Math1 Comput. Modelling, Vol. 11, pp. 253-255, 1988 0895-7177/88 $3.00 + 0.

00
Printed in Great Britain Pergamon Press plc

THE P-VERSION OF THE FINITE ELEMENT METHOD: THEORY AND

APPLICATIONS

ON THE USE OF ESTIMATED STRUCTURAL EQUATIONS FUR PREDICTION: SOME EMPIRICAL RESULTS

Saleh Amirkhalkhali

Department of Economics, Saint Mary's University, Halifax, Canada

Abbas Naini

Department of Economics, Energy Resources Conservation Board, Calgary, Canada

Abstract. The purpose of this simulation study is to address the question of choice
between the reduced-form and the estimated structural equations for purposes of prediction
in simultaneous-equation models.

Keywords. Simultaneous equation model; prediction error variance; constructed variables.

coefficients and ui is a column vector

INTKDDUCTIUN of disturbances with properties

In simultaneous-equation models, E(ui) i) and E(Ui Uj') = oij1.

predicted values of the endogenous
variables are usually made from the Eq.(l) can be expressed concisely in the
reduced-form equations. Accordingly, form:
the direct use of estimated structural
equations appears not to have been Yi’ZiGi t Ui (2)
exploited for purposes of prediction.
The purpose of this simulation studv is where:
to com'pare the performance of-the
estimated structural equations with the Zi = [Vi Xi] and 6i = [ B'iy 'iI'.
COrreSDOndinq reduced-form equations in
the context of prediction. To this end, The corresponding reduced-form of Eq.
the relative predictive efficiency of (2) can be written as:
these two alternatives is examined by
comparing their corresponding estimated yi=x7T i +ui (3)
prediction error variances. Although
most statistical packages can easily where X is the T x K matrix of sample
accorriplish the task of estimation and observations on the predetermined
prediction, they do not provide variables. the vector of
7ifl.'s
information on the prediction error associated toe ficients and vi is a
variance. To overcome this problem, the column vector of disturbances with
use of constructed variables in both the properties
reduced-form as well as the structural
equation is introduced, following E(ui) = 0 and E(u~u~‘) = wijI.
Salkever (1976), Fuller (1980) and Pagan
and Nicholls (1984).
Estimation and Prediction
The paper is developed as follows: the
second section sets out the model. the The coefficients of interest in Eq.(2)
third section discusses the problem of can be consistently estimated by
estimation and prediction. The fourth asymptotically efficient estimators such
section presents the simulation results. as ZSLS, 3SLS and FIML. Consistent
The conclusion is summed up in the fifth estimates of the reduced form
section. coefficients can be obtained either by
aoolvino direct least sauares to Ea. (31
The Model or'by jointly estimating the equation ‘of
interest from (2) and (3) using 3SLS.
Consider a linear model containing G For details see Goldberger, Nagar and
structural relations. The ith Ddeh (1961) and Court (1973).
normalized structural equation can be
written as: In this studv. we examine the oredictive
power of a "s\ngle structuraf equation
yi = Yi&+ XiYit ui; I = 1,2,...G (1) such as (2), with (K - Ki)l_ gi - 1,
estilnated by PSLS with the COrreSDOndino
where Y reoresents a column vectors of T reduced form equation, i.e., Eq. (4j
obser;ations on the ith endogenous estimated by the above mentioned
variable, Yi and Xi are the T x gi and T methods, by comparing their prediction
x ki matrices of sample observations on error variances. To this end, consider
the endogenous and predetermined the following equation:
variables included in the structural
equation (l), respectively. Bi and Yi y=wcY t E (4)
are the vectors of associated

253
254 Proc. 6th Int. Conf. on Mathematical Modelling

matrix of constructed variables and

This equation represents either Eq.(Z)
represents the vector of associated
or Eq. (3) depending on how W is
In the former case: coefficients. The specification of D
defined.
depends on W*. See Pagan and Nicholls
c1=6i=[B;’ $‘I’. (1984).
w=zi=cvi x,1,
Let D=[O'I]', where 0 stands for T x s
E=ui and y=yi,
null matrix and I is an s x s identity
w=x,cr=lri, E=" j matrix. Then Eq. (5) can be written as:
In the latter case:
and y'yi.
Y W 0 a E
= t
Assume that predictions are required for
Yf Wf I 1-1 Ef (6)
yf=(yT+l ,.**yT+s)',
The vector of coefficients in Eq. (6)
with a set of observations for the mat-
can be estimated by minimizing the sum
rix W which is given by Wf. Accordin-
of squares of residuals E*, i.e.,
gly, the estimated future values Of yf
are given by
min(E*‘E*) = min(E'fEf +E'E)
.
y^f * = WfG,
, = min (y-Wcr)'(~-~a)+(yf-Wcc~)'(yf-W4-U)
given
The first part is minimized when
Wf=[v^i,f Xi,f] and& =i:& Ai]'; or
?i =(W'W)-' W'y, and
h
where Wf=Xf and
by Tf,Rl=Wf a" the absolute minimum of zero for the
second part requires
21 = Gil

estimat_ed by djrect least yf-Wf"_P =.cl. Accordingly,

("i
;;uares);or by yf,R2 = Wp2, where Wf = h
p =yf-Wfg=yf-yf = ef.

