iiinova — where your digital self keeps evolving

Reposted by Bert van der Veen

Rob Barber @ra-barber.bsky.social · 25/09/2026

Hi all, I've revived my BlueSky account to share info about our ecological monitoring survey (more below). I don't have many followers, but I'm hoping our survey could be really useful for the UK monitoring community. So please pass on to anyone that might find it interesting! arcg.is/1mn48e5

arcg.is

UK Terrestrial and Freshwater Monitoring Survey

113

Bert van der Veen @vdveenb.bsky.social · 20/09/2026

@bobohara.bsky.social I'm told this is rare. Unfortunately, it's Thundurus.

111

Bert van der Veen @vdveenb.bsky.social · 15/09/2026

Drinking coffee while standing in front of a white board! Potentially in a white lab coat for extra sciencyness.

220

Bert van der Veen @vdveenb.bsky.social · 13/09/2026

dbRDA minimizes the squared euclidean norm, NMDS doesn't. That's good according to some people 😀. It gives you a different ordination.

010

Bert van der Veen @vdveenb.bsky.social · 13/09/2026

For all those infidel unbelievers that don't use model-based ordination, and want something else than Canonical Correspondence Analysis.

110

Bert van der Veen @vdveenb.bsky.social · 13/09/2026

By extension, you can concoct a method for constrained NMDS. It's lived in my head for a while, but now there's some code for it too github.com/jarioksa/nat....

github.com

natto/R/caxNMDS.R at master · jarioksa/natto

An extreme vegan package of random, experimental and alien code - jarioksa/natto

120

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

It is true through that in a GLLVM the species loadings are (literally) components of variance and covariance, which I suppose relates to your point, although you still really don't need to compute it on your way to the data.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Apologies, I did misquote you. To go from the data to the scores you don't need to compute variance across communities. That may be true for dissimilarity based methods, but not for GLLVMs.

110

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Likewise, I am happy to help you understand why some of your statements on double-zeros, the connection of scores in the model and data, and distances, among others, don't hold up, if you are interested in the nitty gritty details at some point. Feel free to drop an e-mail.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Indeed a relevant question, but the wrong way of answering it. Double zeros do not impact a model-based ordination in that way; the mean-variance relationship takes care of that. Richness is predicted from the model directly, so you can check the model and ordination against it directly.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Why do you think that the scores cannot represent the data correctly? This is provably incorrect for many ordination methods, especially model-based ordination.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

This also counts for the ordination: the site scores and species loadings directly reflect the data. The (euclidean) distance in the ordination does that too, but in a very indirect manner, which makes it much less useful.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

The ecology is not reflected by a dissimilarity, it is reflected by the data itself. So I don't need to go through a dissimilarity, as there is a one-on-one connection between the model and the data, and that is what matters.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Fit is measured from the model to the data, which is affected by the model's structure, response distribution, and the link function. Why would I look at distances, if I can straightforwardly measure and check how well the community data is reflected by the model? It doesn't need dissimilarities.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

I missed this, and the answer is yes, but not in the way you seem to want it. The ideas around dissimilarities in classical ordination are not transferable to model-based ordination, because the concept of "ecological meaning" isn't measured by dissimilarity.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

As an aside, this is more of a mathematical given, than that it is a point of discussion. The only exception being Gaussian responses.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

I'm happy for you to disagree, but then bring your point a bit further. If you think that variation in the dissimilarity is relevant, please explain to me how it traces back to species-environment relationships and assembly theory.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

You seem to confuse those. Model-based ordination does not place a distance on the data, that is where the real gain is. Even if there is a distance in the ordination, you cannot analytically trace that back to a distance on the data.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

No worries, I'm clearly persistent myself. I'm pretty sure I agreed that there's an implied distance in the ordination space, that is not the point. What really needs differentiating between is the distance on the data, and the distance in the ordination.

200

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

The discussion should focus on how to use a method or model to capture the data generating mechanism that best represents dynamical ecological processes. We would do good to remember in this discussion that dissimilarities have been a means to an end, but are not ecological processes.

000

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Hence, to compare a model-based ordination to existing dissimilarities seems like a moot comparison, as one cannot prove that dissimilarities capture ecological processes, and the comparison sets a flawed baseline to begin with.

