Cannot run Durbin Watson or Breusch-Godfrey tests

[Sophie Starck]

Hi,

I am trying to run serial correlation (DW and Breusch-Godfrey) tests for my panel data set.
I type in the follwing commands:

reg ROA Kortfrist_sku Långfrist_sku Tot_sku ln_storlek Ålder Likviditet Ränta
bgodfrey

tsset id År
dwstat

reg ROA Kortfrist_sku Långfrist_sku Tot_sku ln_storlek Ålder Likviditet Ränta
estat dwatson

But for all I get the error message "sample may not include multiple panels". How do I solve this?

[Andrew Musau]

You have panel data, so look at

search xtserial

SJ-3-2 st0039 . . Testing for serial correlation in linear panel-data models
(help xtserial if installed) . . . . . . . . . . . . . D. M. Drukker
Q2/03 SJ 3(2):168--177
test for serial correlation in random- or fixed-effects
one-way models that can be applied under general conditions

But as I remarked in your earlier thread (https://www.statalist.org/forums/for...tandard-errors), with a wide panel, you just want to cluster on the panel variable and forget about testing for serial correlation.

[Sophie Starck]

Thank you for your response. I have a wide panel where N>T, where the panel data are for companies so I should write vce( cluster company) then? As per the other thread, I will be using robust standard errors, so how do I go about combining the robust and the clustering?
This is my code right now (with the robust but without the clustering):

qui reg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, vce(robust)
qui outreg using Regression1.doc, nocons addrow(FE,Nej) se bdec (3) starlevels(10 5 1) sigsymbols(*,**,***) starloc(1) summstat(r2 \ N) summtitle(R2 \ Observationer) summdec(2 2 0) title(Tabell X. Regression \ Beroende variabel: ROA) ctitle("", "" \ "Förklaringsvariabler" \ "" "(1)") blankrows nolegend varlabels replace

Kind regards,
Sophie Starck

[Andrew Musau]

For panel data estimators in Stata, \(\text{cluster(panelvar)}\equiv \text{vce(robust)}\), where panelvar is the panel identifier. Try it out yourself as below.

webuse nlswork, clear
xtset idcode year
xtreg ln_wage tenure hours, fe cluster(idcode)
xtreg ln_wage tenure hours, fe vce(robust)

Res.:

. xtreg ln_wage tenure hours, fe cluster(idcode)

Fixed-effects (within) regression               Number of obs     =     28,036
Group variable: idcode                          Number of groups  =      4,698

R-squared:                                      Obs per group:
     Within  = 0.0975                                         min =          1
     Between = 0.1952                                         avg =        6.0
     Overall = 0.1364                                         max =         15

                                                F(2,4697)         =     613.32
corr(u_i, Xb) = 0.1376                          Prob > F          =     0.0000

                             (Std. err. adjusted for 4,698 clusters in idcode)
------------------------------------------------------------------------------
             |               Robust
     ln_wage | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      tenure |   .0342778    .000979    35.01   0.000     .0323584    .0361972
       hours |  -.0003775   .0004802    -0.79   0.432    -.0013189    .0005639
       _cons |   1.584114   .0176755    89.62   0.000     1.549462    1.618766
-------------+----------------------------------------------------------------
     sigma_u |  .39087412
     sigma_e |  .30365406
         rho |  .62363197   (fraction of variance due to u_i)
------------------------------------------------------------------------------

. xtreg ln_wage tenure hours, fe vce(robust)

Fixed-effects (within) regression               Number of obs     =     28,036
Group variable: idcode                          Number of groups  =      4,698

R-squared:                                      Obs per group:
     Within  = 0.0975                                         min =          1
     Between = 0.1952                                         avg =        6.0
     Overall = 0.1364                                         max =         15

                                                F(2,4697)         =     613.32
corr(u_i, Xb) = 0.1376                          Prob > F          =     0.0000

                             (Std. err. adjusted for 4,698 clusters in idcode)
------------------------------------------------------------------------------
             |               Robust
     ln_wage | Coefficient  std. err.      t    P>|t|     [95% conf. interval]
-------------+----------------------------------------------------------------
      tenure |   .0342778    .000979    35.01   0.000     .0323584    .0361972
       hours |  -.0003775   .0004802    -0.79   0.432    -.0013189    .0005639
       _cons |   1.584114   .0176755    89.62   0.000     1.549462    1.618766
-------------+----------------------------------------------------------------
     sigma_u |  .39087412
     sigma_e |  .30365406
         rho |  .62363197   (fraction of variance due to u_i)
------------------------------------------------------------------------------

.

