Porting from R’s robustbase::lmrob

A side-by-side cheat sheet for users coming from R’s robustbase package.

Fit

R	Python
`lmrob(y ~ x1 + x2, data = df)`	`lmrob("y ~ x1 + x2", df)`
`lmrob(y ~ ., data = df, control = lmrob.control())`	`lmrob("y ~ x1 + x2", df, control=Control())`
`lmrob.control(psi = "lqq", setting = "KS2014")`	`Control(psi="lqq", setting="KS2014")`
`set.seed(42); lmrob(...)`	`lmrob(..., seed=42)`

Result accessors

R	Python
`fit$coefficients`	`fit.coef_`
`fit$scale`	`fit.scale_`
`fit$residuals`	`fit.residuals_`
`fit$fitted.values`	`fit.fitted_`
`fit$rweights`	`fit.rweights_`
`fit$cov`	`fit.cov_`
`fit$df.residual`	`fit.df_residual_`
`fit$converged`	`fit.converged_`

Inference

R	Python
`summary(fit)`	`print(fit.summary())`
`summary(fit)$coefficients`	`fit.summary().coefficients`
`summary(fit)$r.squared`	`fit.summary().r_squared`
`confint(fit, level = 0.95)`	`fit.confint(level=0.95)`
`anova(full, red)`	`anova(full, red)`
`anova(full, red, test = "Deviance")`	`anova(full, red, test="Deviance")`
`predict(fit, newdata = nd)`	`fit.predict(nd)`

Control fields

The most-used R control fields and their Python equivalents:

R control field	Python `Control` field
`nResample`	`nResample`
`tuning.psi`	`tuning_psi`
`tuning.chi`	`tuning_chi`
`psi`	`psi`
`method`	`method`
`cov`	`cov`
`setting`	`setting`
`max.it`	`max_it`
`k.max`	`k_max`
`refine.tol`	`refine_tol`
`seed`	`seed`
`subsampling`	`subsampling` (`"simple"` / `"nonsingular"`)
(no equivalent)	`rng` (`"PCG64"` / `"MT19937"` / `"R"`)
(no equivalent)	`n_workers` (parallel resampling)

Default behaviour

Control() with no arguments matches R’s lmrob.control() (default setting=NULL): psi="bisquare", method="MM", cov=".vcov.avar1", tuning_psi=4.685061, tuning_chi=1.547645.

Control(setting="KS2014") matches R’s lmrob.control(setting="KS2014"): psi="lqq", method="SMDM", cov=".vcov.w", tuning_psi=(1.4734061, 0.9822707, 1.5), tuning_chi=(0.4015457, 0.2676971, 1.5).

Control(setting="KS2011") is the same as KS2014 but with the older D-scale tuning.

Things that don’t match exactly

The default fast-S resampling RNG uses NumPy’s PCG64; R uses Mersenne Twister. With the same seed, the two diverge at the first random draw, so basin-of-attraction sensitivity can give small coefficient differences on tiny-n datasets. For tighter agreement, opt in to Control(rng="R") (v0.5.16+): the resample loop runs against R’s exact unif_rand stream and stackloss fits agree with lmrob to rtol~1.7e-5. See the numerical notes for the full list.