Problem with CIs (to close to point estimate)

1 reply [Last post]
m.bartels's picture
Joined: 12/18/2009

Hi all,
Fitting tri-variate models with CIs but CIs look odd.
They are to close to the point estimate and lower point is sometimes higher that point estimate.
Below a selection

MZMmodel.sdAm[1,1] 3.224423e-01 0.32244499 0.39122315
MZMmodel.sdAm[1,2] 4.477139e-01 0.58554106 0.58555366
MZMmodel.sdAm[1,3] 8.285475e-01 0.82850806 0.82855803
MZMmodel.sdAm[2,1] 4.477139e-01 0.58554106 0.58555366
MZMmodel.sdAm[2,2] 4.104425e-01 0.41044353 0.41043733
MZMmodel.sdAm[2,3] 4.554080e-01 0.45540625 0.61216648
MZMmodel.sdAm[3,1] 8.285475e-01 0.82850806 0.82855803
MZMmodel.sdAm[3,2] 4.554080e-01 0.45540625 0.61216648
MZMmodel.sdAm[3,3] 4.554037e-01 0.45540226 0.52462605
MZMmodel.sdCm[1,1] 5.278062e-15 0.01168100 0.07729734

Any suggestions what might cause the problem? (and solution to solve it?)


neale's picture
Joined: 07/31/2009
I would try changing the

I would try changing the function precision; 1.e-12 might be better for continuous data, probably best not to go much lower, how's this look?


(so nothing specific set for this model)... Check the system setting (probably the defaults)

options()$mxOptions$`Function precision`

[1] "1e-14"

Ok let's make it a bit less accurate (this will force larger step-sizes during estimates of the gradients which the optimizer had incorrectly estimated to be zero) and then re-run it

CholAEFit <- mxOption(CholAEFit,"Function precision", 1e-12)
summary(CholAEFit2 <- mxRun(CholAEFit,intervals=TRUE))

snip snip...
confidence intervals:
lbound estimate ubound
MZMmodel.sdAm[1,1] 0.2835146 0.3395857 0.3941184
MZMmodel.sdAm[1,2] 0.6855848 0.8962901 1.1197032
MZMmodel.sdAm[2,1] 0.6855848 0.8962901 1.1197032
MZMmodel.sdAm[2,2] 0.4117450 0.4652302 0.5155276