Standar Errors for variance components of non-linear mixed effect models

Hello!

I am estimating a non-linear mixed effect model using nlmefit, and although Matlab produces standard errors for the estimated betas, it does not produce standard errors for the variance components (the PSI matrix). How can I obtain/estimate the standard errors for the variance components? Would it also be possible to estimate the s.e. for the covariance components of the PSI matrix?

Any help will be really appreciated from the bottom of my heart.

Thank you in advance,

Answer 1

[Jeff Miller]

One way is to use the bootstrp function. For each bootstrap sample, you would write your own function to call nlmefit and save the output PSI matrix values. After you've done that for many bootstrap samples, you can get s.e.'s using the stored PSI matrix values from the individual bootstrap results (e.g., Wikipedia ).

[Alexis Villacis]

Thank you for the tip! Will definitely take it into account.

In R, you can do it by using the confidence interval of these estimates and I was just wondering if there was something similar in Matlab, i.e. Is it possible in Matlab to generate the confidence intervals for the variance components (the PSI matrix)?

I am running 6 different models using LaPlacian approximation and each model takes around 15 minutes in my computer (macbook pro 2.9 GHz intel core i5 8Gb Ram). Generating enough draws using bootstrapping will take a long time.

Any thoughts?

[Jeff Miller]

Sorry, I have no helpful thoughts. I don't know enough about these models to know how the R routines get confidence intervals. Some procedure based on Fisher information maybe?

Did you look at nlmefitsa? It might use the same algorithm as the R package.