Hi,
I am relatively new to Stats in general, so all comments are much
appreciated.
I appear to have stepped, initially unbeknown to me, into the ordinal v
interval data analysis debate.
I have conducted a mail survey, with a sample size of +-150, and
respondents were asked to rate a 60 variables using a Likert scale.
The data is skewed and not normally distributed.
I would like to undertake an exploratory factor analysis of my data.
However, there are 2 issues which need to be addressed.
Firstly, the variables are gathered using a 5 point Likert scale, with
an additional 'Not Applicable' category. The format is based on past
research, however, the 'Not Applicable' category, is not discussed in
the past research. Not applicable is obviously different to Non
Response. Can this data be coded as zero for the purpose of factor
analysis, with the remaining 'Unimportant' through to 'Very important'
being coded with the values 1 to 5?
Secondly, I am aware that Pearson's correlations assume normal
distribution. However, I have also heard that ADF and Polychoric
correlations matrices are only applicable for large sample sizes
(+-2000) and up to 25-30 variables. Is is feasible to generate a
Polychoric matrix in this scenario, or should a conventional' factor
analysis utilising maximum likelihood or general least squares rather be
performed? Is there a commonly accepted approach for relatively small
sample sizes?
Regards,
Chris Walsham
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
|