Skip to contents

Edge-wise post-hoc test for modality-network differences

Source: R/test_modality_edge_differences.R
test_modality_edge_differences.Rd

After a significant global test with test_modality_graph_equality, this function identifies which specific edges contribute most strongly to the observed difference. For each edge, an empirical two-sided p-value is computed from label-permutation distributions of the absolute edge-weight difference, and Benjamini-Hochberg FDR correction is applied across all tested edges.

Usage

test_modality_edge_differences(
  x,
  y,
  n_perm = 500L,
  edges = c("all", "union"),
  p_adjust = "BH",
  strata = NULL,
  seed = NULL,
  verbose = TRUE
)

Arguments

x, y

Two catmodgraph objects on the same variable set, each carrying their $data component.

n_perm

Integer, number of permutations. Default 500.

edges

Character, which edges to test. One of "all" (default, every upper-triangle pair of modalities) or "union" (only edges present in at least one of the two input graphs). Using "union" reduces the multiple-testing burden.

p_adjust

Character, adjustment method for multiple testing, passed to p.adjust. Default "BH".

strata

Optional stratification vector of length equal to the combined sample size. See test_modality_graph_equality.

seed

Optional integer seed for reproducibility.

verbose

Logical. If TRUE (default), prints progress.

Value

An object of class catmodedgetest with components:

edge_table

Data frame with one row per tested edge: columns from, to, weight_x, weight_y, obs_diff, p_empirical, p_adjusted.

n_perm

Integer number of permutations.

n_x, n_y

Sample sizes.

edges

Character, which edge-subset criterion was used.

p_adjust_method

Adjustment method.

strata_used

Logical.

Details

For each edge, the permutation null is that the observed edge-weight difference is no larger than expected under random reassignment of sample labels. The test statistic is the observed edge-weight difference; the null distribution is obtained by recomputing that difference across n_perm random label reassignments of the combined data. Edges not present in the pruned graphs are treated as having weight 0.

The function is designed to be called after a significant global test. Calling it without a significant omnibus inflates the family-wise Type I error beyond the nominal FDR level, because the omnibus test acts as a gatekeeper under the closed-testing principle.

See also

test_modality_graph_equality for the global test.

Examples

# \donttest{
data(survey_health)
df_f <- subset(survey_health, sex == "female")[, -1]
df_m <- subset(survey_health, sex == "male")[, -1]

mg_f <- build_modality_graph(df_f)
mg_m <- build_modality_graph(df_m)

edge_test <- test_modality_edge_differences(
  mg_f, mg_m, n_perm = 200, edges = "union",
  seed = 1, verbose = FALSE
)
head(edge_test$edge_table)
#>                       from                   to    weight_x     weight_y
#> 74       diet_quality=poor   exercise_freq=high 0.031381364 0.1950809523
#> 24           age_group=55+    diet_quality=poor 0.093745693 0.1986175040
#> 89       exercise_freq=low  health_insurance=no 0.207140819 0.0009701169
#> 102      exercise_freq=low health_insurance=yes 0.207140819 0.0009701169
#> 17         age_group=35-54   bmi_category=obese 0.006335803 0.1114963387
#> 90  exercise_freq=moderate  health_insurance=no 0.200021474 0.0156441558
#>       obs_diff p_empirical p_adjusted  abs_diff
#> 74  -0.1636996 0.004975124  0.4228856 0.1636996
#> 24  -0.1048718 0.004975124  0.4228856 0.1048718
#> 89   0.2061707 0.014925373  0.6343284 0.2061707
#> 102  0.2061707 0.014925373  0.6343284 0.2061707
#> 17  -0.1051605 0.019900498  0.6766169 0.1051605
#> 90   0.1843773 0.049751244  0.7611940 0.1843773
# }