Auxiliary function as user interface for fine-tuning 'stergm' fitting.
control.stergm(init.form = NULL, init.diss = NULL,
init.method = NULL, force.main = FALSE,
MCMC.prop.weights.form = "default", MCMC.prop.args.form = NULL,
MCMC.prop.weights.diss = "default", MCMC.prop.args.diss = NULL,
MCMC.init.maxedges = 20000, MCMC.init.maxchanges = 20000,
MCMC.packagenames = c(), CMLE.MCMC.burnin = 1024 * 16,
CMLE.MCMC.interval = 1024, CMLE.control = NULL,
CMLE.control.form = control.ergm(init = init.form, MCMC.burnin =
CMLE.MCMC.burnin, MCMC.interval = CMLE.MCMC.interval, MCMC.prop.weights =
MCMC.prop.weights.form, MCMC.prop.args = MCMC.prop.args.form,
MCMC.init.maxedges = MCMC.init.maxedges, MCMC.packagenames =
MCMC.packagenames, parallel = parallel, parallel.type = parallel.type,
parallel.version.check = parallel.version.check, force.main =
force.main), CMLE.control.diss = control.ergm(init = init.diss,
MCMC.burnin = CMLE.MCMC.burnin, MCMC.interval = CMLE.MCMC.interval,
MCMC.prop.weights = MCMC.prop.weights.diss, MCMC.prop.args =
MCMC.prop.args.diss, MCMC.init.maxedges = MCMC.init.maxedges,
MCMC.packagenames = MCMC.packagenames, parallel = parallel, parallel.type
= parallel.type, parallel.version.check = parallel.version.check,
force.main = force.main), CMLE.NA.impute = c(),
CMLE.term.check.override = FALSE,
EGMME.main.method = c("Gradient-Descent"),
EGMME.MCMC.burnin.min = 1000, EGMME.MCMC.burnin.max = 1e+05,
EGMME.MCMC.burnin.pval = 0.5, EGMME.MCMC.burnin.add = 1,
MCMC.burnin = NULL, MCMC.burnin.mul = NULL, SAN.maxit = 4,
SAN.nsteps.times = 8, SAN.control = control.san(term.options =
term.options, SAN.maxit = SAN.maxit, SAN.prop.weights =
MCMC.prop.weights.form, SAN.prop.args = MCMC.prop.args.form,
SAN.init.maxedges = MCMC.init.maxedges, SAN.max.maxedges = Inf,
SAN.nsteps = round(sqrt(EGMME.MCMC.burnin.min * EGMME.MCMC.burnin.max)) *
SAN.nsteps.times, SAN.packagenames = MCMC.packagenames, parallel =
parallel, parallel.type = parallel.type, parallel.version.check =
parallel.version.check), SA.restarts = 10, SA.burnin = 1000,
SA.plot.progress = FALSE, SA.max.plot.points = 400,
SA.plot.stats = FALSE, SA.init.gain = 0.1, SA.gain.decay = 0.5,
SA.runlength = 25, SA.interval.mul = 2, SA.init.interval = 500,
SA.min.interval = 20, SA.max.interval = 500, SA.phase1.minruns = 4,
SA.phase1.tries = 20, SA.phase1.jitter = 0.1,
SA.phase1.max.q = 0.1, SA.phase1.backoff.rat = 1.05,
SA.phase2.levels.max = 40, SA.phase2.levels.min = 4,
SA.phase2.max.mc.se = 0.001, SA.phase2.repeats = 400,
SA.stepdown.maxn = 200, SA.stepdown.p = 0.05, SA.stop.p = 0.1,
SA.stepdown.ct = 5, SA.phase2.backoff.rat = 1.1, SA.keep.oh = 0.5,
SA.keep.min.runs = 8, SA.keep.min = 0, SA.phase2.jitter.mul = 0.2,
SA.phase2.maxreljump = 4, SA.guard.mul = 4, SA.par.eff.pow = 1,
SA.robust = FALSE, SA.oh.memory = 1e+05, SA.refine = c("mean",
"linear", "none"), SA.se = TRUE, SA.phase3.samplesize.runs = 10,
SA.restart.on.err = TRUE, term.options = NULL, seed = NULL,
parallel = 0, parallel.type = NULL, parallel.version.check = TRUE)
numeric or NA
vector equal in
length to the number of parameters in the formation/dissolution
model or NULL
(the default); the initial values for the
estimation and coefficient offset terms. If NULL
is
passed, all of the initial values are computed using the method
specified by control$init.method
.
