Gujarati Book
Gujarati Book
Gujarati Book
Estimation or VAR
Returning to the Canadian money–interest rate example, we saw that when
we introduced six lags of each variable as regressors, we could not reject the
hypothesis that there was bilateral causality between money (M1 ) and inter-
est rate, R (90-day corporate interest rate). That is, M1 affects R and R affects
M1 . These kinds of situations are ideally suited for the application of VAR.
To explain how a VAR is estimated, we will continue with the preceding
example. For now we assume that each equation contains k lag values of
M (as measured by M1 ) and R. In this case, one can estimate each of the fol-
lowing equations by OLS.12
k
k
M1t = α + β j Mt− j + γ j Rt− j + u1t (22.9.1)
j=1 j=1
k
k
Rt = α + θ j Mt− j + γ j Rt− j + u2t (22.9.2)
j=1 j=1
where the u’s are the stochastic error terms, called impulses or innovations
or shocks in the language of VAR.
Before we estimate (22.9.1) and (22.9.2) we have to decide on the maxi-
mum lag length, k. This is an empirical question. We have 40 observations
in all. Including too many lagged terms will consume degrees of freedom,
not to mention introducing the possibility of multicollinearity. Including
too few lags will lead to specification errors. One way of deciding this ques-
tion is to use a criterion like the Akaike or Schwarz and choose that model
that gives the lowest values of these criteria. There is no question that some
trial and error is inevitable.
To illustrate the mechanics, we initially used four lags (k = 4) of each vari-
able and using Eviews 4 we obtained the estimates of the parameters of the
12
One can use the SURE (seemingly unrelated regression) technique to estimate the two
equations together. However, since each regression contains the same number of lagged
endogenous variables, the OLS estimation of each equation separately produces identical
(and efficient) estimates.
Gujarati: Basic IV. Simultaneous−Equation 22. Time Series © The McGraw−Hill
Econometrics, Fourth Models Econometrics: Forecasting Companies, 2004
Edition
M1 R
R2 0.988154 0.852890
Adj. R 2 0.984034 0.801721
Sum square residuals 4820241. 53.86233
SE equation 457.7944 1.530307
F statistic 239.8315 16.66815
Log likelihood −236.1676 −53.73716
Akaike A/C 15.32298 3.921073
Schwarz SC 15.73521 4.333311
Mean dependent 28514.53 11.67292
SD dependent 3623.058 3.436688
preceding two equations, which are given in Table 22.2. Note that although
our sample runs from 1979–1 to 1988–4, we used the sample for the period
1979–1 to 1987–4 and saved the last four observations to check the forecast-
ing accuracy of the fitted VAR.
Since the preceding equations are OLS regressions, the output of the
regression given in Table 22.2 is to be interpreted in the usual fashion. Of
course, with several lags of the same variables, each estimated coefficient
will not be statistically significant, possibly because of multicollinearity. But
collectively, they may be significant on the basis of the standard F test.
Let us examine the results presented in Table 22.2. First consider the
M1 regression. Individually, only M1 at lag 1 and R at lags 1 and 2 are statis-
tically significant. But the F value is so high that we cannot reject the hy-
pothesis that collectively all the lagged terms are statistically significant.
Turning to the interest rate regression, we see that all the four lagged money
Gujarati: Basic IV. Simultaneous−Equation 22. Time Series © The McGraw−Hill
Econometrics, Fourth Models Econometrics: Forecasting Companies, 2004
Edition
M1 R
R2 0.988198 0.806660
Adj. R 2 0.986571 0.779993
Sum square residuals 5373510. 71.97054
SE equation 430.4573 1.575355
F statistic 607.0720 30.24878
Log likelihood −251.7446 −60.99215
Akaike A/C 15.10263 3.881891
Schwarz SC 15.32709 4.106356
Mean dependent 28216.26 11.75049
SD dependent 3714.506 3.358613