-The more hoops a candidate has to jump through, the more likely they are to bail out of your recruiting funnel. This is especially bad for college/postgrad recruiting when you aren't the #1 employer in your field. Everyone wants to work for the Googles and Facebooks of the world. It's hard getting someone to spend a couple hours for your startup job.
-People cheat. We usually issue a short coding project, grade for correctness, then do a code review over Skype or face-to-face. Many candidates turn in the exact same responses. I've even seen people cheat and have a friend do the Skype session with a totally different guy flying out. Do you proctor your test in a secure center? Use an online service to lock down their machine and record? Both are pretty invasive. Switching up the questions constantly is tough and makes your signal noisier.
-Industriousness and raw intellect trump skills/knowledge most of the time. Sure there's a baseline level of skill required to train someone quickly enough, like I wouldn't hire someone who didn't know basic data structures, but work-sample tests are often biased to those with a very specific background. I don't want employees who are great at doing what I need today. I want ones who will be great at figuring out what to do years down the line.