Avoid seeing an error from statsmodels in kdeplot when data IQR == 0 (#…

…2040)

(cherry picked from commit 09fef02)

doc/releases/v0.10.1.txt

@@ -18,6 +18,8 @@ This is minor release with bug fixes for issues identified since 0.10.0.
 
 - Added the ``showfliers`` parameter to :func:`boxenplot` to suppress plotting of outlier data points, matching the API of :func:`boxplot`.
 
+- Avoided seeing an error from statmodels when data with an IQR of 0 is passed to :func:`kdeplot`.
+
 - Added the ``legend.title_fontsize`` to the :func:`plotting_context` definition.
 
 - Several utility functions that are no longer used internally (``percentiles``, ``sig_stars``, ``pmf_hist``, and ``sort_df``) have been deprecated and marked for future removal.

seaborn/distributions.py

@@ -359,7 +359,16 @@ def _statsmodels_univariate_kde(data, kernel, bw, gridsize, cut, clip,
     """Compute a univariate kernel density estimate using statsmodels."""
     fft = kernel == "gau"
     kde = smnp.KDEUnivariate(data)
-    kde.fit(kernel, bw, fft, gridsize=gridsize, cut=cut, clip=clip)
+
+    try:
+        kde.fit(kernel, bw, fft, gridsize=gridsize, cut=cut, clip=clip)
+    except RuntimeError as err:  # GH#1990
+        if stats.iqr(data) > 0:
+            raise err
+        msg = "Default bandwidth for data is 0; skipping density estimation."
+        warnings.warn(msg, UserWarning)
+        return np.array([]), np.array([])
+
     if cumulative:
         grid, y = kde.support, kde.cdf
     else:
@@ -369,14 +378,7 @@ def _statsmodels_univariate_kde(data, kernel, bw, gridsize, cut, clip,
 
 def _scipy_univariate_kde(data, bw, gridsize, cut, clip):
     """Compute a univariate kernel density estimate using scipy."""
-    try:
-        kde = stats.gaussian_kde(data, bw_method=bw)
-    except TypeError:
-        kde = stats.gaussian_kde(data)
-        if bw != "scott":  # scipy default
-            msg = ("Ignoring bandwidth choice, "
-                   "please upgrade scipy to use a different bandwidth.")
-            warnings.warn(msg, UserWarning)
+    kde = stats.gaussian_kde(data, bw_method=bw)
     if isinstance(bw, str):
         bw = "scotts" if bw == "scott" else bw
         bw = getattr(kde, "%s_factor" % bw)() * np.std(data)

seaborn/tests/test_distributions.py

@@ -6,11 +6,19 @@
 import pytest
 import nose.tools as nt
 import numpy.testing as npt
+from distutils.version import LooseVersion
 
 from .. import distributions as dist
 
 _no_statsmodels = not dist._has_statsmodels
 
+if not _no_statsmodels:
+    import statsmodels
+    import statsmodels.nonparametric as smnp
+    _old_statsmodels = LooseVersion(statsmodels.__version__) < "0.11"
+else:
+    _old_statsmodels = False
+
 
 class TestDistPlot(object):
 
@@ -174,6 +182,29 @@ def test_kde_singular(self):
         line = ax.lines[1]
         assert not line.get_xydata().size
 
+    @pytest.mark.skipif(_no_statsmodels or _old_statsmodels,
+                        reason="no statsmodels or statsmodels without issue")
+    def test_statsmodels_zero_bandwidth(self):
+        """Test handling of 0 bandwidth data in statsmodels."""
+        x = np.zeros(100)
+        x[0] = 1
+
+        try:
+
+            smnp.kde.bandwidths.select_bandwidth(x, "scott", "gau")
+
+        except RuntimeError:
+
+            # Only execute the actual test in the except clause, this should
+            # keep the test from failing in the future if statsmodels changes
+            # it's behavior to avoid raising the error itself.
+            # Track at https://github.com/statsmodels/statsmodels/issues/5419
+
+            with pytest.warns(UserWarning):
+                ax = dist.kdeplot(x)
+            line = ax.lines[0]
+            assert not line.get_xydata().size
+
     @pytest.mark.parametrize("cumulative", [True, False])
     def test_kdeplot_with_nans(self, cumulative):