model.null<- zoib(PDMR~ 1| 1|1|1, data=Dat2, joint = F, random = 1, EUID = Dat2$Stream, zero.inflation = T, one.inflation=F, n.iter = 10200, n.thin = 50, n.burn = 200, link.mu = "cloglog") dic.samples(model.null$MCMC.model, n.iter = 1000, type = "pD")
First time poster here. I'm trying to model the proportional daily movement rate(PDMR in the code below) of several species in relation to several continuous covariates. Given that PDMR adheres to a beta distribution, and is clearly zero-inflated, I decided to use the Zoib package to model this response variable (see code for basic null model).
Assuming I specified everything correctly: the model should indicate that changes in PDMR are strictly a result of the random effect (in this case, Stream which is a complete block and has 4 levels) with no covariates (i.e., xb=1, xd=1, x0=1, x1=NULL, z=1). The "cloglog" link function was specified because the data is highly left skewed.
Here's the data for the response, where each cell represents an independent measurement of PDMR (53 observations total).
For this model, the value for mean deviance information criterion (DIC) is astronomically high (103812) whereas the penalty is extremely low (0.003745). Now I realize that in multi-model selection the relative size of the information criterion is not what matters, but rather it is the difference between these values that allows the user to distinguish which model fits the data the best. However, these results seem quite suspect. I attempted to moderate the number of Markov chains, interations, etc, but that doesn't seem to really influence the size of the mean deviance or penalty. If these results are not a fallacy of incorrect model specification, then I'm assuming they're an attribute of the data?
Also, any general tips on how to digest the results of the model would be thoroughly appreciated!