Insignificant results Fixed and Random Effects Model

[Dung Delime]

https://www.e-jei.org/upload/JEI_38_...2013600288.pdf
I'm doing a specific research with panel data (65 country and 4 year). My topic and model are based on this paper (I will attach this paper (page 75 - 78)
I firstly intend to run POLS in Stata 17., so before that, I checked the normal distribution of the residuals, it' ok. But when I check the normal distribution of other quantitative variables, only some variables are normal distributed, even when I transform to logarit form. However, my main independent variable still had a positive, significant effect on my dependent variable. Then I run REM and FEM models, these models all give better results than POLS, especially running Hausman, FEM is the best choice. However, all independent variables are statistically insignificant. I really don't know how to fix my model. If I still choose POLS, it it will suffer from heteroskedasticity, multicollinearity, autocorrelation, and serial correlation and I intend to use xtscc to fix. But I find it very strange why FEM is better than POLS but the independent variables are not statistically significant, there are still some defects, even if corrected, they are still not statistically significant.
My model has another problem: Time fixed effect and country fixed effect. I use testparm and It said that my model need only country fixed effect. But when I put this into my model, all dependent variables are not statistically significant. Please help me, I would really appreciate any help on the interpretation of these results since I am very unsure about this topic!

[Carlo Lazzaro]

Dung:
welcome to this forum.
There's no way (for me, at least) to comment on what you're facing without taking a look at what you coded and what Stata gave you back (as oer FAQ).
That said, normality is a (weak) requirement for residual distribution only.

[Dung Delime]

Thank you for responding to my question. I will attach 5 pictures: (DTF1 - DTF6 are my main dependent variables ( I expect these variables are positive (+); x1,x3 ...x9 are my control variables.
Picture 1: Result of POLS model. Only 3 main dependent variables are statistically significant, but DTF4 is contrary to the hypothesis.
Picture 2: Result of FEM model. Result has said Fem is suitable than POLS, but non main variables are statistically significant.
Picture 3: Result of REM model. Result has said REM is suitable than POLS, but non main variables are statistically significant.
Picture 4: Result of Hausman test. Result has said REM is suitable than FEM, but in REM model, non main variables are statistically significant. Although the REM and FEM models both have flaws, no matter how hard I tried to fix them, the DTF variables remained statistically insignificant. Because I don't know how to fix, I think I will choose POLS with xtscc.
Picture 5. Result of running XTSCC. After using xtscc, some of DTF variables are statistically significant, but DTF4 is contrary to the hypothesis (it must be positive (+)result). Do you think I should choose POLS and xtscc, or any solutions.

Picture 1:
. regress logfva dtf1 dtf2 dtf3 dtf4 dtf5 dtf6 logx1 x3 x4 x5 x6 x7 x8 log_x9

Source SS df MS Number of obs = 255
F(14, 240) = 30.04
Model 8.88185568 14 .634418263 Prob > F = 0.0000
Residual 5.06861707 240 .021119238 R-squared = 0.6367
Adj R-squared = 0.6155
Total 13.9504728 254 .054923121 Root MSE = .14532


logfva Coefficient Std. err. t P>t [95% conf. interval]

dtf1 .0705629 .0970716 0.73 0.468 -.1206582 .2617839
dtf2 .0476133 .053527 0.89 0.375 -.0578294 .153056
dtf3 .0757482 .0567768 1.33 0.183 -.0360962 .1875926
dtf4 -.2312862 .0710671 -3.25 0.001 -.3712811 -.0912914
dtf5 .0859264 .0445389 1.93 0.055 -.0018107 .1736635
dtf6 .2385201 .0780655 3.06 0.003 .084739 .3923012
logx1 -.034425 .0238293 -1.44 0.150 -.0813662 .0125162
x3 .0124655 .0022985 5.42 0.000 .0079377 .0169933
x4 .002236 .0002191 10.20 0.000 .0018043 .0026677
x5 .0000336 .0003812 0.09 0.930 -.0007173 .0007845
x6 -.0627637 .027466 -2.29 0.023 -.116869 -.0086584
x7 .06392 .0856166 0.75 0.456 -.104736 .232576
x8 -.0035081 .0019316 -1.82 0.071 -.0073132 .000297
log_x9 -.0566343 .0294499 -1.92 0.056 -.1146477 .001379
_cons 1.128491 .1863541 6.06 0.000 .761392 1.495589

Picture 2:

. xtreg logfva dtf1 dtf2 dtf3 dtf4 dtf5 dtf6 logx1 x3 x4 x5 x6 x7 x8 log_x9, fe

Fixed-effects (within) regression Number of obs = 255
Group variable: quốcgia Number of groups = 64

