Hi all,

I’m working on a pharmacokinetic model using rstan. From the initial development I’ve implemented a first-order absorption model using both the analytic and ODE solutions. Both perform fine, but the data suggest an absorption delay is required. A lag-time model performs poorly (irreperable divergences and low ESS) presumably due to numeric stability issues around the transition point.

Instead I’ve implemented a transit compartment model (https://doi.org/10.1007/s10928-007-9066-0) which in principle implements a delayed absorption without an abrupt transition. The system has bolus dosing so I used the analytic solution as in the paper. No matter what I do with it, I receive a ‘non-finite gradient at initial value’ error on attempting to sample from the model.

After some troubleshooting:

- there are no apparent problems with ODE performance; the model appears to be correct as I can simulate from it with Stan, and get equivalent results from deSolve. I used the same initial values. I used the print() function within the Stan code which confirms that the model predicted values appear realistic.
- the ‘grad_log_prob()’ function shows that indeed the mean-transit-time parameter has NaN gradient at any initial value (except very small values in which CVOde fails), but for any values I select, the log_prob apparently exists. I also fixed the ‘number of compartments’ parameter so I think there shouldn’t be an identifiability issue.
- don’t seem to be any programming problems; the model converges just fine if I replace the absorption model with a first-order model with otherwise exactly the same code.

Has anyone made a similar model and encountered this or have any comment about what I might be missing?