Function to perform a meta analysis of a collection of Siena fits.
Estimates a meta analysis based on a collection of Siena fits.
siena08(..., projname = "sienaMeta", bound = 5, alpha = 0.05, maxit=20)
...: names of sienaFit objects, returned from siena07. They will be renamed if entered in format newname=oldname. It is also allowed to give for ... a list of sienaFit
objects.
projname: Base name of report file if required
bound: Upper limit of standard error for inclusion in the meta analysis.
alpha: 1 minus confidence level of confidence intervals.
maxit: Number of iterations of iterated least squares procedure.
A meta analysis is performed as described in the Siena manual, section "Meta-analysis of Siena results". This consists of three parts: an iterated weighted least squares (IWLS) modification of the method described in the reference below; maximum likelihood estimates and confidence intervals based on profile likelihoods under normality assumptions; and Fisher combinations of left-sided and right-sided -values. These are produced for all effects separately.
Note that the corresponding effects must have the same effect name in each model fit. This implies that at least covariates and behavior variables must have the same name in each model fit.
An object of class sienaMeta. There are print, summary and plot methods for this class. This object contains at least the following. - thetadf: Data frame containing the coefficients, standard errors and score test results
projname: Root name for any output file to be produced by the print method
bound: Estimates with standard error above this value were excluded from the calculations
scores: Object of class by indicating, for each effect in the models, whether score test information was present.
requestedEffects: The requestedEffects component of the first sienaFit object in ....
muhat: The vector of IWLS estimates.
se.muhat: The vector of standard errors of the IWLS estimates.
theta: The vector of ML estimates mu.ml (see below).
se: The vector of standard errors of the ML estimates mu.ml.se
(see below).
Then for each effect, there is a list with at least the following.
cor.est: Spearman rank correlation coefficient between estimates and their standard errors.
cor.pval: p-value for above
regfit: Part of the result of the fit of iwlsm.
regsummary: The summary of the fit, which includes the coefficient table.
Tsq: test statistic for effect zero in every model
pTsq: p-value for above
tratio: test statistics that mean effect is 0
ptratio: p-value for above
Qstat: Test statistic for variance of effects is zero
pttilde: p-value for above
cjplus: Test statistic for at least one theta strictly greater than 0
cjminus: Test statistic for at least one theta strictly less than 0
cjplusp: p-value for cjplus
cjminusp: p-value for cjminus
mu.ml: ML estimate of population mean
mu.ml.se: standard error of ML estimate of population mean
sigma.ml: ML estimate of population standard deviation
mu.confint: confidence interval for population mean based on profile likelihood
sigma.confint: confidence interval for population standard deviation based on profile likelihood
n1: Number of fits on which the meta analysis is based
cjplus: Test statistic for combination of right one-sided Fisher combination tests
cjminus: Test statistic for combination of left one-sided Fisher combination tests
cjplusp: p-value for cjplus
cjminusp: p-value for cjminus
scoreplus: Test statistic for combination of right one-sided -values from score tests
scoreminus: Test statistic for combination of left one-sided -values from score tests
scoreplusp: p-value for scoreplus
scoreminusp: p-value for scoreminus
ns: Number of fits on which the score test analysis is based
Snijders, T.A.B, and Baerveldt, C. (2003), A Multilevel Network Study of the Effects of Delinquent Behavior on Friendship Evolution. Journal of Mathematical Sociology 27 , 123--151.
See also the manual (Section 11.2) and https://www.stats.ox.ac.uk/~snijders/siena/
Ruth Ripley, Tom Snijders
print.sienaMeta, funnelPlot, meta.table, iwlsm, siena07
## Not run: # A meta-analysis for three groups does not make much sense # for generalizing to a population of networks, # but the Fisher combinations of p-values are meaningful. # However, using three groups does show the idea. Group1 <- sienaDependent(array(c(N3401, HN3401), dim=c(45, 45, 2))) Group3 <- sienaDependent(array(c(N3403, HN3403), dim=c(37, 37, 2))) Group4 <- sienaDependent(array(c(N3404, HN3404), dim=c(33, 33, 2))) dataset.1 <- sienaDataCreate(Friends = Group1) dataset.3 <- sienaDataCreate(Friends = Group3) dataset.4 <- sienaDataCreate(Friends = Group4) OneAlgorithm <- sienaAlgorithmCreate(projname = "SingleGroups", seed=128) effects.1 <- getEffects(dataset.1) effects.3 <- getEffects(dataset.3) effects.4 <- getEffects(dataset.4) effects.1 <- includeEffects(effects.1, transTrip) effects.1 <- setEffect(effects.1, transRecTrip, fix=TRUE, test=TRUE) effects.3 <- includeEffects(effects.3, transTrip) effects.3 <- setEffect(effects.3, transRecTrip, fix=TRUE, test=TRUE) effects.4 <- includeEffects(effects.4, transTrip) effects.4 <- setEffect(effects.4, transRecTrip, fix=TRUE, test=TRUE) ans.1 <- siena07(OneAlgorithm, data=dataset.1, effects=effects.1, batch=TRUE) ans.3 <- siena07(OneAlgorithm, data=dataset.3, effects=effects.3, batch=TRUE) ans.4 <- siena07(OneAlgorithm, data=dataset.4, effects=effects.4, batch=TRUE) ans.1 ans.3 ans.4 (meta <- siena08(ans.1, ans.3, ans.4)) plot(meta, which=2:3, layout = c(2,1)) # For specifically presenting the Fisher combinations: # First determine the components of meta with estimated effects: which.est <- sapply(meta, function(x){ifelse(is.list(x),!is.null(x$cjplus),FALSE)}) Fishers <- t(sapply(1:sum(which.est), function(i){c(meta[[i]]$cjplus, meta[[i]]$cjminus, meta[[i]]$cjplusp, meta[[i]]$cjminusp, 2*meta[[i]]$n1 )})) Fishers <- as.data.frame(Fishers, row.names=names(meta)[which.est]) names(Fishers) <- c('Fplus', 'Fminus', 'pplus', 'pminus', 'df') Fishers round(Fishers,4) ## End(Not run)