sienaAlgorithmCreate {RSiena}

Function to create an object containing the algorithm specifications for parameter estimation in RSiena

Description

Creates an object with specifications for the algorithm for parameter estimation in RSiena.

sienaAlgorithmCreate() and sienaModelCreate() are identical functions; the second name was used from the start of the RSiena package, but the first name indicates more precisely the purpose of this function.

Usage

sienaAlgorithmCreate(fn, projname = "Siena", MaxDegree = NULL, Offset = NULL,
     useStdInits = FALSE, n3 = 1000, nsub = 4, n2start = NULL,
     dolby=TRUE, maxlike = FALSE, gmm = FALSE, diagonalize=0.2*!maxlike,
     condvarno = 0, condname = "", firstg = 0.2, reduceg = 0.5,
     cond = NA, findiff = FALSE, seed = NULL,
     prML=1,
     maximumPermutationLength=40,
     minimumPermutationLength=2, initialPermutationLength=20,
     modelType=NULL, behModelType=NULL, mult=5, simOnly=FALSE, localML=FALSE,
     truncation=5, doubleAveraging=0, standardizeVar=(diagonalize<1),
     lessMem=FALSE)

sienaModelCreate(fn, projname = "Siena", MaxDegree = NULL, Offset = NULL,
     useStdInits = FALSE, n3 = 1000, nsub = 4, n2start = NULL,
     dolby=TRUE, maxlike = FALSE, gmm = FALSE, diagonalize=0.2*!maxlike,
     condvarno = 0, condname = "", firstg = 0.2, reduceg = 0.5,
     cond = NA, findiff = FALSE, seed = NULL,
     prML=1,
     maximumPermutationLength=40,
     minimumPermutationLength=2, initialPermutationLength=20,
     modelType=NULL, behModelType=NULL, mult=5, simOnly=FALSE, localML=FALSE,
     truncation=5, doubleAveraging=0, standardizeVar=(diagonalize<1),
     lessMem=FALSE)

Arguments

fn

Function to do one simulation in the Robbins-Monro algorithm. Not to be touched.

projname

Character string name of project; the output file will be called projname.txt. No embedded spaces!!!
If projname=NULL, output will be written to a file in the temporary session directory, created as tempfile(Siena).

MaxDegree

Named vector of maximum degree values for corresponding networks. Allows to restrict the model to networks with degrees not higher than this maximum. Names should be the names of all dependent network variables, in the same order as in the Siena data set.
Default as well as value 0 imply no restrictions.
This option is not available for maximum likelihood estimation.

Offset

Named vector of offset values for symmetric networks with modelType = 3 (M.1), and for universal setting in Settings model. Names should be the names of all dependent network variables, in the same order as in the Siena data set. Default NULL implies values 0.

useStdInits

Boolean. If TRUE, the initial values in the effects object will be ignored and default values used instead. If FALSE, the initial values in the effects object will be used.

n3

Number of iterations in phase 3. For regular use with the Method of Moments, n3=1000 mostly suffices. For use in publications and for Maximum Likelihood, at least n3=3000 is advised. Sometimes much higher values are required for stable estimation of standard errors.

nsub

Number of subphases in phase 2.

n2start

Minimum number of iterations in subphase 1 of phase 2; default is 2.52*(p+7), where p = number of estimated parameters; if useCluster=TRUE in the call of siena07, this is divided by nbrNodes.

dolby

Boolean. Should there be noise reduction by regression on augmented data score. In most cases dolby=TRUE yields better convergence, but takes some extra computing time; if convergence is problematic, however, dolby=FALSE may be tried. Just use whatever works best.

maxlike

Whether to use maximum likelihood method or Method of Moments estimation.

gmm

Whether to use the Generalized Method of Moments or the regular Method of Moments estimation.

diagonalize

Number between 0 and 1 (bounds included), values outside this interval will be truncated; for diagonalize=0 the complete estimated derivative matrix will be used for updates in the Robbins-Monro procedure; for diagonalize=1 only the diagonal entries will be used; for values between 0 and 1, the weighted average will be used with weight diagonalize for the diagonalized matrix. Has no effect for ML estimation.
Higher values are more stable, lower values potentially more efficient. Default: for ML estimation, diagonalize=0; for MoM estimation, diagonalize = 0.2.

condvarno

If cond (conditional simulation), the sequential number of the network or behavior variable on which to condition.

condname

If conditional, the name of the dependent variable on which to condition. Use one or other of condname or condvarno to specify the variable.

firstg

Initial value of scaling ("gain") parameter for updates in the Robbins-Monro procedure.

reduceg

Reduction factor for scaling ("gain") parameter for updates in the Robbins-Monro procedure (MoM only).

cond

Boolean. Only relevant for Method of Moments simulation/estimation. If TRUE, use conditional simulation; if FALSE, unconditional simulation. If missing, decision is deferred until siena07, when it is set to TRUE if there is only one dependent variable, FALSE otherwise.

findiff

Boolean: If TRUE, estimate derivatives using finite differences. If FALSE, use scores.

seed

Integer. Starting value of random seed. Not used if parallel testing.

prML

Either one real number, or a vector of 7 numbers. Determines update probabilities used in Metropolis-Hastings routine in ML estimation. Should be nonnegative; if a vector, the sum should be <= 1. See Details.

maximumPermutationLength

Maximum length of permutation in steps in ML estimation.

minimumPermutationLength

Minimum length of permutation in steps in ML estimation.

initialPermutationLength

Initial length of permutation in steps in ML estimation.

modelType

Named vector indicating the type of model to be fitted for dependent network variables. (See the examples below for how to specify a named vector.)
Possible values are:
1=directed standard,
2:6 for symmetric networks only: 2=dictatorial forcing (D.1), 3=Initiative model with reciprocal confirmation (M.1), 4=Pairwise dictatorial forcing model (D.2), 5=Pairwise mutual model (M.2), 6=Pairwise joint model (C.2),
7:10 for directed one-mode only: 7=Double Step model with double step probability 0.25, 8=Double Step model with double step probability 0.50, 9=Double Step model with double step probability 0.75, 10=Double Step model with double step probability 1.00.
Names should be the names of all dependent network variables, in the same order as in the Siena data set.
See Snijders and Pickup (2016) for the meanings of the various models for symmetric networks.
Default NULL implies 1 for directed or two-mode, 2 for symmetric.

behModelType

Named vector indicating the type of model to be fitted for behavioral dependent variables. (See the examples below for how to specify a named vector.)
Possible values are:
1=standard (restricted), 2=absorbing.
Names should be the names of all dependent behavioral variables, in the same order as in the Siena data set.
Default NULL implies values 1.

mult

Multiplication factor for maximum likelihood and Bayes. Number of steps per iteration is set to this multiple of the total distance between the observations at start and finish of the wave (and rounded). Decreasing mult below a certain value has no further effect.
mult can be either a number (which needs to be positive) or a vector of numbers, of length equal to the total number of periods. Note that for multi-group data, the total number of periods is equal to the number of groups times the number of periods per group (if the latter is constant).

simOnly

Logical: If TRUE, then the calculation of the covariance matrix and standard errors of the estimates at the end of Phase 3 of the estimation algorithm in function siena07 is skipped. This is suitable if nsub=0 and siena07 is used only for the purpose of simulation.

localML

Logical: If TRUE, and maxlike, then calculations are sped up for models with all local effects.

truncation

Used for step truncation in the Robbins Monro algorithm (applied to deviate/(standard deviation)).

doubleAveraging

subphase after which double averaging is used in the Robbins Monro algorithm, which probably increases algorithm efficiency.

standardizeVar

Logical: whether to limit deviations used in Robbins-Monro updates to unit variances.

lessMem

Logical: whether to reduce storage during operation of siena07, and of the object produced, by leaving out arrays by iteration and by period of simulated statistics sf2 and scores ssc.
if lessMem=TRUE, it will be impossible to run sienaTimeTest or sienaGOF on the object produced by siena07.

Details

Model specification is done via this object for siena07. This function creates an object with the elements required to control the Robbins-Monro algorithm. Those not available as arguments can be changed manually when desired.
The value prML=1 defines the defaults valid in RSiena up to version 1.3.16.
If prML is given as a vector of 7 probabilities, these are, consecutively: the probabilities of inserting a diagonal step, deleting a diagonal step, permuting, inserting a CCP, deleting a CCP, inserting random missing, deleting random missing; the residual (1 minus the sum) is the probability of a move step.
Further information about the implementation of the algorithm is in
https://www.stats.ox.ac.uk/~snijders/siena/Siena_algorithms.pdf.
Some of the examples use projname=NULL; this is just for the sake of checking the examples, not necessarily intended for normal use.

Value

Returns an object of class sienaAlgorithm containing values implied by the parameters.

Author(s)

Ruth Ripley and Tom A.B. Snijders

References

For modelType:
Snijders, T.A.B., and Pickup, M. (2016), Stochastic Actor-Oriented Models for Network Dynamics. In: Victor, J.N., Lubell, M., and Montgomery, A.H., Oxford Handbook of Political Networks. Oxford University Press.

Also see https://www.stats.ox.ac.uk/~snijders/siena/

See Also

siena07, simstats0c.

Examples

myAlgorithm <- sienaAlgorithmCreate(projname="NetworkDyn")
StdAlgorithm <- sienaAlgorithmCreate(projname="NetworkDyn", useStdInits=TRUE)
CondAlgorithm <- sienaAlgorithmCreate(projname="NetworkDyn", condvarno=1, cond=TRUE)
Max10Algorithm <- sienaAlgorithmCreate(projname="NetworkDyn", MaxDegree=c(mynet=10),
     modelType=c(mynet=1))
Beh2Algorithm <- sienaAlgorithmCreate(projname="NetBehDyn", behModelType=c(mybeh=2))
# where mynet is the name of the network object created by sienaDependent(),
# and mybeh the name of the behavior object created by the same function.