Warton and colleagues’ article has recently been highlighted on Faculty of 1000, a platform providing post-publication peer-review and selecting only the most important articles in biology and medicine. Just 2% of published articles are highlighted on Faculty of 1000 each month.

We strongly recommend reading this article because, beyond technical issues, it stimulates reflections on our consciousness of limits of statistical tools, which is often overwhelmed by our addiction to their routine application.

The article is also accompanied by a free application and a really great video:

This? Is amazing. I really have to think about doing a video like this for the R package I’m working up.

Also, question on mvabund – if you know causal pathways of interactions between species, is there a way to fold that information in? I come at this kind of data from a Structural Equation Modeling perspective, so, I’d take a problem like this and use SEM with multigroup analysis (or its GLM extensions). But, very cool stuff!

Thanks for the comment. Interesting question… mvabund can account for correlation between species when it constructs the test statistic (usually best done using the argument cor.type=”shrink”). But it estimates this correlation from the data with no a priori knowledge, it cannot estimate the correlation using a SEM-style model involving causal pathways etc. An SEM approach would be a very difficult problem in most cases because of the large number of species typically encountered – basically what you are after is a high-dimensional SEM for count data! Certainly would be very interesting if you can pull it off…

Yep – mvabund can handle any fixed effects design – unbalanced data, continuous predictor variables, mixtures of continuous and categorical, etc.

You didn’t ask about random effects, but the main type of design not handled by our software at the moment is designs with random factors. That is part of a broader problem – I don’t believe that designs with random effects are addressed adequately anywhere in the multivariate abundance literature in ecology at the moment. (“manyglmm”, anyone?)

You might need to incorporate covariances in the random effects, which will make thinks more complicated, as you can’t just stick together lots of independent analyses.

This is great!

The bad news is that the Journal of Applied Ecology gazumped us this week by getting themselves into The Onion. Bastards.

This? Is amazing. I really have to think about doing a video like this for the R package I’m working up.

Also, question on mvabund – if you know causal pathways of interactions between species, is there a way to fold that information in? I come at this kind of data from a Structural Equation Modeling perspective, so, I’d take a problem like this and use SEM with multigroup analysis (or its GLM extensions). But, very cool stuff!

Thanks for the comment. Interesting question… mvabund can account for correlation between species when it constructs the test statistic (usually best done using the argument cor.type=”shrink”). But it estimates this correlation from the data with no a priori knowledge, it cannot estimate the correlation using a SEM-style model involving causal pathways etc. An SEM approach would be a very difficult problem in most cases because of the large number of species typically encountered – basically what you are after is a high-dimensional SEM for count data! Certainly would be very interesting if you can pull it off…

That’s great ! Can the mvabund test handle unbalance or asymmetrical experimental designs?

Thanks

Yep – mvabund can handle any fixed effects design – unbalanced data, continuous predictor variables, mixtures of continuous and categorical, etc.

You didn’t ask about random effects, but the main type of design not handled by our software at the moment is designs with random factors. That is part of a broader problem – I don’t believe that designs with random effects are addressed adequately anywhere in the multivariate abundance literature in ecology at the moment. (“manyglmm”, anyone?)

You might need to incorporate covariances in the random effects, which will make thinks more complicated, as you can’t just stick together lots of independent analyses.