Skip to content

v1.2-0

v1.2-0: KRLS 1.2-0: Tier-A polish (formula interface, broom, autoplot, vignette)
@j-hai j-hai tagged this 09 May 23:32 
Tier-A modernization mirroring what landed in ebal/Synth/Inference
work. None of this changes the underlying algorithm or default
behavior; existing krls(X, y) calls continue to work bit-identical.

* Formula interface. krls(y ~ x1 + x2, data = df) now works alongside
  the matrix interface. Drops the intercept column automatically;
  rejects NAs in either side with the same error messages the matrix
  interface uses; expands factor predictors to dummies via
  model.matrix() so factor(region), age * educ, I(age^2), etc. all
  flow through cleanly.

* Tidyverse-friendly extractors. tidy(fit) / glance(fit) /
  augment(fit) registered against the generics package generics so
  library(broom) picks them up.
  - tidy(fit): per-predictor row with the AME (estimate),
    analytical SE, t-stat, p-value, 95% CI, and the pointwise
    quartiles (q25 / median / q75). The pointwise spread is the
    unique KRLS contribution OLS can't tell.
  - glance(fit): one-row summary with nobs, n_predictors, r.squared,
    leave-one-out MSE, lambda, sigma, and effective degrees of
    freedom from the kernel ridge.
  - augment(fit, data): joins .fitted, .resid, and one .dy_d_<x>
    column per predictor (the pointwise derivatives) back to a
    user-supplied data frame.

* ggplot2 autoplot. autoplot(fit) returns a faceted ggplot of the
  pointwise marginal-effect distribution, one panel per predictor,
  with the AME overlaid in blue. Discoverable via library(ggplot2);
  ggplot2 is in Suggests:.

* Quickstart vignette (vignettes/krls-quickstart.Rmd) walks through a
  small simulated example showing the AME-vs-pointwise heterogeneity
  story end-to-end with the new formula + tidy + autoplot pipeline.

Tooling:
* Suggests: + knitr, rmarkdown, ggplot2, generics. VignetteBuilder.
* importFrom stats: + model.frame, model.matrix, model.response,
  na.pass, qt (used by tidy and the formula path).
* 62 tests passing; R CMD check --as-cran clean.

Tier-C (Nystrom approximation for the n^3 scaling story) and the
bigKRLS Rcpp port are independent of this release; Pete Mohanty
replied positively 2026-05-09 and is consulting Robert Shaffer
before picking which of the three options we offered.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Assets 2
Loading