skewness and kurtosis

[Bin Dohry]

hi
in the case of the high value of skewness and kurtosis statistics, what should I do (winsorize) and is the accepted value for skewness and kurtosis (0 and 3) or (3 and 10 ) respectively
thanks

[Felix Bittmann]

I would suggest looking at the actual distribution first, using a histogram. This can easily tell you if there are any severe problems or aspects you should consider.

[Carlo Lazzaro]

Bin:
the answer to your query also depends on what you have planned to do with your data.

[Bin Dohry]

thanks for your reply Mr. Carlo and Mr. Felix
I have panel data for 15 years and will run SGMM regression the number of observations more than 3000 but some of the results I get not sig. so one of my friends suggests me to correct the non-normality. I also stuck in the level of the Winsorize that I should shoes, for example, some variables can corrected at 1% but if I use 3-5% my result will move from sig. at 0.10 to 0.01 strongly sig.

[Nick Cox]

Where does it state that SGMM (what does S mean here?) requires the marginal distribution either of the response or of any predictor to be normal? Your skewness and kurtosis values are features of the data, not problems to be corrected.

[Bin Dohry]

Where does it state that SGMM (what does S mean here?) requires the marginal distribution either of the response or of any predictor to be normal? Your skewness and kurtosis values are features of the data, not problems to be corrected.

thanks for your reply.
Yes, skewness and kurtosis values are features of the data. so can you explain for me why we winsorized the variables that suffer from non-normality?
" system GMM " i just use "S" to refer for system and D to refer for different with the GMM model.

[Nick Cox]

Sorry, but I don't know why you Winsorized your data or why you expect me to know.

[Carlo Lazzaro]

Bin:
It seems that you're hunting for the "best" (whatever that means) model. Unfortunately, data are what they are and any make-up that you can try increases the risk of obtaining a subsample which is miles away from the original one.
A general advice is to give a fair and true view of the data generating process, considering non-significant coefficients as informative as significant ones..

[Bin Dohry]

Bin:
It seems that you're hunting for the "best" (whatever that means) model. Unfortunately, data are what they are and any make-up that you can try increases the risk of obtaining a subsample which is miles away from the original one.
A general advice is to give a fair and true view of the data generating process, considering non-significant coefficients as informative as significant ones..

thanks for your advice. as I mentioned above that I get an advice from my friend, but I think as you said better to present the result as the original to give the real of what we get.