[Clyde Schechter]

I'm afraid I can't help you with this one. I'm not familiar with rank discrimination indexes for multinomial models. What I was thinking of is how well the model discriminates each level of the outcome from all other levels. (I'm assuming your multinomial outcome is not ordinal--if it is then you can just look at Somers D.) Applying -roctab- command separately to an indicator of each outcome level against your model's "xb" will give you that.

[natalia malancu]

i was not familiar with the roc curve, but i did a bit of reading. for mlogit models i found mlogitroc (http://econpapers.repec.org/software...de/s457181.htm), but I am quite confused about the whole idea of multiclass roc curves and the generated output which is not at all helping out (ending up with one curve, where I was expecting one for each outcome). any experience with it? much appreciated

[natalia malancu]

Hi Clyde!

can we go back for a second to the age groups question.
I am now dealing with an age variable that comes in the following format: 17-22-27 up to 67. Given the multitude of groups, I introduce it as a continuous predictor.
my question is...is there a need to rescale it in order to interpret its effect or is it correct to leave it as it is and say that one age band increase (where an age band=5 years, yet that does not mean an increase in age by 5 years; I could take an average of 2.5) has this and that effect.

all suggestions much appreciated,
natalia