2
Appendix A
In addition to obtaining the Durbin-Watson statistic for residuals from REGRESSION,
you should also plot the ACF and PACF of the residuals series. The plots might suggest
either that the residuals are random, or that they follow some ARMA process. If the
residuals resemble an AR1 process, you can estimate an appropriate regression using
the
AREG procedure. If the residuals follow any ARMA process, you can estimate an
appropriate regression using the
ARIMA procedure.
In this appendix, we have reproduced two sets of tables. Savin and White (1977)
present tables for sample sizes ranging from 6 to 200 and for 1 to 20 regressors for
models in which an intercept is included. Farebrother (1980) presents tables for sample
sizes ranging from 2 to 200 and for 0 to 21 regressors for models in which an intercept
is not included.
Let’s consider an example of how to use the tables. In Chapter 9, we look at the
classic Durbin and Watson data set concerning consumption of spirits. The sample size
is 69, there are 2 regressors, and there is an intercept term in the model. The Durbin-
Watson test statistic value is 0.24878. We want to test the null hypothesis of zero
autocorrelation in the residuals against the alternative that the residuals are positively
autocorrelated at the 1% level of significance. If you examine the Savin and White
tables (Table A.2 and Table A.3), you will not find a row for sample size 69, so go to
the next lowest sample size with a tabulated row, namely N=65. Since there are two
regressors, find the column labeled k=2. Cross-referencing the indicated row and
column, you will find that the printed bounds are dL = 1.377 and dU = 1.500. If the
observed value of the test statistic is less than the tabulated lower bound, then you
should reject the null hypothesis of non-autocorrelated errors in favor of the hypothesis
of positive first-order autocorrelation. Since 0.24878 is less than 1.377, we reject the
null hypothesis. If the test statistic value were greater than dU, we would not reject the
null hypothesis.
A third outcome is also possible. If the test statistic value lies between dL and dU,
the test is inconclusive. In this context, you might err on the side of conservatism and
not reject the null hypothesis.
For models with an intercept, if the observed test statistic value is greater than 2,
then you want to test the null hypothesis against the alternative hypothesis of negative
first-order autocorrelation. To do this, compute the quantity 4-d and compare this value
with the tabulated values of dL and dU as if you were testing for positive
autocorrelation.
When the regression does not contain an intercept term, refer to Farebrother’s
tabulated values of the “minimal bound,” denoted dM (Table A.4 and Table A.5),
instead of Savin and White’s lower bound dL. In this instance, the upper bound is