There is a book on Scale Psychometric Properties by Netmeyer et al. 2004 which talks about three stages- EFA, Item Analysis and CFA. The use of EFA/CFA depends on the nature of construct. If the literature talks about the construct and  its dimensions and scale is established then CFA need to be used. And if the scale is yet to be established then EFA need to be used.

Hardip

There is a book on Scale Psychometric Properties by Netmeyer et al. 2004 which talks about three stages- EFA, Item Analysis and CFA. The use of EFA/CFA depends on the nature of construct. If the literature talks about the construct and  its dimensions and scale is established then CFA need to be used. And if the scale is yet to be established then EFA need to be used.

Hardip

If you have theoretical basis for factorial structure of your scale, you can directly use CFA. Please refer the article:

Petchsawang, P. & Duchon, D. (2009). Measuring workplace spirituality in an Asian context. Human Resource Development International, 12 (4), 459-468.

If you are developing a new scale, then you should use EFA. Because it will help you to understand the factor structure.

What is the question you are trying to answer?  EFA addresses the question of an unknown factor structure (empirical or theoretical).  CFA addresses the question of does the tool fit the theory. It, in this case, is a form of item analysis in that latter case, in addition to questions about factor structure. 

Usually the way the tool was constructed helps make the call.  If you used the literature to develop the tool you usually have a content map (written or implied) in mind.  In that case you would typically follow up with CFA.

They have covered the uses of each and Philippe's suggestion of splitting the sample is a very good one interested in early on validation efforts. Another suggestion would be to use exploratory structural equation modeling as it freely estimates the relationships of all variables but allows you to define a bit more about the factor structure at the same time more than with an EFA.

Confirmatory Factor Analsysis (CFA) tests whether a known factor model can predict a set of observed data (DeCoster, 1998). Researchers use CFA to verify or confirm hypotheses or theory (Ruscio & Roche, 2012; Schmitt, 2011), establish the validity of the factor model, compare two models using the same data, test the significance of factor loading, test relationships between factor loadings, test for correlation or lack of correlation of factors, and assess convergent and discriminate validity of measures (DeCoster, 1998).

Exploratory Factor Analysis (EFA) tests the number of common factors that influence measures and tests the strength and relationship between each common factor to the corresponding measure (DeCoster, 1998). Researchers use EFA to identify the nature of constructs that underlie responses given in the questionnaire, determine sets of items that interconnect, demonstrate the depth and breadth of measurement scales, classify the most important features of a group of items, and generate factor scores that represent the underlying constructs (DeCoster, 1998). Researchers commonly use EFA to develop and validate self-reporting assessment instruments, especially when there is little or no a priori knowledge of the structural model being developed (Ruscio & Roche, 2012).

To summarize, EFA allows researchers to discover constructs or concepts that are not directly measurable by collapsing a large number of Likert variables into a few interpretable underlying factors representing such constructs. Researchers use CFA more often than they use its exploratory counterpart. Instead of reducing or identifying the minimal number v of Likert variables to form a set of k factors or latent variables as done by EFA, the CFA procedure determines the extent to which the given set of v Likert questions measures k predefined factors. In other words, EFA is a data reduction procedure or an inductive theory procedure whereas CFA is a procedure for testing hypotheses deduced from that theory. 

REFERENCES

DeCoster, J. (1998). Overview of Factor Analysis. Retrieved from http://www.stat-help.com/factor.pdf 

Ruscio, J., & Roche, B. (2012). Determining the number of factors to retain in exploratory factor analysis using comparison data of known factorial structure. Psychologocial Assessment, 24(2), 282-292. doi:10.1037/a0025697

Hi Carla,

Do you mind if I step in to add to this discussion? Maybe Exploratory Structural Equation Modeling (ESEM) can be used as a possible third way between EFA and CFA. You can read more here: Asparouhov, T. & Muthén, B. (2009). Exploratory structural equation modeling. Structural Equation Modeling, 16, 397-438. Hope you find this useful.

Best wishes to Barcelona,

Andreas

I guess, the practices are in a turn to go more CFA way. The Classical EFA is somewhat restricted from the validations viewpoint and it's quite difficult to find a software for EFA which would allow analysis of non-normal or ordinal variables (like Likert scale variables). The modern ways of combining the EFA and SEM (that is, ESEM as suggested by Andreas) or the most used way of using SEM (that is, to remodel the phanomenon on the basis of the empirical dataset) may be the way the Journals are pushing the practices. So, though the most posts suggest the EFA as the first step, I'd suggest to go to SEM or ESEM way. MPLUS is quite handy for this.

All the best,

Jari 

What is better to use as a first validation attempt: Exploratory Factor Analysis or Confirmatory Factor Analysis?

EFA and CFA are two different world and you should not mixed. EFA is carried out using the Pilot Study Data to determine the dimensionality/grouping of items. In EFA, the study screen-out low factor loading items and redundant items. Thus, EFA would give you the high factor loading items and their dimensions (if any). Here you can compute the measure of reliability of your items in the form of Cronbach Alpha.

