Re: subscales

On 8 Nov 2005 01:47:01 -0800, "jan" <jan_spreemann@xxxxxxxxx> wrote:

> Hi,
>
> I used in a study with 180 persons a 4 item scale to measure
> stereotypes of different outgroups. From the theoretical point of view
> (Fiske, 2002) the scale can be divided in two subscales of two items
> each.
> Question:
> How can I proof, if this two-subscale structure is also true for my
> random sample? 4 Items are two less for a factor analysis and I don't

Is that supposed to read, "Four items are /too few/ for a
factor analysis"? Well, I do see that there might be a big
chance that everything will load only on one big factor,
but, if you require at least two factors, you do get an
indication of (a) whether the variables drop out, two-and-two,
as you expected, and (b) how much of the variance is
associated with a second factor (2nd eigenvalue compared
to 1st eigenvalue).


> think that it would be a good statistically practise just to compare
> the Cronbach´s alpha of the 4 item scale with the correlations between
> the each two items in the subscales.

Cronbach's alpha doesn't say much about unidimensionality.

Basically, you have the inter-item correlations to work
with, and nothing more. (Or, can you work with the means
as meaningful?) If the designated variables do *not* have
the higher intercorrelations, that would be evidence *against*
the theory. How much higher will they be? I suppose you
could test for correlated correlations, if the patterns are
generally what is expected.

Showing scale *validity* depends on showing some
discriminant validity, or convergent validity, or whatever.
Having test-retest data can give evidence without having
outside criteria, but what you can do with the only one
cross-section of the items, without other information on
the sample, is limited.


--
Rich Ulrich, wpilib@xxxxxxxx
http://www.pitt.edu/~wpilib/index.html
.

Relevant Pages

  • Re: _pragmatic_ item/scale analysis
    ... > contribute substantially to scale variance. ... There should not be many dropped, if the correlations were ... > and exclusion does not lower reliability too much. ... two items, or drop one, regardless of what it does to alpha. ...
    (sci.stat.math)
  • Re: Survey scale reliability and validity
    ... > My survey consists of 10 constructs, and each construct consists of 5-6 questions. ... > All items within the same scale has to be significantly correlated. ... > The correlations within scales/constructs have to be low. ...
    (sci.stat.math)
Loading