This is discussed in

help xtreg

[Sophie Starck]

Hi. I've tried running the regressions using both vce(robust) and cluster(id). Although the coefficients are the same, the standard errors are different, meaning the significance is lost when using cluster(id). If they are equivalent to each other, what am I doing wrong?

This is what I am typing:

qui reg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, vce(robust)
qui outreg using Regression1.doc, nocons addrow(FE,Nej) se bdec (3) starlevels(10 5 1) sigsymbols(*,**,***) starloc(1) summstat(r2 \ N) summtitle(R2 \ Observationer) summdec(2 2 0) title(Tabell X. Regression \ Beroende variabel: ROA) ctitle("", "" \ "Förklaringsvariabler" \ "" "(1)") blankrows nolegend varlabels replace 

qui reg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, cluster(id)
qui outreg using Clustertest.doc, nocons addrow(FE,Nej) se bdec (3) starlevels(10 5 1) sigsymbols(*,**,***) starloc(1) summstat(r2 \ N) summtitle(R2 \ Observationer) summdec(2 2 0) title(Tabell X. Regression \ Beroende variabel: ROA) ctitle("", "" \ "Förklaringsvariabler" \ "" "(1)") blankrows nolegend varlabels replace 

Which one should I go with?

Kind regards,
Sophie Starck

[Andrew Musau]

regress is not a panel data estimator in Stata speak. Panel data estimators are those prefixed -xt-, e.g, xtreg, xtpoisson, and xtlogit. If you use regress with panel data, -vce(robust)- is not valid as this calculates White standard errors. You must use -cluster(panelvar)- and include the fixed effects as indicators. Relevant reading is Stock, J. H., and M. W. Watson. 2008. Heteroskedasticity-robust standard errors for fixed effects panel data regression. Econometrica 76: 155–174. https://doi.org/10.1111/j.0012-9682.2008.00821.x

xtset id year
xtreg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, fe cluster(id)

or even adding time-effects:

xtset id year
xtreg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta i.year, fe cluster(id)

[Sophie Starck]

Okay I will read that. At first I did try xtreg but then the R2 would not show, which is why i went with reg instead. Do you know why the R2 did not show?

[Andrew Musau]

xtreg, fe reports 3 different R2 statistics: see the output in #4. The manual explains the calculations of these R2 statistics. If you want the OLS R2 statistic, run regress or areg with clustering.

regress ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, absorb(id) cluster(id)

or with time-effects:

regress ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta i.year, absorb(id) cluster(id)

[Sophie Starck]

Thank you for your help.
I am only planning on including fixed effects in the last regression (I understand how to run this one correctly now), but for the previous ones I am planning on using robust standard errors using vce(robust). If I understand you correctly, you mean that this will lead to the use of White standard errors which is not correct? From what I could tell by the article, I interpret it as it is only combined with fixed effects that I should not use vce(robust) and cluster instead? So do I always have to use fixed effects and cluster when working with panel data then?

[Andrew Musau]

Pooled OLS (POLS) and Random Effects (RE) in general are not recommendable techniques as they do not control for the individual heterogeneity. So the first priority is to get consistent estimates of the parameters before you go ahead and consider other issues. If you use regress, \(\text{vce(robust)= White standard errors}\), as I stated previously. These are not valid for panel data. You will want to cluster even if you estimate POLS or RE.

[Sophie Starck]

Thank you for your response. I just want to double check with you. When I write:

qui xtreg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, fe cluster(id)
qui outreg using RegressionNY.doc, nocons addrow(FE,Ja) se bdec (3) starlevels(10 5 1) sigsymbols(*,**,***) starloc(1) summstat(r2 \ r2_a \ N) summtitle(R2 \ Justerat R2 \ N) summdec(2 2 0) title(Tabell X. Regression \ Beroende variabel: ROA) ctitle("", "" \ "Förklaringsvariabler" \ "" "(1)") blankrows nolegend varlabels replace

Does it use POLS together with FE and clustering or is it just FE and clustering?

[Andrew Musau]

Pooled OLS is just OLS on panel data:

regress ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, cluster(id)

What you have below is the fixed effects model estimated using the within-estimator (or the fixed effects estimator).

xtreg ROA Kortfrist_sku Ålder Likviditet ln_storlek Ränta, fe cluster(id)

[Sophie Starck]

Ok I got it now. Thank you!