If a numeric vector is given, the elements of the vector are
interpreted as follows:
Elements corresponding to
terms enclosed in offset()
are used as the fixed offset
coefficients. These should match the offset values given in
offset.coef.form
.
Elements that do not correspond to offset terms and are not
NA
are used as starting values in the estimation.
Initial values for the elements that are NA
are fit
using the method specified by
control$init.method
.
Passing control.ergm(init=coef(prev.fit))
can be used to
"resume" an uncoverged stergm
run, but see
enformulate.curved
.
Estimation method used to acquire initial values
for estimation. If NULL
(the default), the initial values
are computed using the edges dissolution approximation (Carnegie
et al.) when appropriate. If set to "zeros", the initial values
are set to zeros.
Logical: If TRUE, then force MCMC-based estimation method, even if the exact MLE can be computed via maximum pseudolikelihood estimation.
Specifies the
method to allocate probabilities of being proposed to dyads in
the formation/dissolution phase. Defaults to "default"
,
which picks a reasonable default for the specified
constraint. Possible values include "TNT"
,
"random"
, though not all values may be used with all
possible constraints.
An alternative, direct way of specifying additional arguments to the proposal in the formation/dissolution phase.
Maximum number of edges for which to allocate space.
Maximum number of changes in dynamic network simulation for which to allocate space.
Names of packages in which to look for change statistic functions in addition to those autodetected. This argument should not be needed outside of very strange setups.
Maximum number of Metropolis-Hastings steps per phase (formation and dissolution) per time step used in CMLE fitting.
Number of Metropolis-Hastings steps between successive draws when running MCMC MLE.
A convenience argument for specifying both
CMLE.control.form
and CMLE.control.diss
at once.
See control.ergm
.
Control parameters used
to fit the CMLE for the formation/dissolution ERGM. See
control.ergm
.
In STERGM CMLE, missing dyads in
transitioned-to networks are accommodated using methods of
Handcock and Gile (2009), but a similar approach to
transitioned-from networks requires much more complex methods
that are not, currently, implemented. CMLE.NA.impute
controls how missing dyads in transitioned-from networks are be
imputed. See argument imputers
of
impute.network.list
for details.
By default, no imputation is performed, and the fitting stops with an error if any transitioned-from networks have missing dyads.
The method
stergm{stergm}
uses at this time to fit a series of
more than two networks requires certain assumptions to be made
about the ERGM terms being used, which are tested before a fit is
attempted. This test sometimes fails despite the model being
amenable to fitting, so setting this option to TRUE
overrides the tests.
Estimation method used to find the Equilibrium Generalized Method of Moments estimator. Currently only "Gradient-Descent" is implemented.
Number of
Metropolis-Hastings steps per phase (formation and dissolution)
per time step used in EGMME fitting. By default, this is
determined adaptively by keeping track of increments in the
Hamming distance between the transitioned-from network and the
network being sampled (formation network or dissolution
network). Once EGMME.MCMC.burnin.min
steps have elapsed,
the increments are tested against 0, and when their average
number becomes statistically indistinguishable from 0 (with the
p-value being greater than EGMME.MCMC.burnin.pval
), or
EGMME.MCMC.burnin.max
steps are proposed, whichever comes
first, the simulation is stopped after an additional
EGMME.MCMC.burnin.add
times the number of elapsed steps
had been taken. (Stopping immediately would bias the sampling.)
To use a fixed number of steps, set both
EGMME.MCMC.burnin.min
and EGMME.MCMC.burnin.max
to
the desired number of steps.
