50% off: Unlimited data and AI learning.
State of Data and AI Literacy Report 2025

nethet (version 1.4.2)

diffnet_multisplit: Differential Network

Description

Differential Network

Usage

diffnet_multisplit(x1, x2, b.splits = 50, frac.split = 1/2,
  screen.meth = "screen_bic.glasso", include.mean = FALSE,
  gamma.min = 0.05, compute.evals = "est2.my.ev3",
  algorithm.mleggm = "glasso_rho0", method.compquadform = "imhof",
  acc = 1e-04, epsabs = 1e-10, epsrel = 1e-10, show.warn = FALSE,
  save.mle = FALSE, verbose = TRUE, mc.flag = FALSE, mc.set.seed = TRUE,
  mc.preschedule = TRUE, mc.cores = getOption("mc.cores", 2L), ...)

Arguments

x1
Data-matrix sample 1. You might need to center and scale your data-matrix.
x2
Data-matrix sample 1. You might need to center and scale your data-matrix.
b.splits
Number of splits (default=50).
frac.split
Fraction train-data (screening) / test-data (cleaning) (default=0.5).
screen.meth
Screening procedure. Options: 'screen_bic.glasso' (default), 'screen_cv.glasso', 'screen_shrink' (not recommended), 'screen_mb'.
include.mean
Should sample specific means be included in hypothesis? Use include.mean=FALSE (default and recommended) which assumes mu1=mu2=0 and tests the hypothesis H0: Omega_1=Omega_2.
gamma.min
Tuning parameter in p-value aggregation of Meinshausen et al (2009). (Default=0.05).
compute.evals
Method to estimate the weights in the weighted-sum-of-chi2s distribution. The default and (currently) the only available option is the method 'est2.my.ev3'.
algorithm.mleggm
Algorithm to compute MLE of GGM. The algorithm 'glasso_rho' is the default and (currently) the only available option.
method.compquadform
Method to compute distribution function of weighted-sum-of-chi2s (default='imhof').
acc
See ?davies (default 1e-04).
epsabs
See ?imhof (default 1e-10).
epsrel
See ?imhof (default 1e-10).
show.warn
Should warnings be showed (default=FALSE)?
save.mle
If TRUE, MLEs (inverse covariance matrices for samples 1 and 2) are saved for all b.splits. The median aggregated inverse covariance matrix is provided in the output as 'medwi'. The default is save.mle=FALSE.
verbose
If TRUE, show output progress.
mc.flag
If TRUE use parallel execution for each b.splits via function mclapply of package parallel.
mc.set.seed
See mclapply. Default=TRUE
mc.preschedule
See mclapply. Default=TRUE
mc.cores
Number of cores to use in parallel execution. Defaults to mc.cores option if set, or 2 otherwise.
...
Additional arguments for screen.meth.

Value

  • list consisting of
  • ms.pvalp-values for all b.splits
  • ss.pvalsingle-split p-value
  • medagg.pvalmedian aggregated p-value
  • meinshagg.pvalmeinshausen aggregated p-value (meinshausen et al 2009)
  • teststattest statistics for b.splits
  • weights.nulldistrestimated weights
  • active.lastactive-sets obtained in last screening-step
  • medwimedian of inverse covariance matrices over b.splits
  • sig.lastconstrained mle (covariance matrix) obtained in last cleaning-step
  • wi.lastconstrained mle (inverse covariance matrix) obtained in last cleaning-step

Details

Remark:

* If include.mean=FALSE, then x1 and x2 have mean zero and DiffNet tests the hypothesis H0: Omega_1=Omega_2. You might need to center x1 and x2. * If include.mean=TRUE, then DiffNet tests the hypothesis H0: mu_1=mu_2 & Omega_1=Omega_2 * However, we recommend to set include.mean=FALSE and to test equality of the means separately. * You might also want to scale x1 and x2, if you are only interested in differences due to (partial) correlations.

Examples

Run this code
############################################################
##This example illustrates the use of Differential Network##
############################################################


##set seed
set.seed(1)

##sample size and number of nodes
n <- 40
p <- 10

##specifiy sparse inverse covariance matrices
gen.net <- generate_2networks(p,graph='random',n.nz=rep(p,2),
                              n.nz.common=ceiling(p*0.8))
invcov1 <- gen.net[[1]]
invcov2 <- gen.net[[2]]
plot_2networks(invcov1,invcov2,label.pos=0,label.cex=0.7)

##get corresponding correlation matrices
cor1 <- cov2cor(solve(invcov1))
cor2 <- cov2cor(solve(invcov2))

##generate data under null hypothesis (both datasets have the same underlying
## network)
library('mvtnorm')
x1 <- rmvnorm(n,mean = rep(0,p), sigma = cor1)
x2 <- rmvnorm(n,mean = rep(0,p), sigma = cor1)

##run diffnet (under null hypothesis)
dn.null <- diffnet_multisplit(x1,x2,b.splits=1,verbose=FALSE)
dn.null$ss.pval#single-split p-value

##generate data under alternative hypothesis (datasets have different networks)
x1 <- rmvnorm(n,mean = rep(0,p), sigma = cor1)
x2 <- rmvnorm(n,mean = rep(0,p), sigma = cor2)

##run diffnet (under alternative hypothesis)
dn.altn <- diffnet_multisplit(x1,x2,b.splits=1,verbose=FALSE)
dn.altn$ss.pval#single-split p-value
dn.altn$medagg.pval#median aggregated p-value

##typically we would choose a larger number of splits
# dn.altn <- diffnet_multisplit(x1,x2,b.splits=10,verbose=FALSE)
# dn.altn$ms.pval#multi-split p-values
# dn.altn$medagg.pval#median aggregated p-value
# plot(dn.altn)#histogram of single-split p-values

Run the code above in your browser using DataLab