status of interest in social indica-
Description tors, social monitoring, and social reporting and
described (1) the classical heritage from the
History social indicators movement of the 1960s and
The United States (US) Social Science Research 1970s, (2) the revitalization of interest that
Council opened a Center for Coordination of emerged beginning in the mid-1980s, and
Research on Social Indicators in 1972 in (3) the emergence of quality of life/well-being
Washington, District of Columbia (DC). The as a unifying concept in political, popular, and
Center began circulating its Social Indicators theoretical discourses of these topics in the
Newsletter in March 1973 and ended publication 1990s. The lessons learned, and guidelines
with issue number 18 in September 1983. articulated, from the history and state of