Bridge API naming (ATT_glob) with formal terminology (ATT_loc/ATT_glob)

Update tutorial estimand table, global parameter interpretation, and
binary comparison sections to clarify that the API's ATT_glob/overall_att
identifies ATT_loc under standard PT and equals ATT_glob only under SPT.

Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

Author: igerber

docs/tutorials/14_continuous_did.ipynb

@@ -76,7 +76,7 @@
     "|-----------|---------------------------|\n",
     "| ATT(d) | Total earnings gain from *d* hours of training vs. no training, across all treated workers (requires strong PT) |\n",
     "| ACRT(d) | Earnings gain from **one additional hour** at level *d* (marginal effect; requires strong PT) |\n",
-    "| ATT_glob | Average effect of any training vs. none (binarized; valid under standard PT) |\n",
+    "| ATT_glob | Average effect of any training vs. none (binarized). Identifies ATT_loc under standard PT; equals ATT_glob under strong PT |\n",
     "| ACRT_glob | Average marginal return to an extra hour of training (requires strong PT) |\n",
     "\n",
     "The `ContinuousDiD` estimator from Callaway, Goodman-Bacon & Sant'Anna (2024) handles all three problems. It estimates the full dose-response curve using B-splines and provides valid inference."
@@ -226,10 +226,8 @@
    "source": [
     "### Interpreting the global parameters\n",
     "\n",
-    "- **ATT_glob** is the average total effect of training (at the observed dose levels) compared to no training. It answers: \"On average, how much more do trained workers earn?\"\n",
-    "- **ACRT_glob** is the average marginal return to an additional hour of training. It answers: \"What is one more hour of training worth, on average?\"\n",
-    "\n",
-    "Note: ACRT_glob and the dose-response interpretation of ATT_glob require strong parallel trends (see assumptions above). Under standard PT, the binarized overall effect is still identified.\n",
+    "- **ATT_glob** is the average total effect of training (at the observed dose levels) compared to no training. It answers: \"On average, how much more do trained workers earn?\" The API reports this as `overall_att`. Under standard PT, this identifies the local average effect (ATT_loc); under strong PT, it additionally equals the global average (ATT_glob).\n",
+    "- **ACRT_glob** is the average marginal return to an additional hour of training. It answers: \"What is one more hour of training worth, on average?\" This requires strong PT for causal interpretation.\n",
     "\n",
     "Analytical standard errors are the default (fast, no resampling). For bootstrap-based inference, set `n_bootstrap` (see Section 6)."
    ]
@@ -568,9 +566,9 @@
    "source": [
     "## 7. Comparison to Binary DiD\n",
     "\n",
-    "What if we ignore dose entirely and just run a standard binary Callaway-Sant'Anna estimator? Both approaches should give a similar **overall ATT** (treated vs. untreated), but the binary approach discards all dose information — no dose-response curve, no marginal effects.\n",
+    "What if we ignore dose entirely and just run a standard binary Callaway-Sant'Anna estimator? Both approaches should give a similar **binarized ATT** (treated vs. untreated), but the binary approach discards all dose information — no dose-response curve, no marginal effects.\n",
     "\n",
-    "Note: both estimators compute ATT_glob as a binarized mean difference (treated vs. untreated), so the values should be very close. Any small differences arise from weighting or aggregation choices, control group or base period settings, or finite-sample variation — not from spline smoothing. The continuous approach provides the full dose-response curve on top of the overall effect."
+    "Note: both estimators compute the binarized ATT as a simple mean difference (treated vs. untreated), so the values should be very close. Under standard PT this identifies ATT_loc (the local average); under strong PT it additionally equals ATT_glob. Any small differences arise from weighting or aggregation choices, control group or base period settings, or finite-sample variation — not from spline smoothing. The continuous approach provides the full dose-response curve on top of the binarized effect."
    ]
   },
   {