[Joao Santos Silva]

Dear Giorgio Di Stefano,

As far as I can see in the post you refer to in #208 you are not interested in estimate quantile regressions, so I do not see how xtqreg can be of use to you.

Best wishes,

Joao

[Derick Almeida]

Dear Mr. Joao Santos Silva, I just wanted to double-check the validity of an algorithm in the xtqreg framework.

I've been working with the 'Quantiles-via-Moments' estimator on a panel data set and have recently implemented the jackknife bias correction algorithm that you so kindly shared on your webpage.

In doing so, I've noticed that the algorithm only performs the bias correction in the 5th decile. Is that correct? Using your own example in the do-file, I've tried to run the following code for the 1st decile coefficient, where I substitute, for example, the matrix V=e(V) for V1=e(V_1), as well as matrix Q50 = e(q) for matrix Q10= e(q_1) -- and all other matrices in the subsamples --, following the guidelines of 'help xtqreg' related to how each vector and matrix of the quantiles is stored, but I still get a conformability error r(503). In your opinion, what would be the best way to adapt it?

clear all
ssc install xtqreg, replace

cls
webuse nlswork
xtset idcode

by idcode: egen c=count(idcode)
set seed 123 
g s=runiform()<0.5

*** Simple MM-QR estimation with fixed effects
xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10, ls q(0.10(0.40)0.90)

*** MM-QR results with JK correction and clustered standard errors
qui bs if c>=10, cluster(idcode) rep(20): xtqreg ln_w  age ttl_exp tenure  not_smsa south
mat V1=e(V_1) 
matrix b=e(b_location)
matrix g=e(b_scale)
matrix Q10=e(q_1) 
scalar N=e(N)
qui xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10&s==0  
scalar N0=e(N)
matrix Q010=e(q_1)
matrix g0=e(b_scale)
qui xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10&s==1   
scalar N1=e(N)
matrix Q110=e(q_1)
matrix g1=e(b_scale)
mat bjk1=b+(2*g-g1*(N1/N)-g0*(N0/N))*(2*Q10-Q110*(N1/N)-Q010*(N0/N)) 
ereturn post bjk1 V1
ereturn display 

Thank you so much and kind regards,
Derick

[Joao Santos Silva]

Dear Derick Almeida,

Note that the code only works for one quantile at the time, so you can only have one quentile in the option q() and you have to change the names of the matrices for each quantile. Probably there is a more efficient way to do that...

Also note that selecting observations with c>=10 was specific to that example; you will not need that in your application.

Best wishes,

Joao

[Derick Almeida]

Dear Joao Santos Silva,

thank you for your valuable advice, including the fact that c>=10 is not an essential part of the script when it comes to other datasets . In fact, I reran the code, specifying separately the desired quantiles I needed, and the jackknife correction worked greatly for each one. The code I ran, based again in the do-file, for (let's say) the 1st decile is:

clear all
ssc install xtqreg, replace

cls
webuse nlswork
xtset idcode

by idcode: egen c=count(idcode)
set seed 123 
g s=runiform()<0.5

*** Simple MM-QR estimation with fixed effects
xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10, ls q(0.10)

*** MM-QR results with JK correction and clustered standard errors
qui bs if c>=10, cluster(idcode) rep(20): xtqreg ln_w  age ttl_exp tenure  not_smsa south, ls q(0.10)
mat V1=e(V) 
matrix b=e(b_location)
matrix g=e(b_scale)
matrix Q10=e(q) 
scalar N=e(N)
qui xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10&s==0, ls q(0.10)  
scalar N0=e(N)
matrix Q010=e(q)
matrix g0=e(b_scale)
qui xtqreg ln_w  age ttl_exp tenure  not_smsa south if c>=10&s==1, ls q(0.10)   
scalar N1=e(N)
matrix Q110=e(q)
matrix g1=e(b_scale)
mat bjk1=b+(2*g-g1*(N1/N)-g0*(N0/N))*(2*Q10-Q110*(N1/N)-Q010*(N0/N)) 
ereturn post bjk1 V1
ereturn display 

Best regards,
Derick

[Derick Almeida]

Dear Joao Santos Silva, another question has come up regarding xtqreg.

I want to conduct a linear constraint hypothesis test between two quantile coefficients of the xtqreg regression output, namely the 1st and the 9th deciles, to know if they are statistically different from each other. I have already read earlier posts, in which you recommend bootstrapping the regression, much in the same way as sqreg does. But I've been having som trouble getting the F or chi2 statistics for the test.

Let me first show my regression and hypothesis test under sqreg. This is just a toy example in which log_LP is the dependent variable, log_RD is the only independent variable and I want to calculate an estimate for the 1st and 9th decile coefficients.

sqreg log_LP log_RD, q(0.1,0.9) 

(fitting base model)

Bootstrap replications (20)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
....................