Number of
Metropolis-Hastings steps per phase (formation and dissolution)
per time step used in EGMME fitting. By default, this is
determined adaptively by keeping track of increments in the
Hamming distance between the transitioned-from network and the
network being sampled (formation network or dissolution
network). Once EGMME.MCMC.burnin.min
steps have elapsed,
the increments are tested against 0, and when their average
number becomes statistically indistinguishable from 0 (with the
p-value being greater than EGMME.MCMC.burnin.pval
), or
EGMME.MCMC.burnin.max
steps are proposed, whichever comes
first, the simulation is stopped after an additional
EGMME.MCMC.burnin.add
times the number of elapsed steps
had been taken. (Stopping immediately would bias the sampling.)
To use a fixed number of steps, set both
EGMME.MCMC.burnin.min
and EGMME.MCMC.burnin.max
to
the desired number of steps.
No longer used. See
EGMME.MCMC.burnin.min
, EGMME.MCMC.burnin.max
,
EGMME.MCMC.burnin.pval
, EGMME.MCMC.burnin.pval
,
EGMME.MCMC.burnin.add
and CMLE.MCMC.burnin
and
CMLE.MCMC.interval
.
Multiplier for SAN.nsteps
relative to
MCMC.burnin
. This lets one control the amount of SAN burn-in
(arguably, the most important of SAN parameters) without overriding the
other SAN.control defaults.
SAN control parameters. See
control.san
Maximum number of times to restart a failed optimization process.
Number of time steps to advance the starting network before beginning the optimization.
Logical: Plot information
about the fit as it proceeds. If SA.plot.progress==TRUE
,
plot the trajectories of the parameters and target statistics as
the optimization progresses. If SA.plot.stats==TRUE
, plot
a heatmap reprsenting correlations of target statistics and a
heatmap representing the estimated gradient.
Do NOT use these with non-interactive plotting devices like
pdf
. (In fact, it will refuse to do that with a
warning.)
If SA.plot.progress==TRUE
, the
maximum number of time points to be plotted. Defaults to 400. If
more iterations elapse, they will be thinned to at most 400
before plotting.
Initial gain, the multiplier for the parameter update size. If the process initially goes crazy beyond recovery, lower this value.
Gain decay factor.
Number of parameter trials and updates per C run.
The number of time steps between updates of the parameters is set to be this times the mean duration of extant ties.
Initial number of time steps between updates of the parameters.
Upper and lower bounds on the number of time steps between updates of the parameters.
Number of runs during Phase 1 for estimating the gradient, before every gradient update.
Number of runs trying to find a reasonable parameter and network configuration.
Initial jitter standard deviation of each parameter.
Q-value (false discovery rate) that a gradient estimate must obtain before it is accepted (since sign is what is important).
If the run produces this relative increase in the approximate objective function, it will be backed off.
Range of gain levels (subphases) to go through.
Approximate precision of the estimates that must be attained before stopping.
A gain level may be
repeated multiple times (up to SA.phase2.repeats
) if the
optimizer detects that the objective function is improving or the
estimating equations are not centered around 0, so slowing down
the parameters at that point is counterproductive. To detect this
it looks at the the window controlled by SA.keep.oh
,
thinning objective function values to get
SA.stepdown.maxn
, and 1) fitting a GLS model for a linear
trend, with AR(2) autocorrelation and 2) conductiong an
approximate Hotelling's T^2 test for equality of estimating
equation values to 0. If there is no significance for either at
SA.stepdown.p
SA.stepdown.ct
runs in a row, the
gain level (subphase) is allowed to end. Otherwise, the process
continues at the same gain level.
A gain level may be repeated
multiple times (up to SA.phase2.repeats
) if the optimizer
detects that the objective function is improving or the
estimating equations are not centered around 0, so slowing down
the parameters at that point is counterproductive. To detect
this it looks at the the window controlled by SA.keep.oh
,
thinning objective function values to get
SA.stepdown.maxn
, and 1) fitting a GLS model for a linear
trend, with AR(2) autocorrelation and 2) conductiong an
approximate Hotelling's T^2 test for equality of estimating
equation values to 0. If there is no significance for either at
SA.stepdown.p
SA.stepdown.ct
runs in a row, the
gain level (subphase) is allowed to end. Otherwise, the process
continues at the same gain level.
