What do you mean with "which show gamma oscillations"? If you mean that the percentage values should be modelled bya gamma distribution, you can use a GLM of the gamma family. But percentages are usually modelled as a beta-distributed random variable, so one would use a beta-GLM or a standard linear model based on the logits. But if "oscillations" should indicate that you have time-series data, it would be important to describe the experimental design and the research hypothesis in more detail.

Jochen Wilhelm First, thanks for your reply. Second, I am doing electrophysiology on hippocampal slices. Some of these slices show activity in the gamma range, others don't. I have many groups treated with different drugs. So, how can I compare the percentage of the slices giving the desired effect (let's say gamma oscillations) from one group to the other taking into consideration that group sizes are different. For example,

CTL group size is 15 and 90% of them shows gamma, 5% beta, 5% no activity

A group size is 13 and 30% of them shows gamma, 70% no activity

B group size is 30 and 10% of them shows gamma, 30 % beta, 60% no activity

and so on ....

So which statistical tool should I use to test the significance of the treatment

Hello Amr Elgez. Given your description, you might be able to use a Chi-square test of association. But one important requirement is that each slice is independent of all other slices, even within groups. Are the slices independent in that way? And how many groups are there? Can you share the Group x Activity contingency table with us? This would allow readers to work out the expected cell counts, etc. Thanks.

... or GLM of the binomial family. It might be a mixed model to consider some non-independence as mentioned by Bruce.

As I understood it, you want to predict if Gamma, Beta or no (specific) activity occurs (next question: how did you define "no activity"? Is this an ERP experiment or what else?). Therefore, you might consider a multinomial logistic regression. But I am not sure if it will produce some problems, if one or more groups do not have a realisation of all DV possibilities (e.g. Group A, which has not Beta activity).

Anyone an idea?

If you are still at an exploratory stage of your project, then you would want a visual presentation of your data. For this, you can take account of the different sample sizes for your percentages by using one of the variance-stabilisation techniques. These exist for the binomial distribution or, if in your case the counts are relatively small, those for the Poisson distribution might be better for you as being less technically distracting.