[avehtari]

Summary

lme4 2.1-0 reached CRAN on 2026-10-23. With it, master fails 5 expectations in test_augdat.R and test_datafit.R, and so do the CI runs (see, e.g. #573 to #576). This PR adapts the two tests and changes no package code.

A separate issue will open discussion on what we should do with the code

This PR was assisted by Claude.

Causes

test_augdat.R:244 and :308 compare the traditional and the augmented-data projection of a Bernoulli reference model. For submodels with group-level terms, both projections call lme4::glmer(), which the test setup runs with nAGQ = 30.

lme4 computes the objective for the group-level parameters with binomial()$aic. That function rounds non-integer responses (traditional projection) and non-integer weights (augmented-data projection); lme4/lme4#682 describes the rounding. Up to lme4 2.0-6, nAGQ > 1 used the deviance residuals, which do not round, and the two projections agreed. In 2.1-0, glmerAGQ() uses the same aic() term as the Laplace approximation (lme4 NEWS: "normalizing constants are now included in the log-likelihoods and deviances of models fitted with nAGQ>1"; reported in lme4/lme4#1011). Now the two projections fit different rounded objectives. In the test data the rounded responses separate the data: in 4 of the 6 traditional projections, the largest fixed effect is between 38 and 60 in absolute value, against at most 1.7 with lme4 2.0-6.

nAGQ = 0 does not rescue the comparison. It makes the fixed effects agree in most fits, but the group-level standard deviation still comes from the rounded objective, and one setup still differs.

test_datafit.R:252 checks that the full-data ranking contains only terms of the formula. The default search terms also contain the linear term of each smooth term, for example s.1 for s(s.1). With lme4 2.1-0 and gamm4 0.3-0, the forward search for rstanarm.gamm.brnll picks (1 | z.1) + s.1 at the second step, where it used to pick (1 | z.1) + s(s.1). Both are valid results, and the check for the CV rankings already accepts such linear terms.

Changes

• test_augdat.R: the two comparisons skip submodels with group-level terms. They now cover the submodels fitted by fit_glm_ridge_callback(), which is what their titles say. I removed the extra tolerances for the group-level case. The test setup still runs these projections, so other tests still use them.
• helpers/testers.R: vsel_tester() accepts the linear terms of smooth terms in the full-data ranking, as it already did for the CV rankings.
• NEWS.md entry.

The rounding also affects projections onto binomial submodels with group-level terms outside the tests, with any lme4 version at the default nAGQ = 1. I will open a separate issue for that.

Tests

On master 73cf0ee with gamm4 0.3-0, test_augdat.R and test_datafit.R give 5 failures with lme4 2.1-0 and none with lme4 2.0-6. With this PR, the full suite passes with NOT_CRAN=true under both lme4 versions: 253 tests, 0 failures (55 655 expectations with lme4 2.0-6, 55 401 with lme4 2.1-0). Both runs show the warnings of master plus a CHOLMOD "not positive definite" warning in test_div_minimizer.R. That warning appears with both lme4 versions and did not appear in our earlier runs of master with gamm4 0.2.7. Tested with R 4.6.1, brms 2.23.0, rstanarm 2.32.2, gamm4 0.3-0, and lme4 2.0-6 and 2.1-0, on CSC Roihu (Linux, AMD EPYC).