At the end of each gain level after the minimum, if the precision is sufficiently high, the relationship between the parameters and the targets is tested for evidence of local nonlinearity. This is the p-value used.
If that test fails to reject, a Phase 3 run is made with the new parameter values, and the estimating equations are tested for difference from 0. If this test fails to reject, the optimization is finished.
If either of these tests rejects, at SA.stop.p
,
optimization is continued for another gain level.
Parameters controlling how much of optimization history to keep for gradient and covariance estimation.
A history record will be kept if it's at least one of the following:
Among the last SA.keep.oh
(a
fraction) of all runs.
Among the last SA.keep.min
(a
count) records.
From the last SA.keep.min.runs
(a
count) optimization runs.
Jitter standard deviation of each parameter is this value times its standard deviation without jitter.
To keep the optimization from "running away" due to, say, a poor gradient estimate building on itself, if a magnitude of change (Mahalanobis distance) in parameters over the course of a run divided by average magnitude of change for recent runs exceeds this, the change is truncated to this amount times the average for recent runs.
The multiplier for the range of parameter and statistics values to compute the guard width.
Because some parameters have much, much
greater effects than others, it improves numerical conditioning
and makes estimation more stable to rescale the \(k\)th
estimating function by \(s_k = (\sum_{i=1}^{q}
G_{i,k}^2/V_{i,i})^{-p/2}\), where \(G_{i,k}\) is the estimated
gradient of the \(i\)th target statistics with respect to
\(k\)th parameter. This parameter sets the value of \(p\):
0
for no rescaling, 1
(default) for scaling by
root-mean-square normalized gradient, and greater values for
greater penalty.
Whether to use robust linear regression (for gradients) and covariance estimation.
Absolute maximum number of data points per thread to store in the full optimization history.
Method, if any, used to refine the point estimate at the end: "linear" for linear interpolation, "mean" for average, and "none" to use the last value.
Logical: If TRUE (the default), get an MCMC sample of
statistics at the final estimate and compute the covariance
matrix (and hence standard errors) of the parameters. This sample
is stored and can also be used by
mcmc.diagnostics.stergm
to assess convergence.
This many optimization runs will be used to determine whether the optimization has converged and to estimate the standard errors.
Logical: if TRUE
(the default) an
error somewhere in the optimization process will cause it to
restart with a smaller gain value. Otherwise, the process will
stop. This is mainly used for debugging
A list of additional arguments to be passed to term initializers. It can also be set globally via option(ergm.term=list(...))
.
Seed value (integer) for the random number generator.
See set.seed
Number of threads in which to run the sampling. Defaults to 0 (no parallelism). See the entry on parallel processing for details and troubleshooting.
API to use for parallel processing. Supported
values are "MPI"
and "PSOCK"
. Defaults to using the
parallel
package default.
Logical: If TRUE, check that the
version of ergm
running on the slave
nodes is the same as that running on the master node.
A list with arguments as components.
This function is only used within a call to the stergm
function. See the usage
section in stergm
for details.
Boer, P., Huisman, M., Snijders, T.A.B., and Zeggelink, E.P.H. (2003), StOCNET User's Manual. Version 1.4.
Firth (1993), Bias Reduction in Maximum Likelihood Estimates. Biometrika, 80: 27-38.
Hunter, D. R. and M. S. Handcock (2006), Inference in curved exponential family models for networks. Journal of Computational and Graphical Statistics, 15: 565-583.
Hummel, R. M., Hunter, D. R., and Handcock, M. S. (2010), A Steplength Algorithm for Fitting ERGMs, Penn State Department of Statistics Technical Report.
stergm
. The
control.simulate.stergm
function performs a similar
function for simulate.stergm
.