www.cmm.bristol.ac.uk/forum

I am currently running a four level model using MLwiN version 2.0 (I've just applied for the latest version) and I get the following error message

IGLS/RIGLS numeric warning
SSP matrix for fixed part has gone neagtive 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. A discussion on the possible causes of numeric problems can be fiound under the upic Numeric Problems in the help system.

Continue Estimation? yes/no

I've continuously clicked on yes and at some point I got the same error message for some time and then suddenly it changed the 'fixed part' to the 'random part'. Plus, it would now also give me a numerical error in calculating the likelihhod. However, at the end I obtained some estimates.

I have three questions:
1) what might have caused this problem in my data? I run a similar model specification for another variable and did not have this problem.
2) are the estimates that I obtain at the end reliable?
3) is there anything I can do to correct for this problem.

Thank you in advance for any help and/or comment.

Adriana

Sometimes during estimation some of the matrices will become negative definite. (See this FAQ on the CMM website: http://www.cmm.bristol.ac.uk/MLwiN/tech ... rormessage). If you keep pressing 'yes' and eventually the estimates turn green, then that should be fine: you can have confidence in the results, it was just that they needed to go through some strange combinations to end up with the right answer. I also wouldn't worry that it changed from being about the fixed part to being about the random part- I guess once the fixed part was fine the random part encountered problems, but then the random part became ok too.

However the numerical error in calculating the likelihood is more worrying. I would be wary of results when you have this error. Double check you haven't done anything like overparameterising the model, but note that sometimes there may not be anything actually 'wrong': some combinations of data and model and estimation method just don't work (and there's nothing you can do about it, unfortunately!)

Lydia,
thank you for your reply.

Adriana