Low rationality test

[c912]

Hello,

I would like to do my low rationnality test with:

linreg ev 1968:1 2014:1
# constant intrasrmere

restrict 2
# 1
# 1 0
# 2
# 1 1

But I don't understant if i should use a linreg like this because I make my prevision with an arma model:
boxjenk(constant,ar=0, ma=3, maxl, regressors,define=eveq) ev 1960:01 * resmere2
# pib

UFORECAST(from=1968:1,to=2014:1,equation=eveq,static,noprint) intrasrmere
GRAPH(STYLE=LINE,HEADER="Observation et previsions intra echantillon (sans reestimation)",key=above) 2
# ev
# intrasrmere

Is it a linreg in every case for the test or does it depends of the model?

[TomDoan]

1. TEST is simpler than RESTRICT for that type of test.

linreg ev 1968:1 2014:1
# constant intrasrmere
test(title="Rationality test") 
# 1 2
# 0.0 1.0

2. You should include the CONSTANT in the regressors list, rather than having it automatically included, on the BOXJENK. The simplest way to do that is with BOXJENK with the GLS option.

boxjenk(gls,ma=3,define=eveq) ev 1960:01 * resmere2
# constant pib

That will push the two regressors to the start of the regression (in that same order: CONSTANT PIB) so you can use the same type of TEST instruction to do the test. However,

3. A rationality test should not be done using a GLS estimator (whether it's AR, MA or ARMA doesn't matter). If you have serially correlated errors, use a HAC estimator of the covariance matrix instead. This is covered in quite a few textbooks now such as Hayashi's Econometrics.