Prelim abundance JSDMs

Summary

• Used the same variables as for occurrence, with the addition of slope.

• The models parameterised with a negative binomial distribution showed extreme predicted outliers that were well outside the scope of the original data.

• A full poisson model predicted unexpectedly well onto itself, although there was a tendency to under predict.

• Predictions onto new data were inaccurate, even with the poisson model

• Environmental variables explained a greater proportion of variance than they did for occurrence, which is unexpected.

Covariates

Using colinearity and at least 5% relative influence in BRT models as criteria, 15 variables were chosen. The same set used for occurrence data, but with slope.

fin_abn_df <- dat_abn %>% 
  select(1:10, 
         ph_et_al,
         salt,
         carbon,
         Ca,
         Q_cover,
         elevation,
         percent_over1,
         percent_over2,
         P,
         N_perc,
         aspect,
         texture,
         corr_dN,
         drainage,
         slope)

Negative binomial scaled model

A negative binomial distribution was first used as the data are overdispersed.

Diagnostic plots looked okay.

## NULL

Predictions negative binomial

There was a general pattern where there model would predict most plots fairly well, but then have a few predicted points that are orders of magnitude greater than anything in the dataset. This suggests that the model is misspecified somehow, despite the residuals looking okay.

Using two sites to predict onto the third showed similar outliers.

Maybe there are just a few plots that are causing the outliers?

Doesn’t look like it.

which(pred[,2] > 20) #R_comptonii

## [1] 21

which(pred[,5] > 20) #A_fissum

## [1] 125 132 135

which(pred[,6] > 6000) #A_framesii

## [1] 51 70

which(pred[,8] > 60) #C_staminodiosum

## [1] 126 133

which(pred[,9] > 150) #Dicrocaulon

## [1] 51

which(pred[,10] > 100) #Oophytum

## [1] 90

Poisson scaled model

Used poisson distribution to see if fits better.

## NULL

Predictions poisson

This model predicts surprisingly well onto itself, with no more giant outliers. However, it does tend to under predict.

Predicting onto different sites didn’t go well. Most either an order of magnitude too large or too small. The randomly selected plots were slightly better, but still not particularly impressive.

See the tabs below.

Full model

Sites 2&3

Predict site 1 with site 2 & 3 data.

Sites 1&2

Predict site 3 with site 1 & 2 data.

Random selection

Proportion variance poisson

What is quite cool is that environmental variables seemed more important for abundance than they were for occurrence. This is the opposite of what I expected.

Correlation plots

Correlation due to the environment on the left, due to biotic factors on the right. The environmental relationships are very similar to the occurrence models.

Caterpillar plots

For the poisson model. Confidence intervals tend to be much tighter than for the occurrence models.