analytical weights and logit/mlogit/ologit models

[William Lisowski]

I noticed today that the logistic regression commands mentioned in the title to this topic included the followig in their help files.

fweights, iweights, and pweights are allowed; see weight.

In an earlier topic found at this link there is a suggestion that support for analytical weights for logit was removed in Stata 10.1, and indeed logit throws an error message when analytical weighting is requested.

However, mlogit is happy to provide results using analytical weights for a binary outcome variable - and in both ologit.ado and mlogit.ado the syntax includes analytical weights.

This seems ... inconsistent.

Can anyone point in a direction that will (a) increase my understanding of the issue and/or (b) tell me if the concerns for logit also apply to mlogit and ologit, regardless of the fact that analytical weights do not throw the same error as for logit.

[Leonardo Guizzetti]

I'm not at my computer right now, but might it be that logistic regression executed by -glm- handles analytic weights as an alternative?

[William Lisowski]

That's probably the case. But I should confess that my objective is to get support for not weighting an ologit regression that I'm working on. The hint in the earlier thread, and the removal of aweight from logit, suggests that weighting an ologit may be a similarly bad thing to do. Now I just need to find out if that is true, and if so, why, and then I'm free to not weight.

[Leonardo Guizzetti]

A few observations and thoughts:

* -mlogit-'s allowance of aweights is not documented, and therefore may remain due to an oversight as it is not allowed in -logit- or -ologit-

* -aweight-, as I have only used it, are (proportional to) inverse-variance weights. Therefore, I don't think it's credible to suggest an individual with an observed categorical outcome is associated with an variance. However, it would make sense to use -aweight- if one were adjusting for clustering or using grouped observations (e.g., meta-analysis), in which case -regress- or -glm- would be suitable for a weighted least-squares type of estimation. This idea seems to be suggested by both Paul Allison's and Joseph Hilbe's books on logistic regression.

* -iweight- is allowed, so one can simulated -aweight- by ensuring the weight is scaled to sum to the estimation sample size.