Quantile regression t-stata insignificant

[Aamina Khurraa]

Hi Joao Santos Silva Clyde Schechter

I am running qreg for a panel of 372 firms with a monthly time observation of 237 months each. This makes a total of 88,164 observations in my data. I am focusing on 99 and 95 percentiles of the data. All my t stats and p values are insignificant, although there is a strong relationship between the variables that exist. Any Idea what is going wrong and how this issue can be addressed or resolved?

Best Regards

[FernandoRios]

Hi Aamina
Without known more about the exact syntax that you are applying, I think the problem is related to looking at the very top of the distribution. This means there will be few observations "above" the conditional quantile, which makes it very hard to do a precise identification of the coefficients.
So couple of suggestions:
you could use "(SSC) qreg2 " which provides an alternative estimation of standard errors.
You may want to use "(SSC) xtqreg" to handle the fixed effects.
HTH
Fernando

[Joseph Coveney]

Maybe qreg isn't suitable for a panel dataset? Maybe you misspecified the model or made a syntax error? Maybe distribution tails, like the 95th and 99th quantiles, are hard, even with a lot of data, especially when you have a lot of predictors?

[Joao Santos Silva]

Dear Aamina Khurraa,

I concur with what FernandoRios said. Make sure you use the right specification and the right standard errors, but identifying effects on the extreme tails can be challenging.

Best wishes,

Joao

[Aamina Khurraa]

Tahnk you all FernandoRios Joao Santos Silva

When i used qreg2, the p values became significant for 99 percentile but t values are a way huge numbers, something ;like 3.975e+15......

Can you please explain to me what transformation it made while I used qreg2 to make the estimates significant. Moreover, what would such a big number of t stat depict?

[Joao Santos Silva]

Dear Aamina,

Please show us exactly what you did and the results you get. Anyway, using qreg2 you should probably cluster by company.

Best wishes,

Joao

[Aamina Khurraa]

Dear Joao Santos Silva

The code that I used to run qreg2 for Y (dep var) over X1 X2 X3 (indep.vars) is as below:

program define one_id
qreg2 Y X1 X2 X3, q(0.99)
foreach v of varlist X1 X2 X3 {
gen coeff_`v' = _b[`v']
gen se_`v' = _se[`v']
gen t_`v' = coeff_`v'/se_`v'
gen p_`v' = 2*t(e(df_r), -abs( t_`v'))
}
gen constant_term = _b[_cons]
gen se_constant = _se[_cons]
predict V
exit
end
runby one_id, by(FI) status


The summary of t tests and p values is also presented below:

Variable	stats
t_X1	3.975e+15
t_X2	3.276e+14
t_X3	-2.393e+15
p_X1	.018
p_X2	.016
p_X3	.023

[Joao Santos Silva]

Dear Aamina Khurraa,

Thanks for sharing this. Please check whether you are using the latest version of qreg2, earlier versions did not save e(df_r).

Best wishes,

Joao

[Aamina Khurraa]

Dear Joao Santos Silva

Can you please share the link of the newest version as I downloaded whatever was available in ssc. Also, how would not saving df change the t stats. would you like to comment the huge values of the t stats that I am getting.

Best

[Aamina Khurraa]

I have just checked again through ssc install qreg2 and it returned the syntax as "all files are already installed and upto date"

[Joao Santos Silva]

What is the e(df_r) that you get? What if you use the normal instead of the t?