http://web.mit.edu/xiphmont/Public/theora/demo7.html
Check out the last graph on the page: when measured using PSNR, Theora tweaked to the max is roughly 50% worse than x264's default settings.
This now applies somewhat to Theora as well (in current SVN); they ported over a similar optimization to x264's Variance Adaptive Quantization, which dramatically increases SSIM but lowers PSNR.
Regardless of what video codec you use, the rate-distortion greatly depends on the encoder software and how you use it. When you use MEncoder's x264-compression for example you can tell it to do multi-pass encoding, more accurate motion estimation, use less keyframes, more frames for pixel prediction, ... . Many people don't even use multipass encoding when they convert their videos.
Update: Here's an interesting website about work on the theora encoder http://web.mit.edu/xiphmont/Public/theora/demo8.html