Abstract: We developed a web-based tool to provide students with access to old exam problems and solutions. By controlling the order in which students saw the problems, as well as their access to solutions, we obtained data about student learning by studying old exam problems. Our data suggest that in general students learn from doing old exam problems, and that having access to the problem solutions increases their learning. However, the data also suggest the depth of learning may be relatively shallow. In addition, the data show that doing old exam problems provides important formative assessment about the student’s overall preparedness for the exam and their particular areas of strength and weakness.
In particular, they were trying to address these questions:
How much do students learn from doing a multiple-choice exam problem and getting feedback about the correct answer? Does providing students with a complete solution improve their learning? How accurately does student performance on practice exam problems predict their performance on actual exam problems? In particular, can it help predict areas that a student should spend more time studying?
Going over the paper, I first had a chuckle when looking at Fig. 1 which shows that most of the students tried the practice exams within 24 hours of the exam! Typical last-minute cramming! :)
It appears that these are "multiple choice" tests. Although the authors don't think that this may make a significant difference, I'm not so sure about that, because one also can't discount random selection, or at least, a student making an educated guess without actually knowing the correct answer.
Still, it's an interesting study to read.
 W. Fakcharoenphol et al., Phys. Rev. ST Phys. Educ. Res. v.7, p.010107 (2011).