For example, rows are sites and columns are species, data is count data of those species in those sites. Can I use PCA to analyze the species composition in those sites or is it more robust to use NMDS?
I thought since there is no linearity between species abundances then I couldn´t do a PCA because it violates the linearity assumption. However, I read a few articles that use PCA with count data.
Species counts within a variable can be treated as continuous if the range of count values in a variable is large so that the addition or subtraction of an individual is a small proportion of the total. So, if a species has counts from 20 to 200 then the data for that species is effectively continuous (even though you can't have a value of 34.73 individuals).
If species counts are not like this then there are better alternatives. That is, if the counts are between 20 and 35 (or zero and 10 say), or there are many zeros.
Under these circumstances, if you still want to do PCA, you can: 1, leave out species with many zeros or counts of low range or 2, combine species counts so that the totals meet the criteria outlined above (find a sensible criterion for combining - such as species from the same genus or family or 'all Diptera with aquatic larvae').
Or you can use Categorical PCA (CatOCA) after reducing your counts to categories (discretize the count data). I'd suggest using categories based on the quantiles rather than the standard deviations.
NMDS, however, is a better option if your counts have many zeros.
It may not be the best though. I suggest that you look at Manly's "Introduction to Multivariate Statistics" as a starting point and also, for ideas that may be better for your data, papers like A. Ramette (2007) Multivariate analyses in microbial ecology, FEMS Microbiology Ecology, 62, 142-160.
good luck, Andrew
Species counts within a variable can be treated as continuous if the range of count values in a variable is large so that the addition or subtraction of an individual is a small proportion of the total. So, if a species has counts from 20 to 200 then the data for that species is effectively continuous (even though you can't have a value of 34.73 individuals).
If species counts are not like this then there are better alternatives. That is, if the counts are between 20 and 35 (or zero and 10 say), or there are many zeros.
Under these circumstances, if you still want to do PCA, you can: 1, leave out species with many zeros or counts of low range or 2, combine species counts so that the totals meet the criteria outlined above (find a sensible criterion for combining - such as species from the same genus or family or 'all Diptera with aquatic larvae').
Or you can use Categorical PCA (CatOCA) after reducing your counts to categories (discretize the count data). I'd suggest using categories based on the quantiles rather than the standard deviations.
NMDS, however, is a better option if your counts have many zeros.
It may not be the best though. I suggest that you look at Manly's "Introduction to Multivariate Statistics" as a starting point and also, for ideas that may be better for your data, papers like A. Ramette (2007) Multivariate analyses in microbial ecology, FEMS Microbiology Ecology, 62, 142-160.
good luck, Andrew
Thank you!! Combining abundances in similar groups sounds like a good idea. I will look Manly´s book.
Thanks for answering ;)
- Badges
- Science topic
- Similar topics
- Mathematics
- Statistics