PDP for Poisson GLM

Hi, I am trying to build a model to understand neural spikes of bats as function of their position (I have bunch of Xs and Ys in each timeframe, and spike-count for that timeframe).

I am using Poisson GAM (which is a fancy GLM), to predict spike counts as function of bats positions and the library pygam (GitHub - dswah/pyGAM: [HELP REQUESTED] Generalized Additive Models in Python).
The goal of the model as I mentioned is to make sense of the data, namely to understand the relation between each feature and it affect on the output.
Therefore I am plotting Partial Dependence Plots, but I am not sure on how to interpret the y-axis of these plots. When I was looking deeper into the code, it seems like the PDPs are not being calculated with the inverse of the log link function (which is what I would expect), therefore the y-axis corresponds somehow to the log of the number of spikes… it is not really intutitve. I tried to manually pass it through e^x but the results are still not interpretable to me.

  1. How should I calculate the new confidence intervals after I pass the values through e^x?
  2. Something which is clearly missing is the removal of the effect of other features! I am not sure how to do that.

Thanks ahead!

@itayona: This is quite a specific question which I don’t have experience in. cc @Bruno do you have any experience with these?

Not enough to be able to answer this without some serious digging, which, alas, I can’t afford right now :confused:

1 Like

Its alright… Hope someone else has the experience with this.