bootstrap_network() resamples the cases a network was estimated from,
re-estimates the network on every resample with the same method and
settings, and reports for each edge its bootstrap mean, standard
deviation, confidence interval and a p-value, together with the network
pruned to the edges that pass. It reads any network from
build_network(), built-in method or registered estimator, as well as a
netobject_group or mcml (every network is bootstrapped) and a
wtna_mixed (both of its components). The cases are the sequences of a
transition network (for one-hot data, its actors or sessions) and the
rows of the data otherwise; with actor, whole actors are resampled (see
Nested data and actor).
The p-value comes from one of two tests. With inference = "stability"
it counts the replicates that fall outside a consistency range around
the original weight (by default 75% to 125% of it), so a significant edge
is one whose weight the resamples reproduce within that range. With
inference = "threshold" it counts the replicates whose absolute weight
falls below edge_threshold, so a significant edge stays above that
threshold. In both, the p-value is that count plus one, divided by the
number of replicates plus one, so it is never 0. Edges with a p-value
below ci_level are kept in the pruned network.
The p-value counts replicates, so it cannot fall below
1 / (iter + 1): with too few iterations no edge can pass a small
ci_level. The stability test asks whether a weight is reproduced
within a range, not whether it differs from zero, and because the range
is a proportion of the weight, a weak edge has to be reproduced within a
narrow band; read the interval beside the test. The bootstrap describes
how the estimate varies over resamples of one sample; to compare groups,
use permutation().
Usage
bootstrap_network(
x,
iter = 1000L,
ci_level = 0.05,
inference = "stability",
consistency_range = c(0.75, 1.25),
edge_threshold = NULL,
seed = NULL,
boundary = c("inclusive", "strict"),
ci_method = c("percentile", "basic"),
actor = NULL
)
# S3 method for class 'net_bootstrap'
print(x, ...)
# S3 method for class 'net_bootstrap'
summary(object, ...)
# S3 method for class 'net_bootstrap_group'
print(x, ...)
# S3 method for class 'net_bootstrap_group'
summary(object, ...)
# S3 method for class 'wtna_boot_mixed'
print(x, ...)
# S3 method for class 'wtna_boot_mixed'
summary(object, ...)Arguments
- x
The network: a
netobjectthat carries its data (any network frombuild_network()), or acograph_network, which is converted first. Anetobject_groupormcmlbootstraps every network it holds (for anmcml, the macro network and each cluster's), and awtna_mixedbootstraps both of its components (see Value). A network without its data raises an error, and one that rests on a single sequence, or one-hot data with no actor or session, has nothing to resample and raises an error of classnestimate_single_sequence. For theprint()methods: the result ofbootstrap_network().- iter
The number of bootstrap replicates, at least 2. Default
1000.- ci_level
The significance level, between 0 and 1: the confidence intervals cover
1 - ci_level, and an edge is significant when its p-value is belowci_level. Default0.05.- inference
The test behind the p-values.
"stability"(default) counts the replicates outside the consistency range around the original weight;"threshold"counts the replicates whose absolute weight is belowedge_threshold.- consistency_range
For
inference = "stability": the two multipliers of the original weight that bound the consistency range. Defaultc(0.75, 1.25). For a negative weight the bounds are swapped, so the range still runs from the lower to the upper value. Not used by"threshold".- edge_threshold
For
inference = "threshold": the absolute weight a replicate must reach. DefaultNULL, the 10th percentile of the non-zero absolute original weights. Not used by"stability".- seed
A number that makes the resamples reproducible. The session's own random stream is left as it was. For a
netobject_groupormcml, each network is bootstrapped from the same seed; for awtna_mixed, the seed is set once, before both components. DefaultNULL.- boundary
For
inference = "stability": how a replicate that lands exactly on a bound of the consistency range counts."inclusive"(default, as in tna) counts it as outside (\(\le\) / \(\ge\));"strict"counts only replicates strictly outside (\(<\) / \(>\)).- ci_method
The confidence interval for the edge weights.
"percentile"(default) uses the empirical bootstrap quantiles (Efron)."basic"reflects those quantiles around the observed weight, \((2\hat{\theta} - q_{1-\alpha/2}, 2\hat{\theta} - q_{\alpha/2})\) (Davison & Hinkley 1997, eq. 5.6), which corrects first-order bootstrap bias but can give bounds outside the natural weight range near its limits (for example below 0 for a transition probability close to 0).- actor
The column identifying the actor each sequence belongs to (for example
"student_id"for sessions nested in students,"Group"for students nested in teams), looked up in the network's metadata or its wide sequence data. With it, whole actors are resampled; see Nested data andactor. DefaultNULL: sequences are resampled one by one. It is available for sequence-based transition networks ("relative","frequency","co_occurrence"); a method outside the fast path (see Details) raises an error of classnestimate_actor_unsupported. A value that is not a single column name, or that names fewer than two actors, raises an error of classnestimate_bad_actor; a column that is absent or has missing values, one of classnestimate_actor_missing; and ids that do not give one actor per sequence, one of classnestimate_actor_misaligned.- ...
