Supported models

Which function works on which model. The table is produced by running every function on one model of each kind; a dash means the function raises CheckNotApplicableError (or, for Bayesian models, points you to the functions that do apply).

Columns: OLS / WLS: statsmodels OLS/WLS results. sklearn linear: LinearRegression. sklearn penalized: Ridge, Lasso, ElasticNet and their CV variants. Gaussian GLM, Logistic GLM, Poisson GLM: statsmodels GLM results. sklearn logistic: binary LogisticRegression. MixedLM: statsmodels linear mixed models (random intercepts and slopes). Bayesian: PyMC and bambi models wrapped with bayesian() (Python 3.12+).

Function	OLS / WLS	sklearn linear	sklearn penalized	Gaussian GLM	Logistic GLM	sklearn logistic	Poisson GLM	MixedLM	Bayesian
check_predictions	✓	✓	✓	✓	✓	✓	✓	✓	✓
check_normality	✓	✓	✓	✓	–	–	–	✓	–
check_heteroscedasticity	✓	✓	✓	✓	–	–	–	✓	–
check_collinearity	✓	✓	✓	✓	✓	✓	✓	✓	–
check_outliers	✓	✓	✓	✓	✓	✓	✓	✓	–
check_autocorrelation	✓	✓	✓	✓	–	–	–	–	–
binned_residuals	–	–	–	–	✓	✓	–	–	–
check_overdispersion	–	–	–	✓	✓	✓	✓	✓	–
check_zeroinflation	–	–	–	–	–	–	✓	–	–
check_normality(effects='random')	–	–	–	–	–	–	–	✓	–
check_singularity	–	–	–	–	–	–	–	✓	–
check_convergence	✓	✓	✓	✓	✓	✓	✓	✓	✓
simulate_residuals, check_residuals	✓	✓	✓	✓	✓	✓	✓	✓	–
r2	✓ R2, R2_adjusted	✓ R2, R2_adjusted	✓ R2, R2_adjusted	✓ R2	✓ R2_Tjur	✓ R2_Tjur	✓ R2_Nagelkerke	✓ R2_conditional, R2_marginal	✓ R2
icc	–	–	–	–	–	–	–	✓	–
model_performance	✓	✓	✓	✓	✓	✓	✓	✓	✓

Notes per model type

• scikit-learn estimators need X= and y=, because they do not keep their training data. Penalized models have no likelihood, so model_performance() leaves out AIC, AICc and BIC; leverage and VIFs come from the design matrix, and the checks that use leverage or standardized residuals warn that they are approximate. PoissonRegressor and GammaRegressor behave like the Poisson and other GLM columns (penalized unless alpha=0).
• statsmodels discrete models (Logit, Probit, Poisson, NegativeBinomial) behave like the GLM of the same family and link. A negative binomial fit (NB2 or geometric) is treated like R's glm.nb(); NB1, multinomial, generalized Poisson and zero-inflated models are not supported yet.
• Logistic regression is tested for collinearity, influence, binned residuals and the uniformity of simulated residuals; normality and constant variance are not expected for a binary outcome. Binomial models with several trials per row are supported, but pseudo-R² measures that need 0/1 outcomes raise an error, as in R.
• Poisson regression adds overdispersion (Pearson chi-squared) and zero-inflation. Negative binomial GLMs get both checks from simulated residuals, as R does. Gamma and inverse Gaussian GLMs are checked through simulated residuals too.
• Overdispersion uses Pearson residuals for Poisson GLMs and simulated residuals for every other GLM and for mixed models, following R's rules; residual_type= overrides the choice. Linear models fitted by least squares have no overdispersion check.
• Mixed models use lme4's definitions of leverage, studentized residuals and Cook's distance. Variance components (vc_formula) are not supported yet. If a variance is estimated at (nearly) zero, refit with .fit(method="powell"); see the note on optimizers in Comparison with R.
• Bayesian models are checked through their posterior draws: predictions, convergence, Bayesian R², PSIS-LOO and WAIC. Residual-based frequentist checks do not apply to them.
• Anything else can be supported by implementing the ModelAdapter protocol (five methods) or registering an adapter; see CONTRIBUTING.md.