NEWS | R Documentation |
tergm
packageSTERGM conditional MLE can now be fit to series of more than two bipartite networks. (Previously, for bipartite networks, only a series of two (i.e., one transition) could be fit.)
Miscellaneous CMLE bug fixes.
Some low-priority warnings are no longer printed by default.
A bug in formation TNT proposals has been fixed. The only analyses that may be affected are those on tiny (about 8 nodes) or extremely dense (i.e., high probability of a complete graph) networks.
A new control parameter, SA.par.eff.pow
, now controls the degree to which the magnitude of the gradient afects the rescaling of the estimating functions (see help): higher means that "stronger" parameters effectively get lower gain, making the optimization more stable.
Running with verbose=TRUE now prints gc
output periodically.
Most durational functions can now take a binary option log=, which will result in them using log-ages instead of raw ages. (In a discrete-time model, there are no 0-age ties, so log transformation is safe.)
tergm.godfather example now tests for a few more terms, degrange ages in particular.
Implemented reference to MH_listTNT
in applicable constraints in tergm as well.
Durational statistics can now take NULL
as their "role", disabling role testing.
Running STERGM EGMME with parallel > 0 no longer causes it to eat up all the memory.
Namespace prefixing has been added to some functions called
from external packages to deal with CRAN check warnings, replaced
require
and library
calls with
requireNamespace()
Updated CITATION
file to deal with CRAN check warnings (packageDescription
is no longer used).
Added S3 method registration in NAMESPACE
for all S3
functions.
Handling of networkDynamic
inputs has been improved.
Changed CMLE MCMC default parameters to match those of ergm.
Additions and corrections to documentation
Most formerly undocumented 'internal' functions have been made internal by the NAMESPACE
. As far as we know, this avoids functions in use by reverse-Depending packages. These currently appear as commented items in the NAMESPACE
file.
Some potential memory issues revealed by Valgrind have been fixed.
Speed improvements. Internal data handling has been modified to avoid allocating (and passing around) a large matrix of tie ages unless the formula/model explicitly requires tie duration information. This increases speed of model initialization, which is especially helpful when simulating a network where the model must be re-initialized at each timestep (i.e. models with vital dynamics).
Simplified specification of offset values in formation model, when using stergm.EGMME with target statistics. Edges dissolution approximation is used to initialize the coefficients for estimation when there is an offset. See http://statnet.csde.washington.edu/trac/wiki/tergmUsage for examples.
Offset terms in the target formula are automatically removed. This includes the case where the target is set equal to formation: targets = "formation"
The simulate.networkDynamic and related function now make use of networkDynamics persistant id (PID) functionality. If the network does not have any persistent.ids
defined for vertices, a vertex.pid will be attached in a vertex attribute named 'tergm_pid'
to facilitate 'bookkeeping' between the networkDynamic argument and the simulated network time step.
Parallel functionality: tergm can take user-created clusters as the control.stergm(parallel) control parameter. This is the recommended method for using tergm on a high-performance computing cluster. See ergm-parallel. Functionality is now implemented via the parallel
package, direct dependence on snow
package is removed
Addition of bipartite formation and dissolution Metropolis-Hastings proposals
Addition of a TNT dissolution proposal constraint
Inclusion of a Stergm vignette
-Inf offsets are recoded to a large negative number to be compatible with C code. The fitted coefficients will show the large negative number instead of the -Inf offset.
Namespace prefixing has been added to some functions called from external packages to deal with CRAN check warnings
Bug fix where stergm constraints were not being passed from all the way down into san
Changes to control.stergm
parameters:
init.method
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.
SA.oh.memory = 100000
Absolute maximum number of data points per thread to store in the full optimization history.
Changes to simulate.stergm
parameters: addition of duration.dependent
parameter: Logical: Whether the model terms in formula or model are duration dependent. E.g., if a duration-dependent term is used in estimation/simulation model, the probability of forming or dissolving a tie may dependent on the age the dyad status. If TRUE, the matrix of tie ages will be allocated.
