BUG: make stats.f_oneway work with integer array input.

Closes gh-3376.

scipy/stats/stats.py

@@ -2379,14 +2379,15 @@ def f_oneway(*args):
     .. [2] Heiman, G.W.  Research Methods in Statistics. 2002.
 
     """
-    args = list(map(np.asarray, args))  # convert to an numpy array
-    na = len(args)              # ANOVA on 'na' groups, each in it's own array
+    args = [np.asarray(arg, dtype=float) for arg in args]
+    na = len(args)    # ANOVA on 'na' groups, each in it's own array
     alldata = np.concatenate(args)
     bign = len(alldata)
     sstot = ss(alldata) - (square_of_sums(alldata) / float(bign))
     ssbn = 0
     for a in args:
         ssbn += square_of_sums(a) / float(len(a))
+
     ssbn -= (square_of_sums(alldata) / float(bign))
     sswn = sstot - ssbn
     dfbn = na - 1

scipy/stats/tests/test_stats.py

@@ -2668,22 +2668,25 @@ def test_sigmaclip3(self):
 
 
 class TestFOneWay(TestCase):
-
     def test_trivial(self):
         # A trivial test of stats.f_oneway, with F=0.
         F, p = stats.f_oneway([0,2], [0,2])
         assert_equal(F, 0.0)
 
     def test_basic(self):
-        # A test of stats.f_oneway, with F=2.
-        F, p = stats.f_oneway([0,2], [2,4])
         # Despite being a floating point calculation, this data should
         # result in F being exactly 2.0.
+        F, p = stats.f_oneway([0,2], [2,4])
         assert_equal(F, 2.0)
 
+    def test_large_integer_array(self):
+        a = np.array([655, 788], dtype=np.uint16)
+        b = np.array([789, 772], dtype=np.uint16)
+        F, p = stats.f_oneway(a, b)
+        assert_almost_equal(F, 0.77450216931805538)
 
-class TestKruskal(TestCase):
 
+class TestKruskal(TestCase):
     def test_simple(self):
         x = [1]
         y = [2]