diff --git a/.travis.yml b/.travis.yml
index d9274d8..b6fc8cf 100644
--- a/.travis.yml
+++ b/.travis.yml
@@ -10,6 +10,7 @@ install:
     - pip install --upgrade pip setuptools
     - pip install -r dev-requirements.txt
     - pip install -r requirements.txt
+    - pip install pandas==0.19 ; python_version==3.5
     - pip install -e .
 before_script: flake8 civismlext
 script:
diff --git a/CHANGELOG.md b/CHANGELOG.md
index 6451fde..9623945 100644
--- a/CHANGELOG.md
+++ b/CHANGELOG.md
@@ -8,6 +8,8 @@ and this project adheres to [Semantic Versioning](http://semver.org/).
 ### Fixed
 - Fixed ``DataFrameETL`` transformations of ``DataFrame``s with non-trivial
   index when preserving ``DataFrame`` output type (#32, #33)
+- Add ``pandas`` version restrictions by Python version (#37)
+- Fix code which was incompatible with older ``pandas`` version (#37)
 
 ## [0.1.7] - 2018-03-27
 ### Added
diff --git a/civismlext/preprocessing.py b/civismlext/preprocessing.py
index 72b5be8..24476dc 100644
--- a/civismlext/preprocessing.py
+++ b/civismlext/preprocessing.py
@@ -94,7 +94,7 @@ def _flag_nulls(self, X, cols_to_drop):
     def _check_sentinels(self, X):
         """Replace default sentinel with random values if the default appears
         in the data."""
-        vals = chain.from_iterable(pd.unique(X[col]).tolist() for
+        vals = chain.from_iterable(pd.unique(X[col]) for
                                    col in self._cols_to_expand)
 
         while any(u == self._nan_sentinel for u in vals):
diff --git a/requirements.txt b/requirements.txt
index a27674f..1e13cd9 100644
--- a/requirements.txt
+++ b/requirements.txt
@@ -1,5 +1,7 @@
 scikit-learn>=0.18.1,<0.20
 scipy>=0.14,<2.0
 numpy~=1.10
-pandas~=0.19
+pandas~=0.19; python_version >= '3.5'
+pandas>=0.19,<0.21; python_version == '3.4'
+pandas>=0.19,<=0.23; python_version == '2.7'
 six~=1.10