Newsom

Psy 523/623 Structural Equation Modeling, Spring 2023 1

SEM with Categorical Variables

Definitions and Distinctions

Categorical variables are those with two values (i.e., binary, dichotomous) or those with a few ordered
categories (typically less than five) require special estimation considerations in structural equation modeling.
Examples might include gender, dead vs. alive, audited vs. not audited, or variables with few response
options like “never,” “sometimes,” or “always.” Social scientists generally treat variables measured on a ratio
or interval scale, such as temperature, height, or income in dollars, or ordinal variables with 5 or more
categories as continuous (see the handout “SEM with Nonnormal Continuous Variables” and the handout
from my Univariate Statistics course called “Levels of Measurement and Choosing the Correct Statistical
Test” for references and details). Similar to regression analysis, only measured variables which are not
predicted by other variables in the model (including not predicted by latent variables) do not require special
estimation for categorical variables.

When variables predicted by other variables (an endogenous variable in a model or an indicator of a latent
variable) are measured on an ordinal scale and there are relatively few categories, 2-4 categories,
estimation methods specifically designed for categorical variables are recommended (Finney & DiStefano,
2013). This includes nominal binary variables (e.g., pass/fail, divorced, yes/no, heart attack vs. no heart
attack). For ordinal variables with several values, a categorical analysis approach will have the greatest
advantage (less bias) compared with standard ML when the following conditions hold (Rhemtulla,
Brosseau-Liard, & Savalei, 2012): (1) when the values between categories are not equidistant; (2) when the
relationship between the categorical measured variable and the theoretical variable it is supposed to
measure is not a linear relationship—not entirely unrelated to (1); and (3) when the ordinal variable is
skewed or kurtotic.

Estimators for Binary and Ordinal Variables

WLSMV. There seems to be growing consensus among researchers that the best approach to analysis of
binary and ordinal variables (with few categories) is what is referred most to most commonly as diagonally
weighted least squares (DWLS) approach (Muthén, du Toit, & Spisic, 1997). This estimation method is
called weighted least squares mean and variance adjusted (WLSMV) in Mplus and the R package lavaan
(it is invoked by estimator = WLSMV). The WLSMV approach seems to work well if sample size is 200 or
better (Bandalos, 2014; Flora & Curran, 2004; Muthén et al., 1997; Rhemtulla, Brosseau-Liard, & Savalei,
2012). 1 In Lisrel and EQS, a similar approach that uses WLS together with polychoric correlations and
asymptotic covariance matrices is used.

WLSMV works by first creating a matrix of polychoric correlations. Polychoric correlations estimate the what
the association between two continuous, normally distributed variables would be if they been converted to
dichotomous or ordinal observed variables. 2 Polychoric correlations, which Finney and DiStefano (2013)
refer to as “latent correlations” represent the correlation between the unobserved underlying continuous
variables frequently called y* (“y star”; some authors use eta, but that potentially confuses it with what we
use for multiple indicator latent variables). In the estimation process, the polychoric correlations are used to
create an asymptotic covariance matrix that serves as the weight matrix for the WLS estimation. The
resulting path estimates represent the change in y* on a standardized z scale for each unit change in the
predictor. These are the same as what is obtained with probit regression. The distribution of y* requires an
arbitrary scaling constraint, much like a latent variable. In Mplus, there are two versions of WLSMV
estimates that have different approaches to setting the scaling of the y* distribution, delta parameterization
and theta parameterization (Muthén & Asparouhov, 2002). Delta parameterization is the default and sets
the scaling by setting the measurement residual to 1.0 and theta parameterization sets the scaling of the y*

1
There is another option for weighted least squares, called WLSM in Mplus and lavaan. Both WLSM and WLSMV adjust the standard errors in the
same way, but the WLSM method only does a mean adjustment when correcting the chi-square. Because the WLSMV method seems to work better
(Asparouhov & Muthen, 2010a; Muthen et al., 1997), I do not recommend using the WLSM option.
2
There are three related definintsion for this type of latent correlation, tetrachoric, for binary variables, polyserial for binary and continuous, and
polychoric for ordinal variables. The term polychoric correlation is often used as a general way of referring to any of these three related correlations.
Newsom
Psy 523/623 Structural Equation Modeling, Spring 2023 2

variance to 1.0, estimating the measurement residual variance. Both can be called variants on the probit
model, but theta parameterization corresponds more exactly to the probit regression estimates in which the
y* distribution is assumed to be standardized. These scaling choices are arbitrary in the sense that the chi-
square for the model and the significance tests of the parameter estimates will be equal. WLSMV works well
in many situations, although because WLSMV is not a full information method it has stricter assumptions for
missing data (Asparouhov & Muthén, 2010).

