RRPP and geomorph updates

[Mike Collyer]

Dear geomorph users,

This announcement is in response to the thread started by Lucía last week, regarding the inconsistencies she discovered between lm.rrpp and lm.rrpp.ws ANOVA results, when using within-subject RRPP.  I found the issues, which I can best explain by noting that there are three possible types of analysis: (1) all observations are independent; (2) observations are not independent because of repeated observations on subjects, but subjects are not terms in the linear model; (3) observations are not independent because of repeated observations on subjects, and subjects are terms in the linear model.  A bug in our coding made types 1 and 3 mutually exclusive, so if type 2 was sought, it reverted to type 1.  I have updated RRPP to better discern among these types and I believe the results are now consistent.  Updates are current on Github and I plan to send an update to CRAN within a few weeks, pending some other needed updates.

Additionally, both RRPP and geomorph have been updated (and should be reinstalled from Github) with respect to Box-Cox power transformations for estimating effect sizes, Z.  The reason for updates in recent developments with likelihood ratio tests revealed that RRPP-generated sampling distributions could be strongly negative or strongly positive.  A Box-Cox transformation must be performed on positive values, greater than 0.  A typical step in transformation is to find y’ = y - min(y) + c, where c is a positive value, to shift all values to be positive.  We discovered this was a good idea not only if values were negative, but also grossly positive, which log-likelihoods could be, so our functions had a few kluges to deal with atypical vectors of statistics.  These updates did not work well, however, for statistics constrained to be between 0 and 1, like PLS correlation coefficients.  After some investigation, we discovered that the best solution (In terms of consistency) is to first find z’ = z - min(z) + 1, where z is a vector of the standard deviates of y, as it standardizes the initial variance of statistics prior to the power-transformation,  For most applications, this will have no obvious effect, especially if many RRPP permutations were used.  In some cases, it might change Z-scores, slightly.  In obscure cases, previously illogical Z-scores should now be more reliable.

We decided to elevate a box.cox function to RRPP so that it is available to users.  It differs from the MASS::boxcox function by using the standardization described above.  The RRPP and geomorph packages have been updated to accommodate this change.

I apologize for any hassles that the coding bug caused.  Other updates should be coming soon.

Best,

Mike