How to fix "Domain error" when using posterior predictive sampling?

Questions
1

While I can generate a normal looking traceplot and posterior plot from my model, when I try to run the following:

ppc = pm.sample_ppc(trace, samples=500, model=model)

I get the following error:


    ValueError                                Traceback (most recent call last)
C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\distribution.py in _draw_value(param, point, givens, size)
    397                 try:
--> 398                     dist_tmp.random(point=point, size=size)
    399                 except (ValueError, TypeError):

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\continuous.py in random(self, point, size)
   1074                                 dist_shape=self.shape,
-> 1075                                 size=size)
   1076 

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\distribution.py in generate_samples(generator, *args, **kwargs)
    502         if size == 1 or (broadcast_shape == size_tup + dist_shape):
--> 503             samples = generator(size=broadcast_shape, *args, **kwargs)
    504         elif dist_shape == broadcast_shape:

C:\ProgramData\Anaconda3\lib\site-packages\scipy\stats\_distn_infrastructure.py in rvs(self, *args, **kwds)
    939         if not np.all(cond):
--> 940             raise ValueError("Domain error in arguments.")
    941 

ValueError: Domain error in arguments.

During handling of the above exception, another exception occurred:

ValueError                                Traceback (most recent call last)
<ipython-input-45-10a08c3d7ad4> in <module>()
----> 1 ppc = pm.sample_ppc(trace, samples=500, model=model)

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\sampling.py in sample_ppc(trace, samples, model, vars, size, random_seed, progressbar)
   1127     # draw once to inspect the shape
   1128     var_values = list(zip(varnames,
-> 1129                           draw_values(vars, point=model.test_point, size=size)))
   1130     ppc_trace = defaultdict(list)
   1131     for varname, value in var_values:

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\distribution.py in draw_values(params, point, size)
    319             else:
    320                 try:  # might evaluate in a bad order,
--> 321                     evaluated[param_idx] = _draw_value(param, point=point, givens=givens.values(), size=size)
    322                     if isinstance(param, collections.Hashable) and named_nodes_parents.get(param):
    323                         givens[param.name] = (param, evaluated[param_idx])

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\distribution.py in _draw_value(param, point, givens, size)
    401                     # with theano.shared inputs
    402                     dist_tmp.shape = np.array([])
--> 403                     val = dist_tmp.random(point=point, size=None)
    404                     dist_tmp.shape = val.shape
    405                 return dist_tmp.random(point=point, size=size)

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\continuous.py in random(self, point, size)
   1073         return generate_samples(stats.beta.rvs, alpha, beta,
   1074                                 dist_shape=self.shape,
-> 1075                                 size=size)
   1076 
   1077     def logp(self, value):

C:\ProgramData\Anaconda3\lib\site-packages\pymc3\distributions\distribution.py in generate_samples(generator, *args, **kwargs)
    501     elif broadcast_shape[-len(dist_shape):] == dist_shape or len(dist_shape) == 0:
    502         if size == 1 or (broadcast_shape == size_tup + dist_shape):
--> 503             samples = generator(size=broadcast_shape, *args, **kwargs)
    504         elif dist_shape == broadcast_shape:
    505             samples = generator(size=size_tup + dist_shape, *args, **kwargs)

C:\ProgramData\Anaconda3\lib\site-packages\scipy\stats\_distn_infrastructure.py in rvs(self, *args, **kwds)
    938         cond = logical_and(self._argcheck(*args), (scale >= 0))
    939         if not np.all(cond):
--> 940             raise ValueError("Domain error in arguments.")
    941 
    942         if np.all(scale == 0):

ValueError: Domain error in arguments.

I am very new to using PyMC3, so this is probably some simple mistake, but I haven’t been able to find relevant information online.

My model is the following (attempting to use Kruschke’s multiple linear regression with a student’s t hyperprior for a beta-distributed y variable)
model

How do I fix this? Please let me know what other information would be relevant!

1 Like
2

While the printout of the model is helpful, it would also be nice to see the Python code. Could you add that here as well?

Also, at a first glance, my suspicion is that there is no restraint on the alpha/beta arguments into the beta-distributed response that forces them to be positive as is required for the beta distribution’s parameters.

For example, if the sampling_ppc procedure draws a -1 for beta0 and 0 for beta1, then the resulting input to the beta distribution will be -1 which is invalid and hence the error is triggered.

2 Likes
3

Thank you for your reply! I’ve stopped getting the error now after making other changes to my model, and I didn’t save a copy of the buggy code. Should I close this question or something similar?

For what it’s worth, I had already clipped mu to be between 0 and 1, and as log(kappa) will always be positive (drawn from an exponential distribution), alpha and beta were also guaranteed to be positive. Otherwise, I believe pm.sample would have not worked either?

4

There’s no need to close it, though if you do discover what the issue was, it would be good if you added that information here.

5

Note that zero is also a forbidden value for the beta distribution parameters. If you had clipped the values including zero, then that could have caused your error.

2 Likes