This kind of data has a different kind of selection bias.<p>When you're looking at user actions, you can't discern two different kinds of behavior: (1) The user is doing what he <i>wants</i> to do; versus (2) the user is doing something because he can't figure out the alternative, or that alternative is inconvenient/unwieldy.<p>So to a certain extent, the results of such studies can server to entrench bad design decisions, rather than improve the system. The developers <i>always</i> have to look for every possible alternative explanation for a user's behavior.