In this paper, various aspects of CAT were introduced and reviewed. It was argued that
even though CAT has some important advantages, the cost of development and maintenance
are high. Much higher in general, then the costs of linear testing. To reduce the costs, Bayesian
CAT was introduced. In Bayesian CAT, prior beliefs and observed data are combined to estimate
both item and person parameters. It was demonstrated that, both in the item parameter
estimation and in the person parameter estimation phase, considerable gains can be made
by eliciting empirical priors for both the person and the item parameters, and implementing
them in CAT. Bayesian CAT might therefore be an important future direction of CAT.
The quality of the information is of course very important. If the predictive
power of the model is low, hardly any gains will be made. Moreover, as was also
illustrated by Guyer (2008), inaccurate initialization of CAT will even result in longer
and less informative tests. Another issue is related to the ethical implications of the
use of empirical priors. When they are applied, each candidate is not only scored
based on his/her responses, but background information is taken into account as well. In medical applications it would be no problem to use available information
about the patient to obtain more precise results of testing. But in high-stakes
educational measurement, it would not be accepted. For those applications, it
could be considered to use empirical information during CAT administration, but
to report final scores based on response patterns only.