panelkit 0.2.6__tar.gz → 0.2.8__tar.gz

{panelkit-0.2.6 → panelkit-0.2.8}/Cargo.lock

@@ -462,7 +462,7 @@ checksum = "d6790f58c7ff633d8771f42965289203411a5e5c68388703c06e14f24770b41e"
 
 [[package]]
 name = "panelkit-estimators"
-version = "0.2.6"
+version = "0.2.8"
 dependencies = [
  "criterion",
  "panelkit-linalg",
@@ -471,7 +471,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-geo"
-version = "0.2.6"
+version = "0.2.8"
 dependencies = [
  "panelkit-estimators",
  "panelkit-inference",
@@ -482,7 +482,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-inference"
-version = "0.2.6"
+version = "0.2.8"
 dependencies = [
  "panelkit-estimators",
  "panelkit-linalg",
@@ -491,7 +491,7 @@ dependencies = [
 
 [[package]]
 name = "panelkit-linalg"
-version = "0.2.6"
+version = "0.2.8"
 dependencies = [
  "proptest",
  "rayon",
@@ -623,7 +623,7 @@ dependencies = [
 
 [[package]]
 name = "pypanelkit"
-version = "0.2.6"
+version = "0.2.8"
 dependencies = [
  "numpy",
  "panelkit-estimators",

{panelkit-0.2.6 → panelkit-0.2.8}/Cargo.toml

@@ -3,7 +3,7 @@ resolver = "2"
 members = ["crates/linalg", "crates/estimators", "crates/inference", "crates/geo", "crates/pypanelkit"]
 
 [workspace.package]
-version = "0.2.6"
+version = "0.2.8"
 edition = "2021"
 rust-version = "1.74"
 license = "MIT OR Apache-2.0"

{panelkit-0.2.6 → panelkit-0.2.8}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: panelkit
-Version: 0.2.6
+Version: 0.2.8
 Classifier: Programming Language :: Rust
 Classifier: Programming Language :: Python :: 3
 Classifier: Topic :: Scientific/Engineering

{panelkit-0.2.6 → panelkit-0.2.8}/crates/inference/src/bootstrap.rs

@@ -53,14 +53,13 @@ pub fn multiplier_bootstrap(
 }
 
 /// Jackknife (leave-one-out) standard error from a set of leave-one-out
-/// estimates and the full-sample estimate.
-pub fn jackknife_se(loo_estimates: &[f64], full: f64) -> f64 {
+/// estimates: `sqrt((n-1)/n · Σ(θ_i − θ̄)²)`, centered on the LOO mean.
+pub fn jackknife_se(loo_estimates: &[f64]) -> f64 {
     let n = loo_estimates.len();
     if n < 2 {
         return 0.0;
     }
     let mean = loo_estimates.iter().sum::<f64>() / n as f64;
-    let _ = full;
     let ss: f64 = loo_estimates.iter().map(|x| (x - mean).powi(2)).sum();
     ((n as f64 - 1.0) / n as f64 * ss).sqrt()
 }

{panelkit-0.2.6 → panelkit-0.2.8}/crates/inference/tests/inference.rs

@@ -107,11 +107,11 @@ fn bootstrap_engines_thread_count_invariant() {
 #[test]
 fn jackknife_se_of_constant_is_zero() {
     let est = vec![2.0; 10];
-    assert!(jackknife_se(&est, 2.0) < 1e-12);
+    assert!(jackknife_se(&est) < 1e-12);
 }
 
 #[test]
 fn jackknife_se_positive_for_varying() {
     let est = vec![1.0, 2.0, 3.0, 4.0, 5.0];
-    assert!(jackknife_se(&est, 3.0) > 0.0);
+    assert!(jackknife_se(&est) > 0.0);
 }

{panelkit-0.2.6 → panelkit-0.2.8}/pyproject.toml

@@ -4,7 +4,7 @@ build-backend = "maturin"
 
 [project]
 name = "panelkit"
-version = "0.2.6"
+version = "0.2.8"
 description = "Fast, from-scratch causal-inference estimators for panel/geo experiments (SC, ASC, SDID, DiD, MC-NNM)."
 readme = "README.md"
 requires-python = ">=3.9"

{panelkit-0.2.6 → panelkit-0.2.8}/python/panelkit/design.py

@@ -774,12 +774,13 @@ class GeoDesign:
         if unknown:
             raise ValueError(f"unknown methods {unknown}; choose from {_METHODS}")
 
-        def _fit(method, tr):
+        def _fit(method, tr, Y=None):
+            Y = self.Y if Y is None else Y
             if method == "SC":
-                return _panelkit.fit_sc(self.Y, tr, t0, 0.0, False, level)
+                return _panelkit.fit_sc(Y, tr, t0, 0.0, False, level)
             if method == "ASC":
-                return _panelkit.fit_asc(self.Y, tr, t0, 0.0, None)
-            return _panelkit.fit_sdid(self.Y, tr, t0, 1.0)
+                return _panelkit.fit_asc(Y, tr, t0, 0.0, None)
+            return _panelkit.fit_sdid(Y, tr, t0, 1.0)
 
         treated_series = self.Y[idx].mean(axis=0)
         post_len = self.t - t0
@@ -826,17 +827,25 @@ class GeoDesign:
 
         # --- engine: per-method null att-samples (+ donor placebo paths if used) ---
         if inference == "placebo":
+            # Refit each donor as if treated, on the donor-ONLY sub-panel so the
+            # real treated unit(s) are excluded from every placebo's control pool
+            # (textbook Abadie — no leakage of the actual effect into the null).
             treated_set = set(idx)
-            donors = [u for u in range(self.n) if u not in treated_set]
-            if len(donors) > int(max_placebo):
+            all_donors = [u for u in range(self.n) if u not in treated_set]
+            ydon = self.Y[all_donors]                    # rows = donors only
+            nd = len(all_donors)
+            rows = list(range(nd))
+            if nd > int(max_placebo):
                 rng = np.random.default_rng(int(seed))
-                donors = sorted(int(j) for j in
-                                rng.choice(donors, int(max_placebo), replace=False))
+                rows = sorted(int(r) for r in rng.choice(nd, int(max_placebo), replace=False))
             pb = {m: [] for m in methods}    # per method: list of (att_path, pre_rmspe)
-            for j in donors:
-                for m in methods:
-                    fj = _fit(m, [j])
-                    pb[m].append((np.asarray(fj.att_path, dtype=float), float(fj.pre_rmspe)))
+            # Need >= 2 donors so a placebo (1 treated + >= 1 control) is fittable;
+            # otherwise leave pb empty → inference is reported undefined/low-power.
+            if nd >= 2:
+                for r in rows:
+                    for m in methods:
+                        fj = _fit(m, [r], ydon)
+                        pb[m].append((np.asarray(fj.att_path, dtype=float), float(fj.pre_rmspe)))
 
             def _kept_att(samples, treated_pre_m):
                 keep = [p.mean() for (p, pre) in samples
@@ -895,7 +904,7 @@ class GeoDesign:
         if inference == "placebo":
             treated_pre = sum(wmap[m] * per[m]["pre_rmspe"] for m in order)
             ens_pb = []
-            for di in range(len(donors)):
+            for di in range(len(pb[order[0]])):
                 path = sum(wmap[m] * pb[m][di][0] for m in order)
                 pre = sum(wmap[m] * pb[m][di][1] for m in order)
                 ens_pb.append((path, pre))