In bdabbs13/CIDnetworks: Generative Models for Complex Networks with Conditionally Independent Dyadic Structure

We will be using the CIDnetworks R package to fit single network models. There is another R package that fit latent space models called statnet/latentnet, but we will be using CIDnetworks for several reasons. First, our goodness of fit code works with CIDnetworks objects. Second, CIDnetworks is a bit more flexible than statnet in terms of allowing missing data, a wide variety of tie values, and accommodates random sender/receiver effects.

Loading our network data

load("School1Advice.Rdata")
ls()
## recall network.sch1 is a matrix and network.ties.sch1 is a dataframe
## We could use network.ties.sch1 but we would only want to use the first 3 columns

Fitting Models in CIDnetworks

library(CIDnetworks)
## The first argument is the network; either a sociomatrix or dataframe of tie outcomes is fine
lsm0 = CID.Gibbs(network.sch1, components = c(LSM(2)), draws=300, thin=10)

## Exploring lsm0
class(lsm0)
names(lsm0)

## lsm0$result is a list of 300 draws from the MCMC chain
## after thinning and burnin

summary(lsm0)
LSM.plot(lsm0)

Fitting LSMs with other dimensions

lsm0b=CID.Gibbs(network.sch1, components = c(LSM(3)), draws=300, thin=10)
lsm0c=CID.Gibbs(network.sch1, components = c(LSM(4)), draws=300, thin=10)

Fitting Covariates in LSM

## Note for time, we do not thin these models.  
## When fitting these models in practice, you should thin
## by at least 10 if not 20.
lsm1a=CID.Gibbs(network.sch1, components = c(LSM(2), SenderCOV(teachers), ReceiverCOV(teachers)), draws=300)
summary(lsm1a)
COV.plot(lsm1a)
LSM.plot(lsm1a)

Here are some more to try

lsm1b=CID.Gibbs(network.sch1, components = c(LSM(2), SenderCOV(teachers), ReceiverCOV(teachers), EdgeCOV(edgevars)), draws=300)
summary(lsm1b)
COV.plot(lsm1b)
LSM.plot(lsm1b)


lsm1c=CID.Gibbs(network.sch1, components = c(LSM(2),  EdgeCOV(edgevars)), draws=300)
summary(lsm1c)
COV.plot(lsm1c)
LSM.plot(lsm1c)

lsm1d=CID.Gibbs(network.sch1, components = c(LSM(2),  EdgeCOV(edgevars[,3])), draws=300)
summary(lsm1d)
COV.plot(lsm1d)
LSM.plot(lsm1d)

On Your Own - For Beginners

There is another dataset that you can use called School2Advice.Rdata. Fit the following LSM models.

1. A LSM with 2 dimensions
2. A LSM with 3 dimensions
3. A LSM with 2 dimensions and sender covariates
4. A LSM with 2 dimensions and receiver covariates
5. A LSM with 2 dimension and edge covariates

## Solutions
load("School2Advice.Rdata")
ls()

head(teachers2)
head(edgevars2)
head(sch2net)

## Plotting the Network
par(mar=c(0,0,0,0))
plot.igraph(graph.adjacency(sch2net), layout = layout.fruchterman.reingold, vertex.color = "gold", vertex.label=NA, vertex.size =5, edge.arrow.size=0.25, edge.arrow.width=0.25, edge.width=0.3, vertex.frame.color=NA,
frame = TRUE)

##
fit1=CID.Gibbs(sch2net, components = c(LSM(2)), draws=300)
fit2=CID.Gibbs(sch2net, components = c(LSM(3)), draws=300)
fit3=CID.Gibbs(sch2net, components = c(LSM(2), SenderCOV(teachers2)), draws=300)
fit4=CID.Gibbs(sch2net, components = c(LSM(2), ReceiverCOV(teachers2)), draws=300)
fit5=CID.Gibbs(sch2net, components = c(LSM(2), EdgeCOV(edgevars2)), draws=300)

On Your Own - For Advanced Users

Here is another data set to play with on your own. These are friendship network data from van de Bunt (1999). These come from a longitudinal study of university students. We will be examining data from the 5th time point and looking specifically at network 1, "real friends", see \url{http://www.stats.ox.ac.uk/~snijders/siena/vdBunt_data.htm}

Information about the students was also collected; gender (1=female), program (2-, 3-, or 4- year), and smoking (1=yes).

The R code is below, but try fitting one, two, or all of the following models with 300 samples:

1. A LSM with 2 dimensions
2. A LSM with 2 dimensions and a sender covariate for the smoking
3. A LSM with 2 dimensions and an identical covariate for all three variables
4. A LSM with 3 dimensions and an identical covariate for smoking
5. A LSM with 3 dimensions and a sender covariate for gender and an identical covariate for smoking

The R Code below has some exploratory network stuff along with the 5 models specified above. Scroll down to see the solutions.

library(igraph)
library(CIDnetworks)
load("OnYourOwn.Rdata")

## let's examine the friendship network
# Note the network and node attributes are both matrices
class(friends.net.5)
class(features)

head(friends.net.5)
head(features)

## One thing to note is that this network has
## MISSING data!
## This is actually one cool think about the
## CIDnetworks package - it handles missing data

## Some random useful code
## To change nodes to a specific color quickly
smokecolor = as.character(features[, "Smoking"]) 
smokecolor = gsub("0", "red", smokecolor) 
smokecolor = gsub("1", "blue", smokecolor)

par(mar=c(0,0,2,0))
plot.igraph(graph.adjacency(friends.net.5), layout = layout.fruchterman.reingold, vertex.color = smokecolor, vertex.label=NA, vertex.size =5, edge.arrow.size=0.25, edge.arrow.width=0.25, edge.width=0.3, vertex.frame.color=NA, main = "Smoking (Red =Yes, Blue=No)",
frame = TRUE)

## Solutions
fit1=CID.Gibbs(friends.net.5, components = c(LSM(2)), draws=300)

fit2=CID.Gibbs(friends.net.5, components = c(LSM(2), SenderCOV(features[,3])), draws=300)

fit3= CID.Gibbs(friends.net.5, components = c(LSM(2), IdenticalCOV(features)), draws=300)

fit4= CID.Gibbs(friends.net.5, components = c(LSM(3), IdenticalCOV(features[,3])), draws=300)

fit5=CID.Gibbs(friends.net.5, components = c(LSM(3), SenderCOV(features[,1]), IdenticalCOV(features[,3])), draws=300)

bdabbs13/CIDnetworks documentation built on Nov. 15, 2019, 2:41 a.m.