bossanova 0.1.0.dev19__tar.gz → 0.1.0.dev20__tar.gz

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/.gitignore

@@ -17,7 +17,7 @@ docs/_build/
 docs/performance/
 profile_*.py
 benchmarks/
-!bossanova-docs/engineering/benchmarks/
+!bossanova-docs/benchmarking/
 examples/
 papers/
 paper-summaries/
@@ -84,3 +84,4 @@ bossanova-docs/wip/
 research.md
 201b-ghct-05-models/
 emmeans/
+walkthrough.md

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: bossanova
-Version: 0.1.0.dev19
+Version: 0.1.0.dev20
 Summary: Bridging statistical cultures with some jazz
 Author: Eshin Jolly
 License-Expression: MIT

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/bossanova/internal/fit/dispatch.py

@@ -105,7 +105,7 @@ def dispatch_solver(
     tol: float | None = None,
     verbose: bool = False,
     nAGQ: int = 1,
-    use_hessian: bool = True,
+    use_hessian: bool = False,
 ) -> FitState:
     """Dispatch to the concrete solver implementation.
 
@@ -121,7 +121,7 @@ def dispatch_solver(
         tol: Convergence tolerance (solver-specific defaults if None).
         verbose: Print optimization progress (default: False).
         nAGQ: Quadrature points for GLMER (default: 1).
-        use_hessian: Use Hessian-based vcov for GLMER (default: True).
+        use_hessian: Use Hessian-based vcov for GLMER (default: False).
 
     Returns:
         FitState from the selected solver.
@@ -172,7 +172,7 @@ def fit_model(
     tol: float | None = None,
     verbose: bool = False,
     nAGQ: int = 1,
-    use_hessian: bool = True,
+    use_hessian: bool = False,
 ) -> FitState:
     """Dispatch to appropriate fitter based on model specification.
 
@@ -205,7 +205,7 @@ def fit_model(
         tol: Convergence tolerance (solver-specific defaults if None).
         verbose: Print optimization progress (default: False).
         nAGQ: Quadrature points for GLMER (default: 1).
-        use_hessian: Use Hessian-based vcov for GLMER (default: True).
+        use_hessian: Use Hessian-based vcov for GLMER (default: False).
 
     Returns:
         FitState containing all fitting results.

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/bossanova/internal/fit/glmer.py

@@ -29,7 +29,7 @@ def fit_glmer_pirls(
     tol: float = 1e-7,
     verbose: bool = False,
     nAGQ: int = 1,
-    use_hessian: bool = True,
+    use_hessian: bool = False,
 ) -> FitState:
     """Fit generalized linear mixed model using Penalized IRLS.
 
@@ -70,7 +70,10 @@ def fit_glmer_pirls(
         tol: PIRLS convergence tolerance (default: 1e-7).
         verbose: Print optimization progress (default: False).
         nAGQ: Quadrature points (0 or 1, default: 1).
-        use_hessian: Use Hessian-based vcov (default: True).
+        use_hessian: Use Hessian-based vcov (default: False). The default
+            Schur complement approach matches lme4's ``vcov()`` with
+            ``use.hessian=FALSE`` and avoids expensive numerical
+            differentiation. Set to True for observed-information vcov.
 
     Returns:
         FitState containing:
@@ -91,7 +94,6 @@ def fit_glmer_pirls(
         - bossanova.internal.maths.solvers.glmer: Underlying PIRLS implementation
     """
     from bossanova.internal.maths.solvers.glmer import (
-        compute_hessian_vcov,
         compute_irls_quantities,
         fit_glmm_pirls,
     )
@@ -177,9 +179,17 @@ def fit_glmer_pirls(
     converged = fit_result["converged"]
     n_iter = fit_result["n_func_evals"]
 
+    # Compute IRLS weights once (needed by both vcov paths and for XtWX_inv)
+    working_weights, _ = compute_irls_quantities(y, eta, family)
+    if weights is not None:
+        total_weights = weights * working_weights
+    else:
+        total_weights = working_weights
+
     # Compute variance-covariance matrix of beta
     if use_hessian:
-        # Hessian-based vcov (matches lme4 default)
+        from bossanova.internal.maths.solvers.glmer import compute_hessian_vcov
+
         vcov, _ = compute_hessian_vcov(
             theta=theta,
             beta=beta,
@@ -196,15 +206,7 @@ def fit_glmer_pirls(
             eta_init=eta,  # Use optimal eta as starting point
         )
     else:
-        # Schur complement-based vcov (faster)
-        working_weights, _ = compute_irls_quantities(y, eta, family)
-
-        if weights is not None:
-            total_weights = weights * working_weights
-        else:
-            total_weights = working_weights
-
-        # Build Lambda for Schur complement
+        # Schur complement-based vcov (analytical, matches lme4 vcov(use.hessian=FALSE))
         Lambda = build_lambda_sparse(theta, n_groups_list, re_structure, metadata)
         vcov = compute_vcov_schur_sparse(X, Z, Lambda, weights=total_weights)
 
