The hardware and bandwidth for this mirror is donated by METANET, the Webhosting and Full Service-Cloud Provider.
If you wish to report a bug, or if you are interested in having us mirror your free-software or open-source project, please feel free to contact us at mirror[@]metanet.ch.

Custom Network Comparison Tests

Introdution

A key feature of nct is the ability to test anything hypothesis, as opposed to the set of defaults provided in the R package NetworkComparisonTest (Van Borkulo et al. 2017).

Correlation

This first example compares the correlation between edges

library(GGMncv)

set.seed(1)

main <- gen_net(p = 10)

y1 <- MASS::mvrnorm(n = 500, 
                    mu = rep(0, 10), 
                    Sigma = main$cors)

y2 <- MASS::mvrnorm(n = 500, 
                    mu = rep(0, 10), 
                    Sigma = main$cors)

Correlation <- function(x, y){
  cor(x[upper.tri(x)], y[upper.tri(x)])
}

compare_ggms <- nct(y1, y2, 
                    FUN = Correlation, 
                    progress = FALSE)

compare_ggms
#> Network Comparsion Test
#> (GGMncv Edition)
#> ----
#> Maximum Difference
#> p-value: 0.756 
#> ----
#> Global Strength
#> p-value: 0.904 
#> ----
#> Sum of Squared Error
#> p-value: 0.938 
#> ----
#> Jensen-Shannon divergence
#> p-value: 0.929 
#> ----
#> note: compute p-values manually for custom tests. see vignettes.

By default, the \(p\)-value corresponds to the proportion of the null distribution greater than the observed. For some test statistics, this might not make sense so they are not provided.

A fail safe is to visualize the results and compute the \(p\)-value manually.

hist(compare_ggms$Correlation_perm, 
     main = "null dist: correlation")
abline(v = compare_ggms$Correlation_obs)

Because a higher correlation is a good thing, 1 minus the proportion of the null distribution greater than the observed make sense.

1 - mean(compare_ggms$Correlation_perm > compare_ggms$Correlation_obs)
#> [1] 0.918

Predictability

Next is predictability, or \(R^2\), for the first node in the networks. The first step is defining a function.

# define function
r2 <- function(x, y){
  diag(x) <- 1
  diag(y) <- 1
  
  # network 1
  inv1 <- solve(corpcor::pcor2cor(x))
  beta1 <- -(inv1[1,-1] / inv1[1,1])
  r21 <- cor(y1[,1], y1[,-1] %*% beta1)^2
  
  # network 2
  inv2 <- solve(corpcor::pcor2cor(y))
  beta2 <- -(inv2[1,-1] / inv2[1,1])
  r22 <- cor(y2[,1], y2[,-1] %*% beta2)^2
  
  return(as.numeric(r21 - r22))
}

Next use nct

compare_ggms <- nct(y1, y2, 
                    progress = FALSE,
                    FUN = r2)

hist(compare_ggms$r2_perm,  
     main = "null dist: R2 Difference")
abline(v = compare_ggms$r2_obs)

Note anything can be tested, say, expected influence. As such, this is not limited to only one value returned by the custom function (as in the above examples), but can be any number. There is an example in the documentation.

Further, although not provided here, both tests could have been done at the same time by passing a list to FUN, i.e., list_fun <- list(r2 = r2, Correlation = Correlation).

References

Van Borkulo, Claudia D, Lynn Boschloo, J Kossakowski, Pia Tio, Robert A Schoevers, Denny Borsboom, and Lourens J Waldorp. 2017. “Comparing Network Structures on Three Aspects: A Permutation Test.” Manuscript Submitted for Publication 10.

These binaries (installable software) and packages are in development.
They may not be fully stable and should be used with caution. We make no claims about them.