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

Re: st: RE: convergence problem

From   Maarten buis <[email protected]>
To   [email protected]
Subject   Re: st: RE: convergence problem
Date   Fri, 8 Jan 2010 05:31:15 -0800 (PST)

--- On Fri, 8/1/10, Rosenstock, Summer E. wrote:
> It looks like my estimates are coming in near the boundaries,
> which has shown to be a problem with log-binomial regression
> with continuous independent variables. THere are a couple of
> articles that document this. So, looks like I'll stick to OR
> on that particular variable with a foot note about
> convergence.

There are alternatives. Such non-convergence in these models is 
often a sign that the effect of your continuous variable is not 
linear in the log(risk). One solution is to allow that variable 
to have a non-linear effect.

I like (linear) splines for that, as you can still easily interpret
the coefficients. So in the example below for the model that
converges: the basline risk of union membership for someone from
the north with 0 wage is .04, this risk decreases by 35% 
(=100*(1-.65) if someone comes from the south, the risk increases
43% for every dollar increase in wage if the wage is less 5 $/hour,
increases by 11% for every dollar increas in wage if the wage is
between 5 and 10$/hour, and decreases by 3% for every dollar 
increase in wage if the wage is more than 10$/hour.

*--------- begin example ------------
sysuse nlsw88, clear
gen baseline = 1

*does not converge
glm union south wage baseline,  ///
    link(log) family(binomial)  ///
    iter(50) eform nocons

mkspline w1 5 w2 10 w3 =wage
glm union south w? baseline,    ///
    link(log) family(binomial)  ///
    eform nocons
*----------- end example ------------
( For more on how to use examples I sent to statalist see: )

Hope this helps,

Ps. I know it is an old complaint of mine, but I haven't made
it in a while, so I'll indulge in it again: I think that the
-eform-, -or-, -irr-, -rrr-, etc. options should not suppress
the display of the constant. These are important in judging
the size of the effect. Knowing that the odds, risk, incidence 
rate, or relative risk doubles for a unit change in x can be
meaningfully supplemented by information about the baseline 
value: twice a very small number is still a very small number 
but twice a large number is a huge number.

Maarten L. Buis
Institut fuer Soziologie
Universitaet Tuebingen
Wilhelmstrasse 36
72074 Tuebingen


*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index