diff --git a/src/psycop_model_training/config_schemas/preprocessing.py b/src/psycop_model_training/config_schemas/preprocessing.py
index 035f081b..4ceead47 100644
--- a/src/psycop_model_training/config_schemas/preprocessing.py
+++ b/src/psycop_model_training/config_schemas/preprocessing.py
@@ -30,6 +30,9 @@ class PreSplitPreprocessingConfigSchema(BaseModel):
     convert_booleans_to_int: bool = False
     # Whether to convert columns containing booleans to int
 
+    negative_values_to_nan: bool = True
+    # Whether to change negative values to NaN. Defaults to True since Chi2 cannot handle negative values. Can only be set to True if Chi2 is not used for feature selection.
+
     drop_datetime_predictor_columns: bool = False
     # Whether to drop datetime columns prefixed with data.pred_prefix.
     # Typically, we don't want to use these as features, since they are unlikely to generalise into the future.
diff --git a/src/psycop_model_training/preprocessing/pre_split/processors/value_cleaner.py b/src/psycop_model_training/preprocessing/pre_split/processors/value_cleaner.py
index 932ea019..6ddfc1ad 100644
--- a/src/psycop_model_training/preprocessing/pre_split/processors/value_cleaner.py
+++ b/src/psycop_model_training/preprocessing/pre_split/processors/value_cleaner.py
@@ -65,7 +65,8 @@ def clean(self, dataset: pd.DataFrame) -> pd.DataFrame:
         # In the future, we want to:
         # 1a. See if there's a way of using feature selection that permits negative values, or
         # 1b. Always use z-score normalisation?
-        dataset = self._negative_values_to_nan(dataset=dataset)
+        if self.pre_split_cfg.negative_values_to_nan:
+            dataset = self._negative_values_to_nan(dataset=dataset)
         dataset = self.convert_timestamp_dtype_and_nat(dataset=dataset)
 
         return dataset