Pooled Time-Series Cross-Section Estimation

PROGRAMS ARE AVAILABLE to facilitate calculation of least squares and instrumental
variable coefficients and related statistics from data consisting of time-series observations
for each of several cross-sectional units.' The (FORTRAN) programs2 must be used with
the SPEAKEASY/FEDEASY processor3 and take the form of three FEDEASY commands
that (1)form a moment ("cross product") matrix from an arbitrarily large number of
observations (CPMATRIX), (2) calculate least squares and instrumental variable coefficients
and attendant statistics for any subset of variables included in the moment matrix
(CPOLSQ and CPINST), and (3) decompose a square matrix into two upper triangular
factors by a Cholesky decomposition (DECOMPOSE). The presence of these programs in
the FEDEASY vocabulary allows users to easily construct, using SPEAKEASY matrix
commands, routines appropriate for their particular error covariance assumptions (e.g.,
hypotheses of nth order autoregression, block diagonality, etc.). In addition, three pooling
routines are available both for immediate use and as a guide in coding special applications.
The first (KMENTA), described in [2, pp. 508-5121, as "cross-sectionally heteroskedastic
and time-wise autoregressive" is a standard block diagonal covariance design that is
computable for any number of years and cross-sections. The second (HAVENNER)
emphasizes cross-sectional covariances by calculating a full covariance matrix between all
cross-sections at each point in time, while assuming first-order autoregression of errors in
the same cross-~ection.T~h e third routine (SWAMY) assumes that the coefficients are
random and essentially follows the specification in [3, p. 3121, except that the equation
errors of each cross-sectional unit are assumed first-order a~toregressive.A~l l three
routines are written in the high-level SPEAKEASY/FEDEASY command language and
are easily modified to suit particular cases. All have standard options such as residual and
moment matrix printing, and the first two have iteration limits, convergence criterion
overrides, and restart provisions.
The programs are available on request with the SPEAKEASY/FEDEASY System from
Stan Cohen, Speakeasy Center, Argonne National Laboratory, Argonne, Illinois 60439 or
separately (if the user already has SPEAKEASY/FEDEASY installed) from Ronald
Herman, Board of Governors of the Federal Reserve System, Washington, D.C. 20551.