running CIs on a common path twin model, most of the intervals make sense, but some (see example below) are not.
These are all paths which when dropped significantly reduce fit, but it is as iff the optimiser has stepped over zero looking for a worse value, found itself on a nice slope, and concluded that both, for instance + 0.604 AND –0.604 are fine, without seeing that 0 and nearby values are not.
confidence intervals:
lbound ubound
top.a[1,1] 0.6277528 0.7943150
top.as[3,3] -0.6048829 0.6048829
top.cs[1,1] -0.5030336 0.5030336
top.es[2,2] -0.5443131 0.5443131
#1
Yes, the optimizer will do this fairly reliably. I normally lbound at zero those parameters that are invariant to sign. In a Cholesky this would be the diagonal elements; in a factor loading matrix normally the first loading in each column should be lbounded.
Log in or register to post comments
#2
Sounds like the solution is to put a lbound on the free para
meters.
Log in or register to post comments