Are there alternatives to rarefying for alpha diversity estimation?

Hi Nick !

I disagree with the use of a rarefied table for alpha diversity estimation. Is there a better explanation why one should use a rarefied table for alpha-diversity estimates?

Normalization of some type is absolutely essential for alpha diversity estimates, since sampling depth has a profound impact on species richness (use the alpha-rarefaction method on your own data or see wikipedia for clear examples). Rarefaction is not perfect, but is quite standard in ecology. Other approaches do exist in QIIME 2 ā€” see q2-breakaway.

1 Like

As richness is a poor estimator of diversity, it is generally not useful while talking about alpha diversity. We have found Simpson and Shannon as robust estimators and more importantly these seem independent of sampling depth. So, I consider it is futile to use a rarified table for robust alpha diversity estimation, This off course excludes the poor richness estimator.

I agree, rarefied richness is a poor estimate of total diversity. That is what q2-breakaway attempts to solve. See this preprint (and the original breakaway paper cited there) for some more info.

I am not sure I agree that Shannon and Simpson can be reliably used without normalization, since both would be quite sensitive to singletons (which can often represent erroneous reads in microbiome studies, not real diversity). But I have not benchmarked this. It is certainly achievable in QIIME 2 if that is your preference.


It would be really useful to know (benchmark) how Simpson and Shannon estimators perform with and w/o rarefying mock communities. I am not sure if my previous paper, which demonstrates the robustness of these estimators (especially Simpson), fit into the issue here too. Thanks!

Mock communities would be one good way to test this ā€” I would recommend using as many different mock communities as you can to ensure generalizability over a range of conditions. We have a mock community repository that would be useful for this if you want to benchmark.

Simulated microbiome data should also be used, as mock communities will not really accurately reflect all natural conditions or species distributions, especially rare species.

Thank you for sharing that paper! It looks like that is based on TRFLP data? In which case Iā€™d say not all of the same findings would apply to amplicon sequence datasets. E.g., sequencing methods will be much more sensitive for detecting rare species, and discover higher diversity overall than TRFLP would, so I expect Simpson/Shannon would be less robust to sequencing depth than they would be to, e.g., TRFLP cumulative fluorescence intensity.

1 Like