- object
For the
summary()methods: the result ofbootstrap_network().
Value
A net_bootstrap object. print() lists the five strongest
significant edges (by absolute bootstrap mean, with their interval)
and reports the iterations, the number of significant edges, the
interval level and the test, and with actor the ICC and design
effect; summary() returns one row per edge of the original network;
cograph::splot() draws the bootstrapped network; net_prune() with
method = "bootstrap" removes the non-significant edges from the
original network. The fields:
- original
The original
netobject.- mean, sd
The bootstrap mean and standard deviation of each weight, as matrices.
- p_values
The matrix of p-values.
- significant
The original weights where the p-value is below
ci_level, and 0 elsewhere, self-loops included.- ci_lower, ci_upper
The matrices of interval bounds.
- cr_lower, cr_upper
The matrices of consistency-range bounds (used by
"stability"only).- summary
A data frame with one row per non-zero original edge (one row per unordered pair in an undirected network) and the columns
from,to,weight,mean,sd,p_value,sig,ci_lowerandci_upper, pluscr_lowerandcr_upperwithinference = "stability".- model
The pruned
netobject, whose weights aresignificant: its non-significant edges, self-loops included, are zero. (The networknet_prune()returns keeps every self-loop.)- method, params, iter, ci_level, inference, ci_method
The bootstrap settings.
- consistency_range, edge_threshold
The test settings.
- actor, n_actors
The
actorcolumn and its number of actors;NULLwithoutactor.- clustering
Only with
actor: a one-row data frame ofn_sequences,n_actors,icc,icc_ci_lower,icc_ci_upperanddeff_edges.- clustering_edges
Only with
actor: one row per edge ofsummary, withfrom,to,icc,sd_actor,sd_sequenceanddeff.
A netobject_group or mcml gives a net_bootstrap_group, a named list
of net_bootstrap results: its print() reports each network's
significant and total edges and the significant edges all networks
share (when the networks have edges in common); its summary() is
described with the methods below. A wtna_mixed gives a
wtna_boot_mixed holding the transition and cooccurrence results.
The print() methods return their input invisibly. The
summary() method of a net_bootstrap returns its per-edge table, the
summary field described above. For a net_bootstrap_group
it stacks those tables into one data frame, with a first column group
naming the network each row comes from; for a wtna_boot_mixed it
returns a list of the transition and cooccurrence tables.
Details
The two tests and the inclusive boundary rule follow tna::bootstrap().
Transition networks ("relative", "frequency", "co_occurrence", and
wtna networks, which hold the same one-hot counts) take a fast path:
the count matrix of every sequence is computed once, and each replicate
sums the counts of the sequences drawn, then applies the network's
normalisation, scaling and threshold. With params = list(concat = k),
the consecutive groups of up to k rows merged by the estimator are the
sequences resampled; joins inside each merged sequence are retained.
With actor, a merged sequence must belong entirely to one actor. Every
other method (the association methods, "attention", "ngram", "gap",
"reverse" and registered estimators) is re-estimated in full on
resampled rows. A replicate whose estimation fails, or that loses a
node, is left out of the p-values, with a warning of class
nestimate_resample_failed that counts them, and the bootstrap
statistics are computed over the remaining replicates; if every
replicate fails, every p-value is 1.
A network built from an edge list (for example an mcml from
build_mcml() with from and to) has no sequences, so each
transition is resampled as an independent case, with a warning of class
nestimate_edgelist_resampling that points to the alternatives.
Nested data and actor
The bootstrap resamples sequences as independent units. When sequences
are nested in actors (sessions in students, students in teams),
actor names the column identifying the actor, and whole actors
are resampled with replacement, keeping all their sequences together:
the cluster bootstrap that resamples at the top level only (Davison &
Hinkley, 1997, section 3.8; Field & Welsh, 2007). The number of
sequences per replicate then varies with the actors drawn.
With actor, the result also reports the nesting effect. The ICC
is the proportion of the total variance that lies between actors (Shrout
& Fleiss, 1979); an ICC close to 0 indicates little evidence of a nesting
effect. It is computed as in permutation. The design effect
is the ratio of the variance under the nested design to the variance had
the sequences been sampled independently (Kish, 1965): here, the variance
of the edge weights over actor-level replicates divided by their variance
over sequence-level replicates drawn in the same run, reported as the
median over edges. actor is available for transition networks
("relative", "frequency", "co_occurrence").
