Confidence, prediction, tolerance, and new-group-mean (also
called CInew or new-batch-mean) intervals for linear mixed models fitted
with 'lme4' ('lmerMod' objects), returned through one uniform interface.
Several validated interval engines are provided and selected automatically
for the fitted model, with a manual override: an expected-mean-square,
modified large-sample construction for random-intercept models (Francq,
Lin and Hoyer (2019)
Confidence, prediction, tolerance, and new-group-mean intervals for linear mixed models.
reintervals computes confidence (CI), prediction (PI), tolerance (TI),
and new-group-mean (a.k.a. CInew / new-batch-mean) intervals for linear
mixed models fitted with
lme4, through one uniform
interface. It is a general mixed-model interval package, with particular
support for pharmaceutical stability / shelf-life and bioassay work.
Several tools already do part of the job well:
predict.merMod() (lme4) and emmeans / ggeffects give
point predictions and confidence intervals for the mean / marginal
effects.merTools::predictInterval() gives simulation-based prediction
intervals for merMod fits.tolerance package gives tolerance intervals for many
distributions and some regression settings (for simpler designs,
without the mixed-model degrees-of-freedom treatment).What none of them provides is tolerance and new-group-mean (CInew) intervals across both random-intercept and random-slope / fixed-slope (stability) designs — with the observed-information degrees of freedom of Francq, Lin & Hoyer (2019), the unweighted-lot-mean ANOVA closed form of Montes, Burdick & Leblond (2019), and the generalized-pivotal-quantity method of Oliva-Aviles & Hauser (2025) — behind one uniform interface.
reintervals provides four interval types from four engines, selected
automatically for the fitted model (with a manual override), and returns
them in one tidy lmm_interval data frame. Every interval the package
ships is validated against its published source on continuous
integration. The methods are published and cited; the contribution
here is the validated, unified implementation, not new methodology.
# pre-CRAN: install from GitHub
# install.packages("remotes")
remotes::install_github("percevalsondag/reintervals")
(A CRAN release will be noted here once available.)
library(reintervals)
fm <- lme4::lmer(Reaction ~ Days + (1 | Subject), data = lme4::sleepstudy)
# 95% confidence interval for the mean at Days = 5
ci_lmm(fm, newdata = data.frame(Days = 5))
#> <lmm_interval> ci (method = ems-mls, design = random_intercept_fixed_slope)
#> Days type estimate lower upper level P sides singular boundary vg_floored
#> 5 ci 303.7415 284.6347 322.8483 0.95 NA two.sided FALSE NA NA
#> diagnostics: df, df_method (see attr(., "diagnostics"))
# 99%-content / 95%-confidence tolerance interval at Days = 5
ti_lmm(fm, newdata = data.frame(Days = 5), P = 0.99)
#> <lmm_interval> ti (method = ems-mls, design = random_intercept_fixed_slope)
#> Days type estimate lower upper level P sides singular boundary vg_floored
#> 5 ti 303.7415 142.9097 464.5734 0.95 0.99 two.sided FALSE NA NA
Every verb returns the same lmm_interval data frame, so confidence,
prediction, tolerance, and new-group-mean results all share one column
contract (see ?ci_lmm for the schema).
| Verb | Computes |
|---|---|
ci_lmm() |
confidence interval for the mean |
pi_lmm() |
prediction interval for a future observation |
ti_lmm() |
tolerance interval (content P); over = "group_mean" for the batch-mean target |
new_group_mean_lmm() |
the CInew / new-batch-mean interval: the true mean of a new group (a new innermost cell, or a new lot at t0) |
lmm_predict(model, interval = ...) is a predict()-style wrapper over
the four. lmm_interval_control() supplies the GPQ tuning (seed, M)
and the confidence-interval df rule (ci_df: Satterthwaite by default,
falling back to the SAS-style containment df; Kenward-Roger on request).
The engine is chosen from the fitted model’s structure and can differ
between the verbs (full mapping in ?reintervals and
vignette("choosing-an-interval")). Engines: "ems-mls"
(expected-mean-square / modified large-sample), "anova-mls"
(unweighted lot-mean ANOVA: Montes et al. with a fixed slope, Hoffman &
Kringle without one), "reml-mls" (REML closed form), "gpq"
(generalized pivotal quantity, Monte Carlo).
| Model / situation | CI | PI | TI | New-group mean |
|---|---|---|---|---|
random intercept, one-way, intercept only (y ~ 1 + (1 \| g)) |
ems-mls |
ems-mls |
anova-mls (Hoffman & Kringle) |
anova-mls |
| random intercept, one-way with covariates | ems-mls |
ems-mls |
ems-mls |
ems-mls (Gm) |
| random intercept, nested | ems-mls |
ems-mls |
ems-mls |
ems-mls (Gm) |
| random intercept, crossed with interaction | ems-mls |
ems-mls |
ems-mls |
ems-mls (Gv) |
| random intercept, crossed without interaction or 4+ grouping factors | ems-mls |
ems-mls |
error | error |
| random intercept + fixed slope, clean balanced | ems-mls |
ems-mls |
ems-mls |
anova-mls |
| random intercept + fixed slope, single-observation / staggered | ems-mls (warns with single-observation lots) |
ems-mls (warns likewise) |
anova-mls |
anova-mls |
| random slope, future-observation, non-singular | reml-mls |
reml-mls |
reml-mls |
reml-mls |
| random slope, singular fit or TI Satterthwaite df unavailable | reml-mls |
reml-mls |
gpq (two-sided TI; one-sided stays reml-mls) |
reml-mls |
random slope, batch-mean (over = "group_mean", TI only) |
— | — | gpq |
— |
GPQ is used for the two-sided random-slope TI when the fit is singular
or the closed form’s Satterthwaite df is unavailable. One-sided
intervals (sides = "lower" / "upper") are available from every verb;
one-sided tolerance limits use Hoffman’s (2010) limit for anova-mls
and the package’s validated extension of it for ems-mls and random
slopes, where reml-mls is used on every fit.
Scope: general fixed effects are supported for random-intercept
models; random-slope models are supported for a single continuous
covariate (the stability design). The tier-2 engines ti_anova_raw()
and ti_gpq_raw() compute from raw (y, time, group) vectors without a
fitted model.
Every shipped interval is checked against a published or external-source
oracle on CI: the Francq, Lin & Hoyer (2019) worked numbers (Sections
5.2-5.3); the Hoffman & Kringle (2005, 2007) worked examples; the
Montes-Burdick-Leblond (2019) Appendix A tolerance interval
(57.25, 75.41) and their one-sided limits; the random-slope interval
arithmetic against committed reference values to ~1e-9; and the GPQ
content variance against the Oliva-Aviles & Hauser (2025) closed form
(tau^2 = 8.439). Coverage of the intervals was additionally validated
by simulation.
The package covers general-fixed-effect random-intercept models and
single-time-covariate random-slope (stability) models. Random-slope
models with more than one fixed-effect covariate are not currently
supported; see NEWS.md for the change log.
MIT © Perceval Sondag. A JOSS DOI will be added here on acceptance;
until then, please cite the package (citation("reintervals")) and the
underlying methods: