The function sienaRI computes various effect sizes for
Stochastic Actor-oriented Models.
As its value it returns the relative importance of effects
of a SAOM according to the measure of relative importance described in
Section 3.1 of Indlekofer and Brandes (2013).
This is based on the influence of effects on potential tie change or
behavior change decisions of individual actors at the given observation
moments (waves).
It takes the data as well as the complete model specification into account.
Therefore, necessary arguments are the analysed data given as a
siena data object as well as the complete model
specification represented either by an estimated sienaFit
object or by the triple consisting of a
suitable parameter vector theta and the corresponding
sienaAlgorithm and sienaEffects objects.
Other measures useful for effect sizes are entropy-based
effect sizes as in Snijders (2004) and the within-ego
standard deviations of change statistics.
If getChangeStats=TRUE, the arrays of change statistics are
stored in the sienaRI object.
sienaRI(data, ans=NULL, theta=NULL, algorithm=NULL, effects=NULL,
getChangeStats=FALSE)
# S3 method for sienaRI
print(x, printSigma=FALSE,...)
# S3 method for sienaRI
plot(x, actors = NULL, col = NULL, addPieChart = FALSE,
radius = 1, width = NULL, height = NULL, legend = TRUE,
legendColumns = NULL, legendHeight = NULL,
cex.legend = NULL, cex.names = NULL,...)If the model contains only one dependent variable, sienaRI returns
an object of class sienaRI. Otherwise, it returns a list of objects
of class sienaRI, each corresponding to one dependent variable.
A returned sienaRI object stores the expected relative importances of
effects of one dependent variable at observation moments as defined in
Section 3.1 of Indlekofer and Brandes (2013).
A sienaRI object is a list with the following components.
For the components referred to as lists themselves, these are lists
corresponding to the observation moments.
dependentVariablethe name of the corresponding dependent variable.
effectNamesthe names of considered effects.
RIActorsa list that contains the expected relative importances of effects for each potential actor decision at observation moments. This is equation (3) in Indlekofer and Brandes (2013).
expectedRIa list that contains the expected
relative importances of effects aggregated over all actors for each
network observation. These are the averages of the actor related
values in RIActors.
This is equation (4) in Indlekofer and Brandes (2013).
IActorsa list that contains the expected importances of effects for each potential actor decision at observation moments. This is the numerator of equation (3) in Indlekofer and Brandes (2013).
expectedIa list that contains the expected
importances of effects
aggregated over all actors in each observation.
More precisely, it contains the averages of the actor related values
in IActors.
absoluteSumActorsa list that contains the sum of the (unstandardized) L1-differences calculated for each potential actor decision at observation moments. This is the denominator of equation (3) in Indlekofer and Brandes (2013).
RHActorsa list that contains the degree of certainty,
also called degree of determination, in
the potential ministep taken by an actor at the observation moments;
this is R_H(i,x) of formula (6) in Snijders (2004).
The mean over actors of these degrees of certainty, given by
formula (7) in Snijders (2004), is printed by the
print method for sienaRI objects.
sigmaa list of effects by ego matrices of the values of the within-ego standard deviations of the change statistics.
sigmaseffects by wave matrices of averages
(see above for how this is done; across actors) of sigma.
These are printed if printSigma=TRUE.
meansigmasaverage (see above for how this is done)
over waves of sigmas.
changeStatisticsa list of arrays
(effects by choices by egos)
containing, for each observation wave,
the values of the change statistics for making this choice,
where for one-mode networks the choice is defined by the alters
with alter=ego referring to 'no change',
for two-mode networks the choice is defined by the second-mode nodes
with the last choice referring to 'no change',
and for behavior the choice is defined as
going down, staying constant, or going up;
this output is produced only if getChangeStats=TRUE.
toggleProbabilitiesan array (egos by choices by waves), where "choices" are as directly above, giving the choice probabilities of ego in a ministep, when the data are as in this wave.
siena data object representing the analyzed data and
resulting from a call to sienaDataCreate.
sienaFit object resulting from a call to
siena07. The sienaFit object contains
all necessary information on the model specification,
in particular, the vector of parameter values ans$theta,
the used algorithm for estimation ans$x,
and information on included model effects ans$effects.
If ans is a valid sienaFit object, the calculations
of relative importances are based on ans$theta,
ans$x, and ans$effects.
Alternatively, the necessary information can be given directly
as a suitable parameter vector theta, a
sienaAlgorithm object, and a sienaEffects object.
In this case, ans has to be unspecified
(i.e., ans=NULL).
Vector of parameter values of effects included in the model.
Length of theta has to be equal to the number of
included effects.
sienaAlgorithm object as resulting from a call to
sienaAlgorithmCreate. Works only for estimation
by Method of Moments (i.e., if maxlike = FALSE).
sienaEffects object specifying which effects are
included the model. Note that sienaRI does not yet work
for endowment or creation effect, i.e., included effects have to
be of type eval (or rate).
Boolean: If TRUE, an array of change statistics
is added to the sienaRI object.
sienaRI object resulting from a call to sienaRI.
