I'm using the MNP package to fit Bayesian multinomial probit models to categorical data. I frequently get this error: TruncNorm: lower bound is greater than upper bound. How can I resolve this problem?
My sense is that this might be a convergence issue, so I've tried many different approaches on this front:
- narrower priors / changing priors
- scaling predictors, dropping highly correlated predictors
- start with one predictor then build up to more
But none of these have worked. What I have noticed is that when n.draws is set to a smaller number this Error comes up less frequently. Here's an example using the Iris dataset:
library(MNP)
data("iris")
iris$Species = as_factor(as.character(iris$Species))
res = mnp(Species ~ Sepal.Length,
data = iris,
n.draws = 10000, verbose = TRUE)
summary(res)
The model returns the TruncNorm error at N = 10,000, but will finish at smaller sizes of N.
