Search All of the Math Forum:

Views expressed in these public forums are not endorsed by NCTM or The Math Forum.

Topic: _pragmatic_ item/scale analysis
Replies: 3   Last Post: Jun 20, 2004 10:40 PM

 Messages: [ Previous | Next ]
 Dr. Hans-Christian Waldmann Posts: 5 Registered: 12/18/04
_pragmatic_ item/scale analysis
Posted: Jun 20, 2004 10:01 AM

Hello to all!

A client to our university stats consulting has produced a
questionaire with over 250 items that he considers to be
associated to 12 scales (with 6-30 items for a scale). There
is consent that the number of items is far too high in a
clinical context, and has asked us to reduce the item count
in a "pragmatic" way. Hm. He objects to the idea of re-
evaluating content validity ("they're all reasonable").
Nevertheless, I've been asked to cut 30% of the items while
retaining the 12 scales.

Now, I am not really proficient in this kind of pragmatics.
Having read through the usual references I have a devised
a list of "selection criteria" to eliminate items.

The first consideration _after_ this was: any procedure
meeting the demand of "no more than x items" is likely to
miss the "optimal" scale. I might _have_ to eliminate items
although their withdrawal will lower Cronbachs alpha and
deteriorate other desirable properties.

For the time being, I would kindly ask you to comment on
_sense_ / validity: I have to come up with some formal or
thing. I am _not_ proud of it.

So here is what I have so far: exclude (yes/no)-item if

(1) solution probability is outside range 0.2<p<0.8. The
item does not help differ between people and does not
contribute substantially to scale variance.
(2) item correlates negative with total and mostly negative
with other items of that scale. Assuming correct coding
scheme, the item measures sth. else.
(3) reliability measure Cronbachs alpha rises if item deleted,
this is implemented in both SAS and SPSS routines.
(4) item correlates r>.90 with some (1-3) other items
and exclusion does not lower reliability too much.
Then it is considered "semantically redundant"
(5) the distribution of solution probabilities has a lower
dispersion if item excluded (propto [1], I guess)
(6) all else being equal, switch to linguistics and drop
item with higher lexical and syntactical complexity.

_Technically_, I am confident that I can write a SAS macro
that does some sort of stepwise selection using indices of
(1)-(4). I am not sure what to think of its output.

So it would be of enormous help if you could comment on
that list or provide other hints on "reducing questionaires".

Hans-Christian Waldmann

--------------------------------------------------------------------
PD Dr. Hans C Waldmann
Methodology & Applied Statistics in Psychology & the Health Sciences
located: ZKPR, University of Bremen, Grazer Str.6, D-28359 Bremen
mail: waldmann@samson.fire.uni-bremen.de
web: http://samson.fire.uni-bremen.de/waldmann
friend of: IBM-AIX -- PERL -- POSTGRESQL -- SAS -- LaTeX -- VIM

Date Subject Author
6/20/04 Dr. Hans-Christian Waldmann
6/20/04 Gaj Vidmar
6/20/04 Jim Clark
6/20/04 Richard Ulrich