Agree with Leonhard, if you have a strong conviction, just try the model belonging to the theory. When using SEM-software, it could provide you with options on improving the model (using the modification indexes). In fact, when people design items to belong to concepts, it is strange to first use EFA. One might test the concepts one by one with a one-factor model and later test the structure among the factors.

EFAs require all indicators load on all factors and will always fit the data better than a CFA which typically constrains some loadings to 0. As a result, CFA is a stricter test and if clean and theoretically-based fits can be obtained, will provide stronger evidence than EFA.

It's critical, as Professor Golay mentions, the theory MUST drive the CFA. Too many bad examples exist where a CFA is performed and then is used to "create" the theory.

Hi all of you

Thanks Leonard, Peter, Joseph and Philippe ! I'd strong conviction of the possibility and the appropiateness to directly performe CFA as its aim is to test a theoretical model. As Joseph mention I know many references where CFA is used to create a theory and some reviewers of french journals quote them to critizise submited papers and to ask some methodological modifications (that I personnaly don't understand :-( ). Do you have and example of published paper, that I can quote, that directly uses CFA on the data ?

Thanks and all the best !

Hi all of you

Thanks Leonard, Peter, Joseph and Philippe ! I'd strong conviction of the possibility and the appropiateness to directly performe CFA as its aim is to test a theoretical model. As Joseph mention I know many references where CFA is used to create a theory and some reviewers of french journals quote them to critizise submited papers (that direcly performe CFA) and to ask some methodological modifications (that I personnaly don't understand :-( refering to the definition of CFA). Do you have an or some example(s) of published paper(s), that I can quote, that directly uses CFA on the data ?

Thanks and all the best !

Phil

Hi folks,

my 2 cents: (1) Do a CFA model whenever you have ANY idea about the potential data generating process. Test this idea and adapt of the model will not fit (and do not explain-away misfit)

(2) A CFA model USUALLY comes with a simple structure and fixed-to-zero double-loadings but that is not necessary. No one forbids to estimate double loadings when theoretically appropriate. Hence, inspect the question wordings of the items and think of potential "dual causal antecedents" of factors contained in the model. The issue is to "think and test"

(3) A sentence "An EFA can identify a latent structure" is probably misleading. Any common factor model (may it be EFA or CFA) applies a certain causal structure on the indicators - that the "common cause" model (or in fact a confounding or spurious correlations - model). This will be appropriate in many cases but not necessarily, see

Borsboom, D., & Cramer, A. O. J. (2013). Network analysis: An integrative approach to the structure of psychopathology. Annual Review of Clinical Psychology, 9(1), 91-121. doi:10.1146/annurev-clinpsy-050212-185608

Hayduk, L. A. (2014). Seeing perfectly fitting factor models that are causally misspecified: Understanding that close-fitting models can be worse. Educational and Psychological Measurement, 1-22. doi:10.1177/0013164414527449

Bollen, K. A., & Medrano, J. D. (1998). Who are the Spaniards. Social Forces, 77(2), 587-621.

Furthermore, if the data set contains indicators that solely reflect a latent variable (single indicator) the factor model completely fails to identify this latent cause. The whole thinking about factor analysis is contaminated by the old "data reduction" paradigm.

Best,

Holger

Hi to all

I thank you so much for your clear and constructive insights. Thanks for the reference. It will all be helpfull for my further writings.

All the best! And greatings for the ending/begining year !

Phil

EFA is required to assess the usefulness of items measuring the construct and also to determine dimensionality of items. Poor loading items should be removed. Then, the new set of items will be used for field study. The field study data will be used to validate the constructs through CFA.

The CFA without EFA will results in deleting many items, leading to invalid measurement model of your construct.

With all the respect Zainudin Awang

I disagree with almost everything and I think these are problematic advises:

1) When you have a theory or even a guess, I vote for specifying a theory-based model and test it. Why should one let a exploratory extraction algorithm try to find some meaning when you already have some idea about it? This is even more the case, when the true data generating process is no factor model (but e.g., a network model). I posted some papers which I found very enlightening. Perhaps you will, too.

2) Deleting items in an EFA CAN be problematic when these items solely measure single latent variable. Hence, deleting one item can imply deleting a meaningful latent variable. Factor extraction only applies to sets of variables (as the factor is a supposed underlying cause of this set). The extraction algorithm is not able to identify single-indicator-latent variables.

