I've done some abx tests lately, among them lame vs. other codecs or wav, using WinABX (thnx, great program BTW). When comparing short passages starting at loud points of the samples, sometimes the beginning of the decoded mp3 sounded different to the original wav, I think because of encoding-decoding process adding some samples at the beginning of the file. I tried using --nogap and there was no more difference.

My question: Can I use --nogap for ABX testing or could this introduce problems?

[EDIT]
Nevermind. I noticed using CEP that when encoding and *decoding with lame commandline* there are no samples added or cut at the beginning of the file, no matter if --nogap is used or not. My fault was that I used fb2k (0.43) + totalrecorder for decoding sometimes, mainly because I wanted to apply crossfeed in some cases. Decoding that way --nogap does make a difference. Anyway, if I want crossfeed for abx testing on headphones, it's a better idea to apply it before encoding I think, so I'll do this in future and live happy ever after.
[/EDIT]

If I'm not wrong, some mp3 encoder/decoder combinations introduce a slight offset if the audio files, so that the decoded wav file is a little bit displaced in time over the original wave file, which could be the source of your problems. I believe this is not the case if you use LAME both to encode and decode the files, in this case there should be no offset. In other cases with offset, the only way to fix this actually in WinABX is to manually adjust the decoded file with a wave editor to remove the offset. In case of LAME, I think the --nogap swith won't make anything, and is not neccesary as long as you decode the mp3 also with LAME (using a frontend such as EasyLAME - RazorLAME is fine too).

Edit: your edit read.