Scobit - skewed logistic reg

[Alex Jeffries]

Apologies I wrote a previous post about clogit when it should have been scobit.

I'm doing probit analysis on a very large dataset, and the 0 1 event I'm modelling is rare. I came across King (2001) discussing rare events and the issues with normal probability models, so I have found the skewed logistic regression in stata.

Is it valid to use scobit on panel data? (Theres no xtscobit).
Also the command doesn't allow time serious operators. I've got no issue generating lagged variables, but what is the reason you cant use L.variable notation?

Regards,

[Nick Cox]

In Stata an xt command that allows time series [sic] operators is typically also one where any inferential machinery takes account of that panel structure and/or declared serial structure.

Not supporting such operators is a (subtle) flag is that you can't expect any more recognition of panel and/or time structure than you supply yourself. So, watch out!

The extent of my knowledge is knowing that you are lurking on the border of a dangerous country; I won't be able to answer any supplementary questions on what you might do instead.

[Richard Williams]

I don't trust scobit one bit. The few times I have tried it I have had terrible times getting it to converge. My theory is that the main reason it is in Stata is because somebody happened to write a program for it (kind of like why rreg happens to be in Stata). But maybe I underestimate it. Does it get used very much by researchers?

[Alex Jeffries]

I don't trust scobit one bit. The few times I have tried it I have had terrible times getting it to converge. My theory is that the main reason it is in Stata is because somebody happened to write a program for it (kind of like why rreg happens to be in Stata). But maybe I underestimate it. Does it get used very much by researchers?

Interesting to hear this. I haven't come across anybody using it before so this may well be the reason why. In cases where you thought a scobit was appropriate, what methods have you used instead Richard?

[Alex Jeffries]

Interesting to hear this. I haven't come across anybody using it before so this may well be the reason why. In cases where you thought a scobit was appropriate, what methods have you used instead Richard?

Edit: After running a few different scobits i'm fairly satisfied with the output I am getting, the odds ratios appear to be more realistic given my data and anticipated effects.

[Richard Williams]

The 1994 article by Nagler that the Stata Manual cites has (according to Google Scholar) only been cited 142 times. I don't think that is a whole lot for a piece that old. The manual also warns that convergence can be a problem. Perhaps there are times when something like scobit would have been a good idea, but I've always used methods like logit or heteroskedastic logit/probit instead.

[Rich Goldstein]

well, let's think for a minute about logistic regression - there are not many assumptions but an important one is "symmetry" (i.e., think of a logistic curve with "x" on the X axis; start at the middle of the X axis; as you move in either direction, the effect is the same); this assumption is often not true; my preferred way of looking at this: estimate a clog log; however, this is only non-symmetric in one direction; to check the other direction, "reverse" the 1's and 0's (this has a name also but I seem to be blanking on it <grin>); scobit tests symmetry slightly differently; however, I am not a fan since it often does not converge