Indeed I have some sort of normalization! A slightly different kind of, though:
normalize = (1-s)*(1-Lh)+s*(1-Ls)
events = pm.Poisson('events',(x_1*(1-s)*(1-Lh)+x_2*s*(1-Ls))*exposure/normalize)
May I ask what was the hint that suggested the normalization? What’s happening in the code above is that I’m approximating a sum of two binomials with probabilities x_1 and x_2, and different sample sizes using a Poisson distribution. Is there a better way to parametrize this, or avoid the normalization?
Regarding transform.Order - I’ve tried to find an example and found this one, in which rates is a 3-numbers vector, constrained to be ordered (if I understood the example correctly). I have a slightly different situation, in which x_1 and x_2 are coming from two different (upstream) GLMs, and I want to contrain x_1>x_2. Is this possible?