Hamid.

There are no rigid values for Skewness and Kurtosis. What values did you obtain?

For example, some authors suggest that values should be between -1 and 1, others between - 1.96 and 1.96. yet others say that skewness could be up to 3 and kurtosis up to 10. In other words, without the specifics of your data it is dificult to assist you in finding a good answer.

Hamind

more information is needed. There is is no clearly defined criteria values for Sk and Ku.

For some it should be betwen -1 and 1, for other -1.96 to 1.96 yet other suggest that for Sk < 3 and Ku < 10.

More details on your data is needed in order to be helpful 

 Dear Dr José Vasconcelos-Raposo

Thank you so much, 

Ku for some items is 28, 24, or 20. and Sk is -4.

Dose it cause to dose not fit?  

Dear Hamid,

There are some issues you need to consider in your analysis like try to avoid the Little Jiffy practice (Kaiser, 1970), which is to use Principal component, varimax and K1 as methods in an Exploratory Factor Analysis (EFA). Also, it is recommended to use different samples for EFA and confirmatory factor analysis (CFA). If your items are Likert scaled, I suggest to use Factor software (Lorenzo-Seva & Ferrando, 2006) which alows you to use the polychoric correlation matrix and unweighted least squares as estimation method, which are recommended by Forero, Maydeu-Olivares and Gallardo-Pujol (2009).

The homogeneity of items and internal consistency may be used to establish the statistical behavior of items and if they should be eliminated or not.

For CFA I suggest Mplus or LISREL softwares, which allow you to follow the recommendation of Forero, Maydeu-Olivares and Gallardo-Pujol (2009) about to use de polychoric correlation matrix with Unweighted Least Squares as estimation method.

If you decide keep using AMOS and in case you have big sample size you can use Asymptotically Distribution Free as estimation method, which is robust in front of not normal distributed items.

If the results of CFA show low loadings for some items (below .40; Stevens, 2009) and/or no statistical significance (p>0,05) you should consider them to be eliminated from the model.

I hope some of this issues be useful,

Best regards,

Dear Hamid,

The answer provided by Edgardo Daniel Miranda Zapata is very useful for the technical complexities.  Here is a response about the more general issue - do you omit items from a scale that have been empirically shown to fit that scale? There are a number of reasons why related items might appear to load into a single scale, and not all of them are because they all measure the same thing. For example,

1) EFA can show correspondence when there is a causal relationship between items.  For example, if you ran an EFA between items measuring family income and items measuring student achievement, they would load together - not because income MEASURES achievement,  but because income IS STRONGLY RELATED TO achievement.  So, check your items to see if they measure Xscale or are related to Xscale.

2) EFA is typically run on an assumption of continuous scaling, so that higher-with-higher indicates association.  If you have seriously skewed measures that are, in fact, reflecting categories, the results could well be misleading. You can perform a transformation on most types of skewed measures to make them more normally distributed and try them again in your CFA. Alternately (if transformation doesn't work, as in a solid floor or ceiling on the item), you can categorize the measure and try some of Edgardo's solutions for CFA.

   What you do with the problem items depends entirely on why they are a problem in your scale.  You can remove items IF the reason they are a problem is that they do not in fact measure Xscale. But if you cannot make a strong **conceptual** argument for why measures A, B, C, and D do not measure Xscale, you will get dinged on review for not including them.

Which comes down to - you can remove them if you can also provide an argument for removing them that your readers will accept.  If you can't, you have to figure out what the argument is.

Dear Dr Edgardo Daniel Miranda Zapata and Dr Julia B. Smith 

thank you for your responses.

Hello.

The question is how did you test conformity? You can use tje other way around, meaning test of deviance.

In principle I don't use to try to force reality into a model if it does not dit smoothly

First, your choice of EFA may be inappropriate if your items are at least moderately correlated.  In this case, principal axis factoring with an oblique rotation is necessary and appropriate.  Further, varimax rotation may also be inappropriate if a higher order factor is presumed to emerge.  In this case, use an oblique rotation such as promax. 

Second, if loadings are below .30 to .40 (depends upon researcher as some say .30 (Schmitt, 2011) and some say .40 (Gorsuch, 1983, I believe) then they may be eliminated. You may also eliminate items that are cross-loaded.  You should eliminate items based upon your EFA analyses. Of course, you can 'confirm' you final models using CFA, which I now address.

Third, if your data are non-normal, then it would behoove you to use Robust Maximum Likelihood CFA, which can be obtained via Mplus.  Amos will be problematic for this purpose.  You can also use R Code but this can be cumbersome due to data entry and problems with the program reading copied and pasted glyphs.

Good luck

Dear Dr Florent Lheureux 

Thank you for your comment. I will run by Principal axis factor analysis. I should be assess data based on PAF. 

I think bootstrapping procedure would help on this.

I am agreed with others answer. But if you complete all those process systematically  then you can delete few factors loadings for your final model fit. Because during the EFA some questions can be loading very low less than .3 and all factors can be deleted .