Scribd
Upload
20 views2 pages

Durbin-Watson Test for Autocorrelation

The Durbin-Watson test is used to detect autocorrelation in the errors of a time series regression model. It tests the null hypothesis that the errors are uncorrelated against the alternative hypothesis that the errors follow an AR(1) process. The test statistic d is calculated based on the residuals from least squares estimation. A value of d close to 2 indicates white noise errors, while a value far from 2 suggests autocorrelation. While useful, the Durbin-Watson test has some limitations, such as assuming the model form is known and having low power against alternatives like MA(1) or nonstationary errors.

Uploaded by

Carlos Alas
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd
20 views2 pages

Durbin-Watson Test for Autocorrelation

The Durbin-Watson test is used to detect autocorrelation in the errors of a time series regression model. It tests the null hypothesis that the errors are uncorrelated against the alternative hypothesis that the errors follow an AR(1) process. The test statistic d is calculated based on the residuals from least squares estimation. A value of d close to 2 indicates white noise errors, while a value far from 2 suggests autocorrelation. While useful, the Durbin-Watson test has some limitations, such as assuming the model form is known and having low power against alternatives like MA(1) or nonstationary errors.

Uploaded by

Carlos Alas
Copyright
© Attribution Non-Commercial (BY-NC)
We take content rights seriously. If you suspect this is your content, claim it here.
Available Formats
Download as PDF, TXT or read online on Scribd

THE DURBIN-WATSON TEST

Suppose we have a time series regression model relating a "dependent" time series {yt } to the "independent" time series {x 1t } , . . . , {xpt }. The model is yt = o + 1x 1t + . . . + p xpt + t , t =1,2, . . . , n ,

where {t } is a time series of "errors", or "disturbances". Such models are useful for both explanatory and forecasting purposes. The parameters 0 , 1 , . . . , p may be estimated by least-squares. In practice, it often happens that the errors are not independent (as assumed in standard regression models) but instead are autocorrelated . Such error autocorrelation, or "serial correlation", has many underirable but correctable consequences (e.g., the least-squares estimates sub-optimal, standard condence intervals for are incorrect, the error term is forecastable). Thus, it is highly desirable to try to detect error autocorrelations. The Durbin-Watson Test for serial correlation assumes that the t are stationary and normally distributed with mean zero. It tests the null hypothesis Ho that the errors are uncorrelated against the alternative hypothesis H 1 that the errors are AR (1). Thus, if s are the error autocorrelations, then we have Ho : s = 0 (s > 0), and H 1 : s = s for some nonzero with < 1. To test Ho against H 1, get the least squares estimates and residuals e 1 , . . . , en . The test statistic is d =

(et et 1)2/tet2 t =2 =1

Note that ignoring "end effects", we have d 2(1 r 1), where r 1 is the sample ACF of the residuals at lag 1. If the errors are white noise, d will be close to 2. If the errors are strongly autocorrelated, d will be far from 2. The exact procedure for deciding whether a given value of d is signicant is somewhat complicated, and is described, for example, in Draper and Smith, Page 163. In some cases, the test can be "inconclusive," i.e., Ho is neither accepted nor rejected. Since its development in 1951, the test has been found to be extremely useful, especially for the analysis of economic time series. It does, however, suffer from a number of shortcomings, some of which are as follows. First, the form of the model (i.e., the dimension p and the explanatory variables x 1t , . . . , xpt ) is assumed known. In practice, this is rarely the case, and instead a data based procedure

-2-

must be employed to "identify" the model. Second, the test is sometimes inconclusive, as mentioned above. Third, the AR (1) alternative hypothesis is by no means the only way in which the null hypothesis may fail. Suppose, for example, the errors are in fact MA (1), or perhaps even some nonstationary series such as a random walk. The Durbin-Watson test can have very low power against such alternatives (i.e., it can fail to detect them).

You might also like