3

I am associated with a teaching hospital, and medical students are frequently administered tests and measures to monitor their progress. I am not talking about exam results here, but of psychological measures that do not figure on their transcripts. For "ethical" reasons, researchers typically need to annonymize the results. To be specific, I currently have "burnout" measures from a class of first year students. I also have "burnout" measures from a class of third year students ... the same individuals, two years later. I want to compare the results, but I don't have student ID's. A matched pairs analysis would be nice, but I don't have the matching variable.

The two samples (year 1 and year 3) are dependent, since they involve the same individuals. A two-sample t-test seems dodgy in this case. I imagine that a permutation test might work, but I'm not sure which one to use. Any ideas?

Placidia
  • 13,501
  • 6
  • 33
  • 62
  • [This same question](http://stats.stackexchange.com/a/3639/1036) had an interesting response as to make whatever treatment effect you are interested in by arbitrarily matching the groups to minimize the effect. – Andy W Jul 24 '14 at 17:32
  • 2
    `two-sample t-test seems dodgy in this case` Please say why you think so. For me, using an independent-sample test when you actually know the samples are paired, is fair, only it is less powerful. As long as it is right to stand for H0 (and againts own wish to find effect), it is just a conservatism, not an error. Another approach can be based on random matching, but it is asymptotically the same as independent-sample test. – ttnphns Jul 24 '14 at 17:38
  • @ttnphns Good point about the t-test being conservative in this case. The dodginess, in my view, enters when you build the related confidence interval around the mean difference. Presumably the variance estimate should be smaller. – Placidia Jul 24 '14 at 17:43

0 Answers0