Re: [Gretl-users] Tests for Random Residuals
El Sunday 30 March 2008 16:41:43 Thomas La Bone escribió:
Good Morning,
About two months ago when I first started working with Gretl I asked if
Gretl has tests for randomness of model residuals (accessible from
either the GUI or the command line). I got a number of fairly
non-specific responses like "look at the documentation" and "look at the
menus on the windows". Well, after two months of working with Gretl I am
now a hardened veteran and ready to ask the question again:
The only mention of a test for randomness I can find is the
non-parametric runs test. Can Gretl do a Ljung-Box test (or some other
test I am not familiar with) of the randomness of residuals? Thanks in
advance for any advice you may offer.
If you use OLS, in the model output window you can
select /Tests/Autocorrelation and you will have the Ljung-Box Q statistic in
the last line.
Also, if you right-clik over a time series and select "correlogram" (or
selecting Variable/correlogram in the menu, or using the command "corrgm" in
a gretl console), after selecting a maximum lag, two windows appear, one with
the graphs and the other one is a text window with the values of the simple
and partial correlation coefficients. In this second window, there are five
columns:
LAG, ACF, PACF, Q-stat. and p-value. The fourth column contains the
Q-statistics of Box-Pierce. The k-th element in this column contains the
Q-stat for the joint hipotesis that rho_1=...=rho_k=0.
The manual says it is the Box-Pierce statistic. The Ljung-Box and Box-Pierce
Statistics are asymptotically equivalent. Both of them are not Chi-square in
small samples, but the Ljung-Box statistic has a correction that makes it
closer to the Chi-square (I think gretl should use it in the correlogram
output).
--
Ignacio Diaz-Emparanza
DEPARTAMENTO DE ECONOMÍA APLICADA III (ECONOMETRÍA Y ESTADÍSTICA)
UPV/EHU
Avda. Lehendakari Aguirre, 83 | 48015 BILBAO
T.: +34 946013732 | F.: +34 946013754
www.et.bs.ehu.es