negative numbers and NMS

[T. McCulloch]

Hi Everyone,

I am attempting to use a published index with negative numbers in an ordination (NMS) which is not working well.  The index is from a PCA, and is a leading EOF.  Is there a data transformation that I could use that would retain the information contained in the index, or the information that the index conveys?

Sincerely,

T.

[Bruce McCune]

T:
I think you are saying that you have a data matrix that was produced by PCA, containing negative numbers, and you are attempting to run NMS on it. If not, please clarify. Also, can you clarify the size of the main matrix and the distance measure that you selected?

Bruce McCune

--
You received this message because you are subscribed to the Google Groups "PC-ORD" group.
To unsubscribe from this group and stop receiving emails from it, send an email to pc-ord+un...@googlegroups.com.
For more options, visit https://groups.google.com/d/optout.

[T. McCulloch]

Hi Bruce,

One of the variables in the data (second or explanatory) matrix were produced by PCA, and does contain zeros, positive, and negative numbers.  I have been reviewing Table 6.2 in Analysis of Ecological Communities (2002) trying to determine which distance measure would work with the negative numbers, and trying each distance measure in turn for those listed as "all" under the Domain of x column, with little success.  The main matrix is 10x6, and the second matrix is 10x19 (containing the index created by PCA).

Sincerely,
T. 

[Bruce McCune]

It sounds like you have data on a variety of scales, so that probably needs to be addressed. Having no other information I would suggest relativizing by standard deviates of the columns (i.e. each matrix element is re-expressed as standard deviations from the column mean), then using Euclidean distance.

Good luck.

Bruce McCune

[T. McCulloch]

Hi Bruce,

Thank you for your reply.  I did give that a try, with little success.  However, you are correct regarding the second matrix data being on a variety of scales.

Sincerely,

T.

[T. McCulloch]

Hello All,

I have revisited my species matrix, which has a predominate species in all years and one year is predominate of all other years.  I am able to generate a 3D manual NMS of the species matrix, however the autopilot mode is unable to find a solution.  Relativization by maximum for the species matrix has not found a solution in autopilot mode either, even if there are no outliers.  The average skewdness is low with relativization, however individual species or years can still indicate skewdness greater than 1, and some greater than 2.  Only species that were observed in 25 (5%) or more occurrences are included, resulting in a matrix with just a few species, and low beta diversity.

Any suggestions or pointers would be  greatly appreciated,

Sincerely,

T.

[Bruce McCune]

T., when you say that one species predominates and one year predominates, it sounds like the column totals would be high for that one species and the row totals would be high for that one year. But are years in separate rows (i.e. rows are sample unit-year combinations) or are years in separate columns? (i.e. columns are species-year combinations). I'm guessing the former.

In any case, if your data matrix has only a few columns, then it is not surprising that NMS autopilot can't beat the randomization test, because you should be able to get a low-stress solution with any data matrix with just a few columns, even if they contain random numbers. And it is even easier if there aren't very many rows.

It sounds like your matrix may be small enough that you should skip the randomization test and instead focus on getting a low stress biologically reasonable solution.

Bruce McCune

[T. McCulloch]

Hi Bruce,

Thank you again for your reply and thank you for the advice regarding the autopilot mode.  Your first guess is correct, rows are years and species are columns, and values are average species density per year.