Skip to contents

Edge-weight stability coefficient (case-dropping subset bootstrap)

Source: R/reliability.R
net_casedrop_reliability.Rd

The edge-vector complement to net_stability(). For each drop proportion the network is re-estimated on random case-dropped subsets and the subset edge-weight vector is compared with the full-sample one. The edge-weight CS-coefficient is the largest drop proportion at which the edge-vector correlation stays >= threshold with probability >= certainty (Epskamp, Borsboom & Fried 2018).

Usage

net_casedrop_reliability(
  data,
  method = "glasso",
  drop_prop = seq(0.1, 0.9, by = 0.1),
  iter = 100L,
  threshold = 0.7,
  certainty = 0.95,
  cor_method = c("spearman", "pearson", "kendall"),
  labels = NULL,
  estimator_args = list(),
  ...
)

Arguments

data

Data frame or matrix (rows = observations), resampled exactly as given (see net_boot()), or a psychnet_group (case-dropped per level).

method

Estimator (see psychnet()). Default "glasso".

drop_prop

Proportions of cases to drop. Default seq(0.1, 0.9, 0.1).

iter

Subsets per proportion. Default 100.

threshold

Minimum acceptable edge-vector correlation. Default 0.7.

certainty

Probability the correlation must exceed threshold. Default 0.95.

cor_method

Correlation method for the edge-vector comparison: "spearman" (default, robust to the wide range of edge weights), "pearson", or "kendall".

labels

Optional node labels.

estimator_args

Named list of estimator arguments. Use this for names consumed by the diagnostic itself, such as estimator threshold.

...

Passed to the estimator.

Value

A tidy data.frame (class psychnet_casedrop), one row per metric per drop proportion, with columns metric, drop_prop, mean, sd. The edge-weight CS-coefficient is carried in attr(x, "cs") and shown when the result is printed. Visualise it with plot.psychnet_casedrop().

References

Epskamp, S., Borsboom, D., & Fried, E. I. (2018). Estimating psychological networks and their accuracy. Behavior Research Methods, 50(1), 195-212.

Examples

# `method`, `iter`, and `drop_prop` are chosen so the example runs quickly;
# the defaults (method = "glasso", iter = 100, drop_prop = seq(0.1, 0.9,
# 0.1)) are what a real reliability assessment should use.
net_casedrop_reliability(SRL_Claude, method = "pcor", iter = 5,
                         drop_prop = c(0.25, 0.5))
#> # edge-weight stability: pcor | CS = 0.50 (spearman cor >= 0.70 at 95%)
#>           metric drop_prop       mean          sd
#> 1   mean_abs_dev      0.25 0.02624533 0.005282477
#> 2   mean_abs_dev      0.50 0.03307525 0.011005165
#> 3 median_abs_dev      0.25 0.02232825 0.006143106
#> 4 median_abs_dev      0.50 0.03082581 0.013881714
#> 5    correlation      0.25 0.97818182 0.010141334
#> 6    correlation      0.50 0.96363636 0.030903149
#> 7    max_abs_dev      0.25 0.05938727 0.016592220
#> 8    max_abs_dev      0.50 0.07034004 0.023615060