3) As I also argued, simply repeating a CFA after an EFA is weak evidence for the supposed measurement model. The CFA tests only the restrictions which were not involved in the EFA (fixed error variances and fixed double loadings). This is not useless but not a global test of the overall model structure. In addition, by using a new sample, the CFA tests also if the EFA-results were based ideosyncratic, non-replicable characteristics of the sample. Again, an important goal (like any replication attempt) but not strong evidence of the correctness of the model beyond sampling error.

4) You state that CFA will result in deleting many items. First: when you aforementioned that deleting items is a routine in an EFA, why do you think this is suddenly problematic when it comes to a CFA? But more importantly, why should a CFA imply deleting many items? And third: When you have 5 valid indicators of a factor, will deleting 2 of them "invalidate the measurement model"? I agree that having more indicators has some advantages, _as long as the measurement model is correctly specified_.

Please note that I do not intend to be harsh or unfriendly and that I formulated these sentences with deepest respect!

These are just my opinions and experiences. Perhaps, yours differ. I would like to know if this is the case.

With best regards

Holger

The CFA report require you to assess AVE for Convergent Validity and CR for Compiosite Reliability. No matter how much theory you have to support, it does not make sense if your AVE and CR failed to pass the mark.

The AVE and CR are computed based on Factor Loading. Low factor loading contribute to low AVE and CR. Most of the times, these value failed if the researcher did not filter the items through EFA and removed them prior to collectong the field study data for CFA.

Theoretical arguments is one thing, validity and reliability support from the empirical data is another. The most important thing is what you obtain must be valid and reliable. Meaning what you get did not your theory. You may have overlooked another theory. You know theory may change.

Dear Zainudin Awang

you did not clearly locate your answers to my #1-#4 points made above, so I will try to interpret you answers. If I mis-interpret something, please clarify my misunderstanding.

You said first

- - - - - - -

The CFA report require you to assess AVE for Convergent Validity and CR for Compiosite Reliability. No matter how much theory you have to support, it does not make sense if your AVE and CR failed to pass the mark.

- - - - - - -

I read this as "you have to carefully check if your model is appropriate", right? And I agree that theory is only the underlying rationale but not sufficient because the model can be nonsense. This is exactly what I proposed in my first posting: Test your model and adapt it when necessary. This should be done no matter where the model come from (theory vs. exploration). And it i no argument for a "must-EFA" perspective.

However, neither AVE nor CR are checks of the model structure (and certainly not a strong test of the model). They are measures of the measurement quality of the indicators. And yes, I agree, that checking the quality of the indicators is valuable but the more important goal in any model specification, is to check whether the model structure is correctly specified. Don't you agree?

If the model is nonsense, the extracted (or specified) factor may even be completely spurious and does not reflect anything meaningful/existing. While I agree that closely inspecting the factor loadings can help to judge whether they support the validity and interpretation of the factor, they can only interpreted based on the correct model. I do not see that calculating the AVE/CR can help in this regard. Wrong models can have high AVE/CR values and correct models can have low AVE/CR values.

For instance, as you yourself emphasize, AVE/CR are calculated based on the model parameters (loadings/error variances). Hence, if the model is misspecified, so are these parameters, leading to misleading AVE/CR measures. Hence, calculating AVE/CR does not substitute a focus on the model specification (and test) but does require it as its foundation. It is tragic that some fields (and you ) place their whole focus on these measures and not on the structure.

Likewise, a correct model with a meaningful latent factor can still have low AVE/CR measures. Again, testing the model focuses on the latent variable, AVE/CR focus on the indicator quality. I agree that one should strive for indicators with strong loadings and high CR (I am not so sure with regard to the AVE, as it confounds the loadings with the correlations between two latent variables), but it is a secondary goal. For instance, sometimes, there is a (expectable) conceptual distance between the indicator and the latent variable but the specification is nonetheless correct and the factor meaningful. Again, this has disadvantageous (the factor is difficult to interpret) but does not necessarily "kill the model". The loadings should not be maximized at all costs but match interpretational levels.

