Convergence issue: general question

[qh...@wisc.edu]

Dear all, 

I have a general question about convergence issues using GSEM. We've been encountering a number of convergence problems in our model fitting of ~ 12 traits and 2-5 latent factors. One thing we are wondering is if we could adjust iterations to help the model converge, or if iterations is at all important here. So far I do not see an option to change iterations in any GSEM functions, but I might have missed something. 

Any input is helpful!

Best,

Frank

[Elliot Tucker-Drob]

Hi Frank,

I have tended to find that in those cases, either 1) the model is underidentified (or perhaps overidentified in cases where people accidently put in unit variance and unit loading identification instead of one or the other), 2) the genetic covariance matrix is wonky (e.g. lots of very low genetic correlations; lots of extremely high genetic correlations, nonpositive definite genetic covariance matrices that get dramatically smoothed), or 3) the model is good and the data are well-behaved but the default start values aren't close enough to the optimal solution. If you can rule out numbers 1 and 2, I suggest that you think about whether you can supply more sensible start values.  Supplying better start values will usually obviate the need to increase number of iterations, and will often fix things when increasing iterations doesn't.

Cheers,

Elliot

--
You received this message because you are subscribed to the Google Groups "Genomic SEM Users" group.
To unsubscribe from this group and stop receiving emails from it, send an email to genomic-sem-us...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/genomic-sem-users/d96cb599-af9c-4272-867d-b6a5bda8fbd8n%40googlegroups.com.

[Stephen Wang]

Hi Elliot,

I'm wondering how to supply more sensible start values using Genomic SEM? And what's the default start value if I did not supply a start value?

Best,

Stephen

[Elliot Tucker-Drob]

Hi Stephen,
See starting values here: https://lavaan.ugent.be/tutorial/syntax2.html

I can't tell you off hand what the algorithm for the default starting values is, but I believe in many cases it's going to be loadings and variances of 1.0.

How to choose starting values depends on your model and data. For example if your S matrix has diagonals (variances, i.e. heritabiliteis) that are very small, then you should probably specify start values for variances that are much lower than 1.0. If you have some variables that are negatively genetically correlated with one another, you would probably benefit from having some factor loadings or regression coefficients set to have negative starting values. When using unit loading identification you might also switch the reference indicator to the one with the largest genetic correlations with the other variables in your model. Choosing sensible identification strategies and starting values can be a bit of an art.

Elliot

To view this discussion on the web visit https://groups.google.com/d/msgid/genomic-sem-users/d14614e8-55ad-4d33-8ddd-0f792e236e93n%40googlegroups.com.