ti_lmm() computes the Hoffman & Kringle (2005)
two-sided tolerance interval for an intercept-only model with one
random-intercept factor (y ~ 1 + (1 | g)), balanced or
unbalanced. It is the slope-free branch of the anova-mls
engine and is now the default for that design (previously
ems-mls); method = "ems-mls" still gives the
Francq et al. (2019) interval. The two coincide on balanced data with a
non-singular fit and differ on singular or unbalanced data.
ti_anova_raw() gives the same interval from raw data with
time = NULL.hk_assay: the Hoffman & Kringle (2007)
Table III assay data (6 runs x 3 replicates). Hoffman & Kringle
(2005) Table 6 misprints run 5 / replicate 3 as 1.020 (correct: 0.998);
Francq et al. (2019) Sections 5.2-5.3 used the misprinted value.lmm_interval_control() gains ci_df, the
confidence-interval df rule for random-intercept models:
"satterthwaite" (default) or "kenward-roger"
(requires pbkrtest and errors if it is not installed; random-slope
models keep their own df and reject "kenward-roger"). The
df source is reported in
attr(., "diagnostics")$df_method.ti_lmm() reports its routing reason per
row in attr(., "diagnostics")$route_reason
("none", "singular",
"df_unavailable", "group_mean_target", or
"method").sides = "lower" /
"upper") for ci_lmm(), pi_lmm(),
and new_group_mean_lmm() on every engine: the two-sided
interval’s center, variance, and df with
t_{1 - alpha}.anova-mls (Hoffman-Kringle
one-way and Montes fixed slope) uses Hoffman’s (2010) one-sided limit;
ems-mls and random-slope models use the package’s extension
of Hoffman’s limit to the Francq and REML closed-form bounds, validated
by simulation. For random slopes the one-sided default is
reml-mls on every fit, singular ones included;
method = "gpq" remains available but under-covered near the
variance-component boundary in validation. lmm_predict()
passes sides through.anova-mls tolerance limit used the
two-sided formula with z_P, which is anti-conservative for P < 0.99
(Montes et al. 2019). It now uses Hoffman’s (2010) one-sided limit.new_group_mean_lmm()) was anti-conservative, with coverage
as low as 0.80: it used the observable-variance df and collapsed when a
variance component was truncated to zero. It now targets the true mean
of a new innermost cell (a new group, B-within-A cell, or (a, b) cell; a
new lot at t0 for the fixed-slope design) with
constructions validated by simulation: the unweighted lot-mean (Montes)
construction for fixed-slope and intercept-only one-way designs
(anova-mls), and an EMS construction with the top factor’s
df (nested) or a Satterthwaite df with V_F attached to the
main effects (crossed) (ems-mls). W,
df, and df_rule are reported in
attr(., "diagnostics"). Designs without a closed-form
decomposition now signal an error, and method = accepts
only the design’s own engine for this interval. On single-observation
fixed-slope designs the new-group-mean interval no longer warns (the
warning remains for the confidence and prediction intervals).y ~ z + (1 | g)) was
misclassified as the fixed-slope stability design:
new_group_mean_lmm() and ti_lmm() failed (“no
within-lot time variation”) and ci_lmm() /
pi_lmm() warned. A covariate is now treated as the
fixed-slope time only if it varies within at least one group; otherwise
all four verbs follow the ems-mls path.ci_lmm() and
pi_lmm() on fixed-slope designs now fires only when at
least one lot has a single observation. Staggered designs whose lots all
have two or more observations showed no shortfall in validation and no
longer warn. The tolerance interval still uses anova-mls on
every fixed-slope design that is not clean balanced.t0, when the
Satterthwaite df of the closed form’s between-group or residual variance
is unavailable (its containment fallback fired). Such fits are
non-singular but sit near the variance-component boundary.y ~ grp + (1 | lot) with a 2-level
grp) was misdetected as the fixed-slope stability design.
When every lot had at least two observations, all four verbs failed
(“arguments must have same length”). With single-observation lots,
ci_lmm(), pi_lmm(), and
new_group_mean_lmm() gave a spurious fixed-slope
single-observation warning, and ti_lmm() was misrouted to
anova-mls and failed. The fixed-slope design now requires
that term to be a numeric model-frame column (mapped through the model’s
terms, not the coefficient name). Any other model follows the
ems-mls path. Numeric-covariate routing is unchanged.ci_lmm() now reports singular from
lme4::isSingular(), as pi_lmm(),
new_group_mean_lmm(), and ti_lmm() already
did. Previously the confidence interval always reported
singular = FALSE.ti_lmm() signal an error (they do
not return NA bounds); the random-slope
Cov(phi) is NULL only when the Hessian or its
inverse fails; the boundary flag is an advisory flag on the
point estimate (time-scale dependent) and does not drive the df
fallback; n_E uses the target’s content variance; the
true_value tolerance-interval documentation and errors (in
both the closed-form constructor and the EMS dispatch path) no longer
claim to apply only to the fixed-slope design; the
Cov(theta) and random-slope Cov(phi) Hessians
(and the prediction-interval df built on them) are described as the REML
log-likelihood on the variance-component scale, with beta at its GLS
value, not a profiled REML; and printing a vfun_components
with a NULL Cov(phi) now says “Hessian or its
inverse unavailable” rather than “singular”.method = overrides are limited to each design’s allowed
engines).Initial release: confidence, prediction, tolerance, and new-group-mean intervals for linear mixed models, through one uniform interface.
ci_lmm() — confidence interval for the mean.pi_lmm() — prediction interval for a future
observation.ti_lmm() — tolerance interval (content P);
over = "group_mean" gives the batch-mean tolerance
interval.new_group_mean_lmm() — the CInew / new-batch-mean
interval (prediction for one new group’s mean).lmm_predict() — a predict()-style wrapper
over the four, by interval =.lmm_interval_control() — GPQ tuning (RNG
seed, Monte-Carlo size M).ti_anova_raw(), ti_gpq_raw() — tier-2
direct engines from raw (y, time, group) vectors (no fitted
model required).method overrides)ems-mls — expected-mean-square / modified large-sample,
for random-intercept models (Francq, Lin & Hoyer 2019).reml-mls — REML closed form, for random-slope
models.anova-mls — unweighted lot-mean ANOVA closed form, for
unbalanced single-observation fixed-slope stability designs (Montes,
Burdick & Leblond 2019).gpq — generalized-pivotal-quantity Monte Carlo
(Oliva-Aviles & Hauser 2025); used for the batch-mean target and for
singular random-slope fits.All verbs return a single tidy lmm_interval data frame
with a uniform set of columns (type, estimate,
lower, upper, level,
P, sides, method,
route-independent design, and singular /
boundary / vg_floored flags), the eval-point
column(s) named by the model predictor(s), and engine-specific
diagnostics in attr(., "diagnostics").
General fixed effects for random-intercept models; a single continuous covariate for random-slope models (the stability design). On single-observation / unbalanced fixed-slope data the EMS confidence/prediction/new-group-mean intervals warn that they may be anti-conservative; the tolerance interval is routed to the validated ANOVA closed form there.
Every shipped interval is checked against a published or external-source oracle on continuous integration: the Francq, Lin & Hoyer (2019) worked numbers; the Montes-Burdick-Leblond (2019) Appendix A tolerance interval (57.25, 75.41); the random-slope intervals against committed reference values to ~1e-9; and the GPQ content variance against the Oliva-Aviles & Hauser (2025) closed form (tau^2 = 8.439).