[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]
st: gllapred vs gllasim take two
This question follows up on a previous inquiry about the difference between gllapred and gllasim but with a few additional details. I'm estimating a probit model on panel data and would like to use the results to simulate outcomes over different values of one of the explanatory variables, holding the other variables fixed at their mean. My question is whether gllapred or gllsim is a better tool for this (and my hunch is the latter).
Specifically, I'd like to do the following. 1. Estimate a panel probit
gllamm y age other1 other2, i(persid) family(binom) link(probit) nip(20)
2. Holding other1 and other2 fixed at their mean values, generate
predicted values of y over different integer values of age, say, ranging
between 18 and 65.
To implement step 2, I've created an artificial data set in which other1
and other2 are fixed at their mean values and age varies over the range
of interest. So I first estimate the model on the real data. Then I open
the artificial data and type:
gllasim pred1, linpred fsample
The above seems like the right approach given my objective, but I've noted a few interesting things:
1. the approach doesn't work using gllapred unless the dependent variable is included in the artificial data (which I guess has to do with the fact that gllapred includes empirical Bayes).
2. gllapred produces the same answer when used repeatedly, whereas gllasim always produces a different answer, suggesting that the latter may be sampling from a distribution of the parameter estimates
I guess point 2 would actually be another good reason to use gllsim in my case. I could set up a simple code to implement the command, say 1000 times, and, after taking the mean, I'd have something akin to a Monte Carlo simulation of a predicted value.
One problem is that there is very little documentation of gllasim in the manual or elsewhere, so it's hard to know if my idea is even in the ballpark. Any insights offered would be greatly appreciated.
Colin Vance, Ph.D.
German Aerospace Center