Alternatives to EFA for preliminary data reduction?

[ursula1985]

Dear all,

I’ve just conducted a pilot study to examine individuals reasons for committing crime. Based on literature I pulled out 50 items (5 pt likert response) that I included in my study. Many of these items are similar but are worded differently. The purpose of this study was to determine which commonly reported reasons would be most relevant to my much younger sample, and the best way to approach these items (item wordings).
My original plan was to conduct an EFA on the data. I now have 3 concerns with that plan:

1. I don’t think the data is suitable because a. I only have a sample size of 90 (yes, for 50 variables, I know..) b. because the items are very highly skewed (most people indicate that a majority of the items were not relevant reasons for their previous crime commissions, and I expected this. It varies but people tend to select 1-5 primary reasons - and these reasons vary greatly across people).

2. I am not sure if I ever had the right idea with EFA. I included all these items in the pilot to be comprehensive, but I need to determine which items I should carry forward into my actual study. Considering this is a preliminary look at these reasons would it be acceptable to just boot out items with the lowest reported frequencies? If I report on my pilot findings could I create my own scales (using cronbach’s alpha) to further reduce related items into meaningful constructs?

3. I have found some interesting gender differences but if I conduct a factor analysis on the whole sample (my gender split is 70/30) it appears that items that are more frequent in males are almost “washed out” (i.e. load poorly on all factors).

If anyone has comments or an idea for an alternative please let me know. Your help is (as always) much appreciated!

[trinker]

Do you have some theory about which items belong in certain dimensions? I guess what I'm saying is can you look at content evidence for validity? Pull out items that don't fit based on expert knowledge grounded in theory. Then you may have less items to look at and could actually run CFA if you have a theory behind the latent trait.

Also...

I'd run some IRT stuff on these items and look at the discrimination and item difficulty (particularly the item characteristic curves). If a item discriminates poorly you may want to get rid of it. If it discriminates at either high or low ability you may want to get rid of it (depending on what you are trying to measure).

For your point 3 I don't think I'd use factor analysis for DIF detection. I believe that using the mantel haenszel or better still comparing the Null (the ability only), gender and then the gender x ability. There's debate in the literature about how to compare models, however, I'd look at the difference in in model 3 and model 1 (the interaction minus the null) and I think (R gives me the p values for this so I don't know off the top of my head) the critical value of DIF detection is 5.99. If you have sig there look at model 2 minus model 1 (that's uniform DIF) and also model 3 minus model 2 (non uniform DIF).

[ursula1985]

Dear Trinker,

Thank you for your quick reply and suggestions.

About your first suggestion: Will my sample size and non-normality be less problematic in a CFA? There is theory, much of it conflicts about their expectations (shocking), which is why I was hoping to do my own EFA, but I would certainly have some theoretical basis if I did go for a CFA.

..I am more interested in your suggestion at running some IRT. I was already thinking this might be an avenue to explore, but I wasnt sure if I would still run into the problems mention above..n/item ratio and normality problems? Do you have any additional information/tips on this?

Thanks for your time!