Why would one do CFA instead of EFA?

This is a simple question for which I could not find an answer. CFA and EFA are both methods of factor analysis. It is said that EFA extracts a factor structure from the data whereas CFA is used to test if a factor structure fits the data (or in other words to test a hypothesis).

Given this, why would I want to do CFA instead of EFA if the latter is able to determine the factor structure itself? What extra benefit do I gain by testing my own hypothesis using CFA?

• Strictly speaking, "factor analysis" is EFA. CFA is a label for a subtype of structural equations modeling (SEM), which older name is "path analysis". They are just different tasks. (E)FA is to discover latent factors and CFA is to check (fit, test) statistically those discovered by EFA or those thought out by you or taken from literature. EFA can be used to compare latent factors to some degree, but SEM is more powerfull tool in that respect. Oct 31, 2017 at 15:42
• Ok, but will not EFA already find the best latent factors? What I mean by best is the factors that are found by EFA will maximize the common variance in the data. If that is the case, why would I want to try something else? Oct 31, 2017 at 16:17
• because sometimes you want to test rather than describe. both are legitimate goals, but they are different. Oct 31, 2017 at 18:36
• the best latent factors for fitting this dataset. But who will test how it is reasonably good for the population? Nov 1, 2017 at 8:44
• So in CFA we do not have to create names for latent factors, right? We simply use the names of our existing variables and relate them in a linear manner to explain the data. But in EFA, we obtain a set of initially unnamed factors. Each variable may load onto one or more factors. Based on these loadings we must create meaningful names for our factors. Nov 2, 2017 at 15:37

You are correct; exploratory factor analysis (EFA) and confirmatory (CFA) may be considered "factor analysis methods." Though strictly speaking, as @ttphns pointed out, EFA is technically factor analysis, and CFA is factor analysis constrained to a factor structure of substantive interest.

While EFA can be used to explore the factor structure of the data, it does not automatically give you an answer as to what structure you should test with a CFA. To do this, you must consider item content (i.e., what do the items say?) as well as substantive theory (i.e., what kind of factor structures does data similar to yours look like in the literature). Additionally, you do not necessarily have to run an EFA prior to a CFA. For example, I often work with language proficiency assessments, where items are written explicitly to measure listening, reading, writing, or speaking. Given this, I seldom run EFA before CFA, given I already have a good idea of the factor structure. I still run CFA, however, as testing whether the model fits the data is still important.

This is all to say that EFA does not tell you the correct structure for your research purpose; it is only a tool to explore the structure of your data and cannot be used a test a substantively relevant hypothesis regarding latent factors. In your question, you ask, "What extra benefit do I gain by testing my hypothesis using CFA?". While a good question, it does not make much sense given EFA is seldom ever one's final step when testing a hypothesis concerning latent variables. Your question should instead be, "In your question, you ask, "What extra benefit do I gain by using EFA?" because, in some instances (i.e., when you already have a plausible hypothesis regarding the factor structure), it may not be necessary.

Personally, I find reading the substantive journals that apply the methods I am interested in learning about (i.e., journals that apply factor analysis methods, as opposed to journals that discuss the more technical aspects of factor analysis) to be a great way to understand the rationale behind why certain methods are applied. In the references below, I have included a couple of articles in the patent-reported outcomes (PRO) literature that uses both EFA and CFA together.

References

DeWalt, D. A., Thissen, D., Stucky, B. D., Langer, M. M., Morgan DeWitt, E., Irwin, D. E., ... & Varni, J. W. (2013). PROMIS Pediatric Peer Relationships Scale: development of a peer relationships item bank as part of social health measurement. Health Psychology, 32(10), 1093.

Lai, J. S., Stucky, B. D., Thissen, D., Varni, J. W., DeWitt, E. M., Irwin, D. E., ... & DeWalt, D. A. (2013). Development and psychometric properties of the PROMIS® pediatric fatigue item banks. Quality of Life Research, 22(9), 2417-2427.

Using EFA we can discover the latent factors. Hence, EFA is a theory-generation technique. This generated theory may be tested or confirmed using CFA. Hence, CFA is a theory-testing technique.