Adaptation terminated, GP regression model

Diagonal elements of inverse mass matrix close to 0, implying singularity.

Can assume no issues with Stan code, Standard GP regression model. I’ve used it before with no issues. Data is in “long” format, as if for a survival model. All covariates were standardized, time component was “standardized” between 0 and 1.

# Adaptation terminated
# Step size = 0.506072
# Diagonal elements of inverse mass matrix:
# 0.00213663, 0.00333117, 0.0153331, 0.00724522, 0.0318262, 0.00188058

Here’s everything we need to reproduce. I can email the output.csv if need be.

gp_trees_demo_reduced.input.R (87.3 KB)
gp_regression_ard.stan (2.4 KB)
gp_trees_reproduce_input.R (6.2 KB)
data.csv (20.7 KB)

Could be my fault. I’ll take another look tomorrow. Initial data inspection looks fine to me, though.

Initially data had two treatment effects. I naively pooled them together to simplify the model. I removed a bunch of categorical variables because it was too poorly identified with all of the hierarchies.

I think I realize my mistake.

Usually in a survival analysis or censored data we’ll just “increment log prob”. What I’ve done is just dropped observations from the covariance matrix. Instead, if an observation has dropped out, I’ll include that in covariance matrix and model this in the response accordingly.

This should help with stability. No one spend any time on this, thanks.

2 Likes

Thanks for following up with a diagnosis of the problem.