Hi Hamid, 

I never use the "Eigenvalues > 1" rule to decide how many factors I should extract. Despite course books presenting it as a 'basic' method, I find it almost always greatly overestimates the number of factors you should consider. 

I usually run a parallel analysis instead, which gives me an estimate of the number of factors much closer to what I would choose just by inspecting the scree plot. This usually gives me few very interpretable, mostly stable factors (and, well, it's better than inspecting the scree plot because it's not as subjective). 

Regarding your cumulative variance: 23% is not high at all. However, what you'll want to do now is 'knock off', one by one, the items which have no saturation > .3 on any of the factors you are considering. Always run the EFA again, including the parallel analysis (if you do decide to use it), after removing each item. 

Next, you'll probably want to remove the items with a similar saturation on more than one factor (say, .456 and .412), as it's unclear which latent they are an indicator of. 

Your goal is a simple structure: each item should saturate on one, and only one, factor; and each factor whould have at least three indicators (it will help you down the way, believe me). 

Always use an oblique rotation - I see no reason not to - and don't be afraid to reduce the number of factors if a factor falls below three indicators. Oh, and remember to check the distribution of your variables before running the EFA - you don't want highly skewed data biasing the results, or items with too little variability -  and check that no item pair has a high correlation (I usually use .7 as cutoff, but this depends on your application). 

EFA is as much art as is science, and the steps you will take are often subjective. For this reason, once you determine a simple structure, check its fit using a CFA on an *independent* sample. The factorial structure you'll obtain with the EFA is unreliable is you don't perform this step with success. 

Hope I've been of help - let me know if there are some steps you need some clarification on. I've ran a lot of EFAs and CFAs and I know how much unstable and frustrating they can be :)

Hi, 

Good question to discuss.

In EFA you have different options to explore a new scale. One of them as you did, to get the factors with eigenvalues more than 1. however, as you get 14 factors,, and except the first three gave acceptable variance, the next step is to run EFA with forced number of factors, you can try 3 and 4...,,

It is important to examine the items loading and the items meaning in each factor. The loading is recommended to be .4 and above, clean loading [not shared with a diffrence less than .2 among the other factors].

You also can try to do different types of rotation to get the best items over factors distribution. Read the result and then compare among different analyses, Choose the best one which you get convinced more.

Regards

Hi Hamid, 

I never use the "Eigenvalues > 1" rule to decide how many factors I should extract. Despite course books presenting it as a 'basic' method, I find it almost always greatly overestimates the number of factors you should consider. 

I usually run a parallel analysis instead, which gives me an estimate of the number of factors much closer to what I would choose just by inspecting the scree plot. This usually gives me few very interpretable, mostly stable factors (and, well, it's better than inspecting the scree plot because it's not as subjective). 

Regarding your cumulative variance: 23% is not high at all. However, what you'll want to do now is 'knock off', one by one, the items which have no saturation > .3 on any of the factors you are considering. Always run the EFA again, including the parallel analysis (if you do decide to use it), after removing each item. 

Next, you'll probably want to remove the items with a similar saturation on more than one factor (say, .456 and .412), as it's unclear which latent they are an indicator of. 

Your goal is a simple structure: each item should saturate on one, and only one, factor; and each factor whould have at least three indicators (it will help you down the way, believe me). 

Always use an oblique rotation - I see no reason not to - and don't be afraid to reduce the number of factors if a factor falls below three indicators. Oh, and remember to check the distribution of your variables before running the EFA - you don't want highly skewed data biasing the results, or items with too little variability -  and check that no item pair has a high correlation (I usually use .7 as cutoff, but this depends on your application). 

EFA is as much art as is science, and the steps you will take are often subjective. For this reason, once you determine a simple structure, check its fit using a CFA on an *independent* sample. The factorial structure you'll obtain with the EFA is unreliable is you don't perform this step with success. 

Hope I've been of help - let me know if there are some steps you need some clarification on. I've ran a lot of EFAs and CFAs and I know how much unstable and frustrating they can be :)

Very good advice from Marcello. Plus, be sure that you are not using Principal Components (it is the default in some programs), especially if you are going to do oblique rotation -- which I also recommend.

When dropping items, another thing to look for (in addition to items that do not clearly load on any one factor) is factors that contain only one item, usually with a very high loading for that item. This occurs when an item is not related to the others in the set, so it is given its own factor.

Finally, your sense of looking at the rate that the explained variance "falls off" is a useful approach known as a scree test (although it is not really a test in any formal way).

 Dear Dr Muayyad Ahmad , Dr Marcello Passarelli, and Dr David L Morgan

Your exploration are very good. In continue, I used other rotations.  According to the quartimax rotation and to remove items with communalities < .5, 16 items loaded in four factors. (with 63% The total cumulative variance explained). 

Quartramax is still orthogonal, try oblimin for oblique rotation (i.e., correlated factors).

The corpus of best practice EFA literature generally suggests that rudimentary techniques such as the Kaiser criterion, visual scree test, and/or proportion of variance explained by eigenvalues  should be supplemented with more robust measures for determining the number of factors to retain such as minimum average partials and parallel analysis. I generally use MAP and PA to supplement visual scree and if the results diverge, run the different models and examine the significance and theoretical coherence of the loading's to determine which extraction model to retain. There are syntax programs available in SPSS for both MAP and PA and if you use the R integration package, you can access them directly in the GUI.