Centrality Stability Coefficient (CS-coefficient)
Source:R/centrality_stability.R
centrality_stability.Rdcentrality_stability() estimates how stable a network's centrality
indices are under case-dropping (Epskamp, Borsboom & Fried, 2018). For
each proportion in drop_prop, that share of the sequences (or, for
association networks, of the rows) is removed at random iter times,
the network is re-estimated on what is left, and its centralities are
correlated with those of the full network. The CS-coefficient of a
measure is the largest proportion that can be dropped while the
correlation stays at or above threshold in at least a certainty
share of the draws. It takes a netobject from build_network() of any
method, a bare cograph_network, or a netobject_group or mcml, whose
networks are assessed one by one.
The CS-coefficient says how well the nodes' centralities (their values,
or with method = "spearman" or "kendall" their ordering) survive the
loss of cases; Epskamp et al. (2018) suggest that it should
not be below 0.25, and preferably above 0.5, before centrality
differences are interpreted. It does not say whether two nodes differ in
centrality, nor how precise single edges are: check the edges with
bootstrap_network(), and for glasso networks the centrality
differences with boot_glasso().
Usage
centrality_stability(
x,
measures = c("InStrength", "OutStrength", "Betweenness"),
iter = 1000L,
drop_prop = seq(0.1, 0.9, by = 0.1),
threshold = 0.7,
certainty = 0.95,
method = "pearson",
centrality_fn = NULL,
loops = FALSE,
normalize = FALSE,
invert = TRUE,
normalize_diffusion = TRUE,
seed = NULL
)
# S3 method for class 'net_stability'
print(x, ...)
# S3 method for class 'net_stability_group'
print(x, ...)
# S3 method for class 'net_stability_group'
summary(object, ...)
# S3 method for class 'net_stability'
summary(object, ...)
# S3 method for class 'net_stability'
plot(x, ...)Arguments
- x
The network: a
netobjectfrombuild_network(), a barecograph_network(converted to one), or anetobject_groupormcml, each of whose networks is assessed. Any other object, or a network built without its data, raises an error. For theprint()methods: anet_stabilityornet_stability_group; for theplot()method: anet_stability.- measures
The centrality measures to assess. The default is
"InStrength","OutStrength"and"Betweenness";NULLtakesnet_centrality()'s default ("InStrength","Betweenness","Diffusion"), and"all"every built-in measure:"OutStrength","InStrength","ClosenessIn","ClosenessOut","Closeness","Betweenness","BetweennessRSP","Diffusion"and"Clustering"(net_centrality()defines them)."InCloseness"and"OutCloseness"are accepted for"ClosenessIn"and"ClosenessOut". Other names needcentrality_fn; without it they raise an error.- iter
The number of random subsets drawn at each drop proportion, at least 2. Default
1000.- drop_prop
The proportions of cases to drop, each strictly between 0 and 1. Default
seq(0.1, 0.9, by = 0.1). A proportion that would leave fewer than 2 cases is skipped and counts as not reachingthreshold. The CS-coefficient is the largest proportion that meetscertainty, in whatever order they are given.- threshold
The correlation with the full network's centralities that a draw must reach, between 0 and 1. Default
0.7.- certainty
The share of draws at a drop proportion that must reach
threshold, between 0 and 1. Default0.95. It is a share of all draws: one with an undefined correlation counts as not reachingthreshold.- method
The correlation between the full and the subset centralities:
"pearson"(default),"spearman"or"kendall".- centrality_fn
A function of your own that takes the weight matrix and returns a named list of centrality vectors, such as
list(my_metric = ...). It is called ascentrality_fn(mat)and used only for the requested measures that are not built-in; list their names inmeasures. DefaultNULL.- loops
Whether self-loops (staying in a state) count. With
FALSE(default) they are left out of every measure; the stored matrix is not changed.- normalize
Whether to range-normalise every measure to 0–1, as
tna::centralities(normalize = TRUE)does. DefaultFALSE.- invert
Whether a heavier edge is a shorter step in the shortest-path measures (distance = 1 / weight), as in
tna. DefaultTRUE; withFALSEthe weight is the distance. It affects betweenness and closeness only, not"BetweennessRSP".- normalize_diffusion
Whether to range-normalise
Diffusioneven whennormalize = FALSE. DefaultTRUE.- seed
An integer that makes the random subsets reproducible. Default
NULL. The session's own random stream is left as it was; a group's networks are each started from the same seed.- ...
