weights and two way fixed effects in xtqreg

[Hamidreza Bakhtiarizadeh]

I am using xtqreg command in Stata to implement a quantile regression. I have two questions:

(1) How can I add weight for each observation in the regression? aweight does not work in xtqreg.

(2) How can I have two fixed effects? I want to have both year and id fixed effects.

[Nick Cox]

Cross-posting (as suggested, but without a flag) of https://stackoverflow.com/questions/...ects-in-xtqreg

[FernandoRios]

Hi Hamidreza
try mmqreg. (Ssc) Implements the same estimator but allows for weights and multiple fixed effects
Fernando

[Hideto Koizumi]

Hi FernandoRios,

Thank you for making mmqreg available for us! it appears that pweights are not allowed. I am wondering, is there any theoretical reason to not include it as an option?

[FernandoRios]

None. It’s probably something I need to modify
simply try
w=weight and robust

[Hideto Koizumi]

Thank you FernandoRios for your prompt response! Got it.
I need to use pweights rather than aweights, so unfortunately your suggestion won't work. I will look into the adofile and see if I can tweak around to allow pweights.

[FernandoRios]

I understand that, but pweights and just "weights" with Robust will give you the same result.
The only reason I kept AW on mmqreg, its because its more flexible for other stuff i need to do in the program.
the estimation itself doesnt use the Standard errors from the regressions, it recreates them manually.
F

[Hideto Koizumi]

Hi FernandoRios,
I guess I am misunderstanding your suggestion? so your suggestion,

w=weight

with robust option results in the identical estimates as

aw=weight

without robust option.
Are you saying that the weights option in mmqreg always use pweights??

Rather than using my data, although I know it doesn't have sampling weights in it, here I use the nlswork dataset that you used in your mmqreg page, just for demonstration:

webuse nlswork, clear

. mmqreg ln_w ttl_exp tenure not_smsa south [w=age], abs(idcode) robust
(sampling weights assumed)
(24 missing values generated)
WARNING: some fitted values of the scale function are negative
Consider using a different model specification
17 Observations have negative predicted Scale values
MM-qreg Estimator
Number of obs = 28093
Quantile:  50
------------------------------------------------------------------------------
             |               Robust
     ln_wage | Coefficient  std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
location     |
     ttl_exp |   .0234697   .0008002    29.33   0.000     .0219013     .025038
      tenure |   .0115601   .0009787    11.81   0.000     .0096418    .0134784
    not_smsa |  -.0972124   .0110891    -8.77   0.000    -.1189466   -.0754782
       south |  -.0663628   .0126408    -5.25   0.000    -.0911384   -.0415872
       _cons |   1.556874   .0068966   225.75   0.000     1.543356    1.570391
-------------+----------------------------------------------------------------
scale        |
     ttl_exp |   .0005567   .0004696     1.19   0.236    -.0003637    .0014771
      tenure |  -.0047032   .0005822    -8.08   0.000    -.0058442   -.0035621
    not_smsa |  -.0031209   .0063986    -0.49   0.626    -.0156619    .0094201
       south |  -.0092352   .0073213    -1.26   0.207    -.0235846    .0051142
       _cons |   .1985896   .0039068    50.83   0.000     .1909324    .2062468
-------------+----------------------------------------------------------------
qtile        |
     ttl_exp |   .0235019      .0008    29.38   0.000     .0219338    .0250699
      tenure |    .011288   .0009809    11.51   0.000     .0093654    .0132106
    not_smsa |  -.0973929   .0110596    -8.81   0.000    -.1190693   -.0757165
       south |   -.066897   .0126719    -5.28   0.000    -.0917336   -.0420605
       _cons |   1.568362   .0071953   217.97   0.000     1.554259    1.582464
------------------------------------------------------------------------------


. mmqreg ln_w ttl_exp tenure not_smsa south [aw=age], abs(idcode)
(24 missing values generated)
WARNING: some fitted values of the scale function are negative
Consider using a different model specification
17 Observations have negative predicted Scale values
MM-qreg Estimator
Number of obs = 28093
Quantile:  50
------------------------------------------------------------------------------
     ln_wage | Coefficient  Std. err.      z    P>|z|     [95% conf. interval]
-------------+----------------------------------------------------------------
location     |
     ttl_exp |   .0234697   .0007323    32.05   0.000     .0220343     .024905
      tenure |   .0115601   .0008782    13.16   0.000     .0098389    .0132813
    not_smsa |  -.0972124   .0113173    -8.59   0.000    -.1193939   -.0750309
       south |  -.0663628   .0129058    -5.14   0.000    -.0916576    -.041068
       _cons |   1.556874   .0071701   217.13   0.000      1.54282    1.570927
-------------+----------------------------------------------------------------
scale        |
     ttl_exp |   .0005567   .0004888     1.14   0.255    -.0004014    .0015148
      tenure |  -.0047032   .0005862    -8.02   0.000    -.0058521   -.0035543
    not_smsa |  -.0031209   .0075545    -0.41   0.680    -.0179275    .0116856
       south |  -.0092352   .0086148    -1.07   0.284    -.0261199    .0076496
       _cons |   .1985896   .0047862    41.49   0.000     .1892088    .2079703
-------------+----------------------------------------------------------------
qtile        |
     ttl_exp |   .0235019   .0007341    32.02   0.000     .0220631    .0249406
      tenure |    .011288   .0008814    12.81   0.000     .0095605    .0130156
    not_smsa |  -.0973929   .0113443    -8.59   0.000    -.1196274   -.0751584
       south |   -.066897   .0129368    -5.17   0.000    -.0922528   -.0415413
       _cons |   1.568362   .0074375   210.87   0.000     1.553785    1.582939
------------------------------------------------------------------------------


.

Would you mind clarifying your suggestion a bit?
Thank you!

[FernandoRios]

So, MMQREG uses "AW" just on the surface. Internally, AW were easier to use for some other process like demeaning.
Now Standard errors
The default option (not robust) uses standard errors that correct for a specific type of heteroskedasticity. Think GLS standard errors (for exact formulas, check Machado and Santos-Silva paper in the references)
When you use Robust (with weights) I use an equivalent to the Sandwich Formula Heteroskedasticity consistent. You can also think of it as GMM Standard errors.
Cluster Standard errors also follows GMM clustered Standard errors
For what you need to work with. using PW is the equivalent to using AW with Robust. That is what i do here.

[Hideto Koizumi]

Ah, I see! I now fully understood. Thank you very much for your prompt and detailed explanation FernandoRios!