References
Davison, A. C., & Hinkley, D. V. (1997). Bootstrap Methods and Their Application. Cambridge University Press.
Field, C. A., & Welsh, A. H. (2007). Bootstrapping clustered data. Journal of the Royal Statistical Society: Series B, 69(3), 369-390.
Kish, L. (1965). Survey Sampling. Wiley.
Shrout, P. E., & Fleiss, J. L. (1979). Intraclass correlations: Uses in assessing rater reliability. Psychological Bulletin, 86(2), 420-428.
See also
certainty for the closed-form Bayesian counterpart
(same result layout, no resampling);
build_network
Examples
net <- build_network(data.frame(V1 = c("A","B","C"), V2 = c("B","C","A")),
method = "relative")
boot <- bootstrap_network(net, iter = 10)
# \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")
boot <- bootstrap_network(net, iter = 100)
print(boot)
#> Bootstrap Network [Transition Network (relative) | directed]
#> Iterations : 100 | Nodes : 4
#> Edges : 0 significant / 16 total
#> CI : 95% | Inference: stability | CR [0.75, 1.25]
summary(boot)
#> from to weight mean sd p_value sig ci_lower ci_upper
#> 1 A A 0.1304348 0.1274665 0.08074773 0.6831683 FALSE 0.00000000 0.3022826
#> 2 A B 0.3043478 0.2997861 0.08278592 0.3663366 FALSE 0.14284091 0.4582168
#> 3 A C 0.1304348 0.1418228 0.07482973 0.6732673 FALSE 0.00000000 0.3000000
#> 4 A D 0.4347826 0.4309246 0.11390253 0.3267327 FALSE 0.24720280 0.6570000
#> 5 B A 0.1724138 0.1737516 0.06474250 0.5346535 FALSE 0.07821429 0.3180000
#> 6 B B 0.4137931 0.4205944 0.09336440 0.2970297 FALSE 0.23541667 0.5884848
#> 7 B C 0.2068966 0.2050824 0.07346017 0.4752475 FALSE 0.06981818 0.3420833
#> 8 B D 0.2068966 0.2005716 0.08216875 0.5049505 FALSE 0.05967023 0.3751667
#> 9 C A 0.4000000 0.3857941 0.10757329 0.3861386 FALSE 0.17766798 0.6044118
#> 10 C B 0.2000000 0.1884783 0.09391798 0.6534653 FALSE 0.04761905 0.3828755
#> 11 C C 0.1500000 0.1571277 0.07338105 0.6039604 FALSE 0.00000000 0.2994885
#> 12 C D 0.2500000 0.2686000 0.09868840 0.5544554 FALSE 0.10360963 0.4642308
#> 13 D A 0.1666667 0.1651025 0.08394913 0.6336634 FALSE 0.02261905 0.3436275
#> 14 D B 0.2222222 0.2362980 0.11192342 0.6930693 FALSE 0.00000000 0.4332589
#> 15 D C 0.4444444 0.4507200 0.14203236 0.3861386 FALSE 0.18312325 0.7498039
#> 16 D D 0.1666667 0.1478795 0.11004755 0.8118812 FALSE 0.00000000 0.3750000
#> cr_lower cr_upper
#> 1 0.09782609 0.1630435
#> 2 0.22826087 0.3804348
#> 3 0.09782609 0.1630435
#> 4 0.32608696 0.5434783
#> 5 0.12931034 0.2155172
#> 6 0.31034483 0.5172414
#> 7 0.15517241 0.2586207
#> 8 0.15517241 0.2586207
#> 9 0.30000000 0.5000000
#> 10 0.15000000 0.2500000
#> 11 0.11250000 0.1875000
#> 12 0.18750000 0.3125000
#> 13 0.12500000 0.2083333
#> 14 0.16666667 0.2777778
#> 15 0.33333333 0.5555556
#> 16 0.12500000 0.2083333
# Students nested in teams: resample whole teams
teams <- build_network(group_regulation_long, method = "relative",
actor = "Actor", action = "Action", time = "Time",
group = "Achiever")
bootstrap_network(teams, iter = 100, actor = "Group", seed = 1)
#> Edge High Low
#> --------------------------------------------
#> synthesis→consensus 0.579 0.388
#> cohesion→consensus 0.536 0.451
#> adapt→consensus 0.519 0.458
#> consensus→plan 0.364 0.432
#> discuss→consensus 0.424 0.215
#> ... and 33 more shared significant edges
#>
#> Grouped Bootstrap [2 groups | 100 iterations | 95% CI]
#> High 41 sig / 76 total | Group: 100 actors, ICC -0.003, design effect 1.02
#> Low 44 sig / 75 total | Group: 100 actors, ICC -0.001, design effect 0.98
#> Shared (all groups) 38 edges
# }