fix: check realized never-treated count, not just fraction

Small never_treated_frac values (e.g., 0.01 with n_units=50) floor to
zero never-treated units via int(), silently breaking the conditional_pt
contract. Now checks int(n_units * never_treated_frac) >= 1 and reports
the realized count in the error message.

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: igerber

diff_diff/prep_dgp.py

@@ -1311,8 +1311,8 @@ def generate_survey_did_data(
         Conditional on x1, trends remain parallel (conditional PT holds).
         DR/IPW estimators with covariates recover truth; no-covariate
         estimators are biased. Uses normalized time (t/n_periods) for
-        scale independence. Requires ``add_covariates=True`` and
-        ``never_treated_frac > 0`` (the x1 mean shift only differentiates
+        scale independence. Requires ``add_covariates=True`` and at least
+        one never-treated unit (the x1 mean shift only differentiates
         ever-treated from never-treated units).
 
         .. note:: When used with ``icc``, the ICC calibration is approximate
@@ -1437,12 +1437,14 @@ def generate_survey_did_data(
         )
     if conditional_pt != 0.0 and not add_covariates:
         raise ValueError("conditional_pt requires add_covariates=True")
-    if conditional_pt != 0.0 and never_treated_frac == 0.0:
+    if conditional_pt != 0.0 and int(n_units * never_treated_frac) < 1:
         raise ValueError(
-            "conditional_pt requires never_treated_frac > 0. The x1 mean shift "
-            "applies to all ever-treated units; without a never-treated group, "
-            "treated and control units share the same x1 distribution and "
-            "unconditional parallel trends are not violated."
+            "conditional_pt requires at least one never-treated unit "
+            f"(n_units={n_units}, never_treated_frac={never_treated_frac} "
+            f"yields {int(n_units * never_treated_frac)} never-treated). "
+            "The x1 mean shift applies to all ever-treated units; without a "
+            "never-treated group, treated and control units share the same x1 "
+            "distribution and unconditional parallel trends are not violated."
         )
 
     # --- ICC -> psu_re_sd resolution ---

docs/methodology/REGISTRY.md

@@ -2708,8 +2708,8 @@ The 8-step workflow in `docs/llms-practitioner.txt` is adapted from Baker et al.
   E[x1 | treated] != E[x1 | control], the average time trend differs by group
   (unconditional PT fails). Conditional on x1, trends are identical (conditional
   PT holds). DR/IPW estimators with x1 as covariate recover the true ATT.
-  Requires `never_treated_frac > 0`; rejected otherwise because the x1 mean
-  shift only differentiates ever-treated from never-treated units.
+  Requires at least one never-treated unit (rejected otherwise because the x1
+  mean shift only differentiates ever-treated from never-treated units).
 - **Note:** When `conditional_pt != 0` is combined with `icc`, the ICC
   calibration is approximate. The x1 mean shift creates a mixture distribution
   with marginal Var(x1) = 1 + p_treated * (1 - p_treated) > 1, slightly

tests/test_prep.py

@@ -1969,14 +1969,21 @@ def test_te_covariate_interaction_validation(self):
     # --- conditional_pt parameter tests ---
 
     def test_conditional_pt_requires_never_treated(self):
-        """conditional_pt requires never_treated_frac > 0."""
+        """conditional_pt requires at least one never-treated unit."""
         from diff_diff.prep_dgp import generate_survey_did_data
 
-        with pytest.raises(ValueError, match="conditional_pt requires never_treated_frac"):
+        # Exact zero fraction
+        with pytest.raises(ValueError, match="conditional_pt requires at least one"):
             generate_survey_did_data(
                 add_covariates=True, conditional_pt=0.3,
                 never_treated_frac=0.0, seed=42,
             )
+        # Small fraction that floors to zero never-treated units
+        with pytest.raises(ValueError, match="conditional_pt requires at least one"):
+            generate_survey_did_data(
+                n_units=50, add_covariates=True, conditional_pt=0.3,
+                never_treated_frac=0.01, seed=42,
+            )
 
     def test_conditional_pt_requires_covariates(self):
         """conditional_pt requires add_covariates=True."""