Modify NegativeBinomial threshold for Poisson approximation based on …

…float accuracy, and add tests.

pymc3/distributions/discrete.py

@@ -13,13 +13,15 @@
 #   limitations under the License.
 import warnings
 
+import aesara
 import aesara.tensor as aet
 import numpy as np
 
 from aesara.tensor.random.basic import bernoulli, binomial, categorical, nbinom, poisson
 from scipy import stats
 
 from pymc3.aesaraf import floatX, intX, take_along_axis
+from pymc3.distributions import logpt
 from pymc3.distributions.dist_math import (
     betaln,
     binomln,
@@ -782,7 +784,13 @@ def logp(value, n, p):
         )
 
         # Return Poisson when alpha gets very large.
-        return aet.switch(aet.gt(alpha, 1e10), Poisson.dist(mu).logp(value), negbinom)
+        threshold = 1e10 if aesara.config.floatX == "float64" else 1e7
+
+        return aet.switch(
+            aet.gt(alpha, threshold),
+            logpt(Poisson.dist(mu), value),
+            negbinom,
+        )
 
     def logcdf(value, n, p):
         """

pymc3/tests/test_distributions.py

@@ -1140,11 +1140,13 @@ def modified_scipy_hypergeom_logcdf(value, N, k, n):
             {"N": NatSmall, "k": NatSmall, "n": NatSmall},
         )
 
-    @pytest.mark.xfail(reason="Logp method not refactored yet")
     def test_negative_binomial(self):
         def scipy_mu_alpha_logpmf(value, mu, alpha):
             return sp.nbinom.logpmf(value, alpha, 1 - mu / (mu + alpha))
 
+        def scipy_poisson_approximation(value, mu, alpha):
+            return sp.poisson.logpmf(value, mu)
+
         def scipy_mu_alpha_logcdf(value, mu, alpha):
             return sp.nbinom.logcdf(value, alpha, 1 - mu / (mu + alpha))
 
@@ -1154,6 +1156,15 @@ def scipy_mu_alpha_logcdf(value, mu, alpha):
             {"mu": Rplus, "alpha": Rplus},
             scipy_mu_alpha_logpmf,
         )
+        # Test values around the point where the logp of the NegativeBinomial
+        # is approximated with a Poisson logp
+        self.check_logp(
+            NegativeBinomial,
+            Nat,
+            {"mu": Rplus, "alpha": Domain([0, 1.5e6, 1.5e7, 1.5e9, 1.5e10, inf])},
+            scipy_poisson_approximation,
+            decimal=select_by_precision(float64=2, float32=1),
+        )
         self.check_logp(
             NegativeBinomial,
             Nat,
@@ -1181,7 +1192,6 @@ def scipy_mu_alpha_logcdf(value, mu, alpha):
             n_samples=10,
         )
 
-    @pytest.mark.xfail(reason="Distribution not refactored yet")
     @pytest.mark.parametrize(
         "mu, p, alpha, n, expected",
         [