Changes in version 0.1.0 (2026-11-19) - First release on CRAN. - Three vignettes: getting started (trial-level analysis), time courses (diagnostics, calibration and intervals for full time series) and multiple comparisons (families of tests and pairwise comparisons). - diag_timecourse() now flags a misfit when the largest standardised mean residual (mean / SE across clusters) exceeds a Bonferroni-corrected 5% threshold over all time points and groups, instead of a fixed size of the mean residual. The result gains max_z and threshold, and tidy() a z column. Changes in version 0.0.0.9005 - Parallel workers now attach the package of the model's fitting function (e.g. lme4), so that refits in perm_calibrate() and the other resampling functions no longer fail when that package is not attached in the workers. - perm_calibrate() builds the reduced model before the simulations are sent to parallel workers, so that calibration no longer fails when the model call refers to objects of the calling environment (e.g. a control object). - perm_calibrate() reports a rejection rate of NA instead of failing when no simulation gave a valid p-value. Changes in version 0.0.0.9004 - New perm_pairwise() for pairwise comparisons between the levels of a factor (post hoc tests). Each pair is tested by Freedman-Lane permutation in a reparametrisation of the model where the other levels stay in the reduced model, so an effect of another level does not leak into the test of a pair. Pairs are compared in level or through an interaction (e.g. with a spline of time), adjusted by Westfall-Young max-t or min-p (step-down) or by Holm, with optional bootstrap intervals of each difference. Changes in version 0.0.0.9003 - boot_ci() gains strata: with case resampling, clusters are drawn within the levels of a cluster-level column such as group, so every bootstrap sample keeps the observed number of participants per group. The expanded percentile interval then uses G - S degrees of freedom. The wild bootstrap keeps the design fixed and ignores strata. Changes in version 0.0.0.9002 - Formula helpers (findbars(), nobars()) are now taken from reformulas, where recent versions of lme4 have moved them; this removes a deprecation warning with lme4 >= 2.0. Changes in version 0.0.0.9001 - perm_test() and perm_maxt() with Freedman-Lane: the reduced model now also drops random slopes of lower-order terms contained in the tested term when the other variables of that term are constant within the grouping factor. Previously, testing a between-participant interaction such as group:condition in a model with (1 + condition | participant) kept the condition slope, which absorbed the interaction: the test had no power and its p-values were centred near 0.5 with or without an effect. - boot_ci() gains the expanded percentile interval (ci = "expanded", Hesterberg, 2015), now the default. Plain percentile intervals were too narrow with few clusters (about 87% coverage of 95% intervals with 8 participants and 91% with 15 in simulations); the expanded interval matches a t interval with G - 1 degrees of freedom and changes little with many clusters. Changes in version 0.0.0.9000 - diag_acf() now returns a bias-corrected AR(1) coefficient (phi). Centring each short series biased the mean autocorrelation strongly towards zero (e.g. 0.64 for a true 0.92 with 20 samples per series); the coefficient is now chosen so that the expected autocorrelations of centred AR(1) series of the observed lengths match the observed ones over all lags. The uncorrected lag-1 mean is kept as phi_raw, and correct = FALSE restores the previous behaviour. perm_calibrate() therefore simulates residuals with realistic dependence. - A lower-order term can now be tested in the presence of interactions when every variable it interacts with is numeric and centred, as in growth curve analysis with orthogonal polynomials (poly(time, 3)); it is tested with Freedman-Lane and represents the average difference over the window. - Models whose fixed-effects design is rank deficient (coefficients dropped by lm() or lmer()) are now refused with an informative error, instead of silently testing a different set of coefficients. - The documentation of perm_calibrate() explains that systematic patterns left in the residuals (such as an unmodelled average time course) are simulated as noise, and recommends fitting the time course flexibly first. - Refits of mixed models start from the variance parameters of the original fit and skip lme4's derivative-based convergence check, roughly halving the time per refit on large data; options(lmmresample.fast = FALSE) restores the default settings. - Fixes from an independent review: - Freedman-Lane: fitted values of the reduced model now keep only random effects grouped