Fitting a Binomial distribution with pymc raises ZeroProbability error for certain FillValues in masked arrays about pymc2 HOT 4 CLOSED

This is because you filled the missing values with 999999, which is outside the support of the variable. You need to give it a valid value, but not one that occurs in the non-missing data.

Also, note that you don't have to pass plot=False for observed stochastics. They will not get printed.

from pymc2.

Thanks @fonnesbeck for pointing this out!

So does that means if my observed data contains all possible values inside the support I cannot treat missing data?
E.g. consider the following example:

observed_values = sp.random.binomial(n = 3.0, p = 0.5, size = 100)
print(sp.unique(observed_values))

Output:

[0 1 2 3]

What to choose for the masked value?

mask = sp.zeros_like(observed_values)
mask[0] = True
masked_values = sp.ma.masked_array(observed_values, mask = mask, masked_values = ????)

Is it really necessary that pymc checks for the masked value to be in the support? Because I have the following situation: I do have a model where the n of the binomials actually is a random variable itself. And I'm fitting a number of datasets. That means I would have to check which masked value to use for every dataset which is certainly possible, but sometimes I might not be able to find a masked value. So I would vote for not checking if the missing values are inside the support. Would you agree.

from pymc2.

You can give it a non-integer value in order to avoid the problem that you cite. For example, if you fill with, say, 1.5 that should work.

Its not a matter of PyMC "checking" the the values are inside the support, it is just that PyMC calculates the log-probability at the first iteration, and therefore the values inserted for the missing values at the first iteration have to be valid. If you give discrete values a floating point value, it should end up getting truncated when converted to integers, so that will work.

from pymc2.

Thanks @fonnesbeck, using 1.5 works in my example.

btw: Using the plot = False option for observed stochastics makes makes sense if there are missing data because it will be plotted otherwise.

from pymc2.