Estimates statistical network models using the exponential
random graph modeling (ERGM) framework with extensions for
dynamic/temporal models (STERGM). This is typically the first
step in the network modeling pipeline, followed by netdx
for model diagnostics and netsim for epidemic simulation.
netest(
nw,
formation,
target.stats,
coef.diss,
constraints = NULL,
coef.form = NULL,
edapprox = TRUE,
set.control.ergm = control.ergm(),
set.control.tergm = control.tergm(MCMC.maxchanges = .Machine$integer.max),
set.control.ergm.ego = NULL,
verbose = FALSE,
nested.edapprox = TRUE,
ergm.ego.popsize = 0,
...
)A fitted network model object of class netest. This object is
passed to netdx for diagnostics and to netsim for epidemic
simulation. Use print() to view the model form, including the
formation formula, target statistics, and dissolution model. Use
summary() to view the estimated model coefficients and goodness-of-fit
statistics from the underlying ERGM or STERGM fit.
An object of class network or egor, with the latter
indicating an ergm.ego fit.
Right-hand sided STERGM formation formula in the form
~edges + ..., where ... are additional network statistics
(ERGM terms). This formula specifies which structural features of
the network should be reproduced by the model. Common terms include
edges (overall connectivity), nodematch (homophily by attribute),
concurrent (overlapping partnerships), and degree (degree
distribution constraints). See ergm::ergm-terms for the full
list of available terms.
Vector of target statistics for the formation model, with
one number for each network statistic in the model. These are the
observed (or desired) values for each term in the formation formula.
For example, if formation = ~edges + concurrent, then
target.stats = c(175, 40) means the model should produce
approximately 175 edges and 40 nodes with 2 or more partners. For
an edges-only model, a useful starting value is
mean_degree * network_size / 2. Ignored if fitting via ergm.ego.
An object of class disscoef output from the
dissolution_coefs function. This encodes the average partnership
duration(s) and the corresponding dissolution model coefficients.
For models with vital dynamics (arrivals and departures), the
d.rate argument in dissolution_coefs should be set to adjust
for the competing risk of node departure.
Right-hand sided formula specifying constraints for the
modeled network, in the form ~..., where ... are
constraint terms. By default, no constraints are set.
Vector of coefficients for the offset terms in the formation formula.
If TRUE, use the indirect edges dissolution
approximation method for the dynamic model fit, otherwise use the
more time-intensive full STERGM estimation (see details). The
approximation is recommended for most use cases, especially when
average partnership durations are moderate to long (> 25 time steps).
Direct STERGM estimation (edapprox = FALSE) is slower but may be
preferred for very short durations or when inferential quantities
(standard errors, p-values) on formation coefficients are needed.
For nw of class egor, only edapprox = TRUE is supported.
Control arguments passed to ergm (see details).
Control arguments passed to tergm (see details).
Control arguments passed to ergm.ego (see
details).
If TRUE, print model fitting progress to console.
Logical, only used when edapprox = TRUE. When
TRUE (the default), the dissolution model must be an initial
segment of the formation model; when FALSE, the dissolution terms
need not be nested (see details).
Numeric. The popsize argument passed to
ergm.ego::ergm.ego() when nw is an egor. Shifts the edges
coefficient to reflect a target population size. The default of
0 scales the edges term to the size of the egor sample,
matching the coefficient an equivalent ergm fit would produce.
Set to 1 for per-capita scaling, which lets the fitted model
be applied to a network of arbitrary size without knowing the
population size assumed during estimation. See ?ergm.ego for
other accepted values.
Additional arguments passed to other functions.
The edges dissolution approximation method is described in Carnegie et al. This approximation requires that the dissolution coefficients are known, that the formation model is being fit to cross-sectional data conditional on those dissolution coefficients, and that the terms in the dissolution model are a subset of those in the formation model. Under certain additional conditions, the formation coefficients of a STERGM model are approximately equal to the coefficients of that same model fit to the observed cross-sectional data as an ERGM, minus the corresponding coefficients in the dissolution model. The approximation thus estimates this ERGM (which is typically much faster than estimating a STERGM) and subtracts the dissolution coefficients.
The conditions under which this approximation best hold are when there are
few relational changes from one time step to another; i.e. when either
average relational durations are long, or density is low, or both.
Conveniently, these are the same conditions under which STERGM estimation is
slowest. Note that the same approximation is also used to obtain starting
values for the STERGM estimate when the latter is being conducted. The
estimation does not allow for calculation of standard errors, p-values, or
likelihood for the formation model; thus, this approach is of most use when
the main goal of estimation is to drive dynamic network simulations rather
than to conduct inference on the formation model. The user is strongly
encouraged to examine the behavior of the resulting simulations to confirm
that the approximation is adequate for their purposes. For an example, see
the vignette for the package tergm.
