Breusch-Pagan and Hausman

[Izzuddin Abdulllah]

Hi all, i have encountered a minor problem here where i've read different opinion on utilizing BP test and Hausman test.

This is taken from literature, author test REM and FEM first using Hausman test, and got REM favorable....due to got REM, author said there is need to test REM with POLS aftermath using BP test...

my question is, which test actually (BP/Hausman) should be employed first? because my basic is, we should use BP test first between POLS and REM....then cont with Hausman between REM and FEM..

regarding situation taken from literature above, maybe i can say they did wrong procedure? (because they didn't follow correct order which BP test first, then go to Hausman next)..or is there got any excuses for author to use Hausman first then proceed with BP test???

Thanks a lot.

[Carlo Lazzaro]

Izzudin:
I would follow the approach you suggested ((BP then -hausman-).

[Izzuddin Abdulllah]

Thanks Carlo

Regarding these two tests also, how to decide whether to stop at BP only or continue with Hausman?

I mean that, is it possible if i just employ BP only? and not further with Hausman? or there is compulsory to test these two test together? because literatures ive read just explained on they used Hausman, not for BP

[Izzuddin Abdulllah]

for my situation actually, BP test suggest me to use POLS.. so, do i need to test Hausman also?

[Carlo Lazzaro]

Izzudin:
what does -xtreg, fe- F-test at the foot of the outcome table tell you?
I must also to correct my previous reply: the order of the tests should be reverted: first, compare -fe- vs -re- via -hausman-; if -re- is the specification to be used, then use BP to investigate whether there's evidence for rejectiing the null that variances across entities is zero. You may find the following article interesting: https://www.ncbi.nlm.nih.gov/pmc/art...jhpm-1-163.pdf

[Izzuddin Abdulllah]

. xtreg gdp cf nfdi inf ir ed reer, fe

Fixed-effects (within) regression Number of obs = 55
Group variable: countrynum Number of groups = 5

R-sq: within = 0.1716 Obs per group: min = 11
between = 0.1266 avg = 11.0
overall = 0.1453 max = 11

F(6,44) = 1.52
corr(u_i, Xb) = -0.2639 Prob > F = 0.1945

------------------------------------------------------------------------------
gdp | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
cf | .0241977 .0447395 0.54 0.591 -.0659688 .1143643
nfdi | .3213207 .2190785 1.47 0.150 -.120203 .7628443
inf | -.1979292 .1183671 -1.67 0.102 -.4364823 .040624
ir | -.2436172 .1165544 -2.09 0.042 -.4785172 -.0087172
ed | -.0478039 .0789318 -0.61 0.548 -.2068804 .1112726
reer | .0295066 .0364351 0.81 0.422 -.0439237 .1029368
_cons | 5.153046 4.315112 1.19 0.239 -3.543491 13.84958
-------------+----------------------------------------------------------------
sigma_u | 1.0839499
sigma_e | 1.9505771
rho | .23594751 (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0: F(4, 44) = 1.88 Prob > F = 0.1303

[Carlo Lazzaro]

Izzudin:
go POLS.

[Izzuddin Abdulllah]

. xtreg gdp cf nfdi inf ir ed reer, fe

Fixed-effects (within) regression               Number of obs      =        55
Group variable: countrynum                      Number of groups   =         5

R-sq:  within  = 0.1716                         Obs per group: min =        11
       between = 0.1266                                        avg =      11.0
       overall = 0.1453                                        max =        11

                                                F(6,44)            =      1.52
corr(u_i, Xb)  = -0.2639                        Prob > F           =    0.1945

------------------------------------------------------------------------------
         gdp |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
          cf |   .0241977   .0447395     0.54   0.591    -.0659688    .1143643
        nfdi |   .3213207   .2190785     1.47   0.150     -.120203    .7628443
         inf |  -.1979292   .1183671    -1.67   0.102    -.4364823     .040624
          ir |  -.2436172   .1165544    -2.09   0.042    -.4785172   -.0087172
          ed |  -.0478039   .0789318    -0.61   0.548    -.2068804    .1112726
        reer |   .0295066   .0364351     0.81   0.422    -.0439237    .1029368
       _cons |   5.153046   4.315112     1.19   0.239    -3.543491    13.84958
-------------+----------------------------------------------------------------
     sigma_u |  1.0839499
     sigma_e |  1.9505771
         rho |  .23594751   (fraction of variance due to u_i)
------------------------------------------------------------------------------
F test that all u_i=0:     F(4, 44) =     1.88               Prob > F = 0.1303

[Izzuddin Abdulllah]

thanks. so, no need to test for hausman? right?

[Carlo Lazzaro]

Izzudin:
both specification do not reject the null that individual effect are jointly 0. Hence, you can go POLS.

[Izzuddin Abdulllah]

Thanks Carlo.

I assume there is no need to present for Hausman test in my report then.

Thank you

[Carlo Lazzaro]

Izzuddin:
just for the sake of completeness, I would report -hausman- test outcome, that indicates that -xtreg,re- specification outperforms -xtreg,fe- one, first;
then I would report BP test outcome that shows that POLS outperforms -xtreg, re-

[Izzuddin Abdulllah]

Thanks Carlo.

I have a thinking that, why should I employ Hausman anymore? The BP test already said to use POLS. (it is another situation if the BP shows REM appropriate, then I would have employ Hausman).

I asked in post #1, which test do I need to do first between BP and Hausman. And the literature did wrong procedure regarding this two test.

So, I wonder if there is a need for me to write Hausman result if I just stop at BP since it gives me POLS

[Carlo Lazzaro]

Izzudin:
I do not know what the audience of your research actually is.
Usually, the comparison between -xtreg, fe- and -xtreg,re- entails -hausman- specification test.
Its absence in a paper would sound strange.
Please consider that others might be interested in replication your research: hence, details about what you did are relevant.

[Izzuddin Abdulllah]

Thank Carlo for the suggestion. By the way, it has crossed my mind also since i thought i would be strange not to present Hausman test as other panel studies (in my field) presented it.

With that I think my queries were solved.

Thank you very much sir Carlo for your kind reply