https://doi.org/10.1177/1073191117711020
Assessment
1–17
© The Author(s) 2017
Reprints and permissions:
sagepub.com/journalsPermissions.nav
DOI: 10.1177/1073191117711020
journals.sagepub.com/home/asm
Article
Measurement invariance refers to the case that the same
latent construct is measured by the same observed vari-
ables, in a similar way across different measurement condi-
tions. These conditions include subgroups of a population
(e.g., national cultures, ethnic groups, gender, etc.), occa-
sions of data collection (i.e., repeatedly measured data), or
test-taking situations (e.g., paper/pencil vs. web-based tests;
Meade & Wright, 2012). The establishment of measure-
ment invariance ensures that cross-condition differences in
observed variables reflect true changes among latent con-
structs, rather than alterations in the psychometric proper-
ties of the measures. Lately, measurement invariance has
gained recognition as an important prerequisite for making
meaningful comparisons in clinical assessments (Floyd &
Widaman, 1995; Meredith & Teresi, 2006; Pentz & Chou,
1994).
Multiple-group confirmatory factor analysis (CFA;
Jöreskog, 1971; McGaw & Jöreskog, 1971) has widely
been used to test measurement invariance by assessing the
equivalence of factor models across groups, or factorial
invariance (Meredith, 1993). Researchers have defined
many levels of factorial invariance, primarily in terms of
cross-group equality of model parameters such as factor
loadings, intercepts, residual variances, and so on (e.g.,
Byrne, Shavelson, & Muthén, 1989; Horn & McArdle,
1992; Meredith, 1993; Millsap, 2011; Steenkamp &
Baumgartner, 1998; Widaman & Reise, 1997; see also
Vandenberg & Lance, 2000, for a review).When testing for
a certain level of factorial invariance (such as the equiva-
lence of factor loadings), full invariance is said to be
achieved if all factor loadings can be constrained to be
equivalent across groups; otherwise, when only a subset of
those factor loadings is the same across groups, partial
invariance is said to have occurred. As is often the case, full
invariance is difficult to achieve in real-world research. For
example, approximately half of the 75 reviewed empirical
studies in Schmitt and Kuljanin (2008) included provisions
for some level of partial invariance. Due to the frequent
occurrence of partial invariance in empirical research,
methodologists suggest using a model capable of accom-
modating it in subsequent cross-group comparisons (e.g.,
Byrne et al., 1989; Cheung & Rensvold, 2002; Schmitt &
Kuljanin, 2008; Vandenberg & Lance, 2000).
Whether it is appropriate to use measures or tests with
partial invariance is now an important issue in both empiri-
cal and methodological research. Methodological study has
shed some light on this issue. One line of inquiry concerns
the impact of certain noninvariances on cross-group com-
parisons using composite scores in which composites,
711020ASM XX X 10.1177/1073191117711020AssessmentShi et al.
research-article 2017
1
University of Oklahoma, Norman, OK, USA
2
University of South Carolina, Columbia, SC, USA
Corresponding Author:
Dexin Shi, Department of Psychology, University of South Carolina,
1512 Pendleton Street, Barnwell College, Columbia, SC 29208, USA.
Email: shid@mailbox.sc.edu
The Impact of Partial Factorial Invariance
on Cross-Group Comparisons
Dexin Shi
1,2
, Hairong Song
1
, and Melanie D. Lewis
1
Abstract
This study explored the impact of partial factorial invariance on cross-group comparisons of latent variables, including
latent means, latent variances, structural relations (or correlations) with other constructs, and regression coefficients as
predicting external variables. The results indicate that the estimates of factor mean differences are sensitive to violations of
invariance on both factor loadings and intercepts. Noninvariant factor loadings were also found to influence the cross-group
comparisons of factor variances and regression coefficients (slopes, in the raw metric) with external variables. However,
cross-group comparisons of standardized slopes and interfactor correlations were not subject to noninvariance. Under
conditions of partial invariance, we further compared the performance of four different model specification strategies. In
general, fitting partially invariant models with all noninvariant parameters that were freely estimated yielded more accurate
estimates of the parameters of interest. The implications of the major findings of this work, as well as recommendations
and guidelines for future empirical researchers, are discussed below.
Keywords
partial factorial invariance, multiple group comparisons, CFA