Aggregates several covariance matrices (e.g., from bootstrap samples or model components) into one combined covariance matrix using optional weights.

Signals the intent to merge uncertainty estimates from multiple sources into a representative covariance matrix.

When aggregating covariance estimates from bootstrap resamples, cross‑validation folds, or ensemble model components to obtain a more stable uncertainty estimate.

Statistical modeling, meta‑analysis, ensemble learning, sandwich estimator aggregation, and any setting where multiple covariance estimates need to be pooled.

cov.combine(*covariances, weights=None)

cov (covariance object) . combine ( method call ) ( *covariances: sequence, weights: optional sequence )

cov: covariance‑like object possessing a combine method; *covariances: one or more covariance matrices or objects that support combine; weights: optional sequence of non‑negative floats whose length matches the number of covariances

• statsmodels.stats.sandwich_covariance
• bootstrapped covariance lists
• ensemble model components
• numpy.cov
• scipy.linalg.block_diag

• numpy.cov for a simple pooled covariance
• scipy.linalg.block_diag for block‑diagonal stacking
• manual weighted average using np.average

Assuming equal weighting when heterogeneous variances exist → biased pooled covariance;\nPassing raw arrays instead of covariance objects → AttributeError: 'numpy.ndarray' has no attribute 'combine';\nProviding weights that do not sum to one → unintentionally scaled covariance;\nMixing incompatible matrix shapes → ValueError: shape mismatch;\nForgetting to set weights=None when not needed → silent misuse of default weighting.

numpy.cov – computes covariance from raw data rather than combining pre‑computed matrices;\nscipy.linalg.block_diag – stacks covariances diagonally without mixing information;\nstatsmodels.stats.sandwich_covariance.sandwich – computes a robust covariance from a single model, not a combination.

Coming from MATLAB: may use cov() directly on raw data instead of the object method → use cov.combine() for object‑oriented combination;\nComing from R: may think cov.combine works like cov() with a weights argument → remember the method expects covariance objects, not raw data.

• covariance combination
• ensemble uncertainty pooling

1) Avoid when covariances come from vastly different sample sizes without appropriate weighting – can distort uncertainty;\n2) Computationally cheap for small matrices but O(n³) for large dense covariances due to matrix addition;\n3) Assumes all input covariances are symmetric positive‑semidefinite; feeding indefinite matrices can yield an invalid pooled covariance.

Provides a stable, aggregated uncertainty measure that improves reliability of downstream inference, hypothesis testing, and confidence‑interval construction in ensemble or meta‑analytic contexts.

Think of cov.combine() as a financial analyst blending multiple expert forecasts into a single consensus estimate, weighting each opinion by its confidence.

Consider cov.pool() or cov.aggregate() for advanced weighting schemes or hierarchical combination.