Basically optimizing models is great, but it can have an impact on two things: Vertex lighting and texture size.
As it can be seen in the picture below, the window could have been just one plane with no subdivision but this would have implied being unable to have overlapping UVs.
1- The texture would have ended up being of either lower resolution or I would have had to restort to a larger texture size (the one below is 256x256, and I could in fact use a 128x128 since these are windows the player won't be close to, with some sharpening it's difficult to tell to difference. The panel object is one 64x64 texture).
2- Another problem would be the vertex lighting. While it might not be such a big issue with this window, the less vertices you have on an object the less accurate the lighting will be (unless the engine is using per-pixel lighting). If there was no light touching the bottom right corner of the window (and if the window was just a plane), 1/4 of the whole window would be almost pitch black.
EDIT: Oh and another thing, the more strips you have the more processing power is needed for the engine to "re-stitch" all the pieces togheter, so once again you must be careful here as well and not break it all into individual pieces just to save UV space. If you do this with all models it will definitly slow down the FPS. But to what extent is difficult to judge. You can always flip pieces onto each other without unwelding them when possible.
So basically, it's up to you to make a compromise between the number of tris the model will have and the texture quality+vertex lighting.
This window, with the two panels (which are separate objects so I can break the repetitiveness of the model by rotating each) is 168 tris. If I had optimized the model as much as possible tri-wise, I could have made this a 52 tris object instead. But then I would have needed a texture of at least twice the resolution.
Note: Sorry of the quality of the models, I'm still working on the textures:p