jeffgortmaker
commented
1 year ago Right, so the correct way to do this is actually the hacky workaround that you suggested! You should specify a formulation for X2, probably including any characteristics that your micro moments will depend on. If your micro moments also depend on demographics, you'll have to specify a formulation for them as well.
Then you'll need to specify agent data. If you don't have any demographics (and you're just matching, for example, some second choice statistics, which should absolutely be informative about rho), you can just specify a single row per market with some arbitrary column of nodes, and set sigma to zero. This is actually what PyBLP does in the background for the simple nested logit model, so you won't be slowing anything down. If you do have demographics, you'll have to specify those.
Since you're using micro moments, I'll also note that they're under active development, so you should probably use the "dev" version by cloning this repo and adding it to your PYTHONPATH (and looking at the dev docs), instead of using the "stable" version via pip. I've changed up the micro moments interface a bit in the last few weeks, so if you use the stable version you'll have to modify your code for future versions.
andrew-heinzman
Is there a way to incorporate micro moments into a nested logit model without random coefficients? In a nested logit model rho needs to be optimized and micro moments should help with identification.
As currently written I believe the package does not accept micro moments unless agent data is provided and agent data cannot be provided without formulating X2. A hacky workaround specifies a formulation for X2 and then sets sigma to 0 when solving the problem. However, this likely causes its own issues such as integrating when not necessary.
Thank you for all the hard work that's gone into the package it's a fantastic service!