Re: st: RE: ivreg2 bandwidth

[Mirko <mirko.moro@gmail.com>]

Mark,

I was just wondering whether you could provide some references with
respect to the choice of the bandwidth and the kernel. And how this
choice affects the weak instruments tests. In few words this is what
happens when varying bandwidth and kernels

a) if I run IV regressions with non-truncated kernel and HAC standard
errors, my coefficients are all significant at 1% level (using many
specifications and IV estimators), but the instruments are weak (F
statistics ranging from 3 to 7);

b) if I run IV regressions using truncated kernel, my coefficients are
less robust across different regressions and in some other specific
cases never significant. The (AP) F statistic is now around 9.6 or
10.4, with p-values ranging from 0.001 to 0.01,  suggesting that the
instruments are OK.

I wish to know more about the choice of the kernel and bandwidth as
the story to tell very much depends upon this. Maybe is just better to
show everything?

Mirko


On 13 October 2011 16:51, Schaffer, Mark E <M.E.Schaffer@hw.ac.uk> wrote:
> Mirko,
>
> It depends partly on what you think the source of autocorrelation is.
>
> If the autocorrelation is purely the result of the fact that these are 3-day MAs - so that the autocorrelation will disappear after 3 days - then the right kernel to use is the truncated kernel (a.k.a. "Hansen-Hodrick") with a bandwidth of 3.  It's not guaranteed to be PD in finite samples, but this might not be a problem in practice in your case.
>
> If you suspect that there is autocorrelation beyond the 3rd lag, you could try the automatic bandwidth selection option - just say bw(auto).
>
> HTH,
> Mark
>
>> -----Original Message-----
>> From: owner-statalist@hsphsun2.harvard.edu
>> [mailto:owner-statalist@hsphsun2.harvard.edu] On Behalf Of Mirko
>> Sent: 13 October 2011 15:21
>> To: statalist@hsphsun2.harvard.edu
>> Subject: st: ivreg2 bandwidth
>>
>> Dear all,
>>
>> I am estimating a times-series equation where the dependent
>> variable and the endogenous variable are 3-day moving averages.
>>
>> I am using -ivreg2- with -gmm2s- and -robust- to obtain
>> heteroskedasticity and autocorrelation-robust standard errors such as
>> this:
>>
>> qui count
>> local band = round(r(N)^1/3)
>> ivreg2 y x1 (x2=  z1 z2), gmm2s robust bw(`band') first
>>
>> I am not sure about the correct bandwidth specification in
>> this specific case as I am using moving averages. For a
>> Bartlett kernel function, it is usually suggested to use
>> N^1/3. However, I am not sure whether this is correct
>> specification when using moving averages.
>>
>> I would be grateful to receive any suggestion.
>>
>> Best wishes,
>> Mirko
>> --
>> Mirko Moro
>> Lecturer in Economics
>> Economics Division
>> University of Stirling
>> FK94LA
>> Scotland (UK)
>> t: +44(0)1786467479
>> f: +44(0)1786467469
>> e: mirko.moro@stir.ac.uk
>>
>> *
>> *   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/
>>
>
>
> --
> Heriot-Watt University is a Scottish charity
> registered under charity number SC000278.
>
> Heriot-Watt University is the Sunday Times
> Scottish University of the Year 2011-2012
>
>
>
> *
> *   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/
>

*
*   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/