Structural equation modeling. Structural
equation modeling (SEM) explicitly models latent variables. However,
factor scores tend to estimate what individuals might score
on a factor. If the goal is to attempt to understand how latent variables
(constructs) relate to each other, we have been able to directly model
this for a while now, and this might be a good alternative for some.
However, SEM has many of the same drawbacks in terms of replication
that we just discussed in relation to multiple regression and EFA.
Most notably, it will tend to overfit the model to the sample, so
if the sample is quirky, small, or contains biases or error, the solution
will not be as generalizable as we would hope. So SEM is not a panacea.
You must have an unbiased, large sample in order to hope for replicability—and
then you should test whether the results replicate or not.
Rasch measurement and item response
theory (IRT). Two examples of modern measurement theory,
Rasch measurement and IRT, seek to understand how patterns of responses
across items of different “difficulty” can help estimate
a person’s score on a construct. Rasch and IRT models have
many similarities, but scholars in both groups will tell you they
are different in important ways. However, one important characteristic
underlying both sets of models is that they can provide estimates
of item characteristics that are independent from the sample in which
they were estimated. In other words, a sample low on the construct
of interest and a sample high on the construct of interest can yield
nearly identical item level results. These item parameters can then
be used to compute a consistent score for a construct of interest.
We leave it to you to pursue this topic further, as there are great
books on both. For example, we find Bond & Fox (2006) an accessible
and helpful book on Rasch measurement, and Baker (1985) and Embretson
& Reise (2000) to be accessible and informative books about IRT.