Error in runHelper NLOPT unrecognized error -1

Nice to see that you are trying boundAdj. This option is only available with the latest code.

I just pushed a similar simulation,

inst/models/enormous/wu-neale-2012-lgc.R

This is how I recommend you structure your simulations. Use set.seed(result[rx,'rep']) inside the replications loop to ensure that every replication can be replicated. That way you can point to a specific replication that triggers NLOPT unrecognized error.

As you suspected, NLOPT unrecognized error -1 is only reported by SLSQP. You will not see this error with other optimizers. However, if you are interested in the boundary adjusted CIs then you will need to use SLSQP.

Hm, looking at your script, I think you need to add lbound=0 for the latent factor covariance. But wait, this parameter cannot be less than 0 so the boundary adjustment is not needed. I will email you my draft manuscript to explain better. I suspect you can avoid optimizer failures by estimating the cholesky factor of the correlation matrix. inst/models/passing/bivCorM.R is an example of how to do this.

I revised your code and it works. See below. The key is to set bounds of parameters to restrict the search of the CI algorithm. Note that the bounds I mentioned here has nothing to do with bounds of the parameters discussed in Wu and Neale (2012). Their introduction is just to make the search successful.

The reason that your version did not work out is because you did not specify appropriate bounds for the parameters. The search for CI is very delicate and prone to failure. You need to bound the search within some reasonable range. In your model, the parameter cov is now bound between -0.4 and 0.9; two loadings are bound to be positive to avoid sign change. As long as the search does not touch the artificial bounds, it is OK. If one of them is touched, set a slightly more lenient bound.

Here is the revised code.


for (i in 1:100)
{
print(i)
dat <- read.table(file="dat100NormalNA198.dat")
colnames(dat)<-paste("X",1:8,sep="")
data <- mxData( observed=dat, type="raw" )

resVars <- mxPath( from=paste("X",1:8,sep=""), arrows=2,
free=TRUE, values=rep(1,8), labels=paste("e",1:8,sep="") )
latVars <- mxPath( from=c("F1","F2"), arrows=2,connect="unique.pairs",
free=c(FALSE,TRUE,FALSE), values=c(1,0,1),
labels =c("varF1","cov","varF2") )
facLoads1 <- mxPath( from="F1", to=paste("X",1:4,sep=""), arrows=1,
free=rep(TRUE,4), values=rep(1,4),labels =paste("l",1:4,sep="") )
facLoads2 <- mxPath( from="F2", to=paste("X",5:8,sep=""), arrows=1,
free=rep(TRUE,4), values=rep(1,4),
labels =paste("l",5:8,sep="") )
means <- mxPath( from="one", to=c(paste("X",1:8,sep=""),"F1","F2"), arrows=1,
free=c(rep(TRUE,8),FALSE,FALSE), values=c(rep(1,8),0,0),
labels =c(paste("meanx",1:8,sep=""),NA,NA) )
mxpindx<-"cov"
CIModel <- mxModel("CIModel", type="RAM",
manifestVars=paste("X",1:8,sep=""), latentVars=c("F1","F2"),
data, resVars, latVars, facLoads1, facLoads2, means)
CIModel<-omxSetParameters(CIModel,labels = "cov",ubound = .99,lbound = -.4)
CIModel<-omxSetParameters(CIModel,labels = c("l1","l8"),lbound = 0)
alpha<-.05
CI<-mxCI(mxpindx,interval=1-alpha,type="both")

CIFit <- try(mxRun(mxModel(model=CIModel, CI), intervals = TRUE, silent=TRUE))
if(class(CIFit)=="try-error") print(CIFit)
}