Re: st: re: Troubleshooting 'not sorted' and 'not regularly spaced' errors

[Michael Hanson <mshanson@mac.com>]

Tom:

Kit is, of course, correct that one can use -ivreg2- (from SSC; runs  
in Stata 9) or -ivregress- (Stata 10 only) to mimic the -newey- and - 
newey2- estimators, as all allow you to get Newey-West style  
heteroskedasticity and autocorrelation consistent (HAC) standard  
errors.  There are a few additional points worth noting in order to  
make sure these various Stata commands are delivering the HAC  
estimator that you have in mind.  And when you have time series with  
"gaps" and use lagged regressors, you may need to pay close attention  
to whether you -tsset- or -xtset- to a Stata date variable (with  
gaps) or a linear time trend (without gaps).

1. If m is the number of lags you would specify in -newey- or - 
newey2-, then use m for the -ivregress, vce(hac nw m)- command, but  
n=m+1 for the -ivreg2, bw(n)- command.  As you point out, it takes  
some effort or experimentation to extract this information from the - 
ivreg2- help file.

2. To match the HAC SE estimates of -newey- or -newey2-, you will  
need to add the options -robust small- to an -ivreg2- estimation and  
the option -small- to an -ivregress- estimation.

3. If you -tsset- (or -xtset-) your data to a Stata date variable  
with gaps, you may not get the same HAC SE estimates from -newey2,  
force- as from -ivreg2- or -ivregress- when lagged regressors are  
included.  You can modify my earlier example program to demonstrate  
that fact (left as an exercise).  This problem may or may not occur  
in your dataset.

4. If you instead -tsset- (or -xtset-) to a sequential linear time  
trend (-gen t = _n-), then you will get the same results for all  
three commands -- but you will have, in effect, told Stata that there  
are no gaps in your data.  Indeed, you likely will have included a  
larger set of observations in your regressions, so both the  
coefficient estimates and the SE estimates are likely to differ from  
those in (3).

Which of (3) or (4) above is preferable depends on how you think the  
"gaps" should be handled.  Stata constructs lags by strictly using  
the prior observation at the frequency specified.  For example, if  
you have business daily data (as in the example program in my prior  
message) and -tsset- to a daily Stata date variable, Stata will not  
construct any lags for a Monday, as the prior daily observation  
(Sunday) is missing in that financial data set.  However, you may  
well want Stata to treat Friday as the prior observation -- which  
will require you to pay close attention to the construction of your  
data and then to -tsset- to a time trend.  (Other statistical  
packages that specialize in time series estimation, such as RATS and  
(IIRC) EViews, have an explicit "business daily" frequency as this  
problem arises regularly in applications with financial data.  One  
still has to think about how to handle holidays in that case.)  If  
instead you have data with regular gaps -- such as every two or every  
five years of annual data -- you can use the -delta- option of - 
tsset- or -xtset- to have Stata treat the prior two- or five-year  
observation as the lagged value.

HTH,
Mike

On Oct 13, 2008, at 12:27 PM, Thomas Jacobs wrote:

> Kit,
>
> Thanks for introducing me to ivregress and ivreg2 for use in newey
> west estimation.  I was unfamiliar with both commands.  I confirmed
> they run just as you indicated with both the examples Michael Hanson
> provided as well as the questions from my original post.  No problems
> with either time series commands or gaps.
>
> Is the translation between newey west lags and bartlett kernel
> bandwidth simply bandwidth = lags +1 as it appears?  I was unsure from
> the ivreg2 help file.  Thanks again.
>
> Tom
>
> On Sat, Oct 11, 2008 at 3:07 PM, Kit Baum <baum@bc.edu> wrote:
> > < >
> > Re Michael Hanson's comments on David Roodman's -newey2-:  David  
> > has on more
> > than one occasion suggested that those wanting to calculate Newey- 
> > West
> > standard errors in an OLS regression should just use -ivreg2- of  
> > Baum,
> > Schaffer, Stillman. Despite its name it is happy to estimate OLS  
> > models
> > without any instruments, and it can estimate Newey-West standard  
> > errors as
> > well as a variety of other HAC models. E.g.
> >
> > ivreg2 irx t, robust bw(9)
> >
> > Notice that it reports the presence of gaps but does not choke on  
> > them.
> > -ivreg2- requires Stata 9.2.
> >
> > In fact, contradictory to its syntax diagram, -ivregress- can do  
> > this as
> > well. The -ivregress- syntax states that the (varlist2 = varlist_iv)
> >  component is mandatory, but the program does not enforce this:
> >
> > ivregress 2sls irx t, vce(hac bartlett 8)
> >
> > works, making no mention of gaps.  -ivregress- requires Stata 10.
> >
> > Kit Baum, Boston College Economics and DIW Berlin
> > http://ideas.repec.org/e/pba1.html
> > An Introduction to Modern Econometrics Using Stata:
> > http://www.stata-press.com/books/imeus.html
*
*   For searches and help try:
*   http://www.stata.com/help.cgi?search
*   http://www.stata.com/support/statalist/faq
*   http://www.ats.ucla.edu/stat/stata/