Alternatives to xtgee?

[Jeff Wooldridge]

This is something that has bothered me for many years about built-in Stata commands for panel data. For small T, large N, if I want an alternative to to xtreg, re to allow for an unrestricted T x T within-unit variance-covariance matrix, I seem to be limited to xtgee. There are some important limitations of xtgee. For one, while one can specify corr(uns) to allow an unrestricted T x T correlation matrix, the variances are assumed constant across time. One can get around this by weighting all variables by the inverse of a time-period specific standard deviation, but this is clunky. Even if one does that, the estimation underlying estimation of GEE has its problems. When I run a simulation with high AR(1) serial correlation (say, rho = 0.75), the number of iterations where xtgee fails can be as high as 20%. If GEE were implemented with a two-step feasible GLS procedure, it would almost always run -- the exception being the extremely rare case where the estimated correlation matrix is not positive definite.

I don't think I've seen a user-written command that implements an unrestricted FGLS estimation for the small T, large N case (and allows for fully robust inference). I'd be happy to learn otherwise. Currently, xtgee is closest but doesn't do enough, and computationally, it's unnecessarily hard.

[Stephen Jenkins]

Jeff Wooldridge : please also post this great request to https://www.statalist.org/forums/forum/general-stata-discussion/general/1711202-wishlist-for-stata-19 (Wish list for Stata 19)

[Jeff Wooldridge]

Will do, Stephen! One of the reasons I'm interested in such a command is that it would provide a simple method to improve efficiency over the extended TWFE estimator for staggered interventions, as in xthdidregress twfe. With time-constant controls, it's the same as random effects on an expanded equation. And then rather than RE, I could perform unrestricted GLS to possibly improve efficiency.

I think adding an option to xtgee, such as "fgls," would be very easy. It would be understood that the specific correlation matrix is to be used and, in addition, the variances are estimated separately in each time period, and these are used to construct the T x T variance-covariance matrix. The correlation matrix would be estimated using the usual pairwise sample correlations rather than the more complicated GEE nonlinear moment equations.

[Violet Nguyen]

Jeff Wooldridge please reply my message. I am really struggling of this problem and really hope to see your assistance!

[Sapna Goel]

Jeff Wooldridge hello Professor. I am having some doubts in the fixed effects model. It would be really kind of you could help me out here.
I am facing problem in interpreting the coefficients.
Let me explain my model first:

I am trying to understand the effect of household level monthly migration status on the number days spent by left behind women on farm.
Household level monthly migration status is a binary variable. It takes value 1 if any member in the household migrated in that month, 0 otherwise.

So, the fixed effects coefficient essentially tells me that with-in a given household, if household level migration happens , then how the number of days spent on farm by women would change.

Now, my doubt is : how are these observations treated?

1) those households who never migrated, for them the household level migration status is always zero. How do we treat them?

2) Suppose some household migrated for consecutive months, then do we also consider the change in farm days for these consecutive months?

[Sapna Goel]

hello Professor. I am having some doubts in the fixed effects model. It would be really kind of you could help me out here. I am facing problem in interpreting the coefficients.
Let me explain my model first:

I am trying to understand the effect of household level monthly migration status on the number days spent by left behind women on farm.
Household level monthly migration status is a binary variable. It takes value 1 if any member in the household migrated in that month, 0 otherwise.

So, the fixed effects coefficient essentially tells me that with-in a given household, if household level migration happens , then how the number of days spent on farm by women would change.

Now, my doubt is : how are these observations treated?

1) those households who never migrated, for them the household level migration status is always zero. How do we treat them?

2) Suppose some household migrated for consecutive months, then do we also consider the change in farm days for these consecutive months?

[Carlo Lazzaro]

Sapna Goel:
please start a new thread. Thanks.