Some Metropolis-Hastings proposal functions would sometimes
return incorrect acceptance probabilities when combined with the
bd
constraint. This has been
fixed.
In simulate.networkDynamic
, an error in which
vertices were queried and updated in the simulation has been
fixed.
simulate.networkDynamic
and
simulate.network
functions now take an additional
argument, time.offset
. See help for those functions for
details.
The Tie-NonTie (TNT) proposal has been implemented for dissolution phase models. This should improve mixing and inference for these models.
stergm
and simulate.stergm
now determine the
number of Metropolis-Hastings steps per time step adaptively,
stopping when the formation/dissolution process appears to have
converged.
The previously deprecated start and end attr values attached
to networks returned by simulate.network
and
simulate.networkDynamic
have been removed and
replaced with observation spells recorded as a
net.obs.period
network attribute.
MCMC.burnin
and MCMC.interval
arguments to
control.stergm
, control.simulate.network
, and
control.simulate.stergm
have been replaced by a different
mechanism. See the help for the respective control functions for
more information.
No longer generates deprecation warnings about start and end
attrs when various internal functions use as.data.frame. Resolved
by replacing start and end with a net.obs.period
object.
A bug in the check for whether terms were amenable to being fit using STERGM CMLE for multiple transitions has been fixed.
Some minor documentation typos have been fixed.
Version 3.1.1 has been skipped to ease upgrading for those using a preview release.
This package consists of the dynamic network modeling code that has been
split out of the ergm
package.
Changes listed in the following sections are relative to
ergm
version 3.0.
Although fitting the EGMME for dissolution was possible
before, it was impractical due to nonidentifiability: for example,
one cannot fit both edges
formation and
edges
dissolution with only one edges
target statistic. Four new statistics have been added and
documented that focus on targeting observed and hazard:
mean.age
, edgecov.mean.age
,
degree.mean.age
, degrange.mean.age
,
edge.ages
, edgecov.ages
, and
edges.ageinterval
. This allow jointly fitting
formation and dissolution.
In addition to the progress plot, EGMME fitting routines can now plot the estimated gradient matrix and the matrix of correlations among the target statistics.
EGMME fitting is now more adaptive when determining when to stop the optimization and return the result.
EGMME fitting can now take advantage of multiple CPUs, cores, or cluster nodes for faster and more robust fitting.
EGMME and CMLE fitting can now accommodate a
constraints
argument. However, note that the constraints
apply to post-formation (y^+=y^0\cup\y^1) and
post-dissolution (y^-=y^0\cap\y^1) networks, not to the
final network (y^1). This may change in the future.
CMLE can now be fit to more than two networks. (Not all ERGM terms and constraints can be used in this mode, however.)
A new (sort of) function, tergm.godfather
has
been fixed and documented; it can be used to apply a specific set
of changes to a network, returning statistics of interest as it
evolves. In particular, it can be used to “retrace” the evolution
of a networkDynamic
, calculating statistics at
discrete time points along the way.
tergm
now implements a
formula-based summary
method for networkDynamic
LHS, to compute dynamic
network statistics at specified time points.
Fitting CMLE to network series with transitioned-from
networks having missing dyads is now possible, using automatic
imputation. See impute.network.list
and
control.tergm
.
STERGM simulate
can now be used
for CMLE fits, takes a number of new arguments, and can be used to
“resume” a simulation from a networkDynamic
object.
gof
methods have been implemented for CMLE fits.
stergm
EGMME initial fitting code has been
vastly improved.
The EGMME fitting algorithm has been vastly improved and is now a lot more adaptive and able to recover from problems. Many bugs have also been fixed.
A number of control.stergm
parameters had been
renamed and otherwise changed.
Argument statsonly
to the
simulate.stergm
and related functions has been
deprecated in favor of output
.
The conditional MPLE (CMPLE) for formation and dissolution is now fit correctly. This also means that the starting values for the CMLE are much better.
Bugs in EGMME code related to handling of bipartite networks have been fixed.