[Joro Kolev]

Perfect! Thanks very much.

You are welcome, but wait, dont go so fast

Great investigative work on figuring out what -newey- with the force option does in your post #13 !

Do you see any problem with this approach? Do you think that there is something wrong going on here with using -newey- on panel data or not?

I dont see nothing wrong, but Eric and Sebastian seem suspicious of this whole -newey, force- arrangement.

[Andrew Musau]

Joro Kolev, I seem to remember from a previous post that if you combine the same time periods and give them different values, you need to adjust the lag-length. At least an experiment using xtscc (from SSC) suggests so too. But generally, the approach is fine.

. webuse grunfeld, clear

. xtset company year
       panel variable:  company (strongly balanced)
        time variable:  year, 1935 to 1954
                delta:  1 year

. xtscc invest mvalue kstock, lag(3)

Regression with Driscoll-Kraay standard errors   Number of obs     =       200
Method: Pooled OLS                               Number of groups  =        10
Group variable (i): company                      F(  2,    19)     =     93.20
maximum lag: 3                                   Prob > F          =    0.0000
                                                 R-squared         =    0.8124
                                                 Root MSE          =   94.4084

------------------------------------------------------------------------------
             |             Drisc/Kraay
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .1155622   .0130406     8.86   0.000     .0882679    .1428564
      kstock |   .2306785   .0509161     4.53   0.000     .1241098    .3372472
       _cons |  -42.71437   12.65602    -3.38   0.003    -69.20373   -16.22501
------------------------------------------------------------------------------


. sort company year

. gen t=_n

. xtset company t
       panel variable:  company (weakly balanced)
        time variable:  t, 1 to 200
                delta:  1 unit

. xtscc invest mvalue kstock, lag(3)

Regression with Driscoll-Kraay standard errors   Number of obs     =       200
Method: Pooled OLS                               Number of groups  =        10
Group variable (i): company                      F(  2,   199)     =     76.56
maximum lag: 3                                   Prob > F          =    0.0000
                                                 R-squared         =    0.8124
                                                 Root MSE          =   94.4084

------------------------------------------------------------------------------
             |             Drisc/Kraay
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .1155622   .0107957    10.70   0.000     .0942735    .1368508
      kstock |   .2306785   .0637743     3.62   0.000     .1049183    .3564387
       _cons |  -42.71437   14.88858    -2.87   0.005    -72.07399   -13.35475
------------------------------------------------------------------------------


. xtscc invest mvalue kstock, lag(9)

Regression with Driscoll-Kraay standard errors   Number of obs     =       200
Method: Pooled OLS                               Number of groups  =        10
Group variable (i): company                      F(  2,   199)     =     55.62
maximum lag: 9                                   Prob > F          =    0.0000
                                                 R-squared         =    0.8124
                                                 Root MSE          =   94.4084

------------------------------------------------------------------------------
             |             Drisc/Kraay
      invest |      Coef.   Std. Err.      t    P>|t|     [95% Conf. Interval]
-------------+----------------------------------------------------------------
      mvalue |   .1155622   .0130144     8.88   0.000     .0898983     .141226
      kstock |   .2306785   .0679612     3.39   0.001     .0966619     .364695
       _cons |  -42.71437   15.58115    -2.74   0.007    -73.43973   -11.98901
------------------------------------------------------------------------------

In the above, a lag of 9 is a better approximation. However, the user-written command newey2 (from SSC) does no adjustment. Maybe the author David Roodman can shed more light on this.

[Eric de Souza]

This paper provides a survey of robust standard error estimation for panel models (including Newey-West estimators):
https://www.jstatsoft.org/article/view/2476/0
There are also routines in R with replication material.
I have not gone through it as yet.

[David Roodman]

I wrote newey2 in part to handle panel data properly. But that was a long time ago--2002--and I haven't touched the program and years. Actually I hacked together ivreg and newey to make newey2.

I believe ivreg2 offers Newey-West errors and probably handles panels properly too, and is a much more versatile program anyway. And it can do OLS.

If the panel is wide, it does seem like you should just cluster by individual.

If you are worried about two-way dependence, also consider multiway clustering. waldtest, in my boottest package, can multiway-cluster results after regress and other commands. ivreg2 also offers multiway clustering with a slightly different small-sample correction.

--David

[Joro Kolev]

Thank you for the tips, David.

Apparently the current version of -ivreg2- does everything that appeared in this thread, including HAC with automatic bandwidth selection. ( I just installed the most recent version of -ivreg2- and read the help.)

[Eric de Souza]

@Joro: to use -ivreg2-, you have to also download -ranktest-.
Mark Schaffer has also written a programme called -xtivreg2- which is a panel data wrapper for -ivreg2-. It only supports fixed effect (fe) and first difference (fd) estimators. It does not support factor variables.

[Saunok Chakrabarty]

Hi guys, I had a question here. Does newey-west estimation automatically incorporate time- fixed effects? Or do we have to specify it?

Also, how do you decide what the optimal lag-length is?

[Joro Kolev]

No, -newey- does not automatically incorporate any fixed effects, it is just linear regression with Newey-West standard errors.

If you want to have optimal lag length selection, you need to have a command that knows how to do optimal length selection, e.g., -ivreg2- and -ivregress- know how to choose lag length optimally, -newey- and -newey2- do not know how to do that. In the latter case you can use a mechanical rule of thumb such as Stock and Watson's 0.75*(T)^(1/3).

Hi guys, I had a question here. Does newey-west estimation automatically incorporate time- fixed effects? Or do we have to specify it?

Also, how do you decide what the optimal lag-length is?

[Sally Ahmed]

Hello, I am trying to run Newey and west regression however I assumed that the lag is 0 I really don't understand why sometimes it is 0 and other time is 3? Also I want to have the R2 and adjusted R2 in the regression output but for some reason is not shown here is the command I am using:

newey AQ ESGCombinedScore SIZE_W GROWTH_W RETURNONASSETS_W MTB_W RDINT_W LEV_W CYCLE_W WGI i.year i.IndustryX if year>2001 & year<2019, lag (0)
estimates store NEWcombined

and this is the code I am using for exporting the regression output

esttab NEWcombined using newregression88.rtf, b t parentheses r2(%8.2f) ar2(%8.2f) replace nogap star(* 0.10 ** 0.05 *** 0.01) compress
estimates clear

[Andrew Musau]

It appears that you have panel data. What is your \(N\) and \(T\)? Lag 0 is not meaningful for the newey command, which is linear regression with Newey-West standard errors (there is no such thing as Newey-West regression). It (lag 0) is equivalent to running regress with robust standard errors. The lags specify the maximum order of autocorrelation.

[Sally Ahmed]

The N is 5493 from 2002 to 2018.

[Andrew Musau]

Use xtreg or reghdfe (SSC) and cluster on the panel identifier. Cluster-robust standard errors in wide panels are robust to heteroscedasticity and arbitrary forms of autocorrelation. I do not see the advantage of Newey-West standard errors in this instance.

[Sally Ahmed]

Ok can I ask why Newey-West is not suitable here? Also based on what I decide the lag number?

[Andrew Musau]

As I said, the Newey-West method corrects the standard errors for autocorrelation. You already get this with cluster-robust standard errors in wide panels, plus you are taking into account within-cluster correlations. For optimal lag lengths, see #4 in the linked thread.

https://www.statalist.org/forums/for...lation-command

[Sally Ahmed]

okay thank you for your help.