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).

This first example compares the correlation between edges

```
library(GGMncv)
set.seed(1)
<- gen_net(p = 10)
main
<- MASS::mvrnorm(n = 500,
y1 mu = rep(0, 10),
Sigma = main$cors)
<- MASS::mvrnorm(n = 500,
y2 mu = rep(0, 10),
Sigma = main$cors)
<- function(x, y){
Correlation cor(x[upper.tri(x)], y[upper.tri(x)])
}
<- nct(y1, y2,
compare_ggms 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
```

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

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

Next use `nct`

```
<- nct(y1, y2,
compare_ggms 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)`

.

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.