Dear all,
first I will thank you for any respond and help. I am not a experienced econometrician and no English native, so I apologize if my question might be not very precise, but I did my best
I want to estimate a panel logit model, accounting for fixed effects using a unbalanced panel (T=4 N=56,398). The aim is to investigate the effect of heart diseases (heart=1 if yes, 0 otherwise) on labor force participation (inlf=1 if working, 0 otherwise).
- My first idea was to use the xtlogit command with the option ,fe resulting to my code:
==> As pointed out in http://repec.org/usug2016/santos_uksug16.pdf the estimated marginal effects are conditional on the assumption that the fixed effects are 0, which leads to meaningless results if the assumption does not hold. To assume the fixed effects to be 0 is totally random? and thus the marginal effects are meaningless.
-Then I read about the user written aextlogit command, which solves this problem and transforms the estimated coefficients to "average (semi-) elasticities of Pr(y=1|x,u) with respect to the regressors, and the corresponding standard errors and t-statistics". So I was going with the command
==> The problem here is that my main variable of interest (heart) is a dummy variable and thus there are only discrete changes from 0 to 1 and a interpretation of a elasticity is not meaningful. Also using ln(heart) to get the semi elasticity is not possible, because of the binary nature of heart.
- So both models are not suitable for my investigation, am I right?
-My idea is now to run a random effects logit model with clustered standard errors of the form of
.
which unfortunately does not account for time constant heterogeneity, but in my opinion is the only possibility to use a non-linear model.
-The other alternative would be to use a linear probability model
which is not suitable to predict probabilities (because of constant partial effects, unbounded predicted probabilities and heteroscedasticity) as I want to do.
So my questions are:
- Is it appropriate to first estimate a linear probability model to get a first approximation of the effect and then go to the panel logit regression with random effects?
-Is there any other possibility to control for the fixed effects?
-And can anyone share some citable literature I can use for this argumentation?
I really appriciate any help and comments and thank you in advance.
Best,
Claudio Schiener
first I will thank you for any respond and help. I am not a experienced econometrician and no English native, so I apologize if my question might be not very precise, but I did my best
I want to estimate a panel logit model, accounting for fixed effects using a unbalanced panel (T=4 N=56,398). The aim is to investigate the effect of heart diseases (heart=1 if yes, 0 otherwise) on labor force participation (inlf=1 if working, 0 otherwise).
- My first idea was to use the xtlogit command with the option ,fe resulting to my code:
Code:
xtlogit inlf heart $controlls, fe vce(oim) margins, dydx(heart)
-Then I read about the user written aextlogit command, which solves this problem and transforms the estimated coefficients to "average (semi-) elasticities of Pr(y=1|x,u) with respect to the regressors, and the corresponding standard errors and t-statistics". So I was going with the command
Code:
aextlogit inlf heart $controlls
- So both models are not suitable for my investigation, am I right?
-My idea is now to run a random effects logit model with clustered standard errors of the form of
Code:
xtlogit inlf heart $controlls, re vce(cluster personid)
which unfortunately does not account for time constant heterogeneity, but in my opinion is the only possibility to use a non-linear model.
-The other alternative would be to use a linear probability model
Code:
xtreg inlf heart $controlls, fe cluster (personid)
So my questions are:
- Is it appropriate to first estimate a linear probability model to get a first approximation of the effect and then go to the panel logit regression with random effects?
-Is there any other possibility to control for the fixed effects?
-And can anyone share some citable literature I can use for this argumentation?
I really appriciate any help and comments and thank you in advance.
Best,
Claudio Schiener
Comment