Its more detailed than what I can adequately explain in the limited time I have, but you would be wise to go find an article that talks about the distribution of depth values that results from the near and far values you choose. The naive assumption most people make is that you want near to be as near to 0 as possible, and far to be as far away as practical, but IIRC, that generally gives less than optimal results.
Also, as vstrakh said, you can change them on the fly as you render different parts of your scene. For example, some games render very far-away objects as a means to enhance the "skybox" (perhaps because the objects are animated, dynamically lit, or are far away but near enough for perspective to matter) and those would probably benefit from a different distribution of depth values. Its also a handy technique in space games, where the distances are great and the space is sparsely populated. In general, the maxim seems to be that this is helpful when you mix very large objects at great distances together in a scene with relatively smaller objects at much smaller distances. I imagine that the borders between level-of-detail transitions might also make a good candidate, but it probably depends on how near or far the borders are.