R-squared: Obs per group:
Within = 0.2272 min = 3
Between = 0.5394 avg = 4.0
Overall = 0.5374 max = 4

F(14,177) = 3.72
corr(u_i, Xb) = 0.1492 Prob > F = 0.0000


logfva Coefficient Std. err. t P>t [95% conf. interval]

dtf1 .0128274 .0556282 0.23 [IMG]file:///C:/Users/Administrator/Pictures/Screenshots/Screenshot%20(41).png[/IMG] -.0969526 .1226073
dtf2 -.0276993 .0256203 -1.08 0.281 -.0782599 .0228613
dtf3 .0090874 .0198548 0.46 0.648 -.0300952 .04827
dtf4 .0371234 .0255062 1.46 0.147 -.013212 .0874588
dtf5 -.0112031 .0185394 -0.60 0.546 -.0477898 .0253837
dtf6 .0204818 .0351562 0.58 0.561 -.0488974 .089861
logx1 -.0506932 .3199963 -0.16 0.874 -.6821923 .5808059
x3 .0060748 .0039525 1.54 0.126 -.0017252 .0138748
x4 .00141 .0003409 4.14 0.000 .0007372 .0020827
x5 -.0000495 .000069 -0.72 0.474 -.0001857 .0000866
x6 .0268641 .0193582 1.39 0.167 -.0113385 .0650667
x7 .0475457 .1001055 0.47 0.635 -.1500081 .2450995
x8 .0011949 .0010178 1.17 0.242 -.0008137 .0032034
log_x9 -.0303707 .0168804 -1.80 0.074 -.0636834 .002942
_cons 1.400875 2.331728 0.60 0.549 -3.20069 6.002441

sigma_u .16059298
sigma_e .02511175
rho .97613235 (fraction of variance due to u_i)

F test that all u_i=0: F(63, 177) = 124.77 Prob > F = 0.0000

Picture 3:

. xtreg logfva dtf1 dtf2 dtf3 dtf4 dtf5 dtf6 logx1 x3 x4 x5 x6 x7 x8 log_x9, re

Random-effects GLS regression Number of obs = 255
Group variable: quốcgia Number of groups = 64

R-squared: Obs per group:
Within = 0.2213 min = 3
Between = 0.5747 avg = 4.0
Overall = 0.5719 max = 4

Wald chi2(14) = 130.57
corr(u_i, X) = 0 (assumed) Prob > chi2 = 0.0000


logfva Coefficient Std. err. z P>z [95% conf. interval]

dtf1 .0176399 .053646 0.33 0.742 -.0875042 .1227841
dtf2 -.0321077 .0247951 -1.29 0.195 -.0807052 .0164898
dtf3 .0091752 .0191456 0.48 0.632 -.0283494 .0466998
dtf4 .0327812 .0245336 1.34 0.181 -.0153039 .0808662
dtf5 -.0041733 .0177705 -0.23 0.814 -.0390029 .0306562
dtf6 .0301217 .0323778 0.93 0.352 -.0333377 .0935811
logx1 -.0379132 .032929 -1.15 0.250 -.102453 .0266265
x3 .007656 .0027492 2.78 0.005 .0022677 .0130444
x4 .0017333 .0002532 6.85 0.000 .0012371 .0022296
x5 -.0000441 .0000686 -0.64 0.520 -.0001785 .0000902
x6 .0164448 .0171427 0.96 0.337 -.0171544 .050044
x7 .0214958 .0719827 0.30 0.765 -.1195877 .1625792
x8 .0005773 .0009462 0.61 0.542 -.0012771 .0024318
log_x9 -.0339612 .0162174 -2.09 0.036 -.0657467 -.0021757
_cons 1.278815 .2434473 5.25 0.000 .8016673 1.755963

sigma_u .15530221
sigma_e .02511175
rho .97452057 (fraction of variance due to u_i)

Picture 4:

Note: the rank of the differenced variance matrix (13) does not equal the number of coefficients being tested (14); be sure this is what
you expect, or there may be problems computing the test. Examine the output of your estimators for anything unexpected and
possibly consider scaling your variables so that the coefficients are on a similar scale.