MLR (robust marginal maximum likelihood). Marginal maximum likelihood estimation (or sometimes just “full
maximum likelihood”) is a special maximum likelihood approach for binary and ordinal variables (Bock &
Atkin, 1981; Christoffersson, 1975; Muthén & Christoffersson, 1981). This method, which is less commonly
employed with SEM models than the WLSMV method, uses the frequency tables in the analysis so can be
distinguished from the ML estimation process used for continuous variables. Commonly used in Item
Response Theory (IRT) measurement analysis, the default categorical ML estimation yields logistic
parameter estimates that can be converted to odds ratios. 3 This approach is not available in many SEM
software programs, but Mplus uses the marginal maximum likelihood estimation approach this when
ESTIMATOR=ML is used in conjunction with dependent variables identified as categorical on the
CATEGORICAL statement. A robust version of ML for categorical variables, which uses robust standard
error estimates, is called MLR in Mplus. The R package lavaan currently has limited categorical ML
estimation capabilities. An important drawback to the MLR estimation is that the traditional chi-square nor
most of the usual alternative fit indices are available to judge fit. Instead Mplus prints a Pearson chi-square
and likelihood ratio chi-square based on the contingency tables for the model implied and obtained
frequencies. WLSMV probit and ML logistic estimates will often be quite similar in terms of their statistical
conclusions. Work by Bandalos (2014) indicates that robust MLR performs better than the unadjusted ML
and that MLR performed similarly to the WLSMV method. Compared with WLSMV, MLR has somewhat less
power but better control of Type I error in smaller samples. Bandalos's work also suggests that sample
sizes of 150 may be too small with either method, especially where distributions of the categorical variables
are asymmetric. Newsom and Smith (2020) also found that the two estimators performed similarly but that
with MLR had important convergence advantages and that WLSMV had some advantages for the small
sample sizes in the context of binary growth curve models.

Bayesian. Amos does not use either of the above estimation approaches for categorical variables, but the
most recent editions allow a Bayesian approach (Lee, 2007). The Bayesian approach, also available in
Mplus and the R package blavaan, requires an iterative process known as the Markov Chain Monte Carlo
(MCMC). To date, there is less information on the performance of this approach with SEM with respect to fit
estimation, the optimal algorithms to use, and standard errors under various conditions (cf. Lee & Yang,
2006). The Bayesian estimation process depends upon the distributional priors used and some artful
judgment in the testing process. The Bayesian structural modeling approach has not become a popular
alternative thus far but may increase in popularity at least for certain circumstances in the future (see
Kaplan & Depaoli, 2012 for an introduction; see also Depaoli, 2021; and the handout from this class
“Alternative Estimators”).

Also see the handout from this class “Alternative Estimation Methods” for more details on categorical
estimation.

Fit Indices
Less is known about how fit indices perform with WLSMV and categorical MLR under various
circumstances—certainly not with the same level of precision on which Hu and Bentler based their
recommendations about fit with continuous variables. The WLSMV chi-square used by Mplus (see
Asparouhov & Muthén, 2010a) seems to perform pretty well, even for sample sizes as small as 100 (Flora &
Curran, 2004), although there is still likely to be a practical problem with using chi-square as a sole measure
of fit because of its sensitivity to sample size. Although some work has supported use of RMSEA, TLI, and

3
Probit parameter estimates can also be requested with link = probit.
Newsom
Psy 523/623 Structural Equation Modeling, Spring 2023 3

CFI with categorical model estimation (WLSMV; Beauducel & Herzberg, 2006; Hutchinson & Olmos, 1998;
Yu & Muthén, 2002), Savalei (2021) provides evidence that fit is overestimated with these indices and
suggests some computational adjustments (using continuous ML with polychoric correlations) that appear to
work well.

Until the most recent version (Version 8), Mplus reported the Weighted Root Mean Square Residual
(WRMR) for fit of models with categorical observed variables (Yu & Muthén, 2002 recommended WRMR of
less than 1.0 as indicative of good fit). The most recent versions of the program no longer use WRMR
because of data suggesting poor performance with larger samples and larger models (DeStefano, Liu,
Jiang, & Shi, 2018) and use a modified computation of the SRMR instead (Asparouhov & Muthén, 2018),
where the usual cutoff of < .08 is suggested to indicate good approximate fit. Recent evidence (Savalei,
2021; Shi, Maydeu-Olivares, & Rosseel, 2020) suggests that SRMR is not subject to the overfit bias of the
CFI and RMSEA.

