Dear Karl Lundström,

I think you can try for a diet similarity analysis using cluster analysis, SIMPER and ANOSIM in PRIMER.

Regards,

Sileesh Mullasseri

Thank you Dr Sileesh Mullasseri

Would you the recommend doing a 3-way ANOSIM (or 3-way PERMANOVA) using year, season and areas as factors, or would you recommend me to divide my data into more balanced groups and do separate comparisons e.g. between years for the season I have more samples from and from one area at a time, and e.g. also compare the two areas separately only for the years (2) and season (1) I have samples from both areas? Or can I look at all the samples and factors at the same time, e.g. using a 3-way crossed ANOSIM or 3-way PERMANOVA? I am interested in getting an understanding of how my data may differ between years, seasons and areas and then see if I can motivate pooling samples e.g. from several seasons and/or from the two areas. The samples are opportunistically collected and I don't have samples from all years, seasons and areas. The material is biased to a few years and 1-2 seasons. From one of the areas I have samples from several years and seasons (but still very unbalanced), from the other area I only have samples from two years and from the same season both of the years.

I doubt that you can compensate by statistical means the inherent bias in your data. I used physical means (Folsom Plankton splitter) to accurately count thousands of planktonic organisms. There are no short-cuts.

Your experimental design must relate to the frequency of occurrence of what you are trying to sample.

Best to start again!

Sandy

Sandy Bartle Thank you.

Could you perhaps develop a bit further of what you mean? Not sure that I understand...

ANOSIM is good for skewed data. But you must look at all samples and factors at the same time because of interactions between factors and to avoid bias. If there are only slight differences among sample sizes between treatment combinations, you could impute missing values using the mean or median of the treatment levels. However, this approach should be used with caution and should only be used when sample sizes are nearly equal to begin with. If the sample sizes are not equal and the assumption of equal variances is violated, you could instead perform a non-parametric equivalent to an ANOVA such as the Kruskal-Wallis test :)

Hi Andrew,

Thank you for shining some light into dark places! You expressed the issue much better than I did. Statistical methodology cannot compensate fully for poor sample design. But your suggestion is helpful.

Regards,

Sandy

ANOSIM is pretty robust to a lack of balance in a design. You may find this relevant (and it includes a fish diet study):

Article Analysis of similarities (ANOSIM) for 3‐way designs

Are your samples individual fish? In which case some judicious pooling is probably sensible.

Thank you Paul Somerfield . I'll have a look at that paper.

Yes, my samples are individual'fish" (actually seals).

If your model is too unbalanced I recommend using a PERMANOVA, just don't forget to separate well the categories of the samples, either dam type, sites, and years. Also if you could do an nMDS before doing your ANOSIM or PERMANOVA it would show you a clearer idea of the key categories in your sample data.

Look at this article

Hey Karl, just an idea here. Could you subset your dataset to have a less skewed data for analysis? I have done this in the past where I needed a more balanced dataset (i.e. equal number of samples for each factor), but in reality I had many more samples for specific seasons and places. I did 100 random subsets of my dataset, only allowing a maximum number of samples per factor to be extracted for a analysis at each subletting, so to have a balanced dataset. At the end I summarized the results of 100 tests and these were my main result. This allows you to analyse your data including all factors and without throwing away samples.