200

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Dissimilarities discard all information on species, and communities are made up of species, so there is no connection to an original data generating process. Those dissimilarities have also been proven to fail without possibility of diagnostic and I don't see a reason to hold onto that bygone era.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Feel free to explore such a comparison, if that convinces you. In my mind, this argument is the other way around: on occasion, dissimilarities approximate the statistical process accurately that is represented by a certain distribution (corresponding to the view in Warton et al. linked before).

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Community ecology should be moving away from dissimilarities, not reinforcing it. The distribution in a statistical model is based on information theoretic grounds, which in my view is much stronger backing than ordination from "dissimilarities with well-known properties".

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

True. My definition of multi-species doesn't cover GDMs, as they don't return anything on the species.

010

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

But this is not a matter of distances. This is a matter for accommodating properties of data, so that the ecological signal in the data is captured by the model.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Because the method can predict, fit to the data is also straightforward to quantify. In essence, there are a range of statistics and tools that can be used to that end.

100

Bert van der Veen @vdveenb.bsky.social · 10/09/2026

Indeed, this is a far statement. If the model is misspecified, the ordination may not carry much meaning. There are indeed ways to verify this; there are limited suitable choices for the distribution of the response, but residual diagnostics can be computed for validation.

110

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

That makes it weirdly akin to model-based ordination, where you choose the distribution and model to accommodate data properties, instead of a dissimilarity. That has incredible benefits, such as flexibility, diagnostics, and more, but I think it's clear where I stand on that. #gllvm

000

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

So at zero stress, the ordination space has the same rank order as the data dissimilarity.

000

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

NMDS estimates the scores by minimizing the normalised square root of squared differences for the distances in the euclidean space, and the (estimated) distance constrained by the data's rank order.

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

For the record, this is how NMDS works. You choose a dissimilarity, and NMDS estimates the ordination scores in a chosen coordinate system (for vegan this is euclidean).

231

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

The ordination space is always euclidean, and that's completely fine, just as in NMDS. Unless you're saying NMDS also doesn't reflect ecological dissimilarity?

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

The connection between the ordination and the data isn't set by a dissimilarity as in a classical ordination, but by the chosen distribution for the data. The validity of that choice is what determines whether the ordination accurately reflects the community, together with the model structure.

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

If you change the space, the coordinates are estimated to reflect that. As I said, the dissimilarity of the space is irrelevant, and I'm not sure why you're convinced otherwise.

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

That's just not the case. You seem too caught up in the way that classical ordination methods are considered.

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

I understood the question as "can you change dissimilarity to accommodate the data".

000

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

It may require reconsidering the concept of "ecological dissimilarity". The model makes the ordination represent the community in a best fit manner, akin to how NMDS does. It's completely irrelevant what coordinate system this is in, but euclidean is convenient. Happy to chat further.

000

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

The distances are represented by coordinates implicitly, but the coordinates are directly estimated without distances. Naturally, an ordination has distances, if that is your point. Any issues that classical ordination has (double zeros, arch, whatever you can think up) are resolved by the model

110

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

If it helps, you can take vegan's NMDS as an example: dissimilarity on the data is not euclidean, but the ordination is in euclidean space.

100

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

That's the problem with euclidean distances calculated from data, not with an ordination in euclidean space.

100

Reposted by Bert van der Veen

Bob O’Hara @bobohara.bsky.social · 09/09/2026

Your question seems to be predicated on the idea that dissimilarity indices are ecologically meaningful :-) The distances will be in Euclidean space, so we won’t need other indices. The ecological interpretation should be as good/bad as any other JSDM/GLLVM.

111

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

Instead, they work more similarly to GLMs: a distribution and model structure are chosen to match the properties of the data.

000

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

Multi-species models don't work with dissimilarities at all; they're a fundamentally different concept incompatible with dissimilarity-based thinking. See, e.g., www.researchgate.net/profile/Davi....

researchgate.net

300

Bert van der Veen @vdveenb.bsky.social · 09/09/2026

Same richness, but different composition. That's the point of a multi-species model: richness and composition are connected, and should be analysed as such. Come to my ##NSOGFÖ2026 workshop next week if you want to learn how to do this with #gllvm! nordicsocietyoikos.glueup.com/event/joint-...

030

Bert van der Veen @vdveenb.bsky.social · 08/09/2026

This is written a lot more politely than my LinkedIn criticism of the paper.

110