[Rasch] analysing CTT construct using IRT
man.xu at education.ox.ac.uk
Tue Oct 27 17:40:57 EST 2009
Thank you very much for your reply. My question was about Test Information Function. Is it necessary to evaluate the precision of the measures? I was reading a book Data Analysis Multivariate Social Science Data and found that many examples of IRT it uses for analysing Likert scale data have not even been given any information about the Test Information but fit index such as Chi-Square. I understood that the Test Information is supposed to be one of the advantages of the IRT compared to CTT, but is it used much in practice for purposes other than designing a ability test?
From: Mike Linacre (RMT) [rmt at rasch.org]
Sent: 27 October 2009 00:25
To: rasch at acer.edu.au
Subject: Re: [Rasch] analysing CTT construct using IRT
Thank you for telling us about your research. Please help us to understand the situation better ....
Are you talking about the Test Characteristic Curve = raw-score to measure transformation (higher measures for higher scores)
the Test Information Function = precision of the measures (usually peaks near the center of the test) ?
At 10/26/2009, you wrote:
I am very new to the IRT field so please forgive me if my question is too simplistic.
I am using the PISA 2006 data for my doctoral dissertation and so far I have been using CFA methods for analysing the construct properties the Likert scale data in PISA. All constructs in my study have high level of reliability and mostly demonstrate good multigroup invariance properties under CFA/SEM methods. Recently, I was introduced to the IRT methods and start to development new perspectives on measurement issues of the data at hand.
In PISA, many of the motivational constructs are measured through Likert scales and were measured through well established constructs developed under the framework of CTT. Using IRT method, I found that the difficulties of items in these constructs are often clustered at middle or high levels of the trait. The Test Characteristic Curves are always peaked at one place rather than spread all over the trait levels. I guess this means that the latent variable is not realiable enough for all the levels. So I was wondering if these constructs are valid enough to use for further analysis or whether I should just stick to my old way of analying this data, i.e. using CFA/SEM and treating the likert data as if they were continuous.
Editor, Rasch Measurement Transactions
rmt at rasch.org www.rasch.org/rmt/<http://www.rasch.org/rmt/> Latest RMT: 23:2 Autumn 2009
Please consider the environment before you print
-------------- next part --------------
An HTML attachment was scrubbed...
More information about the Rasch