If the data collection is verbose enough, dead reckoning from speed/acceleration information can be used to unmask some non-driving information. It can't be generally unmasked completely, but even small correlations, like "they start slow in the morning but go faster after xx seconds" -> "they live near a freeway" -> <regional information> -> "they are slightly more likely to be rich" -> "they under-report accidents" can be inferred, leading down the same path. This is the sort of thing that I was indicating above -- a lot of "good" drivers can be lumped in with the "bad" ones to make Class B, if you can establish an attributes that appears to be causal but whose accuracy is low.