Thanks @junpenglao! I would have never caught that. Does that mean pm.Bernoulli cannot see data as observed proportion but rather a “single datapoint”? I find that fact a bit strange because I thought pm.Bernoulli(p, observed) would have tried to match p to “observed p”…
Is it even possible to pass in the whole data to Bernoulli? I thought about perhaps using the beta(a,b) distribution where a = success and b = failures, but I won’t be able to use sigmoid anymore, so that would not work.