Mixed Effect Operators with mixed_regress() • multivarious

1. Operator-valued ANOVA

mixed_regress() treats each named fixed-effect term as a multivariate object rather than a scalar test. The package-level idea is:

fit the row-side geometry once,
extract a named effect,
analyze that effect with the existing multivarious verbs.

For a term H, the effect matrix is

$M_H = P_H^{(\Omega)} Y B$

where:

Y is the stacked observation-by-feature response matrix,
P_H^{(\Omega)} is the covariance-weighted projector for the term,
B is an optional shared feature basis.

The returned effect_operator behaves like a bi_projector, so you can call components(), scores(), truncate(), reconstruct(), perm_test(), and bootstrap() directly.

2. Simulate a repeated-measures design

We generate a simple low/mid/high repeated-measures dataset with a between-subject group factor, a random intercept, and a random slope on the ordered within-subject effect.

set.seed(1)

n_subject <- 18
levels_within <- c("low", "mid", "high")

design <- expand.grid(
  subject = factor(seq_len(n_subject)),
  level = factor(levels_within, levels = levels_within),
  KEEP.OUT.ATTRS = FALSE
)

subject_group <- rep(c("A", "B"), length.out = n_subject)
design$group <- factor(subject_group[as.integer(design$subject)])

level_num <- c(low = -1, mid = 0, high = 1)[as.character(design$level)]
group_num <- ifelse(design$group == "B", 1, 0)
subj_idx <- as.integer(design$subject)

b0 <- rnorm(n_subject, sd = 0.7)
b1 <- rnorm(n_subject, sd = 0.3)

n <- nrow(design)
p <- 8

Y <- cbind(
  b0[subj_idx] + 1.2 * level_num + rnorm(n, sd = 0.2),
  0.8 * group_num + rnorm(n, sd = 0.2),
  1.4 * level_num * group_num + rnorm(n, sd = 0.2),
  -0.9 * level_num + rnorm(n, sd = 0.2),
  b1[subj_idx] * level_num + rnorm(n, sd = 0.2),
  rnorm(n, sd = 0.2),
  rnorm(n, sd = 0.2),
  rnorm(n, sd = 0.2)
)

dim(Y)
#> [1] 54  8

3. Fit the model

The current implementation supports one grouping variable and a shared row covariance across features. You can supply either a single random-effects bar such as ~ 1 + level | subject or multiple bars that collapse to the same grouping variable, such as ~ (1 | subject) + (0 + level | subject).

fit <- mixed_regress(
  Y,
  design = design,
  fixed = ~ group * level,
  random = ~ 1 + level | subject,
  basis = shared_pca(ncomp = 4),
  preproc = center()
)

print(fit)
#> mixed_fit object
#> 
#> Observations: 54
#> Features: 8
#> Terms: group, level, group:level
#> Basis rank: 4
#> Row metric: grouped_lmm
#> Grouping variable: subject
summary(fit)
#>                    term df_term   scope exchangeability
#> group             group       1 between between_subject
#> level             level       2  within  within_subject
#> group:level group:level       2   mixed   whole_subject

The fit stores:

the design matrix,
the grouped row metric,
the shared feature basis,
metadata for named fixed-effect terms.

4. Extract a named effect

Now extract the interaction effect as a first-class multivariate object.

E <- effect(fit, "group:level")

print(E)
#> effect_operator
#> 
#> Term: group:level
#> Components: 2
#> Term df: 2
#> Scope: mixed
#> Exchangeability: whole_subject
#> Basis rank: 4
ncomp(E)
#> [1] 2
components(E)[1:8, ]
#>             [,1]        [,2]
#> [1,]  0.05109814 -0.06294787
#> [2,]  0.05330174  0.30134100
#> [3,] -0.97145296 -0.05483448
#> [4,]  0.12712397  0.52753036
#> [5,]  0.16843403 -0.73658797
#> [6,]  0.01281331  0.04287406
#> [7,]  0.06526767  0.03733842
#> [8,]  0.04327207 -0.27953870

Because E is an effect_operator and inherits from bi_projector, the familiar decomposition grammar carries over:

components(E) gives feature directions,
scores(E) gives observation-side effect scores,
truncate(E, k) keeps only the first k axes.

