reconf 0.2.1

• Builds with Eigen 5 (#4, @eddelbuettel).
• Supports the lme4 2.0 covariance structures diag(), cs(hom = TRUE), and cs() and ar1() with two effects. A variance or covariance shared by a term’s effects is one parameter, named var_*|g or cov_*|g. With three or more effects, cs() with separate variances and ar1() are an error. Version 0.2.0 misread all of these terms and could return wrong intervals without an error.
• vc_model() stops if the covariance structure does not reproduce the fitted covariance matrix, or if lme4 reports a residual standard deviation that is not finite, as lme4 2.0-6 can.
• vc_test() and vc_ci() warn when an internal maximization stops away from a root of the score.
• No spurious “NaNs produced” warnings in nearly singular fits.

reconf 0.2.0

• First CRAN release.

New interface

• vc_ci() computes confidence intervals, vc_test() score tests, and vc_loglik() the log-likelihood, score, and information as functions of the covariance parameters. Each takes an lme4 fit or a model from the generic vc_model(), which does the precomputation once.
• vc_ci() replaces ci_lmer() and ci_all_lmer(), and vc_test() replaces score_test_lmer(). parm selects parameters by name or index.

Changes that may affect existing code

• vc_ci() returns an object of class vc_ci with columns estimate, lower, and upper. Code that indexed the old two-column output by position should select columns by name.
• Parameters are named var_x|g for a variance, cov_y.x|g for a covariance, and var_Residual for the error variance.
• test_idx and known_idx are renamed parm and known. Both index the covariance parameters, also for maximum likelihood fits. 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 htest object instead of a named vector, and takes null_value, the values of the tested parameters (default 0), in place of theta_null. The arguments efficient and profile are removed. The statistic is the profile score statistic, standardized by the efficient information.
• Log-likelihood values include all normalizing constants and equal logLik() of the lme4 fit, up to half the sum of the log weights in weighted fits.

New features

• vc_ci(statistic = "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 extend below zero, with a warning. Requires onestep = FALSE. The default is statistic = "score".
• method selects the computational path. The default "auto" uses a sparse q-by-q path, a dense n-by-n path when q >= n and Z is dense, or a spectral path with cost O(n) per evaluation when r = 2. On a dense genomic model (n = 742, q = 1484) one evaluation takes 0.07 s instead of 6.8 s.
• expected = FALSE gives the observed information also for the restricted likelihood. Previously the expected information was used, with a warning.
• Prior weights and offsets in lme4 fits are supported.
• print() and tidy() methods for vc_ci objects.
• The interval search evaluates the statistic at the estimate. If the estimate is outside the confidence set, the search warns and restarts from the maximizer over the extended parameter set.
• With nonneg = TRUE the interval for a variance is intersected with [0, Inf). An empty intersection gives NA bounds and a warning.
• If no bound is found within the search distance, the warning reports the statistic where the search ended. The bound is infinite if the statistic levels off below the critical value, as for a variance whose grouping factor has few levels.
• The search continues from a feasible starting point when the nuisance parameters from the previous step are infeasible at the next value.
• vc_ci() and vc_test() reject misnamed optimizer arguments and invalid flags. The default parm excludes the parameters in known.

Bug fixes

• The log-likelihood is -Inf wherever Z Psi Z' + psi_r I is not positive definite. It could be finite there before, which corrupted interval searches.

Performance

• vc_ci() on the fev1 example takes about 0.6 s instead of 43 s, with identical intervals.

reconf 0.1

• Initial release.
• ci_lmer(): score-based confidence interval for a covariance parameter of an lme4 fit.
• ci_all_lmer(): intervals for all or some covariance parameters.
• score_test_lmer(): score test for covariance parameters.