Nested Tests
Nested tests (likelihood ratio test) require special attention for robust estimation, including WLSMV (i.e.,
estimator = WLSMV in Mplus and lavaan). The scaling correction factor (scf) must be used to weight the
difference (Satorra, 2000; Satorra & Bentler, 2001). Asparouhov and Muthén (2006; 2018b) have adapted
the tests developed by Satorra (2000) and Satorra and Bentler (2001) that computes the estimated ratio of
the weighted likelihoods of two models using WLSMV estimation for ordinal variables. Mplus provides
automated nested tests with the DIFFTEST command that can be used for several estimation or robust
methods (Asparouhov & Muthén, 2013, 2018b; Bryant & Satorra, 2012; Satorra & Bentler, 2010). See the
handout “Examples of Chi-square Difference Tests with Nonnormal and Categorical Variables” for an
illustration.

Missing Data Estimation with Non-normal or Categorical Data

WLSMV (method = WLSMV) estimation uses a partially pairwise deletion estimation and is not a full
information estimation approach, and therefore it may not work as well with missing data as categorical
MLR unless there is only one dependent variable and values are least MAR (Asparouhov & Muthén, 2006;
Asparouhov & Muthén, 2010b).

Other Types of Categorical Variables

In addition to binary and ordinal variables, Mplus also has estimation approaches for count variables,
including Poisson, negative binomial, zero-inflated Poisson and negative binomial, nominal (multnomial
logistic regression), and continuous survival analysis. Few other SEM packages have features for these
types of variables and, although the extension from regression is relatively straightforward, there has been
little simulation work thus far to examine how these algorithms perform with model fit and standard error
estimation in the SEM context.

References
Asparouhov, T., & Muthén, B. (2010a). Simple second order chi-square correction. Unpublished manuscript. https://www.statmodel.com/download/WLSMV_new_chi21.pdf
Asparouhov, T., & Muthén, B. (2010b). Weighted least squares estimation with missing data. Unplublished technical report, retrieved from
https://www.statmodel.com/download/GstrucMissingRevision.pdf
Asparouhov, T. and Muthen, B. (2013). Computing the strictly positive Satorra-Bentler chi-square test in Mplus. Unpublished manuscript. https://www.statmodel.com/examples/webnotes/SB5.pdf
Asparouhov, T., & Muthén, B. (2018a). SRMR in Mplus. Unpublished manuscript. https://www.statmodel.com/download/SRMR2.pdf
Asparouhov, T., & Muthén, B. (2018a) Nesting and equivalence testing in Mplus. https://www.statmodel.com/chidiff.shtml
Babakus, E., Fergnson, C. E. & Joreskog, K G, (1987). The sensitivity of confirmatory maximum likelihood factor analysis to violations of measurement scale and distributional assumptions. Journal
of Marketing Research, 24, 2228.
Bentler, P.M., & Wu, E.J.C. (2002). EQS for Windows user’s guide. Encino, CA. Multivariate Software, Inc.
Bock, R. D., & Atkin, M. (1981). Marginal maximum likelihood estimation of item parameters. An application of an EM algorithm. Psychometrika, 46, 443–459.
Bryant, F. B., & Satorra, A. (2012). Principles and practice of scaled difference chi-square testing. Psychometrika, 66, 507–514.
Curran, P. J., West, S. G, & Finch, J. F. (1996). The robustness of test statistics to nonnormality and specification error in confirmatory factor analysis. Psychological Methods, 1, 16-29.
Christoffersson, A. (1975). Factor analysis of dichotomized variables. Psychometrika, 40, 5–32.
DeCarlo, L. T. (1997). On the meaning and use of kurtosis. Psychological Methods, 2, 292-307.
Depaoli, S. (2021). Bayesian structural equation modeling. Guilford.
DiStefano C., Liu J., Jiang N., & Shi D. (2018) Examination of the Weighted Root Mean Square Residual: Evidence for Trustworthiness? Structural Equation Modeling, 25, 453-466.
Dolan, C. V. (1994), Factor analysis of variables with 2, 3, 5, and 7 response categories: A comparison of categorical variable estimators using simulated data. British journal of Mathematical and
Statistical Psychology, 47, 309-326.
Finney, S.J., & DiStefano, C. (2013). Non-normal and categorical data in structural equation modeling. In G.R. Hancock & R.O. Mueller (Eds.), Structural equation modeling: A second course, 2nd
Edition (pp. 439–492). Charlotte, NC: Information Age Publishing.
Fouladi, R.T. (1998, April). Covariance structure analysis techniques under conditions of multivariate normality and non-normality—modified and bootstrap based test statistics. Paper presented at
the annual meeting of the American Educational Research Association, San Diego, CA.
Hancock, G.R., & Nevitt, J. (1999). Bootstrapping and the identification of exogenous latent variables within structural equation models. Structural Equation Modeling, 6, 394-399.
Newsom
Psy 523/623 Structural Equation Modeling, Spring 2023 4