After pilot study, you go to the field and collect the Field Study Data. Using this Field Study Data, you carry out the CFA to determine Unidimensionality, Validity, and Reliability as required prior to running SEM.

I would echo Andreas - ESEM provides a useful alternative. This has been well demonstrated by Herb Marsh and colleagues:

Marsh, H. W., Nagengast, B., Morin, A. J. S., Parada, R. H., Craven, R. G., & Hamilton, L. R. (2011). Construct validity of the multidimensional structure of bullying and victimization: An application of exploratory structural equation modeling.Journal of Educational Psychology, 103(3), 701-732. doi:10.1037/a0024122

John

Markus Mattsson gives some practical tips for using ESEM in

Investigating the factorial invariance of the 28-item DBQ across genders and age groups: an Exploratory Structural Equation Modeling study.

Accident; analysis and prevention 09/2012; 48:379-96.

According to Skrondal & Rabe-Hesketh (2004) EFA and CFA are complementary strategies, the latter could be seen as a particular case of the former as this approach restrict a priori to zero loadings those not theoretically connected to the hypothesized factors. Another book (Hancock, G. R. & Mueller, R. O. (Eds.), 2010. The Reviewer´s Guide to Quantitative Methods in the Social Sciences. New York: Routledge.) discuss when using one or other approach and their opinion are the same as Chahal has already commented. However, if there are some empirical evidence about the construct that you could not confirm in your study, you have to keep in mind that if you change your model based on the modification indices you turn back to the exploratory models. It is also possible to perform exploratory factor analysis using the SEM approach, the Exploratory Structural Equation Modeling. 

CFA is theory based. As I interpret from your statements, the factorial properties of the model is already explored  and yours is the first application then you can use CFA. Alternatively you can always first use EFA and then CFA. There is no harm in using CFA irrespective of the case. If your result did not show up as you expected, you can always pursue CFA in EFA mode.  

It seems pretty controversial to me.  Some experts like Bengt Muthen of Mplus are ok with using EFA like this.  Others would never use EFA under any circumstances.  They think of it as being like stepwise regression - letting the computer tell you what your model is instead of theory. 

If you have a strong theoretical model you could find literature to support both approaches (the direct to CFA approach or the EFA first approach).

I choose EFA or CFA dependig the strength of the construct, while the theory is more stronger and have more evidence, i use CFA.

Although, strictly the specific usage of EFA and CFA  has very different and distinct purpose, but some scholars argue that there is no hard and fast rule; you can also use EFA in CFA mode and  CFA in EFA mode. My suggestion would be, at first, you need to be very clear about the goal of your research. The appropriateness in the use of EFA  and/or CFA is very much dependent on the goal of your research. For example, if your  goal is to evaluate psychometric properties and factorial validity of the instrument for which sufficient theory has already been developed in the literature, you now know the nature and characteristics of the construct(s) then use CFA. CFA is a hypothesis testing mode of a measurement construct and the instrument. Else, if the literature suggests a very vague theoretical sense of the measurement construct, you use EFA. In general, you can start with EFA then switch into CFA if you have a large sample split them into two halves; use one of the halves for EFA and the other halves for CFA. For cross validation you need two different samples meaning thereby drawn from two different populations. This is a conventional way of doing. For accuracy of your result CFA needs larger sample than EFA. In much of the contemporary research, scholars tend to use CFA first then if the theory do not hold for the data  they switch into EFA mode in the very CFA framework.  I hope this would help. 

I agree with Senapaty. Just to complement, it helps thinking how it works... In EFA, every item is correlated with every hypothetical factor made available by the variance of the data extracted. What it means is that as you can see in your output for factor loadings, every item has a factor loading for each extracted factor right? So there are no imposed restraints between items and factors, unless you impose in the software how many factors you want, otherwise it is going to extract – again – as many factors as the variability makes possible.

On the other hand, in CFAs, you apply restraints to the relationships. You have to indicate a specific number of factors. After that you have to indicate in the model the specific relationships between items and factors, in a way that the items will be loading in specific factors. That's exactly why CFA is theory based, because you need a hypothesized model (theoretical) to actually specify and infer about the relationships.

Imagine how hard would be to conduct a confirmatory approach in a set of items that you don’t have a previous theory/prior findings to guide it. Without a background theory, information on the factor loadings of an EFA are valuable as they indicate which items go altogether aligning under common factors. It gives you a first glimpse at how your empirical data behaves in terms of structure.

Going through a CFA without theory or without a preliminary information on the latent structure, would make you go through endless trials until you reach a model that fits. So EFA is usually ideal at least as a preliminary approach, but still you need to think about your research issues. For example, EFA is also a valuable procedure to reduce a huge set of items into a smaller and simple one. You can do that with a CFA as well, but in terms of being practical, I’d go with a EFA. Than think about the objectives of your study as well.