Problem of Hausman Test for RE&FE by programming

[SC Peng]

Hello,

I am using Eviews 8. I am using pool object and unbalanced panel data w 19,240 firms over the sample period for 2000 ~2011, a total 46,658 obs. The pool object I idenify is YEAR. And I try to use Hausman test to see whether the fixed effect or random effect is appropriate or not. So, I set two eqs as following.

eq1: show p_1.ls(cx=f,p) Tobin? c {%fz} {%cz} {%zz} {%z1a}?
eq2: show p_2.ls(cx=r,p) Tobin? c {%fz} {%cz} {%zz} {%z1a}?

Eviews shows the error message to me said that "Random effects estimation requires number of cross sections>number of coefs for between estimator for estimate of RE innovation variance." while I run the eq2.

However, the equation is efficient if I revise the equation option to eq3.
eq3: show p_2.ls(cx=r,rancalc=wh,p) Tobin? c {%fz} {%cz} {%zz} {%z1a}?
The difference btw eq2 & eq3 is rancalc=wh!!!! I cann't understand why :(

Code: Select all

for !eqno=1 to {%model} pool p_{!eqno} for !YEAR=1 to {%p3}-1 p_{!eqno}.add {!YEAR} next p_{!eqno}.describe(m,c,i,p) Tobin? next %cz="MV? LevF? " %fz="gdpg? DIS? " %zz="Ind1b? Ind2b? Ind3b? Ind4b? Ind5b? Ind6b? Ind7b? Ind8b? Ind9b? Ind10b?" show p_1.ls(cx=f,p) Tobin? c {%fz} {%cz} {%zz} {%z1a}? show p_2.ls(cx=r,p) Tobin? c {%fz} {%cz} {%zz} {%z1a}? table result setcolwidth(result,1,20) setcell(result,1,1,"Hausman test for fixed versus random effects") setline(result,2) !a=3 for !eqno=1 to 1 setcell(result,!a,1,"{!eqno}") vector beta_f = p_{!eqno}.@coefs matrix covar_f = p_{!eqno}.@cov scalar nbeta_f = @obs(beta_f) - {%p5} vector b_fixed = @subextract(beta_f,2,1,nbeta_f,1) matrix cov_fixed = @subextract(covar_f,2,2,nbeta_f,nbeta_f) !w=2 vector beta_r = p_!w.@coefs matrix covar_r = p_!w.@cov scalar nbeta_r = @obs(beta_r) - {%p5} vector b_random = @subextract(beta_r,2,1,nbeta_r,1) matrix cov_random = @subextract(covar_r,2,2,nbeta_r,nbeta_r) matrix b_diff = b_fixed - b_random matrix var_diff = cov_fixed - cov_random matrix qform = @transpose(b_diff)*@inverse(var_diff)*b_diff if qform(1,1)>=0 then !df = @rows(b_diff) setcell(result,!a+1,1,"chi-sqr(" + @str(!df) + ") = ") setcell(result,!a+1,2,qform(1,1)) setcell(result,!a+2,1,"p-value = ") setcell(result,!a+2,2,1-@cchisq(qform(1,1),!df)) else setcell(result,!a+1,1," Test undefined: Quadratic form is negative" ) endif !a=3+!a next

Thanks in advance.

SC

[EViews Glenn]

Am I understanding correctly that your pool is defined over the 11 years?

If so then the reason is exactly what the error suggests. The default Swamy-Arora method requires the estimation of the between estimator which has as many observations as you have pool identifiers. It sounds as though you have more regressors than the number of pool identifiers.

In contrast, the other two estimators do not require this auxiliary regression. As we note in the manual

The Swamy-Arora estimator of the component variances, cited most often in textbooks, uses residuals from the within (fixed effect) and between (means) regressions. In contrast, the Wansbeek and Kapteyn estimator uses only residuals from the fixed effect (within) estimator, while the Wallace-Hussain estimator uses only OLS residuals.

[SC Peng]

Hello,

Thank you very much. You are right. I made the error as Eviews Error Message and you said. I am very happy to Know what error I made. Very appreciated. Thanks again. :D

SC