Factor analysis with non-normal distribution?

[Wiggy]

Hi all,

Does anyone know an alternative to factor anlysis for non-normal distrbutions which explores factors or clusters on a set of 64 questionnaire items each with a 5-point Likert scale (363 respondents). I did read a recent post by Analysis factor who recommended latent class analysis (LCA) for catergorical data. However, I believe I can treat my data as interval* so would LCA be appropriate or is there another alternative?

Any help appreciated

Sue

*Jaccard and Wan (1996: 4) summarize, "for many statistical tests, rather severe departures (from intervalness) do not seem to affect Type I and Type II errors dramatically." Jaccard, James and Choi K. Wan (1996). LISREL approaches to interaction effects in multiple regression. Thousand Oaks, CA: Sage Publications. "

 

[owenpediatrica]

Can I have a second question--
Why has non-normally distributed data been a problem in EFA? I know there is some assumption in OLS regression or so, but it has never been a practical problem before unless you have a really wierd distribution. It seems to me just a data exploration (or reducing) technique. Unless you are really using the factor calculated from the factor analysis (like a propensity scale or some instrumental variables), it does not seem to me an issue.

 

[vinux]

Hi Sue,
When you use Principal component method for estimating factors, you don't require any normality assumption.

 

[Wiggy]

Factor analysis with non-normal distribution

Hi
and thanks for your responses, you have solved a problem

Firstly, in response to owenpediatrica, at this stage in my research I do not want to use the resultant factors for a scale or instrumental variable; the excercise is simply data exploration. So I guess I should be fine. After reading your response I grabbed my copy of Tabachnick and Fidell (2007) and they say what you have confirmed... 'as long as PCA and FA are used descriptively as convenient ways to summarize the relationships in a large set of observed variables, assumptions regarding the distributions of the variables are not in force'.
Secondly, in response to Vinux, I really should use Exploratory FA and not PCA as I am literally exploring the data and not trying to confirm or support any theory as such. In other words, no theories or hypotheses were generated as to what factors would emerge, so EFA is more appropriate.

The data I have will eventually be subjected to Rasch analysis which checks for 'good' and 'bad' items as well as 'good' and 'bad' respondents...but that's another story.

Thank you again for you help guys; it really is appreciated.

Sue