My data:

Tracking forest communities (via species abundances) in various forest plots across time.

My approach: Non-metric Multidimensional Scaling ordination

I performed NMDS (using nmds() in the R package Ecodist), and then added change vectors to my ordination.

  • The ordination was performed using 1000 separate ordinations (i.e., argument nits was set to 1000), so I would think the results should be fairly robust

  • Note: each point in my resulting ordination represents a forest plot in a given year, and change arrows connect the points of each consecutive sampling year for each given plot.

I also performed a hierarchical cluster analysis (w/ flex beta), and subsequently colored my points in the ordination based on the cluster analysis results.

The resulting colored NMDS ordination (with change vectors) looks like this:

enter image description here

The point

So if you'll notice, the forest plot in the bottom right corner changes minimally in all but a single year. In that year, the NMDS point is drastically different than in all other years -- it's located all the way toward the top of the NMDS output plot.

Normally, I would see this trend and assume something drastic happened that year. I would assume that the plot's overall species characteristics matched the plots represented by red dots on the top more so than the plots associated with blue dots toward the bottom of the ordination graph in that one given year.

However ...

When I actually examine the raw data for this forest plot in the years leading up to, including, and after the massive shift in NMDs space, I find that basically nothing changed between years (especially between the drastic change year and the sampling years immediately before and after).

  • In other words, the same individual trees (of the same species and with the same relative distribution of sizes) existed in this massively different year as in the years before and after it.

My Question(s):

So I know that NMDS isn't perfect and can get stuck in "local" minima etc etc ... However, after running 1000 ordinations, I would think my results would be robust enough to avoid such issues. But maybe it's still possible?

  • So, is this anomaly possibly still due to some random chance issue in the algorithm itself?

  • If so, is there any way for me to determine the cause??

Lastly, I have re-run this ordination a hand-full of times (each time with slightly different input), but each time some plot seems to have a weird "outlier " point that doesn't make sense like above. So I'm beginning to think maybe this is something that I won't be able to eliminate or account for.

  • If that's the case ...

    • Is there a way to describe the phenomenon?

      • (the phenomenon that a random ordination point just won't cooperate for no apparent reason)
    • More importantly, is there precedence for simply excusing an anomaly like this as an artifact of the NMDS approach and just moving forward with the present result?

      • examples or suggestions of how to word such an explanation would be appreciated.
  • $\begingroup$ The ordination was performed using 1000 separate ordinations (i.e., argument nits was set to 1000) What's that - for somebody not using same software as you do? What is the model of your NMDS? One-matrix or repeated model or weighted model (INDSCAL)? $\endgroup$ – ttnphns Aug 23 '17 at 20:55
  • $\begingroup$ Also, I feel (maybe wrongly) that your questions would be hard to answer without being in your research shoes. $\endgroup$ – ttnphns Aug 23 '17 at 20:59
  • $\begingroup$ @ttnphns so essentially it's 1000 different starts of the NMDS calculation to help avoid the NMDS result finding local minima. I do not know what you mean in your second question (One-matrix or repeated model or weighted model (INDSCAL), so I cannot answer that). $\endgroup$ – theforestecologist Aug 23 '17 at 21:07
  • $\begingroup$ @ttnphns I don't think you need to be in my shoes to answer it. I've told you the "science" side of it: there is no data/scientific difference b/w the points in question, so the resulting "anomaly" must be due to some oddity in the NMDS algorithm itself and not due to the data. Because the data do not appear to be the problem, I've covered my bases as the researcher. Also, I'm more curious what others do with this situation: do they simply excuse it as a regular occurrence when using NMDS, or do they know the cause/solution to finicky anomaly points? $\endgroup$ – theforestecologist Aug 23 '17 at 21:10
  • $\begingroup$ UPDATE: By changing the seed (i.e., using the seed() function in R) and rerunning the code I was able to get the point to "correct" itself. So, for me, the solution was finding a seed that allowed the randomizations to sort things out on their own... $\endgroup$ – theforestecologist Jun 8 '18 at 14:30

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.