nct() tests whether two networks estimated from two samples of the same
variables differ, at the three levels of van Borkulo et al. (2022):
network structure (statistic M, the largest absolute edge
difference), global strength (statistic S, the difference in the
summed absolute edge weights, or with abs = FALSE in the summed signed
weights, the global expected influence) and individual edges
(statistic E, the absolute difference of each edge). Each network is an
EBIC graphical lasso network of the Pearson correlations, and each
p-value comes from re-estimating both networks on permuted group labels
(within-pair swaps with paired = TRUE).
A p-value counts the permutations whose statistic is at least the
observed one, with the observed split added:
(sum(perm >= observed) + 1) / (iter + 1), so the smallest attainable
value is 1 / (iter + 1). A non-significant test does not
show that the networks are equal; with small samples the test has little
power (van Borkulo et al. 2022). The edge tests are many tests at once;
set p_adjust before reading single edges.
Both networks are regularised estimates, so check how stable each one is
with boot_glasso().
Usage
nct(
data1,
data2,
iter = 1000L,
gamma = 0.5,
paired = FALSE,
abs = TRUE,
weighted = TRUE,
p_adjust = "none",
na_method = c("pairwise", "listwise")
)
# S3 method for class 'net_nct'
print(x, n = Inf, ...)
# S3 method for class 'net_nct'
summary(object, ...)
# S3 method for class 'net_nct'
as.data.frame(x, row.names = NULL, optional = FALSE, ...)Arguments
- data1
The observations of group 1: a numeric matrix or data frame, one row per observation. Non-numeric columns are dropped; at least two numeric variables and three rows are needed.
- data2
The observations of group 2, with the same variables as
data1in the same order; columns that differ in name or order raise an error of classnestimate_bad_input. An unnamed matrix is matched by position and takes the other group's names, orV1,V2, ... when neither group has names.- iter
The number of permutations, a whole number of at least 1. Default
1000.- gamma
The EBIC tuning parameter of the glasso, a non-negative number; larger values give sparser networks. Default
0.5.- paired
Whether row
iof both groups is the same unit (such as a person measured twice). Each permutation then swaps the two rows of a unit at random instead of shuffling all rows, and the groups need the same number of rows, otherwise an error of classnestimate_bad_inputis raised. DefaultFALSE.- abs
Whether global strength sums the absolute edge weights. With
FALSE, it sums the signed weights (global expected influence). DefaultTRUE.- weighted
Whether the tests compare the weighted networks. With
FALSE, both networks are binarised (an edge is present or absent) before the statistics are computed. DefaultTRUE.- p_adjust
The multiplicity correction for the per-edge p-values, any method of
stats::p.adjust(). Default"none".- na_method
How missing values are handled.
"pairwise"(default) keeps every row, permutes rows as usual, and estimates each sample or permuted split from its pairwise-complete correlation matrix, with the median pairwise co-observation count as the sample size; a variable pair sharing fewer than two observed rows there raises an error of classnestimate_degenerate_sample."listwise"drops incomplete rows first (withpaired = TRUE, a pair is dropped when either member is incomplete). Complete data are unaffected by the choice.- x
For the
print()andas.data.frame()methods: anet_nctobject.- n
For the
print()method: the number of edges to list, largest difference first. DefaultInf(all).- ...
For the
print(),summary()andas.data.frame()methods: ignored.- object
For the
summary()method: anet_nctobject.- row.names, optional
For the
as.data.frame()method: ignored, kept for compatibility with the base generic.
Value
A net_nct object. print() reports the three tests, each
group's global strength and the edges, largest difference first, and
returns the object invisibly; summary() returns the M and S tests as
a data frame with one row per test. The object's fields:
- nw1, nw2
The estimated weighted adjacency matrices of the two groups (binarised with
weighted = FALSE).- structure
The network-structure test (statistic M, the largest absolute edge difference):
observed,perm(the statistic in each permutation) andp_value.- strength
The global-strength test (statistic S): the same, plus
by_group, each network's global strength (NCT'sglstrinv.sep).- E
The per-edge tests:
observedandp_valuehold one value per node pair (the upper triangle),permis aniterby node-pair matrix, andedge_namesis a data frame with the columnsfromandtonaming each node pair.- n_iter
The number of permutations.
