Thanks.
I used an equivalent formulation of the inverse logit:
p = exp(a[actor_] + g[block_] + b[treatment_]) / (1 + exp(a[actor_] + g[block_] + b[treatment_]))
Numbers almost work when I plug them in like this:
az_res = az.summary(trace_13_4)
# treatment for 0th datapoint is 0
exp_res = np.exp(az_res.loc["a[0]", "mean"] + az_res.loc["b[0]", "mean"] + az_res.loc["g[0]", "mean"])
p0_res = exp_res / (1+exp_res)
p0_res is 0.339 while az_res.loc[“p[0]”, “mean”] is 0.344. So quite close but not exact. What could be the cause of the difference?