Coefficients ----
(b) (B) (b-B) sqrt(diag(V_b-V_B))
fe re Difference Std. err.

dtf1 .0128274 .0176399 -.0048126 .0147177
dtf2 -.0276993 -.0321077 .0044085 .00645
dtf3 .0090874 .0091752 -.0000878 .0052593
dtf4 .0371234 .0327812 .0043422 .0069761
dtf5 -.0112031 -.0041733 -.0070297 .0052839
dtf6 .0204818 .0301217 -.0096399 .0136979
logx1 -.0506932 -.0379132 -.01278 .3182976
x3 .0060748 .007656 -.0015812 .0028397
x4 .00141 .0017333 -.0003234 .0002283
x5 -.0000495 -.0000441 -5.40e-06 7.62e-06
x6 .0268641 .0164448 .0104193 .0089926
x7 .0475457 .0214958 .02605 .0695672
x8 .0011949 .0005773 .0006175 .000375
log_x9 -.0303707 -.0339612 .0035905 .0046843

b = Consistent under H0 and Ha; obtained from xtreg.
B = Inconsistent under Ha, efficient under H0; obtained from xtreg.

Test of H0: Difference in coefficients not systematic

chi2(13) = (b-B)'[(V_b-V_B)^(-1)](b-B)
= 15.60
Prob > chi2 = 0.2714
(V_b-V_B is not positive definite)

Picture 5:
. xtscc logfva dtf1 dtf2 dtf3 dtf4 dtf5 dtf6 logx1 x3 x4 x5 x6 x7 x8 log_x9, lag(1)

Regression with Driscoll-Kraay standard errors Number of obs = 255
Method: Pooled OLS Number of groups = 64
Group variable (i): quốcgia F( 14, 3) = 23.72
maximum lag: 1 Prob > F = 0.0120
R-squared = 0.6367
Root MSE = 0.1453


Drisc/Kraay
logfva Coefficient std. err. t P>t [95% conf. interval]

dtf1 .0705629 .0188894 3.74 0.033 .0104484 .1306774
dtf2 .0476133 .0115057 4.14 0.026 .0109972 .0842294
dtf3 .0757482 .0241137 3.14 0.052 -.0009922 .1524886
dtf4 -.2312862 .0487318 -4.75 0.018 -.3863726 -.0761998
dtf5 .0859264 .0154736 5.55 0.012 .0366825 .1351703
dtf6 .2385201 .0059737 39.93 0.000 .219509 .2575312
logx1 -.034425 .0207053 -1.66 0.195 -.1003186 .0314686
x3 .0124655 .0006431 19.38 0.000 .010419 .014512
x4 .002236 .0001477 15.14 0.001 .0017659 .0027061
x5 .0000336 .0002677 0.13 0.908 -.0008182 .0008854
x6 -.0627637 .0058983 -10.64 0.002 -.0815347 -.0439927
x7 .06392 .0114298 5.59 0.011 .0275454 .1002946
x8 -.0035081 .0003097 -11.33 0.001 -.0044938 -.0025224
log_x9 -.0566343 .0067057 -8.45 0.003 -.0779749 -.0352938
_cons 1.128491 .1688034 6.69 0.007 .591283 1.665698


I'm really sorry, I wanted to attach a photo but didn't know how. Can you try to see these hard to read lines and help me?
.

[Carlo Lazzaro]

Dung:
1) I would discard your first -regress- as it does not take the panel structure of your datsaset into account;
2) given the number of your panels, you should add the -robust- or the -vce(cluster panelid)- option to adjust your standard errors for heteroskedasticity and/or autocorrelation;
3) if you add, as you should what recommeded at point 2), you should switch from -hausman- to the community-contributed module -xtoverid-;
4) unless you detected across panel correlations in a T>N panel dataset (you are currently dealin with a N>T one), I can safely forget -xtscc-:
5) last but not least:

please use code delimiters when you share what you typed and what Stata gave you back (see the FAQ on how to use them). Thanks.

[Dung Delime]

Dear Carlo Lazzaro,
Thank you for your advice. I know that my model has heteroskedasticity and autocorrelation, so according to you advice, I add robust/ vce(cluster panelid) into FEM model. However, all my main dependent variables (DTF) are still not statistically significant. You had said that I should use xtoverid instead of hausman (for choosing between random or fixed effects ?). I will attach 3 pictures about my results (I don't know if I do right).
If I'm not mistaken, whether I run the Fixed or Random effect model, my DTF variable is statistically insignificant (P-value >0.05).
I wonder if there is any other method, because now I only know using POLS with xtscc will make my model better (I'm not sure but there are some sources said that we can use xtscc for N>T too). But I find it quite odd, because as you say, POLS is clearly not as suitable as the Fixed and Random effect Models (the results of running the model also say the same), but why only in POLS are my variables statistically significant?