You are basically looking at different reference points and vectors to derive the voltages. I would not see the value in artifact rejection in one montage then analysis in another.   Wouldn't consitencey be better?  Most artifact is lessened in an averaged referenced montage.   It would seem that you should have a theoretical basis for such a change, as you are going to generate a ton of data;  always worry about p vales relevence to  the amount of statistical tests.

Dear Gaffney and  Alexander,

I have re-referenced my data before from earlobe to average reference but i did not see much difference in the artifact rejection. I wonder the effect of changing the recording montage from one another does it have the same effect with re-reference? Does the average reference reduce p value?

You can think of rereferencing as a multiplication with a matrix (i.e. a linear operation). Hence, if you're chaining linear operations, the order doesn't matter (for example, a linear filter and rereferencing can be switched without affecting output). If you're using nonlinear operations (some artifact rejection/reducing techniques are nonlinear, e.g.) then the order matters. 

Appears that your questoin  asks if you can 'convert' the data after artifact rejection.  I would say sure, you have simply eleimaterd dirty epochs.  I would think you could manipulate the data any way you wish after cleaning it up.

1) Filtering

2) Segmentation

3) Artefact Rejection

4) Montage Operation

Thanks for your answers dear colleagues...

@ Gary & Ahmed: sorry, I didn't understand your point about the p-value dealing with montage convertion in EEG

@ Paulo: Have you probed to do montage convertion in other order? for example: 1) Filtering 2) Montage convertion 3) Segmentation 4) Artifact rejection ?? Do you think that there would be any difference?

Yes, it is the form used by eeglab for processing EEG data.