Differential edges between two groups

Tests which edges differ between two groups of an ssjgl fit, by refitting ssjgl() with the same settings on permuted or bootstrapped data.

Usage

diff_edges(
  fit,
  groups = c(1, 2),
  score = c("pcor", "precision"),
  method = c("permutation", "bootstrap"),
  B = 100,
  alpha = 0.05,
  Y = NULL,
  v0_index = NULL,
  ...
)

# S3 method for class 'ssjgl_diff_edges'
print(x, n = 10, ...)

Arguments

fit

An ssjgl object.

groups

Two distinct group indices (or group names) to compare. The difference is group 1 minus group 2.

score

Scale of the comparison: "pcor" (partial correlations, default) or "precision" (precision-matrix entries).

method

"permutation": (default) Rows of the two groups are pooled and the group labels shuffled (group sizes kept; any other groups are left as they are), and ssjgl() is refitted B times. The null scores $|\Delta|$ are pooled across all edges and permutations. Each edge's p-value is $(1 + \#\{null \ge |\Delta_{obs}|\}) / (1 + \#null)$, and the p-values are adjusted by Benjamini-Hochberg. selected is p_adj < alpha.
"bootstrap": Rows are resampled with replacement within every group, ssjgl() is refitted B times, and lower/upper are percentile $1 - \alpha$ intervals for $\Delta$. selected means the interval excludes 0. This gives no multiplicity control.

B

Number of permutations or bootstrap resamples. Default 100.

alpha

BH level (permutation) or 1 - interval coverage (bootstrap). Default 0.05.

Y

The list of data matrices fit was estimated from. If NULL (default), the Y argument of fit$call is evaluated in the calling environment; supply it explicitly if that object has changed or is not available.

v0_index

Ladder step whose estimates are compared. Default: the last step.

...

Arguments passed to the ssjgl() refits, overriding the settings stored in fit (e.g. maxitr.em to make the refits faster).

x

An ssjgl_diff_edges object.

n

Number of selected edges to print. Default 10.

Value

A data frame of class "ssjgl_diff_edges", one row per edge $i < j$ in upper.tri() order, with columns

i, j: Variable indices.
node1, node2: Variable names.
est_g1, est_g2: Estimates in the two groups on the score scale.
diff: est_g1 - est_g2.
pvalue, p_adj: Permutation p-value and its BH adjustment (method = "permutation").
lower, upper: Bootstrap percentile interval for diff (method = "bootstrap").
selected: Logical: called differential.
prob_dissimilar: $P(\xi = 1)$ (only if the fit is doubly).

Attributes: method, score, groups, group_names, B (successful refits), alpha, v0_index, and null_scores (the pooled null, permutation) or boot_diffs (a B x edges matrix, bootstrap).

Details

Refits use the arguments stored in fit$args, so they repeat the whole v0 ladder of the original fit. The resampling indices are drawn in the calling R session, so results are reproducible with set.seed() and do not depend on how the refits are run.

Parallel refits. If the future.apply package is installed, the refits are run with future.apply::future_lapply(), so they are parallel under any future::plan() other than sequential, e.g. future::plan(future::multisession, workers = 4).

prob_dissimilar. With doubly = TRUE the table also has prob_dissimilar, the EM's posterior probability $P(\xi_{jk} = 1)$ that edge $(j, k)$ is in the slab of the cross-group similarity penalty, i.e. is only weakly shrunk towards equality across groups. It is shared by all groups, comes from the point-estimating EM, and is not a calibrated probability that the edge differs between groups. In practice it can be uninformative: in one real fit (TCGA PI3K pathway, p = 28, group penalty), every edge with $P(\delta = 1) > 0.5$ also had $P(\xi = 1) > 0.5$. Use the permutation p-values for inference.

Column names. Where the concepts match, the columns follow multiGGMr's diff_summary() and select_diff_edges() (i, j, lower, upper), so the shared columns of the two packages' tables can be bound together.

Examples

# \donttest{
sim <- simulate_ssjgl_data(K = 2, p = 8, n = 60, seed = 1)
fit <- ssjgl(sim$data_list, lambda0 = 1, lambda1 = 1, lambda2 = 1,
             standardize = TRUE, impute = FALSE)
set.seed(1)
de <- diff_edges(fit, Y = sim$data_list, B = 20)
de
#> <ssjgl_diff_edges>
#>   group 1 vs group 2, pcor scale, permutation (B = 20), v0 step 3
#>   4 of 28 edges selected at BH-adjusted p < 0.05
#>  i j node1 node2 est_g1  est_g2   diff  pvalue  p_adj selected
#>  1 2    V1    V2  0.579 -0.0574  0.636 0.00178 0.0125     TRUE
#>  4 5    V4    V5  0.193 -0.3917  0.584 0.00178 0.0125     TRUE
#>  2 3    V2    V3 -0.269  0.2658 -0.535 0.00178 0.0125     TRUE
#>  2 4    V2    V4 -0.182  0.2619 -0.444 0.00178 0.0125     TRUE
plot(de)

# }

Usage

Arguments

Value

Details

See also

Examples