Embedding collusion secure fingerprints as watermarks depends on several factors, not only the robustness and imperceptibility.
First of all, one of the main issues of collusion secure codes is the length, so you should think about the payload in the same scale as robustness and imperceptibility. Of course, the length the depends on several factors such as size of coalition, number of users, max error, and even the algorithm (if you're using Tardos codes, or the Boneh & Shaw approach, etc...).
The more levels you use, the more robust the watermark will be but at the same time, the less data available to embed you'll have. For example, if your image has a resolution of 1024x1024 pixels, then decomposing it to a DWT to embed the WM in the LL sub-band will give you only an image of 512x512 pixels to do so (and now the payload constraints make all the difference).
From my own experience, while using DWT transform to embed fingerprinting codes as watermarks in images I used only a single level DWT transform and embedded the watermark in the L-L sub-band of the image.
If your images are somewhat small, I suggest you then to use 1 level decomposition, because DWT embedding provides great imperceptibility (even on the 1st level) but if you have large images and/or the fingerprinting code is not so long then you can go for 2 or more levels.