Dear Basma,

in general I am against excluding outliers. Data is data. If you're data is too noisy you should add more animals to increase your n.

The only time I think it is reasonable to exclude animals if they do not swim but just float. These animals are not even trying. Normally this only happens to 2% percent of the animals. If it happens more and it is not evenly distributed between the two groups it could be considered a phenotype.

best

Nils

Thank you doctor .@ Nils Borgesius

yes i have already increased the n of the gp such that each experimental gp is about 15, but there sometimes weird numbers in the data set , that do affect significance , i dont know if they are true outliers or not ? should i remove or not?

so there is no exclusion criteria in water maze .like if an animal never reached the platform in training and hidden platform should i exclude or not.

There is no exclusion criterium other than not swimming as far as I know. However, there is a maximum time, usually 60 seconds. When an animal has not reached the platform in that time it is put on the platform by the experimenter during training. During the probe trial the animal is simply taken out and the latency is recorded as 60 seconds.

Better not to exclude, as NB said "data is data". You may consider:

1) analyses of individual data (randomization and permutation tests

2) nonparametric approaches

3) analyses that take into account the variability of data (they give more weight to the average score, less to that of an outlier, e.g. multilevel modeling

4) instead of usual central tendency measures, you may think about for example trimmed means in following analyses

5) if you decide to exclude outliers, always check if you would arrive to the same conclusion without excluding outliers

6) this might be helpful if you decide to exclude outliers:

http://www.ucl.ac.uk/~smgxprj/pdfs/statout_v0_1_0_jou.pdf

7) always report the screening procedure!