Hello, PyMC community!
Imagine that I am modelling the impact of X on Y for different countries.
And I am doing it using the following piece of code:
beta_country = pm.Normal(f"beta_country", mu=0, sigma=3, dims="country")
As I understand, this will give the same Normal priors for each of the country with specified parameters.
But imagine that for some one specific country “C” I have a separate belief. (i.e. for all countries I would set mu=0, but for that country I would set it to mu=10).
Can you please recommend the best way to put it to code?
One version I have in mind is to use a for-loop, something like:
(pseudocode)
for country, country_prior in zip([countries,country_priors]):
beta_{country} = pm.Normal("beta_{country}", mu=country_prior, sigma=3)
Please let know if there are better versions.
Thank you!
You can provide a sequence to mu, sigma, or both, and PyMC will “broadcast” the parameteres across the RVs that are created. For example, here are 9 countries with mu of 0, and one with mu of 10:
n_countries = 10
mu_country = np.zeros(n_countries)
mu_country[3] = 10
beta_country = pm.Normal('beta_country', mu=mu_country, sigma=3, dims="country")
Look at some draws:
with np.printoptions(suppress=True, precision=3):
print(pm.draw(beta_country, 10))
>>>Out: [[ 5.973 -5.269 3.147 7.184 -1.387 0.398 -2.57 0.23 0.08 0.291]
>>> [ 0.042 1.767 4.018 8.41 -3.219 0.786 3.871 0.99 1.003 1.992]
>>> [ 3.58 -3.631 4.387 13.197 4.132 1.609 -3.308 1.881 -0.717 -0.658]
>>> [ 0.057 -3.223 -1.516 11.424 -6.257 3.19 -2.073 0.682 2.305 1.862]
>>> [ 4.132 0.885 0.783 11.652 -1.128 0.276 -1.531 -6.94 1.327 3.237]
>>> [-3.827 -5.264 0.452 9.152 -1.022 1.621 0.407 1.508 6.123 -1.106]
>>> [ 1.843 0.676 -1.493 7.441 -0.922 3.346 -3.009 -0.124 -3.32 -3.799]
>>> [ 1.251 -2.288 0.307 7.046 1.591 -1.67 -2.838 1.829 -2.188 1.75 ]
>>> [-0.052 2.99 1.198 7.059 -0.276 1.99 2.898 -4.114 1.749 0.47 ]
>>> [-3.075 0.637 -2.235 11.917 2.998 -0.811 -3.973 1.682 2.09 -2.479]]
You can see the 4th column has a different mean than the others, as expected.
1 Like
Thanks a lot, @jessegrabowski !
Sounds like a solution!