EViews 8 New Econometrics and Statistics: Testing and Diagnostics
EViews 8 features a number of additions and improvements to its extensive set of diagnostics and tests.
EViews 8 extends the existing Chow and Quandt-Andrews structural break test tools to allow for multiple breakpoint testing (Bai, 1997; Bai and Perron, 1998, 2003). You may now, for a regression model estimated using linear least squares specified by list, ask EViews to test for multiple unknown breakpoints up to a specified maximum.
EViews offers the following test methods:
- Sequential L+1 breaks vs. L.
- Sequential tests all subsets.
- Global L breaks vs. none.
- L+1 breaks vs. global L.
- Global information criteria.
You may test against “pure” breakpoint specifications in which all of the regressors have regime specific coefficients, or specifications in which only some coefficients vary with the regime.
For more details on multiple breakpoint testing, see our examples.
For panel equations estimated by GMM, EViews 8 computes the first and second order serial correlation statistics proposed by Arellano and Bond (1991) as one method of testing for serial correlation. The test is actually two separate statistics, one for first order correlation and one for second. If the innovations are i.i.d. we expect the first order statistic to be significant (with a negative auto-correlation coefficient), and the second order statistic to be insignificant.
EViews 8 extends the existing Granger Causality tests to perform panel data specific testing.
EViews 8 supports two of the simplest approaches to causality testing in panels. The first is to treat the panel data as one large stacked set of data, and then perform the Granger Causality test in the standard way, with the exception of not letting data from one cross-section enter the lagged values of data from the next cross-section. The second approach, the Dumitrescu-Hurlin (2012) approach, makes an extreme opposite assumption; it allows all coefficients to be different across cross-sections.
EViews 8 now offers heteroskedasticity and autocorrelation consistent (HAC) covariance computation in equations estimated by GLM.
EViews 8 now reports the robust Wald test of the null hypothesis that all non-intercept coefficients are zero in cases where you specify a robust coefficient covariance method.
Previously, EViews only reported the residual based F-statistic for testing the null hypothesis. This F-statistic statistic depends only on the coefficient point estimates, and not their standard errors, and was valid only under the maintained hypotheses of no heteroskedasticity or serial correlation. For ordinary least squares without conventionally estimated standard errors, this statistic is numerically identical to the Wald statistic for the hypothesis that all non-intercept coefficients are equal to zero. However, the numerical equivalence between the two test statistics breaks down if robust standard errors are employed.