Bookmark and Share

Notice: On April 23, 2014, Statalist moved from an email list to a forum, based at

[Date Prev][Date Next][Thread Prev][Thread Next][Date Index][Thread Index]

Re: st: Re: XTGEE or XTLOGIT with rare events

From   Nick Cox <>
To   "" <>
Subject   Re: st: Re: XTGEE or XTLOGIT with rare events
Date   Tue, 4 Jun 2013 14:14:00 +0100

My main purpose was to get you to sharpen up your question and now
others who are expert in this territory may find it a bit easier to
say something useful. But there remains some fuzziness over what kinds
of comments you think we can deliver.

"[A]re the results reliable and is it justified to use these models?"
An oracular "Yes" or "No" is presumably not what you are expecting
here. I think you would have to post some results for people who have
used these models to comment.

The probability of rare events is just what it is; you only need
-summarize- for that, apart from sampling issues.


On 4 June 2013 14:03, Kamyar Baradaran <> wrote:

> Thank you very much Nick, sorry for not clearly explaining my issue.
> the issue is as Gary King and Langche Zeng discuss,  "most popular
> statistical procedures, such as logistic regression, can sharply
> underestimate the probability of rare events". I have used -xtlogit-
> and -xtgee-, my models converge and I do not have any problem, but are
> the results reliable and is it justified to use these models?
> As per the selection layer here is the logic: we first model the
> probability of an event (a 1 in the dependent variable), and then
> conditional on that, estimate the effects of interest on occurrence of
> an event. I am just a student and trying to learn and I hope this
> makes sense. If it makes sense, is there anyway to do it in Stata with
> longitudinal data?

 From   Nick Cox <>

> You cannot apply programs that do not exist, but have you tried using
> -xtlogit- or -xtgee-?
> In broad terms, the rarity of events tends to mean that models are
> more difficult to fit, but not necessarily impossible. I wasn't aware
> that adding a selection layer made anything easier.
> By -relogit- you perhaps meant to refer to
> Without knowing anything personally, it seems a fair guess that that
> Stata program is not going to be developed further by that group.

On 4 June 2013 12:59, Kamyar Baradaran <> wrote:

>> I have a rare binary dependent variable (most of the time zero and
>> rarely 1). My dataset is longitudinal and to my knowledge "relogit"
>> and "heckman" selection models are not yet developed for longitudinal
>> (Am I correct?). Could you please advice me how to deal with this
>> issue?
*   For searches and help try:

© Copyright 1996–2018 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   Site index