reconf 0.2.0

New interface

• The functions are vc_ci() for confidence intervals, vc_test() for score tests, and vc_loglik() for the log-likelihood as a function of the covariance parameters. Each accepts an lme4 fit or a model object from vc_model(), which does the precomputation once and can be reused across calls. vc_model() is generic, so other model sources can be added as methods.
• vc_ci() replaces ci_lmer() and ci_all_lmer(): vc_ci(fit) computes all intervals, and parm selects parameters by name or index. Only the requested intervals are computed.
• vc_test() replaces score_test_lmer().
• vc_loglik() returns the value, score, and information, named by the parameters, at any covariance parameters in the parameter set.

New features

• New statistic argument for vc_ci(). The default "score" is the previous behavior; "rlrt" inverts the profile likelihood ratio statistic, restricted for restricted likelihood fits, with nuisance parameters and the reference maximum taken over the extended parameter set. If the maximizer has a negative variance, the interval for that variance can lie below zero, and a warning is issued. Requires onestep = FALSE.
• The outward CI search now evaluates the statistic at the search origin instead of assuming it vanishes there. If the estimate is outside the confidence set, the search warns and restarts from the extended-set maximizer. The nonneg clamp is applied as a set intersection; an empty intersection gives NA bounds and a warning.
• When no bound is found within the search distance, the warning reports the statistic at the end of the search. The bound is infinite if the statistic levels off short of the critical value, as for a variance whose grouping factor has few levels.
• vc_ci() and vc_test() reject misnamed optimizer arguments and invalid flags instead of passing them on, and the default parm excludes the parameters in known.
• Prior weights and offsets in lme4 fits are supported, handled exactly by transforming Y, X, and Z with the square-root weights.
• The outward search chooses step lengths by secant extrapolation, each at most twice the previous one. If the nuisance parameters from the previous step are infeasible at a proposed value, a feasible starting point is constructed instead of ending the search.
• New method argument selecting the computational path. The default "auto" uses the sparse q-by-q path, a dense n-by-n path when q >= n and Z is dense, or an O(n)-per-evaluation spectral path when r = 2. On a dense genomic model (n = 742, q = 1484) one evaluation drops from 6.8 s to 0.07 s.
• expected = FALSE now works with the restricted likelihood: all three computational paths return the observed information, where previously a warning was issued and the expected information used. The added cost is of the same order as the score.
• print and tidy() (broom-style) methods for the confidence intervals.

Changes that may affect existing code

• Intervals are returned as an object of class vc_ci with an added estimate column (columns estimate, lower, upper). Code that indexed the previous two-column output by position should select columns by name.
• Interval rows are named var_x|g for a variance, cov_y.x|g for a covariance, and var_Residual for the error variance, following lme4’s names for standard deviations and correlations.
• test_idx and known_idx are renamed parm and known; both index the covariance parameters, also for maximum likelihood fits, and parameters in known are held at their estimates. The REML argument is removed; the likelihood is the one the fit maximized.
• vc_test() returns an object of class htest with components statistic, parameter, p.value, and null.value, instead of a named vector with elements stat, p_val, and df. It takes null_value, the values of the tested parameters (default 0), in place of the full parameter vector theta_null. The arguments efficient and profile are removed; the statistic is the profile score statistic, standardized by the efficient information.
• For maximum likelihood fits the fixed effects are profiled out of the likelihood by generalized least squares, which leaves the statistics unchanged.
• The log-likelihood value now includes all normalizing constants and equals logLik() from an equivalent lme4 fit (weighted fits: up to half the sum of log weights).

Bug fixes

• The log-likelihood could be finite at parameters where Sigma = Z Psi Z' + psi_r I is not positive definite, corrupting CI searches. Feasibility is now decided by an exact Cholesky test.

Performance

• Large speedups; vc_ci() on the FEV1 example runs in about 0.6 s instead of 43 s, with identical intervals. Sparsity-aware solves via Matrix, no dense q-by-q matrices in the restricted information, cached symbolic factorizations, and reuse of precomputed quantities.

Infrastructure

• GitHub Actions R CMD check (Linux, macOS, Windows).
• Internal loglik() and loglik_res() take the matrix A and its log-determinant as arguments. Dead code removed and pure-R reference implementations moved to the test suite.

reconf 0.1

• Initial release.
• ci_lmer(): score-based confidence interval for a single covariance parameter in a linear mixed model fitted with lme4.
• ci_all_lmer(): score-based confidence intervals for all (or a subset of) covariance parameters.
• score_test_lmer(): score test for covariance parameters against a user-supplied null hypothesis.