Attachment | Size |
---|---|

CI.pdf | 144.31 KB |

Hi, all.

I have a mxAlgebra of a parameter multiplied by a constant, say new_x = 2*x. When I construct an LBCI on both x and new_x, I expect that the CIs on new_x are the same as the CIs on x multiplied by 2. It turns out that they are not exactly the same (see the "diff" in the following output). When the constants are larger, say 5 or 10, the CIs on the new_x even become NA.

Any ideas why this happens? Thanks.

## Multiplied by 2 ## variance Two.two_variance.1.1. variance_x2 diff ## lbound 0.7693044 1.537972 1.538609 -0.0006373491 ## estimate 1.0030992 2.006198 2.006198 0.0000000000 ## ubound 1.3427832 2.688370 2.685566 0.0028032732

## Multiplied by 5 ## variance Five.five_variance.1.1. variance_x5 diff ## lbound 0.7693044 3.840211 3.846522 -0.006311445 ## estimate 1.0030992 5.015496 5.015496 0.000000000 ## ubound 1.3427832 NA 6.713916 NA

## Multiplied by 10 ## variance Ten.ten_variance.1.1. variance_x10 diff ## lbound 0.7693044 NA 7.693044 NA ## estimate 1.0030992 10.03099 10.030992 0 ## ubound 1.3427832 NA 13.427832 NA

Best,

Mike

Hi, Mike. I reproduce what you report. If you pass argument

`verbose=TRUE`

to`summary()`

, you can see a CI details table which explains the`NA`

-valued confidence limits. They are all due to the change in -2logL from the MLE to the confidence limit differing too much from the expected value of about 3.841. For instance, the change in -2logL corresponding to the upper limit of 'ten_variance' is 4.1023.However, by switching to SLSQP as the optimizer at the beginning of the script, I get no confidence limits reported as

`NA`

, and much smaller differences between the calculated and estimated confidence limits. OpenMx's default behavior is to use an inequality -constrained representation of the confidence-limit optimization problem with SLSQP, but to use a quadratic-penalty representation with NPSOL and CSOLNP. My guess is that multiplying the variance as your script does serves to make the quadratic-penalty representation ill-conditioned.Hi, Robert.

Thanks for the comments and suggestions.

I have compared the performance of the three optimizers. Attached is a real problem I have in using OpenMx to conduct a meta-analysis. I am interested in calculating the CI on Tau/(Tau+s2) where Tau is a parameter and s2 is a constant (0.08486598 in this example).

SLSQP does not work well on both the

`lbound`

and`ubound`

.CSOLNP works okay in the

`ubound`

but not the`lbound`

.NPSOL works similarly as that of CSOLNP.

If we look at the CIs on the mxAlgebra, it is hard to tell which ones, if any, are the correct CIs. Any suggestions? Thanks.

Best,

Mike

You can always validate a profile-likelihood confidence limit as follows. First, make a new model in which you constrain the reference quantity (the thing you provide as value to

`mxCI()`

argument`reference`

) to the confidence limit. If the reference quantity is a free parameter, that can be done by fixing it to the limit; otherwise, you'll need an MxConstraint. Then, run the new model. Finally, compare the -2logL from the fitted new model to that of the original model. If the difference is sufficiently close to 3.841 (for a 95% interval), then you've validated the confidence limit.See the attached script. SLSQP and NPSOL both validate their lower limits. CSOLNP has trouble running the model to validate its lower limit, but I think its lower limit is still trustworthy, since it's approximately the same as the other two optimizers' lower limits. However, SLSQP's upper limit does not validate.

An OpenMx function to automatically attempt to validate confidence limits is a planned feature.

Thanks, Robert.

It's very helpful.

Mike

You're welcome!

hi mike,

can you post the output using

`mxSummary(..., verbose = TRUE)`

? That will allow abetter look a the diagnostics for usHi Timothy,

Here it is. Thanks.

Mike

We need to see verbose summary output for the model from your 05/20/2019 post. This trivial model you have posted is not helpful; the output looks fine.

Please see the attached output with verbose on the 05/20/2019 post. Thanks.

So if we line up the relevant info then it looks like SLSQP is getting stuck in a local minimum,

I don't really see anything wrong here, in terms of bugs. Gradient-based optimizers cannot always find the global minimum.