vegandevs/vegan

…lots

…l-master

if 'permutations' has wrong number of columns, R can crash or
permutations can be wrong. This is a user error, but easily done,
and we should not crash R but handle user errors more smoothly.

issue #239 in github

Now we assign to the parent.frame(). As orditkplot() is only used
interactively, this is the same as the global environment, but
passes --as-cran tests.

- further modernization of adonis2 object: Ybar must be in the top
  level, and it is safe to use initDBRDA instead of hand-made Gower
- anova.ccabyterm failed when all terms were not in scope (e.g., in
  models with interaction terms)

anova(cca ~ A1*Management, dune.env) failed because only one term
A1:Management was in scope, and the same for all other case where
scope did not include all terms in marginal models (also in adonis2).

- anova.cca by = "terms" uses directly permutest(..., by="term") and
  runs mainly in compiled code (hence is faster)
- by = "margin" redesigned: no update() of formula, no partial term
  (hence robust against scoping problems), works with NA data
- by = "axis" uses now 'forward' tests instead of marginal (which
  were biased) at the price of being 3-4 times slower
- anova.cca lists of models mangled long model formulae
- adonis2 bug fix: was not adapted to new permutation design
- model.frame is more robust with subsets
- model.matrix finds result directly from the result object and avoids
  scoping problems; works even when model.frame fails with scope
- ordConstrained gained new method="pass" that can refit a model instead
  of update() of formula (which has scoping issues)

Added average distance matrix function.

Partial models gave inconsistent results in dbrda() because
ordiYbar(..., "CA") did not know input was distance-based. This
also appeared in anova(..., by="axis") which uses partial models.

earlier we used the supplied method argument, but after allowing
method = "pass" this was insufficient and made distance-based
methods wrongly analysed. Consequently anova by = "margin" and
by = "axis" were wrong. It was also necessary to make permutest.cca
identify distance-based methods from the x$method item as an
alternative to class.

This is somewhat hacky, and more fundamental change may be needed
if ordConstrained(..., method="pass") is  used more widely.