You then said:

- - - - - - -

The AVE and CR are computed based on Factor Loading. Low factor loading contribute to low AVE and CR. Most of the times, these value failed if the researcher did not filter the items through EFA and removed them prior to collectong the field study data for CFA

- - - - - - -

Agreed, but again this does not imply doing an EFA first, because you can eliminate poor loading items based on the (fitting) CFA. Simply picking items with higher loading from an EFA runs the risk of deleting meaningful latents (single-indicator-latent variables).

Finally, you said:

- - - - - -

Theoretical arguments is one thing, validity and reliability support from the empirical data is another. The most important thing is what you obtain must be valid and reliable. Meaning what you get did not your theory. You may have overlooked another theory. You know theory may change.

- - - - - -

I partially agree: I agree that validity is the *core issue* and concerns the correctness of the model and the meaningfulness of the factor and its specified causal effect on the indicator. Reliability has advantages (high power) but low reliability does not invalidate the model. Especially SEMs with measurement models allow estimating effects despite measurement error. By saying this, I interpreted your use of "reliability" specifically focusing on random indicator error. If you mean reliability such that the model must be replicable, than I agree. As I said, checking the danger of a idiosyncracy (sampling error) or robustness against different contexts is important.

Overall, you made valued points but - in my perspective - focus too much on indicator error instead of the model structure. EFA can be helpful but is not necessary for all of this and I could not find an argument defending this practice as a norm.

All the best,

Holger

Strong argument is one thing, poor factor loading is another. What if, you have strong arguments but your CFA produce poor loading for that items? You need to delete them. You know, deleting more than 20 percent of the items during CFA will cause it to be invalid?

That is why I said doing EFA is important to filter the rubbish items prior to doing CFA with field study data. I mean, collect pilot study data anf perform EFA. Then, collect field study data and perform CFA snd SEM.

Hello Zainudin,

I can only guess that you start your EFA by throwing a huge number of items with varying content and no clue which and how many factors may be involved. Only this can explain why you seem to have to delete such a huge amount of poorly loading indicators Again, in my personal view, this is another argument for start with some at least rough ideas about the potential factors and go to CFA. Starting with a conceptual idea leads to the creation of a focused set of indicators which may work well in the CFA or not. If they don't work well - again - my starting point is testing the structure. Loadings are only reasonable as a source of any validation check, IF THE MODEL HOLDS. Otherwise, they can be hugely biased. In the extreme, you may have highly valid indicators and - because their (biased) loadings seem to indicate invalidity- you throw them away.

When you say that

" You know, deleting more than 20 percent of the items during CFA will cause it to be invalid? "

I guess that the model fit was a desaster and not the poor loadings. As mentioned in my last post, sometimes low loadings is what one would expect and if the model holds, these indicators may well be valid indicators. In my point of view, you are too much concerned with "loadings as high as possible" but this is not always reasonable.

You can do whatever you want of course, but I think that creating an unfocused mass of indicators and hoping that the exploratory algorithm will do the work, will almost never lead to a correct model. And we haven't talked at all about that that overall idea of a common factor model holds at all is possibly wrong (see the network model approaches by Borsboom et al. or the sequence models by Bollen about indicators influencing each other...

Best,

Holger

Informative! this is even beyond the question raised by Mr Wanlin. Thank you Holger Steinmetz and Zainudin Awang this educated and Knowledge based argument provide a complete instrument validation process using EFA and CFA. I go through all the discussions and I got answers to many questions

Put my two cents in: here is how I did EFA using R: Article Factor Analysis by R Programming to Assess Variability Among...

Scholars suggest that adopted scales with sufficient empirical and theoretical evidence can be taken directly to CFA without running EFA beforehand (Hurley et al., 1997).

Hurley, A. E., Scandura, T. A., Schriesheim, C. A., Brannick, M. T., Seers, A., Vandenberg, R. J., & Williams, L. J. (1997). Exploratory and confirmatory factor analysis: Guidelines, issues, and alternatives. Journal of organizational Behavior, 667-683.

Hassen Houache But you know by doing CFA without running EFA data it would be challenging to some time because Data are changing due to the environmental factors such as time , location etc.

I do think that if you are developing a newly model and questionnaire you need to first conducting EFA to find out the loading of items to their constructs than CFA to evaluate the measurement of you model and scale while adopted questionnaire and model you need to conduct CFA Alone