www.cmm.bristol.ac.uk/forum

Hello,

I carried out a multinomial multilevel analysis in MLwiN, by using 2nd order PQL. My dependent variable is contraceptive use and consists of three categories:
(1) using no method (n = 3,328)
(2) using natural methods (n = 1,450)
(3) using hormonal/barrier methods (n = 8,693)

The analysis runs just fine when I estimate my models with the third category (using hormonal/barrier methods) as the reference group. However, if I run the same models with the first/second category as the reference group, the models only configure if use 1st order MQL. If I switch to 2nd order PQL, I get the following error: "IGLS/RIGLS NUMERIC WARNING. SSP matrix for fixed part has gone negative definite - a reconstruction to the nearest non-negative form has been used. Numeric warnings can be suppressed by setting an option on the Estimation Control, IGLS/RIGLS panel."

I tried the following solutions:
- Click "Yes" to continue the estimation, but it goes on until my estimates reach number over one billion.
- When I activate the option "suppress numeric warning", my estimates look like this: 1.#Q0(0.000)
- Allowing negative variances (also in the IGLS/RIGLS panel) doesn't work either, I get the same error
- I looked for a solution on your website (http://www.bristol.ac.uk/cmm/software/s ... rrors.html), but I cannot get my head around it.

Can anyone help me with this?

Thank you in advance!
Rozemarijn

Are you using the 1st order MQL estimates as starting values (i.e. using the More button)? This can sometimes help with models that are having problems with convergence. Have you tried fitting the models using MCMC to check that the estimates look sensible there?

Dear Chris,

Thank you for your reply.
I tried using the MQL values as starting values, but that doens't seem to work either. I estimated the models with the MCMC procedure a while ago (only with the third category as the reference group), but that took days to converge. Moreover, the results were very similar to those estimated with 2nd order PQL. Do you maybe have any other suggestions?

Rozemarijn

These models can sometimes be a bit sensitive to starting values, although if you were able to send us an example of a model where this occurs we could check to see whether there is anything else causing the problem. Otherwise you should be able to transform the results using the base category that did work to give the answers you should get for the others. I am not sure why MCMC took so long to converge to a distribution. What settings were you using for this?