Zero-adjusted Bray-Curtis dissimilarity not in vegan?

[MonaFuhrmann]

Hei,
I stumbled across the bray0 (x, ...) dissimilarity measure (R package ecole) which I am using before calling metaMDS on my community data. It creates a distance object of class 'dist', but as far as I can see its not implemented as a "method" in vegdist (similar to method="bray", etc.). Is there a reason for this?

[jarioksa]

vegdist is a vegan function that implements some distance and dissimilarity functions. There are other packages that implement other dissimilarities. It seems that ecole is one such package that implements other dissimilarities. These functions from other packages are not methods of vegan::vegdist and they cannot be invoked with a distance argument to vegdist. However, you have two choices:

1. Calculate first a dissimilarity matrix with ecole::bray0 and submit that to metaMDS instead of the raw data. If you want to add species scores, use vegan function sppscores after the analysis. However, you must check that ecole::bray0 returns dissimilarities and not similarities which would give grossly wrong results.
2. Use argument distfun = "bray0" in metaMDS call. However, this requires ecole::bray0 to have a method argument (which it may not have). Moreover, just like above, you must check that ecole::bray0 returns dissimilarities and not similarities which would give grossly wrong results.

[MonaFuhrmann]

Thank you! The Zero-adjusted Bray-Curtis dissimilarity (Clarke et al. 2006) basically just works on the raw data, by adding a dummy species to all sites, and then applying bra-curtis measure. Yes, my workaround is simply providing the bray0 dissimilarity matrix to metaMDS. My problem arises when I want to do some simulations with the vegan::oecosimu function which calls directly to metaMDS and needs community data which are then permuted. It doesnt accept the "distance" argument to be "bray0". Example:

oecosimu(comm = round(species.data), method = "quasiswap_count", nestfun = metaMDS, autotransform = FALSE, k = 2, distance = "bray", nsimul = reps, parallel = 2, statistic = "stress", alternative = "less")

There is probably another way of coding this operation, maybe somebody has some ideas?

[jarioksa]

You should not pass distance but the name of the distance function as distfun = "bray0" to metaMDS. I did not find ecole package in CRAN, but following the documentation I wrote my own version:

### bray0 of Clarke et al. (2006) J Exp Marine Biol & Ecol 330:55-80
### adds a dummy species of smallest non-zero value to every sampling
### unit and calculates Bray-Curtis dissimilarity. Please do not do this!
`bray0` <-
    function(x, ...)
{
    delta <- min(x[x>0])
    d <- vegdist(cbind(x, delta), method="bray") # ignore dots
    attr(d, "method") <- "bray0" # only needed for strict vegan compatibility 
    attr(d, "maxdist") <- NA     # ditto
    d
}

This could be passed as distfun in the oecosimu call. It makes sense to set trace = FALSE in oecosimu.

[MonaFuhrmann]

Thanks a lot! This fixes my problem!

But you advice not using this? > ("Please do not do this!")

My data has lots of 0s and is very similar to what Clarke et al. 2006 describe (gradients, few species, species absent from samples for similar reasons). An NMDS based on Bray_Curtis makes no sense (looks very collapsed, largely seperating a few samples). While when I apply zero-adj Bray Curtis the NMDS looks much more realistic and seperates communities along extepected environmentals.