Simultaneous quantile regression Number of obs = 5,237
bootstrap(20) SEs .10 Pseudo R2 = 0.0786
.90 Pseudo R2 = 0.0364

------------------------------------------------------------------------------
| Bootstrap
log_LP | Coef. Std. Err. t P>|t| [95% Conf. Interval]
-------------+----------------------------------------------------------------
q10 |
log_RD | .3659615 .0139983 26.14 0.000 .338519 .393404
_cons | 2.760837 .0301514 91.57 0.000 2.701728 2.819947
-------------+----------------------------------------------------------------
q90 |
log_RD | .1975607 .0181288 10.90 0.000 .1620207 .2331007
_cons | 4.826183 .0286634 168.37 0.000 4.769991 4.882376
------------------------------------------------------------------------------

Then, to test whether the 1st and 9th coefficients are statistically the same:

--------------------------------------------------------------

test [q10=q90] : log_RD 

( 1) [q10]log_RD - [q90]log_RD = 0

F( 1, 5233) = 5.94
Prob > F = 0.0148
---------------------------------------------------------------

Now, conducting the bootstrap regression for panel data through the xtqreg command, separately for each quantile, we get:

bootstrap, reps(20): xtqreg log_LP log_RD, q(0.1) 

Bootstrap replications (20)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
....................

Bootstrap results Number of obs = 5,237
Replications = 20

------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
log_LP | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
log_RD | .4139783 .0252642 16.39 0.000 .3644614 .4634953
------------------------------------------------------------------------------

bootstrap, reps(20): xtqreg log_LP log_RD, q(0.9) 

Bootstrap replications (20)
----+--- 1 ---+--- 2 ---+--- 3 ---+--- 4 ---+--- 5
....................

Bootstrap results Number of obs = 5,237
Replications = 20

------------------------------------------------------------------------------
| Observed Bootstrap Normal-based
log_LP | Coef. Std. Err. z P>|z| [95% Conf. Interval]
-------------+----------------------------------------------------------------
log_RD | .3451415 .0179813 19.19 0.000 .3098987 .3803843
------------------------------------------------------------------------------

At this point, differently from sqreg that offered me equations named q10 and q90 to conduct the F test on the log_RD quantile coefficients, I can't find the right name or reference to use the information of both regressions in xtqreg to conduct the same test. I'm afraid it's because xtqreg does not estimate the equations simultaneously. I've looked into "help xtqreg" to figure out the right way to reference it in the test, but I had no success. Do you have an idea of how to carry on this exercise?

Thank you so much again.

All the best,
Derick

[Joao Santos Silva]

Dear Derick Almeida

It may be enough to run xtqreg for any quantile with the option ls. If a variable is significant in the scale function, its coefficients will vary across quantiles.

To you the bootstrap, what you have to do is to create a small program that estimates the two quantiles and computes the difference between the two coefficients. You then bootstrap that programme.

Best wishes,

Joao

[Justin Matz]

Joao Santos Silva Dear professor,

I was wondering why, when I implement ". xtqreg Return CPI", why do I obtain :

"MM-QR regression results

It is not possible to identify the conditional quantiles with T<3; the estimator is valid onl
> y for large T"

Knowing that I have 166 observations. What Can I do to solve this problem?

Kind regards,
Matz Justin

[Joao Santos Silva]

Dear Justin Matz,

I assume you have a panel; what are the values of N and T in your dataset?

Best wishes,

Joao

[Justin Matz]

Joao Santos Silva,

I have an excel sheet with 80 variables, in monthly basis, spread over 13 years.
Those variables are macroeconomic variables and I want to regress Returns (dependant variable) on each independent variable.

Kind regards,
Matz Justin

[Justin Matz]

Joao Santos Silva,

I have an excel sheet with 80 variables, in monthly basis, spread over 13 years.

This sheet is an average of 13 other sheets, each representing a different country.

[Joao Santos Silva]

Dear Justin Matz,

Then, just use xtset to define your panel and then use the xtqreg command.

Best wishes,

Joao

[Justin Matz]

Dear professor Joao Santos Silva,

Yeah, it is what I wanted to do, but obtain this error message : "It is not possible to identify the conditional quantiles with T<3; the estimator is valid onl
> y for large T"

My complete code : ". xtset Date

Panel variable: try (weakly balanced)
Time variable: Date, 01jan2008 to 01dec2021
Delta: 1 day

. xtqreg Return CPI



MM-QR regression results

It is not possible to identify the conditional quantiles with T<3; the estimator is valid onl
> y for large T"

Do you know how I can solve this issue?

Kind regards,
Matz Justin

[Joao Santos Silva]

What is the variable "try"?

[Justin Matz]

Sorry, it is an error (somethings that I tried)

complete code : ". xtset Date

Panel variable: Date (balanced)

. xtqreg Return CPI



MM-QR regression results

It is not possible to identify the conditional quantiles with T<3; the estimator is valid onl
> y for large T
"

[Joao Santos Silva]

When using xtset, you need to define the id identifier and then the sate; you are not including the id identifier.