IMO, it doesn't seem remotely comparable to a coin flip because the odds aren't 50/50. The dogs are only brought out in certain circumstances when they think a positive might be possible, meaning that lots of people who would be negatives aren't tested and thus not included in that 93%. The cases making up the 93% only consist of people whom they thought may get a positive, meaning the sample is biased to having lots of positives.

I'd compare it to the fact that lots of very hard certification tests actually have fairly high pass-rates (80% or above), simply because people who don't think they could pass don't take the test to begin with - The 80% pass-rate is not representative of the average person's chances of passing the test, nor the difficulty of the test itself.

What's really important is the percentage of those positives which were true positives, vs. false positives. In this case, you would expect the number of positives to be high - The number of true positives vs. false positives would tell you how accurate the dog actually was in the positive case. If this number is also high, then the dog is doing a reasonable job. Of course, to derive a real argument from this, you'd need to know the number of true and false negatives as well, but obviously we don't have that information.

Sure, because a coin is evaluated on its capacity to map to an even distribution of outcomes. Drug possessions, however, are almost surely not an even distribution. If, say, 95% of the people that the dog evaluates have drugs, then raising an alarm 93% of the time is probably pretty good.

In other words: If we wanted a drug dog to be as close as possible to a coin flip, we'd just flip a coin instead.