Skip to contents

Partial Least Squares Correlation (PLSC) with Inference

PLSC.Rmd

1. What this vignette covers

This walk-through shows how to:

  1. Fit a Behavior-style PLSC model (plsc()),
  2. Inspect latent variables (LVs) and scores,
  3. Run a permutation test to decide how many LVs are significant, and
  4. Use bootstrap ratios to identify stable loadings, with simple visualizations.

The workflow is intentionally small and fast so the vignette runs quickly; for real analyses, increase the number of permutations/bootstraps.

2. Simulate coupled X/Y blocks 

n  <- 80   # subjects
pX <- 8    # brain/features
pY <- 5    # behavior
d  <- 2    # true latent dimensions

# orthonormal loadings
Vx_true <- qr.Q(qr(matrix(rnorm(pX * d), pX, d)))
Vy_true <- qr.Q(qr(matrix(rnorm(pY * d), pY, d)))

F_scores <- matrix(rnorm(n * d), n, d)         # latent scores
noise    <- 0.10

X <- F_scores %*% t(Vx_true) + noise * matrix(rnorm(n * pX), n, pX)
Y <- F_scores %*% t(Vy_true) + noise * matrix(rnorm(n * pY), n, pY)

3. Fit PLSC 

fit_plsc <- plsc(X, Y, ncomp = 3,                   # request a few extra comps
                 preproc_x = standardize(),         # correlation-scale
                 preproc_y = standardize())

fit_plsc$singvals
#> [1] 4.15551071 1.76186530 0.02831984
fit_plsc$explained_cov
#> [1] 8.475956e-01 1.523650e-01 3.936602e-05

4. Inspect scores (brain vs behavior) 

scores_df <- data.frame(
  LV1_x = scores(fit_plsc, "X")[, 1],
  LV1_y = scores(fit_plsc, "Y")[, 1],
  LV2_x = scores(fit_plsc, "X")[, 2],
  LV2_y = scores(fit_plsc, "Y")[, 2]
)

ggplot(scores_df, aes(x = LV1_y, y = LV1_x)) +
  geom_point(alpha = 0.7) +
  geom_smooth(method = "lm", se = FALSE, color = "firebrick") +
  labs(x = "Behavior scores (LV1)", y = "Brain scores (LV1)",
       title = "Score association for LV1") +
  theme_minimal()
#> `geom_smooth()` using formula = 'y ~ x'

5. Permutation test: how many LVs?

Shuffle rows of Y to break the X–Y link and build an empirical null for the singular values.

set.seed(123)
pt <- perm_test(fit_plsc, X, Y, nperm = 199, comps = 3, parallel = FALSE)
pt$component_results
#> # A tibble: 3 × 5
#>    comp observed  pval lower_ci upper_ci
#>   <int>    <dbl> <dbl>    <dbl>    <dbl>
#> 1     1   4.16   0.005   0.269    1.07  
#> 2     2   1.76   0.005   0.0819   0.509 
#> 3     3   0.0283 0.205   0.0140   0.0396
cat("Sequential n_significant (alpha = 0.05):", pt$n_significant, "\n")
#> Sequential n_significant (alpha = 0.05): 2

6. Bootstrap ratios for stable loadings

Bootstrap resamples subjects, re-fits PLSC, sign-aligns loadings, and reports mean/SD (ratio ≈ Z).
Here we keep it light for the vignette; use ≥500–1000 in practice.

set.seed(321)
boot_plsc <- bootstrap(fit_plsc, nboot = 120, X = X, Y = Y, comps = 2, parallel = FALSE)

# X-loadings bootstrap ratios for LV1
df_bsr <- data.frame(
  variable = paste0("x", seq_len(pX)),
  bsr      = boot_plsc$z_vx[, 1]
)

ggplot(df_bsr, aes(x = reorder(variable, bsr), y = bsr)) +
  geom_hline(yintercept = c(-2, 2), linetype = "dashed", color = "grey50") +
  geom_col(fill = "#1f78b4") +
  coord_flip() +
  labs(x = NULL, y = "Bootstrap ratio (X loadings, LV1)",
       title = "Stable variables exceed |BSR| ≈ 2") +
  theme_minimal()

Do the same for Y loadings if needed:

7. Practical tips

  • Use standardize() to run PLSC on correlation scale (default here); switch to center() for covariance scale.
  • Increase nperm (e.g., 999) and nboot (≥500) for publication-grade inference.
  • Interpret loadings only where bootstrap ratios exceed ~|2|, and only for LVs that pass the permutation test.
  • To visualise higher-dimensional loading maps (e.g., imaging), replace the bar plots with your spatial plotting routine applied to boot_plsc$z_vx.