Wed, 06/18/2014 - 09:17

Hi, all.

I have some issues with the CIs in OpenMx_1.4-3532 and OpenMx_2.0.0-3575. I am fitting a random-effects meta-analysis with two effect sizes. My model was generated from the metaSEM package (metaSEM.model in the attached file). Another version was specified directly in OpenMx (mymodel in the attached file). This model worked fine. It was only used for comparisons. Both versions were supposed to be equivalent. The loglikelihoods and the parameter estimates were the same, while there were some issues on the CIs.

The files of test4a were based on OpenMx_1.4-3532. The lbound and the ubound of the CIs in metaSEM.model were the same.

The files of test4b were based on OpenMx_2.0.0-3575. The metaSEM.model worked fine with CSOLNP as the optimizer. The metaSEM.model generated NA in both the lbound and ubound with NPSOL as the optimizer. When I rerun the model, it threw an error.

Although these two models are equivalent, the model generated by the metaSEM package is more complicated in terms of model specification and number of matrices involved. This may partly explain the differences in the behaviors.

Any suggestions are welcome. Thanks in advance.

Mike

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

test4b.R | 2.02 KB |

test4a.R | 1.58 KB |

test4a.txt | 6.79 KB |

test4b.txt | 7.34 KB |

test4b output.txt | 15.93 KB |

test4a output.txt | 7.63 KB |

Regarding the first 2 models in test4b, I obtain the same -2LL but very different CIs. You comment in the file that "It seems fine." Is this expected?

confidence intervals:

lbound estimate ubound

Tau2_1_1 0.004626727 0.004727034 0.004668791 ** estimate outside of CI

Tau2_2_1 0.001193085 0.003934128 0.005295539

Tau2_2_2 0.008317209 0.008413367 0.008421468 ** estimate outside of CI

Intercept1 -0.017700266 0.001350473 0.028866220

Intercept2 0.035016073 0.068825948 0.096814068

-2 log likelihood: -161.9216

confidence intervals:

lbound estimate ubound

Tau2_1_1 0.002101854 0.004727078 0.009489479

Tau2_2_1 0.001193054 0.003934105 0.008349120

Tau2_2_2 0.004601975 0.008413355 0.015237170

Intercept1 -0.026791136 0.001350212 0.028866220

Intercept2 0.035016079 0.068826046 0.102484186

-2 log likelihood: -161.9216

With respect to test4b.R,

It is hard for me to believe, but if you clear the lower bounds on the Tau matrix before you re-run the NPSOL metaSEM.fit model then it works,

## No CI reported.

metaSEM.fit <- mxRun(metaSEM.model, intervals=TRUE)

summary(metaSEM.fit)

metaSEM.fit$Tau$lbound[,] <- NA # add this line!

## Throw an error when rerun the model

metaSEM.fit <- mxRun(metaSEM.fit, intervals=TRUE)

Maybe somebody with the NPSOL source code can look at lssetx_ and see if there is something obviously wrong with the bounds checking.

Thanks for pointing out this.

No, I don't expect them to be different. They should be the same in theory.

test4b also seems to work when you change the lower bound to 1e-6, -1e-10, or 0.

However, re-running the original model with intervals=FALSE still generates the error. I think this is a re-running problem.

When it runs without an error but gives no confidence limits, are they all

`NA`

? If so, what are the confidence interval codes? When I checked them, they were all -1, indicating that optimization was aborted prematurely. This might be happening if the initial solution is at or very near the boundary of the parameter space. Then, the optimizer just finds`NA`

s on the first iteration of every attempt to find confidence limits.Zero definitely does not work. Take care to only set the lbound on the diagonal elements,

metaSEM.fit$Tau$lbound[1,1] <- 0

metaSEM.fit$Tau$lbound[2,2] <- 0

-1e-10, -1e-3 both fail, but -1e-2 works.

Thanks for looking into the matter.

Mike