Good morning,
I am not an expert on Factor Analysis, so I hope the explanation of my problem makes sense.
My current task is to perform analyses on an older data set from experiments my lab conducted a couple of years ago. We have 212 individual items, consisting of a dozen or so demographic questions and items from a total of 24 different scales that measure separate constructs. Given the large number of items and constructs, I would like to reduce the number of dimensions to achieve a more clear starting point for theory development. Obviously, an exploratory factor analysis is a good choice for this.
My question is whether I have to input the 200 or so individual non-demographic items in the EFA, or whether I can instead just use the 24 composite variables/constructs and reduce the number of dimensions from there. My hesitation using all the individual items is that an EFA would simply return something very similar to the composite variables, as the constructs generally have a quite high internal consistency and are fairly distinct from each other based on theory. The obvious caveat with using the composite variables is that it is not something I have seen done much, and as I am not an expert on EFA, I am unsure as to whether there is a major methodological road block to using composite variables that I am unaware of.
Thank you for your help!
Best wishes,
Pascal
I'm not sure I understand your hesitation. If the EFA simply confirms the structure/dimensions that you expect, you would have nothing to loose. The only problem could be that 200 variables are a lot for an EFA, so the question would be whether your sample size is sufficient for stable results when analyzing all 200 items simultaneously.
The composite approach would lead to a loss of a lot of information. In addition, and maybe more importantly, it would address a completely different question (i.e., about the dimensionality of the composites rather than the dimensionality of the items).
Hey Pascal
200 items is a lot for an EFA. But why EFA - you have some ideas about the structure of the data as you have "24 different scales that measure separate constructs". You have a lot of scales with relatively few items. What not address this question in separate stages. First, you could test the dimensionality of each of the 24 scales separately using CFA to check that each scale is measuring only one latent variable. Second, when you have sorted out the structures at the level of the scales you can look at how they are correlated, or indeed factor analyze composite scores for all the scales together.
Mark
Good morning,
In your situation, you can perform an EFA on the 24 composite variables/constructs instead of the 200+ individual non-demographic items. This approach helps you reduce the number of dimensions and is suitable when the constructs have high internal consistency and are theoretically distinct. It's important, however, to ensure the composite variables are appropriately calculated (e.g., by averaging or summing the items) and to verify that the assumptions for EFA are met. Additionally, consider interpreting the results cautiously and further validating them, as using composite variables in EFA is less common compared to individual items.
If I understand correctly, you have 200 variables with about 10 loading separately on about 20 composites, and are not want to consider crossing loadings (e.g., item 5 loading on constructs 6 and 7). This seems a case for the Structural-After-Measurement (SAM) model, which is implemented in lavaan (https://search.r-project.org/CRAN/refmans/lavaan/html/sam.html). I think the first step though is looking at the measurement of each of the 24, and making sure that there is nothing problematic about these. And as the above commentators point out, this is a lot of variables. One of my stats mentors (David Bartholomew) used to stress how the more variables the less trust in the results. And as the commentators note, you likely have some beliefs and things to evaluate wrt the relationships of the 24, so a confirmatory approach may be better (and sam function in lavaan allows this).
- Badges
- Science topic
- Similar topics
- Mathematics