SP.Inf {NetworkReg}

Fitting Linear Regression Models on Network-Linked Data

Description

SP.Inf is used to the regression model on network-linked data by subspace project and produce the inference result.

Usage

SP.Inf(X, Y, A, K, r = NULL, sigma2 = NULL, thr = NULL, alpha.CI = 0.05,
boot.thr = TRUE, boot.n = 50)

Arguments

X

the covariate matrix where each row is an observation and each column is a covariate. If an intercept is to be included in the model, the column of ones should be in the matrix.

Y

the column vector of response.

A

the network information. The most natural choice is the adjacency matrix of the network. However, if the network is assumed to be noisy and a better estimate of the structural connection strength, it can also be used. This corresponds to the Phat matrix in the original paper. A Laplacian matrix can also be used, but it should be flipped. See 'Details'.

K

the dimension of the network eigenspace for network effect.

r

the covariate-network cofounding space dimension. This is typically unknown and can be unspecified by using the default value 'NULL'. If so, the user should provide a threshold or resort to a tuning procedure by either the theoretical rule or a bootstrapping method, as described in the paper.

sigma2

the variance of random noise. Typically unknown.

thr

threshold for r estimation. If r is unspecified, we will use the thereshold to select r. If this is also 'NULL', aa theoretical threshold or a bootsrapping method can be evoked to estimate it.

alpha.CI

the 1-alpha.CI confidence level will be produced for the parameters.

boot.thr

logical. Only effective if both r and thr are NULLs. If FALSE, the theoretical threshold will be used to select r. Otherwise, the bootstrapping procedure will be used to find the threshold.

boot.n

the number of bootstrapping samples used when boot.thr is TRUE.

Details

The model fitting procedure is following the paper exactly, so please check the procedure and theory in the paper. If the Laplacian matrix L=D-A is the network quantity to use, notice that typically we treat the smallest values and their corresponding eigenvectors as network cohesive space. Therefore, one should consider flip the Laplacian matrix by using cI - L as the value for A, where c is sufficiently large to ensure PSD of cI-L.

Value

A list object with

beta

estimate of beta, the covariate effects

alpha

individual effects

theta

coefficients of confounding effects with respect to the covariates

r

confounding dimension

sigma

estimated random noise variance

cov.hat

covariance matrix of beta

coef.mat

beta and the confidence intervals according to alpha.CI and the p-values of the significance test

fitted

fitted value of response

chisq.val

the value of the chi-square statistic for the significance test for network effect

chisq.p

the p-value of the significance test for network effect

Author(s)

Can M. Le and Tianxi Li.

Maintainer: Tianxi Li <[e-mail]>

References

Le, C. M., & Li, T. (2022). Linear regression and its inference on noisy network-linked data. Journal of the Royal Statistical Society Series B: Statistical Methodology, 84(5), 1851-1885.

Examples

library(randnet)
library(RSpectra)
### data generating procedure in Section 5.3 of the paper

n <- 1000
big.model <- BlockModel.Gen(lambda=n^(1/2),n=n,beta=0.2,K=4)
P <- big.model$P
big.X <- cbind(rnorm(n),runif(n),rexp(n))

eigen.P <- eigs_sym(A=P,k=4)
X.true <- big.X
X.true <- scale(X.true,center=TRUE,scale=TRUE)*sqrt(n/(n-1))
X.true <- cbind(sqrt(n)*eigen.P$vectors[,1],X.true)
X.svd <- svd(X.true)
x.proj <- X.svd$v%*%(t(X.svd$u)/X.svd$d)
Theta <- X.svd$v%*%(t(X.svd$v)/(X.svd$d^2))*n
R <- X.svd$u
U <- eigen.P$vectors[,1:4]
true.SVD <- svd(t(R)%*%U,nu=4,nv=4)
V <- true.SVD$v
r <- 1
U.tilde <- U%*%V
R.tilde <- R%*%true.SVD$u
theta.tilde <- matrix(c(sqrt(n),0,0,0),ncol=1)
beta.tilde <- matrix(sqrt(n)*c(0,1,1,1),ncol=1)
Xtheta <- R.tilde%*%theta.tilde
Xbeta <- R.tilde%*%beta.tilde

theta <- solve(t(X.true)%*%X.true,t(X.true)%*%Xtheta)
beta <- solve(t(X.true)%*%X.true,t(X.true)%*%Xbeta)
alpha.coef <- matrix(sqrt(n)*c(0,1,1,1),ncol=1)
alpha <- U.tilde%*%alpha.coef

EY <- Xtheta+Xbeta + alpha


#### model fitting


A <- net.gen.from.P(P)
Khat <- BHMC.estimate(A, K.max = 15)$K  ### estimate K to use

## model fitting
Y <- EY + rnorm(n)
fit <- SP.Inf(X.true,Y,A,K=Khat,alpha=0.05,boot.thr=FALSE)
### In general, boot.thr = T works better for small sample but is slower.
### It was used in the paper.
fit$coef.mat  
### notice that beta1 inference is meaningful here. Check the paper.
beta
fit$chisq.p

## find a parametric estimation of the network. This is generally not available.
rsc <- reg.SP(A,K=Khat,tau=0.1)
est <- SBM.estimate(A,rsc$cluster)
Phat <- est$Phat
fit2 <- SP.Inf(X.true,Y,Phat,K=Khat,alpha=0.05,boot.thr=FALSE)
fit2$coef.mat  
### notice that beta1 inference is meaningful here. Check the paper.