On Mon, 22 Jan 2001 15:05:22 -0500, "Lee Creighton"
<[EMAIL PROTECTED]> wrote:
> I have some data to analyze for my dissertation, but I'm not sure what
> method to use to answer the question I am investigating. This may not be the
> correct newsgroup for it, but there are rumors that some statisticians read
> it.
>
> First, the data set is from the Third International Mathematics and Science
> Study, and consists of (among other things) a student's imputed mathematics
> ability score, the region of the US where this student lives, as well as the
> response to each question on the test. There were about 15 forms of the test
> given, so each student didn't answer each question.
>
> In other words, I have something like this:
>
> MathScore Region Question 1 Question 2 ... Question n
> 100 NorthEast correct incorrect ... correct
> 130 Central incorrect incorrect ... correct
> 115 NorthEast incorrect correct ... incorrect
> . . . . .
< snip, detail. 350+ Q in all, no single question in common.
6000 students. >
Okay, you have determined that regions differ, grossly, on the overall
score.
Do you need to investigate that score, or do you accept it as Fine?
- There are 15 forms of the test. Presumably, they swap sections of
items at once. How many Blocks of items are there?
- Are there supposed to be equivalences? (are they?) - are there
equivalences, item-by-item, or block-by-block?
The obvious display is the one of ANOVA F-tests for each block it
items, testing across Regions, while using the overall math score as
covariate. Those block-overall scores would be followed by the
ANCOVAs that control for the block-total as covariate, performed on
each of the items.
Somewhere in there, it would be useful to look at the internal
reliability of the blocks -- some items are more reliable, and would
show differences more poignantly merely because they are better. The
items that are apt to reflect differences in teaching-content are the
few which discriminate better than you should expect if you judged by
reliability.
Of course, the study should try to control more closely than just by
"region." Aren't there clusters within school? Aren't there
potential clusters based on the textbooks that the schools use?
> So my question is, how do I determine which questions are good at this
> discrimination? I've been trying to cast it as a regression problem, but I
> don't think that's it. Is this a case for something like Fisher's
> Discrimination analysis? I've only got passing familiarity with the
> technique.
Figure out what to you might want to say about a SINGLE variable
before you start worrying about 350.
--
Rich Ulrich, [EMAIL PROTECTED]
http://www.pitt.edu/~wpilib/index.html
=================================================================
Instructions for joining and leaving this list and remarks about
the problem of INAPPROPRIATE MESSAGES are available at
http://jse.stat.ncsu.edu/
=================================================================
