Holtz-Eakin-Newey-Rosen example

[TomDoan]

This is an example of the techniques described in Holtz-Eakin, Newey and Rosen(1988), "Estimating Vector Autoregressions with Panel Data," Econometrica, vol. 56, no 6, pp 1371-95. This is an example provided by the authors, but isn't the same data set as used in the published paper. It estimates a bivariate VAR with panel data on local government employment data (full- and part-time workforce per capita) with wage rates for each type of employee treated as pre-determined variables. This allows for individual effects, that is, the intercepts are allowed to vary across individuals, though the lag coefficients and the covariance matrices are assumed to be fixed. It includes fixed time effects in the form of dummies.

It's well known that the straightforward LSDV estimation (what you would get with PREG(METHOD=FIXED) is subject to bias that is a function of the T dimension only, which is here quite small. To avoid this problem, the authors apply what's now known as the Arellano-Bond estimator; using a large number of instrumental variables in a regression on first differences.

This does single equation estimates using 2SLS, single equation GMM with allowing for general serial correlation, 3SLS and systems GMM allowing for general serial correlation, all with a reduced set of Arellano-Bond instruments. This is a data set with a fairly large N dimension (161) and small T (six usable data points per individual). If you have a data set with a much smaller N, the GMM estimators may not be feasible since the number of orthogonality conditions may exceed N.

This requires a revised version of the @ABLAGS procedure (http://www.estima.com/forum/viewtopic.php?f=7&t=1116).

ratslabor.rpf

Program file

(2.46 KiB) Downloaded 184 times

labor.xls

Data set

(639.32 KiB) Downloaded 473 times

[luxu1983]

dear
the author use three steps to get the result

[TomDoan]

That's what the 3SLS is at the end. You don't have to do the steps yourself.

[luxu1983]

That's what the 3SLS is at the end. You don't have to do the steps yourself.

in your meaning
3sls at the end is to get the gls estimator
am i right?

is it necessary to add the option "update=continuous,zudep"?

[TomDoan]

That's what the 3SLS is at the end. You don't have to do the steps yourself.

in your meaning
3sls at the end is to get the gls estimator
am i right?

Yes. The SUR at the end is the GLS estimator.

is it necessary to add the option "update=continuous,zudep"?

They aren't "necessary". You get a consistent estimator without them. It is, however, what they recommend.

[luxu1983]

can i get single equation GLS estimator rather than using SUR

[TomDoan]

The single equation GMM (not GLS) estimators are done with:

Code: Select all

linreg(inst,optimal) dpart
# constant pdummy dfwage{0} dpwage{0} dfull{1} dpart{1} dfwage{1} dpwage{1}
linreg(inst,optimal) dfull
# constant pdummy dfwage{0} dpwage{0} dfull{1} dpart{1} dfwage{1} dpwage{1}

[cv003h]

Hi Tom. Thanks so much for posting an example from Holtz-Eakin et al. I am relatively new to RATS, and am wondering how I would alter your code to accommodate an unbalanced panel with missing observations? Think 10,000 firms, 30 years of annual obs, and differing start/end dates per firm with not always every year in between.
Thanks!
Chris

[TomDoan]

There's nothing about the way that this works that requires a balanced sample. You just have to get the data into RATS in a balanced form with missing values padding the short time series. The missing values within an individual record are more of a problem. First, off you're estimating a VAR, so you would lose (with one lag) two observations for each of those, since you are missing the lag as well. Then, the first difference operator to get rid of the individual effects costs another one. The instruments for an individual like that are going to be a patchwork as well. It might make more sense to filter out the individuals that have embedded missing values.

[luxu1983]

dear
How to set instrumental variable, so that each equation is just identified?

[TomDoan]

dear
How to set instrumental variable, so that each equation is just identified?

That's the Anderson-Hsiao estimator which has very poor properties. The whole point of the Arellano-Bond instruments is that the lagged values are all fairly weak instruments, so you need many of them to get good results.

[cv003h]

There's nothing about the way that this works that requires a balanced sample.

I guess I was thrown by the line "calendar(panelobs=8) 1973". Does this imply that every "firm" begins in 1973 and runs for exactly 8 years?

[TomDoan]

There's nothing about the way that this works that requires a balanced sample.

I guess I was thrown by the line "calendar(panelobs=8) 1973". Does this imply that every "firm" begins in 1973 and runs for exactly 8 years?

In the RATS panel data structure, each individual has the same size block of entries. However, they don't all have to have data at the same time periods.

[cv003h]

Ok, got it. Thanks Tom. I'll give it a shot!

[cv003h]

This does single equation estimates using 2SLS, 2SLS with heteroscedasticity robust standard errors and joint estimation using 3SLS, all with the Arellano-Bond sets of instruments.

Tom: Do I need to leave in both 2SLS estimates, or do I choose only 1 for my final code? And when I run the code, I get an error stating:

## CP18. ABLAGS is not the Name of a PROCEDURE. (Did you forget to SOURCE?)
>>>>@ablags(<<<<

Any idea what is going wrong? Thanks.