diff --git a/target_csv.py b/target_csv.py
index c372545..aff869f 100755
--- a/target_csv.py
+++ b/target_csv.py
@@ -16,6 +16,9 @@
 from jsonschema.validators import Draft4Validator
 import singer
 
+DEFAULT_PARENT_KEY = ''
+SEP = '__'
+
 logger = singer.get_logger()
 
 def emit_state(state):
@@ -25,21 +28,33 @@ def emit_state(state):
         sys.stdout.write("{}\n".format(line))
         sys.stdout.flush()
 
-def flatten(d, parent_key='', sep='__'):
+def flatten(d, parent_key=DEFAULT_PARENT_KEY, sep=SEP):
     items = []
     for k, v in d.items():
-        new_key = parent_key + sep + k if parent_key else k
+        new_key = generate_key(parent_key, sep, k)
         if isinstance(v, collections.MutableMapping):
             items.extend(flatten(v, new_key, sep=sep).items())
         else:
             items.append((new_key, str(v) if type(v) is list else v))
     return dict(items)
-        
+
+def get_headers(schema, parent_key=DEFAULT_PARENT_KEY, sep=SEP):
+    headers = []
+    for k, v in schema.get('properties', {}).items():
+        new_key = generate_key(parent_key, sep, k)
+        if v.get('type') == 'object':
+            headers = headers + get_headers(v, new_key, sep=sep)
+        else:
+            headers.append(new_key)
+    return list(headers)
+
+def generate_key(parent_key, sep, key):
+    return parent_key + sep + key if parent_key else key
+
 def persist_messages(delimiter, quotechar, messages, destination_path):
     state = None
     schemas = {}
     key_properties = {}
-    headers = {}
     validators = {}
 
     now = datetime.now().strftime('%Y%m%dT%H%M%S')
@@ -64,19 +79,10 @@ def persist_messages(delimiter, quotechar, messages, destination_path):
 
             flattened_record = flatten(o['record'])
 
-            if o['stream'] not in headers and not file_is_empty:
-                with open(filename, 'r') as csvfile:
-                    reader = csv.reader(csvfile,
-                                        delimiter=delimiter,
-                                        quotechar=quotechar)
-                    first_line = next(reader)
-                    headers[o['stream']] = first_line if first_line else flattened_record.keys()
-            else:
-                headers[o['stream']] = flattened_record.keys()
-
+            headers = get_headers(schemas[o['stream']])
             with open(filename, 'a') as csvfile:
                 writer = csv.DictWriter(csvfile,
-                                        headers[o['stream']],
+                                        headers,
                                         extrasaction='ignore',
                                         delimiter=delimiter,
                                         quotechar=quotechar)
diff --git a/target_csv_test.py b/target_csv_test.py
new file mode 100644
index 0000000..078f444
--- /dev/null
+++ b/target_csv_test.py
@@ -0,0 +1,37 @@
+import target_csv
+import unittest
+
+
+class TestTargetCsv(unittest.TestCase):
+
+  def setUp(self):
+    pass
+
+  def test_get_headers(self):
+    schema = {"properties": {
+      'a': {'type': 'string'},
+      'b': {'type': 'array', 'items': {'type': 'string'}},
+      'c': {'type': 'object', 'properties': {'d': {'type': 'string'}}}
+    }}
+
+    assert target_csv.get_headers(schema) == ['a', 'b', 'c__d']
+
+  def test_get_headers_matches_flatten(self):
+    schema = {'properties': {
+      'a': {'type': 'string'},
+      'b': {'type': 'array', 'items': {'type': 'string'}},
+      'c': {'type': 'object', 'properties': {'d': {'type': 'string'}}}
+    }}
+
+    record = {
+      'a': 'alpha',
+      'b': ['beta'],
+      'c': {'d': 'delta'}
+    }
+
+    keys_from_schema = target_csv.get_headers(schema)
+    keys_from_records = list(target_csv.flatten(record).keys())
+    assert keys_from_schema == keys_from_records
+
+if __name__ == '__main__':
+  unittest.main()