Regression in ordiplot(<rda-class>, scaling=1)

[jarioksa]

The problem was found in reverse depedency tests for BiodiversityR examples:

> ### Name: PCAsignificance
> ### Title: PCA Significance
> ### Aliases: PCAsignificance ordiequilibriumcircle
> ### Keywords: multivariate
> 
> ### ** Examples
> 
> library(vegan)
> data(dune)
> Ordination.model1 <- rda(dune)
==== clip ====
> plot1 <- ordiplot(Ordination.model1, choices=c(1,2), scaling=1)
Error in match.call(definition, call, expand.dots, envir) : 
  formal argument "scaling" matched by multiple actual arguments
Calls: ordiplot ... plot -> plot.rda -> NextMethod -> plot.cca -> match.call
Execution halted

A short example is:

ordiplot(rda(dune), scaling = 1)
# Error in match.call(definition, call, expand.dots, envir) : 
#  formal argument "scaling" matched by multiple actual arguments

The bug does not occur in the release branch cran-2.6-6. It does not influence cca but it influences dbrda and capscale.

The problem was bisected to commit ff75668 but fix may be needed in a file that was not touched in that commit.

[jarioksa]

The error really seems crop out on line 81 of plot.cca:

    dots <- match.call(expand.dots = FALSE)$...

which was added in ff75668 to handle dot arguments. However, this seems not work:

• all plot.cca arguments that are not ordiplot arguments will cause an error with rda, dbrda & capscale (but not in cca): scaling, const, hill, correlation.
• ordiplot arguments that are not plot.cca arguments are ignored in plot.cca also in cca but do not give an error: arrows, optimize.

Seems to need re-thinking.

[gavinsimpson]

Shouldn't ordiplot have scaling, hill, correlation, and const? Right now it doesn't seem like you can change what scores get plotted except on which axes, and what types of scores, and ... is not being passed on to scores() in ordiplot(). (I appreciate that this is not really the source of the problem here, but still.)

[jarioksa]

I don't think so: scaling, hill, correlation and const only make sense with cca methods which are a marginal case for ordiplot and in most cases these arguments would be ignored. With "cca" and "decorana" classes we short-cut to plot.cca and plot.decorana and skip most of ordiplot, but for ordiplot proper and most ordination methods those arguments make sense.

[gavinsimpson]

Ignore the above I missed the second part of the conditional

if (inherits(ord, "decorana") || inherits(ord, "cca")) {

[gavinsimpson]

I have pushed an alternative fix to issue-685 branch as discussed in cafdc03 - I need to test this though - was there a specific test you were running to check this @jarioksa? I'm not 100% clear on how the new code all works for the piping behaviour. I think optimize worked where I passed it through ordiplot() on the example that failed from BiodiversityR, but it would be good to get a second opinion on that.

[jarioksa]

I used only ordiplot(rda(dune ~ Management, dune.env), scaling = 1) in git bisect (found later that rda(dune) is sufficient). I found the problems with optimize and arrows later and incidentally and didn't test or serch them. However, if you run R CMD check you should get tests for most cases with pipes and configurable plots.

[jarioksa]

I checked the use of NextMethod in vegan and it seems that everyone of them incorrectly uses .... They should be fixed.

$ grep -n NextMethod R/*R
R/add1.cca.R:12:    out <- suppressMessages(NextMethod("add1", object, test = "none", ...))
R/anova.prc.R:22:        NextMethod("anova", object, ...)
R/biplot.rda.R:12:        NextMethod("biplot", x, ...)
R/drop1.cca.R:11:    out <- suppressMessages(NextMethod("drop1", object, test = "none", ...))
R/goodness.metaMDS.R:5:        return(NextMethod("goodness", object, ...))
R/plot.cca.R:147:    NextMethod("plot", x, ...)
R/print.cca.R:118:    NextMethod("print", x, ...)
R/print.mso.R:4:    NextMethod("print", x, digits = digits, ...)

[gavinsimpson]

Some checking suggests that optimize passing from ordiplot() -> plot.cca() currently is not happening with NextMethod().

Rereading ?NextMethod suggests that my description in the commit comment is not quite correct and what you are doing should work, but the evidence is to the contrary and scaling etc are getting duplicated. I'll need to look into this further when I get back from the UK after the weekend.

[jarioksa]

The logics and arguments of ordiplot and plot.cca are non-concordant. If plot.rda is changed so that dots are removed in NextMethod("plot", x) then in the following command

ordiplot(rda(dune), scaling=2, type="t", bg="white", optimize=TRUE)

scaling, type and bg are passed to plot.cca, but optimize is not. However, in

ordiplot(rda(dune), scaling=2, type="t", bg="white", spe.par=list(optimize=TRUE))

optimize is passed to species in plot.cca. The arguments differ in the following ways:

• type is an argument of both ordiplot and plot.cca and it is passed.
• optimize is an argument of ordiplot but not of plot.cca and it is not passed.
• spe.par is not an argument of ordiplot but it is an argument of plot.cca and passed.
• neither ordiplot nor plot.cca have argument bg and it is passed as a dot argument. It is an argument of text.ordiplot and will be available there.

Inspect the following command:

ordiplot(rda(dune), scaling=2, type="t", bg="white", spe.par=list(optimize=T), arrows = TRUE)

This will be passed to plot.cca as:

### with NextMethod("plot", x) or empty NextMethod() match.call()
plot.cca(x = ord, choices = choices, scaling = 2, type = type, 
    xlim = xlim, ylim = ylim, spe.par = ..3, ... = pairlist(bg = "white"))
### with NextMethod("plot", x, ...) sys.call()
### match.call() fails with multiple definitions, args after ... are dot args
plot.cca(ord, choices = choices, type = type, xlim = xlim, ylim = ylim, 
    ..., scaling = 2, bg = "white", spe.par = list(optimize = TRUE))

argument	ordiplot	plot.cca	NextMethod no dots	NextMethod w/dots
choices, type, xlim, ylim	yes	yes	pass	pass
scaling, spe.par	no	yes	pass	pass as dot arg with conflict
bg	no	no	pass as a dot argument	pass as a dot argument
arrows/optimize	yes	no	do not pass	do not pass

The minimum nuisance strategy is to add optimize and arrows among arguments of plot.cca and have NextMethod without dots

I focussed on plot.rda with NextMethod, but ordiplot(cca(dune), ...) works like NextMethod("plot", x). That is, it works OK but does not pass optimize and arrows which are not among formal arguments of plot.cca.