Quick Question about p and t-value

[Kasper Rudy]

Hello gents.

I just have a quick question about p & t-value and significance.

I have done a fixed effects analysis in stata (screenshot attached), but I have problems interpreting the significance levels. The t-value is estimated to 1.88, which in a t-table (two-tailed, 59 degrees of freedom) should translate to a significance p-value on the 10% level. But stata indicates that it is only significant on a 15,7% level. Is there a good explanation for this?

Hope for a positive answer.

Best Regards,
Kasper

[Carlo Lazzaro]

Kasper:
as you imposed vce(robust), the degrees of freedom (4-1=3) are calculated on panels (4) instead of observations. You can test it with the following example:

. di ttail(3,1.88)
.07834754
. di .07834754*2
.15669508

the latter figure well approximates 0.157.

[Marcos Almeida]

Hello Kasper,

Apart from Carlo's insightful suggestions, I wish to underline that your model has huge CIs for most predictors. Even it they hadn't crossed the zero line (then, providing significant p-values), I fear you'd still have to face a "difficult" model, to say the least, in terms of precision. Maybe it's due to the small sample size. Last but no least, the model failed to provide the value for the F "omnibus" test.



Best,

Marcos

[Kasper Rudy]

Thank you for the really insightful answers.

But is there any way to run a Heteroskedasticity-Robust command without adjusting the degrees of freedom?


Best regards,

Kasper

[Carlo Lazzaro]

Kasper:
not that I know.

[Kasper Rudy]

Ok, thanks Carlo.

As our dataset contain heteroscedasticity, do you then recommend to run the regression without vce(robust)? If we use the robust command, our results will apparently be wrong.


Best regards,

Kasper

[Carlo Lazzaro]

Kasper.
i was probably unclear in my previous reply.
You're right in using vce(robust) if you suspect that your residuals are heteroskedastic.
However, in -xtreg- (but not in -regress-, for instance) this option is equivalent to -vce(cluster clusterid).
Hence, as in your example, the degrees of freedom are calculated referring to the number of clusters, instead of observations.

[Kasper Rudy]

Hello again, I just have one more quick question.

I also use First Differences in my calculations (alongside Fixed Effects as in my first post). Does anyone know why I get different answers between the two models (screenshot is attatched)? Can it be because of the fe's demeaning?

In the screenshot, the red marked coefficient is Fixed Effects, and the blue one is First Differences