Page 2 of 2
Re: WTP space results - means are negative or NANs produced
Posted: 09 Mar 2022, 16:00
by Simone R
Hi Stephane,
I ran the model in preference space with 3000 Sobol draws without any issues; LL: -2733.08
Simone
Re: WTP space results - means are negative or NANs produced
Posted: 11 Mar 2022, 13:06
by stephanehess
Simone
looking at this, it's clear that your models with 1000 and 4000 draws just didn't really move away from the log-likelihood at starting values, so they immediately got stuck in a hole. Were you using the same starting values in all these models?
Stephane
Re: WTP space results - means are negative or NANs produced
Posted: 14 Mar 2022, 10:16
by Simone R
Hi Stephane,
yes I used the same starting values for the models with 1000, 2000, 2100, and 4000 draws. Since I only got good results without NANs or saddle point convergence for the models in the 2000-2100 draws range, I then used the results of the model with 2100 draws as starting values to run the model with 3000 draws to test whether that would improve the results when I increase the draws.
Simone
Re: WTP space results - means are negative or NANs produced
Posted: 14 Mar 2022, 18:38
by stephanehess
Simone
more draws is always better. If your SLL with the same values is worse with more draws, then that doesn't mean that more draws is worse. It just shows that the approximation of the true LL (remember that you are using the SLL, simulated LL) is less accurate with fewer draws
Stephane