Result interpretation in vignette6
Closed this issue · 5 comments
Hi.. vignettte6 gives an interesting study about predictor variables to multivariate community data. how we can interpret the result in summary(CGLLVM)? some coefficient estimates give negative values and some are positive, for example, soil.dry(LV2) 1.5032 and fallen.leaves(LV1) -0.6957. Is it can be concluded that most spider species prefer in more dry soil and they did not prefer in many fallen leaves habitats?
Thanks.
Thanks for posting you question!
In general, I would be cautious with inferring on the effect of predictor slopes for specific latent variables.
In the example, the slopes for soil dry mass are both positive. In this case, I might interpret that as indicator of a positive relation between spider abundance and soil dry mass, though I am not an expert on spiders nor on soil dry mass! Soil dry mass reflects the organic matter content of soils, and is not a reflection of moisture content. It might, however, be an indicator of how productive a soil is.
Similarly with fallen leaves, but now we have a predictor that might reflect a measure of cover for hunting spiders. The slope is negative on both latent variables, so that there the model seems to indicate that there is a negative relationship between hunting spider abundance and places with cover? That does seem rather unintuitive, but I'm not a spider expert!
If you are interested in species-specific effects, those can (for a linear response model as is the case here) be retrieved as CGLLVM$params$LvXcoef%*%t(CGLLVM$params$theta), since these effects are likely to differ on a species-specific basis. Alternatively, you can have a look at the ordination plot to see how species relate to these effects ordiplot(CGLLVM, biplot = TRUE).
I hope that answers your question.
Thanks for your great explanations. I think the inference from ordiplot was more favorable. But, when I use my dataset, is it any explanation for why the result always changes if I rerun the analysis? is it because I set inappropriate lv.formula ? for example I just define lv.formula = ~DO + salinity + pH + temperature. and I observe that applying family="negative.binomial" produce a more inconsistent result than family="poisson". Is it must use family="poisson" in here?. I have no issue with using NB in mvabund package.
Also, is it possible to put a legend of environmental predictor gradient in ordiplot (for example legend for pH in vignette2)?
Unlike in classical constrained ordination methods, the algorithm used here is sensitive to the initial values. In case you know non-metric multidimensional scaling, think of it like that: you will have to re-run the analysis to find an optimal solution.
Jenni has written a paper on this, but there are some instructions in the vignettes and help pages too. Once you have found an optimal solution (use e.g. the n.init argument), the seed is stored in model$seed so you can reproduce your model, without having to re-run it multiple times frequently.
You can add a legend to your ordination plot by using the legend function in for base R plotting.
Details on the method are available in our preprint: https://t.co/XgW6qTR81w?amp=1.
Great. thanks for sharing the preprint and explanations.