- groups
The display names of the two groups: the argument names when they are plain variables, otherwise
"group 1"and"group 2".- n
The rows used in each group.
- paired
Whether a paired test was used.
- params
The settings used:
gamma,abs,weighted,p_adjustandna_method.
The letters follow van Borkulo et al. (2022) and
NetworkComparisonTest. Before Nestimate 0.9.27 the fields M and S
held the global-strength and network-structure tests the other way
round; reading them now raises an error of class nestimate_renamed.
For the summary() method: a data frame with one row per global
test and the columns test ("network structure", and
"global strength", or "global expected influence" when
abs = FALSE), statistic ("M", "S"), observed, p_value and
significant (p_value < 0.05, a fixed level). The as.data.frame()
method gives one row per node pair, with the columns from, to,
weight_1 and weight_2 (the edge in each group's network),
difference (statistic E), p_value and significant
(p_value < 0.05).
Details
nct() follows NetworkComparisonTest::NCT() with its defaults
abs = TRUE, weighted = TRUE and paired = FALSE. The network
estimator is EBIC-selected glasso (psychnets::ebic_glasso()) applied to
a Pearson correlation matrix made positive definite with
Matrix::nearPD() (NCT's NCT_estimator_GGM default). Under the same
seed both functions permute the same rows, so the permutation p-values
agree; the glasso solver is not the Fortran one NCT wraps, so the test
statistics agree to independent-solver precision (of the order of
1e-4) rather than bit for bit. The Matrix package (in Suggests)
must be installed.
The two groups must hold the same variables in the same order;
non-numeric columns, such as an id or a group label, are dropped first.
Infinite values are refused; use finite observations or NA for missing
data.
A variable without variation in a group, or in a permuted split (a rare
binary item can be constant in a split), enters that network as an
isolated node. The statistic is then defined on every split, so the
permutation test stays exact. A constant variable in an observed group
raises a warning of class nestimate_constant_variable.
Malformed iter, gamma, paired, abs, weighted or data raise an
error of class nestimate_bad_input.
References
van Borkulo, C. D., van Bork, R., Boschloo, L., Kossakowski, J. J., Tio, P., Schoevers, R. A., Borsboom, D., & Waldorp, L. J. (2022). Comparing network structures on three aspects: A permutation test. Psychological Methods. doi:10.1037/met0000476
Examples
set.seed(1)
x1 <- matrix(rnorm(100 * 4), 100, 4)
x2 <- matrix(rnorm(100 * 4), 100, 4)
colnames(x1) <- colnames(x2) <- paste0("V", 1:4)
# iter = 20 keeps the example fast; a real analysis uses 1000 or more.
res <- nct(x1, x2, iter = 20)
res
#> Network Comparison Test: x1 vs x2
#> 20 permutations, independent groups, n = 100 and 100
#>
#> Test Statistic p
#> M network structure 0.000 1.000
#> S global strength 0.000 1.000
#> x1 0.000 vs x2 0.000
#> E edges differing 0 of 6
#>
#> Edges, largest difference first
#> Edge x1 x2 |diff| p
#> V1 - V2 0.000 0.000 0.000 1.000
#> V1 - V3 0.000 0.000 0.000 1.000
#> V1 - V4 0.000 0.000 0.000 1.000
#> V2 - V3 0.000 0.000 0.000 1.000
#> V2 - V4 0.000 0.000 0.000 1.000
#> V3 - V4 0.000 0.000 0.000 1.000
#>
#> * p < 0.05
summary(res)
#> test statistic observed p_value significant
#> 1 network structure M 0 1 FALSE
#> 2 global strength S 0 1 FALSE