Thanks for the investigative work tracking down the troublesome data point - I have a couple of things here:
- Can you confirm that the model component which is blowing up the logp is the likelihood? See here for how to do that.
- Can you try changing alpha individually on its own and then changing the hyperparameters for
mu
on their own and see whether or not the resulting error occurs? - Finally, can you make a plot of the initial logp for values approaching the bad Y_OBS? If you post a complete minimal viable reproduction I can help with this.