In my data for acoustic features of songbirds, I have unequal sample size across groups. I had been using Sigmaplot 14.

In 6 birds, I have given aCSF (control), 100 and 200ng/ml of drug. Of these one bird did not receive 100ng treatment. For each dose I did the song recording 3-4 times and analysed the songs in SAP (sound analysis pro) and got the data (acoustic features like pitch, frequency etc). So the problem is that the number of chosen (clear noiseless) syllables (specific sound notes) for analysis are different for each bird in each group of treatment resulting in highly unbalanced data. Please see the attached table for syllables per bird per group.

So by experimental design, it is repeated measures and RM ANOVA should be performed. But for RM ANOVA rows should be equal for each group and if performed anyway would compare only rows equal to the smaller group size. Furthermore, this data is always non-parametric and does not pass equal variance test. Therefore the option left is RM ANOVA on ranks (Friedman’s test) but that also requires equal group size. The option now left is to ANOVA on ranks (Kruskal-Wallis). Is that a better test? I heard that this test is often misused or not preferred although I have used it in one of our paper (Kumar et. al., 2019) on similar data.

Another option, I found that in Graphpad, Mixed effects model does not mind missing values/data. Although this is not really the case of missing values in repeated measures but the sample size is different for each group and for each bird within that group. I do not want throw data just to make group size equal or near equal. Mixed effects model can run stats on this data, but I have my doubts since it is designed as a replacement for RM ANOVA to test for fixed and random effects whereas it deals with missing values by some complex regression analysis. But having that much of unbalanced data, I am bit hesitant using this method whether it would compute the stats well. Additionally mixed effect model sometimes doesn’t not calculate and give me F and P-values with a message that data is not that complex and mixed effect model might not be needed (see the screenshot of Graphpad attached). Also attached the data sheet (excel file) of acoustic data for pitch-goodness if somebody wants have a better look for trying.

Thanks