This establishes the conclusion that the

and a2 = Gi2
prediction errors are the coefficients
on the dummy variables. Accordingly,
(vi estimated jointly with6 i using the
the corresponding estimated variances of
3SLS estimator). Note that Xi f and X
these coefficients give the estimated
represent futu_re values of Xi an dX
Yi f are estimates of fu- prediction error variances, i.e.,
respectively.
ture values of tHe endogenous variables
on the right hand side of Eq. (2). (The Var( ) = Var(ef).
subscripts S and R denote the use of
structural and reduced form equations
respectively). In this simulation Simulation Results
study, the s dimentional column vector
Thus the corresponding For purposes of experimentation the
yf is known.
estimated prediction errors are given by basic model considered is given by the
the s x 1 vectors following two behavioral equations:

= Yf - y^f,S, Y' = Y'O + y11x1 + y12x2+ 511y2+u1

ef,S

ef,Rl'yf-~f,Rl, anu Y2 = Y2D + Q'Y' + u2

wherey's are endogenous variables,x's

ef,R2=yf - y^f,f?2' are predetermined variables and u's are
random disturbances. Numerical values
Let Var (ef S), Var (ef,,R1) and Var are specified for all parameters and
(ef,R2) repretent their estimated varia- also for the two predetermined
rices respectively. The objective is to variables. The data on endogenous
examine the predictive power of variables are generated from the reduced
form coefficients and the values of
yf,S> yf,Rl and yf,R2 predetermined variables together with a
specified sample of reduced form
by comparing these variances. Consider disturbances. The experiments employ
the expanded form of Eq. (4) as: 100 generated samples of 24 observations
each. 20 observations are used in
y" = W*C(+ DP+E* (5) estimating the structural and reduced
form parameters. The 21st through 24th
where values on the endogenous variables serve
as the true psot-sample values the
Y* =[y' y'f]', w* =[W' w'f]', model-structure has generated and with
which the post-sample predictors made by
E*=[E' Ef’l’ and D is a (T t s)x s alternative procedures are compared. In
each experiment,
 Proc. 6th Int. Conf. on Mathematical Modelling 255

Var(ef,s), Var(ef,R1) and Var(ef,P2) References

are directly calculated through the use Court, R. (1973). "Efficient Estimation
of constructed variables, and then ave- of the Reduced Form from Incomplete
raged over the number of experiments. Econometric Models", Review of
The results are given in the following Economic Studies, 40, 411-417.
table.
Fuller, W.A. (1980), "The Use of
Prediction Error Variance Indicator Variables in Computing
Predictions", Journal of Economet-
Post- Reduced-Form Equation Structural rics, 23, 231-243.
Sample Unrestricted Restricted Equation
Prediction Var(ef,R1) Varlef,Kp) Var(ef,S) Goldberyer, Nagar and Odeh (1961), "The
CovarianCe Matrices of Reduced Form
Coefficients and of Forecasts for a
yZ,T+l 6.76 4.01 2.74 Structural Econometric Model",
Econometrica 29, 556-573.
Y2,T+2 9.69 4.12 2.77
Pagan, A.R. and D.F. Nicholls, (1984),
y2,T+3 9.01 4.41 2.79 "Estimating Predictions, Prediction
Errors and Their Standard Deviations
Y2,T+4 9.61 4.45 2.71 Using Constructed Variables",
Journal of Econometrics 24, 293-310.

Salkever, D.S. (1976), "The Use of Dummy

The estimated prediction errors and Variables to Compute Predictions,
their corresponding standard deviations Prediction Errors and Confidence
speak for themselves. The restricted Intervals", Journal of Econometrics,
reduced- form outperforms the 4, 393-397.
unrestricted one and this is in accord
with the general rule that estimators
that use more information are more
efficient. However, what is new and
noteworthy is the use of the estimated
structural-equation for prediction which
outperforms the use of the restricted
reduced-form procedure.

Conclusions

We conducted sampling experiments to

compare the relative performances of the
estimated structural equations with the
corresponding reduced-form equation in
view of the mathematical complexities
which hamper analytical derivations.
Further. desirable asymptotic properties
are no guide to the applied researcher
who works with small samples.
Accordingly, we examined the performance
of each approach relative to the other
by the magnitude of their respective
prediction error variances within the
Monte Carlo framework. As shown,
variances of error predictions resulting
from each of these alternatives can be
directly estimated by using constructed
variables. The simulation results
reported here suggest the superiority of
structural equation for prediction. Any
conclusions based on Monte Carlo
experiments are, by their very nature,
model-structure specific. In other
words, we should try to increase the
generality of the conclusions of our
study by conducting a large number of
experiments with various feasible
combinations of models. Nevertheless,
the use of constructed variables is
clearly a great advantage particularly
from an empirical point of view because
it permits a check whether to go for
reduced-form approach or estimated
structural equation for predicting the
future values of endogenous variables in
a simultaneous equation model.