Rendering a frame is a complex issue. The render time is highly dependant on what you have in the scene, what renderer features you're using, what kind of simulation stuff (if any) you're using etc. If you can break down your scene (in your mind) into the various tasks that the renderer has to do, then you know what areas you can optimize and how...

It's all about the amount of data vs the computing power at hand. Triangle count is the first place to start with... including nurbs/subd/displacement tessellation settings... From the renderer's point of view everything is tessellated to triangles at the end. The fewer you have, the faster it'll render.

Texture and shadow resolution is one critical point too like Caligraphics said. Same goes with raytracing. If you know that your image looks ok with only one reflection ray, then there's no point on allowing the renderer to do 10 reflections.

Again repeating what Caligraphics said; use light linking whenever it might make sense.

Another often overlooked thing is the antialiasing / sampling quality settings. These might be a little hard to understand but tweaking them can really improve the render times. Read the documentation on these buggers...

Also with MR, tweaking the BSP settings can improve the render times when raytracing.

With huge scenes with lots of high resolution textures and perhaps a massive particle disc cache the speed of you hard disk might come into play as well. If it takes 10 seconds per frame to load everything for one frame, then for a 60 frame animation that's one extra minute to wait... an like we all know, 60 frames is only two seconds of animation in NTSC format...