0
$\begingroup$

Edited following helpful feedback.

I have vegetation species data for a number of grassland habitat sites, and am preparing to begin Exploratory Data Analysis.
Data was collected in 100 quadrats over 9 sites in percent cover to attempt to establish one or more species which indicate improved quality of grasslands over time through their presence or abundance (%).

Quadrat   Site   Soil.type  Mean.quadrat.height.cm  Species.1%  Species.2% Species.3%...
1          2       3               40                  5           10         25
2          3       2               50                  0           0          10
3          2       1               25                  10          0          0

I am confident in the steps I need to take to ensure rigorous data analysis, and I'm not attempting to shirk any long winded jobs, but I wanted to see if I should be reducing the number of species variables before going further. I have over 100 separate species, some of which only occur on one occasion. Certain papers I have seen recommend removing any species which only occur in <5% of quadrats, but I don't know how to back this up. To be on the safe side should I conduct EDA (boxplots etc.) on every single variable, or is there a robust method of species reduction? I had originally thought PCA should do the trick but some literature I have read says other forms of EDA should come before PCA.

Thanks

$\endgroup$
9
  • $\begingroup$ Not everybody here is familiar with vegetation research. Please formulate your question such that it is clear for as many people as possible. What exactly is your data? You have over 100 species. You "measure" them in different places? Are there "quadrats"? What are these measurements? How many places are there? For complete clarity, can you maybe show a small part of your data table? And finally: what is the purpose of the analysis? $\endgroup$
    – amoeba
    Jan 8, 2015 at 11:46
  • 1
    $\begingroup$ EDA has nothing to do with dimensionality reduction (other than giving you useful information that you should use in subsequent phases). So, after you're done with EDA, proceed with one of dimensionality reduction approaches. If you don't care about latent structure of your model, then PCA should do the trick, otherwise you need to perform factor analysis (at least, EFA, but, potentially CFA as well). $\endgroup$ Jan 8, 2015 at 13:12
  • 2
    $\begingroup$ @AleksandrBlekh: This is an aside, but I find the strong distinction you are making between PCA and FA ("if you don't care about latent structure") not really warranted. See my answer to the question "Is there any good reason to use PCA instead of EFA?" (search for it and it pops up immediately; I intentionally don't give a link, as the questions are not very connected, whereas giving a link makes them officially "related"). $\endgroup$
    – amoeba
    Jan 8, 2015 at 14:31
  • 2
    $\begingroup$ @AleksandrBlekh: It is a complicated issue with many people having strong (and often opposite) opinions about. So take my word as just one opinion. I did not mean that PCA and FA are approximately equal; they can produce quite different outcomes (but it is very unlikely with 100+ species, as in the OP here). But I insist that they are not so different conceptually as many people often think. $\endgroup$
    – amoeba
    Jan 8, 2015 at 16:28
  • 1
    $\begingroup$ Thank you both for your interesting discussion. I believe from your discussion that it sounds like further research would be beneficial here, and so if time allows I will use both FA and PCA on the same data and compare the results. I think I will cover all bases and perform EDA on all variables without further strict guidance. Many thanks. $\endgroup$
    – Sam Hall
    Jan 9, 2015 at 9:52

0

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service and acknowledge you have read our privacy policy.