Boolean: If TRUE,
average within-actor standard deviations of change statistics
('sigma') are included in the print.
vector of integers: set of actors to be included in the plot;
if NULL, all actors.
Colors used in the plot. If col=NULL a default color
scheme is used.
Boolean: If TRUE, pie charts of aggregated relative
importances for the complete set of actors will be added to the plot.
Radius of pie charts. Only effective if addPieCharts = TRUE.
Width of the plot. If width=NULL a default value is used.
Height of the plot. If height=NULL a default value is used.
Boolean: if TRUE a legend is added to the plot.
x$effectNames are used as labels.
Number of columns in legend. If legendColumns=NULL
a default value is used. Only effective if legend=TRUE.
Height of legend. If legendHeight=NULL a default
value is used. Only effective if legend=TRUE.
Specifies the relative font size of legend labels.
Specifies the relative font size of bar graph labels.
Other arguments.
Natalie Indlekofer, additions by Tom Snijders
sienaRI takes the data as well as the complete model specification
into account. Therefore, necessary arguments are the analyzed data given as a
siena data object as well as the complete model specification represented
either by an estimated sienaFit object or by the triple consisting of a
suitable parameter vector theta and the corresponding
sienaAlgorithm and sienaEffects objects.
A sienaFit object contains all necessary information on the model
specification, in particular, the vector of parameter values ans$theta,
the used algorithm for estimation ans$x, and information on included
model effects ans$effects.
If ans is a valid sienaFit object, the calculations of relative
importances are based on ans$theta, ans$x, and
ans$effects. Alternatively, the necessary information can be given
directly as a suitable parameter vector theta, a sienaAlgorithm
object, and a sienaEffects object. In this case, ans has to be
unspecified, i.e., ans=NULL.
Note that sienaRI works only with Method of Moments
(i.e., for sienaAlgorithm objects with maxlike = FALSE)
and that it does not yet work for endowment or creation effects
(i.e., included effects have to be of type eval),
and also not for models with interaction effects.
For two-mode (bipartite) networks as dependent variables,
it works only if the number of second-mode nodes is less than the
number of actors.
If the network is non-directed, the relative importances and entropy-based
'degrees of certainty' are calculated for modelType=2
('forcing'; see sienaAlgorithmCreate).
For dependent behavior variables, behModelType=1 ('standard') is assumed.
If there are any missing tie values in the network data set, they are imputed by
initial zeros and Last Observation Carried Forward. Structural zeros and ones
are replaced by NA and treated as impossible choices in the probability
vectors and ignored in the standard deviations; but the change statistics
for these dyads still are given in changeStatistics (if requested).
The averages reported in the components sigmas (average across actors)
and meansigmas (average across waves) are obtained by averaging at the
variance level and then taking square roots.
Indlekofer, N. and Brandes, U. (2013), Relative Importance of Effects in Stochastic Actor-oriented Models. Network Science, 1, 278--304.
Snijders, T.A.B. (2004), Explained Variation in Dynamic Network Models. Mathematics and Social Sciences, 168, 31--41.
myalgorithm <- sienaAlgorithmCreate(nsub=1, n3=50, projname=NULL)
mynet1 <- sienaDependent(array(c(tmp3, tmp4), dim=c(32, 32, 2)))
mydata <- sienaDataCreate(mynet1)
myeff <- getEffects(mydata)
myeff <- includeEffects(myeff, density, recip, outAct, inPop)
(myeff <- setEffect(myeff, reciAct, parameter=1))
ans <- siena07(myalgorithm, data=mydata, effects=myeff, batch=TRUE)
RI <- sienaRI(mydata, ans)
RI
print(RI, printSigma=TRUE)
# average within-ego standard deviations of change statistics by wave:
RI$sigmas
# sigma averaged over waves:
RI$meansigmas
# semi-standardized parameters by wave:
ans$theta * RI$meansigmas
if (FALSE) {
plot(RI, addPieChart=TRUE)
plot(RI, actors=1:20, addPieChart=TRUE, radius=1.08)
}
myalgorithm <- sienaAlgorithmCreate(nsub=1, n3=50, projname=NULL)
mynet2 <- sienaDependent(array(c(s502, s503), dim=c(50, 50, 2)))
mybeh <- sienaDependent(s50a[,2:3], type="behavior")
mydata2 <- sienaDataCreate(mynet2, mybeh)
myeff2 <- getEffects(mydata2)
myeff2 <- includeEffects(myeff2, density, recip, transTies)
ans2 <- siena07(myalgorithm, data=mydata2, effects=myeff2, batch=TRUE)
# Use only the parameters for the evaluation function:
theta.eval <- ans2$theta[myeff2$type[myeff2$include]=="eval"]
RI <- sienaRI(mydata2, theta=theta.eval, algorithm=myalgorithm,
effects = myeff2)
RI
if (FALSE) {
plot(RI[[2]], col = c("red", "green"), legend=FALSE)
plot(RI[[1]], addPieChart = TRUE, legendColumns=2)
}
Run the code above in your browser using DataLab