Hi everyone,
In our study (N=300) focused on cooperation between a human operator and an autonomous system, we administered a six-item scale repeatedly (eight measurements after each condition in the experiment). Now we want to run the EFA and CFA. Our idea is to split the data matrix into halves, one for EFA and one for CFA. The question is: can we use all the data we obtained (300 participants * 8 measurement points = 2400 measurements), or only one measurement point (the first time participants filled in, leaving us with 300 measurements)?
Many thanks!
Dear Nikol Figalová
For the precise reason you suggest, performing an EFA and a CFA on the same data is often a terrible idea: a factor structure obtained from an EFA will nearly always fit extremely well in a CFA using the same data. Because EFA and CFA are so tightly associated, it is not surprising that this is the case.
Best Regards
Dr. Fatemeh Khozaei
Dear Fatemeh Khozaei ,
I am aware that we can't run both EFA and CFA on the same data. That is why I mentioned that we would split the dataset into halves and run EFA on one half of the dataset and CFA on the other half of the dataset (as is the common practice). Therefore, the same data will not be used twice.
My question is if we can use all 8 measurements from our participants (hence 300 participants times 8 measurements, giving us 2400 measurements), or just one measurement (e.g., the first measurement, giving us 300 measurements).
Best, Nikol
In my opinion it would be most useful for you to use only CFA on the complete data set with all measurement occasions, as CFA is a lot more flexible in handling longitudinal (repeated measures) data than EFA. EFA cannot easily deal with, for example, indicator-specific (method, wording, etc.) effects, which often occur in longitudinal data and lead to a complex factor structure. Unless you really need to explore the number of factors, I would recommend using only CFA.
If you need to explore the number of factors, then I would recommend running an EFA separately for each time point, as combining all time points would most likely lead to uninterpretable EFA results due to the complex facet structure of longitudinal data.
Dear Christian Geiser ,
Many thanks for your answer. Do you think it is a feasible option to run EFA on the first measurement point only, and CFA on the remaining seven measurement points?
Best, Nikol
Nikol Figalová
I cannot tell whether your reply came after or before Christian Geiser 's so this may be redundant, but in addition to his remarks I would add that you also need to take into consideration if you already have a clear item-factor-structure in mind. If so, CFA would be preferable. Personally, I do not believe I have ever seen a longitudinal EFA, let alone with eight measurement points (which is fantastic!), but I am sure they do exist.
Best
Marcel
Nikol Figalová Unless you think that the factor structure could have changed over time it seems OK to run EFA on the Time-1 data only--if you feel that EFA is necessary.
Nikol Figalová Randomly splitting a data set into two parts is not a good idea, because both samples - by the randomized split - should be identical. Hence, whatever result you get in the first sample, a correct factor structure or an incorrect factor structure, will be "validated" in the second sample. Therefore, if your data are based on a theory, just proceed with the total sample.
I would test the factor structure at the 1st measurement time point first and then perform invariance tests over the following measurement time points.
Kind regards,
Karin
The exploratory factor analysis is just employed to reduce the data for better understanding . CFA is used to confirm the the factor structure and to analyse the reliability and validity of the measurement variables
There's a really helpful thread on the Mplus discussion boards that seems the most appropriate approach in this case: http://www.statmodel.com/discussion/messages/9/1326
This article may also help: Wu, A. D., Liu, Y., Gadermann, A. M., & Zumbo, B. D. (2010). Multiple-Indicator Multilevel Growth Model: A Solution to Multiple Methodological Challenges in Longitudinal Studies. Social Indicators Research, 97(2), 123–142. http://www.jstor.org/stable/40649360.
Hope that helps
Kane
Yep, but you need to control for the lack of independence of observations. Multilevel cfa is probably the best way to go here. You could also use a random intercepts model (which is sort of the same thing). Would be interested to know which approach you took 😊
- Badges
- Science method
- Similar topics
- Mathematics
- Statistics
- Statistical Techniques
- Structural Equation Modeling