"Measurement Invariance in Longitudinal Studies of Trauma and Psychological Distress"
Demonstrating the integrity of a
construct is recognized as a prerequisite to the quantification of change in
longitudinal studies. Methods for evaluating this measurement invariance are
available in the context of structural equation models. These methods allow
some flexibility when some indicators of the construct differ across time. Shifting
indicators may be common in trauma research, where disruptions are typical concomitants
of a traumatic event such as community violence or a natural disaster. We illustrate
a two-step approach to assessing measurement invariance prior to evaluating
change that requires a single scale to remain constant across time. At the first
step, the common scale is evaluated for invariance at the item level. At the
second step, the common scale is used to set the metric and the construct’s
factorial invariance is assessed. Using data from a study of 151 Kuwaiti children
exposed to the Gulf war, assessed in 1993 and again in 2003, we illustrate the
approach with a measure of posttraumatic stress symptoms using Davidson’s
self-rating scale as the common scale over time. Fourteen out of 17 items were
shown to have invariant loadings on a 3-factor confirmatory analysis, although
items tended to have greater reliability when participants were older. Scales
to measure depression and anxiety were also administered at both times, but
different scales were used at each time. In 1993, depression was assessed with
the Children’s Depression Inventory and anxiety was assessed with the
Revised Children’s Manifest Anxiety Scale. In 2003, depression was assessed
with the Beck Depression Inventory and anxiety was assessed with the Spielberger’s
Trait Anxiety scale. Using the three latent variables from step 1 and the two
additional measures of depression and anxiety, a second-order factor of psychological
distress was specified at each time. Invariance in the loadings of the common
scales was established, allowing the loadings of the different scales to be
estimated freely at each time. The means of the second-order latent variable
may now be compared over time. The presentation of the two-step approach will
be embedded in a broader presentation that will cover an introduction to measurement
invariance, how invariance may be assessed in the context of structural equation
models, latent models for dealing with shifting indicators over time, and other
methodological issues associated with measurement invariance in the context
of a longitudinal study.