- object
For the
summary()methods: anet_stabilityornet_stability_group.
Value
A net_stability object. print() shows the CS-coefficient of
each measure; summary() returns one row per measure and drop
proportion, with the mean and SD of the correlations (over the draws
where it is defined) and prop_above, the share of all draws that
reach threshold; plot() draws the mean correlation against the
proportion dropped, one line per measure with a band of one SD and a
dashed line at threshold, and returns the ggplot invisibly. For a
netobject_group or mcml, the result is a net_stability_group, a
named list of net_stability objects, one per network, whose print()
shows the CS-coefficients as a networks-by-measures table and whose
summary() stacks the per-network summaries with a leading group
column. The print() methods return their input invisibly. The
net_stability's fields:
- cs
The CS-coefficients, a named numeric vector with one per retained measure.
- correlations
A named list, one per retained measure, of
iterxlength(drop_prop)matrices of correlations.- measures
The measures assessed (see the rule for measures without variance under Details).
- drop_prop, threshold, certainty, iter, method
The settings used.
Details
For transition methods, per-sequence count matrices are computed once
and each draw only re-sums them; the network's scaling and threshold
are applied to every re-estimated network as to the original. For the
other methods the estimator is re-run on each subset with the network's
params (after the multilevel decomposition, when the network was built
with a level and an id column and its estimator is undirected).
The built-in measures are computed from the weight matrix
directly, without igraph, as in net_centrality(): strength from row
and column sums, betweenness and closeness from Floyd-Warshall shortest
paths. A network that rests on one sequence, or one-hot data with no
actor or session, has nothing to drop and raises an error of class
nestimate_single_sequence.
Every draw counts towards certainty: a draw whose correlation is
undefined (the measure is constant in it), and a draw whose network
could not be estimated, count as not reaching threshold. Failed
estimates are reported in a warning of class
nestimate_resample_failed.
Measures without variance are handled by two different rules, both
long-standing behaviour. When some requested measures do not vary
across the nodes of the original network (or are undefined on it), they
are dropped, and the result covers only the retained ones. When every
requested measure is so, a warning of class nestimate_constant_measure
is issued and all requested names
are returned with a CS-coefficient of 0 and correlations that are all
NA. A measure counts as constant when its values span no more than
the square root of the machine epsilon relative to their size, so
"OutStrength" of a "relative" network with loops = TRUE, which is
1 for every node up to rounding error, is dropped rather than correlated
on that error. The same rule makes a measure constant in a draw an
undefined correlation.
References
Epskamp, S., Borsboom, D., & Fried, E. I. (2018). Estimating psychological networks and their accuracy: A tutorial paper. Behavior Research Methods 50(1), 195-212. doi:10.3758/s13428-017-0862-1
Examples
seqs <- data.frame(
T1 = c("plan", "code", "debug", "plan", "test", "code"),
T2 = c("code", "debug", "code", "plan", "code", "test"),
T3 = c("debug", "code", "plan", "code", "debug", "plan"),
T4 = c("test", "plan", "test", "debug", "plan", "code")
)
net <- build_network(seqs, method = "relative")
cs <- centrality_stability(net, iter = 10, drop_prop = 0.3, seed = 1)
# \donttest{
set.seed(1)
seqs <- data.frame(
V1 = sample(LETTERS[1:4], 30, TRUE), V2 = sample(LETTERS[1:4], 30, TRUE),
V3 = sample(LETTERS[1:4], 30, TRUE), V4 = sample(LETTERS[1:4], 30, TRUE)
)
net <- build_network(seqs, method = "relative")
cs <- centrality_stability(net, iter = 100, seed = 42,
measures = c("InStrength", "OutStrength"))
print(cs)
#> Centrality Stability (100 iterations, threshold = 0.7)
#> Drop proportions: 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9
#>
#> CS-coefficients:
#> InStrength 0.00
#> OutStrength 0.30
# }