diff --git a/lenskit/algorithms/bias.py b/lenskit/algorithms/bias.py
index c33e1182b..2fa7e7b5f 100644
--- a/lenskit/algorithms/bias.py
+++ b/lenskit/algorithms/bias.py
@@ -122,13 +122,15 @@ def transform(self, ratings, *, indexes=False):
         if self.item_offsets_ is not None:
             rvps = rvps.join(self.item_offsets_, on='item', how='left')
             rvps['rating'] -= rvps['i_off'].fillna(0)
+            rvps = rvps.drop(columns='i_off')
         if self.user_offsets_ is not None:
             rvps = rvps.join(self.user_offsets_, on='user', how='left')
             rvps['rating'] -= rvps['u_off'].fillna(0)
+            rvps = rvps.drop(columns='u_off')
         if indexes:
             rvps['uidx'] = self.user_offsets_.index.get_indexer(rvps['user'])
             rvps['iidx'] = self.item_offsets_.index.get_indexer(rvps['item'])
-        return rvps.drop(columns=['u_off', 'i_off'])
+        return rvps
 
     def inverse_transform(self, ratings):
         """
@@ -142,10 +144,12 @@ def inverse_transform(self, ratings):
         if self.item_offsets_ is not None:
             rvps = rvps.join(self.item_offsets_, on='item', how='left')
             rvps['rating'] += rvps['i_off'].fillna(0)
+            del rvps['i_off']
         if self.user_offsets_ is not None:
             rvps = rvps.join(self.user_offsets_, on='user', how='left')
             rvps['rating'] += rvps['u_off'].fillna(0)
-        return rvps.drop(columns=['u_off', 'i_off'])
+            del rvps['u_off']
+        return rvps
 
     def transform_user(self, ratings):
         """
diff --git a/tests/test_bias.py b/tests/test_bias.py
index e506b9da3..db459427d 100644
--- a/tests/test_bias.py
+++ b/tests/test_bias.py
@@ -8,7 +8,7 @@
 import pandas as pd
 import numpy as np
 
-from pytest import approx, raises
+from pytest import approx, raises, mark
 
 from lenskit.util.test import ml_test
 
@@ -225,6 +225,29 @@ def test_bias_transform_indexes():
     assert denorm['rating'].values == approx(ratings['rating'].values, 1.0e-6)
 
 
+@mark.parametrize(['users', 'items'], [(True, False), (False, True), (False, False)])
+def test_bias_transform_disable(users, items):
+    algo = Bias(users=users, items=items)
+    ratings = ml_test.ratings
+
+    normed = algo.fit_transform(ratings)
+
+    assert all(normed['user'] == ratings['user'])
+    assert all(normed['item'] == ratings['item'])
+    denorm = algo.inverse_transform(normed)
+    assert denorm['rating'].values == approx(ratings['rating'], 1.0e-6)
+
+    n2 = ratings
+    nr = n2.rating - algo.mean_
+    if items:
+        n2 = n2.join(algo.item_offsets_, on='item')
+        nr = nr - n2.i_off
+    if users:
+        n2 = n2.join(algo.user_offsets_, on='user')
+        nr = nr - n2.u_off
+    assert normed['rating'].values == approx(nr.values)
+
+
 def test_bias_item_damp():
     algo = Bias(users=False, damping=5)
     algo.fit(simple_df)