Hutchinson, S, R., & Olmos, A (1998). Behavior of descriptive fit indexes in confirmatory factor analysis using ordered categorical data. Structural Equation Modeling: A Multidisciplinary Journal, 5,
344-364.
Hu, L., & Bentler, P.M. (1999). Cutoff criteria for fit indexes in covariance structure analysis: Conventional criteria versus new alternatives. Structural Equation Modeling, 6(1), 1-55.
Hu, L., Bentler, P.M., & Kano, Y. (1992). Can test statistics in covariance structure analysis be trusted? Psychological Bulletin, 112, 351-362.
Johnson, D.R., & Creech, J.C. (1983) Ordinal measures in multiple indicator models: A simulation study of categorization error. American Sociological Review, 48, 398-407.
Kaplan, D., & Depaoli, S. (2012). In R.H. Hoyle (Ed.). Handbook of structural equation modeling (pp., 650-673). New York: Guilford Press.
Lee, S-Y., & Tang, N.-S. (2006). Bayesian analysis of structural equation models with mixed exponential family and ordered categorical data. British Journal of Mathematical and Statistical
Psychology, 59, 151–172
Lee, S-Y. (2007). Structural Equation Modeling: A Bayesian Approach. New York: Wiley.
Muthén, B., & Christoffersson, A. (1981). Simultaneous factor analysis of dichotomous variables in several groups. Psychometrika, 46, 407–419.
Muthén , B.O, du Toit, S., & Spisic,, D. (1997). Robust inference using weighted least squares and quadratic estimating equations in latent variable modeling with categorical and continuous
outcomes. Unpublished manuscript.
Muthén, B., & Asparouhov, T. (2002). Latent variable analysis with categorical outcomes: Multiple-group and growth modeling in Mplus. http://www.statmodel.comexamples/webnote.shtml#web4.
Nevitt, J., & Hancock, G.R. (2001). Performance of bootstrapping approaches to model test statistics and parameter standard error estimation in structural equation modeling. Structural Equation
Modeling, 8, 353-377.
Newsom, J.T., & Smith, N.A (2020): Performance of Latent Growth Curve Models with Binary Variables. Structural Equation Modeling: A Multidisciplinary Journal, 1-20. DOI:
10.1080/10705511.2019.1705825
Rhemtulla, M., Brosseau-Liard, P. É., & Savalei, V. (2012). When can categorical variables be treated as continuous? A comparison of robust continuous and categorical SEM estimation methods
under suboptimal conditions. Psychological methods, 17, 354-373.
Satorra, A., & Bentler, P.M. (1988). Scaling corrections for chi-square statistics in covariance structure analysis. 1988 Proceedings of the Business and Economic Statistics Section of the American
Statistical Association, 308-313.
Satorra, A., & Bentler, P.M. (1994). Corrections to test statistics and standard errors in covariance structure analysis. In A. von Eye and C.C. Clogg (eds.), Latent Variable Analysis: Applications to
Developmental Research (pp. 399-419). Newbury Park: Sage.
Savalei, V. (2021). Improving fit indices in structural equation modeling with categorical data. Multivariate Behavioral Research, 56(3), 390-407.
Shi, D., Maydeu-Olivares, A., & Rosseel, Y. (2020). Assessing fit in ordinal factor analysis models: SRMR vs. RMSEA. Structural Equation Modeling: A Multidisciplinary Journal, 27(1), 1-15.
Yu, C.-Y., & Muthén, B. (2002). Evaluating cutoff criteria of model fit indices for latent variable models with binary and continuous outcomes. Doctoral dissertation. Retrieved from http://
www.statmodel.com/download/Yudissertation.pdf