5. Reconstruct the effect

You can reconstruct the fitted contribution of the effect on different scales.

E_proc <- reconstruct(E, scale = "processed")
E_orig <- reconstruct(E, scale = "original")

dim(E_proc)
#> [1] 54  8
dim(E_orig)
#> [1] 54  8
round(E_orig[1:6, 1:4], 3)
#>        [,1]   [,2]   [,3]   [,4]
#> [1,] -0.035 -0.046  0.698 -0.105
#> [2,]  0.035  0.046 -0.698  0.105
#> [3,] -0.035 -0.046  0.698 -0.105
#> [4,]  0.035  0.046 -0.698  0.105
#> [5,] -0.035 -0.046  0.698 -0.105
#> [6,]  0.035  0.046 -0.698  0.105

Typical choices:

scale = "whitened" for the covariance-adjusted effect geometry,
scale = "processed" for the response scale after preprocessing,
scale = "original" for the final effect contribution on the original variables.

6. Omnibus and rank inference

perm_test() works directly on the extracted effect object.

set.seed(2)
pt <- perm_test(E, nperm = 99, alpha = 0.10)

print(pt)
#> 
#> Effect operator permutation test
#> 
#> Term: group:level
#> Method: Reduced-model residual permutation test for effect_operator with sequential deflation
#> Exchangeability: whole-subject trajectory permutation within equal block-size strata
#> Omnibus statistic (trace_ratio): 1.447
#> Omnibus p-value: 0.01
#> Selected rank: 1
#> 
#>   comp statistic effective_rank   lead_sv2       rel   observed pval
#> 1    1  lead_sv2              2 344.460800 0.9932418 344.460800 0.01
#> 2    2  lead_sv2              1   2.343768 1.0000000   2.343768 0.32
pt$component_results
#> # A tibble: 2 × 7
#>    comp statistic effective_rank lead_sv2   rel observed  pval
#>   <int> <chr>              <int>    <dbl> <dbl>    <dbl> <dbl>
#> 1     1 lead_sv2               2   344.   0.993   344.    0.01
#> 2     2 lead_sv2               1     2.34 1         2.34  0.32

The permutation result provides:

an omnibus trace test,
a sequential rank test based on relative singular-value statistics,
ncomp(pt) as the selected number of significant effect axes.

k <- ncomp(pt)
E_sig <- truncate(E, k)

k
#> [1] 1
ncomp(E_sig)
#> [1] 1

7. Stability by bootstrap

Permutation asks whether an effect exists. Bootstrap asks whether the geometry is stable under subject resampling.

set.seed(3)
bres <- bootstrap(E, nboot = 49, resample = "subject")

print(bres)
#> Bootstrap stability for effect_operator
#> 
#> Term: group:level
#> Bootstrap samples: 49
#> Resampling unit: subject
#> Mean singular values: 19.0326, 2.3598
bres$singular_values_mean
#> [1] 19.03259  2.35981

The bootstrap result contains means and standard deviations for:

singular values,
feature loadings,
full loading arrays across resamples.

8. Array input

Repeated-measures arrays can be supplied directly. Internally they are normalized to the same stacked representation.

Y_array <- array(NA_real_, dim = c(n_subject, length(levels_within), p))
idx <- 1
for (i in seq_len(n_subject)) {
  for (j in seq_along(levels_within)) {
    Y_array[i, j, ] <- Y[idx, ]
    idx <- idx + 1
  }
}

fit_array <- mixed_regress(
  Y_array,
  design = design,
  fixed = ~ group * level,
  random = ~ 1 | subject,
  basis = shared_pca(ncomp = 4),
  preproc = center()
)

effect(fit_array, "level")$term
#> [1] "level"

9. Current scope

The present implementation is intentionally narrow:

Gaussian multivariate responses,
one grouping variable,
random intercepts and random slopes,
grouped permutation and subject bootstrap,
shared feature basis or identity basis.

The calibration harness used for the empirical checks in this vignette lives in experimental/mixed_effect_operator_calibration.R. Batch outputs can be written to experimental/results/ for larger Monte Carlo runs outside the package test suite.

Still to come:

richer exchangeability schemes for repeated-measures contrasts,
broader support beyond one grouping variable,
more extensive mixed-model examples across design families.