diff --git a/sharrow/aster.py b/sharrow/aster.py
index 806553a..f0e655b 100755
--- a/sharrow/aster.py
+++ b/sharrow/aster.py
@@ -886,6 +886,20 @@ def visit_Call(self, node):
                 )
                 result = ast.BinOp(left=left, op=ast.BitAnd(), right=right)
 
+        # change XXX.isna() [with no arguments] to np.isnan(x)
+        if (
+            isinstance(node.func, ast.Attribute)
+            and node.func.attr == "isna"
+            and len(node.args) == 0
+            and len(node.keywords) == 0
+        ):
+            apply_args = [self.visit(node.func.value)]  # convert XXX into argument
+            result = ast.Call(
+                func=ast.Name("isnan_fast_safe"),
+                args=apply_args,
+                keywords=[],
+            )
+
         # if no other changes
         if result is None:
             args = [self.visit(i) for i in node.args]
diff --git a/sharrow/flows.py b/sharrow/flows.py
index 3a23437..35c2881 100644
--- a/sharrow/flows.py
+++ b/sharrow/flows.py
@@ -1073,7 +1073,7 @@ def __initialize_2(
                 "from contextlib import suppress",
                 "from numpy import log, exp, log1p, expm1",
                 "from sharrow.maths import piece, hard_sigmoid, transpose_leading, clip, digital_decode",
-                "from sharrow.sparse import get_blended_2",
+                "from sharrow.sparse import get_blended_2, isnan_fast_safe",
             }
 
             func_code = self._func_code
@@ -1342,7 +1342,16 @@ def load_raw(self, rg, args, runner=None, dtype=None, dot=None):
                     if arg.startswith("__aux_var"):
                         arguments.append(arg_value)
                     else:
-                        arguments.append(np.asarray(arg_value))
+                        arg_value_array = np.asarray(arg_value)
+                        if arg_value_array.dtype.kind == "O":
+                            # convert object arrays to unicode str
+                            # and replace missing values with NAK='\u0015'
+                            # that can be found by `isnan_fast_safe`
+                            # This is done for compatability and likely ruins performance
+                            arg_value_array_ = arg_value_array.astype("unicode")
+                            arg_value_array_[pd.isnull(arg_value_array)] = "\u0015"
+                            arg_value_array = arg_value_array_
+                        arguments.append(arg_value_array)
                 kwargs = {}
                 if dtype is not None:
                     kwargs["dtype"] = dtype
@@ -1421,8 +1430,15 @@ def iload_raw(
                         arguments.append(argument)
                     else:
                         if argument.dtype.kind == "O":
-                            argument = argument.astype("unicode")
-                        arguments.append(np.asarray(argument))
+                            # convert object arrays to unicode str
+                            # and replace missing values with NAK='\u0015'
+                            # that can be found by `isnan_fast_safe`
+                            # This is done for compatability and likely ruins performance
+                            argument_ = argument.astype("unicode")
+                            argument_[pd.isnull(argument)] = "\u0015"
+                            arguments.append(np.asarray(argument_))
+                        else:
+                            arguments.append(np.asarray(argument))
                 kwargs = {}
                 if dtype is not None:
                     kwargs["dtype"] = dtype
diff --git a/sharrow/sparse.py b/sharrow/sparse.py
index 444e526..2adc8f0 100644
--- a/sharrow/sparse.py
+++ b/sharrow/sparse.py
@@ -197,12 +197,21 @@ def blenders(self):
         return b
 
 
-@nb.njit
+@nb.generated_jit(nopython=True)
 def isnan_fast_safe(x):
-    if int(x) == -9223372036854775808:
-        return True
+    if isinstance(x, nb.types.Float):
+
+        def func(x):
+            if int(x) == -9223372036854775808:
+                return True
+            else:
+                return False
+
+        return func
+    elif isinstance(x, (nb.types.UnicodeType, nb.types.UnicodeCharSeq)):
+        return lambda x: x == "\u0015"
     else:
-        return False
+        return lambda x: False
 
 
 @nb.njit
diff --git a/sharrow/tests/test_relationships.py b/sharrow/tests/test_relationships.py
index 2a771fa..216fc7b 100644
--- a/sharrow/tests/test_relationships.py
+++ b/sharrow/tests/test_relationships.py
@@ -4,7 +4,7 @@
 import numpy as np
 import pandas as pd
 from numpy.random import SeedSequence, default_rng
-from pytest import mark, raises
+from pytest import approx, mark, raises
 
 import sharrow
 from sharrow import Dataset, DataTree, example_data
@@ -773,3 +773,34 @@ def test_nested_where(dataframe_regression):
         check_names=False,
     )
     dataframe_regression.check(result)
+
+
+def test_isna():
+    data = example_data.get_data()
+    data["hhs"].loc[data["hhs"].income > 200000, "income"] = np.nan
+    tree = DataTree(
+        base=data["hhs"],
+    )
+    ss = tree.setup_flow(
+        {
+            "missing_income": "((income < 0) | income.isna())",
+            "income_is_na": "income.isna()",
+        }
+    )
+    result = ss.load()
+    assert result[0, 0] == 1
+    assert result[0, 1] == 1
+    assert result[:, 0].sum() == 188
+    assert result[:, 1].sum() == 188
+
+    qf = pd.DataFrame({"MixedVals": ["a", "", None, np.nan]})
+    tree2 = DataTree(
+        base=qf,
+    )
+    qf = tree2.setup_flow(
+        {
+            "MixedVals_is_na": "MixedVals.isna()",
+        }
+    )
+    result = qf.load()
+    assert result == approx(np.asarray([[0, 0, 1, 1]]).T)