It has recently been found that subtracting a modified version of the
dissolution coefficients from the formation coefficients provides a more
principled approximation, and this is now the form of the approximation
applied by netest for both settings of nested.edapprox. The modified
values subtracted from the formation coefficients are equivalent to the
(crude) dissolution coefficients with their target durations increased by 1.
The nested.edapprox argument does not change this approximation; it only
controls how it is implemented. When TRUE (the default), the dissolution
terms are an initial segment of the formation model, so the correction
values are subtracted in place from the corresponding initial formation
model coefficients. When FALSE, the dissolution terms need not be nested
in the formation model: they are appended to the formation model and the
corresponding correction values are appended to the vector of formation
model coefficients.
The nesting requirement is a property of the approximation, so it is only
checked when edapprox = TRUE. A direct STERGM fit (edapprox = FALSE)
estimates formation and persistence jointly and places no such requirement on
the dissolution terms. Note, however, that tergm derives the starting
values for its EGMME fit from the same approximation, so a direct fit whose
persistence terms have no formation analog needs starting values supplied
through set.control.tergm, with control.tergm(init = ...) or
control.tergm(init.method = "zeros").
The ergm, ergm.ego, and tergm functions allow control settings for the
model fitting process. When fitting a STERGM directly (setting
edapprox to FALSE), control parameters may be passed to the
tergm function with the set.control.tergm argument in netest.
The controls should be input through the control.tergm() function,
with the available parameters listed in the tergm::control.tergm help
page in the tergm package.
When fitting a STERGM indirectly (setting edapprox to TRUE), control
settings may be passed to the ergm function using set.control.ergm,
or to the ergm.ego function using set.control.ergm.ego. The controls
should be input through the control.ergm() and control.ergm.ego()
functions, respectively, with the available parameters listed in the
ergm::control.ergm help page in the ergm package and the
ergm.ego::control.ergm.ego help page in the ergm.ego
package. An example is below.
The network modeling pipeline in EpiModel typically follows these steps:
Initialize a network: network_initialize
Specify formation and dissolution: a formation formula (e.g.,
~edges + concurrent) with target statistics, and dissolution
coefficients via dissolution_coefs
Estimate the network model: netest()
Diagnose model fit: netdx
Simulate the epidemic: netsim with param.net, init.net,
and control.net
netest is a wrapper function for the ergm, ergm.ego, and tergm
functions that estimate static and dynamic network models. Network model
estimation is the first step in simulating a stochastic network epidemic model
in EpiModel. The output from netest is a necessary input for running the
epidemic simulations in netsim. With a fitted network model, one should
always first proceed to model diagnostics, available through the netdx
function, to check model fit. A detailed description of fitting these
models, along with examples, may be found in the
Network Modeling for Epidemics
tutorials.
Krivitsky PN, Handcock MS. "A Separable Model for Dynamic Networks." JRSS(B). 2014; 76.1: 29-46.
Carnegie NB, Krivitsky PN, Hunter DR, Goodreau SM. An Approximation Method for Improving Dynamic Network Model Fitting. Journal of Computational and Graphical Statistics. 2014; 24(2): 502-519.
Jenness SM, Goodreau SM and Morris M. EpiModel: An R Package for Mathematical Modeling of Infectious Disease over Networks. Journal of Statistical Software. 2018; 84(8): 1-47.
Use dissolution_coefs to compute dissolution coefficients
before estimation. Use netdx to diagnose the fitted network
model, and netsim to simulate epidemic spread over a simulated
dynamic network consistent with the model fit. Parameterize the
epidemic simulation with param.net, init.net, and
control.net.
# Initialize a network of 100 nodes
nw <- network_initialize(n = 100)
# Set formation formula
formation <- ~edges + concurrent
# Set target statistics for formation
target.stats <- c(50, 25)
# Obtain the offset coefficients
coef.diss <- dissolution_coefs(dissolution = ~offset(edges), duration = 10)
# Estimate the STERGM using the edges dissolution approximation
est <- netest(nw, formation, target.stats, coef.diss,
set.control.ergm = control.ergm(MCMC.burnin = 1e5,
MCMC.interval = 1000))
# View the model form (formation, targets, dissolution)
est
# View the estimated coefficients
summary(est)
if (FALSE) {
# Model with homophily on a nodal attribute
nw2 <- network_initialize(n = 500)
nw2 <- set_vertex_attribute(nw2, "risk", rep(0:1, each = 250))
formation2 <- ~edges + nodematch("risk")
target.stats2 <- c(175, 110)
coef.diss2 <- dissolution_coefs(dissolution = ~offset(edges), duration = 50)
est2 <- netest(nw2, formation2, target.stats2, coef.diss2)
est2
# Direct STERGM estimation (slower, for short durations or inference)
est3 <- netest(nw, formation, target.stats, coef.diss, edapprox = FALSE)
}
Run the code above in your browser using DataLab