Sampling from multilevel multivariable regression model is very slow

[mishooax]

@junpenglao

So use non-centered parameterization instead, and also get rid of the for-loop:

    # group slopes
    beta_base = pm.Normal('beta_base', 0., 1., shape=(n_predictors, n_groups))
    beta = pm.Deterministic('beta', mu + pm.math.dot(L_Omega, beta_base))

@junpenglao Gotcha - thanks!! However, the sampling is still slow even with the vectorised + QR + non-centered param version. One chain with 2000 samples takes about 10 hrs to draw. I reckon this has to do with the complicated shape of the posterior - any ideas on how to speed this code up further? I’ve tried to increase the no. of cores when calling pm.sample() but then I run into another issue.