Need more information here to even make a guess. Are you comparing three groups? If so, what constitutes "locomotion" data?

Indeed, are your locomotion data continuous or variable? Do you have single or repeated measurements? You need to describe shortly your experimental design and, equally important, what is your research question. I assume is something along the line of "is there any difference between wild type and mutant flies?".

Hey ,

So here is the experimental design.

I have a mutant that shows locomotion defect compare to wild type and when I did a genetic screen I found out that there are some other mutants that can rescue this locomotion defect.

I am taking Ten flies of each genotype and measuring its locomotion activity (i used a pteri plate and it has defined lines and I am measuring the lines crossed by the flies in two minutes) three times each and then plotting a box plot. I need to do statistics and need to show if the rescue of locomotion is significant. I used a t-test but the reviewers said its not an appropriate test. what would you guys suggest?

Could you tell us what was the objection of the reviewer(s) (beside rejecting t-test)? Did (s)he/they provide any clue about why a t-test would not be appropriate? It might be related to the small sample size, because 10 is quite small (but the t-test WAS developed for small samples) or to the repeated measurements design. The small sample size will increase the bias but I am not convinced that another statistical test will overcome that small sample bias. It is also not clear to me whether you compare 30 measurements with 30 measurements or 10 average measurements with 10 average measurements.

they said it's not a good test to use for multiple comparisons. they advised using a Non-parametric test. I have an average of 3 measurement so 10 measurement for each sample.

Robert's questions are very important and his comments about Gosset's t are dead-on, yet I am still not sure of the question. In this case, "multiple comparisons" would typically refer to a situation in which there are multiple contrasts between groups, and the need, therefore, to handle alpha appropriately with that in mind. Yet, from your last post, I think that I see two groups, thus only one contrast.

Now, it seems that you do not view this as repeated measures as you are averaging the three measures; this is for you to decide, and involves your research question more than statistics. I would have no difficulty using a t-test with an N of 10 in each group, unless there is another problem. For instance are there distribution irregularities? I say this because I see a recommendation of a non-parametric test. If it is so simple as heterogeneous variances, use of a Welch-type t-test is a solution.

Finally, if all else fails, you can avoid all statistical distribution theory by making your own distribution with a re-sampling plan such as bootstrapping.

I fully agree with John Morris's thought, as if they were mine. It is indeed very important (although some statisticians challenge that idea) to control for the alpha spending in the case of multiple comparisons, but here it is not the case. It sounds as if a reviewer knows something about the danger of multiple comparisons and type I error inflation and applies it in the wrong. place. But a non-parametric test is in no way a remedy for multiple correction. It may be for non-normal distribution. A permutation test or bootstrapping would be indeed such a non-parametric test and would be valid (but still it has nothing to do with multiple contrasts).

I should add that the Mann–Whitney U test (aka Mann–Whitney–Wilcoxon (MWW), Wilcoxon rank-sum test, or Wilcoxon–Mann–Whitney test) is the simplest and classical non-parametric alternative to the (indepdent) t-test.

Thanks Robert, I think we are kindred statistical spirits.

I would add that an added complication, that I do not see discussed much, is that adoption of a nonparametric test due to distributioinal difficulties, caries with it the fact that the alternative test is testing a different null hypothesis. This may be OK, and in some cases perhaps unavoidable, but the researcher needs to realize this. Then too, as I suggested previously, the alternative resampling meaning of "nonparametric" can be invoked allowing testing the hypothesis originally intended.

Indeed, John. And an additional aspect is that almost every non-parametric test comes at a cost in power (i.e. they tend to be less powerful than the parametric equivalent).

Quite true, but a difference in power testing different hypotheses is a bit difficult to catalog. We usually think of differential power in respect to testing the same null.

In any case, I believe that knowledge of these issues with choice of a nonparametric test are not always clearly attended to; unintended consequences often are not. So, I judge this of importance to note.