Skip to contents

Estimates the stability of centrality measures in a network using subset sampling without replacement. It allows for dropping varying proportions of cases and calculates correlations between the original centralities and those computed using sampled subsets.

Usage

estimate_cs(
  x,
  loops,
  normalize,
  invert,
  measures,
  iter,
  method,
  drop_prop,
  threshold,
  certainty,
  progressbar = NULL,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

estimate_centrality_stability(
  x,
  loops,
  normalize,
  invert,
  measures,
  iter,
  method,
  drop_prop,
  threshold,
  certainty,
  progressbar = NULL,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

# S3 method for class 'tna'
estimate_cs(
  x,
  loops = FALSE,
  normalize = FALSE,
  invert = TRUE,
  measures = c("InStrength", "OutStrength", "Betweenness"),
  iter = 1000,
  method = "pearson",
  drop_prop = seq(0.1, 0.9, by = 0.1),
  threshold = 0.7,
  certainty = 0.95,
  progressbar = FALSE,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

# S3 method for class 'tna'
estimate_centrality_stability(
  x,
  loops = FALSE,
  normalize = FALSE,
  invert = TRUE,
  measures = c("InStrength", "OutStrength", "Betweenness"),
  iter = 1000,
  method = "pearson",
  drop_prop = seq(0.1, 0.9, by = 0.1),
  threshold = 0.7,
  certainty = 0.95,
  progressbar = FALSE,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

# S3 method for class 'group_tna'
estimate_cs(
  x,
  loops = FALSE,
  normalize = FALSE,
  invert = TRUE,
  measures = c("InStrength", "OutStrength", "Betweenness"),
  iter = 1000,
  method = "pearson",
  drop_prop = seq(0.1, 0.9, by = 0.1),
  threshold = 0.7,
  certainty = 0.95,
  progressbar = FALSE,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

# S3 method for class 'group_tna'
estimate_centrality_stability(
  x,
  loops = FALSE,
  normalize = FALSE,
  invert = TRUE,
  measures = c("InStrength", "OutStrength", "Betweenness"),
  iter = 1000,
  method = "pearson",
  drop_prop = seq(0.1, 0.9, by = 0.1),
  threshold = 0.7,
  certainty = 0.95,
  progressbar = FALSE,
  detailed = NULL,
  actor = NULL,
  edges = FALSE
)

Arguments

x

A tna or a group_tna object representing the temporal network analysis data. The object should be created from a sequence data object. Grouped models with scaled edge weights, and their networks, are not resampled (an error of class tna_error_scaling).

loops

A logical value indicating whether to include loops in the network when computing the centrality measures. The default is FALSE.

normalize

A logical value indicating whether to normalize the centrality measures. The default is FALSE.

invert

A logical value indicating whether the weights should be inverted for distance-based measures. The default is TRUE.

measures

A character vector of centrality measures to estimate. The default measures are "InStrength", "OutStrength", and "Betweenness". See centralities() for a list of available centrality measures.

iter

An integer specifying the number of resamples to draw. The default is 1000.

method

A character string indicating the correlation coefficient type. The default is "pearson". See stats::cor() for details.

drop_prop

A numeric vector specifying the proportions of cases to drop in each sampling iteration. Default is a sequence from 0.1 to 0.9 in increments of 0.1.

threshold

A numeric value specifying the correlation threshold for calculating the CS-coefficient. The default is 0.7.

certainty

A numeric value specifying the desired level of certainty for the CS-coefficient. Default is 0.95. The certainty is the share of all subsamples whose correlation reaches threshold; a subsample in which the correlation is undefined (the measure is constant in it) counts as one that does not reach it.

progressbar

A logical value. If TRUE, a progress bar is displayed. Defaults to FALSE.

detailed

Deprecated. This argument is ignored and will be removed in a future version.

actor

An optional actor identifier for sequences nested in actors: a character string naming a non-sequence column of the data the model was built from, or a vector with one identifier per sequence (per sequence of the data the grouped model was built from, for a group_tna object). When given, drop_prop is the proportion of actors dropped, and each actor's sequences are dropped together. The default NULL drops single sequences.

edges

A logical value. If TRUE, the stability of the edge weights is also estimated and returned as the element Edges. The default is FALSE.

Value

A tna_stability object which is a list with an element for each measure (and an element Edges when edges = TRUE) with the following elements:

  • cs_coefficient: The centrality stability (CS) coefficient of the measure.

  • correlations: A matrix of correlations between the original centrality and the resampled centralities, with one row per resample (iter rows) and one column per drop proportion.

The drop proportions and the correlation threshold are stored in the attributes drop_prop and threshold. Use summary.tna_stability() for the CS-coefficients as a table.

If x is a group_tna object, a group_tna_stability object is returned instead, which is a list of tna_stability objects.

Details

The function works by repeatedly resampling the data, dropping varying proportions of cases, and calculating centrality measures on the subsets. The correlation between the original centralities and the resampled centralities is calculated for each drop proportion. The stability of each centrality measure is then summarized using a centrality stability (CS) coefficient, which represents the proportion of dropped cases at which the correlations drop below a given threshold (default 0.7).

With edges = TRUE, the stability of the edge weights is estimated in the same way: in each subsample the network is estimated again and its edge weights are correlated with the edge weights of the original network (Epskamp et al., 2018). Only the edges present in the original network are compared, self-loops included. An edge that is absent from the original data cannot appear in a subsample, so including the absent edges would add pairs of zeros that always agree and inflate the correlation. The loops, normalize, and invert arguments do not affect the edge weights.

The results can be visualized by plotting the output object showing the stability of the centrality measures across different drop proportions, along with confidence intervals. The CS-coefficients are displayed in the subtitle.

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

# A small subset and few iterations
estimate_cs(
  tna(head(group_regulation, 200)),
  drop_prop = c(0.3, 0.6),
  measures = "InStrength",
  iter = 5
)
#> Centrality Stability Coefficients
#> 
#> InStrength 
#>        0.6 
# \donttest{
model <- tna(group_regulation)
# Small number of iterations and drop proportions for CRAN
estimate_cs(
  model,
  drop_prop = seq(0.3, 0.9, by = 0.2),
  measures = c("InStrength", "OutStrength"),
  iter = 10
)
#> Centrality Stability Coefficients
#> 
#>  InStrength OutStrength 
#>         0.9         0.9 
# Stability of the edge weights alongside the centralities
estimate_cs(
  model,
  drop_prop = seq(0.3, 0.9, by = 0.2),
  measures = "InStrength",
  iter = 10,
  edges = TRUE
)
#> Centrality and Edge Stability Coefficients
#> 
#> InStrength      Edges 
#>        0.9        0.9 
# }