Python API reference¶

Import public functions directly from rparity. Formula arguments are strings, and data is a pandas data frame or a Polars frame converted to pandas internally. This page lists the main call arguments and the result methods you use in an analysis. Detailed option comparisons are in the migration guides.

Mixed models¶

lmer(formula, data, reml=True, weights=None, offset=None,
     contrasts="treatment", control=None)

glmer(formula, data, family="binomial", *, link=None, weights=None,
      offset=None, nAGQ=1, contrasts="treatment", control=None)

lmer() fits Gaussian mixed models; reml=False selects ML. glmer() supports binomial and Poisson families with Laplace integration (nAGQ=1). Binomial links are logit, probit and cloglog. Weights and offsets can be vectors or data-column names. contrasts accepts "treatment" or "sum".

Result method	Returns or purpose
`fixef()`	Named fixed-effect coefficient series
`vcov()`	Fixed-effect covariance data frame
`ranef(cond_var=True)`	Random-effect data frames by group
`VarCorr()`	Random-effect covariance data frames by group
`summary()`	Printable model summary
`predict(newdata=None, re_form=None, allow_new_levels=False, offset=None)`	Conditional or population prediction array
`fitted()`, `residuals(type="response")`	Fitted values and residuals
`logLik()`, `AIC()`, `BIC()`, `deviance()`	Fit criteria
`is_singular(tol=1e-4)`	Random-effect singularity check

For lmer(), summary(ddf="satterthwaite") chooses the coefficient degrees-of-freedom method. refit(reml=False) refits the model with ML. For both fit types, re_form="NA" predicts with random effects set to zero. See mixed models for examples.

Generalized additive models¶

gam(formula, data, *, family="gaussian", link=None, method="GCV.Cp",
    weights=None, offset=None, sp=None, scale=0, gamma=1, control=None)

gam_check(model, *, k_rep=200, k_sample=5000, seed=0)

Supported families are Gaussian, binomial, Poisson and Gamma. Smoothing methods are "GCV.Cp", "ML" and "REML". Formula smooths use s(), te() or ti(). Negative entries in sp request estimated penalties; nonnegative entries fix the corresponding penalty.

gam() returns a GamResult. gam_check() returns a dictionary of convergence and residual diagnostics, including the "k.check" table.

Result method or attribute	Returns or purpose
`coef()` or `coefficients`	Named coefficient series
`vcov(freq=False)`	Bayesian covariance by default; frequentist covariance with `freq=True`
`smooth_edf`	Effective degrees of freedom by smooth
`summary_tables()`	Dictionary including `"parametric"` and `"smooth"` tables
`summary()`	Printable summary
`predict(newdata=None, type="link", se_fit=False, terms=None, exclude=None, offset=None)`	Predictions, term effects or linear-predictor matrix
`partial_effects(term=None, n=100, newdata=None)`	Numerical effect tables
`check(k_rep=200, k_sample=5000, seed=0)`	Equivalent model diagnostics
`fitted()`, `residuals(type="deviance")`	Fitted values and residuals
`logLik()`, `AIC()`, `deviance()`	Fit criteria
`converged`	Fit convergence status

Prediction types are "link", "response", "terms", "iterms" and "lpmatrix". With se_fit=True, ordinary or term predictions return {"fit": ..., "se.fit": ...}. Standard errors condition on the smoothing parameters. See the GAM guide.

Distributional models¶

glmmTMB(formula, data, family="gaussian", *, ziformula="~0",
        dispformula="~1", link=None, weights=None, offset=None,
        contrasts="treatment", control=None)

Supported families are "gaussian", "poisson", "binomial", "nbinom1", "nbinom2" and "beta", with their default links. Conditional random effects use the main formula. Zero-inflation and dispersion formulas contain fixed predictors. Fitting uses ML.

glmmTMB() returns a GlmmTMBResult:

Result method or attribute	Returns or purpose
`fixef()`	Coefficient series keyed by `"cond"`, `"zi"`, `"disp"`
`vcov(component="cond", full=False)`	Component covariance or full parameter covariance
`ranef(cond_var=True)`, `VarCorr()`	Component dictionaries of group results
`predict(newdata=None, type="response", re_form=None, allow_new_levels=False, offset=None)`	Response or component prediction array
`model_matrix(component="cond", newdata=None)`	Component design matrix
`sigma()`	Intercept-only dispersion summary
`summary()`	Printable model and component summaries
`fitted()`, `residuals(type="response")`	Fitted response means and residuals
`logLik()`, `AIC()`, `BIC()`	Fit criteria
`is_singular(tol=1e-4)`, `converged`	Fit diagnostics

Prediction types include "response", "conditional" (or "cond"), "link", "zprob", "zlink" and "disp". Response means combine the conditional mean and zero inflation. See the zero-inflated-model guide.