@@ -218,13 +220,8 @@ def fit_glmer_pirls(
     # Degrees of freedom
     df_resid = float(n - p)
 
-    # Compute final IRLS weights and XtWX_inv for cluster-robust SEs
-    final_working_weights, _ = compute_irls_quantities(y, eta, family)
-    if weights is not None:
-        final_irls_weights = weights * final_working_weights
-    else:
-        final_irls_weights = final_working_weights
-    XtWX = X.T @ (X * final_irls_weights[:, np.newaxis])
+    # XtWX_inv for sandwich/cluster-robust SEs (reuse total_weights)
+    XtWX = X.T @ (X * total_weights[:, np.newaxis])
     XtWX_inv = np.linalg.inv(XtWX)
 
     return build_fit_state(
@@ -240,6 +237,6 @@ def fit_glmer_pirls(
         u=u,
         converged=converged,
         n_iter=n_iter,
-        irls_weights=final_irls_weights,
+        irls_weights=total_weights,
         XtWX_inv=XtWX_inv,
     )

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/bossanova/internal/marginal/slopes.py

@@ -312,6 +312,15 @@ def compute_slopes_finite_diff(
 
     validate_focal_var(bundle, focal_var)
 
+    # --- drop rows with missing values in model variables -----------------
+    # The caller may pass raw data that includes rows dropped during fitting.
+    # Null/NaN values propagate through matrix multiplication and corrupt the
+    # AME estimate, so filter to complete cases on the model's columns.
+    model_cols = [c for c in bundle.X_names if c in data.columns]
+    if bundle.y_name in data.columns:
+        model_cols.append(bundle.y_name)
+    data = data.drop_nulls(subset=model_cols)
+
     # --- step size --------------------------------------------------------
     focal_col = data[focal_var].cast(pl.Float64).to_numpy()
     var_range = float(focal_col.max() - focal_col.min())
@@ -443,6 +452,12 @@ def compute_slopes_crossed(
             "(needed by evaluate_newdata)."
         )
 
+    # --- drop rows with missing values in model variables -----------------
+    model_cols = [c for c in bundle.X_names if c in data.columns]
+    if bundle.y_name in data.columns:
+        model_cols.append(bundle.y_name)
+    data = data.drop_nulls(subset=model_cols)
+
     # --- step size --------------------------------------------------------
     focal_col = data[focal_var].cast(pl.Float64).to_numpy()
     var_range = float(focal_col.max() - focal_col.min())

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/bossanova/model/core.py

@@ -1028,7 +1028,7 @@ class model:
             "tol": kwargs.pop("tol", None),
             "verbose": bool(kwargs.pop("verbose", False)),
             "nAGQ": int(kwargs.pop("nAGQ", 1)),
-            "use_hessian": bool(kwargs.pop("use_hessian", True)),
+            "use_hessian": bool(kwargs.pop("use_hessian", False)),
         }
         if fit_kwargs["max_iter"] is not None:
             fit_kwargs["max_iter"] = int(fit_kwargs["max_iter"])

{bossanova-0.1.0.dev19 → bossanova-0.1.0.dev20}/pyproject.toml

@@ -1,7 +1,7 @@
 [project]
 name = "bossanova"
 requires-python = ">=3.11"
-version = "0.1.0.dev19"
+version = "0.1.0.dev20"
 authors = [
     {name = "Eshin Jolly"},
 ]
@@ -167,6 +167,10 @@ setup-kernel = { cmd = "python -m ipykernel install --user --name=\"$(basename $
 _tutorials-export = { cmd = "bash scripts/tutorials_export.sh", description = "Convert marimo notebooks to MyST markdown" }
 _tutorials-import = { cmd = "bash scripts/tutorials_import.sh", description = "Convert MyST markdown to marimo notebooks" }
 
+# === Benchmarks ===
+# `pixi run benchmark lm`, `pixi run benchmark glm`, etc. Runs --all-backends by default.
+benchmark = { cmd = "bash scripts/run-logged.sh benchmark-{{ model }} python benchmarking/benchmark_{{ model }}_methods.py --all-backends", args = [{ arg = "model", choices = ["lm", "glm", "lmer", "glmer"] }], description = "Run model benchmarks (writes markdown to docs)" }
+
 # === Build & Release ===
 _build = { cmd = "uv build --clear", description = "Build sdist + wheel (clears old artifacts)" }
 release = { cmd = "python scripts/release.py {{ target }}", args = [{ arg = "target", choices = ["pypi", "testpypi"], default = "pypi" }], description = "Full release: review logs, bump version, build, test, publish" }