Bronwyn H. Hall

                                                        March 1993



                    Documentation for GMMPANEL Example. 



     This example comprises 6 TSP runs, plus 3 TSP input files which 

are used by these runs.  All of the runs have been tested on TSP 4.2B 

on a Compaq 386/25 computer.

     

     All of the runs use the same sample dataset, which is included 

with the examples in compressed form due to its size.  Before running

any of these examples, give the command



      USBAL



at the DOS prompt, which creates USBAL.DAT.  Then USBAL.EXE may be

deleted.



     The sample dataset consists of data 

on 495 US manufacturing firms for 10 years (1979-1988). The data is

sorted with one record per firm-year, first by firms, and then by year

within a firm. This enables it to be read either as a pooled time 

series-cross section with one year per observation, or as a panel with

all years together (see the Panel Data chapter in the TSP User's Manual

for a discussion of how to read panel data). The file contains data

on sales, labor, capital, R&D, etc. for each of the firms, but these

examples only use deflated sales, labor, and capital in a very simple

constant-returns-to-scale Cobb-Douglas production function estimation.

This example is not intended to be realistic, or a reasonable model of

firm behavior; it is solely for illustrative purposes. 



     The examples typically only use the last 3 or 6 years of the data. 

A description of each example follows:



     USBALFE reads the data as a pooled TS-CS and estimates the standard 

panel data models with and without lagged X's.  There are tests for the 

constancy of the parameters across firms, and correlated effects. This

run also shows how to create and store variables which are constant

across time, and how to remove time means from all the variables. 



     USBALPI performs basic PI matrix estimation of a simple fixed

effects model with a single x and time-varying coefficients on the

firm effect. This example is discussed in Chamberlain, Journal of

Econometrics (1982), and also in the Panel Data chapter in the Handbook

of Econometrics (North-Holland).



     USBALGMM estimates the same models as USBALPI using the GMM 

technique popularized by Hansen (Econometrica 1982). Like the PI matrix

method, these estimates are asymptotically efficient even if the 

errors are heteroskedastic. 



     USBALPIM uses the PI matrix methodology to estimate a model with

measurement errors, as an example. This may not be the easiest way to

estimate such a model - the example is included as an illustration. The

estimated variances of the measurement errors in the case here are

negative (insignificantly different from zero), so the estimates do 

not compare well to those in the next run.



     USBALME uses GMM to estimate with serially uncorrelated measurement

error in the X's. The model is the same as for USBALPIM, but it is 

easier to set up, and the estimates differ in some cases, probably

because of the variance problem mentioned above. The methodology here

is from Griliches and Hausman, Journal of Econometrics (1986). 



     USBALGM4 uses GMM to test for strong and various kinds of weak

exogeneity of the X's in the model given in the earlier runs.  It 

illustrates how to use the MASK= option in GMM to exclude some instruments

from some equations.



     The supporting input files for these jobs are the following:



     RDUSBAL reads the data as a panel; it is used by all of the runs

except USBALFE. It also constructs the dependent and independent variables

and removes their year means (equivalent to including time dummies in a

linear model). 



     FITEST calculates a goodness-of-fit statistic after estimation

of a PI matrix model, and returns the chi-squared value and its degrees

of freedom, so that these can be used in further testing. This PROC

also displays the covariance matrix of the residuals computed at the 

estimated parameter values.



     GMMCHISQ calculates the chi-squared test of the overidentifying

restrictions after a GMM estimation, and returns it along with the

degrees of freedom, so these can be used in further testing (of one

set of restrictions which are nested within another). It also displays

the covariance matrix of the residuals at the estimated parameter 

values.