Data requirements for Factor analysis

[Haijie Ding]

Dear Listers,
I have a data set which contain data in different scale. Most question scale
is 1-5 likert scale. We also have other scale for some questions. For
example, we have category question, like location of work. Another example
is we have question to distribute 100% of time to different locations. My
question is: what's creteria to do explorative factor analysis? Can I do
exporative factor analysis with all of the data i got?

Could you guys please give me some advice and/or reference on this topic?

Thank you very much.


Bests,
Haijie

[Maguin, Eugene]

Haijie,

Ok, so you have some questions that have a 1-5 likert scale format.
You have one or more questions that are purely nominal categorical--the
response format order doesn't mean anything. It may be that some of these
questions are also 'choose all that apply'. For instance, if there are
several work sites and people do work at multiple sites, then they might
legitimately respond that they work at sites 1, 3, and 7. On the other hand,
some of the questions might be the standard single response question e.g.,
gender.
Lastly, you have a one or more questions whose response format is such that
the sum of responses equals some number, in your case, 100%.

I would use only the likert scale questions in the factor analysis. Partly
this is because conventional factor analysis with categorical items is
statistically incorrect. But more importantly, I would expect that the
likert items were (should have been) constructed to measure certain
concepts. Ideally, each item measures an aspect or exemplar of the concept.
For the single response nominal categorical questions, you could construct
dichotomous indicators and factor those. BUT, the question is what evidence
can you offer that the items were selected apriori to measure a construct.
If you can, then I think spss now has a specialized routine(s) for
categorical items as the spss factor routine is not correct for dichotomous
or ordinal items. However, I'd prefer to consider the categorical variables
as predictors or covariates of the likert scale factors. Finally, I have no
idea of the analysis possibilities for multiple response categorical items,
except to also treat them as covariates. I can't comment on your third type
of item. I'm skeptical of such items meeting my underlying measurement
criteria and I'm unsure of whether the factor routine is mathematically
capable of factoring such items. Perhaps somebody that knows about these
kinds of items will comment.

Gene Maguin



If you could convince me that your

[Haijie Ding]

Gene,
Thank you very much for your advice. I will do factor analysis on only those
questions with likert scale.

Bests,
Haijie

On 5/11/07, Gene Maguin <[hidden email]> wrote:

>
> Haijie,
>
> Ok, so you have some questions that have a 1-5 likert scale format.
> You have one or more questions that are purely nominal categorical--the
> response format order doesn't mean anything. It may be that some of these
> questions are also 'choose all that apply'. For instance, if there are
> several work sites and people do work at multiple sites, then they might
> legitimately respond that they work at sites 1, 3, and 7. On the other
> hand,
> some of the questions might be the standard single response question e.g.,
> gender.
> Lastly, you have a one or more questions whose response format is such
> that
> the sum of responses equals some number, in your case, 100%.
>
> I would use only the likert scale questions in the factor analysis. Partly
> this is because conventional factor analysis with categorical items is
> statistically incorrect. But more importantly, I would expect that the
> likert items were (should have been) constructed to measure certain
> concepts. Ideally, each item measures an aspect or exemplar of the
> concept.
> For the single response nominal categorical questions, you could construct
> dichotomous indicators and factor those. BUT, the question is what
> evidence
> can you offer that the items were selected apriori to measure a construct.
> If you can, then I think spss now has a specialized routine(s) for
> categorical items as the spss factor routine is not correct for
> dichotomous
> or ordinal items. However, I'd prefer to consider the categorical
> variables
> as predictors or covariates of the likert scale factors. Finally, I have
> no
> idea of the analysis possibilities for multiple response categorical
> items,
> except to also treat them as covariates. I can't comment on your third
> type
> of item. I'm skeptical of such items meeting my underlying measurement
> criteria and I'm unsure of whether the factor routine is mathematically
> capable of factoring such items. Perhaps somebody that knows about these
> kinds of items will comment.
>
> Gene Maguin
>
>
>
> If you could convince me that your
>
>
>
>
>