www.cmm.bristol.ac.uk/forum

I am trying to run negative binomial models and have three questions:

1) I would like to add my weight variable in the equation. I am currently running the analyses on individual level only (thus one level). Any idea how to add the weight variable in the models?
2) When more than one level: how to add the weight variable in this case?

3) Also, when running the negative binomial models, I sometimes get the following error-message:
“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. A discussion on the possible causes of numeric problems can be found under the topic Numeric Problems in the help system
Continue Estimation? yes/no?”

Is the result still reliable if I press yes until it turns green?


NB: The demonstration files on the website regarding weights doesnt seem to exist anymore: see http://www.bristol.ac.uk/cmm/software/s ... ights.html


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

Soenita

1) Go to Model -> Weights. If you only have one level you should see just one pair of drop-down boxes. In the one on the left choose your weights variable and in the one on the right any free column (i.e. the default that's already there is probably fine). I'll confess to not being entirely sure whether there's any difference between using raw or standardised weights for a single level model, but if you select 'Use raw weights' you should be fine. Then click 'Done'. If you now run your model your weights will be applied.

2) When you have more than one level you'll need more than one weights variable. With two levels, you'll have one reflecting unequal selection probabilities at level 2 and another reflecting unequal selection probabilities for the individual cases within a level 2 unit once that level 2 unit has been selected; and similarly with more than two levels. Once you have those weights variables adding them to the model is quite straightforward: proceed as described in 1) except that you'll now need to specify your weights variable for both levels and choose a free column for each, and you should choose 'Use standardised weights'. The tricky part is getting the right weights in the first place. If you collected your data yourself you should know how you sampled your cases at each level, and you can calculate the appropriate weights variables using this information and Section 1.7 of the document on multilevel weighting available here. If you're using secondary data it's very common for only one set of weights to be provided even when the sampling was multi-stage, and this is unfortunate. There's often not much you can do about that, except to try writing to the people who originally collected the data and asking them for the two sets of weights (possibly referring them to the document linked above). (Of course, in some cases, if all the variables that determined the unequal selection probabilities are included in the dataset or if the selection probabilities are equal for all units at all but one of the levels, and the sampling design is, as it should be, clearly explained, you may be able to calculate the weights for yourself.) Good luck!

3) Yes, that's ok. The estimates won't turn green until it's stopped being negative definite (and converged), and it doesn't matter that the matrix had to go negative definite along the way so long as it ends up not being negative definite.

Hope I've understood what you were asking - let me know if I've missed something!