Generalized least squares¶

gls(formula, data, *, correlation=None, weights=None, method="REML",
    REML=None, control=None, na_action="omit")

Choose method="ML" for maximum likelihood. weights is a variance structure here, rather than a vector of observation weights.

Structure	Main arguments
`corAR1`	`value=0.0, form="~1", fixed=False`
`corARMA`	`value=None, form="~1", p=0, q=0, fixed=False`
`corCompSymm`	`value=0.0, form="~1", fixed=False`
`corSymm`	`value=None, form="~1", fixed=False`
`varIdent`	`value=None, form="~1", fixed=None`
`varPower`	`value=0.0, form="~fitted(.)", fixed=None`
`varExp`	`value=0.0, form="~fitted(.)", fixed=None`

For example, corAR1(form="~time | subject") specifies within-subject AR1 correlation. GLS results provide coef(), fixef(), vcov(), summary(), predict(newdata=None, se_fit=False), fitted() and residuals(type="response"). They also provide logLik(), AIC(), BIC(), anova(type="sequential") and intervals(level=0.95, which="all"). See the nlme guide.

Marginal means and contrasts¶

emmeans(model, specs, *, by=None, at=None, weights="equal", type="link",
        pairwise=False, adjust=None, lmer_df="kenward-roger", **kwargs)

emtrends(model, specs, *, var, delta_var=None, at=None, by=None,
         weights="equal", type="link", lmer_df="kenward-roger", **kwargs)

contrast(object, method="pairwise", *, by=None, adjust=None, ref=1,
         max_degree=None)
pairs(object, *, reverse=False, **kwargs)
joint_tests(model, *, by=None, at=None, weights="equal",
            lmer_df="kenward-roger", **kwargs)

specs names the predictors to retain in the result. at sets reference values, for example {"time": [0, 5]}. weights accepts "equal", "proportional", "cells", "outer", "flat" or explicit weights. lmer_df accepts "kenward-roger", "satterthwaite" or "asymptotic". For distributional models, pass component="cond", "zi" or "disp" through kwargs.

emmeans() and emtrends() return an EmmGrid, or a dictionary with "emmeans" and "contrasts" when pairwise output is requested. contrast() and pairs() return an EmmGrid; joint_tests() returns a data frame.

grid.summary(type=None, level=0.95, adjust=None, infer=None, null=0.0) returns estimates, intervals and optional tests. infer=(True, True) requests confidence intervals and tests.
grid.confint(level=0.95) returns confidence intervals.
grid.grid exposes reference-grid rows.

For explicit reference-grid control, use ref_grid(model, at=None, cov_reduce=..., cov_keep=2, lmer_df="kenward-roger", component="cond", ...). regrid(grid, transform="response") transforms grid cells before averaging. See marginal means for the distinction from ordinary response-scale summaries.

Hypothesis tests and model comparison¶

Anova(model, type=2, test_statistic=None, error_estimate="pearson",
      component="cond")

anova(*models, type=None, ddf="satterthwaite", refit=True, adjust_sigma=True)

Both return pandas data frames. Anova() performs Type II or III term tests. Its statistic depends on the fitted model; distributional models use component-specific Wald chi-squared tests. For supported linear mixed models, anova(model, type=...) accepts 1, 2 or 3 for sequential or marginal tests; ddf selects the degree-of-freedom method. anova(model1, model2) compares fitted likelihoods, refitting REML models with ML unless refit=False.

Type III main-effect hypotheses in interaction models require suitable contrast coding; use contrasts="sum" when that hypothesis is intended. See the car migration guide and numerical limitations.