tnam: Fit (temporal) network autocorrelation models

In leifeld/tnam: Temporal Network Autocorrelation Models (TNAM)

Fit (temporal) network autocorrelation models

Description

Fit (temporal) network autocorrelation models.

Usage

tnam(formula, family = gaussian, re.node = FALSE, 
    re.time = FALSE, time.linear = FALSE, time.quadratic = FALSE, 
    center.y = FALSE, na.action = na.omit, ...)

tnamdata(formula, center.y = FALSE)

Arguments

formula	A formula where the left-hand side specifies either a vector containing the outcome variable (for a cross-sectional model) or a list of such vectors (for modeling the outcome at multiple time steps) or a data frame with one time step per column (also for longitudinal models of behavior). The right-hand side of the formula consists of tnam-specific model terms like netlag, structsim and other terms which are described on the help page of tnam-terms.
family	The link function for fitting the generalized linear model or the mixed effects model, for example gaussian or binomial. The options are the same as in the glm and glmer functions. For details on the family argument, see the family help page.
re.node	If multiple time steps are present: should a random effect for the nodes be added to the model? This results in the estimation of a mixed effects model.
re.time	If multiple time steps are present: should a random effect for the time steps be added to the model? This results in the estimation of a mixed effects model.
time.linear	If multiple time steps are present: should a linear effect for time be added to the model? This can be estimated in the standard GLM framework.
time.quadratic	If multiple time steps are present: should a squared effect for time be added to the model? This can be estimated in the standard GLM framework.
center.y	Center the dependent variable by subtracting the mean from the actual value within each time step?
na.action	How should missing values be treated? By default, they are omitted. See the na.omit help page for details.
...	Further arguments that should be passed to the glm, lmer, or glmer function, which is used under the hood for estimating the model.

Details

The tnam function serves to estimate temporal or cross-sectional network autocorrelation models. Model terms such as spatial lags, temporal lags, spatio-temporal lags, centrality etc. can be specified in the formula argument. Details on the model terms can be found on the tnam-terms help page.

The tnamdata function accepts a formula (like in the tnam function) and returns a data frame with the response variable and the covariates for estimation with any estimation function. tnam first calls tnamdata internally and then hands over the resulting data structure to a glm, lmer, or nlmer call. If models such as tobit, multinomial or multilevel models should be estimated, one can leave out the estimation step and feed the results of tnamdata manually into any type of model.

See Also

tnam-package tnam-terms knecht

Examples

# The following example models delinquency among adolescents at 
# multiple time steps as a function of (1) their nodal attributes 
# like sex or religion, (2) their peers' delinquency levels, (3) 
# their own and their peers' past delinquency behavior, and (4) 
# their structural position in the network. See ?knecht for 
# details on the dataset. Before estimating the model, all data 
# should be labeled with the names of the nodes such that tnam 
# is able to merge the information on multiple nodes across time 
# points.

library("tnam")
data("knecht")

# prepare the dependent variable y
delinquency <- as.data.frame(delinquency)
rownames(delinquency) <- letters

# replace structural zeros (denoted as 10) and add row labels
friendship[[3]][friendship[[3]] == 10] <- NA
friendship[[4]][friendship[[4]] == 10] <- NA
for (i in 1:length(friendship)) {
  rownames(friendship[[i]]) <- letters
}

# prepare the covariates sex and religion
sex <- demographics$sex
names(sex) <- letters
sex <- list(t1 = sex, t2 = sex, t3 = sex, t4 = sex)
religion <- demographics$religion
names(religion) <- letters
religion <- list(t1 = religion, t2 = religion, t3 = religion, 
    t4 = religion)

# Estimate the model. The first term is the sex of the respondent, 
# the second term is the religion of the respondent, the third 
# term is the previous delinquency behavior of the respondent, 
# the fourth term is the delinquency behavior of direct friends, 
# the fifth term is the delinquency behavior of indirect friends 
# at a path distance of 2, the sixth effect is the past delinquency 
# of direct friends, the seventh term indicates whether the 
# respondent has any contacts at all, and the last term captures 
# the effect of the betweenness centrality of the respondent on 
# his or her behavior. Apparently, previous behavior, being an 
# isolate, and religion seem to have an effect on delinquency in 
# this dataset. There is also a slight positive trend over time, 
# and direct friends exert a minor effect (not significant).
# Note that a linear model may not be the best specification for 
# modeling the ordered categorical delinquency variable, but it 
# suffice here for illustration purposes.

model1 <- tnam(
    delinquency ~ 
    covariate(sex, coefname = "sex") + 
    covariate(religion, coefname = "religion") + 
    covariate(delinquency, lag = 1, exponent = 1) + 
    netlag(delinquency, friendship) + 
    netlag(delinquency, friendship, pathdist = 2, decay = 1) + 
    netlag(delinquency, friendship, lag = 1) + 
    degreedummy(friendship, deg = 0, reverse = TRUE) + 
    centrality(friendship, type = "betweenness"), 
    re.node = TRUE, time.linear = TRUE
)
summary(model1)

# for nice table output, use the texreg package
library("texreg")
screenreg(model1)

leifeld/tnam documentation built on July 7, 2023, 7:54 a.m.