avlm 0.2.0

New Features

av.lm() gains a cluster argument for CR1 cluster-robust standard errors (the sandwich scaled by G / (G - 1), without Stata/fixest’s extra (n - 1) / (n - k)). When cluster is supplied, the anytime-valid p-values, F-test, and confidence intervals use G - 1 (clusters minus one) as the denominator degrees of freedom, rather than n - p (observations minus coefficients). cluster cannot be combined with vcov_estimator. The method is described in Lindon (2026), “Anytime-Valid Inference in Linear Models with Clustered Outcomes: Exact Multivariate Tests and Cluster-Robust Confidence Sequences”, https://arxiv.org/abs/2610.25613.
cluster accepts a vector of cluster identifiers, a one-sided formula (cluster = ~ cyl), or a column name (cluster = "cyl"). Formulas and column names are looked up in the model’s data and aligned with the observations used in the fit.
With cluster, the anytime-valid sample size is the number of clusters G, so g is on the cluster scale, and the joint F-test uses the cluster-robust e-value with exponent G/2. The F-test p-value is 1 when the cluster-robust covariance of the tested coefficients is singular (e.g. fewer clusters than tested coefficients).
vcov() on an avlm object returns the covariance used for inference (classical, HC or CR1).
optimal_g() gains a nu argument, so it can tune g for cluster-robust inference (optimal_g(G, nu = G - 1, alpha = 0.05)).
The summary prints the value of g.

The heteroskedasticity-robust (vcov_estimator) F-test now uses the Wald statistic of the tested coefficients, inverting their covariance rather than taking a block of the full precision matrix, and stores it as Q / d. Previously the evidence was overstated when testing more than one coefficient.
confint() no longer misplaces intervals when the fit has aliased (NA) coefficients; previously every coefficient after an aliased one got NA or the wrong standard error. Heteroskedasticity- and cluster-robust summaries of such fits no longer fail with a singular-matrix error.
Heteroskedasticity- and cluster-robust standard errors now account for weights (rows and residuals are scaled by sqrt(w); zero weights are dropped). Previously the weights were ignored, giving wrong standard errors for weighted fits.
Intercept-only models now work with vcov_estimator and cluster.
cluster now accepts one observation per cluster, which gives the HC0 sandwich scaled by n / (n - 1).
optimal_g() now searches below g = 1, where the optimum lies for small samples.
Anytime-valid ANOVA p-values use the number of observations as the sample size; models without an intercept were off by one.
av() checks that g is a single positive number and gives informative errors for glm, multivariate (mlm) and multi-stratum (Error()) fits instead of failing later or returning meaningless output.

The classical, heteroskedasticity-robust and cluster-robust code paths share one covariance routine, and unused helpers were removed.

First release of the avlm package, which implements anytime-valid inference for linear models and ANOVA.
Provides a drop-in replacement for summary.lm() and summary.aov() with sequential p-values and confidence intervals.
Includes support for:
- av(): applies the anytime-valid transformation to lm and aov objects.
- confint(): computes anytime-valid confidence intervals for regression coefficients.

Based on the methodology described in
“Anytime-Valid Linear Models and Regression Adjusted Causal Inference in Randomized Experiments”
available at: https://arxiv.org/abs/2210.08589