Sequence of nested models to test for measurement invariance of categorical items

118 views
Skip to first unread message

Timothy Wong

unread,
Mar 29, 2023, 9:59:04 PM3/29/23
to lavaan
Dear lavaaners,

I have a dataset consisting of 30 ordered categorical indicators (2-5 response categories) measured at two timepoints (60 items in total). EFA conducted on each time point separately suggests a solution of 3 factors (7-11 items loading on to each factor).

I now want to test the 3-factor structure for measurement invariance using Wu and Estabrook's (2016) method for categorical indicators. I have read their article and the measeq.syntax example syntax.

From my understanding, I would first set up the configural model:
config.syntax <- measEq.syntax(configural.model = config.mod,
                               data = complete(MI2, action = 1),
                               parameterization = 'theta',
                               ID.fac = 'std.lv',
                               ID.cat = 'Wu.Estabrook.2016',
                               longFacNames = list(
                                 slp = c('slp1', 'slp2'),
                                 dep = c('dep1', 'dep2'),
                                 man = c('man1', 'man2')
                               )) %>%
  as.character()

Then, as per Wu and Estabrook, I would set thresholds equal across timepoints before testing other forms of invariance:
thresh.syntax <- measEq.syntax(configural.model = config.mod,
                               data = complete(MI2, action = 1),
                               parameterization = 'theta',
                               ID.fac = 'std.lv',
                               ID.cat = 'Wu.Estabrook.2016',
                               longFacNames = list(                                                slp = c('slp1', 'slp2'),
                                 dep = c('dep1', 'dep2'),
                                 man = c('man1', 'man2'),
                               long.equal = c("thresholds")
                               )) %>%
  as.character()

However, it is unclear to me which model(s) to test next, among the following:
  • long.equal = c("thresholds", "loadings")
  • long.equal = c("thresholds", "intercepts")
  • long.equal = c("thresholds", "loadings", "intercepts")
  • long.equal =  c("thresholds", "loadings", "intercepts", "residuals")
I have avoided using names (weak, metric, scalar, strong, strict etc) as these seem more relevant for continuous indicators.

Any guidance would be much appreciated!

Terrence Jorgensen

unread,
Mar 30, 2023, 6:08:44 AM3/30/23
to lavaan
it is unclear to me which model(s) to test next, among the following:
  • long.equal = c("thresholds", "intercepts")
 Intercepts are not really comparable when slopes differ (think of homogeneous slopes with ANCOVA).
Once you equate latent-response scales with equal thresholds, then the usual (metric, scalar, strict) sequence follows from there:
    • long.equal = c("thresholds", "loadings")
    • long.equal = c("thresholds", "loadings", "intercepts")
    • long.equal =  c("thresholds", "loadings", "intercepts", "residuals")

      Terrence D. Jorgensen    (he, him, his)
      Assistant Professor, Methods and Statistics
      Research Institute for Child Development and Education, the University of Amsterdam
      http://www.uva.nl/profile/t.d.jorgensen


      Timothy Wong

      unread,
      Mar 30, 2023, 10:48:58 PM3/30/23
      to lavaan
      Thanks for that Terrence, your help is always much appreciated.

      A small follow-up question, so I can better understand the syntax. Say I generate measEq.syntax for two levels of measurement invariance, e.g.
      1. long.equal = c("thresholds") vs.
      1. long.equal = c("thresholds", "loadings")
        As expected, the first model constrains only the thresholds, while the second model constrains thresholds and item loadings.

        However, what I find interesting is that syntax 2 also releases some constraints which are present in syntax 1. Specifically, syntax 2 allows all common factor variances at the second occasion to be freely estimated.

        Comparing a subset of the syntax (highlighted are the lines of syntax which differ between the two):

        1.   ## COMMON-FACTOR VARIANCES:

          slp1 ~~ 1*slp1 + psi.1_1*slp1

          slp2 ~~ 1*slp2 + psi.2_2*slp2

          dep1 ~~ 1*dep1 + psi.3_3*dep1

          dep2 ~~ 1*dep2 + psi.4_4*dep2

          man1 ~~ 1*man1 + psi.5_5*man1

          man2 ~~ 1*man2 + psi.6_6*man2

        2.   ## COMMON-FACTOR VARIANCES:

          slp1 ~~ 1*slp1 + psi.1_1*slp1

          slp2 ~~ NA*slp2 + psi.2_2*slp2

          dep1 ~~ 1*dep1 + psi.3_3*dep1

          dep2 ~~ NA*dep2 + psi.4_4*dep2

          man1 ~~ 1*man1 + psi.5_5*man1

          man2 ~~ NA*man2 + psi.6_6*man2
        My understanding of measurement invariance (which is admittedly, quite elementary) is that when testing increasingly strict levels of invariance, constraints are added and not released.

        So why is it the case that when going from 1 > 2, certain constraints are released?
        N.B. I have also noticed this behaviour when comparing long.equal = c("thresholds", "loadings") with  long.equal = c("thresholds", "loadings", "intercepts")syntax.

        Thanks

        Terrence Jorgensen

        unread,
        Apr 3, 2023, 5:38:32 AM4/3/23
        to lavaan
        Wu and Estabrook's (2016) introduction discusses the distinction between identification constraints and hypothesized equation constraints.  When you leave variances fixed (an identification constraint in a configural model) after imposing hypothesized equality of loadings, then you are not testing metric invariance.  You are instead simultaneously testing equality of loadings AND factor variances.  Likewise with factor means when you constrain intercepts: the fixed means would no longer be identification constraints when they are not needed for identification.  There are also several threads on this forum you could search for my discussions of invariance in general and Wu and Estabrook (2016) specifically.  Notice also that relative to the configural model, equating thresholds also releases some constraints that are no longer needed for identification (check latent-response intercepts and residual variances per indicator).

        Timothy Wong

        unread,
        Apr 4, 2023, 1:33:17 AM4/4/23
        to lavaan
        Thanks for the detailed explanation and pointers Terrence
        Reply all
        Reply to author
        Forward
        0 new messages