Are there instances when reparameterizing a model makes it slower?

[DanWeitzenfeld]

@Brideau

@DanWeitzenfeld

How many groups are there, just two? Because two is too few to use a hierarchical distribution.

Can you elaborate on this? Looking at the prior it generates once the model is run, it seems to be behaving just fine, even if it is a bit slow.

It’s hard to infer the mean and standard deviation of a distribution if you only have two draws from it!

@Brideau

@DanWeitzenfeld

My guess is the centered model is running slowly because all_groups_mean is given low prior likelihood of a value that extreme (-4).

Not sure if I follow here either. The prior has a lot of weight around -4 because that’s what the data informs it to be. E.g. inv_logit(-4) = 0.018 or 1.8% , which is in the ballpark of what the expected conversion rate is.

The posterior has a lot of weight around -4 because that’s what the data informs it to be. The prior, with a mean of 0 and a std dev of 1.5, puts very little weight on -4.