Hello!
I conducted a mutlilevel analysis with a binary outcome (0/1) in HLM 7 and read that the Adaptive Gaussian Quadrature (AGQ) method leads to better estimates than REML. The model converges, but below the table with the final estimations of variance components in the output its written "Unable to compute AGQ T(hat) for likelihood".
Does that mean I should use a higher number of quadrature points?
So far I've used 100 iterations and 12 quadratur points.
Which estimations might still be biased? The manual of HLM 7 does not provide an explanation for this error term.
Thank you in advance!
Kind regards, Josefine
Do you think it can be estimated eg not a very rare event for the response - no collinearity in the predictors, not very bad imbalance? If you then think things are are Ok , incease the limits incremetality to see wthat happens.
If no joy, consider swiching to MCMC estimation methods rather likelihood based; see
Article Mixed Effects Models are Sometimes Terrible
as to why this may help.
The suggestion given by Kelvyn Jones will be helpful.
Thanks to Kelvyn Jones.
Best wishes to Josefine Atzendorf.
- Badges
- Science topic
- Similar topics
- Statistics
- Survey