[Date Prev][Date Next][Thread Prev][Thread Next][Date index][Thread index]

st: Re: Weak instruments

From   "Austin Nichols" <[email protected]>
To   [email protected]
Subject   st: Re: Weak instruments
Date   Wed, 9 Jan 2008 14:56:45 -0500


The Kleibergen-Paap statistic is indeed the generalization of the
Cragg-Donald statistic.  There is no table of critical values, but you
can use the critical values for the Cragg-Donald statistic from Stock
and Yogo until someone generates a newer version.  A Cragg-Donald
statistic of 100 is good news, assuming you have not generated
instruments that have artificially high correlations with the
endogenous variables.  You can also do OverID tests if you construct
new excluded instruments by taking products and squares of your
existing excluded instruments, or products with exogenous variables.
See these 3 working papers:
for more.

I am copying the Statalist, which I recommend you join, if you
anticipate having more questions about -ivreg2- (Kit Baum and Mark
Schaffer are regular contributors).  You should be aware that all
Statalist posts are available on the web, including originating email
addresses, which can dramatically increase your volume of spam if you
are unprotected.  A free address is a good solution to this

On Jan 9, 2008 2:09 PM, <snip> wrote:
> Dear Austin,
> I wish to test for weak instruments using Stata and found a presentation of
> yours on Stata's website, and thought I might email you to get some advice
> from you. Basically I wish to estimate a regression of the type
> y = c + az + bx + e
> where z are two endogenous variables I need to instrument, and x is a set of
> exogenous variables. My model is just identified as I only consider two
> excluded instruments to instrument z.
> I just uploaded the new version of ivreg2 which contains much more
> information than in the previous version, such as the Stock and Yogo
> critical values for the Cragg-Donald statistic, which makes things simpler
> as I was not sure whether I should use Table 1 or 2 in Stock and Yogo for my
> critical values.
> From what I understand, if the Cragg-Donald statistic is larger than the
> critical values, then I can reject the null hypothesis of weak instruments.
> My calculated statistic is always huge, and depending on specifications it
> is even larger than 100. Do you think this is plausible and makes sense?
> I also understand that this statistic, as well as the Stock and Yogo
> critical values are valid only under homoskedasticity, i.e. when I do not
> include the robust command. However, in my regressions I wanted to control
> for heteroskedasticity by adding the robust command. In that case, ivreg2
> instead reports the Kleibergen-Paap statistic. Is this the equivalent to the
> Cragg-Donald statistic but under heteroskedasticity? I am asking this as I
> would prefer to control for heteroskedasticity but the Cragg-Donald does not
> allow me to do so.
> Apart from that I am also using the ivendog command to obtain the Wu-Hausman
> and Durbin-Wu-Hausman statistics to test for endogeneity. I am not sure
> whether I should also test for other things, especially as I am a bit
> concerned by the huge value of the Cragg-Donald statistic I obtain.
> Many thanks for your suggestions!!!
> Best wishes,
> Alice.
*   For searches and help try:

© Copyright 1996–2024 StataCorp LLC   |   Terms of use   |   Privacy   |   Contact us   |   What's new   |   Site index