above the resampled units, so effects can no longer be absorbed by participant or trial intercepts when those units are permuted. - Freedman-Lane permutation of residuals between units now requires time and matches samples by time, refusing units with different time points. - perm_maxt() with min-p includes the observed data in the reference set of each test, which removes an anti-conservative bias for small B. - data supplied in a different order from the fitted model, and models with weights or offsets, are refused. - Series identifiers that repeat across participants are refused in diag_acf() and perm_calibrate(). - Default coefficients in boot_ci() and diag_loso() exclude terms of the cluster factor; diag_agreement() handles any contrast coding; groupings that include the cluster are no longer reported as crossed. - Random-effects terms left empty in the reduced model are dropped; calibration is reproducible with any RNG kind; the sharp null warns when there is no random slope to remove. - New boot_ci() computes bootstrap confidence intervals for fixed effects by resampling whole clusters (case bootstrap, with repeated clusters and nested units relabelled) or by cluster wild bootstrap with Rademacher weights. Percentile, BCa (jackknife acceleration over clusters) and basic intervals; confint() returns any type without refitting. Random effects crossed with the cluster trigger a warning. - perm_maxt() gains method = "freedman-lane", so families of tests can include interactions and covariate-adjusted effects, and different terms in different models, and combine = "min-p", which combines statistics on different scales (e.g. t and chi-square). The adjustment procedure is now chosen with adjust. - New diag_loso() refits the model leaving out each participant in turn and flags influential participants; diag_agreement() compares the model coefficient of a two-level condition with the mean participant-level difference. - New diagnostics: diag_acf() estimates the autocorrelation of residuals within series (its lag-1 value can be passed to perm_calibrate()), and diag_timecourse() shows mean residuals over time by condition to detect a misspecified time course or a missing condition by time interaction. - The package now supports linear models only (lm() and lme4::lmer()). Generalized linear (mixed) models are refused with an informative message, because residual permutation does not apply to them. - perm_test() and perm_calibrate() gain method = "freedman-lane": residuals of the reduced model are permuted between whole units or sign-flipped by block, which allows tests of interactions (e.g. condition:time), of terms with several coefficients (joint Wald chi-square) and of effects in the presence of nuisance terms. Random slopes of the tested term are removed from the reduced model. method = "auto" chooses relabelling for simple main effects and Freedman-Lane otherwise. - sim_blocks() gains effect_onset (condition effects that appear later in the trial) and ar1_trials (correlation between consecutive trials). - New perm_maxt() and perm_spec() test the same variable in several models at once (e.g. one model per feature) and control the family-wise error rate with the Westfall-Young max-t procedure (single-step or step-down), using one shared relabelling of units per permutation. - New perm_calibrate() checks the type I error of a permutation test for the user's own design by simulating data under the null from the fitted model (mean or sharp null, optional AR(1) residuals within series), and compares it with the model's Wald test. Includes p-value calibration and rejection-rate plots. - New perm_test() tests a fixed-effect term by permuting it across whole units (trials or participants) and refitting the full model. Supports lm() and lme4::lmer() models, runs in parallel through the future framework, and records failed, non-convergent and singular refits. Results have print(), summary(), tidy() and plot() methods, including a trace plot of the running p-value. - New exch_within(), exch_signflip() and exch_free() declare which units are exchangeable: within blocks (e.g. trials within participants), by swapping the two levels of the tested variable for whole blocks (sign-flipping, for the mean effect when participants differ in their effect), freely, or within strata. - sim_blocks() gains prop_condition (unbalanced conditions, e.g. oddball designs), residual_df (heavy-tailed residuals), and per-participant trial counts through a vector n_trials. - New sim_blocks() simulates block-structured repeated measurements (full time series, trial-level or participant-level data) for within-participant, between-group and mixed designs, with known effects, random intercepts and slopes, and AR(1) residuals within trials.