diff --git a/bids/analysis/tests/test_transformations.py b/bids/analysis/tests/test_transformations.py
index 892721a9e..3179cb3a5 100644
--- a/bids/analysis/tests/test_transformations.py
+++ b/bids/analysis/tests/test_transformations.py
@@ -4,6 +4,7 @@
 from bids.variables.entities import RunInfo
 from bids.variables.kollekshuns import BIDSRunVariableCollection
 from bids.layout import BIDSLayout
+import math
 import pytest
 from os.path import join, sep
 from bids.tests import get_test_data_path
@@ -16,14 +17,30 @@
     import mock
 
 
+# Sub-select collection for faster testing, without sacrificing anything
+# in the tests
+SUBJECTS = ['01', '02']
+NRUNS = 3
+SCAN_LENGTH = 480
+
+cached_collections = {}
+
+
 @pytest.fixture
 def collection():
-    layout_path = join(get_test_data_path(), 'ds005')
-    layout = BIDSLayout(layout_path)
-    collection = layout.get_collections('run', types=['events'],
-                                        scan_length=480, merge=True,
-                                        sampling_rate=10)
-    return collection
+    if 'ds005' not in cached_collections:
+        layout_path = join(get_test_data_path(), 'ds005')
+        layout = BIDSLayout(layout_path)
+        cached_collections['ds005'] = layout.get_collections(
+            'run',
+            types=['events'],
+            scan_length=SCAN_LENGTH,
+            merge=True,
+            sampling_rate=10,
+            subject=SUBJECTS
+        )
+    # Always return a clone!
+    return cached_collections['ds005'].clone()
 
 
 @pytest.fixture
@@ -77,8 +94,8 @@ def test_convolve(collection):
 
 
 def test_rename(collection):
-    dense_rt = collection.variables['RT'].to_dense(10)
-    assert len(dense_rt.values) == 230400
+    dense_rt = collection.variables['RT'].to_dense(collection.sampling_rate)
+    assert len(dense_rt.values) == math.ceil(len(SUBJECTS) * NRUNS * SCAN_LENGTH * collection.sampling_rate)
     transform.Rename(collection, 'RT', output='reaction_time')
     assert 'reaction_time' in collection.variables
     assert 'RT' not in collection.variables
@@ -132,9 +149,10 @@ def test_demean(collection):
 def test_orthogonalize_dense(collection):
     transform.Factor(collection, 'trial_type', sep=sep)
 
+    sampling_rate = collection.sampling_rate
     # Store pre-orth variables needed for tests
-    pg_pre = collection['trial_type/parametric gain'].to_dense(10)
-    rt = collection['RT'].to_dense(10)
+    pg_pre = collection['trial_type/parametric gain'].to_dense(sampling_rate)
+    rt = collection['RT'].to_dense(sampling_rate)
 
     # Orthogonalize and store result
     transform.Orthogonalize(collection, variables='trial_type/parametric gain',
@@ -174,7 +192,8 @@ def test_split(collection):
 
     orig = collection['RT'].clone(name='RT_2')
     collection['RT_2'] = orig.clone()
-    collection['RT_3'] = collection['RT'].clone(name='RT_3').to_dense(10)
+    collection['RT_3'] = collection['RT']\
+        .clone(name='RT_3').to_dense(collection.sampling_rate)
 
     rt_pre_onsets = collection['RT'].onset
 
@@ -208,14 +227,22 @@ def test_split(collection):
 
 
 def test_resample_dense(collection):
-    collection['RT'] = collection['RT'].to_dense(10)
+    new_sampling_rate = 50
+    old_sampling_rate = collection.sampling_rate
+    upsampling = float(new_sampling_rate) / old_sampling_rate
+
+    collection['RT'] = collection['RT'].to_dense(old_sampling_rate)
     old_rt = collection['RT'].clone()
-    collection.resample(50, in_place=True)
-    assert len(old_rt.values) * 5 == len(collection['RT'].values)
+    collection.resample(new_sampling_rate, in_place=True)
+    assert math.floor(len(old_rt.values) * upsampling) == len(collection['RT'].values)
     # Should work after explicitly converting categoricals
     transform.Factor(collection, 'trial_type')
-    collection.resample(5, force_dense=True, in_place=True)
-    assert len(old_rt.values) == len(collection['parametric gain'].values) * 2
+
+    new_sampling_rate2 = 5
+    upsampling2 = float(new_sampling_rate2) / old_sampling_rate
+
+    collection.resample(new_sampling_rate2, force_dense=True, in_place=True)
+    assert len(old_rt.values) == math.ceil(float(len(collection['parametric gain'].values) / upsampling2))
 
 
 def test_threshold(collection):
@@ -346,12 +373,13 @@ def test_filter(collection):
     transform.Filter(collection, 'RT', query=q, by='parametric gain')
     assert len(orig.values) != len(collection['RT'].values)
     # There is some bizarro thing going on where, on travis, the result is
-    # randomly either 1536 or 3909 when running on Python 3 (on linux or mac).
+    # randomly either 1536 (when all subjects are used) or 3909 when running
+    # on Python 3 (on linux or mac).
     # Never happens locally, and I've had no luck tracking down the problem.
     # Best guess is it reflects either some non-deterministic ordering of
     # variables somewhere, or some weird precision issues when resampling to
     # dense. Needs to be tracked down and fixed.
-    assert len(collection['RT'].values) in [1536, 3909]
+    assert len(collection['RT'].values) in [96 * len(SUBJECTS), 3909]
 
 
 def test_replace(collection):