You generally can't transform scales with a (relatively) low number of
points to normality. If a lot of people have the top (or bottom)
score, no transformation will get rid of that pileup.
Your best bet is to use an estimator that treats the data as ordinal
(such as WLSMV), then distributions are irrelevant. Your sample is on
the low side for that, but is probably OK. (It looks like you're using
AMOS, I'm not sure if AMOS can do this).
Jeremy
On 11 August 2013 13:32, Sam Jee <[log in to unmask]> wrote:
> Hi everyone,
>
> I am really struggling to deal with my survey data. I collected data from
> about 350 respondents on three scales. The questions were rated on 5 point
> scales. I have found that most respondents were scoring very highly and I
> was getting quite a high number hitting the 'ceiling'. It really skewed my
> data negatively and I can't normalise it (I have tried a number of
> transformations (http://www-01.ibm.com/support/docview.wss?uid=swg21479677)
> with no luck). I think it is unlikely I will be able to salvage this data to
> normalise it. Unfortunately, I need to test the factor structure for the
> scales which have been shown to have a number of factors ranging from 2-5
> for the three scales. Provisional EFA analyses has shown terrible pattern
> matrices, usually extracting 1-2 factors, and I can only imagine the data
> will not fit the factor structure in CFA laid out by previous researchers.
>
> How best is it to proceed?
>
> If anyone has any ideas on normalising this data that would be welcomed! I
> suppose my plan is to try to fit the model which will not work no doubt and
> then go with EFAs to make my own model. I can always test this in CFA after.
> I am breaking assumptions like univariate and multivariate normality but I
> have read that there is some room for violations like this...
>
> Thanks for your help!
|