Skip to contents

Changelog

Source: NEWS.md

oalasso 1.0.0

Initial release, implementing the outcome-adaptive lasso of Shortreed & Ertefaie (2017), Biometrics 73(4):1111–1122, doi:10.1111/biom.12679, and the generalized outcome-adaptive lasso (GOAL) of Baldé, Yang & Lefebvre (2023), Biometrics 79(1):514–520, doi:10.1111/biom.13683.

  • oal(): outcome-adaptive lasso propensity scores. Adaptive penalty weights |b_j|^-gamma from an unpenalized outcome model Y ~ A + X on the full sample (gaussian() validated; binomial() allowed with a one-time warning, GLM theory per Baldé 2025, doi:10.1016/j.spl.2025.110379); tuning over Shortreed & Ertefaie’s exponent grid lambda_n = n^delta with the paired penalty exponent gamma = 2*(gamma.factor - delta + 1); selection by the weighted absolute mean difference (wAMD) with |b_j| weights, ATE (default) or ATT weighting; ties take the first minimum in documented grid order.
  • Exact reproduction of the published objective on glmnet: every grid point is evaluated at s = mean(pen) * lambda_n / n with coef(..., exact = TRUE), undoing glmnet’s internal sum-to-nvars penalty-factor rescaling and its per-observation loss scale (the rejoinder recipe of Jones, Ertefaie & Shortreed 2023, doi:10.1111/biom.13681). The KKT conditions of the target objective are the acceptance criterion in the test suite, with a drift guard against the installed glmnet version.
  • method = "goal": Baldé’s elastic-net generalization via the Zou–Hastie data augmentation with the (1 + lambda2) rescale of ALL coefficients including the intercept (Baldé 2025 supplement: expit(cbind(1, X) %*% (1 + lambda2) * coef)); on augmented grid points (lambda2 > 0) the raw penalty constant is (n + q)^delta with q = p augmentation rows, matching the author’s adaptive.lasso(lambda = n.q^(il)) with n.q = n + q. lambda2 is selected jointly with (delta, gamma) by the wAMD — the flat first-minimum, equivalent to Baldé’s nested per-lambda2-then-over-lambda2 rule — and lambda2 = 0 exactly nests plain OAL (no augmentation on that code path). The default lambda2 grid c(0, 10^c(-2, -1.5, -1, -0.75, -0.5, -0.25, 0, 0.25, 0.5, 1)) is the author’s published grid (Baldé 2025 supplement, “taken from Zou and Hastie (2005)”), verified against the official GOAL code on 2026-07-02.
  • Baldé’s reference code runs UNCLIPPED IPW weights; the default clip = c(0.01, 0.99) is a Shortreed–Ertefaie-lineage safety choice, and clip = c(1e-12, 1 - 1e-12) effectively disables clipping to reproduce the reference behavior.
  • Fixed standardization protocol: covariates are standardized once (training statistics, all columns including dummies); the outcome model and glmnet fit on the same matrix with standardize = FALSE; coefficients returned on both scales.
  • Options with provenance labels printed on every fit: refit = TRUE (post-selection unpenalized refit, Schnitzer et al. 2025, doi:10.1002/sim.70316, adapted from a longitudinal setting; convergence recorded per grid point), scalar gamma (Schnitzer-style fixed exponent crossed with the delta grid), and the outcome.coef hook (user-supplied weights on the standardized scale — screening use only; zeros become hard exclusions via glmnet’s Inf-to-exclude).
  • Ecosystem integration: fit$ps (numeric, named by rownames(data), clipped to clip = c(0.01, 0.99)) drops into MatchIt::matchit(distance = ), WeightIt::weightit(ps = ), and psAve::psave(ps.append = ); oal_match() / oal_weight() wrappers reuse the stored formula and data; cobalt::bal.tab() works on oal objects.
  • Full S3 suite: print() (provenance line first; retained vs excluded covariates with the instrument-exclusion rationale of Brookhart et al. 2006 and Myers et al. 2011; the literal next call; near-positivity diagnostic when > 5% of scores sit at a clip bound), summary(), plot() ("wamd", "coef", "balance"), coef() (both scales), predict() (works without keep.fits), weights().
  • Exported criterion: oal_wamd() scores any candidate propensity score on the exact S&E wAMD formula (single source of truth; never delegated to cobalt).
  • Policies: any NA in used variables is an error naming the variables (never a silent drop); zero-variance and aliased/rank-deficient covariates error; fully deterministic pipeline (no seed argument); English-only messages.
  • Documented divergences from the author’s legacy fAL/fOAL code (which is superseded, not reproduced): the fAL gamma transcription bug, the varlist wAMD bug, global-n scoping, per-subset scale(), and the archived lqa dependency. See vignette("method-details", package = "oalasso").
  • Three vignettes: Getting Started (matching and weighting workflows, reading the output), Using oalasso with psAve (the ps.append composition and its caveats), and Method details and provenance (all formulas, the exact penalty-scale correction, provenance table, relation to other software, honest nonlinear-extension notes).