Reaster with negative binomial distribution

[José Waterton]

Hello!

I have been using aster for phenotypic selection analysis of emergence timing under multiple different environmental manipulations (e.g. with and without herbivore exclusion), and my measure of fitness (total mass of seeds produced) is heavily right-skewed. The predecessor node is just whether or not the plant produced seeds (bernoulli), as I did not collect data on number of fruits etc. A zero-inflated negative binomial distribution seemed appropriate, and I recently tried using a modelling seed mass (rounded to integers) predicted by emergence time and seed mass and exclusion treatment, with block and maternal lines as random effects. the graphical model:

root -> produced seeds (bernoulli) -> seed mass (truncated negative binomial)

When I ran the model I got an error message saying that the negative binomial distribution is incompatible with random effects in reaster. Firstly, I wondered if there are any solutions to this issue other than dropping the random effects from the model? Secondly, I wondered if anyone could advise me as to the best way to extract and plot the residuals from aster/reaster models, as I have struggled with this so far.

Thank you for your time!

José

[geyer]

Sorry.  But math can be obnoxious.  Yes, this error was put in because the negative binomial distribution, which has canonical parameter space that is restricted (- infinity, 0).  But random effects (at least the normal random effects everyone uses) are not restricted.  The "random effects" way to get overdispersion is to put in individual effects, but then one runs afoul of the other warning on the reaster help page, the one about individual effects not working well with Breslow-Clayton approximation.   AFAIK this is a problem that has no solution in the existing random effects literature.