[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: Re: clogit
The command -clogit- means Conditional Logit... this is not exactly
fixed-effects models as we have in the linear regression [Check the math in
a textbook]. The logit function allows us to "remove" the fixed-effects
component in the cases that you observe different pattern in your dependent
variable (let's say y). This means that all cases where y does not change
over time are useless (as you saw in the error message).
Why? y can be 0 or 1, simplify your life using a linear model with fixed
effects: y(i,t)=b*x(i,t)+a(i)+e(i,t). In the within variation you will see
that y(i,t)-y(i,.) is zero for the individuals with not change in y.
Then -clogit- drops these observations as -xtreg- will drop too.
If your x(i,t) is uncorrelated with a(i) you could use random-effects
model -help xtlogit-, remember to check the quadrature after the estimation
to be sure that the numerical approximation is stable, -help quadchk-
----- Original Message -----
From: "Caroline Xu" <firstname.lastname@example.org>
Sent: Saturday, April 22, 2006 11:29 PM
Subject: st: clogit
I use "clogit" to control some individual fixed effects and got this
"note: multiple positive outcomes within groups encountered.
note: 1378 groups (1399 obs) dropped due to all positive or
all negative outcomes."
The number of observations dropped is almost 1/3 of the total number
of observations. Will this cause any problem on the estimated results?
After control for the individual fixed effects, the PseudoR2 increase
from 0.12 (logit model) to 0.60. I'm not sure what this implies? It
means unobserved individual effects are very important for making
prediction? or Is this increase in PseudoR2 caused by the drop of # of
If I want to control for both year and individual fixed effects in the
logit model, what should I do?
Thanks a lot,
* For searches and help try: