Maintenance and verification ledger

Reviewed 2026-10-01 against statsmodels 0.15.0, with Python 3.13.3, NumPy 2.5.3, SciPy 1.18.1, pandas 3.0.6, matplotlib 3.11.2 and scikit-learn 1.9.1. The package itself supports Python 3.10+; this exact numerical stack requires 3.12+. No remote service endpoint, authentication or pagination is part of these statistical workflows.

The current release notes and tagged source were checked with the installed wheel. Forty-seven source files covering the model, diagnostic, graphics and treatment imports matched the release tag byte for byte. The API audit included constructors, fitting options, result attributes and prediction semantics in every reference. Source inspection is distinct from numerical execution.

The quick start is executable in order. Other references contain contextual fragments: provide their named inputs and retain the appropriate model result for each section; they are not single scripts to concatenate. Native synthetic checks cover the corrected contracts, linear/GLM/discrete/time-series models and diagnostics. They establish API behavior and numerical identities, not estimator calibration, power for a real study, causal identification, or validity of a fitted scientific model. Optional custom MLEModel subclasses, manual proportional-odds tests and arbitrary user datasets remain illustrative and require separate validation.

Primary API sources and consequential contracts

• OLS, robust covariance, RecursiveLS, MixedLM.fit: intercept/design alignment; recursive estimates are expanding-window estimates; ML versus REML and covariance assumptions matter.
• GLM.fit, GLM prediction, pseudo R-squared: specify robust covariance during fit, request the intended prediction target, use likelihood-based pseudo-R² and bic_llf, and preserve exposure for new data.
• Marginal effects, ZIP prediction, OrderedModel, ConditionalLogit, hurdle: overall gives AME; atexog keys are column numbers; inflation differs from observing zero; ordinal/conditional models prohibit a constant; ordinal threshold increments need transformation. Hurdle zero/count processes share exog.
• ADF, KPSS, SARIMAX forecasting, Ljung-Box, VARMAX: use named result objects, distinguish test nulls, supply future exog, account for fitted dynamic parameters and initialization, and avoid unidentified unrestricted VARMA specifications. MarkovRegression.order does not add AR coefficients.
• RESET, AnovaRM, pairwise comparisons, proportion effect size, SciPy Levene: RESET returns a result object; repeated-measures ANOVA requires balanced data and lacks sphericity corrections; 0.15 adds Games-Howell via use_var="unequal"; proportion power uses Cohen's h, and CompareMeans has no test_equal_var method.
• TreatmentEffect and its released implementation: OLS outcome modeling remains the supported route. ipw_ra/aipw_wls GMM still slice the final six selection parameters; non-six-parameter designs raise a shape error. The point-estimate-only route does not provide standard errors. Do not change a scientific design to accommodate that implementation bug.

Executed checks and limits

The isolated suite runs eight native tests, including all 173 Python fenced blocks with explicit synthetic contexts (some blocks only import or explain a workflow). Assertions cover design shapes, OLS interval width, GLM mean uncertainty and robust covariance, AME derivatives, ZIP mixture probabilities, ordinal thresholds, multinomial results, selected ARIMA convergence and preserved forecast dates, and treatment-effect identities plus the known GMM failure.

The broad ARIMA grid emits expected nonconvergence warnings for some candidates; selection excludes those candidates and retains the selected result. KPSS reports table-boundary p-values. Gamma inverse/binomial log links emit domain warnings and need an explicit support check on real data. Johansen testing emitted complex-to-real cast warnings in the tested NumPy stack; the synthetic eigenvalues had exactly zero imaginary parts and finite trace/max-eigenvalue statistics. That fixture check is not a general justification to discard complex components on other data. Plotting was exercised with the noninteractive Agg backend; no publication-layout certification or empirical interval coverage study was performed.