Good question, but I am not aware of anything that says you cannot do this. You should run Cronbach's Alpha stats to make sure the transformed measures are reliable. Also, you could run the CA using the transformed and non-transformed variables to see which have the highest CA values. If you are adding the measures into a composite measure weighted by Eigenvalues from the FA this may alleviate the skew and make prior transformation unnecessary. So I would check that too.

I hope this helps,

RJS

Hello, la. Could you please share what type of question you are using in the survey? Also, is your data set normal or non-normal? While it is not inherently problematic to conduct factor analysis after transformations, caution is required in interpreting results w.r.t. the nonlinear relationship between transformed and non-transformed items which occur within the same factor. I agree with the Mr. Stringer that Cronbach's Alpha is essential to gauge reliability. I'd suggest one further step of also using Pearson's correlation coefficient. Norris & Aroian (2004) find that transformations may not be necessary in the case of using both Cronbach's Alpha and Pearsons correlation coefficient. Here's a link to their paper: http://www.ncbi.nlm.nih.gov/pubmed/14726780 Tashi delek! 

Technically, you can use factor analysis on any correlation matrix. What would concern me is the possibility that all or most of the transformed variables would load on the same factor. This is known as a "methods factor" because the items share a degree of correlation due to be treated with the same method.

The best way to find out if this is an issue to try an Exploratory Factor Analysis.

Hi Cathryn, Are you still at GCBS? Thank you for your answer and also the link to a paper. I had a 9-point Likert type validated scale beginning with 'nothing' to 'a great deal'. So I applied transformation to some of the non-normal items.

I agree with David L Morgan. You should try exploratory factor analysis before conducting CFA.

It is important to realize the effect of a log-transformation: It changes an inherently multiplicative relationship between raw items into an additive relationship. Factor analysis is assuming additive relationships so the transformation is even a must when the items concerned are in multiplicative relationships to each other and to the other items that were not transformed.

This is an important question. There can be theoretical reasons for the transformation (such as the one mentioned in the previous paragraph) and there can be distribution motivated reasons to select a normalizing transformation. If there are no theoretical reasons to expect multiplicative relationships it is often better to resort to non-parametric methods.

From your description I gather that your items are Likert items. Normally the distributions that benefit from a log-transformation concern items that ratios or proportions that are inherently truncated items. Since your items are not of that kind it suggests that your respondents use a proportional model internally. There might be something in the wording of your questions that stimulates the respondents to use such an internal model in their replies. When that is the case I would consider the log-transformation proper. But you need to investigate why some items are in need of transformation while others are not, else you could apply the transformation to all items! If not, I would suggest non-parametric methods.

This may be helpful - it provides an SPSS like  interface to R and allows you to undertake a factor analysis on ordinal variables

https://www.jstatsoft.org/article/view/v046i04/v46i04.pdf

here is the abstract

Exploratory factor analysis is a widely used statistical technique in the social sciences.

It attempts to identify underlying factors that explain the pattern of correlations within a set of observed variables. A statistical software package is needed to perform the calculations. However, there are some limitations with popular statistical software packages, like SPSS. The R programming language is a free software package for statistical and graphical computing. It offers many packages written by contributors from all over the world and programming resources that allow it to overcome the dialog limitations of SPSS. This paper offers an SPSS dialog written in the R programming language with the help of some packages, so that researchers with little or no knowledge in programming, or those who are accustomed to making their calculations based on statistical dialogs, have more options when applying factor analysis to their data and hence can adopt a better approach when dealing with ordinal, Likert-type data.

Keywords: polychoric correlations, principal component analysis, factor analysis, internal reliability.

The R package suggested by Kelvyn Jones is indeed vsuitable if you want to apply non-parametric methods. A good suggestion!

Hi, Kezang, sir! Thanks for your reply to give us more information on your project. You have a lot of high quality information here to help you decide how to approach this challenge! I agree that non-parametric tests may be most suitable. Have you had a chance to look at the dataset's normality if some of the problematic items are removed? That may be just the info you need to have a statistically sound basis for how to proceed. What do you think will work best for you and your research's assumptions/intended outcomes? I'm not physically at Gedu just now, but I remain close at heart with the place and my colleagues! How is Paro town? Best wishes for the end of semester chaos! Here's a link to another discussion on FA and non normality that may be helpful for you: http://stats.stackexchange.com/questions/34753/are-data-transformations-on-non-normal-data-necessary-for-an-exploratory-factor Also, here is an old, but sound paper on the cross-products dilemma outlined by Niemeijer above: http://link.springer.com/article/10.1007%2FBF02294108#page-1 Again, tashi delek, la!

If the transformation is applied to mitigate negatively skewed distributions, shouldn't it be applied to the rest of the variables as well? Otherwise, could the correlations for transformed variables be all negative coefficients?