Major difference between frequentist and HB results for apollo_mnl component
Posted: 15 Dec 2021, 15:49
Hi David and Stephane,
I am estimating two ICLVs (different data and model specifications) where at least one of the indicators I'm interested in is a binary outcome. The measurement function for that indicator is a binary logit, and the "utility" is equal to U = c_0 + z_1 * LV. I'm also using ordered indicators with ordered probits and a "utility" equal to z_i * LV.
For one model, frequentist estimation gives z_1 = 2.2, which is reasonable and consistent with the indicator and the scale of LV. Using Bayesian estimation, z_1 "shoots up" to very high values and never stabilizes, no matter the number of iterations I give it. Something similar happens to the other ICLV.
I've tried estimating this model with and without an error term, fixing the parameters of the structural equation, etc., and this still keeps happening. My gut feeling is that this behavior is most likely due to something going on either in Apollo (running 0.2.6) or RSGHB (1.2.2).
I can share the code and data offline.
Thanks!
I am estimating two ICLVs (different data and model specifications) where at least one of the indicators I'm interested in is a binary outcome. The measurement function for that indicator is a binary logit, and the "utility" is equal to U = c_0 + z_1 * LV. I'm also using ordered indicators with ordered probits and a "utility" equal to z_i * LV.
For one model, frequentist estimation gives z_1 = 2.2, which is reasonable and consistent with the indicator and the scale of LV. Using Bayesian estimation, z_1 "shoots up" to very high values and never stabilizes, no matter the number of iterations I give it. Something similar happens to the other ICLV.
I've tried estimating this model with and without an error term, fixing the parameters of the structural equation, etc., and this still keeps happening. My gut feeling is that this behavior is most likely due to something going on either in Apollo (running 0.2.6) or RSGHB (1.2.2).
I can share the code and data offline.
Thanks!