 |
 |
Hi Fiona, The wilcoxon test will be fine. The test does not compare medians, it compares differences in mean ranks, so don't feel obliged to mechanically report the medians just because that's what's normally done. In an ideal world, if you had more mistakes, I would use something like a Poisson regression, but it's going to be complicated, because you will almost certainly have zero inflation, and that means you really need to use a negative binomial regression. And then it's repeated measures, and that means either a sandwich estimator or a multilevel model. And that would make me stop and think whether I really needed that model, and I get excited about this sort of thing. I would be tempted to say that the data are so horribly skewed, that people either made mistakes, or they didn't make mistakes, so they are coded 0, 1 and it's a sign test. (And sign tests are the world's easiest statistical test. You don't need a computer to do a sign test, in fact you can do a sign test without paper. In your head.) Was it the same people, or different people making errors? jeremy On 23 February 2011 14:04, Fiona Essig <[log in to unmask]> wrote: > Hi all, > > Apologies in advance if this is a silly question - I'm not at all good with > stats so have probably confused myself unnecessarily with what should be a > simple question. I have a dataset with 28 participants and I want to analyse > whether the number of errors made in two versions of the same task are > significantly different. The type of error is unusual, so there aren't very > many instances of it - most people don't make this type of error, as > follows: > > Task A - 3 of the participants make errors (a total of 5 errors) > Task B - 6 people make errors (a total of 24 errors) > > Would it be appropriate to carry out a Wilcoxon signed-rank test? Obviously > as most people score zero errors in both tests they are excluded from the > ranking - is that problematic? Also, it looks quite dodgy reporting a median > of zero! I've been faffing around with this for ages - I just don't know how > best to deal with data where most people have no score. I'd be very happy > going with a non-parametric test if I can - I've had enough of it! > > Fiona > > -- > Tel: 01707 284 761 > E-mail: [log in to unmask] > > Fiona Essig > Research Student > Room E384 Research Huts > School of Psychology > University of Hertfordshire > College Lane > Hatfield > Herts AL10 9AB > -- Jeremy Miles Psychology Research Methods Wiki: www.researchmethodsinpsychology.com